Explore every episode of the podcast Latent Space: The AI Engineer Podcast

A few years ago I stumbled upon Leash Bio’s Substack, and was hooked. Be it their Kaggle competition that showed that no, we really haven’t solved protein-ligand binding, or their crazy discovery that knowing a molecule’s author is perhaps all you need to know to predict how good it is. Their substack — or at least Owl Posting’s deep dive on them — should be required reading for any scientist trying to do chemistry with AI.

So I was super excited to finally be able to sit down with Ian Quigley from Leash Bio, and pick his brain. Leash specializes in DNA Encoded Libraries (aka “DELs”), a way of collecting ridiculously high-throughput interaction data between proteins and small molecules. Before I discovered Leash, I had heard nothing but horror stories from DEL users. And yet, Ian and friends somehow found a way to tame the beast, extracting clean and reproducible data that is just perfect for large scale modeling. Without teasing too much, the answer comes down to epistemic rigor, and an incredible conviction that you just have to Do The Thing.

It gets even better! Near the end of this episode, Ian teases something his team was just playing with when we recorded: an AI playground for agents to do verifiable scientific reasoning. They discovered the data they had spent the last few years collecting could provide tens of thousands of “lab-in-the-loop” simulations, just what an agent needs for scientific RL. Two months later, it’s matured into Minos, and today they’re releasing a full preprint, blog, and open dataset. We’ll be releasing a lightning podcast in a few days with an update. Stay tuned!

We had a blast with Ian, probably one of our most fun episodes to date. For both complete chemistry novices and expert AI x Bio practitioners, we think you’ll get something out of this.

Enjoy!

We covered:

* Why DELs traumatize everyone who touches them, and how Leash puts dozens of barcodes on each molecule to catch proteins that bind the DNA instead of the drug.

* Ian’s rule that biology only scales two ways: taking pictures and sequencing DNA.

* The Kaggle competition: about 4B measurements from a basement and about 2,000 teams, with nobody beating chance on an unseen chemical library. Also why Leash quietly feared co-folding would make them obsolete.

* Clever Hans in ChEMBL: models can guess a molecule’s chemist from its structure, and “who made it” predicts activity about as well as the structure does.

* Hermes, trained only on Leash’s own data, keeping pace with Boltz on public benchmarks. Plus the BRD4 molecule Boltz had already seen in the PDB.

* Breaking the pocket: engineering protein mutants at scale after catching Hermes memorizing “nicknames” for proteins instead of binding sites.

* Lumacaftor-like hits as an accidental detector for misfolded proteins.

* The agent sandbox: LLMs that reason about SAR in plain English, hunting for the best binders among 6.5M molecules on a budget of 100.

* The Recursion mafia, and why its alumni companies all rhyme.

* The Kosuri paradox: if you can collect the data, why do you need the model?

* Going from a DEL hit to good mouse PKPD in six months on an inflammation program, and the case for pills over antibodies on cytokine targets.

* What stops a better-funded lab from outscaling them, and Ian’s by-fiat wish for massively parallel purification.

Watch along AIE NYC and join us at AIE CODE in SF next month!

This episode includes a live demo of the Taku robot that is best enjoyed on YouTube.

Last month, Dyna made waves by announcing Dyna-2, a world-action model pre-trained on one million hours of human video, claiming several new scaling laws on human data across four orders of magnitude. They have now followed up the model with a physical robot agent, Dyna-2.1, that has to be seen to be believed:

In his talk at AIE this year, Jason talked about how the real challenge for robotics in the real world is reliable, long-horizon autonomy, and how Dyna got its models to a 99.4% success rate folding restaurant napkins for 24 hours straight, including recovering after it pulled over the whole stack.

From training a world-action model on over one million hours of human video to deploying robots that fold thousands of napkins daily at Din Tai Fung, Dyna Robotics is pushing physical AI beyond impressive demos and into real-world production. In this episode, Dyna Robotics co-founder and Chief Scientist Jason Ma joins swyx on-site to unpack how robots can learn new tasks from as little as 45 minutes of demonstrations, why predicting the future of the physical world may be the key to general-purpose robotics, and what it actually takes to build robots that work reliably for real customers.

We go deep on Dyna-2, Dyna’s million-hour world-action model that combines future video prediction with robot actions, the scaling laws behind human-to-robot transfer, diffusion transformers and flow matching, and why synthetic environments and self-generated training data could unlock the next wave of robotics progress. Jason also explains why Dyna builds its own hardware instead of relying on off-the-shelf humanoids, how agentic harnesses enable robots to execute hour-long workflows, why dexterous hands and missing sensory data remain major bottlenecks, and what separates real commercial deployments from the graveyard of robotics startups. We also cover world-action models versus vision-language-action models, GPT-6 and coding agents in robotics, and Jason’s journey from AlphaGo and Penn to NVIDIA, Google DeepMind, and founding Dyna.

We discuss:

* Why Dyna-2 was trained on over one million hours of human video

* World-action models: predicting future video to guide physical robot actions

* Diffusion transformers, flow matching, and real-time robotic control

* How robots learn new tasks with as little as 45 minutes of demonstrations

* Why robotics has no equivalent of the internet and data remains its biggest scaling challenge

* Human egocentric video, robot demonstrations, and cross-embodiment transfer

* Synthetic environments, simulation, and models that generate their own training data

* Why robots need on-device inference, model distillation, and custom hardware

* Why Dyna chose wheeled semi-humanoids instead of traditional bipedal robots

* The hardware bottleneck: why dexterous robot hands still overheat

* Best Western, Din Tai Fung, and the difference between robotics pilots and scaled deployments

* The Bitter Lesson of Robotics: scaling data instead of engineering specialized solutions

* Agentic harnesses and robots executing multi-step, hour-long physical workflows

* GPT-6, coding agents, and AI-generated simulation environments for robotics research

* World-action models vs. VLAs, and why vision alone may not be enough

* Jason’s journey through AlphaGo, Penn, NVIDIA, Google DeepMind, and the Eureka project

* Dyna’s research culture, deployment-first philosophy, and vision for millions of robots

Jason Ma

* LinkedIn: https://www.linkedin.com/in/jason-ma-742224a2/

* X: https://x.com/JasonMa2020

Dyna Robotics

* Website: https://www.dyna.co/

* X: https://x.com/DynaRobotics

Timestamps

00:00:00 Introduction and Dyna-2’s World-Action Model

00:03:08 Diffusion Transformers, Flow Matching, and Generalist Robots

00:08:19 The Robotics Data Bottleneck

00:13:50 Synthetic Data and Self-Improving World Models

00:16:51 Real-Time Inference and Model Distillation

00:19:45 Full-Stack Robotics: Hardware and Supply Chains

00:25:02 Why Dyna Built a Wheeled Semi-Humanoid

00:29:35 Teaching Robots New Tasks and Evaluating Performance

00:31:59 Autonomous Workflows and Hardware Limitations

00:36:06 Real-World Deployments at Best Western and Din Tai Fung

00:41:07 The Bitter Lesson of Robotics

00:46:13 Agentic Harnesses for Physical AI

00:49:07 GPT-6, Coding Agents, and AI-Native Robotics

00:52:02 World-Action Models vs. VLAs

00:59:17 From AlphaGo to Founding Dyna Robotics

01:04:20 Dyna’s Research Culture and the Future of Robotics

Transcript

Introduction: Jason Ma and Dyna Robotics

Swyx [00:00:00]: All right, we’re here on site with Jason Ma, the co-founder, chief scientist of Dyna Robotics. Thanks for having us in your office.

Jason Ma [00:00:06]: Of course. Thank you for coming.

Vibhu [00:00:07]: It’s been a blast. We got to see—you’ll see plenty of B-roll—drinks being pulled, a lot of evals, robots with knives for the first time. So under the hood of all this you have a world action model. What is that?

Dyna-2: A World Action Model That Predicts the Future

Jason Ma [00:00:22]: So last month, we released our world action model called Dyna-2. So it’s the first world action model trained on more than one million hours of data. So going back to your question first, world action model is a model that’s able to predict the future video, the future state of the world, so hence the name world model. But in addition, you can also use that future generation to also produce actions robots can execute in the physical world to control the robot. So why do we do that? It’s a very powerful architecture, because if you think about how humans act in the physical world, in our brain, we are implicitly thinking about how the world will evolve based on our actions. So now we have a model architecture that’s able to do that to guide robot actions. So what we find is that by doing so, the model is much more dexterous, much more robust, and much more capable of generalizing to environments, to scenarios it’s never seen before.

World Action Models: Inputs, Outputs, and Visual Reasoning

Vibhu [00:01:17]: I wanna prompt this question a bit. I think a lot of people hear world model, and they think you’re outputting video or world or 3D state. You’re also doing action. I think most people don’t understand what models are under the hood of robotics. So what are the inputs and outputs? What’s the training data? What is the architecture? Give us all the details.

Jason Ma [00:01:39]: So, for inputs, the most important input is video, right? You can think about videos that the robot is capturing with its cameras, right? So video or images of the past many seconds, that’s the input. But for robots, what’s really interesting is that robot also has a lot of actuators and joint encoders on itself. So the robot also knows, the angles of its joints. So those can also be inputs to this world model. That’s something that’s very different from just a pure video model that’s only looking at videos, right? So that’s the input. And additionally, for robots, you want to be able to talk to it, to instruct it to do different tasks, right? So you can also feed it language instruction. So those are what I consider the basic set of inputs that go into any robotics models, including world action models. And once we have that the output, at the basic level, it needs to output actions the robots can execute on the hardware. But in the case of world action model, what’s really interesting is that in addition to the actions, it can also output future video, right? So basically, you can almost output the future video or some representation of it first, and then use that to guide robot actions. So the way I think about it is this is kind of analogous to chain of thought, you know. And basically, the models are reasoning in the visual, the physics space, and then outputting actions. So that’s the key difference between world action models outputs versus other kind of models, robot models outputs.

Diffusion Transformers, Flow Matching, and Smooth Actions

Vibhu [00:03:08]: You put out a good blog post around your model. I think a lot of people understand language models or transformers, next-token predictors. In this case, you shared quite a bit. There’s diffusion, flow matching, you’re predicting next world state plus action. One thing you say is a million hours. How do you determine the architecture? Like, what are the choices there? What is the architecture? I feel like in language models, there’s a lot of take something off the shelf- Take trillions of tokens, do prediction, SFT, RL, and you have a model. What about this? How do you go from nothing to your model? Is there vanilla architecture off the shelf? Do you determine all this and what’s, what is the current state?

Jason Ma [00:03:49]: Yeah, I think, yeah. So the mental framework to think about how do we build models for robotics is that there are essentially a couple things to look at. One is data, and the second is how do we make a model learn as much as possible from that data? So let’s first talk about the model actually, right? So the blog post we released is our Dyna-2 world action model. So under the hood is a diffusion transformer. The model uses an encoder-only transformer architecture.

Vibhu [00:04:20]: Right, yeah.

Jason Ma [00:04:20]: But what’s really interesting is that in visual domains, right, if you think about generating images, generating videos, and in this case, robots looking at images to generate actions, people have found the diffusion kind of architecture to work really well. So what is a diffusion model? A diffusion model is a model that takes some random noise and through iterative computation can go from noise to something that’s really high-dimensional, that’s something really highly defined and have very fine-grained feature. If you think about generating images, right, you want sharp edges, sharp corners, color contrast. And for robotics, even though you’re not generating image, you’re generating actions that needs to be very precise, very smooth, right? So it has kind of the same kind of consideration and desiderata as generating image. So that’s why from first principles, the whole field, not just us, have been using this kind of diffusion transformer architecture to train robotics foundation models. You need the transformer aspect because transformer is the thing that can just ingest the most amount of data, can learn. Then diffusion helps you to refine the outcome to get to the smooth, precise, fine-grained motion you need to actually execute on the robot. So that’s the architecture that goes behind a lot of our recent breakthroughs.

Swyx [00:05:37]: And we’ve covered flow matching before on the podcast. It is mostly an efficiency speed-up. Like, let’s say you normally would take forty to a hundred diffusion steps. You can reduce it to four.

Jason Ma [00:05:48]: Yeah, that’s right. So with the diffusion model, one thing that’s very hard in practice is that you need to tune the scheduler, like by how much you’re tuning or converting the initial noise input into the final output. So flow matching is kind of like a special case of diffusion that makes the transformation much simpler. In practice, you don’t have to tune as much, and it works just as well. That is why I think about two to three years ago, most generative modeling for image and video has shifted from a diffusion model to a flow matching model. But the concept is the same, and flow matching is just a special, simpler to tune, variant of diffusion. Yeah.

Swyx [00:06:26]: It actually makes me think that. So flow matching obviously has trade-offs in quality. But actually, it’s very suited for action diffusion and robotics because you get to reevaluate in real time. Whereas in images, you are— it’s basically one shot.

Jason Ma [00:06:40]: That’s right.

Pretraining a Generalist Robot and Adapting to New Tasks

Vibhu [00:06:41]: Which is pretty fun. We got to mess with the laundry folding robot … throw a random towel in. The models adapt, right? I think one key thing about your models is that they’re tr— they’re made to be generalists, so one foundation model and you adapt them a lot. How does that work? Basically, in 45 minutes of mapped data, you can generalize towards. We saw an example of pulling drinks out of a fridge. What’s the what’s the mindset there of one general model, and then how do you adapt to different tasks?

Jason Ma [00:07:14]: So the way we think about it is that for robotics, we need to build a very general, strong base model, right? And then— Because if you think about how a human interacts and behaves in the physical world, right, we have very strong base knowledge. We can just readily do a lot of tasks. But if you want to become an expert at any given task, there is some amount of training that needs to happen. And I think to train a general-purpose robot like the one we’re building here, the same principle has to apply. So our approach is to first train a general model that’s training on as much data as we can find. So that is why in our Dyna-2 report, our model is trained over a million hours of human data as a starting point. And then given that strong foundation, what we found is we can quickly make the robot or make the model do many different tasks with a very minimal amount of fine-tuning to just tell the model, “Here’s the task I want it to do, and here’s very few amount of demonstrations,” and the model can readily do bunch of the tasks that you have seen in office. Yeah.

A Million Hours of Data: Egocentric Video and Robot Demonstrations

Swyx [00:08:19]: We actually have quite a few people who deal in data, who sell data. What is good data? What is bad data? What do you want to buy? I think that these are all open questions. Like, obviously, I think the top, important thing is egocentric data of humans doing tasks. What else would you modify on that basic statement?

Jason Ma [00:08:42]: So I think data is the most important thing in robotics today, because unlike language model, there is no internet, right? So every company that’s building model, building hardware has to think about data. So there’s, roughly couple of different ways of operating here. One is that you can collect a lot of data yourself, which is what we are doing as well.

Swyx [00:09:02]: You have employees here doing that.

Jason Ma [00:09:04]: There are robots everywhere in the office, right? But if you think about how much data you need for robotic foundation model, I think the amount of data any individual company can collect by themselves is just nowhere near sufficient, right? It’s kind of like, you’re trying to build an internet by just having Reddit or just having Facebook. They’re great websites, a lot of users, but it’s not enough, right? So when we think about data, there has to be some sort of data acquisition happening, right? You buy data, you look at data that’s already out there. And then there’s many different tiers, right? You mentioned data quality. I think data quality in robotics is a non-stationary problem in the sense that what is high-quality data for your model changes over time. Like, when your model doesn’t know anything, like when it’s zero, any data is useful, right? You can use a lot of video data. In the case of egocentric data, I think it’s actually a very good way of pre-training robots because human data is so diverse. Like, it covers every single physical activity you can possibly imagine. And that gives robot, the model such a strong prior on how things, how the world can evolve around a embodied actor, human. But the issue with human data is that until your robot looks exactly like a human, there’s gonna be some gap between human actions and the robot actions, right? So just by training on human data, today, it’s not very likely you can just execute robot actions directly. So you still need some amount of adaptation using the robot’s own data, right? But the robot’s own data collected on a robot is gonna be much more expensive and much harder to collect. So the kind of recipe we’re thinking about is like, how do we take the largest source of data we have today, like egocentric data, to give the model a strong prior over everything physical manipulation can cover, and then combine that with a small amount of robotics data to make the robot master whatever task we’re interested in. So data quality and getting data in robotics is really, I think it’s a fascinating, multifaceted problem. But the reality is that there’s not a single winning strategy like, “Oh, here’s one kind of data. I’m only gonna collect on that,”

Jason Ma [00:11:25]: Which will give you a strong foundation model. You really need to think about, “This human data covers this part of robot cognition I need. Robot data covers low-level reflex dexterity the robot also needs,” and put them into one model. And I think that’s what gives a really adaptable, really strong model at the end. Yeah.

Data Partnerships and Specialized Training Examples

Swyx [00:11:44]: I’ll prompt you, just for again, for the people listening who want to work with you. What, d— what data vendors do you want. This is your call for data. What-

Jason Ma [00:11:55]: Yeah, no, that’s a great question. And yeah, so in our Dyna-2 blog, at the end, we actually have an email address called datapartnerships@dyna.co. So if you’re a data vendor that has a lot of high-quality, let it be egocentric data, robot data, video data, or anything you think could help robots, send it to that email. Our team will take a look, and we’ll get back to you.

Swyx [00:12:16]: But at your stage, are you. You’re probably not buying any more egocentric data off the shelf, right? That you probably want to be like, “Here’s what I need. You go-”

Vibhu [00:12:26]: Is there a niche that’s lacking?

Swyx [00:12:28]: You need a Scale AI, you need a Mercor for doing this. I see. The custom data-

Vibhu [00:12:32]: The folks that don’t work in robotics are probably thinking like, “Okay, some niche factory.” Like … you showed putting tubes in. There’s only so much data you’ll find about that but like would a lot of that help you?

Jason Ma [00:12:43]: Yeah, no. If you really want to nail a task, having data from that task will certainly help you a lot, right? So for us, it’s like we still think, just a million hours of data is not enough. It’s just the beginning. Like, we showed the first model on a million-hour scale, but we need more than that right? So we are still looking at a lot of just raw, diverse data, that we can get. But we’re also thinking about how do we get more custom data either collected in-house or by our partner for really, specific niche tasks? Because I think you still need that to get to excellence that any model needs to actually be useful in the physical world, right? It’s kind of like humans, you know. If, Sean or me, we go work at Rosewood Hotel tomorrow, we are not gonna be able to make the bed the way they want, right? We need a lot of training. So that kind of data, like special data, that’s kind of like what Mercor is doing today for language model, I think will be very valuable for robotics as well.

Vibhu [00:13:42]: So hotels, strap a GoPro on your employees’ heads and.

Jason Ma [00:13:48]: And then there’s probably more of that.

Vibhu [00:13:49]: There’s the

Synthetic Data, Simulation, and the Eureka Project

Swyx [00:13:50]: Put themselves in their way. Yeah. So one thing that we discussed is the closing this interview, and what’s very interesting is now you can basically fabricate environments on the fly. Have you tried that with the frontier models? What’s your experience? Are they good enough?

Vibhu [00:14:06]: Are you using any world model robotic. Like, world model companies to generate these and train on them?

Jason Ma [00:14:13]: So, I think generating synthetic environments, synthetic data, simulation data for robotics, it is going to be one of the next frontiers, right? ‘Cause today people have looked at egocentric data, robot data. Those are scaling, but I think synthetic data will be a new dimension of scaling. So going back to your question yeah, we have been exploring using frontier models to generate synthetic, environments for us on our data, on our environments that we care about. And surprisingly, I think now it’s at a point where this technology is actually very much worth going really deep into. So for some context, so, in my PhD, I primarily focused on simulation-based training for robotics. Back then, simulation wasn’t ready, but we had this paper called Eureka, which was the first paper taking a large language model. At the time, it’s GPT-4, which is how to self-generate environment, generate reward functions or reinforcement learning, and just do things in simulation. And at the end, I think, we had this really nice demo of the robot spinning pencils, right? That was really cool at the time. But that kind of thing was still very limited at the time, because back then the model didn’t really have good vision capability, didn’t have a innate 3D understanding, right? So it was basically just generating code and using very limited feedback from text to try to come up with new simulation environments. But today the vision capability of these frontier models are crazy. They can do real to sim, do CADing, there’s 3D reconstruction and all of that. So I think now it’s at a tipping point where

Jason Ma [00:15:52]: This technology could really accelerate robotics. And for us, right, we think about that as a way to accelerate our internal progress, right? Because we’re building models, we’re building hardware. So if a frontier model can give us more data that way, like generate really tailored data for ourselves, that’s extremely useful for us. And another thing we’ve been thinking about is instead of using these models to generate 3D environments, we have also used our own world model, right, the world action model I was telling you about, because the model not only generates action, it can also generate videos. We— The model can also generate its own training data. So that becomes a self-fulfilling flywheel that I think will also be powerful. So overall, I’m just really excited about this direction because I think even though we have scaled to a million hours of data, we are deploying a bunch of robots. The robots are act— already starting to do very cool things. We’re just at the beginning, just the tip of iceberg for what is there to come. Yeah.

On-Device Inference, Latency, and Model Distillation

Vibhu [00:16:51]: For those on the research side, we recently had Runway on the podcast. One of their things is robotic world models. They’re, shifting a little towards there, so … nudge people to look at it. One of the questions I have about you making your own models is what is the constraints you operate in? So if you draw a parallel to language models, you see GPT-OSS 20B is very fit for your average MacBook. Then there’s a larger one that’s fit for a single GPU. Do you work in the constraints of robot, of model on on device, in cloud, a hybrid? And then how does that affect your scaling law of data? A million hours is great, but like do we just scale up parameters, data?

Jason Ma [00:17:38]: No, that’s a great question. So I think what’s really different about robots versus language model and other applications is that for coding agents, language model, it’s okay if it’s— it has latency, right? Well today, whenever I ask ChatGPT, it takes like couple of seconds before it spits out anything. But for robots the inference latency has to be very short, right? The robot needs to act. Like, it can’t just hold a knife and not do anything for three seconds. That’s gonna freak people out, right?

Vibhu [00:18:07]: Yeah, we saw someone put their real hand under a knife.

Jason Ma [00:18:10]: Our model’s, pretty smart, so that’s okay. But in general if you think about very dynamic tasks, like for example, you want your robot to hold a I don’t know, a tennis racket and juggle the ball, right? It’s just very dynamic, real reactive. You can’t just like. The physics, gravity doesn’t wait for you, right? So for that reason, robotics model needs to be able to inference very fast. And the fastest way to do that is by serving on-prem, like on device, right? Because if you call a model-

Swyx [00:18:39]: Network latency.

Jason Ma [00:18:40]: There’s no guarantee-

Vibhu [00:18:41]: Self-drive a car.

Jason Ma [00:18:42]: There’s no guarantee, right? Like, so for that reason, the model that gets deployed eventually has to be of reasonable size that can be served locally. But that doesn’t mean that the initial model you train also has to be small, right? Like in language model-

Vibhu [00:18:57]: You can distill.

Jason Ma [00:18:58]: People do distillation, right? Like, the best way to make a small model very smart is by doing distillation, right? You train the biggest model, then you distill, right? So I think in robotics, we’re getting there, but I think that’s what people are going to converge on. Yeah.

Vibhu [00:19:14]: How about inference? You have your own inference stack or. A lot of this is architecture. You’ve pulled diffusion and whatnot. I’m guessing you can pull off of something. Anything you wanna share?

Jason Ma [00:19:26]: So we build our own inference stack for whatever model we deploy, because that’s— like I was saying latency is so important. So owning the whole stack is extremely important for robotics, right? You want to optimize the inference speed for the kind of model you’re running, and there’s a lot of work in there. So that’s what we do.

Building the Full Robotics Stack: Sensors and Actuators

Swyx [00:19:45]: You’re really full stack, like- Everything from the actuators all the way to in the industrial design ID, which I- Didn’t really know as a acronym. Obviously to the models and data as well.

Vibhu [00:19:57]: I’m curious are these your own actuators, or are they. How much of this is your own? Like, of course, the design, it’s not like humanoid- Actuators, like how far down the robotics stack?

Jason Ma [00:20:08]: So the way we think about, what we need to own versus what we can buy is like if the Input-output and of a component is very clear, and like there’s already extremely good mature solution, then there’s no point of building in-house. For example, on this robot, you see the LiDARs, right? So those, black circle, those are LiDARs to help the robot detect obstacle, navigate surroundings. So that’s a very mature technology that people have been using for now decades. It’s on self-driving car. So we don’t need to build LiDAR from scratch, right? Same goes with some cameras. But for anything that we think is extremely crucial for the whole system’s performance at the end, our vision is to be able to own that and build that ourselves. So one thing is, you mentioned actuator. So that’s something we’re working on and we’re going to own ourselves. The reason is that like, the quality and how you design the actuator really matters for how well you can model the robot, right? Actuator is what drives the robot. To be able to model your robot very well in simulation or to do any kind of modeling, you need to select actuators very good that can actually be easily modelable. So that’s what we care about eventually. So we’re gonna own that. Same thing with the robot’s proportion. This robot’s proportion is we measure it against real people. Why do we do that? ‘Cause we’re pre-training on human data.

Swyx [00:21:36]: Oh, yeah.

Jason Ma [00:21:36]: Best transfer, right? So there’s a lot of these kind of decision that goes into hardware that’s very much tied to like. Because eventually, it’s hardware that carries intelligence. Intelligence is the extremely important thing. We want to make sure the hardware serves for that.

Swyx [00:21:53]: I would just highlight, I don’t know if this is relevant to you, but one of the most interesting robotics discussions I’ve had is someone who was very stressed out about actuators and specifically magnets, ‘cause apparently we don’t make magnets in the US.

Manufacturing, Rare-Earth Magnets, and Global Supply Chains

Jason Ma [00:22:06]: No, we don’t.

Swyx [00:22:07]: And there’s only two countries that make magnets.

Jason Ma [00:22:09]: Rare earth material.

Swyx [00:22:10]: Rare earth, right? I don’t know what the. Wh-what’s the chemi— what’s the material? I don’t know what the thing is, but- so yeah, if anyone’s interested in that’s, apparently a bottleneck that we have.

Jason Ma [00:22:20]: No, not bottleneck.

Swyx [00:22:21]: It’s a bottleneck. So what do we do?

Jason Ma [00:22:23]: It’s a good question, right? I think, there’s a lot of exciting discussions on bringing manufacturing back to the US, but for robots especially, that’s actually much harder than what it seems. And I would say even now, there isn’t a company in the US that’s truly building everything from scratch in the US, right? Like, even when they say they assemble robot, they build whatever they’re actually doing in the US, they’re still sourcing their raw materials from other places in the world. And I think that’s something to really be aware of as a robot builder, right? Like, I think, like as a full-stack robotics company, we realized just how long the supply chain is and how long the whole problem is. It’s actually extremely hard. And I think these days, a lot of the discussion centers around a robot foundation model data, but I think hardware and just how hard it is to build good hardware is very under-emphasized. But I think that’s gonna be extremely important.

Swyx [00:23:27]: This is a platform, right?

Why Dyna Co-Designs Its Hardware and Models

Jason Ma [00:23:29]: Platform, yeah.

Swyx [00:23:30]: There are some companies that choose to be, let’s call it platform-less. They only do the model, and they will say, “We deploy on any platform.” We deploy on Optimus.

Vibhu [00:23:37]: Off-the-shelf robotics.

Swyx [00:23:38]: And so, it’s a choice, right? Like, obviously, you own everything, so you can do that. Do you see yourself building more in the US, or do you see building your own global supply chain? What’s the direction that is likely?

Jason Ma [00:23:52]: So we are currently exploring many different options, right? Because with global supply chain, then there is some, geopolitical things, right?

Swyx [00:24:02]: Shipping.

Jason Ma [00:24:02]: Shipping. Also shipping delays and all that stuff.

Vibhu [00:24:04]: Literal ships for robots.

Jason Ma [00:24:06]: And in the US, it’s just that going from zero to one will take some time, right? So we’re trying to think about a solution that can both allow us to build more in the US while leveraging global supply chain, right? But I do think, going back to your previous question on some companies choose to just build model, we build everything, I think it’s very important to get this right. I feel that for robotics, the hard thing is. Like building the whole system. What are the necessary things to build a whole system? Like, previously, I was at DeepMind, I was at NVIDIA, primarily just building models, and I saw the upper bound of the system when you don’t control the hardware, when you don’t get to optimize that right? Because if you think about the value creation for robots, eventually robots are here to deliver a service to solve some physical tasks. So if you wanna solve physical tasks, you wanna be able to optimize everything, and that’s important.

The Semi-Humanoid Robot: Wheels, Torso Movement, and Dexterity

Vibhu [00:25:02]: We have the robot. You gotta tell us about the design. Your own custom platform. What do you call this?

Jason Ma [00:25:11]: What do we call it?

Vibhu [00:25:12]: Yeah, so it’s not humanoid. It’s very smooth, but what’s the what’s the rotating arm?

Jason Ma [00:25:16]: Yeah, so we think of this as generally like a semi-humanoid. So basically, the upper body is, like a humanoid. It’s got, of course, shoulder, got arms. It also got four degrees of freedom on the waist. So you can basically do all the upper body motion that a human can do. But it’s a semi-humanoid because the lower body is wheels. All right, so why do we choose wheels? Because I think there is a difference. There’s, optimal practicality versus performance trade-off. I think with legs, the energy con— the energy consumption would be a lot higher. It would be a lot more unstable. Safety will be extremely hard to guarantee. So right now, for legged robots, there’s actually no guarantee that the robot will stay upright, balanced the entire time. I’m sure you guys have seen a lot of those, robot dancing videos. Like the robot dancing thing.

Swyx [00:26:08]: Seeing dancing, yeah. They all, they fall over very easily.

Vibhu [00:26:10]: No, but you also see the demos that you can tell are not production, where you can box the robot and you can push it over and it’s very good. But that robot is just good at balance, right? Like- there’s a.

Swyx [00:26:21]: I think the more interesting thing that you guys have innovated. So we have, one of the prototypes, which is the Mewtwo back here.

Jason Ma [00:26:29]: You can call it that.

Swyx [00:26:30]: This is just like an up-down, axis, and you’ve chosen to then move towards something like this which, has the ability to lean. And we’re gonna show the laundry demo where you have to lean into machines. Right? That’s another degree of freedom that I don’t even have the words for it. It’s not pitch, it’s not yaw. It’s something else.

Jason Ma [00:26:47]: So yeah, how did we think about that right? Like,. So actually, it’s very used for exercise. Like, I think sometimes, a couple years ago when I was exercising, I accidentally like, pulled a muscle. Right? I’m sure you guys have that experience too, right? Then the next few days you just couldn’t move, or like you pull it where you slept in the wrong position, so your neck is very stiff. We have all experienced that. And then whenever you experience that you kind of realize, wow, so many muscles, so many degrees of freedom that we use in our daily life that we take for granted. But when it’s disabled it’s very awkward for you to do some tasks that you feel are very natural, right? So that’s also how we think about when we design robots. When you have a— this kind of design, which is this purely elevator, could only go up and down, we realize a lot of the human torso movement, like we use our waist, back a lot for day-to-day physical activity, right? Even if I wanna get this bottle I’m not just moving my arms. It’s actually very strange for me if I have to intentionally think about fixing my upper body. I would always do this right? So there’s just a lot of subconscious way that humans move their body that I think will be equally useful if robot has a degree of freedom. And if you further think about, okay, we want to learn from human data. Like, humans are moving their torso all the time. You wanna be able to ingest that into the model to control the robot. But if you have a form factor that doesn’t allow for that then the transfer becomes less and the robot becomes less dexterous and it’s just more awkward.

Transferring Human Pretraining to New Robot Hardware

Vibhu [00:28:22]: What’s the data labeling side of this? Like, I— you’ve seen people,. Is this the last post-training fine-tuning side of actually the hour of using the robot tele-operated or remote-controlled mapping motion? Or does some of it generalize over?

Jason Ma [00:28:41]: When you say labeling, you mean?

Vibhu [00:28:43]: Like, taking the model you have that predicts state to mapping it to your specific hardware.

Jason Ma [00:28:50]: Oh, yeah. It generalizes very well. So for all the demos that I have shown you on the new robot, ‘cause the new robot, we, it got ready not so long ago. So we quite literally do not have a lot of data compared to the stationary arms we showed you, which had much more data, right? But a new robot, the model has not seen this in pre-training, in some of our pre-trained model. It’s able to learn the task that you saw very quickly. And I credit a lot of that to like, even before the hardware is ready, we already committed to pre-training on humans because we knew at some point our robots will be like human. You can get ahead on the training side while hardware is getting ready. And I think now we’re seeing that the transfer is actually there and is actually very helpful.

Evaluating New Skills and Learning from Real-World Feedback

Vibhu [00:29:35]: Talk about the eval side of this. So you can have predictions, you can have it do stuff, and every now and then we see it miss stuff. You test on very niche things as well. It’s like filling a pill bottle. How long does that process take? How do you think about the aspect of actually deploying on a new task, evaling it, iterating? What does that look like?

Jason Ma [00:29:56]: Yeah, so if we wanna learn a new task, we basically do a little bit of robot data collection. Then the model, these-

Swyx [00:30:04]: What is a little bit like?

Jason Ma [00:30:05]: Yeah, like couple hours. Yeah, we’re— In some of the tasks, like just tens of minutes, right? Yeah. And then-

Swyx [00:30:11]: The chopping one was 45 minutes.

Vibhu [00:30:13]: The open the fridge was on-

Swyx [00:30:14]: And that’s an interview— That’s like an onboarding task for one of your engineers.

Jason Ma [00:30:18]: So,.

Vibhu [00:30:20]: Fun thing if you wanna work here, you know. Find the task, have the robot do it.

Jason Ma [00:30:24]: Pitch me a new task and I’ll come to the office, collect some data, train a model and see what happens.

Vibhu [00:30:29]: Tie your shoes.

Jason Ma [00:30:31]: Truly, it’s tying. Yeah, that would be cool. But yeah. So, you know. A little, a small amount of data collected, train a model, then we just deploy the model in the real world. And then we look at the performance, then-

Vibhu [00:30:41]: To fill in collected means you are— you have a person actually operating-

Jason Ma [00:30:46]: Operating.

Vibhu [00:30:48]: Doing the task for a bit. And then you train? And it does okay, you improve. How do you measure?

Jason Ma [00:30:56]: So when you’re training the model, there’s, certain metrics you can look at right? Just like training a model, it converges nicely. Now, after that for robotics today, we just run the model in the real world, right, on the same setup that the data is collected to see how the model does. Sometimes the model just works great out of the box, right? You don’t really need to do much more. But when the model’s maybe like seventy percent, eighty percent there, then you see how the model behaves. You figure out what it’s not good at. Then you can collect a bit more data, then iterate, right? I think of this as the robotics version of reinforcement learning from human feedback, right? The robot rolls out. You can score it good, bad S- somewhat good. And then that data, the robot’s own experience can be used to train the model. And you can also just collect more data yourself to train a model, right? Eventually, our goal is like, this robot is gonna be our platform. Anyone should be able to train tasks on it.

Long-Horizon Tasks and the Bottlenecks to General Robotics

Vibhu [00:31:59]: I wanna get a bit of a prediction and timeline to where that goes. Right now, the tasks are, somewhat simple, like- Take laundry out. It’s a towel. You can’t crush it. Chop some food. There’s also the side of like, okay, set up this filming studio, right? Move stuff around, bring a table, set up lights. I think-

Swyx [00:32:21]: Manage my data center. And also, ref— like, recharge another robot.

Vibhu [00:32:29]: When do we see the factory or what do we need to scale up? Is it model? Is it everything? It seems like your hardware is set for that but where do we see it go to a point where I can go from instead of just cut one thing to, “Okay, here’s a recipe.”

Jason Ma [00:32:49]: I would say to get to truly general-purpose robots like basically a robot or a model that does anything, like-

Vibhu [00:32:58]: Not that. Like, okay, so to keep it somewhat grounded on a technical level, like- when do we start to see us pushing that feedback, right? So-

Jason Ma [00:33:06]: I see.

Vibhu [00:33:06]: Instead of very simple chopping- Maybe a five to ten-minute sequence, right? Where now the eval step, it can break at many more points. Instead of just fill one pill bottle- Ten different pills all in a sequence, right? What’s limiting us? What are we pushing on?

Jason Ma [00:33:20]: Yeah, I think. Yeah, I think a couple of things. One is that the laundry demo that you have seen, yeah, it’s gonna be one hour to finish a whole loop, right? So the longest duration the robot can do a full job is already in the hour range. Right? But to be able to do that for any task, I think the bottleneck is coming back to data and hardware, right? So for some of the tasks, for example, setting up the space, you might need more than a the parallel jaw gripper, right? You might need something more dexterous, like a hand, like a human hand, right? But then if you wanna train a model on the hand, then you’re bottlenecked by the hardware stability. Like, hand— Today, the best hands, dexterous hands are still not, durable enough. They overheat pretty quickly, right? So that becomes a bottleneck, right? Like, you— We— you quite literally have to put a ice bag-

Swyx [00:34:15]: Ice. To cool.

Jason Ma [00:34:16]: Human.

Vibhu [00:34:16]: We’re not just auto-researching, make better model we have. We have hardware limitations.

Jason Ma [00:34:20]: Yeah, so hardware limitation is a big thing. But then if hardware is limited, then that will slow down everything, right? Because to know that you’re making progress on any task, you eventually have to eval on the hardware, right? So then without the eval being very spontaneous, very fast, that also limits your research progress, right? Because then how do we collect the right data, right? I think, how do we collect the right data to do extremely complex, like setting up this studio, right? Because you can imagine, like today, most of the data is vision-based, right? It’s like video streams and some, encoding of your hand motion where the robot’s, joints. But Humans have a lot more sensors. We hear things. We feel things.

Swyx [00:35:05]: Smell.

Jason Ma [00:35:06]: Yeah, you smell things, right? So we subconsciously use all of those to help us be so good at manipulating the physical world. But today if we’re not capturing that at all, then even when we scale to millions of hours, you might just be missing critical dimensions of the task, so you’ll never be able to solve the task, right? So I think those are what I consider what’s dragging on the timeline. But for tasks or for workflows that you can collect data today to high fidelity, then I don’t ne— I don’t necessarily think there is a fundamental bottleneck, right? Just like what we are demonstrating, to the public this week. We have a physical agent that can do the entire laundry workflow, right, as a human would, right? Because for that task, we don’t really have like, oh, you actually need more than a parallel jaw gripper. You actually need some sensing modality that’s just not available today, right? So I do think the biggest bottleneck To have a much more general thing is still hardware and data.

Real-World Deployments: Best Western and Din Tai Fung

Swyx [00:36:06]: I want to take this chance to talk— let you brag about your customers and partners. You have some announced, relationships. We can start with Best Western and then maybe some of the others that you’re, that you’re proud of. Cool stories, anything like that.

Jason Ma [00:36:20]: So we have a deployment blog post, on our website that goes into much more detail. I really recommend watching the video because I think the video is-

Swyx [00:36:29]: Which one? The Best Western one?

Jason Ma [00:36:30]: No, we have one video for all the customers we’re working with. It’s a really nice video. It shows what we’re doing. And yeah, I think it’s great. But yeah, we have customers, Best Western. That’s in the hotel space. We fold a lot of laundry towels for them. Every hotel needs it, every gym needs it, every fitness center needs it. -

Swyx [00:36:51]: What’s the deployment so far? How— Like, where can I go to— Which Best Western do I go to see it?

Jason Ma [00:36:56]: So if you wanna go somewhere, I actually recommend going to our other deployment, Din Tai Fung. Have you guys been to-

Vibhu [00:37:04]: Very thin wrapper.

Swyx [00:37:05]: We should have filmed there.

Jason Ma [00:37:09]: Just filmed there.

Vibhu [00:37:10]: The B-roll.

Jason Ma [00:37:11]: But in Din Tai Fung if you go to Din Tai Fung or other somewhat upscale restaurant, there’s napkins on your table, right? So our robots today fold thousands of napkins for many different locations, many different Din Tai Fung locations in the United States. So that’s actually our biggest scaled customer, right? So I wanna make a distinction here. Today, there’s a lot of robotics companies that’s talking about deployments, talking about pilots. But what’s really hard is not putting your robot in one location or doing a pilot. What’s really hard is that the customers believe or validate that your robot creates enough value that they wanna deploy your robots across many of their sites. So that’s what’s happened for us with Din Tai Fung, which, to the best of my knowledge, hasn’t really happened for other companies doing similar general purpose robots. So at Din Tai Fung, our robots fold thousands of napkins per day. It’s — The quality standard is extremely high, ‘cause they want like, okay, the second fold or the seam of the napkin is only like a couple centimeters below the I don’t even know how you call it, the pocket, so you can stick the chopsticks in. So it’s actually very fine-grained, high-precision task. I can show you.

Vibhu [00:38:26]: At some level, that’s what robotics are good for robots are good for too, right? Humans are not the most consistent- Especially in small details, whereas a robot can somewhat do it perfectly. It’s interesting that’s the example you bring for general purpose robotics versus, at some level, I think people would assume- you can have a specialized robot that will fold napkins perfectly to the exact, millimeter as opposed to something like this but.

Jason Ma [00:38:53]: Yeah, I’m sure, maybe you can, right? But the hard thing about manipulating deformable object, right, like napkin, towels, everything’s- Deformable. There’s just infinite- infinitely many possible configuration. So that’s really hard for a specialized robot, right? And what we find is that by developing a general purpose model in a robot and adapting it to these use cases, you can actually get to a production quality faster. And what’s really nice is that now that Din Tai Fung has scaled, has signed a much bigger contract with us, they not only want our robots to fold napkins. Now they want our robots cut. That’s why we had the- Example, to cut food to the-

Swyx [00:39:31]: Is that actually the second use case?

Jason Ma [00:39:33]: That’s one of the use case they are thinking. Din Tai Fung needs to cut so many vegetables per day.

Swyx [00:39:38]: They need to do the dumplings.

Jason Ma [00:39:39]: Exactly. To the fold. That’s the final boss. That’s-

Vibhu [00:39:41]: That’s the final one.

Swyx [00:39:43]: I think you can. I think you have the dexterity and the precision to do this.

Vibhu [00:39:46]: You have a little robot.

Swyx [00:39:47]: But no, don’t break the dumplings.

Vibhu [00:39:50]: That’s the obvious benefit is you can adapt them anymore.

Generalist Robotics, Speed, and Consistency at Scale

Vibhu [00:39:54]: I think the other thing you really touch on is the speed, right? To make a custom napkin folding with the exact ply, whatever- you have to make the machine. You can’t just Amazon this. But this can do it pretty fast, right?

Jason Ma [00:40:07]: Yeah, this can do it ex— very fast. And now it’s actually al— it’s at human parity. And this robot doesn’t get tired, doesn’t need to you know. You can. And I think I really like the fact that you mentioned consistency, right? Like, one issue with Din Tai Fung is like, their LA location with New York location, Seattle location, the they fold napkins slightly differently, but from their perspective, they want consistency. But that’s exactly what these models are good at, ‘cause models are, just, it’s a set of weights. You can just put the weights in the different robots across different-

Vibhu [00:40:43]: You can OTA update them as well.

Jason Ma [00:40:45]: Yeah, you can just update all of them at once. It just. Once it works, I think it’s truly, like- Incredible.

Vibhu [00:40:52]: Final boss is Cheesecake Factory. If the story there, they’re from scratch. Everything is very consistent, down to the measured gram. Every sauce is made. Like, anywhere you go in the world, you’ll have the exact same thing. If you don’t know, you don’t know. But that’s the that’s the final.

The Bitter Lesson of Robotics: Scale the Right Data

Swyx [00:41:07]: The term that comes to mind is kind of like bitter lesson for robotics, right? You, instead of the specialized robots, which is a lot of industrial engineering, you build a generalized platform that can- Do a last-mile fine-tuning and-

Vibhu [00:41:18]: That’s right.

Swyx [00:41:19]: And do all those things.

Vibhu [00:41:20]: Give the expert add-on outside of your. What is more bitter lesson robotics?

Jason Ma [00:41:25]: Yeah, I think there’s several bitter lessons for robotics. One is that robotics is going to scale just like any other domains that machine learning has solved, right? So what does that mean? That means having as much data as possible. But in robotics, the one thing that’s a lot harder is what is the right data is much harder to figure out than other domains, right? Language is text. In video games, just video game data. But for robotics what sensors you choose determine what kind of data you have, right? But if you have the right set of sensors and just scale that data, I think you’ll get predictable return on model performance. So in that way, robotics isn’t so different. So that’s the bitter lesson. As a corollary of that is that like there’s a lot of research in robotics that’s coming out with very specific priors. Like, for example, we need to make the robot see objects as they are, draw some boxes. That will give robot better perception, right? But I think the bitter lesson here is that just keep the algorithm simpler and scale the right kind of data. Then generalization, better behavior will emerge. So that’s not so different. And the second. I think that’s the main bitter lesson, which is at the end of the day, scaling will work, but in robotics, to scale and how to scale is just much harder ‘cause there is no internet. You have to come up with your hardware, come up with the sensors, get all of those right, and then scale. Then I think,.

Humanoids, Practical Iteration, and the Path to Generality

Vibhu [00:42:53]: There’s the argument that the world is very made for humans. It should be a human, right? As hard of a hardware problem that is, think ten, twenty, a hundred, two hundred years down the line, can we make a really efficient human? But you have to be practical with what we have now, right?

Jason Ma [00:43:12]: So the way I think about it is that yes, eventually the robot should. If we wanna build a general-purpose robot, right, like a machine that can be adapted to as many use cases as possible, I think human form factor will be the best one, just because that’s what we have been doing, for ages, right? But I think the best way to actually build a humanoid robot is to take an iterative approach. Because robotic is so hard, the whole problem, the whole stack is so long, it’s really hard to know a priori what’s the right decision to make, right? And because hardware decision is much more expensive and much more costly, just ‘cause hardware iteration is a lot slower, you actually want to take an incremental iterative approach, not going all the way to the final form factor. Because the chances are the thing you build actually doesn’t last. It’s not good enough. So you just end up scratching it anyway, you know. Like, it’s public news Optimus, already gone through many iterations, and they had to scratch their designs, right? And I think by iterating with a semi-humanoid form factor and actually putting them out to real customers or in getting models really deployed, you get that feedback loop that you wouldn’t get if you’re just trying to build like, warehouse humanoid and you spend like five, ten years building it. I think the iterative approach will actually get us closer to the final robot faster.

Vibhu [00:44:33]: And I will say, for those listening and not watching, it’s probably the smoothest robot we’ve seen. Like-

Continuous Action Spaces and the Importance of Context

Vibhu [00:44:38]: … every humanoid you’ve seen is jerky and even humans don’t walk as smooth as this thing.

Swyx [00:44:43]: Because, yeah, the model is. Can move every single motor at once.

Vibhu [00:44:48]: In more than two cohort, yeah.

Jason Ma [00:44:50]: Yeah, and that’s actually a very hard problem, because if you just think about the model, it’s outputting action. There’s the action, dimension, right? The bigger the dimension is, the harder to fit, the model to the data. But here, the action space is quite large, but we did very good job to make sure the model fits well and gives you a smooth-

Swyx [00:45:10]: Can you bound the order of magnitude of the action space, or is that-

Jason Ma [00:45:14]: Yeah, so it’s on the order of, about a hundred action dimensions.

Swyx [00:45:19]: Yeah, and these are all continuou- continuous, right?

Jason Ma [00:45:21]: Continuous, right. So the model is just basically predicting at very high frequency, like tens, like at least fifty-dimensional vector, like ten, twenty times a second to get that continuous motion.

Swyx [00:45:34]: And potentially predicting one to two seconds out, so that is another hundred frames or whatever.

Jason Ma [00:45:39]: Of course.

Vibhu [00:45:40]: There’s a bit of a history, similar to context of previous state, I guess.

Jason Ma [00:45:45]: So context is a very important thing in robotics, right? Because otherwise, if you’re just making decision based on the current observation, sometimes it’s really hard to tell what to do next, right? For example if you think about opening a microwave, like in the middle part, it’s unclear whether or not you’re-

Vibhu [00:46:01]: Closing

Jason Ma [00:46:01]: … closing or trying to open it, right? So context is also really important.

Swyx [00:46:05]: And I have a screenshot of the fridge demo where it is, there’s agentic planning and tool holding and all those things. So we can show that for listeners.

Vibhu [00:46:13]: I think that’s an interesting side. After the model-

Agentic Harnesses and Planning Physical Workflows

Vibhu [00:46:16]: … how do you wrap it around into agentic harness? Like, what is-

Swyx [00:46:22]: Claude Code

Vibhu [00:46:22]: … Claude Code deployment? Claude Code UI. You have a lot of cool UIs, but what goes into wrapping model output towards final thing? Is this custom per use case? Is it you have your own SDK?

Jason Ma [00:46:37]: So, eventually we will have our own SDK, but the way we think about is we want a general harness layer for robotics foundation model. Today, that is our own model, right? The way I think about is for any Real task, long task in the physical world, there is a lot of a high-level decision-making you have to do, right? It’s not just like robot, only reasoning about the physical world in the action, continuous trajectories, right? Because that’s a very low level, the thing you’re actually executing in the physical world, but that’s not the intent. But like if. When we are doing decision-making, we think about steps of a recipe. We think about, oh I finished this I need to do that. While I’m waiting for this part to finish, I need to go somewhere else to do something else. There’s a lot of high-level cognition, decision-making that’s happening. And I think for robots to handle even, the laundry task we are showing you, it has to have that capability. Because, it needs to reason that okay, the washer is running, so I should go do something else while the washer’s finishing. But now the washer is done, I should take, I should take the towels from the washer and put it in dryer immediately, right? Because if not, then that becomes a bottleneck in the whole workflow, right? So there is a lot of reasoning that needs to happen for robotics. And I think in theory, you could just, try to learn that with a low-level model that’s outputting actions, but it’ll be extremely inefficient. So that is why having a hierarchical system, like putting a harness, like an agentic system around our world action model is what allow the model to compose different capabilities, different skills to solve a much longer horizon task. It’s very much like in language model, right? Like two years ago, three years ago, the model was only good at autocomplete. Like, humans have to tap to get the model to keep spitting out code. But today, just an agentic system just carries the end-to-end task. And that’s. I, and I think that’s important because I think today in robotics, people are still thinking about autonomy at the very task level like one minute, thirty-second task. But really, the value is about tasks or workflows that go on unsupervised for hours. And that’s what we’re building towards. And that’s why I think thinking about. The robotics problem, not just as a modeling problem, but rather as a system problem, you need good agent, good harness engineer, good model, good hardware, everything. It’s like how we think general-purpose robots ought to be solved.

How Coding Agents Accelerate Robotics Research

Vibhu [00:49:07]: Do you guys see benefits from language model coding tools getting better? So new Opus 5.5, new Astra. Do you have language models in your agent harness? How does AGI being developed in language model space affect you guys?

Jason Ma [00:49:24]: No. Yeah, no, absolutely. Like, it affects us very deeply in many ways. One is that we are extremely AI-native company, right? Like, coding agents, all of that is a very fundamental part of our work, and we’re very welcoming of that. Like, whenever there’s a new model, we’re the first one to test it, right? Like, we being. Like, when GPT-6 came out, we had a company-wide hackathon the same week, and the thing that our engineers built was incredible. And many of them are— Many of the features we built is already in main, in production, right? So we really like using coding agents to accelerate our work.

Swyx [00:50:00]: Just for the people interested, anything surprising in GPT-6 that you found?

Jason Ma [00:50:05]: … it’s about the ability for it to construct virtual environments, right? Turns dissemination. We talked that twenty minutes ago. I think that’s really exciting.

Vibhu [00:50:17]: It’s interesting. There were so many CAD, Zillow to 3D-ish.

Swyx [00:50:21]: He lost the hackathon to the CAD, -

Vibhu [00:50:24]: I won the. I did the CAD b******t. Other way.

Swyx [00:50:27]: No. Oh.

Vibhu [00:50:28]: That was. I won a GPT hackathon just for fun. I was like, “I know nothing CAD.” Might as well play around with it ‘cause every demo is doing it. I’m like, “It’s gonna be garbage.” And I’m like, “Well, it’s kinda grounded, it’s kinda good.” But it’s just hackathon stuff.

Deployment Milestones and Competing Robotics Strategies

Vibhu [00:50:45]:.

Swyx [00:50:45]: Cool. I wa— I was gonna round out, some of the use cases and then go into a little bit of company culture and your personal background.

Vibhu [00:50:52]: I have two technical questions. One is, are there any goals you wanna set for yourself, like six months, one year, two year timeline that you see robots should be able to do, fill in the blank, that you’re kinda working towards?

Jason Ma [00:51:09]: Yeah, in the six months to one year frame, our goal is definitely scaling up our robots and our model and truly getting them to deploy-

Vibhu [00:51:18]: Like a task.

Jason Ma [00:51:18]: … completely different.

Vibhu [00:51:19]: Like, right now it’s cutting or opening a fridge. Is there. Like, do you set yourself to by this timeline, we should, to stay on track, be able to do whatever?

Jason Ma [00:51:29]: I think, in terms of deployments, we’re exploring many different industries. There isn’t a particular task we think we have to nail, right? Because our goal is to build a general model, a general platform, and we think of each different task as validation of our platform as a way, as a wedge for us to figure out how to build everything better. So today, we are doing a lot of that in laundry, in restaurants. We have already deployed tens of robots, but our goal next year is overall just deploying hundreds of robots.

Vibhu [00:52:02]: My part two on the last technical question is, how do you see the rest of the industry? So fill in for people, what other approaches are people taking to world models? Why are they wrong? What is different but good about their approaches as well? Just what is the landscape of robotic world models in this space and why— how have you come to your current?

Jason Ma [00:52:25]: No, that’s a great question. So I think robotics is the most exciting space today, and there are just so many, amazing companies that I truly respect. They’ve been doing great frontier work like us, and it’s very exciting. And I think what we are seeing today is that there’s still many different approaches to robotics. Like, we believe that hardware needs to be built in-house. We need to own the whole, chain. We believe that scaling human data is extremely important. But there are companies that’s betting on just scaling robots’ data or scaling some sort of intermediate data collection devices, right? All of them are in play, and I think it’s different approaches. We’ll just see what happens. I think right now, it’s still a bit early to tell, but we feel pretty good about the way we’re building things just because we as a company, culture-wise, mission-wise, we’re not here to just build a model. We’re here to truly get robots deployed. So we deploy as much as possible, work on the hard problems. So we identify problems much earlier on. So then I think that really helps us identify what the true bottlenecks are and then solve them. And I think for folks, for people who are only working on one part of the problem, it’s kind of like fog of war. You don’t really see the whole picture, and you might be building something that’s extremely cool technologically, but may not be what people need or want at the end, right? I think there’s a whole graveyard of robotics company, and I think all of them had very shiny, very cool robots, but just figuring out what people actually want is actually. It’s very easy to say “Build something people want.” everyone says that but it’s actually really hard to execute a robotics company.

World Action Models vs. Vision-Language-Action Models

Vibhu [00:54:07]: On the technical side, you have your world action model. Is that standard across industry? Is there other directions people are taking?

Jason Ma [00:54:15]: Yeah, I would say that’s also not standard. So we’re building world action model on top of world model. So there is many companies building world model. You mentioned Runway. They’re also getting to adapt-

Vibhu [00:54:26]: Video world models.

Jason Ma [00:54:27]: There are also companies building a vision language action model.

Vibhu [00:54:31]: VLA.

Jason Ma [00:54:31]: So that’s like taking vision language models and adapting them to robotics. So I think there’s many different ways to go about it. But I think at the end of the day what we call the models is less important than what we believe the model needs to be able to do at a first principle level. So I wanna explain that a bit. I think for robotics, the most important thing is that you need to be able to predict the future. Right. Like robotics, any physical system is fundamentally about based on the information you have at the present moment and before, what should you do next? So it’s really about predicting what comes next. So predicting robot actions is like predicting what the robot should do next. But if you think about it, robot is just a part of the world, right? Robot influences the world, but the world also influences how the robot should behave in that environment. So if you think about it this way, then the maximal, the most first principle thing you can do is just predict as much of the world that you can possibly predict, right? So that’s why we believe predicting the future is extremely important. And today, the instantiation is world action model. And VLA only predicts robot actions or some more abstract representation of the world. But whenever you’re predicting abstract representation of the world, you’re encoding your own bias into what’s important to predict. And the bigger lesson tells you, let’s remove human biases. Let’s just take the most raw representation, the camera inputs, and just predict all of that right? So that’s how I think about, modeling for robotics today.

Swyx [00:56:12]: So just to press you a little bit on this you’re, basically you’re saying your approach is better than the VLA, right? Because you’re emphasizing prediction. I fully buy that. I— Like, you’ve sold me on it. If you were wrong, what is the most likely number two-

Beyond Vision: Tactile, Audio, and Infrared Sensing

Swyx [00:56:26]: Someone else is doing?

Vibhu [00:56:27]: Interesting question.

Jason Ma [00:56:28]: So yeah, there’s a caveat in what I’ve said before, and I think that’s where I could be wrong, right? I said that you need to predict the world as much as possible, right? But what that doesn’t say is that does your sensor have everything about the world that you need to predict? I think I touched on this a little bit, right? Like, if you’re training a world model that’s purely based on videos, I think it will be a very powerful world model, but it may not be a sufficiently rich enough world model because you’re missing critical aspects of the world, like sound, like touch, right? So I think where we could be wrong is like, okay, I have the best world model on vision-based world model, but tomorrow someone else comes out with a vision plus tactile world model that’s just like a hundred X more data efficient just because, vision is a very high information or a highly redundant, input because, every other pixel, they don’t change that much. You can see— you can understand the world without perfect vision. But tactile, other modality might be much denser in information, right? So maybe having that predicting that would just allow you to use much less vision. So you can get to the same performance with much less compute, with much less data, right? But-

Swyx [00:57:51]: I think you will be able to adapt very easily.

Jason Ma [00:57:53]: No, exactly.

Vibhu [00:57:53]: There’s also a level of how much data.

Jason Ma [00:57:56]: Right. So the reason why I bring this up is because I think if a vision-tactile world model or vision-tactile-audio, whatever world model, is the final thing that solves robotics, then I think robotics companies will be able to do that better than video generation companies that’s trying to pivot to robotics. The reason is to understand sensing is actually very complicated. Like, we’ve been building that since the company started, so we know the bottleneck. We know how to combine them. But for— if you’re just trying to adapt a video model to robotics action, then if we were wrong, if, video model is not enough, then you don’t really have the know-how to get tactile to work very well. But we have that.

Swyx [00:58:39]: This is a side question, but I think it’s a quick one. Heat seems very important. Is infrared just another color to you, or is that a.

Jason Ma [00:58:47]: Yeah, it could be another-

Swyx [00:58:48]: Just another color.

Jason Ma [00:58:48]: It could be another color, another input to the model. Could be used in cooking scenario, actually. Like, I know companies that’s using infrared to know how cooked food is. Because that’s really hard to tell from RGB camera, right?

Vibhu [00:59:01]: My very naive pushback on your last point is maybe we don’t need to understand hardware, cloud code, or Codex can just, vibe code it in. There’s a lot you can do without understanding, but of course-

Swyx [00:59:15]: It’s just not in the distribution.

Jason Ma’s Journey: AlphaGo, Penn, NVIDIA, DeepMind, and Dyna

Swyx [00:59:16]: So I wanted to bring it a little bit to your personal story or your background. We both went to Penn.

Jason Ma [00:59:24]: I just found out today. Crazy.

Swyx [00:59:26]: Yes. I, my big regret is I didn’t join the self-drive, the Docker challenge. Which, and then I could have been early Waymo or something. You were at Grasp. And then you’ve also worked at Google and NVIDIA. What are the different focuses of let’s say, academia versus Google versus NVIDIA? You’ve seen all these places, right? And then that obviously caused you to say “Screw all that. I’m gonna do my own thing,” right? You could have joined OpenAI’s robotics lab. So may-just take us through chronologically, but also one thing you learned from each.

Jason Ma [00:59:58]: So chronologically, yeah, maybe it’s useful to talk about how I got into AI in the beginning, right? So, as a kid, so I spent time in Japan and China. I grew up playing Go, the Go in AlphaGo. So when AlphaGo came out, I was a senior in high school, and I was mind-blown because I’ve been playing this game for many years. It’s so complicated. Like, I think before AlphaGo came out, people were like, “Okay, Go is one of the game that AI will never solve.” The search space is larger than the total number of atoms in the world. How could you possibly search for solutions in that? Then RL, reinforcement learning just cracked it, right? And I think that was how I got into AI initially. Like, I was just really interested in models or systems, agents that can self-improve, right? So when I got to Penn as a PhD student, I just really wanted to work on reinforcement learning. So I’m co-advised. Both my advisors were interested in RL at the time. One of my advisor was more in robotics. He did a postdoc at Berkeley Robotics Lab. So that’s how I got into robots. At Penn, we have GRASP Lab, I think one of the best places to do robotics. And so that’s how I got into doing RL for robotics in my PhD. And then, yeah, I did a bunch of research thinking about how to connect reinforcement learning, with human data, how to do reinforcement learning, theory in a very rigorous way. And then I went to NVIDIA. That’s when GPT-4 came out. At the time, we were very interested in just solving a pen spinning task. I don’t know why. Like, -

Swyx [01:01:35]: It’s the dexter- it’s the dexterity limit, right? That you can solve Rubik’s cubes, and then you-

Jason Ma [01:01:39]: At the time, I was working with, -

Swyx [01:01:41]: Jinpeng

Jason Ma [01:01:41]: … Jinpeng at UCO, which you guys probably also know. Like, yeah, we were just like, “Okay, let’s try to do dexterous pen spinning.” The initial idea was learning it from human videos, right? We were trying to watch humans do it. But then I realized, okay, today we don’t have a dexterous hand that can spin pencil at all. So how would I even be able to evaluate the algorithm, even if learning from human worked at all? So I thought, “Okay, we need to first try this in simulation.” Then as soon as I tried to set up a simulated environment for a robot hand with a pen, I immediately saw what the issue is. The issue is like, if I had to come up with this simulation environment, the reward function for reinforcement learning to solve pen spinning, that’s impossible. Like, I wouldn’t be able to come up with all the terms that you need to get RL to learn the right behavior. And then I was just thinking, “What if a model that can write code can do that?” That’s what became the Eureka project, right? We basically. Like, in my opinion, we had the first recursive self-improving robotic agent in simulation, ‘cause we really used that to solve a new task that no one has demonstrated before. And that allowed me to see the potential of using language model, using coding models to really accelerate robotics. So then I went to Google DeepMind, ‘cause at the time, even today, they have one of the frontier models in Gemini, right? So I was looking at how to use Gemini for robotics. But eventually I realized like, I had a lot of fun doing academic research, but there was always a question in the back of my mind “Does any of this matter?” like, when you’re working on this kind of thing in a lab you ask me, “How do we come up with so many interesting, fun tasks?” You’re basically doing your own eval. You’re grading your own homework. Like, I wanted something more real. I wanted something that could actually be used by millions of people in the world. And I felt twenty-twenty-four was when we started Dyna, like end of twenty-twenty-four. That was the right time that we can take some of these technology to actually putting them out in the real world. So that’s why

Jason Ma [01:03:45]: I started Dyna with Lindon Gao and York Yang. And fast-forward two years, we built our platform, our model. Our robots are in Din Tai Fung. Go check it out and, -

Vibhu [01:03:55]: Go eat.

Jason Ma [01:03:56]: Yeah, go eat. It’s been a journey. And,. But what still gets me really excited is, like some of— a lot of the ideas that I temporarily abandoned when we started Dyna, because, to build a company, you have to be very practical. Like, I think simulation wasn’t ready in twenty-twenty-four. But now with like the new frontier models and what we’re seeing internally this might change. And that’s very exciting.

Product-Led Research, Experimental Evals, and Company Culture

Swyx [01:04:20]: And so then also let’s talk about company culture. What principles you’re building the company with. Obviously the— you’re responsible for the science and the research, agenda. And it just seems like when we meet the engineers, they are, they’re very engaged and then they have a lot of ownership. They have a lot of motivations.

Vibhu [01:04:38]: I’ll say it’s interesting to see how many side projects people just work on. So some of the demos we’ve seen, it’s just the engineers that work on it were like “Yeah, this seems fun. Might as well do it.” To your point on fun evals, right here, like two feet behind the camera-

Swyx [01:04:55]: Can’t talk about that.

Vibhu [01:04:57]: … just the e— the eval is just there’s highlighters, four highlighters being rubber banded together, right? It’s like, who comes up with, “ yes, we need to test putting rubber bands on high.” It’s just-

Swyx [01:05:07]: Fissionary-ish

Vibhu [01:05:08]: … fun stuff. People are very … creative here. It’s a beautiful space.

Jason Ma [01:05:12]: Yeah, I think culture. Yeah, so let’s first talk about, how we run a company, right? Like, we think of ourself as like a like a eventually like a research plus product company. It’s kind of like OpenAI where you do frontier research, but eventually all of them make their way into Codex through ChatGPT. That’s how we think about it too. We’re not here to just do open-ended whatever thing that interests us at the time. We’re always product-driven. Like, we wanna deploy our robots, our platform to a new customer. What are the technical problems we have to solve, right? We have to solve generalization. We have to solve whole body control. We have to solve how to adapt quickly. We have to solve how to make the robots follow language. We have to solve dexterity. So once all those problems are very clear, then, like. Our engineers are very capable. They’re very smart. So they just create instances of the task. They create tasks that will stress test those capabilities, right? Like the j— the grabbing a drink task is just how we test language following, right? You can just specify what you want. If the model follows intent, follows instruction, it should be able to do it. So that’s how we think about, like. At the top level, we identify what are the things we have to solve. Then we have a lot of bottom sub-project to actually drive execution to solve them. And you hire the right people. And everyone here is super passionate about robots. They’ve been doing this for a long time. They love it. And we just create a culture where it just naturally happens. And it’s a really fun place to work.

The Team: Hardware, Software, Research, and Learning

Vibhu [01:06:44]: What’s the split between technical research software versus hardware side of the company?

Jason Ma [01:06:51]: We recently grew so much, so I don’t have the perfect number. But I would say on the hardware side, we have, I would say probably like thirty percent, forty percent of hardware, just because hardware you need a lot more people. And then we have around ten people working on models. And then we have software engineers, robotics engineer. We have a lot of very talented people on the business side. The most fun thing about building a robotics company is you see everything. Everything under the sun that could possibly be helpful will be helpful, right? And so you just get so much exposure working with so many different people from many different backgrounds who are all very passionate about robots. And I think it’s been really eye-opening for me as a founder. ‘Cause when I was a researcher at a lab or a PhD student, I was talking to researchers every day, and that was great. That was really good. I still get to do that every day, but now I also get to learn just from so many different people from different walks of life. Like, some of our people who collect data, they had no robotics experience. They were working in retail, working in some other jobs, and they came here, you know. They also became very well-trained in robotics, and I think that’s just pretty amazing to see. And I think of myself as I’m still learning, you know. I think of myself as just always learning in the robotics space, and I haven’t learned so much as I am now. Even though, you think about me now, I’m not a student anymore. I’m here running a company, managing a lot of things, but I still feel like the day-to-day learning is just so much.

Dyna’s Vision and the Origin of Its Name

Swyx [01:08:29]: And then just to close off, where’s Dyna going? Two years, five years, what’s, the vision?

Jason Ma [01:08:37]: Yeah, I think our vision is to truly— The high-level mission is truly bring robots and intelligence, right, into the physical world. Like, actua— That means actually seeing our robots deployed thousands of customers, even more, millions of robots. That’s where we wanna get to.

Vibhu [01:08:58]: What’s the what’s the name from Dyna? Dynamic? Is there a is there a story?

Jason Ma [01:09:03]: Yeah, there’s many stories. So I na— I named Dyna. There were a lot of internal options.

Vibhu [01:09:09]: Take the credit.

Jason Ma [01:09:11]: … internal, different options. But how the name came to be was, there are two things. One is so, Turing Award winner, Richard Sutton, who was the pioneer of reinforcement learning, wrote a paper called Dyna many years ago. It’s about basically an agent that self-improves, world model. And I felt even when we started the company we weren’t doing world model yet, but I felt that’s how robotics will get solved. And now our technology direction is converging to that. So I felt that was a really neat inspiration, also because I worked in reinforcement learning, I have a lot of respect for Richard. The second thing is that it’s a bit more personal. Like, I grew up in Japan, and the year I was born, there was just a cartoon show. I don’t know if you guys have seen Ultraman.

Swyx [01:10:00]: Yeah, of course.

Jason Ma [01:10:01]: The yeah, the Ultraman that year was called Dyna. So, yeah, it’s a very subtle thing not people know, but I guess now everyone would-

Vibhu [01:10:10]: The team is finding out too.

Swyx [01:10:13]: But I think Richard would appreciate it. I don’t know if you ever talked to him, but-

Jason Ma [01:10:16]: No, I’ve seen him at RL conferences.

Swyx [01:10:18]: You should just be like, “Hey, this company-”

Jason Ma [01:10:20]: Yeah, I’ll invite him, come-

Swyx [01:10:21]: Inspired by you.

Vibhu [01:10:22]: Everyone knows Richard for a better lesson, but you know.

Jason Ma [01:10:24]: Yeah, also better lesson.

Swyx [01:10:27]: Okay, any parting thoughts or calls to action?

Closing: Deploying Real Robots and Building the Future

Jason Ma [01:10:31]: Calls to action. Yeah, I think I do feel like as an industry, robotics is still very demo-oriented. Like, I have done a lot of demos in the past in the lab, even here we had to show some demos. But I think we need more people that’s truly, mission-focused on doing whatever it takes to get robots deployed, get, robotic foundation models truly integrated with hardware and seeing them at real customer sites. I think we’re a pioneer on this regard, and I wanna see more in this field. And for the audience that’s watching, if this mission inspires you, if, solving the hardest problems today is what makes you wake up every day and you wanna contribute, we’re always hiring and yeah, looking forward to that.

Swyx [01:11:14]: Bay Area? Anywhere else?

Jason Ma [01:11:16]: Yes, just Bay Area today.

Swyx [01:11:18]: Bay Area.

Jason Ma [01:11:18]: But as company grows, never say never to other locations. Yeah, but that’s what I wanna say, and I wanted to see more robots in the world.

Swyx [01:11:27]: Cool. Congrats on everything. This is incredible progress for two years. Yeah. Amazing.

Jason Ma [01:11:34]: Yeah, I really appreciate it. Thank you for visiting us. I hope it’s been fun.

Swyx [01:11:40]: I don’t regret doing this on site. We do a lot of interviews in our studio, but this was worth it.

Jason Ma [01:11:48]: I think for robotics, gotta do it on site.

Vibhu [01:11:51]: For people listening in, watch the video version. We have fun B-roll.

From the Bitter Lesson of AI scaling to the unsolved mysteries of protein folding, Google DeepMind’s Pushmeet Kohli and Biohub’s Sal Candido are rethinking what it takes to build AI that truly understands biology. In this special panel moderated by Brandon Anderson, they explore why AlphaFold’s breakthrough was only the beginning, why scaling compute and data alone won’t solve biology, and how the next generation of AI models could transform our understanding of proteins, cells, and human disease.

We go deep on the future of AI-driven biology: finding scaling laws in biological data, the tradeoffs between scientific intuition and general-purpose architectures, why protein structure prediction is far from solved, and what it would take to build predictive models of living systems. Pushmeet reflects on the lessons behind AlphaFold, the limits of human interpretability, and whether future frontier models could understand other AI systems better than we can. Sal explains why protein language models may already contain scientific knowledge we haven’t unlocked, how biological modeling must move beyond individual proteins, and why achieving Biohub’s mission to cure all disease requires thinking in terms of 10x breakthroughs rather than incremental improvements.

We discuss:

* The Bitter Lesson for biology: why scaling compute and data isn’t enough

* Why finding the right scaling law matters more than blindly increasing model size

* How low-quality metagenomic data can improve protein language models

* Why AI researchers optimize for available data instead of the most important scientific problems

* Lessons from DeepMind on balancing modeling, data generation, and scientific expertise

* Why building a virtual cell requires fundamentally different datasets

* AlphaFold’s handcrafted architecture and the role of scientific intuition

* Why good data matters more than simply having more data

* Inductive biases, scaling laws, and the future of specialized AI architectures

* Why we aren’t in a post-Transformer world, but architectures are evolving

* Why AlphaFold didn’t actually solve all of protein folding

* Protein dynamics, disorder, and the limitations of static structure prediction

* How cryo-EM micrographs could unlock richer biological representations

* Moving from models of individual proteins to whole biological systems

* Feynman’s famous principle and why AI can now create things we don’t understand

* The hidden biological knowledge inside protein language models

* Why trustworthiness and uncertainty calibration matter more than full interpretability

* Whether frontier AI models could interpret other AI systems better than humans

* When AI could deliver 10x–100x acceleration in drug discovery

* Why curing all disease requires thinking about 10x breakthroughs instead of 10% improvements

Pushmeet Kohli — Google DeepMind

* X: https://x.com/pushmeet

* LinkedIn: https://www.linkedin.com/in/pushmeet-kohli-4838994/

Sal Candido — Biohub

* Biohub: https://biohub.org/team/salvatore-candido/

* X: https://x.com/salcandido

* LinkedIn: https://www.linkedin.com/in/salcandido/

Brandon Anderson — Moderator

* LinkedIn: https://www.linkedin.com/in/brandon--anderson

Timestamps

00:00:00 Introduction: The Bitter Lesson for Biological Data

00:01:00 Finding Scaling Laws and the Right Data for Biology

00:04:54 DeepMind’s Bitter Lesson: Solving Problems vs. Scaling Models

00:06:50 AlphaFold, Data Limitations, and Building the Virtual Cell

00:09:52 Handcrafted Architectures vs. Scaling Compute

00:11:53 Good Data, Inductive Bias, and Model Design

00:13:39 Beyond Transformers: The Future of AI Architectures

00:14:38 Why AlphaFold Hasn’t Solved Protein Folding

00:17:39 Protein Dynamics, Design, and Cryo-EM

00:19:09 From Individual Proteins to Whole Biological Systems

00:20:43 Feynman’s Principle: Creating Without Understanding

00:22:18 The Hidden Knowledge Inside Protein Language Models

00:23:48 AlphaFold, Trustworthiness, and Interpretability

00:25:56 Could AI Understand Other AI Models Better Than Humans?

00:27:43 When Will AI Revolutionize Drug Discovery?

00:29:47 Why Curing All Disease Requires 10x Breakthroughs

Transcript

Introduction: Is There a Bitter Lesson for Data?

Brandon Anderson [00:00:04]: Great to be here. What an exciting morning. So many cool announcements. I think the future of bioscience is being announced right here. This is the modeling session. We’re all modelers, so of course it’s natural for us to talk about data.

One of the things I like to think about when it comes to data is how to scale it properly. This has brought me to the question of the bitter lesson, but recast in the frame of data. The bitter lesson, for those of you who are not AI people, is the statement that methods that scale win eventually. If you can scale enough, it wins. So my question, starting with Sal, is: Is there a bitter lesson for data?

Sal Candido [00:01:00]: For sure. You obviously need the right data in order for it to work. One misconception of scaling laws is that scaling laws are everywhere and they always exist. A lot of the work is actually finding that scaling law. A lot of what we do is trying to figure out: What’s a situation where, if you put more compute into it, if you put more data into it, you get a better result out?

That’s a great situation because, once that happens, you can turn the crank. It becomes an engineering problem, which is something I like. That has to do with architecture, but it also very much has to do with data. If you don’t have data with the right information and statistics to solve the problem you want, you’re not going to get a model that has the capabilities and understanding you want. You can only really pull information from the data you have and use that to generalize beyond it.

Choosing the Right Data: Availability vs. Scientific Impact

Brandon Anderson [00:02:09]: When it comes to data collection, how do you think about which modalities are best? A related question: Do modelers tend to work on problems where the data is available, rather than the problems that best serve their goals or have the greatest impact on translational medicine?

Sal Candido [00:02:32]: For sure. I won’t speak for all modelers, but I’m lazy, so I’m going to work with what’s available. There’s a good side and a bad side to this.

The good side is that, when you’re doing conventional machine learning, you’re often looking for the most pristine, high-quality data examples you can find. But if you’re like me, you’re rooting around in the back room, looking through people’s junk to see what’s there. A concrete example is training a protein language model on metagenomic sequences, which are not the highest-quality data. In fact, much of that data, I can guarantee you, isn’t even a real, whole protein. And yet it makes the performance of the model go up for designing real proteins that work and understanding proteins that we know exist.

That’s the positive side. But it can lead you to a negative place where you say, “Let’s just scale up the data that we can generate easily.” I don’t think that’s necessarily the way to do it. That’s one of the things that’s exciting to me about what we’re talking about here today with the BBI. It’s going out and asking, “What is the data that we need to solve the problem?”

It’s about the right resources, but also the right community. One thing we try to do at Biohub is work in the open, work with the community, and move the whole community forward. That’s critical because, if you just have people building models, we’re going to lean toward the data that exists. If you just have people generating data, they’re going to lean toward the things that can be generated. But if you can work together as a community, every step of the way, in an open fashion, you can figure out what data you actually need. Then you can find that scaling law.

The Problem Comes First: Modeling, Data, and Expertise

Brandon Anderson [00:04:50]: Pushmeet, what do you think about the bitter lesson for data?

Pushmeet Kohli [00:04:54]: I was at DeepMind when Rich was with us and thinking about this idea of the bitter lesson. What I took from Rich’s original lecture at DeepMind wasn’t about the specific notion of whether data is useful or how we should think about machine learning. My take was that he was talking about something more conceptual.

Sometimes when we’re looking at problems, we think about solutions in a very religious way: “I’m a modeler,” or “I’m a data generation person.” I think that is the bitter lesson. If you approach a problem with that mindset, you might not succeed. The problem comes first, and you should be flexible in your solution space. You should try both things. You should understand the problem you’re trying to solve. If it requires modeling effort, do the modeling effort. If it requires collecting more data, then collect data.

What happened in machine learning at the time was people saying, “We’re machine learning researchers. The dataset is there. Here’s some training data, here’s some test data, and we’ll just optimize the model.” That is broken in the sense that, if your eventual goal is to solve the problem, you have to look at both aspects of what goes into the process.

And what goes into the process is not just data or modeling. It’s also expertise. With AlphaFold, for example, we looked at what was possible with existing datasets because we didn’t have the core expertise, or even the resources, to say, “Let’s augment the PDB by a significant order of magnitude.” The investment that organizations and scientists across the world had put into constructing that data was invaluable. So there, you had to focus on getting the biggest bang for your buck by investing in modeling.

But in other areas, say cell genomics, we took the same approach and asked, “What can you do with cell-by-gene data?” After a lot of work, it was very clear that the data was not there yet to pursue that grand ambition of building the virtual cell. This brings people together to focus on the actual challenges of advancing science rather than religiously following advances in data generation or modeling.

Brandon Anderson [00:08:23]: I really like that answer. What’s the actionable takeaway? Always define the problem first and then figure out which solution space to search over. But more broadly, how should the community think about this as we move toward the next generation of translational medicine?

Pushmeet Kohli [00:08:47]: The advice I give to anyone starting in this area is to think of yourself as a multidisciplinary person. Understand the problem first. Why are you working on it? What are you trying to achieve? Then think about what’s needed, whether that’s modeling, compute scaling, or data generation. Understanding the problem is extremely important, and of course you need to build your expertise.

There are constraints, too. Maybe there’s only a certain amount of data you can generate, or a certain model size you can afford to train. Understand those constraints, try to fail fast, and look at which approaches will be feasible in the long term to get you to the level of impact you’re aiming for.

Scientific Inductive Bias vs. Scaling: The Craft of Building Models

Brandon Anderson [00:09:52]: This leads right into my next question. When I look at the evolution of modeling, AlphaFold 2 was essentially a work of art, with a lot of carefully handcrafted features. There was very intentional thought put into every part of the solution. Some of Google’s or Alphabet’s work still stays in that space, but the general consensus seems to be moving toward more scalable, general strategies.

Do we still need artisanal, craft solutions for certain problems? Or are resources better spent focusing on scale first? With fixed resources and money, you can invest in compute, talent, or data. How should we think about that trade-off?

Pushmeet Kohli [00:10:48]: Let’s approach it from first principles. The art and craft of constructing a better model wasn’t accidental. We did a lot of experimentation, but there was a vision behind it. Scientific intuition from biophysics and biochemistry told us that amino acid residues are not just doing their own thing. They’re influenced by other residues. So let’s bake that in. If you’ve learned something from the scientific community, use that information. Give the model that unfair advantage.

That makes the model much more data-efficient because it doesn’t have to relearn everything. I also have to mention that curating good data is an art in itself. It’s not as if you can just say, “Put in more data.” If you replicate the same data, you’re not going anywhere. It’s not just about big data; it’s about good data and understanding the coverage of the data necessary to make progress. That’s a much more interesting and challenging problem in itself.

Brandon Anderson [00:12:20]: It looks like you have a thought, Sal. What do you think?

Sal Candido [00:12:24]: I very much agree. It depends on the problem to be solved. If you have a smaller amount of data, having more inductive bias in the model is going to help you. You actually need that to get results at smaller data scales. As you get more and more data, sometimes the model finds things you didn’t necessarily know about, and sometimes that inductive bias, if it wasn’t exactly correct, can hold you back. There’s a tipping point as things get better.

But I also object to your question a little because there’s a lot of craft in the scaling part of things as well. There’s a lot of algorithmic work that goes into taking models, training them on more data, putting more compute into them, and making them bigger. And that isn’t only from an infrastructure perspective or about making the models run inference faster. We’re seeing things go beyond standard transformers to more bespoke architectures. I don’t think we’re in a post-transformer world in any way, shape, or form, but we’re modifying those architectures to make them more fit for purpose and work better, even with internet-scale data.

As more data comes in, the challenge isn’t only curating that data or selecting the next batch of information the models need. It’s also asking, at every step and at every scale, what the right architecture is to get the most out of that information.

Why AlphaFold Didn’t Solve All of Protein Biology

Brandon Anderson [00:14:36]: If you think about the state of protein structure prediction right now, the news will say that protein structure prediction has been solved. But if you talk to my friends, they’ll say, “We have so much left to do here.” Function, dynamics, and design are still wide-open problems. It’s been about five years since AlphaFold 2 was announced, and progress has been made.

Pushmeet, do you think there’s another big leap coming? Are there blockers to solving these problems? Do we not have the right data, algorithms, or ideas? Or is it just a matter of time?

Pushmeet Kohli [00:15:28]: Science operates by isolating something and then making progress step by step. When people say the protein-folding problem has been solved, at a conceptual level there have been advances. But think about the narrative of proteins being the building blocks. Proteins aren’t blocks, and they don’t act as blocks. I say proteins are the building blocks all the time, but I don’t actually believe it.

Proteins are extremely complex. They’re disordered. Their shape might change depending on the context. John Jumper and I used to discuss this: What are we trying to solve? We don’t know the actual true ground state that proteins take, or the actual distribution of structures that proteins take. What we were trying to do was replicate a structure that somebody had obtained and deposited in the PDB. That’s what we did. And it just so happens that it’s useful.

But that doesn’t mean we’ve understood all of protein dynamics. At the top level, it’s easier to communicate that we’ve made progress, but the scientists in this crowd know how much remains to be done. There’s a lot to celebrate, but let’s not stop funding protein structure prediction and protein dynamics, because we’re just getting started.

Cryo-EM, Molecular Dynamics, and the Missing Information

Brandon Anderson [00:17:39]: What’s the biggest blocker? If you could wave a magic wand and say, “We have more of this,” and that would accelerate function, dynamics, or design, what would you bring into existence?

Pushmeet Kohli [00:17:54]: My background is quite eclectic. I started as a security researcher, then went into computer vision, Bayesian theory, discriminative machine learning, deep learning, AI for coding, and finally science. So when you ask me that question, the computer vision researcher in me gets very excited about cryo-EM micrographs.

I thought, “What is this PDB data? I should be working at the source. I should be looking at cryo-EM micrographs. I don’t want those structures. They must be missing out on all the data. I should just operate directly on cryo-EM micrographs.”

Getting models that can scale at that level, with the right amount of data, and extract the dynamics and distributional information captured there would be amazing. I tried it, but it requires more work.

Brandon Anderson [00:18:57]: There’s still work to do, but you believe that’s a route that could give you that information?

Pushmeet Kohli [00:19:00]: Yeah. I think at some point maybe some people better than me will take a stab at it, and we’ll get somewhere.

From Protein Structures to Larger Biological Systems

Brandon Anderson [00:19:07]: What do you think, Sal?

Sal Candido [00:19:09]: It’s interesting because these models are quite useful, but they’re not exactly the problem that most people want to solve. They solve a very specific purpose, and you can also use them to do other things. For example, you can use them to design new proteins, which isn’t necessarily what you would start with.

I think of the models we’re building now as someone who’s trying to understand how a bicycle works, but is modeling a spoke on it. Those models can get better and better over time, but what you really need to do is move from models of spokes to wheels to whole bicycles, because that’s what people want to understand. To continue the analogy, it seems like people want to use a model of the bicycle to design a part for a pickup truck.

As we put these models into the particular biological context in which they’re operating, we’ll be able to learn more about these interactions on a broader scale. That’s where I think things are going. But I agree that we should keep working on folding models, because they’re going to keep getting better.

Protein Design vs. Scientific Understanding

Brandon Anderson [00:20:43]: With regard to design, there’s a famous Feynman quote: “That which I cannot create, I do not understand.” Now we’re in a world where it’s really easy to create things without understanding them at all. How important is it to have models that help humans understand things, versus magical black boxes that can effectively one-shot a picomolar binder or something like that?

Sal Candido [00:21:18]: We were designing things with magical black boxes long before AI came around. In some sense that’s still useful. But understanding is really important, and it’s one of the big things I think about with AI.

There’s so much for these models to learn. As intelligence gets cheaper and more available, you can deploy it to learn more and more about what’s going on in the world. But how do you pull that knowledge out of the machine so that I can understand it? Maybe that’s just my esoteric curiosity. I think there are so many things to learn.

Brandon Anderson [00:22:05]: A lot of scientists really want to understand things, and the endpoints may not be as important to them. But we’re here to solve translational medicine as a problem, right?

Sal Candido [00:22:18]: I think the more you dig into things, the more you find the right way to keep pushing them forward. One thing that’s really salient to me is that these models have a lot more information in them than we know.

We’ve worked a lot on interpretability for our models, for example, and you find a lot of information there. People know that protein language models learn some notion of structure within their representations, but we find information about functions and motions as well. There’s a lot still to be unlocked, even from the models we have now. That’s important for us to understand as we raise their capabilities.

At the end of the day, you expect a world model to emerge from compressing all this information into a model. How does it do its job? How does it design a protein? It has compressed information from evolution into that model. In addition to being able to produce something useful to us, there’s certainly something to learn just by looking at what’s inside.

AlphaFold Confidence, Calibration, and Interpretability

Brandon Anderson [00:23:48]: What do you think, Pushmeet? Design versus understanding?

Pushmeet Kohli [00:23:53]: I have a different take in the sense that some level of understanding is necessary. Let me explain what level I mean.

AlphaFold isn’t perfect. AlphaFold 2 had a GDT score of about 90 on that set at the time. But even if it had a GDT score of 95, if its pLDDT score were completely uncalibrated, who would trust it? Imagine that it magically gave good answers but told you it was very confident, and then you worked on it for the next year only to find out it was completely wrong. Calibration of the uncertainty measure was extremely important.

In that sense, we do understand AlphaFold 2, and we made a lot of progress in understanding how it behaves. That’s different from understanding how it worked internally to find the solution. There I agree with Sal that interpretability asks: Why did it work? Why did it give this answer?

At the highest level, AlphaFold 2 was interpretable in terms of its behavior and ability to generalize, and we didn’t discover everything about that before launching. When we launched AlphaFold 2 and made the weights available, people found that it was a great disordered-protein predictor. It could figure out which elements of the protein are disordered. That shows it generalizes.

But interpretability asks another question: Interpretable by whom? If you’re saying interpretable by a human rational system, with the cognitive and computational limitations of the human brain, then no, AlphaFold 2 is not interpretable. But if you’re asking whether AlphaFold 2 is interpretable to a much larger, more sophisticated model in terms of how it works, maybe it is. We just don’t get it.

As users of AlphaFold 2, we do need to understand what it can and cannot do. Understanding its behavioral characteristics, strengths, and limitations is extremely important. We can’t just use these models without that characterization, because otherwise, rather than being helpful, they can harm us.

Brandon Anderson [00:26:53]: So your take is that interpretability, strictly speaking, isn’t necessary, but ensuring that a model is trustworthy so humans can make actionable decisions is what people should focus on.

Pushmeet Kohli [00:27:04]: Exactly. Interpretability is also in the eye of the beholder. Who is interpreting it? If it’s a human scientist trying to interpret how the model makes a prediction, that’s a different question from giving another, much larger LLM access to the activation layers and saying, “Can you predict what AlphaFold will do?” Maybe those frontier models of the future will be able to predict that and come up with a theory of how AlphaFold 2 was interpreting and producing its results.

When Will AI Transform Drug Discovery and Clinical Medicine?

Brandon Anderson [00:27:43]: We have a two-minute warning, so one last question for both of you. There’s a real chance that AI will make dramatic improvements in human health in the immediate future. I like quantitative predictions. Best guess: How long until we start seeing AI results in the clinic? Pushmeet, do you want to start?

Pushmeet Kohli [00:28:08]: I think “AI results in the clinic” is an ill-posed question. AI is already being used today in every part of the drug discovery process. From that perspective, it’s already there.

But if you’re asking when we’ll see a 10x or 100x acceleration in timelines, then the next question is: What are we accelerating? Is it lead optimization? Target discovery? Preclinical work or toxicology?

Over the next few years, and this is why the effort announced today is extremely important, we need to tackle some of the hard challenges of biology. Only then will we get the true unlocks of acceleration that dramatically transform drug discovery. AI will continue to be used in things that go into the clinic all the time, but larger acceleration will only be unlocked with a better understanding of the biological models this effort is trying to create.

Brandon Anderson [00:29:43]: All right, thanks. Sal, a quick answer, if you can. We’re almost out of time.

Sal Candido [00:29:47]: I’m tempted to literally put on my biohacker hat to answer this question. It’s important to think about what it means to push the field forward. What we really want to see is outcomes being affected. When will a drug be made entirely by AI? I don’t know. That’s hard to predict. But I do think we’re going to see rapid progress very quickly, because all these tools are already being used.

Going back to your point about basic research and basic understanding, one thing I learned a long time ago in my career, back at Google, is that sometimes it’s easier to approach a problem by asking what it would take to make a 10x improvement rather than a 10% improvement. That’s not because the 10x path is necessarily easier. It’s because it allows you to take a broader view and see solutions you haven’t been approaching. You go back to first principles and ask, “How are we going to do this? How are we going to really push this?”

Both approaches, the 10% and the 10x, are valuable, and we should do both. I’m very happy that at Biohub we have a beautiful and lofty mission statement: to cure all disease. If you want to do that, you really need to figure out what the 10x approach is. It’s a great opportunity to be able to go do that.

Brandon Anderson [00:31:46]: Awesome. Thank you both. Thank you for being in the literal hot seat. Very interesting.

It’s hard to believe that Periodic was only launched last September:

One year later, it is considered one of the pre-eminent AI scientist labs, with dizzying talent density and astonishing progress in the autonomous lab buildout:

Most people are familiar with the standard credentials of Liam and Dogus, but we found an incredible “talent slope” while learning more about Periodic, where each successive employee seems more impressive than the last:

From building AI systems that reason over noisy physical experiments to creating laboratories where every instrument can become intelligent, Periodic Labs is betting that the next frontier of AI won’t come from simply training on more internet data, it will come from letting models experiment with the real world. In this episode, Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk join swyx and Brandon to explain why scientific discovery is fundamentally different from math and coding, and what it takes to build AI scientists that can actually discover new materials.

We go deep on Periodic’s vision for “synthesis superintelligence”: reinforcement learning grounded in physical experiments, AI-powered materials characterization, simulations and density functional theory, high-throughput labs, and systems that learn from the entire process of doing science rather than only its published results. Liam and Dogus also explain why frontier models will still need experiments, why failed experiments may be some of the most valuable training data, what it means to give every piece of lab equipment “140 IQ,” and how autonomous experimentation could compress decades of scientific trial-and-error into months.

We discuss:

* Why intelligence alone isn’t enough for scientific discovery

* How reinforcement learning changes when the environment is the physical world

* Why science requires reasoning under uncertainty, noise, and missing information

* Prediction, synthesis, and characterization in the materials discovery loop

* Why physics and materials science are still far from “solved”

* The “matter compiler” and Periodic’s goal of synthesis superintelligence

* Phase transitions, X-ray diffraction, and AI-powered materials characterization

* DFT, simulations, and why experiments remain the ultimate ground truth

* Room-temperature superconductors, new magnets, batteries, and more efficient compute

* Why quantum computing may not automatically solve materials discovery

* What it means for every piece of lab equipment to have “140 IQ”

* Why data quality and negative results matter more than simply throwing more compute at science

* Training models on the process of doing science rather than the final answer

* Why even future frontier models will still need to physically experiment

* Scaling autonomous labs across AI, chemistry, physics, and custom hardware

* How automated experimentation could massively increase the “surface area for luck” in discovering new materials

Liam Fedus

* LinkedIn: https://www.linkedin.com/in/liam-fedus-26547811/

* X: https://x.com/LiamFedus

Ekin Dogus Cubuk

* LinkedIn: https://www.linkedin.com/in/ekin-dogus-cubuk-9148b8114/

* X: https://x.com/ekindogus

Timestamps

00:00:00 Introduction

00:02:49 AI and Reinforcement Learning in the Physical World

00:09:17 The End-to-End Materials Discovery Loop

00:13:28 Why Physics Isn’t “Solved”

00:19:04 The Matter Compiler and AI Characterization

00:30:22 DFT, Simulation, and Experimental Ground Truth

00:42:27 Dream Materials, Compute, and Quantum Computing

00:45:57 Giving Every Lab Instrument “140 IQ”

00:50:02 Automating the Lab

00:53:16 Data, Models, and Negative Results

00:58:13 Training AI on the Process of Science

01:00:20 Why Frontier AI Still Needs Experiments

01:06:06 Scaling Autonomous Labs

01:10:45 Building the Team and Deploying to Industry

01:17:01 From AI Copilots to Scientific Outcomes

01:20:57 Automating the Search for Superconductors

Transcript

Introduction: Intelligence Must Meet Reality

Swyx [00:00:00]: Okay. We’re here at Periodic. We’re here today with Liam and Doğuş from Periodic. Welcome, and thanks for having us at yours.

Liam Fedus [00:00:11]: Yeah, great to be here, and yeah, thanks for coming in.

Swyx [00:00:13]: I want to start off with one of these quotes that I love from your website. It says, “Intelligence is necessary but not sufficient. New knowledge is created when ideas are found to be consistent with reality.” It seems so straightforward, but why is it non-obvious?

Liam Fedus [00:00:29]: I think that’s sort of the thesis and premise behind Periodic that Doğuş and I thought when we were building this, which is you can’t just think your way to a solution. The universe is so complicated that in order to actually push the frontier of knowledge and to make progress, you need to create these conjectures and then actually see whether or not it holds. No amount of rereading that textbook or paper, thinking, or disappearing into a room is going to allow you to think through all possible experimental outcomes, expected and unexpected, and you really need this iterative process. And that’s our core belief for putting this together, and that’s why from the very beginning, we’re like, “We need to have the AI systems, the simulations of the physical world, but then also to build up the physical high-throughput experiment.” And we think a different type of intelligence emerges from that.

Swyx [00:01:21]: I think it’s also notable, like with OpenAI and Google, you don’t have those resources in those big labs. You had to come out and do your own thing, because you’re kind of inventing your own playbook as you go along.

Building an Interdisciplinary Lab for AI Science

Ekin Doğuş Çubuk [00:01:34]: Yeah, and a team like this has never existed, I think. We try to bring together solid-state chemists, solid-state physicists, experimentalists, theorists, hardware engineers, LLM experts, computer scientists, and some of these technologies are very recent. The high-throughput experimentation, the robotic arms have only been tried in the last three years or so. The force field expertise, some of these force fields are also very recent. But yeah, we felt like having a lab is very important. Having this focus and bringing all these people together to work together is very important. And I guess the closest examples from history are places like Bell Labs, where incredible theorists and experimentalists and chemists worked together and achieved incredible things. So we are trying to do the same here, yeah.

Physical-World Reinforcement Learning and Uncertainty

Brandon [00:02:17]: I want to talk about force fields in a bit, but

Swyx [00:02:19]: Yes

Brandon [00:02:19]: Before we get to that. Yeah what are they? But before we get to that, For context, some of our audience are AI engineers, and some are scientists. For the engineers: what’s the difference between working at a big lab like OpenAI and what you’re doing here? I think you’ve already hinted at this but maybe explicitly, how do you have to reframe your thought process?

Liam Fedus [00:02:49]: Well, I think one interesting thing is now our reinforcement learning environments literally derive from the environment, from our physical labs. Our data comes from our physical labs, and this is sort of our ultimate truth. It’s not enough just to do optimization against, some answers that were known in some papers or textbooks because we’re going beyond that, and I think that’s one of the biggest differences. But along with that, there’s issues of say, decision-making under uncertainty. So when you’re doing optimization against math, there’s a high precision to it. You’re not really dealing with kind of like variance or uncertainties or aberrant measurements, whereas that’s very key to our process. For example, when we are doing a materials discovery loop, things don’t come out of the furnace labeled, right? Even the labeling process can be stochastic and noisy. And sometimes you’ll have aberrations between different machines. Maybe we’ll have insufficient telemetry. So we thought we ran it at this temperature, but really the temperature was, some delta away. And an intelligence that can take in this noisy data and make intelligent decisions like a scientist would, is a sort of a different set of reasoning strategies. So I think there’s a huge amount of commonalities, so the standard mid-training, reinforcement learning, the construction of tool-using agents, variance reduction, making sure infra doesn’t have mismatches between training and inference. But we have to go beyond that and really think about how do you do accurate, work when there’s a high amount of uncertainty? How do you make really incredibly efficient use of a limited set of data? So we’re scaling things up significantly, but still, it’s very different than a digital environment where you can arbitrarily add more environments or more rollouts. We don’t have that same capacity. So sample efficiency is another key thing.

Why Experiments Are Noisy and Incomplete

Brandon [00:04:48]: This seems a little bit abstract to me, and it is a theme which has been on the science pods for a bit, but I think it might be helpful when you talk about experimental uncertainty, to explain what would a specific process that you were doing experimentally look like, and how would a human approach this problem, before we get into the AI side? Do you have a specific example of a type of material you make and what sort of uncertainty in the measurements you would experience in the process?

Ekin Doğuş Çubuk [00:05:17]: I think there are many dimensions to this. So one of them is I think every scientific experiment has to do some dimensionality reduction, and this is very different than coding or math. So, when you’re doing math or coding, all the context could be available to the human or the LLM. It’s a bunch of axioms, some corollaries people have derived from them. It’s a bunch of functions and API calls. So everything you need to reason through is available to you, and then you just have to be very smart and figure it out. In physics, as you know, we start with more atoms than we could ever store on a computer. So clearly, we’ll have to go from the original number of dimensions and bits that represent a system to the number of bits we can fit in the computer. And this was, I guess, the original premise of thermodynamics. They were originally really confused about how steam engines worked, but then turns out you can find five or six thermodynamic variables that explain what’s going on pretty well, which is incredible. That’s the beauty of physics.

Brandon [00:06:13]: You have a room full of atoms, and there are 10²³ or 10²⁷

Ekin Doğuş Çubuk [00:06:19]: Exactly

Brandon [00:06:19]: Or something atoms in this room, and yet you need temperature and pressure and simple variables, and that tells you everything. No, it tells you mostly what you need

Ekin Doğuş Çubuk [00:06:28]: A lot

Brandon [00:06:29]: To know for most cases, right?

Ekin Doğuş Çubuk [00:06:30]: Yeah.

Brandon [00:06:30]: So let’s say you make a material, and you could do something like look at an X-ray structure, and this thing is a very lossy, projection, right? It’s not actually a structure, or it doesn’t tell you the true structure. It tells you some piece of principal components of that. So, how would a human take these simple principal components and then, kind of reason about them? And then how do you extend that to an AI?

Ekin Doğuş Çubuk [00:06:58]: So on the theoretical physics side, right, we’ve decided that energy, the average value of energy and the fluctuations of energy is very important. So a lot of thermodynamics has been derived from this. And then we realized reaction barriers are also very important. The kinetics is very important. So a human can try to reason through this extremely complex system by coming up with some, reduced dimensionality descriptors. People also really like atomistic picture. You think about what the local neighborhoods look like, whether they’re from octahedra, tetrahedra, and that especially chemists really benefit from that. Okay, so that’s on the theoretical side. And on the experimental side, you would try to collect as much data as possible. So you’d have a lab notebook, and you’d write everything you see, et cetera, but with the understanding, of course, that you’ll miss a lot. So like as Liam mentioned, there are all these issues that come up. For example, your furnace will degrade over time because as you’re using the furnace, some of the things you’re baking will evaporate and cover the heating element. So every time you use the furnace, it gets worse and worse. Another issue is you have a furnace, but the temperature isn’t perfectly uniform, so where you put in the furnace will affect the result. We don’t have this issue yet, but we will maybe someday. The optical instruments, suffer a lot from vibrations. So every time somebody walks. I remember when I was doing my PhD, one lab had this big issue because sometimes their results would be very different than other times, and they brought it down to at a certain time in the night upstairs, somebody would be walking and that vibration affected the laser setup. So yeah, science is very hard, but that’s what makes it special, right? The other thing Liam and I have been talking about is a lot of the current improvements focus on math and coding and theoretical computer science because it’s easier for LLMs. But in real life, most things that require intelligence are actually more like science. There’s a lot of uncertainty, there’s a lot of noise, there’s a lot of missing context, but you have to be the intelligent being and figure out what to do next.

Liam Fedus [00:08:53]: I think a really interesting point to build on that too is the optimal reasoning strategies to do math, to do theoretical computer science, to do these types of things may not be the optimal reasoning strategies to do science.

Ekin Doğuş Çubuk [00:09:04]: Yeah.

The Materials Discovery Loop: Synthesis and Characterization

Brandon [00:09:05]: How do you define an RL reward function when your output is a noisy crystal structure or some number of I don’t know collective variables or something? Like.

Liam Fedus [00:09:17]: Maybe, I’ll give, one part of the loop, and I’ll just kind of dig into that. So, as part of the materials discovery loop, we have to first figure out, what to make. So that’s the stability of the atoms together, as well as do we expect it to have the properties we want. So we don’t just want, a novel configuration of the atoms. We want them, when put together, to have the property of interest. Next is how do you synthesize it? How do you actually make the thing? That’s also non-trivial. So even if you know something will hold together, the process of actually, what are the processing conditions to actually bring that forth is highly non-trivial. But again the thing. Once you’ve done those two steps and you’ve made something, it doesn’t come off labeled, so you actually have to characterize it and figure out what you made. So in that case rather than thinking about the RL environment of we’re just going to kick off an experiment, wait a couple days, and try to do a update on did you find a room-temperature superconductor or not, that’s completely infeasible.

Brandon [00:10:15]: Yeah. I was going to ask that. Yeah you can’t leave your GPU sitting for a week or something.

Liam Fedus [00:10:19]: It’s. Yeah, it’s completely infeasible

Brandon [00:10:20]: Yeah

Liam Fedus [00:10:20]: Because you don’t have enough agents to reduce the variance sufficiently. The rollout time is going between the different instruments. You’re waiting. There’s just some physics limits of the experiments, for the actual synthesis or furnace time. So it’s just too slow, too noisy. So the way we think about the AI program is basically constructing agents around this set of data that we’ve produced. And so in the case of characterization, you’re looking for RL environments that are rewarding the identification of the phases actually present. Are you able to given the raw experimental data, fit that effectively? So the way we do this is we’ll shoot X-rays, at the material. X-rays have a frequency or a wavelength roughly measurable or commensurate to the spacing between atoms, so you get nice diffraction patterns that kind of give a fingerprint of the crystal structure. And doing this, identification of what actually is present is highly non-trivial. And so some of our early AI work has been on basically producing systems to do this. And this allows us to move much more quickly too, because if you’re running so many experiments, you really quickly become bottlenecked on your ability to understand what was produced. But again, now the reinforcement learning environment is fairly straightforward, which is you’re going to reward the identification of the phases present. You’re going to penalize spurious phases or things that don’t have any real chemical plausibility. So that’s like a little succinct thing where you can kind of do this for the different pieces. And when you stitch this together, that’s the end-to-end discovery loop.

Training on Novel Experimental Data, Not Memorized Answers

Brandon [00:12:02]: Okay. You

Liam Fedus [00:12:02]: Maybe another

Brandon [00:12:03]: Yeah. Oh, sorry

Liam Fedus [00:12:03]: Another piece would be as we build up, a large basket of experimental data, you can think of another thing of basically timestamping the state of the world. So you could say, “At this date, this was our experimental evidence up until that point.” And you can construct reinforcement learning environments where you say, “Okay, given that body of experimental evidence and given, the choices, what was the next choice that the scientist made, or what was an outcome of that experiment?” And. This is really interesting too, because this allows us to do types of programs that would be harder externally. Because if the pre-trained model has already memorized some of this data and then you try to do reinforcement learning on that, if it already knows the answer and you construct reinforcement learning tasks that depend on that answer, it can fake work. So, and then, because it’s like, oh, well, it already knows the answer, so it doesn’t actually have to do the hard physics reasoning. It doesn’t have to actually do the calculation simulations. And it gets to the right answer, and then you reinforce that policy and you up-weight those reasoning strategies. Those reasoning strategies will not generalize to novel systems, so it’s not useful to us. But as we build up baskets of experimental data and stitching things together, having the lineage through the whole process, that gives us the ability to construct new types of environments that are just not plausible elsewhere.

Why Known Physics Does Not Mean Perfect Simulation

Swyx [00:13:27]: I want to ask my question

Brandon [00:13:28]: Seriously

Swyx [00:13:28]: But I’m worried that I’m taking it off the rails. I’m just going to go for it because I, again, I come from the non-scientist, but engineering sort of. I’m familiar with the ML side, but not the not the physics side. Okay, there’s a few versions of this, but I think the basic question is, don’t we have most of the laws of physics worked out? And how come we don’t have perfect simulators already?

Brandon [00:13:50]: Oh, that’s a great question.

Swyx [00:13:52]: Like, this is very basic, like. And he was like, “Ha, that’s cute.” But like

Brandon [00:13:56]: No, it’s a great question

Swyx [00:13:57]: Dude like I have all these physics books. What do you mean, what do you mean we’re not done?

Ekin Doğuş Çubuk [00:14:03]: Yeah, so there are a couple of things going on here, right? So

Swyx [00:14:07]: Yeah

Ekin Doğuş Çubuk [00:14:07]: I think first of all, we’re definitely not done with any of the laws of physics. Even the

Swyx [00:14:12]: At quantum scale, okay, but. And maybe very large scale, but at human scale, at material scale, we’re, we. What’s left?

Ekin Doğuş Çubuk [00:14:22]: We’re not. So it’s a it’s a really good question. And it’s very interesting why we’re not done yet, right? So there’s a very, common story people talk about. When Dirac was figuring out quantum mechanics, like late 1920s, he wrote a textbook about quantum mechanics, and he kind of framed it as it’s all done. And then there’s a phrase people like to quote. I’m not sure actually how accurate it is, but apparently Dirac said, “The rest is chemistry.” The idea being he could solve the hydrogen atom

Swyx [00:14:50]: Just compose everything.

Ekin Doğuş Çubuk [00:14:51]: Yeah. He could maybe solve a chain of 1D hydrogen atoms, but then he cannot currently solve, say, nitrogen interacting with oxygen, but that’s okay, that’s just chemistry.

Swyx [00:15:00]: Yeah.

Brandon [00:15:00]: Physics physicists really love single atom rep. Really simple systems.

Ekin Doğuş Çubuk [00:15:05]: Simple systems.

Brandon [00:15:05]: And that’s all you can solve.

Ekin Doğuş Çubuk [00:15:07]: Spherical cow.

Brandon [00:15:07]: Yeah, spherical cow.

Ekin Doğuş Çubuk [00:15:08]: But it turns out, I think what we learned in the last hundred years is, first of all, that’s not true. It’s not like a simple extension of what it was to do just hydrogen. And it’s not just. The rest is not just chemistry. There’s actually a lot of physics there. I guess one of the things we still haven’t figured out is high-temperature superconductivity. But there are a lot of other

Swyx [00:15:25]: And high for you is 175, 200?

Ekin Doğuş Çubuk [00:15:28]: Oh, when people say high-temperature superconductivity, what they mean there is unconventional superconductivity. So there’s the conventional superconductivity that’s just mainly driven by electron-phonon coupling. And in that case, you can kind of see the isotope effect. If you take the same system but just different weight for one of the elements, you can see the superconducting temperature drops at the rate you expect. So that one is conventional. But then higher-temperature superconductors like cuprates, like the ones that are above 77 kelvin, like 93 kelvin, turns out don’t obey that physics. But we don’t know what physics they obey. They’re just incredible superconductors. But it’s not just that. So that definitely is one of these very, popular topics we don’t know the theory for yet. But there’s so many other things we don’t understand. For example, even strong correlation in simple quantum mechanical systems, we cannot simulate yet. Density functional theory is a very powerful tool. It’s incredibly accurate on some things, but as long as. As soon as there’s some strong electron correlation, it actually fails to capture some of the effects. There’s so much to figure out. That’s why I’m really excited for AI to be applied to this field, because theoretically, computationally, and experimentally, there’s so much to figure out. And every improvement we make here should improve human life, because, the better we can understand materials, solid-state physics, the better devices we can make for them.

Brandon [00:16:41]: There’s this famous quote by Phil Anderson, which is, “More is different,” which is basically you may understand all the basic laws of something, but when you add many things, they behave qualitatively distinct, differently from how the simple physics should tell you. So understanding how lots of things behave together, it seems like it should be simple. Beginning laws are simple, but the collective behavior is so complicated that it’s really hard to model.

Ekin Doğuş Çubuk [00:17:07]: It’s emergent and universal, which is crazy. And we’re also seeing this with deep learning models, right? We see power laws everywhere in deep learning. We don’t understand, but it is very reminiscent of physics where there is, emergence and universality.

Brandon [00:17:20]: There’s been many theory papers in the physics world about emergence in neural networks.

Ekin Doğuş Çubuk [00:17:24]: Yeah, absolutely, yes.

Brandon [00:17:25]: Yeah.

Synthesis Prediction and the Matter Compiler

Swyx [00:17:26]: And then my other follow-up question was also just on the process of let’s call it the yeah, process and the end result. Let’s call it like, you want to have these properties that you’re targeting for materials, and then you have to figure out the process to get there. Is it worth separating these things so that you can have. You just train a model that perfectly reverse engineers any process whatsoever given, whatever theoretical end state you want? Is that, meaningful?

Ekin Doğuş Çubuk [00:17:54]: Maybe going back to the conversation we had about how since there are more than Avogadro’s number of atoms, we’ll never know the exact context of what happened. I think that might be a reason why we’ll never be able to reverse engineer everything perfectly. But we’re just trying to reverse engineer sufficiently to be able to improve the performance of the materials, basically.

Swyx [00:18:13]: Yeah.

Brandon [00:18:13]: What you’re asking is there a forward model where you put in an input and you can predict the output, and you want to under

Swyx [00:18:19]: Yeah.

Brandon [00:18:20]: You want to create a like a surrogate model which can understand

Liam Fedus [00:18:22]: The input is the final state of that, this configuration of atoms, and then the prediction is

Swyx [00:18:27]: Because then you can, you can split up the work almost. You can have one team do the process side, and then the other team do the property side, and then

Liam Fedus [00:18:33]: Yep

Swyx [00:18:33]: You know, just race them.

Liam Fedus [00:18:36]: Yeah, I think that there’s, a lot of progress we can make in terms of these. Splitting up these agentic workflows to these different areas, I think there’s a lot of empirical work we can make, on synthesis prediction as well.

Swyx [00:18:49]: Yeah.

Liam Fedus [00:18:49]: Whereas taking in that configuration of atoms, what tools can you use? How, given this experimental evidence, given prior literature, prior papers, how do you actually get to these things? And you ultimately are testing, are you able to actually replicate it or not?

Swyx [00:19:04]: The reason I say this maybe also is just thinking about it in terms of a company. You know, in the way that TSMC is kind of like a fab for semis, but they don’t design the semis, you could be the TSMC of materials where people come to you with like, “Well, here’s what I want,” and then you can figure out the process and

Liam Fedus [00:19:21]: Yeah, it’s like a matter compiler.

Swyx [00:19:23]: Yeah.

Liam Fedus [00:19:24]: So I guess given these requirements

Swyx [00:19:26]: Very Star Trek.

Liam Fedus [00:19:27]: Yeah. But it’s like, given these requirements, is this actually a valid set of configurations? Can this, can this exist?

Swyx [00:19:33]: Is it a good objective function? I don’t know. It’s too big.

Liam Fedus [00:19:37]: Yeah. We kind of talk about one of our missions as synthesis superintelligence, so I think it’s in line with this.

Phase Transitions, X-ray Diffraction, and RL Rewards

Swyx [00:19:44]: Yeah.

Brandon [00:19:44]: I want to double-click on something that you said a little while back, which was you talked about phases. And I think this is, going to Doğuş’s comment about, what is emergence. The concept of a phase transition I think might be foreign to a lot of the audience. So can you explain what is a phase transition, and then why is this a helpful signal for something like an RL environment?

Ekin Doğuş Çubuk [00:20:08]: Phase transitions are fascinating. I think the maybe the one that people can most easily relate to in their life is, ice melting probably, or water boiling. You increase the temperature of ice, and it still looks like ice and it behaves like ice, but there’s a certain temperature at which it just stops raising its temperature and then turns into liquid. So it suddenly goes from this solid phase to liquid phase. Another one that’s very relevant to us, of course, is the Ising model. And I think computer science and mathematicians also study this maybe under different names, but basically, you can have spins up and down, and they have different interaction terms if they’re both up versus one is up, one is down. And at high enough temperature, they’re usually just randomly up or down. Entropy wins. You lower the temperature, it’s still looking the same. But then at some point, suddenly they perfectly start aligning. Yeah, the phase transition. Phase transitions are really nice because it makes physicists’ life easier for studying certain things. They tend to have certain, spatial correlations that really help us, as well. Okay, so in our lab, phase transitions come about usually because we mix the precursors that are different, crystals basically, but then we raise the temperature or do something else to encourage them to react, and then atoms start reacting and they form into this new crystal. And the this shows up in the XRD pattern because usually if the geometry of the atoms change drastically, the XRD pattern changes a lot. But what really happens in a typical practical materials discovery campaign is when you first try something, it doesn’t work, and the outcome isn’t just a clear yes or no, but it’s usually a very mixed phase. It will usually have some of the precursors. It might have some amorphous phase, and then it will have a bunch of phases that you maybe didn’t try to make. So that becomes a challenge, and that’s why we find it very difficult to bring together simulations and AI to do the characterization. So the simulations can always say, “Oh, this phase doesn’t look like what we predicted, but this is a slight variation of it. Let me now do force field calculation on it to see if it’s still stable.” Or the AI can say, “Based on the previous experiment and this one, this is probably not the phase we want to make, so let’s change the conditions.”

Liam Fedus [00:22:18]: Yeah, and in some cases

Swyx [00:22:19]: For listeners, XRD is X-ray diffraction, which you already described.

Liam Fedus [00:22:22]: Yeah. In some cases, there’s a phase that just hasn’t been recorded before. It doesn’t exist in any papers or database. And so the AI has to actually go and sort of use these tools to say, “Well, what configuration atoms would actually explain these phases?” But this becomes incredibly important for directing the scientific process because, let’s say we have a phase target in mind. We really need to sort of hill climb that to get better measurements on that. So if it’s, one percent and we’re like, actually, we want to increase this phase purity, we want to have a very good measurement on this experimental data to say, what is present and what’s not.

Brandon [00:22:57]: So one of the advantages is you have just a distinct variable that you can just observe. Going back to the statement about how do you deal with uncertainty, one of your, it seems like part of your answer is you make your observable signature somewhat unambiguous. There’s no interpretation there. Is that right? Is that the logic behind this? Or is this literally you just needed to find a new phase, and so this is the thing you care about?

Liam Fedus [00:23:20]: I think there’s still ambiguity in a lot of these cases. A very dumb thing is replicates.

Brandon [00:23:25]: Yeah.

Liam Fedus [00:23:26]: We run replicates, of course.

Brandon [00:23:27]: Yeah. Yeah.

Liam Fedus [00:23:28]: But no, I think there — You know, it’s a challenging task because it’s not just a fully deterministic, and there are well, there are ambiguities, and other things. Two different, phases can actually be consistent with the same pattern.

Brandon [00:23:42]: Yeah.

Liam Fedus [00:23:42]: And so I think that’s why it’s really important for the system to use some chemical intuition to say “Okay, well, this would be highly unlikely given the synthesis conditions, given this, the prior knowledge.” and so that can be used to help disambiguate.

Chemical Priors, Multimodal Measurement, and Ambiguity

Swyx [00:23:58]: So you’re, so you’re somewhat injecting priors there, based on what you expect.

Ekin Doğuş Çubuk [00:24:03]: Yeah. I think thermodynamics is the biggest prior, right?

Swyx [00:24:06]: It is.

Ekin Doğuş Çubuk [00:24:07]: And then physics is a big prior. Another one that we really benefit from here is multimodality or materials characterization. So we can do XRD, and yes, some phases might look similar on the XRD pattern, but then we can also measure their electrical properties, magnetic properties. We can measure their morphology, like some kind of electron microscopy. And then some of the phases that look similar on XRD will look different in some of these dimensions. And that kind of multimodality really helps. And again, AI is really helpful here because humans have limited context and computation power as well. So if you give humans 10 different modalities at the same time and say, “Analyze these all consistently,” it’s a bit difficult. But it’s actually not super intelligent for an AI to do it. It’s great, yeah.

Liam Fedus [00:24:46]: Yeah, it’s like the super intelligence is it can just do so much calculations. It can look at everything. And this is, kind of gets to that, phrase about better decision-making under uncertainty, where now it’s stitching together, the signal across many different instruments longitudinally, across different experiments, across the replicates, and then kind of getting to this underlying better model

Swyx [00:25:08]: Yeah

Liam Fedus [00:25:08]: Rather than just over-indexing on one measurement from one instrument.

Information Limits, Telemetry, and Maxwell’s Demon

Swyx [00:25:12]: And let me just, like. So you know, you started, I think a while ago also saying that there’s just more data than you can fit in any reasonable computer or process or anything. So at some point, you have to throw out data even though you have all these replicates, even though you have presumably ten different ways to measure the same thing. It’s. ‘Cause what if one of your temperature things is wrong? So when do you throw that out? How do you decide what to do? As a data-hungry ML guy, I just want everything, right? And then I just, you know

Ekin Doğuş Çubuk [00:25:37]: So it’s, yeah.

Swyx [00:25:37]: Waddle, learn everything. I don’t know what I don’t know, and it’s just throw it, throw it, the machine at it.

Ekin Doğuş Çubuk [00:25:42]: Yeah. So we don’t throw any data. I think maybe what I meant was potentially. If God was observing the experiment

Swyx [00:25:50]: Yeah.

Ekin Doğuş Çubuk [00:25:50]: There is more data in the experiment than humans can fit in a computer. Unfortunately, us mortals cannot observe all the atoms. So even though there’s more data than we could store, we actually can’t access that data. For example, I cannot trace all the atoms’ positions. So unfortunately, we’re still in a regime where any data we collect is very precious, and we don’t delete any of it. Yeah.

Brandon [00:26:11]: Maybe a way of thinking of it to connect to the AI side is think about probing, like linear probing of a of a foundation model, right? This model has all this data, so God has this entire picture of what the material is doing, and then the experiment is this little, tiny linear probe which has eight bits or something of output. And so you’re taking this humongous thing and coarse-graining it down to a little bit of information, and then it’s now, either traditionally a human’s job or now your agent’s job to reconstruct from those eight bits of information

Liam Fedus [00:26:43]: It’s an analog, yeah.

Brandon [00:26:44]: What is actually happening inside.

Liam Fedus [00:26:45]: It’s an analog where you can’t have full visibility over every parameter weight, over every activation for that input. Yeah, you’re getting some sort of subset on it or some coarse-grained features across the whole thing.

Ekin Doğuş Çubuk [00:26:58]: Yeah. But so you’ll probably find that. You probably, do find, Maxwell’s demon argument very interesting because that kind of

Swyx [00:27:05]: Can you, can you res-recite it? Because I don’t. He’s mentioned it before. I keep forgetting it.

Ekin Doğuş Çubuk [00:27:10]: So, in the 1800s, people realized that entropy of a system has to increase or a for a closed system

Swyx [00:27:18]: Yeah.

Ekin Doğuş Çubuk [00:27:18]: Or stay the same, but it can’t go down.

Swyx [00:27:20]: Yeah.

Ekin Doğuş Çubuk [00:27:21]: It’s the second law of thermodynamics. And there was a thought experiment that was asking, what would happen if there was a very tiny, all-knowing demon that could kind of look at atoms, and then whenever an atom had low energy, kind of open a door and let it into this other room and keep doing this to reduce entropy of the system, which would break second law of thermodynamics? And for a long time, I think there wasn’t a very satisfying answer for why that was the case. But then there was a more recent paper in the 1900s, more recently to us, Landauer came up with the argument that you’d have to spend energy to delete information, and Maxwell’s demon would basically have access to so much information by doing this thing he’s doing that he’d have to delete some. And for. When he’s deleting information, he’d have to spend energy, which would then make sure entropy increases. So even though we’re not at the level where we can be Maxwell’s demon, I think your, question is very visionary towards a future where we store so much data we have to delete some.

Swyx [00:28:17]: Yeah, no, I wasn’t, I wasn’t obviously going to that, to that extent. It does remind me of air conditioning. That’s kind of how air conditioning works. But also, I think it’s just a question of why don’t you just buy every sensor in the world and just over. Ridiculously over-instrument, over-replicate everything?

Liam Fedus [00:28:33]: Oh, we do that.

Swyx [00:28:34]: Right? Yeah, okay, yeah, okay. There we go.

Liam Fedus [00:28:35]: Yes, yeah.

Swyx [00:28:35]: Okay, it works out. Right. All right.

Liam Fedus [00:28:37]: Yeah, I think, I think it’s a matter of

Swyx [00:28:39]: Confirmed.

Liam Fedus [00:28:39]: Yeah, exactly. Are you getting visibility to individual atoms, which is infeasible? But no, we add a huge amount of telemetry, and that basically reduces the number of hidden variables, if you will.

Swyx [00:28:49]: Yeah. Yeah. And just out of curiosity, you guys do you worry that you’re just in California? Don’t you want to do this in Nepal and Australia?

Ekin Doğuş Çubuk [00:28:59]: Yeah, we could. We have plans to open more labs.

Swyx [00:29:02]: Because obviously gravity and where you face and where you. Where you are matters.

Ekin Doğuş Çubuk [00:29:07]: Also expertise can change, right? Here we have certain kinds of expertise. For example, Stanford Physics Department, Engineering Department has certain kinds of expertise that really helps us because we can hire from there. Our researchers can collaborate. As you said geography can matter, but then different cities can have some differences of expertise within physics, chemistry.

Swyx [00:29:27]: Yeah. Oh, we have Zoom

Ekin Doğuş Çubuk [00:29:28]: Yeah.

Swyx [00:29:28]: For that.

Ekin Doğuş Çubuk [00:29:29]: Yeah. But still

Swyx [00:29:31]: It’s not like your experiments are going to be affected by where you are.

Liam Fedus [00:29:34]: Yeah. Yeah, the future of Periodic is not a single lab.

Swyx [00:29:38]: Yeah. You gotta head up your private jet and go to space and you know.

Liam Fedus [00:29:42]: Right. Yeah, I think it’s necessary to have this sort of network of labs. And like, yeah, I think as Doğuş was just kind of pointing out there’s an optimal place for each of these based on expertise or materials constraint. There’s many different constraints, permitting.

Connecting Simulations With Physical Experiments

Brandon [00:29:56]: Okay, so one of them is. So one thing I-I’ve been curious about is how do you deal with computational tools? Or how do you integrate computational tools into a decision-making workflow, especially when, let’s say, computation and experiment don’t necessarily, meet? It’s also theory in terms of high-level understanding. And yeah, how does that sort of work together?

Ekin Doğuş Çubuk [00:30:22]: Yeah. So, one thing that’s really helpful here is there are certain things that are easier to do in simulation and more accurate in simulation. There are certain things that are easier to do in experiment and more accurate in experiment. In general, of course, experiments are more accurate, but certain measurements are very hard to make with experiment. So you make a bad version of it, so you get inaccurate data. Let me just give you some examples. So one of them is, we use density functional theory to estimate formation enthalpy of materials and

Brandon [00:30:51]: Sorry, quick pause. Can you explain densital density functional theory and formation entropy?

Density Functional Theory and Formation Enthalpy

Ekin Doğuş Çubuk [00:30:55]: Yeah, of course.

Brandon [00:30:56]: Sorry, enthalpy, yeah.

Ekin Doğuş Çubuk [00:30:57]: Yeah. So, density functional theory is probably the most commonly used simulation method for materials. It kind of comes from Hohenberg–Kohn theorem, and Kohn got a Nobel Prize for this. He. When he realized that, sure, quantum mechanics is very expensive because you basically have this exponential Hilbert space for every wave function, so you have to solve this very difficult problem, especially as the system size grows. Kohn and Hohenberg realized that turns out you don’t need to do it at exponential space. All you have to do is think about the charge density, at least for the ground state of quantum mechanical systems. So this it’s actually the theorem is pretty simple. So even I think an undergraduate in physics can understand. But the idea is all the properties of a quantum mechanical system ground state is a functional of the charge density, which is amazing because charge density is just this three-dimensional object, whereas the Hilbert space and wave functions are exponential. So that was a really interesting observation. And then Kohn published another paper, I think with his postdoc, Sham, Kohn–Sham, wave functions, kind of had a practical way to solving some approximation of quantum mechanical properties of materials. And today, we use it very much, both in our lab and externally. And then formation enthalpy is one of the measurements we can make to understand the energy of the material. Basically, we want to ask, what is the energy that this material has? Because if that energy is higher than other materials these atoms can come into, then it’s unlikely that this material will be able to be made. So stable just means like it’s on the convex hull of other materials, and their energies. Was that clear enough, or should we.

Brandon [00:32:42]: Maybe a simple high-level point. So density functional theory is a way of taking something which is exponentially hard, in number of electrons

Ekin Doğuş Çubuk [00:32:51]: Right

Brandon [00:32:52]: Or number, in your system size, and then reducing it approximately with some introduced error to maybe usually an n-cubed approximation. So now you can compute quantities which are, you know. You can now compute things which are just uncomputable, but at some cost, right? If you could do it perfectly, you would — we wouldn’t need a lab. But we can’t. This might be a good callback to our, actually our second episode ever. Heather Kulik had said, she famously said, “There is no AlphaFold for materials, not just computationally, but the ground truth just doesn’t exist.” Like, we don’t really know what material crystal structures look like, and DFT is generally our best route towards.

Ekin Doğuş Çubuk [00:33:34]: Yeah, I think so. Maybe two things I would add is DFT by definition doesn’t have to be an approximation. The Hohenberg–Kohn theorem shows it

Brandon [00:33:42]: It doesn’t

Ekin Doğuş Çubuk [00:33:42]: It can be exact. But I think, as you’re pointing out, the exchange correlation functional we don’t yet have access to. And then for the charge density-based models, we don’t know what the kinetic energy functional is. And then even if we had perfect DFT, I still think we’d need a lab because we. Even if we had perfect DFT, we cannot fit ten to the twenty-three atoms in a computer.

Brandon [00:34:03]: Yeah.

Swyx [00:34:04]: You’re very obsessed with that number.

Brandon [00:34:06]: But. No. Well, okay. Yeah. But with exponentially growing compute, you can argue that in n-cubed scaling size, eventually you could, in principle, calculate things. You just wait till the pause long enough.

Ekin Doğuş Çubuk [00:34:17]: Oh, yeah. We need massive computers, yeah.

Brandon [00:34:19]: But yeah. But as long

Ekin Doğuş Çubuk [00:34:20]: It’ll take a while.

Brandon [00:34:20]: Yeah. But as long as n-cubed is a good scaling, you can, you can.

Swyx [00:34:24]: As the non-scientist, I will make an observation, which I don’t know if you want to throw around or anything like that. So I came from finance. We had the Gaussian copula, which was a way of pricing credit default swaps, by doing correlations and reducing everything into a single kernel. And it’s very. It feels also spiritually similar to the VAE, in terms of like, condensing all these things into a single sort of parameter. I wonder if just there’s the same trick everywhere. Yours sounds, a bit more high-dimensional than mine, which VAE is just like, one sigma thing. But sounds the same?

Liam Fedus [00:35:01]: Yeah, the charge density file is still quite a big file.

Swyx [00:35:04]: Yeah. Right. You hand wave a bit, but it’s good enough. This is, this is the highest order bit. Yeah.

Ekin Doğuş Çubuk [00:35:10]: I personally don’t know of a better method to predict the stability of a new material. It’s definitely not perfect, but it’s definitely better than other methods I can think of. And yeah, so going back to your question, the reason we use them is for certain things it’s easier to simulate than experiment, and they kind of complement each other well. And then we can do this in a loop. We feel like simulations will never be enough by themselves, but in the loop of simulations, AI, and experiments, I think we can make progress much faster than before.

When Simulations Fail: Microstructure and Calibration

Brandon [00:35:41]: One thing that simulations let you do is scale up much more quickly. Because now compute is, much easier to bring online

Ekin Doğuş Çubuk [00:35:48]: Yeah

Brandon [00:35:48]: Than bringing in bringing on new labs. How do you avoid. Or let’s say you scale up a bunch of DFT, how do you avoid in having your models over-index on that and still be grounded that the real world is ultimately the truth you care about? If you. You know, how do you sort of balance that?

Ekin Doğuş Çubuk [00:36:05]: Well, we scale up the lab as well.

Brandon [00:36:06]: Yeah, you scale up the lab.

Ekin Doğuş Çubuk [00:36:07]: Right, yes.

Brandon [00:36:07]: But you could. I could imagine scaling up DFT to millions, hundreds of millions given modern compute, whereas your lab, I would imagine, your hundreds, thousands a day? I don’t know. Something still fairly. Well, even in a really high-throughput case you’re still fairly limited by comparison, right?

Ekin Doğuş Çubuk [00:36:27]: Yeah. One thing is humans, but also LLMs, are kind of pretty aware of the limitations of DFT. For example, even if we can do, as you said, hundred million trials with DFT

Brandon [00:36:39]: Yeah

Ekin Doğuş Çubuk [00:36:39]: We could never actually try their microstructure. And microstructure is this idea that DFT usually simulates the perfect crystal, but crystals usually actually aren’t perfect, and their microstructure, which is the structure at a medium order, can really affect the properties. Other issues, of course, like some properties like superconducting temperature cannot be easily simulated by DFT. Again even if you do a hundred million trials with DFT, it won’t be enough. So we have to rely on heuristics, experiments, et cetera.

Liam Fedus [00:37:08]: And so there’s a huge filter from all those calculations to actually what gets executed in the lab.

Swyx [00:37:13]: Is that human-mediated or LLM or a mix?

Ekin Doğuş Çubuk [00:37:17]: Can be a mix.

Swyx [00:37:17]: Yeah.

Ekin Doğuş Çubuk [00:37:18]: For example, for a long time, Materials Project, which was the best open source DFT database, Use experimental calibration for DFT. So they would actually use experimental data to calibrate DFT. We also do this internally here. Whenever there’s a certain chemical system we’re interested in you get experiments, you get simulations, and you can calibrate them.

Why Inorganic Materials Differ From Biology

Brandon [00:37:37]: This is another thing. I’ve talked to a lot of biologist friends, and they are always very confused when you say materials, you can’t just do an XRD structure and actually know what the structure is. Because if you’re in the biology world, you can look at a crystal structure of a protein and typically more or less with sub A well order angstrom accuracy reconstructed. What is the difference between the materials XRD, X-ray disfor — X-ray diffu- no

Ekin Doğuş Çubuk [00:38:05]: Diffraction.

Brandon [00:38:06]: Diffraction, sorry. Experimental XRD for materials versus, let’s say, the biology world. What information do you lose, and why is this sort of like a lossy projection?

Ekin Doğuş Çubuk [00:38:16]: It’s hard not to sound crazy when answering this question. But do you not feel like, organic chemistry and biology almost stem from a lower VC dimension, like lower Kolmogorov complexity model? Where like, you can almost represent as a one D sequence, and it can almost be compiled. So there’s a sense in which it’s, simpler in the Kolmogorov complexity sense.

Swyx [00:38:38]: One D?

Ekin Doğuş Çubuk [00:38:40]: Like just DNA is just

Brandon [00:38:43]: Yeah.

Ekin Doğuş Çubuk [00:38:44]: Or RNA.

Brandon [00:38:44]: I think most people would argue it’s more two D, but

Ekin Doğuş Çubuk [00:38:47]: Okay, well, yeah. I’m not a biologist, so

Brandon [00:38:49]: Yeah.

Ekin Doğuş Çubuk [00:38:50]: I’m sure they’re right. But with inorganic chemistry, virtually it’s crazy. It’s clearly not coming from such a low Kolmogorov complexity model because that three D inorganic crystals can host metals, insulators, superconductors, diamond. It’s really crazy different. People tried, but we could never find a one D representation. It’s very hard to find a SMILES string for three D inorganic crystals. And the way the atoms interact with each other is quite strange, like covalent bonds, ionic bonds, metallic bonds. But yeah, this is, it’s very interesting. I wonder sometimes have you ever looked into issues with solid electrolytes? Why people cannot replace liquid electrolytes in batteries with solids? There are all these dendrites that form. Basically the lithium will form these structures that go into a solid electrolyte and crack it. And you look at that, and you think this is so primitive compared to biology. In biology, these systems can do such complex nano technology- technological progress, whereas we can’t even make two interfaces work with each other because lithium will break it. Yeah, but then I guess inorganic stuff can be very resistant to high temperature. You can make space shuttles. You can make silicon, compute Moore’s law. So yeah, it’s a trade-off.

Brandon [00:40:05]: Yeah, we’ve talked about this with several other guests. I think one of the big differences in my mind is that biology, we have a toolkit that you can basically borrow from millions of years of evolution

Ekin Doğuş Çubuk [00:40:19]: Yeah

Brandon [00:40:19]: That has already given us all the tools, and we can just reproduce those. But also evolution also constrains the yeah, the VC dimension to be fairly low for proteins and so on. The emergent beyond- behavior of large-scale systems like cells can be much more complicated.

Ekin Doğuş Çubuk [00:40:36]: Yeah, and consciousness.

Brandon [00:40:37]: Yeah. Yeah, but maybe even just more specifically for. If you just look at an X-ray, you had a perfect crystal, and you shoot an XRD at it, and you get the you get the spectrum. That still doesn’t actually tell you uniquely what the crystal structure is, right?

Swyx [00:40:54]: No.

Brandon [00:40:55]: So, and that’s something that

Swyx [00:40:56]: What?

Ekin Doğuş Çubuk [00:40:57]: It’s just average.

Brandon [00:40:58]: Yeah.

Ekin Doğuş Çubuk [00:40:58]: It’s just average behavior. But like, for example, disproportionation is a very complicated thing, right, where it’s possible that you have the perfect crystal, and on average, it looks like this perfect XRD. But in reality, there is some other pattern going on where atoms slightly shift to the right or left, but on average, they’re in the middle. Yeah, it’s a difficult life, yeah.

Brandon [00:41:16]: Whereas with proteins, you basically know there’s some sort of underlying protein follows some rules, so you can, you can combine a forward model with that.

Liam Fedus [00:41:23]: Right. And then you can kind of move and move towards single crystal XRD, and that can help.

Swyx [00:41:29]: When we started the Science Pod, there were all these elements of AI in science, AI in math, AI in physics, and all these things. And we there was this kind of pissing contest which is harder. And I think you can make the scientific argument that materials science is the hardest.

Liam Fedus [00:41:43]: Well, we didn’t pick it necessarily for hardness. There’s actually a lot of areas where it is easier. We have these great simulators for large classes of materials, whereas in biology, it’s incredibly difficult to model a cell or an organ or a full organism. So that’s a huge advantage.

Swyx [00:42:01]: Yeah. That, isn’t that interesting that it, at the smallest level, it is one-dimensional or two-dimensional, but and then it just has many orders more scale and structure, and that actually introduces the complexity. It’s, that’s weird. Whereas maybe complexity in materials is more the microstructure.

Liam Fedus [00:42:20]: Yeah, I still — you still get different complexity at the different, length scales for materials as well. So yeah, microstructure and beyond.

Dream Materials: Superconductors, Magnets, and Batteries

Brandon [00:42:26]: Yeah.

Swyx [00:42:27]: Yeah. Just sort of a brief sort of palate cleanser. In terms of science fiction, let’s say you can invent whatever materials that you want. What is more valuable? You know, I have, I have a list here. There’s obviously room temperature super- superconductors. But also, you mentioned batteries. I’ve always thought batteries. Just do better than lithium-ion, you’re good. That’s stood around for a hundred years. And then the last one is carbon nanotubes, mostly for the space elevators. I don’t know if, like. Those are, those are the three that come to mind. Are there any that people talk about in your world that are the dream?

Ekin Doğuş Çubuk [00:43:03]: Definitely superconductors

Brandon [00:43:05]: Yes

Ekin Doğuş Çubuk [00:43:05]: Magnets.

Brandon [00:43:06]: Right.

Ekin Doğuş Çubuk [00:43:06]: Especially if we can lower the dependence on rare earth or transition metals that are hard to source, like cobalt, reducing cobalt in batteries. One of the things I think is very fundamentally important is, if we can get close to Landauer limit for compute energy efficiency.

The Landauer Limit, Reversible Computing, and Energy

Brandon [00:43:24]: Can you, explain Landauer limit?

Ekin Doğuş Çubuk [00:43:26]: Yeah. So if you look at how much energy we spend per, like FLOP or compute, it has followed the Moore’s Law type behavior, but it’s gotten exponentially more efficient. So we’ve been spending exponentially less energy per FLOP or compute. I haven’t checked recently. This was the case ten years ago when I last

Brandon [00:43:44]: Yeah, that’s great. I remember when I first learned about it, looked at that, and we were, what ten, twenty orders of magnitude away from it or something.

Liam Fedus [00:43:50]: Beyond the limit, yes, right.

Brandon [00:43:50]: And the last time I looked, it’s. We’re weirdly close. We’re within a factor of I don’t know three or four orders of magnitude away.

Ekin Doğuş Çubuk [00:43:59]: Yeah.

Brandon [00:43:59]: It’s actually like

Liam Fedus [00:44:00]: I just exponentially rolled out over many years.

Brandon [00:44:01]: Yeah, exponentially over many years are

Ekin Doğuş Çubuk [00:44:03]: Four or five, yeah.

Brandon [00:44:04]: Yeah, pretty impressive.

Swyx [00:44:05]: But in your lifetime, you

Brandon [00:44:06]: Yeah, no I. Yeah, since, just since I discovered this, what, ten or fifteen years ago? Yeah, no.

Ekin Doğuş Çubuk [00:44:11]: And one of the things we have to do is dissipate heat, because as these, computations are happening, it’s producing heat, which has to happen, but then we have to dissipate it. So if we can get close to the Landauer limit for how efficiently we do computation, that’s amazing for humanity, right? Because that means now we’re doing computation, which is probably one of the most fundamental things we do as humanity, but as efficiently as possible from energy perspective, which is a real occurrence in the universe.

Swyx [00:44:34]: That get affected by quantum computing? I’m just going to throw it out there. Theoretically, massively, embarrassingly parallel compute.

Ekin Doğuş Çubuk [00:44:41]: So it definitely gets affected by reversible computing. But honestly, I’m not an expert. I don’t understand. So if you can do reversible computing now, you’re not. You don’t actually have to spend energy to do computation because you’re not actually deleting any information. It’s reversible. But I don’t know if that works. I don’t know. Quantum computing, I assume, has to obey these laws somehow because they are still, like. They have to obey thermodynamic

Brandon [00:45:06]: It’s unitary, so

Ekin Doğuş Çubuk [00:45:07]: Yeah, exactly

Brandon [00:45:07]: It’s reversible, so for now.

Swyx [00:45:10]: One of my most memorable conversations with Elad Gil was he actually was a very skeptical person about quantum computing.

Ekin Doğuş Çubuk [00:45:15]: I see.

Swyx [00:45:15]: He’s like, he’s like, “Even if you had it today, there’s no applications.” I’m like, “Whoa.”

Ekin Doğuş Çubuk [00:45:19]: I

Swyx [00:45:19]: It breaks, it breaks the it breaks the security RSA. That’s it.

Ekin Doğuş Çubuk [00:45:22]: I agree with that. Like for them, people talk about how if you had a quantum computer today, you could simulate things so much better. And I asked them, “Okay, let’s accept that. What would you simulate?” Like, you’re still simulating perfect crystal. You still have the issues that DFT has if DFT was perfect in its prediction ability. So yeah, I do think people are kind of glossing over some things because quantum computing is so exciting. If we can compute in a quantum logic space instead of classical logic, it’s just so exciting that people are glossing over what it would actually do when it’s made. Maybe that’s fine because maybe once it’s made, it will do amazing things we can’t even imagine.

Putting Intelligence Into Every Laboratory Instrument

Swyx [00:45:57]: Good. Okay. I was going to move to the automating the lab side, where you’ve closed the loop on your experimentation. You can talk about robotic arms, talk about. But it’s basically just everything you’ve done here. My favorite quote from you was that every piece of equipment in your lab is going to have a hundred and forty IQ. Okay, what does that mean?

Liam Fedus [00:46:17]: In the early days, as we were scaling things up, we realized that we had huge bottlenecks imposed just by operating machinery. So, for one set of machinery, we would have technicians and scientists looking for particular morpho- morphology and trying to see, okay, what actually were we making, see if this is consistent with our intentions. And we really quickly, as we scaled up the lab, came into these bottlenecks where it just wasn’t keeping up. So we started doing some programmatic approaches to capturing the data on SEM. And the data that was being surfaced from this very simple program where it sort of takes a field of view, captures, zooms in captures, was just not all that useful. It was a little too dumb for what we really wanted to get. And so at that point, it was then, really pertinent to build AI systems directly onto the machines to start controlling these things. And now they have the full context as to what we were trying to achieve. So what was the intent of the experiment? What were we trying to synthesize? What were some other experimental evidence? And it’s actually looking, in the machine and capturing that data. And this is really valuable now because we’ve been talking about these hidden variables not being able to capture everything, but if you can do more intelligent data capture at the time of that experiment, your data for future AI systems and future computational predictions is that much better. You’ll have just a richer set of data. And so that’s sort of what we mean by just, everyth- everything on the lab has to be incredibly intelligent to just make the data as useful as possible. So that was, that was sort of the motivation and inspiration behind that.

Swyx [00:47:57]: What’s the state of the art there in terms of putting intelligence on every device, right?

Liam Fedus [00:48:02]: Well, I think one interesting aspect, is what’s the latency

Swyx [00:48:08]: Exactly

Liam Fedus [00:48:09]: Of controlling that instrument?

Swyx [00:48:10]: Because there’s cloud latency, but then there’s also device compute, which

Liam Fedus [00:48:13]: That’s right

Swyx [00:48:13]: Also.

Liam Fedus [00:48:14]: And then there’s sort of a time scale associated with different physical processes. And if calling out to an API or something is simply too slow, so the amount of reasoning or tokens or tool calls, is just not matched to the latency of that actual process, then that’s infeasible.

Swyx [00:48:33]: Okay. That is right.

Liam Fedus [00:48:34]: It’s almost like, self-driving cars, right? So.

Brandon [00:48:37]: I. Real quick. I just am actually surprised to hear that because the latency I’d imagine for reasoning seems small compared to a lot of these things take hours to run, right? Or maybe, you know. Or maybe I. Maybe your experiments are much faster or high-throughput or something. I’m just surprised to hear that, actually.

Liam Fedus [00:48:52]: So ultimately, the experiments do take, many hours, days, but there could be particular steps where latency would matter.

Ekin Doğuş Çubuk [00:49:00]: Yeah. You might want to have finer control. Another thing is, how expensive it is because the same reason some of these models can be really slow in analyzing data also makes them very expensive. And then finally, the human patience. If a human wants the analysis of a certain XRD pattern and they have to wait two hours before they get a good result, that’s very different, I think, than if they can get it in two minutes.

Brandon [00:49:25]: Oh, okay.

Liam Fedus [00:49:26]: Yeah. And then the reasoning, it’s. You know, you’re not doing a single call to that model, right? You’re doing, potentially many tokens, to kind of get to these patterns.

Ekin Doğuş Çubuk [00:49:35]: Run simulations in the loop.

Liam Fedus [00:49:37]: Exactly.

Ekin Doğuş Çubuk [00:49:38]: Deep research in the

Liam Fedus [00:49:39]: It’s not a single tool call. It’s not a single inference thread. So then the latency can blow up quite a bit.

Brandon [00:49:44]: I see what you mean.

Liam Fedus [00:49:45]: Yeah.

Brandon [00:49:46]: How much of your latency is tool calls? And I assume tool calls is largely DFT or computational- … and therefore. Yeah. So how much of your latency is derived from those versus actual reasoning?.

Liam Fedus [00:49:59]: It’s really process dependent.

Pragmatic Robotics, Automation, and Lab Reliability

Brandon [00:50:01]: Okay. Yeah.

Liam Fedus [00:50:01]: Yeah.

Swyx [00:50:02]: And then the other thing I think about is also, I guess, building up from small things to bigger things, where I assume that the general temptation or the typical development is incremental, where everything is human, operated, and then you find ways in which to automate it, and then you sort of build up from there. I worry that sometimes that is the way that people evolve things, but that’s a local minima, optima.

Ekin Doğuş Çubuk [00:50:25]: Like short-horizon optimization.

Liam Fedus [00:50:26]: Right.

Swyx [00:50:26]: Exactly. When actually you should get a humanoid and just put them in there.

Liam Fedus [00:50:31]: We. I think we’re of the opinion that solving humanoids would actually be slower to kind of getting to some of our goals.

Swyx [00:50:37]: Just checking. Again a lot of this is just like you do this every day. We, like

Liam Fedus [00:50:41]: Yes

Swyx [00:50:41]: See this, but we. I don’t know the reality of the situation. A lot of people having humanoids.

Liam Fedus [00:50:46]: Yeah. I think maybe one process is, by having a mix of humans and automation, you can identify really quickly what are some of the bottlenecks in the experimental process, and you start alleviating those bottlenecks one at a time. And so for example, if there’s some really tricky dexterity task that humans are excellent at but you’d have to spend months automating machine, maybe don’t spend a ton of time there. And maybe a more promising thing is something very routine that takes up a huge amount of scientists’ and technicians’ time, and it’s easy to automate. So I think it’there’s this pragmatism to it. But ultimately what we want from the lab is a huge quantity of data, high-quality data, diverse data, and those are our goals. And full autonomy is a non-goal. It’s sort of in the service that we use automation in the service of achieving the goals on the data.

Brandon [00:51:39]: Yeah.

Liam Fedus [00:51:40]: But also, I think another aspect to automation is, robots will just be. Make fewer mistakes potentially in the lab. And it’s been really helpful having AI systems with a full view over all of our data, because sometimes if there’s a permutation in data, then it can identify it. So for example, one of our steps at one point had, a cyclic error because one of the machines was loaded incorrectly, and the patterns were inconsistent. So the AI was reading through these things and says, “Well, given what was run, this is not expected.” And it’s looking at this basket of data longitudinally, and then it realized, “If I do this cyclic permutation and reverse it, everything is consistent.” And then we were able to go back to the physical infrastructure and understand that a mistake had been made in the loading. And I think managing data quality is so foundational to doing AI in the physical world. And these are the types of things that we’re building in. And so you can improve the operating process, but at that point you’re like, “Okay, this is another great opportunity for automation. How do we make this just so reliable, so durable that we never have those types of mistakes again?”

Ekin Doğuş Çubuk [00:52:51]: Yeah, the reducing the noise floor of experimental data I think is very important. Most experimental data in the literature have such a high noise floor that it’s actually usually worse than DFT accuracy. Because, when different people do experiments, different labs, different parts of the world, different times, it really introduces so much fluctuation to the result. So we are hoping that by standardizing these workflows, one of the biggest benefits will be that the noise floor will be lowered.

Open Models, Proprietary Data, and Compute Efficiency

Swyx [00:53:16]: You’ve been pretty public about how you use open source models and fine-tune them for stuff like this. Is it better. Is it basically. I can imagine a situation where, you have maybe the dumber models do one task, and then you optimize for that task, and then sort of like a generalist frontier model that supervises everything. Is that a good mental model to have? Are there more stages to this that I can think about?

Liam Fedus [00:53:37]: We basically use a mix of open source models and closed source models. You know, we’re huge beneficiaries of this from our simulation side, building up the ML stack, the simulation stack. We don’t need to push on that axis. But there’s a lot of areas where, the latency is too high, the cost is too high. But in some cases we can actually. You know, by having access to this data, we can actually push beyond the frontier of what these systems can do under even some the highest reasoning efforts. And so it’s not necessarily just about cost or Pareto efficiency, but in some cases you can push beyond it when you have access to data that no one else has. One, way to characterize this is in terms of compute efficiencies. So by having access to this data, you can be that much more compute efficient compared to some of the frontier models. And that’s been able. Been a really instrumental thing in our program.

Swyx [00:54:33]: Yeah. I’m so. Data quality or data access as a trade-off of compute I think there’s some amount of exchange rate of dollars for compute, for dollars for data that people. I feel like the pendulum might be swinging now towards data. Obviously you would, you would agree with that, but like.

Liam Fedus [00:54:50]: Yeah, especially like, talking about the noise floor of experimental data. We need to have high-quality data. If you throw a huge amount of compute against a bunch of noise, you’re not going to have a good thing emerge.

Swyx [00:55:00]: And it’s harder for you because you actively try to be sparse. You actually try to get null results and learn from that and

Liam Fedus [00:55:06]: That’s right.

Swyx [00:55:06]: Yeah.

Negative Results and Learning From Failed Experiments

Brandon [00:55:07]: Yeah, you’ve Yeah, you’ve talked about null results in several other venues. What does a null result look like for materials, and how do you use that effectively? Especially when I think your overall signal is probably quite sparse usually in terms of success.

Liam Fedus [00:55:19]: A null result could be we intended to produce some structure, and then all of the evidence points to us not producing that structure.

Brandon [00:55:28]: Do you ever intend to not produce a structure? Of course.

Liam Fedus [00:55:30]: Yeah.

Brandon [00:55:30]: Okay.

Liam Fedus [00:55:31]: Absolutely.

Brandon [00:55:31]: Okay. Cool. You do the negative controls

Liam Fedus [00:55:34]: Absolutely, yes.

Brandon [00:55:34]: Regularly. Okay.

Ekin Doğuş Çubuk [00:55:35]: There are impurity cases that will kill your property. Yeah.

Brandon [00:55:39]: Okay.

Ekin Doğuş Çubuk [00:55:40]: Can be toxic.

Liam Fedus [00:55:41]: Yeah, exactly. 100%.

Brandon [00:55:43]: Okay.

Liam Fedus [00:55:43]: And a negative result, in some cases is actually. You know, we intended to make this, the following thing, and then we’re actually able to identify a new structure that hadn’t previously been identified. So it’s. Negative in some respect. We intended to do something else, but something else emerged from the data.

Ekin Doğuş Çubuk [00:56:02]: Yeah. And also in general when you’re training a machine learning algorithm, especially like at a very basic level, if it’s a classification algorithm, if you don’t have negative samples, you can’t really train if everything is positive.

Brandon [00:56:13]: Yeah. Yeah.

Ekin Doğuş Çubuk [00:56:13]: And this is a particularly bad problem in material science because people usually publish crystals they could synthesize, but they usually don’t publish if they fail to synthesize a crystal. Sometimes they might. And also we never know really if a crystal is not synthesizable ever, right?

Brandon [00:56:30]: It could just be a skill issue.

Ekin Doğuş Çubuk [00:56:31]: Exactly.

Brandon [00:56:31]: Yeah. Yeah.

Ekin Doğuş Çubuk [00:56:31]: It could be a skill issue.

Brandon [00:56:32]: Especially with the literature.

Ekin Doğuş Çubuk [00:56:33]: Synthesis method issue, technology. Yeah. So, I think it really helps us when we do our own experiments and get negative results in the context of what we tried, so then we can even train a classification algorithm.

Liam Fedus [00:56:46]: I think there’s also another valid thing too

Brandon [00:56:48]: Yeah

Liam Fedus [00:56:49]: Of when you have this sort of string of negative results and then finally through process iteration, you’re able to get to that positive result, it’s a really interesting set of I’ll call it process engineering-type data. So it’s like through iteration, how did you actually get to that correct result? Because so much of material science has this ambiguity in actually how something was made. And so, yeah. And so there’s actually. Even if there’s a known material, it can be highly non-trivial to replicate that. Some things are, in like, a high school textbook, but other things are really at the frontier and we’re building up that know-how as well. And then building up a system that, given this string of negative results, how do you actually get to that

Brandon [00:57:35]: Yeah

Liam Fedus [00:57:35]: Positive case?

Brandon [00:57:36]: And going to the classifier result, I would almost assume that if you are doing everything in-house, most of your results will actually be negative rather than positive, which is kind of ironic because the literature only gives you positive results. So if you are sort of cold starting this, it seems like the problem is actually the reverse, that you

Ekin Doğuş Çubuk [00:57:52]: Yeah

Brandon [00:57:52]: Have an abundance of negative results and not enough positive.

Ekin Doğuş Çubuk [00:57:56]: And as you said if we were doing per experiment labels, it would be mostly negative.

Brandon [00:57:59]: Yeah.

Ekin Doğuş Çubuk [00:57:59]: But if we were doing per campaign and assuming campaigns end when we succeed, then it could be a bit more balanced.

Training on the Process of Science, Not Just Its Results

Brandon [00:58:08]: I see. So your reasoning traces could almost be over entire campaigns

Liam Fedus [00:58:12]: Absolutely.

Brandon [00:58:12]: And not. Okay.

Liam Fedus [00:58:13]: This is, this is the key thing. This type of data basically doesn’t exist anywhere else, and we spend so much of our time getting the full lineage of the scientific process into the model. And so tracking all this data, like the conversations, the intuitions, what was executed in the lab, where the computations run, where was the code written, stitching all this together is so valuable. And I think the overall goal is rather than training on the final output of science, you’re training on the process of doing science.

Brandon [00:58:44]: This feels very much like you. What you would try to do with RSI, where you are having a model trained to train better models. You’re now having a model trained to make better experiments. But

Liam Fedus [00:58:56]: Absolutely

Brandon [00:58:56]: The difference is, it’s not going back into the core model for. So it’s

Ekin Doğuş Çubuk [00:59:00]: Unless the physics models you’re developing are improving chips, which then improve AI. So this is the wider

Brandon [00:59:07]: The bigger. Yeah, the higher level RSI.

Liam Fedus [00:59:09]: That’s the big data flywheel. Yeah.

Brandon [00:59:10]: But does that mean that. Is it plausible that in Fable 6 or something, or GPT-8 could just have a. Which has been tuned on these sort of higher order, reasoning traces, could actually just do this without any

Liam Fedus [00:59:30]: Mm

Brandon [00:59:30]: Logic? Because it’s sort of the same thinking process, right?

Liam Fedus [00:59:33]: Yeah. I think there’s, decision-making under uncertainty. Obviously in machine learning, there’s noise, in running the AI loops. But we do think there’s a different set of challenges when you’re actually interfacing with the physical world. But also I think there’s another piece too, which is getting the compression of everything into weights is still really valuable. If inference time reasoning was sufficient, all of the frontier labs would have stopped training at GPT-4 and were like, “Okay, every. From now on out, we’re going to get really good at inference time improvements.” and so we think that by. You know, because of the differences between, physical sciences and machine learning, getting that compressed into our own weights will lead to different types of systems and different types of capabilities.

Ekin Doğuş Çubuk [01:00:20]: But also, we feel like even if Fable 7 gets really good at. Even better than what it is today, it will still have to run experiments to get results. And the reason for it is, right machine learning is really good at what it’s been trained on but scientific discovery is almost by definition what you haven’t been trained on. And that’s why we’re building these labs, so that, whether open models or closed models can use these labs to tinker with the universe, because we don’t feel like you can make a big discovery without trying things.

Liam Fedus [01:00:50]: Yeah, there’s

Brandon [01:00:51]: Yeah

Liam Fedus [01:00:51]: There’s not going to be like. No one’s going to zero shot the room-temperature superconductor.

Ekin Doğuş Çubuk [01:00:54]: Think theoretically

Brandon [01:00:55]: That would be pretty cool. Yeah. As someone who has worked closely with wet labs before, certainly I am more skeptical about zero shotting scientific results than some.

Liam Fedus [01:01:05]: Yeah.

Brandon [01:01:05]: But it is something that I think, some people might ask, so.

Liam Fedus [01:01:08]: Yeah. I think it’s really important to distinguish the results we see in math and theoretical physics from the physical world. Right? Like

Brandon [01:01:18]: Yeah

Liam Fedus [01:01:18]: These are two different things.

Brandon [01:01:19]: Two very different things.

Liam Fedus [01:01:20]: Yes.

Ekin Doğuş Çubuk [01:01:20]: I’m not even theoretical physics. So far it’s been theoretical computer science

Liam Fedus [01:01:22]: Yes

Ekin Doğuş Çubuk [01:01:22]: Coding and math.

Liam Fedus [01:01:23]: Right.

Brandon [01:01:24]: Yeah.

Ekin Doğuş Çubuk [01:01:25]: Maybe theoretical physics is next.

Liam Fedus [01:01:26]: Don’t come. Yeah.

Brandon [01:01:28]: Yeah.

Levels of Scientific Abstraction and the Human Role

Ekin Doğuş Çubuk [01:01:29]: Oh, yeah.

Swyx [01:01:29]: Can I get a sort of mental model of the layers, levels of extraction that you can go? So for example, in terms of automation, and I’m just on this theme again where we talked about the campaign, talked about individual essays and running tests and how they’re. Humans are bad at it, so we should stop humans from doing it. Are there others? So for example, one level higher than campaign could be, a physics theory that you’re testing? Or one level lower than campaign is what? And like, I just. I like to think about it from that point of view and Think about it from like, okay, well, this is an API call now. You never have to touch this again. And this one is. Yeah, no, this is still ninety percent human. And maybe we can sort of draw the map of the territory that way. Is there — are there other levels?

Ekin Doğuş Çubuk [01:02:17]: I can tell you some levels. I feel like you asked a good question. I don’t have a very systematic answer, but let’s talk about some levels. So one level is the atomistic structure. There’s an abstraction, right? There’s no perfect atomistic structure in anything we do, but it’s one approximation. Another one is the continuum model. So this one is not atomistic anymore, but it’s, a continuous mesh representing a material.

Swyx [01:02:39]: Okay.

Ekin Doğuş Çubuk [01:02:39]: And then in a different dimension, there’s the thermodynamics. Assuming, you do this experiment forever, what would be the final state? And then there’s another layer of abstraction, which is kinetics, acknowledging that we’re not doing this experiment forever, so time matters. So how quickly will a reaction happen or not, even if it’s lower energy or not? So thermodynamics, kinetics, atomistic, continuum. What else? Are there other levels of abstraction?

Liam Fedus [01:03:04]: Well, I think your point about okay, there could be an overarching new theoretical advance that informs multiple campaigns right?

Swyx [01:03:14]: It makes a difference. Kind of like as an investor, I want to go here’s the bottleneck, guys. And when we get this, we get everything else. I don’t know.

Ekin Doğuş Çubuk [01:03:22]: To me, from the beginning, our hypothesis has been that one of the big bottlenecks is automated characterization, because it’s not that hard to mix powders to get to try stuff. But if you can’t characterize and analyze it and then decide what the next step should be intelligently, you don’t really benefit much from mixing powders randomly. So we focused a lot on automated characterization, closing the loop with simulations, because that’s how you try a lot of things, make an informed decision, and then decide what to do the next day.

Brandon [01:03:51]: Maybe as a follow-up question, what is how do humans live in this loop? Where at what points do you pull drop the human in? I’m guessing you probably

Swyx [01:04:02]: You could, you could randomly drop in at any part

Ekin Doğuş Çubuk [01:04:04]: Of course.

Swyx [01:04:04]: And add value.

Liam Fedus [01:04:05]: Yes.

Swyx [01:04:05]: But where do you decide to spend your time?

Brandon [01:04:08]: It seems like it might be all levels. You have you have people in the lab who are physically moving materials, and then you also have scientists who are guiding decisions about what campaign, and then you have humans who are. What’s that? I’m just asking a question for you. Yeah.

Liam Fedus [01:04:21]: Yeah. One hundred percent. I think it’s always shifting, too.

Brandon [01:04:23]: Yeah.

Liam Fedus [01:04:24]: So, I think even just using the guidance of campaigns, in the early days, that was fully human-driven. Now, increasingly, it’s a mix of AI-driven and human-driven. And then Doğuş was saying early on we identified that characterization was a big bottleneck. In the early days, it was human-driven, and the scientists were very much overwhelmed, and the balance has shifted significantly towards AI-driven. And this has allowed the scientists who were previously spending all of their time doing these refinements to now elevate their work to something else. But I think we’re. It’s really sort of a function of time. There’it’s always changing at each of these levels.

Search, Optimization, and Design of Experiments

Swyx [01:05:04]: Yeah. In terms of just general search, are there things that perform well in let’s say, the physical world that don’t perform well in the other worlds or vice versa? Like so evolutionary search is pretty popular in LLMs. I don’t imagine it works well here.

Ekin Doğuş Çubuk [01:05:21]: So especially in simulations, people have been using evolutionary search, for a long time. It’s pretty good at structure prediction, for example, finding low energy structures.

Swyx [01:05:31]: Okay, so it works. Yeah.

Ekin Doğuş Çubuk [01:05:33]: And then process engineers in semiconductor industry will do DOE, design of experiments, and they’ll often use some kind of zeroth order whether it’s Bayesian optimization or evolutionary search or. I don’t think they use RL, but similar idea, yeah.

Swyx [01:05:47]: You have to have some kind of algorithm

Ekin Doğuş Çubuk [01:05:48]: Yeah

Swyx [01:05:49]: To just figure it out.

Ekin Doğuş Çubuk [01:05:50]: And all zeroth order optimization behaves the same anyway, right? So.

Scaling Labs and Reusing Data for Reinforcement Learning

Brandon [01:05:54]: Yeah. Okay, I want to go back to the question I had earlier, which was one of scaling. And so I-I’m curious, what is your vision for scaling the lab? So there’s different ways. Coming from the biology world there’s a lot of cool tricks you can do in biology. You can tag things, you can use DNA sequencing for all sorts of interesting readouts. And if you can map your complicated assay onto sequencing, you can now scale to millions or hundreds of millions or something. Can you do tricks like this? Ultimately, every assay has its own runtime. I’ll, I’I have a favorite blog post I would want to. I’ll quote in the show notes. But what does the runtimes look like for these labs? Is it. Is your scaling just for mater-materials just fundamentally linear in terms of you do you just need more resources to do more things, or can you paralyze things in clever ways and combine things, and are there tricks there?

Liam Fedus [01:06:52]: Well, I think one way to think about it is, the RL environment is not quite trivially run the experiment, wait for that agent to fully roll out, doing the calculations, going through all the instruments to the final outcome. We basically start producing data across all the different instruments and all these different campaigns. And one way you can begin to expand that for machine learning is now reinforcement learning environments can be constructed based on again dates for the experimental campaigns or computational campaigns at that point. Or you can also, start taking, subsets of instruments, and you’re like, “Okay, we’re going to create reinforcement learning environments on the basis of this instrument alone.” Or maybe it’s, across these different instruments to kind of get to some reward state. So that’s a way where. This finite basket of data, when viewed in different ways, can then be expanded, for training purposes. So it doesn’t get at the actual experiment run, but from an ML perspective.

Ekin Doğuş Çubuk [01:07:55]: Maybe I can give you some analogs for the bio examples you gave. So, one thing people have tried is combinatorial sputtering. I don’t know if you’ve ever heard of this. You have these sputtering targets, and you create this compositional gradient. So when you look at the final product, because there is gradient from each precursor, it creates different crystals at different spatial locations. So in one go, you maybe try hundreds, thousands of crystals. That’s an that’s similar, right, to your example.

Liam Fedus [01:08:23]: Oh, yeah, that makes sense.

Ekin Doğuş Çubuk [01:08:24]: And as far as I know, this hasn’t worked very well so far because it turns out there is diffusivity, so things move around. Another one that we thought about, but we haven’t done yet, but a lot of people talk about this is superconductivity measurements are a bottleneck because unlike XRD, there’s no high-throughput superconductivity measurement, and it can take one hour per measurement. So people have thought about taking the different candidates, mixing them all into one sample, and putting it in. And if, it has superconductivity, you know that one of the sources that. And then you can do a bit like how people are doing COVID testing to. So there are ideas like this exist. Some of them work well, some of them don’t. Yeah.

Custom Hardware, New Labs, and Scaling Ambition

Brandon [01:09:00]: So you, I think, are about to announce a big fundraise. How are you thinking about, scaling?

Liam Fedus [01:09:06]: We’ve been doing our design of our labs for a while, and so I think the resources allow us to continue to scale up the labs significantly, but also the compute, both from the AI side and the computational side. But I think really importantly too, it’s what are the new types of labs we’re able to build? So I think we’ve been proving out this loop, in our initial labs here in Menlo Park, but we’ll be continuing to expand to new types of labs and repeat the same process.

Ekin Doğuş Çubuk [01:09:33]: Yeah, like the kind of lab we’re trying to build, I think haven’t been built before at this scale or at this approach. So we have learned a lot from our own build-ups so far, and those lessons guide us for the next version and the next version. And then we can, as you said increase the scale, the ambition, the quality of instruments. Some instruments are very expensive. As we get a really good understand for what we get a lot of return from we can invest into those more.

Liam Fedus [01:09:58]: And it makes. Hardware engineering has become so core to it. So for example, in the early days, just for speed, we’d buy some off-the-shelf instruments. And as we kind of push the scale of the lab

Brandon [01:10:08]: You make your own.

Liam Fedus [01:10:09]: We have to make our own. So we realize, okay, for this instrument, it’s pretty fast. These components are actually pretty quick, but this weighing machine actually becomes the bottleneck. Or there can be things pertinent to data quality where, well, the resting position of this robotic arm actually is above the plate of previously mixed things, so there’s a contamination risk. So in our hardware design, we’re going to redesign it so that the resting position lies away from the-those things. These are the kind of subtle details that allow us to kind of push the noise floor down for our experimental campaign. So, basically take these learnings and scale it up.

Assembling the Team Behind Synthesis Superintelligence

Brandon [01:10:45]: You ship your org chart. You have many teams. I think the. We were surprised when we looked at your jobs page, and we were like, “Okay, we don’t think we fully map what Periodic does.”

Liam Fedus [01:10:56]: Maybe some organizing principles, so to kind of make sense of the job page. So we hire for AI research and infrastructure, computational, experimental roles, and then hardware engineering roles, and then product roles.

Ekin Doğuş Çubuk [01:11:10]: We want to achieve synthesis superintelligence, right? And we feel like for that, we need the right chemistry expertise, the right physics expertise, simulation theory, and thin films, powder. So hardware engineering these seem like very diverse roles, which they are, but they’re all actually, coherent in what they’re trying to do, which is a synthesis for intelligence. And this also applies to the LLM researchers, the infra, even the product engineers, because the product engineers kind of make sure all the research gets made into a product that the experimentals in the lab can use.

Liam Fedus [01:11:42]: Yeah. So in order to do the end-to-end loop with the physical world, it really requires your ability to have agency over the physical world to build these labs, to run the campaigns effectively, to create AI systems against that, to make them good users of the computational tools. That’s like the difficulty, but also the opportunity of Periodic. It’s like this group of people has just never been brought together before. It’s irreducibly a multidisciplinary problem.

Ekin Doğuş Çubuk [01:12:07]: Yeah. And one other guiding principle for us that is really important for us is we want the people who are very good at what they do and experienced to do it hands-on. You know the modern life has gotten us to this place where, especially in academia, when someone is really good at research in some area, we give them so much responsibility for grant writing, teaching, that they stop having the time to do research in that area hands-on. But if you look back at Bell Labs, IBM, institutions that made really good progress, it was really experienced people doing hands-on work. Like Bardeen was in the lab every day, even though he’s a theorist. Alex Müller was in the lab doing experiments, even though he was the lab lead. So we try to do that here too. So we have some of the world’s leaders in different fields, but they’re doing hands-on work.

Brandon [01:12:52]: Do you want to brag or just call them out? You know.

Ekin Doğuş Çubuk [01:12:54]: Oh, I would love to yeah. So, when I was doing my PhD, my favorite computational material scientist who’s kind of from my age group is Muratan Aykol. He did his PhD with one of our advisors, Christopher Wolverton. And he is kind of like our computational material science expert. He’s incredible. He’s sometimes, a really good experimentalist, even though he does simulations. He understands systems really well. Our lab lead is Joe Checkelsky, who is a professor at MIT, but he’s on leave to work with us full-time. And Joe is this incredible physicist, understands superconductivity really well, but he also was a professor in Japan for a bit, so learned synthesis and chemistry really well, especially for a physicist. We have Daniel Chica, who was a graduate student in Mercury Canasides’ group, which is maybe the best solid-state chemistry group in the world. And Daniel was one of his synthesis experts, so he’s like a magician with his fingers. There’s so many examples like this.

Liam Fedus [01:13:50]: No, I know. I think Dima Bahdanau, so he’s been leading a lot of our AI and LLM efforts. He was the inventor of neural network attention.

Brandon [01:13:59]: Oh, Bahdanau. The Bahdanau attention.

Liam Fedus [01:14:01]: Exactly. Yeah.

Brandon [01:14:03]: Yeah.

Liam Fedus [01:14:03]: Yeah. Yeah, so very hands-on in sort of getting into the traces, the training distribution, how to connect it to the lab.

Ekin Doğuş Çubuk [01:14:10]: The chemistry.

Brandon [01:14:11]: Chemistry.

Ekin Doğuş Çubuk [01:14:11]: He’ll go deep.

Liam Fedus [01:14:12]: He was very deep. He actually has solid-state synthesis books at his lab in Montreal. Ray Nakano, so he used to tech lead some of the operator work at OpenAI. He’s in the lab. So when we were doing some lab tours, we were initially, we’re like, “Oh, this new, scientist, this new technician really looks like Ray.” And no, it was literally this was Ray actually in the lab, with the scientists automating the machines. And it’s like this is sort of the DNA and the type of people we need to kind of pull this off.

Forward-Deployed Engineering for Semiconductor Partners

Swyx [01:14:43]: So, we also saw that you have forward-deployed engineering, roles. We happen to be spinning up a forward-deployed engineering podcast because there’s so many engineers that would do that role. They may not necessarily know that there even is a role for them at Periodic. What is it, and how do they work with customers?

Liam Fedus [01:15:01]: So we’ve been building our own products, our own tools, our own AI and computational

Swyx [01:15:06]: For yourself. Yeah.

Liam Fedus [01:15:06]: For ourselves. So we’ve been customer zero. Now what we’re doing is we’re taking these same tools to different industries, and we’we have a huge focus right now in the semiconductor industry. However, in order to do this these are very, private companies. We need to be able to operate within some of the most secure environments. And so our Periodic forward-deployed engineers and researchers will actually be on site working and actually integrating our AI systems, our computational systems to help our partners get to their goal states more quickly. So basically the tools, the know-how that we’ve built in doing our own materials discovery, materials engineering, we’re now bringing to industry. And we think this is the most effective way to get partners to their end states and to their goals, rather than just throwing some technology over the wall and telling them to figure it out. Then these forward-deployed engineers will also do the inference locally, but then also can, train on the data. So again, once we take our system that understands these different areas and we deploy, we can make it expert on the customers or the partners’ data so that they can own their own intelligence and have systems that understand this much more effectively than just hitting some API or some untrained model. And so this is a much expanded, forward-deployment engineering role than typical because this really requires some very deep, machine learning expertise. You have to be incredibly precise from a data training perspective, infrastructure perspective, but also a physics perspective. So we’re working with highly technical industries. They have to understand, different areas of chemistry, material science, some of the devices. And so that, those are the roles we’re building out right now.

Swyx [01:16:53]: Part of it is willingness to spend extended periods on site in Taiwan. It’s like, okay, well, I know that kind of customer.

Swyx [01:17:01]: It’s interesting that you’re monetizing. That is the primary way that you’re monetizing now, but it might change in the future. But like, that’s like a model that I think people don’t really get about APL Research Lab. It’s. And when you compare yourself to Bell Labs, that’s what people are thinking, right? Which is.

Commercializing Scientific Discovery

Liam Fedus [01:17:19]: I think a really great analog would be software engineering.

Swyx [01:17:22]: Right.

Liam Fedus [01:17:22]: So in the early days of software engineering we had GitHub Copilot, then we had early versions of ChatGPT. People were using these things as copilots to help them get to their solutions. And as the automation improved, now we have things like Codex, and very few of our engineers are writing code the way they used to. And we think a very similar thing could play out for Periodic as well, where you can have systems to accelerate the researchers, the material scientists, the materials engineers, the process engineers. But as the autonomy, intelligence, and capability grows, you can begin to price outcomes. So help me get to this kind of goal state, and I think that’s, a really interesting area for us.

Ekin Doğuş Çubuk [01:18:03]: Yeah. We kind of feel like our best contribution to solid-state physics and science could be if we made these tools and topics profitable, similar to how ChatGPT made CS and LLM majors way more popular in colleges before and after. We’d love to show that solid-state physics, material science research can make a big impact commercially, and then that will attract more attention. Young people will want to study physics, which will be, a dream. And Bell Labs did make a huge commercial impact, right? So they failed to commercialize some of their incredible advances. They, of course, did some research that just couldn’t be commercialized like the cosmic, background. But they did benefit a lot from the vacuum tube connecting, East Coast to West Coast by phone lines.

Liam Fedus [01:18:49]: Yeah. And I think it’s really also interesting to think that technology and capital are incredibly intertwined. So if you look at what was the progress on chatbots for many years imagine recounting, okay, what was the progress from chatbots from I’ll say 2010 to 2015? We’d be kind of at a loss to say over that five-year period, how much did they improve? Whereas if you look at the period from, 2021 to 2026, it’s night and day. And what happened was ChatGPT and these other systems were able to achieve a product market fit, and it changed the capital landscape entirely. This changes the landscape for hiring, compute, data, et cetera.

Swyx [01:19:31]: The virtuous cycle.

Liam Fedus [01:19:32]: Exactly. And so

Swyx [01:19:32]: Like success breeds success.

Liam Fedus [01:19:33]: Exactly. So technology, it’s incredibly coupled to the capital and those resources, and we want to achieve the same thing in the physical world.

Open-Source Contributions and Academic Research

Swyx [01:19:41]: Yeah.

Brandon [01:19:42]: So one thing I’ve been seeing is there’s this move from science being funded from government grants and so on to VCs and private funding. Do you all plan on making anything you’re doing, open source, like either releasing data sets or actual models? Is this something in the future even. You know, there are many models which may be not well, your state of the art, but which could still be you know, useful releases.

Ekin Doğuş Çubuk [01:20:07]: Absolutely. So, we have a couple of things. So we have been contributing to open source to kind of like PyTorch and the Materials Project, code base, Custodian, their DFT runner. TorchSim was actually created by one of the researchers here, Abhijit Ganggan, and he maintains it still. JAX-MD, Abhijit maintains it. So we do a lot of open source contributions, including LLMs.

Liam Fedus [01:20:31]: XLANG, Megatron.

Ekin Doğuş Çubuk [01:20:32]: Yeah.

Liam Fedus [01:20:32]: Yeah. So we’ve been very active contributors back into open source.

Ekin Doğuş Çubuk [01:20:36]: And we also have an academic grant program where we give academic gift grants to academic groups in universities that we feel like are really advancing this direction towards synthesis superintelligence. So that’s also been great to see. I think the first paper from this funding is about to come out, so it’ll be exciting to just get more papers come out like this. Yeah.

The Path to Better Superconductors

Brandon [01:20:57]: Okay. Let’s steelman this, say that you’ve achieved 100% lab automation, which does everything, characterizes everything correctly. You have all of the fun agents which can do everything. I still am somewhat unclear about how you actually get to superconductivity. There’s a lot of hard problems to solve. What is the path there when there is no sort of theory for model for most of the strongly correlated systems or high-temperature systems?

Ekin Doğuş Çubuk [01:21:23]: Yeah. So, our view is that it’s not hard to think of chemical spaces that would host superconductivity. What’s really hard is to synthesize them. That’s why we’re really emphasizing synthesis superintelligence. So maybe one example to give historically is Alex Müller, when he thought that transition metal oxides might be a good path to superconductivity, he was actually trying nickelates, the nickel, oxygen, and some, cations, and that didn’t work. And then he tried cuprates, which is copper, oxygen, and some cations, and that worked, and then he got a Nobel Prize. But then we know that later on turns out nickelates was a good idea. Nickel and copper are next to each other on the periodic table.

Brandon [01:21:59]: But only as two layers, right?

Ekin Doğuş Çubuk [01:22:00]: Yeah. Like there are only so many 3D transition metals.

Brandon [01:22:03]: Yeah.

Ekin Doğuş Çubuk [01:22:03]: And then, Harold Hwang, one of our, Stanford professors here who’s incredible, he realized he can make nickelates in thin film form and show the superconductivity. So, if we have a really good synthesis superintelligence in the lab and we can really scale up these things, I think we won’t run out of ideas, or the LLM won’t run out of ideas about what directions to try, like 3D transition metals, oxygen. There are very related ideas here. Many people have been trying cobaltates, which is like instead of copper or nickel, you try cobalt. So I think that’s super exciting. Maybe one historical example to also study is the Japanese group that discovered magnesium diboride. As you know, that’s the highest temperature, ambient pressure, conventional superconductor, and they discovered it by just trying a bunch of materials. I think they tried 30,000 different things. Thirty of them seemed to host interesting superconductivity, and MgB₂ was one of them. But magnesium diboride sat on people’s shelves as a precursor for decades before then, and BCS theory had been invented back in 1957. So the way people found these materials wasn’t projected from theory. It wasn’t because they couldn’t synthesize until then. It was just like they tried a bunch of things. So now imagine if we had this perfect automation that you’re describing, which sounds like a dream, and we tried the 30,000 the Taketo group tried in over his career in a month. We just really increased the surface area for luck.

Closing: Increasing the Surface Area for Discovery

Swyx [01:23:30]: No, but yeah, it’s very inspiring what you’ve done. Congrats on all your success. I feel, I feel like you’re creating, yeah, the modern playground. I imagine recruiting must be super easy for you, so I’m just a little bit jealous, but you’ve worked hard for it to get here, so yeah.

Ekin Doğuş Çubuk [01:23:44]: Yeah.

Swyx [01:23:44]: Yeah. Well, thank you so much. Yeah, it was great chatting with you both.

Ekin Doğuş Çubuk [01:23:46]: Yeah, super fun. Yeah.

Swyx [01:23:47]: Yeah. Thank you.

Last call for regular tickets for AI Engineer NYC! As an exclusive for Latent Space subscribers, the first 30 of you can take a 30% off code if it helps - for new tickets only, no refunds! See you in 2 weeks!

While we tend to cover industry on the pod, every so often we celebrate a clearly emerging superstar PhD. In 2024 we featured Shunyu Yao, who went on to build Operator at OpenAI and is now Chief AI Scientist of Tencent. In 2025 we featured Jack Morris, who went on to cofound Engram at $600m and is now a leading voice on continual learning.

This year we are proud to feature the work of Alex Zhang of MIT.

From GPU kernels and KernelBench to Recursive Language Models, Mismanaged Geniuses, and massive multi-agent swarms, Alex Zhang is exploring how much capability we’re leaving on the table by wrapping increasingly powerful models in primitive systems.

RLMs took over the timeline early this year:

and an RLM based harness was the first to ~solve ARC-AGI-3 before OpenAI’s Astra:

and is even today, influencing new research that has more extreme implications than RLMs:

We go deep on GPU Mode and AI-written kernels, research taste and why academics should take bets industry labs won’t, GEV and alternatives to the standard autoregressive language model, and the idea of harnesses as compositional generalizers. Alex explains RLMs, context offloading, programmatic subagent calling, Prime Agent, persistent subagents, and why the “language model” of the future may actually be an invisible swarm of agents underneath a simple interface. We also discuss OpenAI’s massive agent experiments, Kimi swarms, open-ended research at Sakana AI, speculative programmatic tool calling, capability overhang, Neuralese, and where Alex thinks the next big research opportunities may lie.

We discuss:

* Why AI-generated GPU kernels still leave substantial room for human expertise

* How one expert insight can potentially replace enormous amounts of brute-force token search

* Why PhD students should take research bets that initially look trivial, weird, or pointless

* What SWE-bench, RLMs, ReAct, and Quiet-STaR reveal about research taste

* GEV and why a language model does not have to mean an autoregressive text-to-text decoder

* Why Claude Code, Codex, and Pi are structurally more similar than they look

* How harness design can improve compositional generalization across tasks and domains

* RLMs: context offloading, code execution, recursive subagents, and shared memory

* Prime Agent, continual harnesses, and persistent agent-to-agent communication

* Why the model you query in the future may secretly be an entire swarm or scaffold

* OpenAI’s 10,000-agent experiment, 130B output tokens, and ~$40M-equivalent problem solving

* Why much of an agent swarm may be wasted search — and why convergence is still hard

* Kimi versus OpenAI and different approaches to multi-agent systems

* Open-endedness, Sakana AI, and finding hidden gems in enormous amounts of generated work

* Why current frontier models may already have a large capability overhang

* Speculative programmatic tool calling and overlapping tool execution with generation

* Whether English, code, or an entirely new “Neuralese” constrains how models reason

* AI for science, fast-moving benchmarks, and how Alex chooses what research problems to bet on

Alex Zhang

* Website: alexzhang13.github.io

* X: @a1zhang

Timestamps

00:00:00 Introduction

00:00:49 GPU Mode, KernelBench, and AI-Written Kernels

00:07:38 Human Expertise vs. Brute-Force AI Search

00:13:20 Research Taste and Taking Big Bets

00:19:28 GEV and Rethinking the Language Model

00:29:03 Video Game Agents and the Harness Problem

00:31:01 Why Claude Code, Codex, and Pi Are So Similar

00:36:42 Harnesses as Compositional Generalizers

00:44:24 RLMs Explained

00:52:01 Prime Agent and Persistent Subagents

00:57:41 RLMs in the Wild

01:00:30 OpenAI Swarms and the Future of Language Models

01:07:26 Open-Endedness and Sakana AI

01:15:52 Kimi vs. OpenAI Agent Swarms

01:20:06 Capability Overhang and Speculative Tool Calling

01:28:19 Neuralese, Future Research, and AI for Science

Transcript

Introduction: Alex Zhang, RLMs, and GPU Mode

Swyx [00:00:00]: All right, we’re here in the studio with Alex Zhang, I guess most famously of, RLMs, but you have a few other affiliations. Welcome to the show.

Alex Zhang [00:00:12]: Yeah. Thank you for having me.

Swyx [00:00:13]: Yeah. I guess GPU Mode as well?

Alex Zhang [00:00:15]: Yes, GPU Mode as well.

Swyx [00:00:16]: You were shepherded in by Mark Saroufim. Not everyone gets that kind of welcome.

Alex Zhang [00:00:19]: Yep. Yeah. Yeah. I’m very close to all the people in GPU Mode, so yeah

Swyx [00:00:23]: Yeah

Alex Zhang [00:00:23]: We often end up working together in various capacities, like even beyond just GPU Mode itself, so.

Swyx [00:00:29]: Yeah. Can we explain, so people who are not that close Don’t know about this. It’s just a-- it’s a Discord. It used to be focused on, I guess, CUDA Mode, and then generalized a little bit. it was started by Mark.

Alex Zhang [00:00:41]: Yep.

Swyx [00:00:42]: It was basically like. To me, it’s like the hiring pipeline of the PyTorch team.

Swyx [00:00:45]: And then you left PyTorch.

Alex Zhang [00:00:47]: Yep.

Alex Zhang [00:00:49]: Yeah. So it used to be, I think, it started actually around when I was in college, in like 2023. I think it was started by Mark, Andreas, and Jeremy Howard. The original premise was just like, it was a GPU-- or it was a Discord dedicated to learning how to write GPU kernels, and they had, like, lectures. That was basically the extent of it. and I got interested in it because I was writing GPU kernels. It was actually out of, like, pure chance. I was interning at Snapchat at the time, and

From CUDA Mode to GPU Mode

Swyx [00:01:22]: Rexis.

Alex Zhang [00:01:23]: Yeah, I was very bored with Rexis. So, they had a project where, like, they were interested in writing. It was this paper called Infinite Attention. It was like a Google paper.

Swyx [00:01:35]: Yes, we’ve covered it on Paper Club.

Alex Zhang [00:01:36]: Yes, yeah. So I was interested in whether or not you could write specialized kernels for it at Snapchat. it didn’t. Nothing really came of it, but I joined GPU Mode. At the time, it was called CUDA Mode, I think for, like, legal reasons or something, they changed the name. But I met Mark, I met Matei, I met a bunch of other people that were very involved in the community. And then Mark had pitched this idea called Popcorn, which was now what you see as the leaderboard today. But the general idea was like. I think all of us had this, like, intuition that GPU programming is, like, very similar to if you guys have done, like, competitive programming. It’s a, it’s a not. I don’t mean to say, like, they’re transferable skills.

Popcorn, KernelBench, and Automating GPU Kernels

Swyx [00:02:17]: You have constraints. You code golf a little bit.

Alex Zhang [00:02:19]: Yep.

Swyx [00:02:19]: Yeah.

Alex Zhang [00:02:19]: Yeah. And there’s like. There’s actually a surprisingly small space of optimizations that people do. and there’s actually not that many kernels per se that people are interested in optimizing. And so we kind of had this thought that, like, if you had enough data, like, in the same way that Codeforces, there’s millions of problems. If you could do this with GPU code, like, you could scale and automate kind of GPU kernel development, which for researchers is a huge deal. ‘Cause I think one of the bigger bottlenecks. Like, if you look at, like, Mamba, for example, like, they release the paper with kernels because otherwise, like, you can’t really use it in any meaningful way? And not everyone has, like, a Tri Dao on their team. So we’re very interested in this. KernelBench kind of spawned from that too, of like, can we get LLMs to automate, GPU kernel code? And I think that was like a. It was a very fun time. It was like between college and my PhD, and yeah, I had a really pleasant time doing stuff with GPU Mode. Now I kind of just help with the lectures sometimes. I’m not as involved, and I think in general, like, we don’t have as many competitions as we used to. But, yeah, I still keep in touch a lot with everyone there.

Swyx [00:03:27]: Is there a friendly rivalry? Because I think the previous community that used to do this was like MLSys, MLPerf

Alex Zhang [00:03:32]: Yeah

Swyx [00:03:32]: Kind of thing. Is there a friendly rivalry? Is this like just new generation MLPerf, or what’s going on?

Alex Zhang [00:03:38]: The nice thing about GPU Mode is that it is also a community in the sense that, like, a lot of the lectures are very easy enough for a beginner to follow and ask questions and things like that. And, like, the competitions are, like, somewhat, not secondary, but, like, you can participate in them to learn. I think with a lot of, like, MLSys, MLPerf kind of benchmarks, like, for the most part, like, only serious labs and companies participate. Like, seriously in them, at least. That was my understanding of it. I could be wrong. But I think also beyond GPU Mode now, one thing that has been really exciting is there’s a lot more websites and, like, people that work on hosting competitions. Like, I think there’s this.

Alex Zhang [00:04:21]: I think there’s this website called, like, LeetGPU or something, and it’s like leet code for GPU problems.

Swyx [00:04:26]: Wow.

Alex Zhang [00:04:26]: There’s, like, other ones that I’ve. Like, we’ve, we’ve seen. Like, there’s many that have kind of spawned and, like, talked on GPU Mode, and like, it’s very exciting in the sense that I think GPU programming used to be super niche, like when I was interested in it. And the only reason I got interested in it was Tri Dao gave a talk at Princeton because he was, applying for faculty. he is faculty there now, but I listened to his talk on FlashAttention in like 2023, and I was like, “Wow, this is like the coolest thing ever.”

Alex Zhang [00:04:56]: And I was like, “This is like. This is what everyone should be working on.” I guess, like, vLLM and stuff had come out too, and it was like, “Oh, we should be writing kernels.” But now it’s like, everyone writes kernels. Like, everyone. It’s, it’s. I think it’s actually almost saturated in some sense, as a field.

AI-Written Kernels and the Verification Gap

Vibhu [00:05:11]: Any interesting takes for people that wanna get into it? So I think one of the biggest news is GPT-5.6 Wrote more efficient kernels, so Terra and, Luna could be 80% cheaper.

Alex Zhang [00:05:24]: Yeah.

Vibhu [00:05:25]: And then we’ve seen other competitions where people are, like, setting records, and they’re like, “We’re doing some auto research loop,” and these are people that don’t have a background

Alex Zhang [00:05:34]: Yep

Vibhu [00:05:34]: In any kernel writing, right?

Alex Zhang [00:05:36]: Yeah. So even on the GPU Mode leaderboard, if you look at, like, a lot of the recent problems, almost all the solutions are AI generated. However, you’ll notice on the leaderboard. So there’s this guy named Gauners who’s, like, a very, like, regular member of GPU Mode. We’ve always known for a long time that he’s, like, a super cracked, like, GPU kernel writer. One thing we discovered on this leaderboard is, like, almost ev-- Like, he also used AI to help him with these solutions, but for the most part, like, he helped prompt and move it in certain directions. we found that, like, his kernel was, like, basically the only one in, like, the top 10 that was actually stable in, like, actual, like. end-to-end systems. And it does bring into question, like, it’s not. Like, GPU kernels have a verification problem. Like, we’ve kind of known this. It’s been a problem since KernelBench was released. Like, there’s a lot of reward hacking that goes on. But you al-- Yeah, you also notice, like, the lines of code is a lot smaller, but

Vibhu [00:06:33]: Yeah, I was gonna ask, is that noticeable, or is it just

Alex Zhang [00:06:36]: Yeah, no, it’s, it’s

Vibhu [00:06:36]: Okay

Alex Zhang [00:06:36]: It’s definitely, like, very important, and I think, like, it’s, it’s really interesting that still there’s a lot of alpha in being good at writing GPU kernels.

Vibhu [00:06:44]: Okay, so there is a gap from verifying

Alex Zhang [00:06:45]: There definitely is, yeah. I think, like. And this applies to a lot of AI systems as well. Like, I think, even with the most recent, like, math proofs and stuff, like, it doesn’t necessarily mean mathematicians are obsolete. these companies still hire mathematicians, like, to do, whether it be, like, data labeling work or even just, like, steering the models to solve problems. Like, there is still a lot of alpha in being, like, knowledgeable in these things, so.

Swyx [00:07:12]: Is it just knowledge, or is it also there is just more planning, and is there, an emergent style of planning that works better?

Alex Zhang [00:07:22]: I think it’s, it’s a mix of. Maybe this is what you mean, like intuition for

Swyx [00:07:27]: Something like that

Alex Zhang [00:07:28]: How to solve the problems.

Swyx [00:07:29]: Like, for example, I always diagram my code.

Alex Zhang [00:07:31]: Yeah.

Swyx [00:07:31]: Right?

Alex Zhang [00:07:31]: Yeah.

Swyx [00:07:31]: And then, like, if there’s a part of the diagram I don’t understand, I work until I understand it. Otherwise, I, it’s not allowed.

Alex Zhang [00:07:37]: Yeah.

Swyx [00:07:38]: Yeah.

Alex Zhang [00:07:38]: So I think it’s, like, it’s a mix of those things of, like, the people who work. Like, the people who know how to look at these problems and how to solve them, like, also know how to use AI to do them. Because, like, you’re acting as a very strong verifier. Like, if you are knowled- or if what to do and you are also. Like, I think the thing that we’ve kind of discovered with all these agent swarms and things like this is, like, when you throw enough compute at a problem, you, like, can sufficiently explore solutions to that problem. But oftentimes, like, maybe you can burn, like, 100 billion or a trillion tokens on something, but if you bring in someone who knows something about the problem, they can uncover something for the model that would, like, erase that one trillion token spend. it’s not, it’s not super clear, like, what exactly the trends are here. But I think, like, there are so many problems in the wild still right now that we want to solve, and, like, we can’t afford to just always, throw as much compute as possible at it. Like, there is still an efficiency aspect of all of these things that is super important.

Speed-of-Light Limits, Memory, and Megakernels

Swyx [00:08:41]: Is there, like, a theoretical right answer that you just calculate based on physics, and then you just get close to the physics limit?

Alex Zhang [00:08:49]: Yes. So for GPU kernels, you can compute. It’s actually not that easy to compute sometimes, like, depending on how complex the problem is. Like, for matrix multiplication, it’s very easy to compute, this, like, speed-of-light kind of, estimate of what the fastest kernel can be. And, like, this is also assuming, like, maybe all of your, all your data starts on the CPU, or maybe it starts in DRAM, on the GPU, et cetera. Like, this changes these numbers slightly, but

Swyx [00:09:18]: The transfers and all these things, yeah.

Alex Zhang [00:09:20]: Yeah. But I will say, like, it’s not clear, though, like, in a lot of cases if it’s even possible to hit this theoretical number, if that makes sense. Like, this is assuming, like, perfect overlapping and transfer of data, and, like, there’s maybe some bottleneck that you can’t get around. But often, the kernels are not even close. Like, that we write are not nearly close enough to this number to be, like, meaningful at all.

Swyx [00:09:43]: Yeah. And is it speed that matters? Do you also care about, obviously memory, which

Alex Zhang [00:09:48]: Mm

Swyx [00:09:48]: Feeds into speed? Do you care about power consumption? So one of my, one of our top pods of the year was Geoff Dean, who was like, “Actually, I just tracked the microjoules or, like, the nanojoules, picojoules.”

Alex Zhang [00:09:59]: Yeah, it’s often picojoules today.

Swyx [00:10:01]: Picojoules.

Alex Zhang [00:10:01]: Yeah.

Swyx [00:10:01]: Do you care about that?

Alex Zhang [00:10:03]: So I don’t.

Alex Zhang [00:10:04]: Yeah. I guess maybe I’m not, I’m not as

Swyx [00:10:06]: But everything here is speed, right? Like

Alex Zhang [00:10:07]: Everything here is speed

Swyx [00:10:08]: Nobody’s counting picojoules.

Alex Zhang [00:10:09]: But I-- There’s a caveat here, which is, I think, like, there is speed in the context of a single kernel, and there is speed in the context of a larger problem, like maybe the N10 model. Because, like, one thing to consider, and this is why it’s important to talk about what speed-of-light is referring to, because in these cases for the kernels, like, we always start with everything in, like HBM, for example, right? But you can imagine that, like, an end-to-end like an end-to-end model, what you might wanna do between two layers is, like, you might sacrifice the speed of the first operation to keep things in the cache for the second operation. And, like, these are things that, like, you can’t really get out of, in isolation with, like, these kinds of kernels. And, people call this, like, the fusion or, like, the

Swyx [00:11:01]: Megakernel

Alex Zhang [00:11:02]: Kernel fusion problem. Yeah, or, like, megakernel stuff. And it generally only applies, like, when you are, like, memory-bound in most cases. But this is something that, like, also there is this question of, like, as these models get better, like, should we just be generating like, megakernels? Is that, like, what we want?

Vibhu [00:11:20]: What’s, what’s your take?

Alex Zhang [00:11:21]: I think that this is really difficult because you need the data to do this. And I think, like, I have yet to see an example in the wild of, like, we bootstrap the ability to solve a very difficult class of problems without any examples. and I think, like, the other reason why I think maybe this isn’t that interesting is that at the level of an individual kernel, A, like, they’re not that, they’re not as complex, but B, you’re somewhat confident that there’s not as much structure in a single kernel. But, like, in a megakernel, like, I would be more inclined to believe that, like, a compiler would be better here. Like, some compiler over, like, higher level- Ops makes sense, because in general, like actually, I think mega kernels are very, like the pieces are very composable of like the individual kernels. There’s some areas where you might wanna do like weird fusions and everything, but in general, I think these are cases that like a compiler can probably handle. And there is a company that’s working on this from what I understand that has given some talks on GPU mode as well.

Swyx [00:12:28]: Yeah, I wanna basically cluster all the GPU mode discussions here because obviously there’s other parts

Alex Zhang [00:12:32]: Right.

Swyx [00:12:32]: That we need to move on to.

Vibhu [00:12:33]: I think there is something to plug. You guys do host a lot of really good lectures. They’re all on YouTube. People can follow. And you

Alex Zhang [00:12:39]: Yes

Vibhu [00:12:39]: Lead quite a bit of it. You’re still quite involved.

Alex Zhang [00:12:41]: I used to. sometimes I still do. I think they’re mostly Mark. Mark is the one who usually does them. Matei does sometimes as well, but, yeah, I highly recommend them. They are extremely good resources. Like, I think it’s kind of crazy how much people share on there, so yeah.

Swyx [00:12:59]: ‘Cause like if you’re there, like you’re very, like you’re exactly the right audience?

Alex Zhang [00:13:03]: Yes, exactly.

Swyx [00:13:03]: Like this isn’t gonna reach the mainstream.

Alex Zhang [00:13:04]: And there’s a lot of like introductory material as well, that we’ve put on, that I think is useful for people.

Benchmarks, Princeton, and Research Taste

Swyx [00:13:10]: KernelBench was kind of influential. I just wanna see like, that was last year.

Alex Zhang [00:13:13]: Yep.

Swyx [00:13:14]: What other ongoing work do you wanna shout out that people should pay attention to? ‘Cause obviously you’re involved in this

Alex Zhang [00:13:20]: Yep

Swyx [00:13:20]: Field.

Alex Zhang [00:13:20]: Yeah. I will give maybe the background story of like I am actually involved in a lot of benchmarks, or I used to be, maybe prior to my PhD. It started because I was at Princeton. I worked with the SWE-bench team there.

Swyx [00:13:33]: John, Carlos

Alex Zhang [00:13:33]: John, Carlos

Swyx [00:13:34]: Ofir

Alex Zhang [00:13:34]: And Ofir. They’re all great. Like

Swyx [00:13:36]: Karthik

Alex Zhang [00:13:36]: I love them. Yeah.

Swyx [00:13:37]: There’s basically this, like I think people don’t understand how much benchmarks come from the same group

Alex Zhang [00:13:42]: Yeah

Swyx [00:13:42]: At Princeton.

Alex Zhang [00:13:44]: It is crazy.

Swyx [00:13:45]: Do Xun Yu?

Alex Zhang [00:13:45]: Yes. Yeah.

Swyx [00:13:46]: We had him on a pod before. Now he’s like running Tencent.

Alex Zhang [00:13:48]: Yeah, now he’s like, he’s like a superstar. when I met him, so he was advising my friend Michael Tang, who is now at Anthropic, but they worked together a lot. We were like the two undergrads in Karthik’s lab. I-- And then some others joined later as well. But yeah, Xun Yu is great. I did not know he was like such a superstar until like later on, like after I left, but

Swyx [00:14:12]: Yeah. like, okay, so there are very few PhD students. Like yours is like the next one. Like once a year, we feature someone like who is like basically entire PhD, has been like on target.

Swyx [00:14:25]: There’s not that many of them. Xun Yu was like clearly one of them. And, Jack Morris is another one. And like, we talked, before the show, we talked about research taste.

Swyx [00:14:33]: Right? Like somehow some grad students just have a very blessed career where like, yeah, mostly like, yep, this is like going to stick around, relevant, everyone should know this.

Alex Zhang [00:14:42]: Yeah.

Swyx [00:14:42]: And then others, just nothing.

Alex Zhang [00:14:44]: I think this is also true of like even people within like industry labs as well. I think it’s just like grad students are a lot more visible. So you just see, like you see, like, there are some people who

Swyx [00:14:56]: Yeah, you can publish

Alex Zhang [00:14:57]: Who really like get lucky and like, or it’s, it’s a mix of being lucky and also being very smart and things like that. I think like with research taste as well, like I think it gets developed through opportunities, at least in my case. Like I got-- I was very fortunate to have like taken the path that I took, like working at Princeton and then like later, like finding my like Omar at MIT. Like he’s a fantastic advisor. I will say, though, I find that the most successful research from grad students or like in academia comes when people care about problems that maybe like most people in industry are not looking at. I think this is the issue that like a lot of grad students work on things that benefit, like that look good to an industry lab. Like for example, they’ll work on some, like some benchmark that’s really popular now. I think benchmarks in 2023 were a very different story than benchmarks now. There are a lot of people that work on like harnesses and like meta-harnesses and like specific harnesses for XYZ task. And when you really think about it, the reason someone would work on this is like maybe there’s like a clear goal shaped around the models that we have today of like, this is what I want to see. But like, I’ll give, I’ll give the like RLM, like the recursive language model paper as an example, because I think like it’s a super simple idea. I think when it came out as well, like there were a lot of people that were like, when they see something like that, they’re like, “What is even the purpose of this?”

Swyx [00:16:27]: Or like too cool.

Alex Zhang [00:16:28]: Yeah. Like why, like what? This is just subagents or something, right?

Alex Zhang [00:16:32]: And I think it’s like when you get a reaction like that, it’s almost like a good sign in the sense that like it’s clear that people aren’t thinking about what the purpose of this is. And I will give another example of like SWE-bench. When SWE-bench came out, Ofir loves to tell this story. When it came out, like nobody cared. Like everybody was like, “This is an impossible task. Like why would we ever even consider this as a benchmark?” And it wasn’t until Devin came out that everyone was like, “Whoa, like this is something we wanna hill climb.” And I think this is, this rings true for. You tend to see that a lot of ideas. I think like the. My favorite, I guess, example of this is Eric Seligman’s work, with like STaR and like Quiet-STaR. Like I think when you read the paper, at least when I first read the paper, I was like, “Is this not like an obvious idea?” Or maybe not. I don’t know. I was like, “Oh, this seems really simple.” Or like chain of thought, and the same thing. Or like Xun Yu’s react. It’s like, okay, like, yeah, sure. But then like when you really think about it’s like why. What is the value of the paper? And I think it comes from like, it tells a bit of a story as to like what you want the field to look like. And that is something that it’s very hard to do this in academia because if you look at all these papers, Quiet-STaR, ReAct, RLMs, SWE-bench, none of these papers are. It’s not like a GPT-6 Astro release? It’s not like everyone’s like, “Oh my gosh, like I’m gonna use this now and this is the best thing in the world.” Like academia just can’t afford to do this, at least right now. I. There’s a whole slew of reasons why I think that should change, but I think it’s like. If you don’t have. As a PhD student, I think you’re in such a unique position where you can work on literally whatever you want for the most part. If you’re not taking advantage of that and working on things that, like, nobody cares about or, like, people see as some trivial thing, like, “Oh, I thought about this, but, like, I don’t use it,” I just think, like, in the end, the research is just never gonna be that interesting because you kind of need to take big bets if you’re gonna be in academia. Because otherwise, I think, like, just go to an industry lab. Like, they have tons of resources, tons of talent. Why constrain yourself in an area where you don’t have a lot of resources and, like, there’s not even that many people around? And I think it’s just. it literally just comes down to, like, big bets. like, you just have to take big bets, and, like, a lot of them will fail? Like, that’s just. it’s, it’s natural. But I think

Alex Zhang [00:18:55]: That is, as a PhD student, like, that’s the biggest advantage you have over any single person at another lab because you don’t have to deal with bureaucracy and all these other things.

Swyx [00:19:07]: Fair enough.

Alex Zhang [00:19:07]: Yeah.

Swyx [00:19:08]: I ask a lot of people this question, and usually they hand-wave away. So I think. I appreciate that you’re actually giving a thoughtful response on Like, no, like, this is your unfair advantage because everything else is biased against you, basically.

Jev and Breaking the Autoregressive Decoder Paradigm

Alex Zhang [00:19:20]: Yeah, exactly. And so, like, it’s honestly. I will bring up Jev as an example because it’s, it’s not an academic

Swyx [00:19:26]: Wow, okay.

Alex Zhang [00:19:27]: It’s not an academic project.

Swyx [00:19:28]: Yes.

Alex Zhang [00:19:28]: I want to bring this up because this also happened with RLMs and, it happens with many other works. Like, things get overhyped, right? To an extent, like, something gets overhyped and then people are like, “Why is this overhyped?” Like, “This is trivial. This is stupid.” And I saw the same thing with Jev because I think the release was like. there is this whole thing about, like, academics, or they’re not an academic group, but, like, people have to do branding and they have to, like, kind of market their research. And so, like, I understand, but I think there was a lot of discourse about Jev just being, like, something we’ve known for years. And I think it’s kind of missing the point of, like, why is such a system so interesting? It’s why is it not just some stupid NLP classifier that, like, we’ve, we’ve been doing, back in our intro ML classes or something? I think what’s really interesting about Jev is that it kind of opens up this question of, are language models correct? Like, in the form that they’re in, can we consider a different design space other than text-to-text? Because what they’re doing is they’re basically saying like, “I will take advantage of this language model backbone. Like, I know it captures a lot of information about language, but I’m going to change the output space of the model to give you a trade-off, which is I will do very fast inference over.” Like, if you have some prior about this problem, like, let’s say I only need to make a binary classification. Am I gonna ask my language model to do this and pay, like, a 400X cost? Like, no. That’s-- it’s, like, silly, right? and I think for the longest time, because the labs are the only places that control, you’re never gonna use something other than, like, GPT-4 or GPT-6 or Fable because they’re the best models. But because of that, like, people have gotten kind of accustomed to this idea that a language model is just a autoregressive decoder. Like, we have accepted this. And I think when RLMs came out, it was the same thing. Like, one of the comments, like a very frequent criticism I got was like, “This is not a language model.” Or like, “When I look at this, like, I thought it was a new architecture, but it’s actually not.” And my response to that is like, “Well, a language model is just modeling language. It doesn’t have to be this transformer decoder,”? and Jev is really interesting in that, like, we now have a new

Alex Zhang [00:21:49]: Thing to tune, which is like, what is the output space and how does this affect inference latency? and I think we can actually start asking this about various parts of the language model itself. we are seeing this too with, like, loop transformers. It’s a similar idea of a lot of the attention around it was like, “This is a silly idea.” Like, “Why? Who cares about this?” But it’s like, it is a simple idea, but it’s actually. it opens up a whole new set of questions that I think, like, especially if you’re a PhD student, these are the things that you wanna answer. Because I think it’s like we don’t know. For Jev, for example, we don’t know how far we can take this. for loop transformers, we also don’t know how far we can take this. What if you loop only a subset of the model? what if you route to, like, only. like you have some router to different parts of the model? Like, can you mimic what you do in a harness inside of the model? What can you bridge between a harness choice and a model architecture choice? These are all questions I think that open up with works like this, and that’s, like, really exciting. Jev in particular, when I saw it, I was like, “This is actually really useful for RLMs.” Like, I think it’s, it’s. it makes sense ‘cause the biggest bottleneck in RLMs or swarms or systems like these is they’re slow. When you do multiple language model calls all the time, you’re not distributing your compute correctly because, like, maybe there’s something trivial that you just want a simple model to do, but you can’t do it because your language model is just this bulky thing? So I’m very excited. I think we will start to see new types of models emerge beyond just the bog-standard frontier model, and that is like. there’s so many things that you can do with these, like, new trade-offs.

Swyx [00:23:34]: I’ll also shout out Thinky with their interaction models.

Alex Zhang [00:23:36]: Yes. Yeah. Another great example.

Swyx [00:23:38]: Yeah. So, like, basically try to break the paradigm from sequence to sequence, decoder only, and just literally do anything else.

Alex Zhang [00:23:46]: Yeah.

Vibhu [00:23:46]: I think there’s, there’s a level of if you’re trying to compete, you’re not gonna compete with a Frontier lab doing an autoregressive

Alex Zhang [00:23:54]: No.

Vibhu [00:23:54]: Decoder on. Like, the amount of compute scaling resources they Even Thinky will not. okay, there may be one of the handful that can, but you’re not really gonna do much in that at least.

Alex Zhang [00:24:06]: I don’t know too much about Thinky, or I don’t wanna say anything either, but it’s like if their strategy is just to replicate OpenAI or Anthropic, like that’s a horrible strategy.

Alex Zhang [00:24:15]: Because, well, because, like, they just don’t have. Like, you kinda just have to think of it in terms of, like, what advantage do you have? And if you’re going to use the same setup. I’m sure they’re not, but it’s like if you’re going to do the same setup, like you’re basically competing on the things you can control, which is data and compute, and obviously they cannot compete with the Frontier Labs on that. So yeah, it makes sense that, like, if you are a neo lab, like. Actually, I don’t know if you would consider them to be a neo lab, but I guess, like

Vibhu [00:24:40]: Yeah. That’s why they’re, they’re in there.

Alex Zhang [00:24:42]: I guess they’re kind of a weird one, yeah.

Vibhu [00:24:43]: They’re in their list. They shipped Inkling. Like, they can’t.

Alex Zhang [00:24:46]: Anything other than OpenAI or Anthropic, maybe like Meta and GDM, like you just, you gotta do something else? Like, it just. It’s the sad reality, but I think. I actually think it’s a good thing. I’m very happy that, like, scale works and these companies will just keep doing it because it opens up, like, potentially new players, like if they uncover something really interesting. Because I sort of have my doubts that this is, like, seriously going on at Frontier Labs, ‘cause it’s like why would you do that? like, why would you take the risk of allocating a large amount of compute to new bets when the old bet already works? So

Vibhu [00:25:24]: And I think that’s what spins off a lot of neo labs, right?

Alex Zhang [00:25:27]: Yeah.

Vibhu [00:25:27]: You have a side bet and you don’t get compute, and you’re like

Alex Zhang [00:25:30]: Yeah

Vibhu [00:25:30]: “Okay, I’ll go, I’ll go do that.”

Alex Zhang [00:25:31]: Exactly.

Vibhu [00:25:31]: And, your example of the potential upside is something like Jev, which is X hundred times cheaper, comes out, and maybe it is language model.

Calibration, Fast Classification, and New Model Trade-offs

Alex Zhang [00:25:41]: Yeah.

Vibhu [00:25:41]: In this case, it’s just different.

Alex Zhang [00:25:42]: Yeah.

Swyx [00:25:43]: Yeah.

Swyx [00:25:44]: So no speculation on what Jev actually is?

Alex Zhang [00:25:46]: I guess I have some guesses for what it might be. I have seen some people say like, “Oh, it’s like a diffusion thing.” I guess that

Swyx [00:25:56]: Which is the parallel decode, right?

Alex Zhang [00:25:58]: Yeah, parallel decode. Honestly, I think regardless of what it actually is, ‘cause I think you can. I’ve seen some, like, open-source replications of it. What is really exciting to me about what they did is I’m not entirely sure what their optimization objective was and how they trained it. And I think, like, this is a thing for RLMs that we’ve also been thinking about, which is like, okay, like RLMs are a very simple idea. If I come out with this paper, like anyone can use it now. But what distinguishes The actual value of an RLM is whether or not you can train it properly, and whether or not maybe you can mold some architecture around the system to make it really good. And that’s something that, like, I’m actively working on, I guess. But I think for them, like, they figured out a way to train the system, which is completely non-trivial. Like, I actually don’t really know how they did it. And I’ve seen some comparisons online of, some people are claiming they used Qwen, or they post-trained on top of Qwen, but every open-source Qwen that you use is gonna be worse, ‘cause whatever they did to train it clearly works very well. And so that’s, that’s very exciting.

Swyx [00:27:04]: There’s one element of calibration Which, is a rare topic that I don’t think people even knew about or understood. We covered it with, our conversation with Clementine Foley of Hugging Face, and she used to run the evals, at Hugging Face, which is basically the idea that, models are attuned to give you the most likely next token. But, they’re gonna lie to you when you ask them, “How confident are you?” Because they’re just gonna give you the most likely next answer instead of, like, actually, like, no, let’s calibrate. Like, I am actually fifty percent sure, or I am twenty percent sure, and, like, let’s try to calibrate that. I would say, like, if anything, I think that actually that’s pretty easy to generate synthetic data around Because you can sort of see the truth and then synthetically generate a bunch of answers, have Jev classify it, and then compare with ground truth.

Alex Zhang [00:27:52]: Oh, I see.

Swyx [00:27:52]: That would be my reverse engineering of this.

Alex Zhang [00:27:54]: Yeah.

Swyx [00:27:54]: I’ve actually. I think calibration is probably the under. Like, people are just using it as a very fast classifier But they’re actually not even using the probability or calibration estimates.

Alex Zhang [00:28:04]: Yeah.

Vibhu [00:28:04]: I think it’s also still just misunderstood to reiterate. When you ask a model, “How confident are you?” it will spew out what, forty-three percent. the big delta is this is a grounded classification, right?

Alex Zhang [00:28:16]: Yeah. Yeah, I’m, I’m very excited to see what people do with this model. is it gonna solve everything? Like, no, of course not. But I think it solves a class of problems that we traditionally struggled with, which is low-latency things. So I love the examples with games. That’s actually like. I

Swyx [00:28:34]: Yeah, the Doom example

Alex Zhang [00:28:35]: Yeah

Swyx [00:28:35]: Was very good.

Alex Zhang [00:28:35]: I have a benchmark on language models playing video games. I’ve always been fascinated by whether you can have an intelligent system play new games and things of this nature. And so I think it’s really cool that they have sort of a unique way to do this, to capture language and understanding in, like, a fast, a very fast model.

Language Models Playing Video Games

Vibhu [00:28:59]: Oh, while you’re on the topic, anything you wanna point out for video games?

Alex Zhang [00:29:02]: Oh, yeah.

Vibhu [00:29:03]: This is. you did do a benchmark on any project, right?

Alex Zhang [00:29:05]: So, yeah. I guess these numbers are very outdated

Vibhu [00:29:08]: Yeah.

Alex Zhang [00:29:08]: Because a lot of the models are very different. And I’ve seen actually people run. There are some folks out there that are running newer models on these games, which is really cool. I guess the general premise of this benchmark was we just want to see if vision language models are good enough at just, like, plugging into games with the latency constraint included. ‘Cause this actually. I came out with this right after Claude Plays Pokémon came out.

Alex Zhang [00:29:33]: So this was, like, two years ago, which I guess is, like, ancient now. But I think what’s really cool about this suite of tasks, it’s very diverse in terms of what games they are. And also, I think most of the games are games that people know or, like, have seen before. I saw, yeah, Jeff playing Doom. I will say I don’t think. I think they were just playing, like, really simple levels and stuff. But honestly, like, most models still can’t really do. Or I don’t actually think any models can solve these games very meaningfully. Like, there are some games that they can. I think I’ve seen Astra be able to solve the Kirby game. And we also, for this benchmark, we intentionally designed a really minimal harness. And I will get to this point about harnesses because I think there’s a whole conversation to be had about, like, what is the value of a harness? Like, what is even the purpose of a harness for a model? But in general, like, I think it’s. yeah, I hope to see very quickly or very soon, like, all of these games beaten by newer models.

Vibhu [00:30:35]: Yeah, it’s interesting. Like, the old Cloud Place Pokemon, they, like, read state from RAM and saw what tiles are walkable and whatnot. We did a podcast with them A long time ago.

Alex Zhang [00:30:45]: Gotcha.

Vibhu [00:30:46]: Yeah. Just fun.

Swyx [00:30:46]: Yeah, and it’s similar. Like, Jeff doesn’t have vision

Alex Zhang [00:30:48]: Yes

Swyx [00:30:48]: So you have to kind of feed in,

Harnesses as Compositional Generalizers

Alex Zhang [00:30:50]: Yeah

Swyx [00:30:50]: These, like, game state and all these things. let’s go right into the harness stuff

Alex Zhang [00:30:54]: Awesome

Swyx [00:30:54]: Because you brought it up. Language model harnesses are compositional generalizers.

Alex Zhang [00:30:59]: Yes.

Vibhu [00:31:00]: You struggled to read that one.

Alex Zhang [00:31:01]: Explain. Yes. Okay. So I have been a little unsatisfied maybe with how people think about harnesses, because people compare like, “Oh, like, I love Claude Code, I love Codex, I love Pi.” Like, “No, I love Oh My Pi, I love Prime Agent.” To be honest, I think all of them are the same. Most of the design decisions or, like, the design choices around these harnesses are the same. Maybe Prime Agent is a little bit different because it’s, like, inherently an RLM. But in general, like, I think we can be a lot more creative with harnesses. And what by that is if we think about this from the perspective of what exactly is the harness doing for the model? Well, basically, when you’re trying to solve a problem and you want to use a language model to solve it, like a very difficult task, one thing that we have discovered is that next token prediction is a really awkward form to do a lot of these tasks. So for example, take SWE-bench. When you’re navigating a code base, like, are you going to be able to figure out how to do all of this with a single language model call? Like, you just say, “Solve code,” or like, “Solve my query over this code base.” No. So we rely on a harness to help you do these kinds of things. And I think what is interesting is, like, a harness is a very opinionated program over how you want a language model to be form fit over a problem. And I bring this up because when we think about the, what a harness is doing, we should really think about, like, what choices in the harness let the language model solve this task? And can I actually just have a language model that just does this? Because a harness, if you think about it, now that loop transformers are a thing, I think what’s really interesting about it is you can model a looped transformer in some ways, like, with a harness as well, right? You’re just looping over the model. Now you can say like, “Oh, I’m not decoding,” so it’s, like, a little bit different. But in general, we, for whatever reason, have stuck with the same model architecture choice forever. And I. And there’s many arguments for why, but clearly, like, we are now training models around harness rollouts. And so there is this very awkward way of doing training over harnesses, which is that we train a language model to act within a harness, but, like, now it’s like a really long, maybe, like, multiple agent rollout that we’re doing. And there’s, like, really awkward, hacky ways of doing this. So what this blog talks about is like, well, one way you can think about what is going on here is if the harness is basically helping the model solve a particular task, can different harness design choices actually do something a little bit more meaningful beyond just, “Here are some tool calls that will help you. Here is a way to grep through your code base.” And so this actually. The idea for this blog came with the RLM idea as well. We just didn’t package it that way. And I think this is actually true. There are many other ideas around RLMs that, like, we will be coming out with, but were all there from the beginning. these are all design decisions around. I think with what is. What I like about the RLM is that there were many iterations and versions of different abstractions that I was interested in doing, and ultimately the RLM made the most sense. But there’s a lot of reasons that aren’t public as to why that’s the case. you will see. So in this example, one of the things that we see with an RLM is that if you sufficiently offload context and write. ask the model to write code over that context, you get this really weird but useful property, which is that when the model recognizes during training how to solve a task, it turns out that the solution to many tasks is very similar across, like, tasks where you don’t even. Like, it’s not even that clear to you that the solutions are similar. So in this example, we have, like, a retrieval task and we have, like, an aggregation task, and they’re very different query. Like, the domain is just completely different. And when you train a regular language model over these two tasks, the trajectories look very different. And so what you’re relying when you. Like, let’s say you use Pi or Claude Code or something, which is not in this blog, but we do have these results. You’ll find that, like, these harnesses distinguish too much between these problems, even though the solutions are the same. And so one thing that we find when training RLMs is that, like. When it sees these problems, it’s the same. And the reason it sees these problems as the same is the sub-agent sees different problems, but the sub-agent is solving an easier sub-task, and so you’re confident it’s smart enough to do it. But for the base overall strategy, they end up looking the same. And so when you train on the left task, for example, it can immediately solve the right task. And so if you go down to, like, the plots that we have, one thing you’ll kind of observe is that the. as you just naively train your RLM on these tasks, they naturally learn to generalize, for example, to longer tasks, because the strategy is basically the same. You’re just modifying, like, a length variable. And this actually also holds for tasks that are different, and they’re, it’s not even-- they’re not different across length. They’re completely different tasks, math tasks versus writing tasks. But the solution, the, like, meta high-level solution is the same. And so when you train the RLM on one of them, it generalizes this behavior to the second one. And there’s no magic here. I guess maybe that’s the thing that I wanna kind of stress. Like

Vibhu [00:36:38]: How would you kind of verbalize what they are learning? So I think in here you say you train it on short tasks, they generalize to stuff 8–30x longer.

Alex Zhang [00:36:47]: Yeah.

Vibhu [00:36:48]: They are learning how to solve these type of problems, or what’s the, what’s the core thing they’re actually learning?

Alex Zhang [00:36:52]: Yeah. They’re learning how to solve these types of problems at a certain length. And it turns out that when you take the strategy that they learned, it is directly transferable to the longer length. Like, they’re

Vibhu [00:37:05]: Yeah.

Alex Zhang [00:37:05]: Effectively the same program. And this is something that is really exciting because what this kind of implies is that when you have a corpus of data or environments that you train your model on, the hope is that, like, A, you can train on less environments and generalize to more than what existing models can do through or harnesses through existing kind of, like, naive training. But B, also, you still wanna use all the data you have. So when you train on these tasks, like, hopefully it generalizes to a wider class of problems. And why this is also even more exciting, at least in the context of RLMs or any recursively calling system, is this argument holds inductively. So, like, I’ll give you an example, because I said, I claimed that competitive programming and GPU optimization use very similar skill sets. The model can. the harness potentially, you might have to nudge it a certain way, but it can learn that, like, “Okay, how I’m gonna go about solving this GPU programming task is very similar to what I learned for competitive programming. So I’m gonna list out a set of solutions. I’ll, I’ll, like, spawn subagents to list out promising solutions, and then I’ll, like, write this loop to go through and check these solutions, maybe evolve them, and, like, evolve them against a verifier.” And between these two tasks, this looks the same. But what the subagents are doing are maybe, like, unique and something, like, different. But even what the subagents are solving might actually also be of the same form, right? Because it’s like a, it’s a recursive argument. And so what I’m trying to get at with this whole blog post is just that, like, we should rethink what the role of the harness is, because harnesses can actually greatly increase the generalization capability of your model and the amount of data that it’s given. And this is not exclusive to RLMs. I think there is a wide class of harnesses that are yet to be discovered that actually can yield similar properties. And to extend this argument a little bit further, I think what you can also extrapolate from this is, like, if I look at an RLM, what are the components of an RLM that are actually necessary, and can I actually just directly train a model to do this? Can I train a model to act as an RLM implicitly in its forward pass? It’s a really weird thing to think about because, like, you might say like, “Oh, code is non-differentiable, blah.” But there are many approximations of this behavior that we will start to uncover. And I think, like, we will see beyond just, like, I’m gonna design a new coding harness that uses a special form of compaction or something. I think we can be a lot more creative here. Like, there’s, there’s so much we can do with these language models that I think we are just not doing. And I’m, like, very excited about this because I think, like, I think we can get serious gains from very opinionated and good harness design that lends itself better to scale. And what by this is, like, the RLM, for example, is a very primitive inductive bias. Like, there’s nothing super special about the design other than the fact that it’s very different than what we currently do. But this may potentially scale much better with, like, the data and the environments that we have available to us.

Vibhu [00:40:20]: I guess the, opposite thing that people would probably ask is current harnesses Are very generalized towards coding, which people see works for a lot of domains. Cloud code is being used for design, presentations

Alex Zhang [00:40:33]: Yep.

Vibhu [00:40:33]: Everything. MuseSpark, Grokbot.

Vibhu [00:40:36]: These are very simple, non-opinionated harnesses that are good at code, and that is also scaling out. what’s the example of how we improve those, I guess?

Alex Zhang [00:40:48]: Yeah. let me bring up another paper, which came out very recently. It’s like the harness tax paper. I think it’s by Arena. I really like this paper because it puts forward a prior that I had, which is basically that, like, most

Vibhu [00:41:07]: So it confirms the prior.

Alex Zhang [00:41:08]: Yeah. Like, most harness choices don’t matter because

Vibhu [00:41:12]: Yeah.

Alex Zhang [00:41:13]: All of these harnesses are the same. But I will say, like, Grokbot, for example, is actually quite different, I think, from my understanding, than how some of these other harnesses have been designed, and I like that a lot. and I think it’s clear from here at least that, like, I’m pretty sure. Anthropic or OpenAI are exclusively training on their harnesses. They’re probably not training on their competitor’s harness. I’d assume not, because I don’t know why they would do that. But

Vibhu [00:41:38]: But, this is a thing you see in open models, right? Like, Qwen is really good at using open code.

Alex Zhang [00:41:43]: Yes.

Vibhu [00:41:43]: They need to train in harnesses. Old Gemmas were notoriously bad at this.

Alex Zhang [00:41:47]: Yeah.

Vibhu [00:41:48]: Models are good, but you need to train in a harness.

Alex Zhang [00:41:50]: I think, though, as models get smarter, or, like, as they get better, this distinction becomes, not that important in the sense that, like, if you take Astra and you put it inside of open code, like, it’s not gonna go crazy, because I think it’s, like, just sufficiently good. And so I say this because the only benefits between these different harnesses is just cost, for the most part. And I think, like, what I’m getting at, Sugru, is if you plug these models into RLMs, though, they’re not that good still. They’re okay. And I think it’s mainly because the types, like the class of harness that we are training around is this class of harness, this, like, pi loop, this, like. I like to call it trajectory as a prompt, which just means, like, you keep the whole trajectory of the rollout as the context that your main model is using. Even if you use subagents, it’s still, like, kind of this form. And I think we’re going to. If we want to explore new harnesses, like, there needs to be teams that are dedicated to actually running meaningful experiments over, like, scaling out new harnesses, like maybe post-train scaling out on different harness designs. Like, I think we can actually get very meaningful knowledge or gains from doing this kind of thing, whether it’s an RLM or whether it’s something different. And that’s exciting, ‘cause I think, for example, if you train a lot on. Fable for a long time was the best model for RLMs because they had dynamic workflows, and it was pretty obvious that, like, this was a capability that was somewhat trained in. Even if the model was still, like, a little dumb, like, in the RLM harness, it still worked a lot better than other models did. Astra is now also, like, good enough at doing these things. But

Swyx [00:43:33]: Wait, is this where we see that Fable is the best for RLMs, or is there some other

Alex Zhang [00:43:38]: Oh, no, these are all internal results, I guess.

Alex Zhang [00:43:40]: Yeah, I don’t, I don’t have them

Swyx [00:43:41]: Okay

Alex Zhang [00:43:42]: Public right now. But in general, like, I think you can, You can very easily tell that we have not optimized for RLM, like, workflows yet. And I think if we get models that do this correctly, they will be a lot more efficient as well. you can kind of just think through, like, why this is the case, right?

Vibhu [00:44:03]: I think this is the point where you have to give the ten-second what are RLMs.

What Is an RLM?

Alex Zhang [00:44:07]: Oh, yes.

Vibhu [00:44:07]: Because there’s a lot of listeners here that

Swyx [00:44:09]: Yes. We’re, we’re assuming a lot of knowledge.

Alex Zhang [00:44:10]: Yes.

Swyx [00:44:11]: Also, I think you. But you have set some context

Vibhu [00:44:13]: Yes

Swyx [00:44:13]: So you can. Like, with everything we just said

Vibhu [00:44:15]: Yes

Swyx [00:44:15]: Can we have a clean, crisp definition of RLMs?

Alex Zhang [00:44:18]: Yes. Okay. I want to go back to the blog, the

Vibhu [00:44:21]: Yes

Alex Zhang [00:44:21]: The compositional generalizers blog. This one. Okay.

Swyx [00:44:24]: Okay.

Alex Zhang [00:44:24]: This is, like, the best. I wish I had this in the original paper. An RLM is basically just a harness design where the only tool in the harness is code, which is this programmatic subagent calling thing, where it has the option to call itself as a tool, and it has other tools. But all of these things are functions in code, and the context that it’s dealing with is always stored in some memory inside of this code environment. So this could be a file system. Like, this could be, like. I’ll give you an example, Prime Agent. The trajectory of Prime Agent, like the context, even the. when you compact and do all these things, is stored on disk. So the model can always reference its original context, even if it’s compacted, and all of its tools are run inside of, let’s say, like, a Python REPL or a Bash REPL. And so this. it’s like this very primitive abstraction. And I would say, like, where most harnesses differ is, A, context offloading is not done that, like that. if you look at Prime Agent, by the way, Prime Agent does context offloading, but not all the way. So, like, it still maintains the standard cod code, Codex loop of, like, trajectory as a prompt where you compact, but it has the additional kind of, like, the context is offloaded, and it only has. The unique point of Prime Agent is that the only tool is IPython. So this is, like, the very kind of generic abstraction around, like, RLMs. Yeah.

Vibhu [00:45:59]: Concretely, what’s the core thing RLMs are trying to solve?

Long Context, Composition, and Locally In-Distribution Tasks

Alex Zhang [00:46:02]: Yeah.

Vibhu [00:46:02]: At one point, I think when it first came out, it was context.

Vibhu [00:46:06]: I’ll pass the question.

Alex Zhang [00:46:06]: Yeah. So the original problem was harnesses have a really bad time dealing with long context. They typically were used to only really do them for, like, specific things like code. Like, they could deal with your code base because it was trained on it. But now it’s more around what this blog is talking about, which is Compositionality and the fact that, like, I think harnesses. We want to have language model systems that have much more control over the actions they make at every step. And what by this is tool calls are very limited because you have to invoke them every turn. Like, you have to invoke tool A, then tool B, then tool C, and there’s no central context that you can kind of draw back from. And RLMs are specifically, like, designed around composition and having, like, a central context that you can always draw from. and this context is, like, designed around the existing language models. Another, like, very similar example actually in design is, like, agent swarms, for example, the Hugging Face incident. Like, these agent swarms have, like, a message board that they learn to communicate over. And this message board, in some sense, is the shared context that they, like, act over. And RLMs basically say that, like, the best way to communicate through this is in code. Like, you write the code to do this, and it’s, it’s because these models are so good at writing code. like, we wanna take advantage of that fact.

Swyx [00:47:37]: And then for this compositional thing, up to and including generating your own harness, specific for the task.

Alex Zhang [00:47:44]: Exactly.

Swyx [00:47:45]: Right?

Alex Zhang [00:47:46]: I think we will start to see that if you go up to, this figure. Okay. So we talk about this idea of locally in-distribution tasks for a harness, and it is like a, an idea on top of, like, in-distribution tasks. When we think about language models, an in-distribution task is just a task where, like, the prompt is something that the model has either seen before or, like, has seen some version of it. Most harnesses work like the one on the right, which is they keep appending the trajectory as a prompt. And so eventually, unless you’re Anthropic or OpenAI and you train on, like, kind of these, like, user trajectories, most of these things end up being out of distribution for the most part. But locally in distribution is basically the compositional argument of if an RLM breaks down its computation into, like, kind of a meta-harness of sorts or, like, a program that involves subagents that, like, look at a local problem, every individual language model call over the course of this task is in distribution, even if the entire task itself is out of distribution. And this is a very desirable property, I think, for obvious reasons. Like, if every task is in distribution for each individual language model call, you will probably get to the right answer. so.

Swyx [00:49:00]: The logical limit of RLMs is RLLMs where, like, you not just, you don’t just write the harness, you also train a custom model for

Training RLMs and Smarter Harnesses

Alex Zhang [00:49:11]: Exactly.

Swyx [00:49:11]: You collect data, everything.

Alex Zhang [00:49:14]: Yeah.

Swyx [00:49:14]: Like, it’s a fully automated AI researcher inside of your harness.

Alex Zhang [00:49:16]: Yeah. We will see where the training of RLMs goes. I will say, as an academic, I am not working on this at MIT, or at least in the scaled sense, because I can’t afford to. but there are companies out there that are working on this. I think Prime and Select is very clearly working on this, and it’s very cool. Like, I’m, I’m very excited to see. Maybe we’ll observe, I don’t know, but maybe we’ll observe better, like, post-training scaling laws with when you train around a smart harness. Maybe we’ll even see smarter harnesses that come out and, like, they work better around these kinds of principles.

Swyx [00:49:50]: What is a smarter harness? Like, that doesn’t mean anything. You just said they’re all the same.

Alex Zhang [00:49:54]: No. What, more of what is, like, Claude Code, Codex, Pi, et cetera, are all the same in that, like, when you break down the logic of the harness, it’s, like, virtually the same thing.

Swyx [00:50:06]: Yeah, two calls in a loop or

Alex Zhang [00:50:07]: Yeah

Swyx [00:50:08]: Whatever.

Alex Zhang [00:50:08]: But With RLMs and with other harness abstractions, it looks very different. And this is where I think you really distinguish. it’s, it’s in the same way that, like, I think with language model architecture choices, a lot of architecture choices end up kind of looking the same when you, like, scale it out or, like, it. The differences end up being, like, somewhat minor in terms of. for a lab it’s not minor, but, maybe one model converges better than the other one, like, slightly. But in general, like, if you were. So for example, pre-training scaling laws only hold because the architecture choices we have are somewhat stable right now. But if you were to completely change the architecture, pre-training scaling laws probably don’t hold. Or, like, these kinds of. This, like, power law is gonna look very different. And it’s like the same thing with harnesses. Like, I think all the harnesses we have right now, for the most part, roughly look the same, but there are some exceptions to this, I think, that are coming out.

Swyx [00:51:06]: I was gonna say, I actually, one of the things that I’ve been more interested by, like, talking about PhD students who take big risks, is that people have been. People also pursuing the other side, which is pre-training scaling laws don’t hold if you change data. right now it’s just raw, unstructured text, corpus of internet. What if you had a better data representation to train on? Then yes, your scaling law would change as well.

Alex Zhang [00:51:28]: Yeah.

Swyx [00:51:28]: So there’s architecture, there’s data, and, whatever else, you can think about. Well, so I just wanna get back to this. it all makes sense. It’s, it’s very interesting how you sort of recurse up and down the stack from, like, very conceptual to, like, not like, well, this is where we are today.

Prime Agent and Opinionated Harness Design

Swyx [00:51:43]: But, like, yeah, obviously, it can scale up and down. I guess, I’m curious, how did you start working with Prime? Is Prime taking on more work with this? Is this their answer to Hermes agents?

Alex Zhang [00:51:56]: Mm.

Swyx [00:51:56]: You mentioned Grokbot is a little bit different. I just wanted to, like, namecheck all these guys and get your thoughts on each.

Alex Zhang [00:52:01]: Yeah. I got involved with Prime, after they released a blog post, by the way, not affiliated with me at all, about, like, how they believed RLMs were kind of the future. And I had a friend that was working there, GPU Mode, Matei. Like, we got in touch, and I think I agreed with a lot of the researchers there and, like, what they believed about harness design. Like, I was very impressed, I think, that, like, they understood the purpose of the RLM paper, which is not necessarily just to say that, like, we’re solving long context tasks, but actually, like, we want more opinionated harness designs.

Swyx [00:52:39]: Yeah. There’s always, like, the result of the paper that you choose to highlight

Alex Zhang [00:52:42]: Yes

Swyx [00:52:43]: Versus the actual point.

Alex Zhang [00:52:44]: Yes. As I would love to talk about, like, the incentives of academia and, like, the things around, like, why it’s kind of flawed and all the issues, and we’ll get back to that. Yeah. So anyways, I love the guys at Prime. So we kind of had been. After we decided to work together, we decided to look into training in RLM and also build this kind of RLM harness and kinda see where we can take it. That is how, like, Prime Agent came about, and I think the reception for Prime Agent has been pretty good. Like, the one thing I was worried about with Prime Agent is that none of them, at least at the time when we were building it, none of the models were that good at doing RLM stuff. So this was, like, pre-Fable, pre-Astra.

Swyx [00:53:27]: I guess, I think to take a step back, can you explain what Prime Agent is, how it’s different than

Alex Zhang [00:53:32]: Yeah

Swyx [00:53:33]: A traditional, Claude Code, what people would expect harness?

Alex Zhang [00:53:36]: Yes. So Prime Agent, I think I mentioned this a little bit earlier

Swyx [00:53:40]: Yeah, there was the diagram. Yeah

Alex Zhang [00:53:41]: Is basically. it is a. A harness on top of Pi, like Pi Mono, which is-- Pi Mono, for context, is like the, like a

Swyx [00:53:50]: Core agent

Alex Zhang [00:53:51]: A minimalist

Swyx [00:53:51]: Yeah.

Alex Zhang [00:53:52]: Yeah, like harness. I use Pi as the reference for everything because I think all other harnesses are basically just Pi.

Alex Zhang [00:53:57]: But it is Pi, except we explicitly restrict IPython to be the only tool that’s available to it. Every other tool gets loaded in as, like, a Python module, or like a Bash kind of script that it can run. So it uses the core RLM abstraction on top of Pi, and then it also has this continual harness thing, which is, Seth, he’s another PhD student. This is a thing that he used to get language model harnesses to play games. Like, so he worked a lot with Joel, who is the, like, Gemini plays Pokemon guy. And continual harness is also, by the way, very simple. I quite like it. It basically is this design, principle around, like, what parts of the harness can you let the harness itself modify? There are certain pieces that, like, you’ll let it modify its own skills, the subagents available to it, what the system prompt to the model is. And continual harness is available basically as a tool inside of the IPython kernel. And so that’s what Prime Agent is, like, how, what it’s designed around. Everything else in Prime Agent is like

Swyx [00:55:05]: Standard.

Alex Zhang [00:55:06]: Standard.

Swyx [00:55:06]: Standard.

Alex Zhang [00:55:06]: Right? Yeah. I think what is, what I really liked about it, and we got kind of lucky, is that, like, a lot of the new frontier models actually work really well inside. And actually, even a lot of the open-source models work really well, at least some of the newer ones. And there is another thing in Prime Agent I should highlight, which is that, like, we have a very particular agent-to-agent communication system or, like, framework, which is because RLMs tend to spawn many subagents, we want a way for subagents to communicate with maybe the root or with each other. And so there are some design decisions around, like, what each subagent is allowed to talk to, how it does it. Again, everything is in code, so it writes the code to do this kind of communication, which I think is really cool. And then there’s, I guess, persistent subagents is another thing that was kind of added, which is the subagents, they can last beyond, like, the standard runtime of the actual, like, original agent. And you can go into that subagent, you can prompt it more, like, you have more visibility and flexibility into what is kind of going on.

Swyx [00:56:10]: This is my number one pain with Codex right now. They, their subagents are just very ephemeral And they actively discourage you from using it for long-running things.

Alex Zhang [00:56:17]: Yeah. Yeah. Which I think it makes sense. Yeah. I

Swyx [00:56:21]: So the trick is just externalize to a file system, right?

Alex Zhang [00:56:24]: Yes. Yeah. That’s

Swyx [00:56:25]: Like, that’s the trick.

Alex Zhang [00:56:25]: That is the big trick.

Alex Zhang [00:56:27]: Yeah.

Swyx [00:56:27]: And well, and also, like, force everything to run through code. trust the model

Alex Zhang [00:56:30]: Yep

Swyx [00:56:30]: That can write code, and it’s gonna write its own harnesses itself. So is Prime gonna take on, like, training, post-training custom models for this? Is this a one-off collaboration between you guys, that’s it? Like, what’s

Alex Zhang [00:56:41]: Yeah. They are training a model, intern-- I think they were pretty public about this actually

Swyx [00:56:46]: Yeah

Alex Zhang [00:56:46]: Back in March.

Swyx [00:56:47]: Clearly it is their business.

Alex Zhang [00:56:49]: Yeah.

Swyx [00:56:49]: Yeah, so.

Alex Zhang [00:56:50]: Yeah. they’re, they’re showing that they can train it on their kind of hosted training stack. But no, so for model training, I’m, I’m not involved with them on that. The main reason is just I have other things in the PhD I wanna work on. I think, like, there are many other big bets to take,

Swyx [00:57:03]: Ooh

Alex Zhang [00:57:03]: Outside of just RLMs. some

Swyx [00:57:06]: Ooh

Alex Zhang [00:57:06]: Some I don’t know how much I can share yet. but in general, like, I think, I actually think one of the luxuries of being a PhD student, genuinely, is that there’s so many big bets to take. most of them will probably yield nothing, but it’s a really exciting time to be in research, especially because I think most progress in the field has been a little bit boring. Like, I’m not saying the outcome is boring, but the process of doing these things tends to be quite boring. and so there is kind of this question of, like, what do we wanna do next? but

Swyx [00:57:39]: Yeah

Alex Zhang [00:57:39]: We can talk about that later.

Third-Party RLM Work: Harvey, Headlong, DSPy, and ARC-AGI-3

Swyx [00:57:41]: Yeah.

Alex Zhang [00:57:41]: Yeah.

Swyx [00:57:41]: Okay, I wanna close out a little bit more of your research, and then we can,

Alex Zhang [00:57:44]: Cool. Yep

Swyx [00:57:44]: Start putting it out. since you released RLM, a lot of excitement about it. Any secondary third-party work that you wanna shout out as, like, that you guys should take a look at this?

Alex Zhang [00:57:53]: Oh, yeah. So, Harvey, the legal AI company, released a blog post, not affiliated at all, but they post-trained an RLM on their, like, legal work, which often involves a lot of, like, sifting through documents and kind of looking through, like, a variety of, specific information that maybe is not so easy to retrieve with, like, a pure retrieval system. And they show, like, really good results. It’s very exciting. I was shocked that they worked on this. They did not tell me, so when this came out, I was like, “Oh, that’s awesome.” So there’s this one I think is super cool and what they’re doing there. I think this is a collaboration with Base 10, by the way, as well.

Swyx [00:58:33]: Yes, this was Base 10.

Alex Zhang [00:58:34]: Headlong, which is law, the Law Institute’s kind of. it is their, like, persistently running harness. it’s very cool that

Swyx [00:58:42]: Oh, they renamed it? They used to call it something else.

Alex Zhang [00:58:45]: It was like Auto

Swyx [00:58:46]: Terminus.

Alex Zhang [00:58:46]: Yeah, I know. They’ve gone through. Yeah.

Swyx [00:58:49]: All right.

Alex Zhang [00:58:49]: So this is Andy Konwinski’s big project. it’s super cool. I love Andy. I don’t want to downplay what they’re doing because they’re using the RLM abstraction, but they’re doing something much cooler than the RLM, which is like they have a system that kind of what they call, like, thinks persistently. So even when you don’t query it has a way to, think through problems that it has in its context.

Swyx [00:59:16]: Oh, so it’s just like a always-on type thing.

Alex Zhang [00:59:18]: It’s like an always-on thing, but it’s, like, not that expensive. Like, they control the token costs, to make sure it’s not, like, burning through all your credits. This is super cool. I’m trying to think. There are many. Actually, if you go to the RLM, GitHub page, there’s a bunch of things I’ve linked, below. There’s a ton of really cool kind of things that people have been doing. Axe is another really cool one that I think it’s just by this one guy. It’s like a harness around DSPy and RLMs. DSPy also has an RLM. Oh, the last thing I’ll shout out is on ARC-AGI-3, I believe, there were a lot harnesses on their, like, Kaggle competition, like the official one, not the, like, public primates, like one that, or like what people have evaluated on. They all, like, claim to use or they reference, like TUFA, for example, some form or some inspired form of the RLN abstraction in their harness, which is really cool. I think it’s, This is where-- this is exactly the setting where you would see a lot of benefits from composition and using code and combining, like, neuro symbolic systems with AI. And so

Swyx [01:00:27]: Yeah.

Alex Zhang [01:00:27]: Yeah. Very cool.

Swyx [01:00:28]: We love a good neuro symbolic reference.

Agent Swarms, Unsolved Math, and What the User Should See

Alex Zhang [01:00:30]: Yeah.

Swyx [01:00:30]: You also, mentioning ARC-AGI-3, OpenAI comes out and says, “We’re at 99.9% on this.”

Alex Zhang [01:00:36]: Yep.

Swyx [01:00:37]: They also say, “We solved Navier–Stokes. We just threw a model at it.”

Swyx [01:00:40]: There’s some debate around whether or not it’s just model.

Swyx [01:00:44]: Are they using an RLM? Do?

Alex Zhang [01:00:46]: I would guess probably not, unless you say, like. I’ll, I’ll be, I’ll be careful here because, people debate what is an RLM, what is not an RLM. It’s somewhat clear that what they used is some kind of swarm of agents with a shared, some shared context, like some shared file system. And, like, this is very much in the spirit of RLM stuff, but I think there’s a, there’s a lot of, like, more clever things that they did that’s not maybe related to the RLM itself. I agree a little bit with the idea that, like, a harness is not that necessary for what they did. The way that I would put this is that I think a model, like a GPT-6 Astra type thing, is technically smart enough, conditioned on the right information, to come up with a proof for these very difficult problems. Now, how you get to that information is a giant question mark. And in their case, it probably came down to, like, a very long search over, like, many of these sub-age-- or many of these, like, agents in the swarm and maybe also, like, researchers cond-- I’m, I’m actually not sure about this part, but putting in, like, their kind of intuition as to, like, what you should explore and things like this. And ultimately, like, this produced some information that some agent was able to take to finish the proof. And so in that sense, like, I think, was the harness that important? No. And I think what this is pointing at is, like, the specific details of a harness do not really matter, and I think that’s also what that, what the harness task paper is pointing at, which is that, like, beyond the user’s feeling of the harness, realistically all that matters is just, like, how are you composing these agents in a meaningful way to get to the final answer? And maybe that’s what, like, swarms and all these things are really about. And so from my POV at least, if we start to think about, like, for user use cases, what do we want out of harnesses and things like this? Like,

Alex Zhang [01:02:51]: We want to take the good parts out of these, like, the Claude codes, the Codexes, like the stream that people like to see. But, like, under the hood, whatever is running can be some really weird, complicated swarm of agents that, like, ultimately come up with an answer. The user doesn’t wanna see that, though, obviously, right? Like, it’s, it’s not legible information. And so I-- this was another kind of thing in the spirit of RLMs, like recursive language model. It sounds like it’s a language model and, but it’s not a language model architecture. But the reason for this is, like, I think we will start to see in the future probably one day, and I wrote a blog about this, what we think of as a language model, like the thing that we query, might actually be like a swarm or like a scaffold or like some Weird harness design that scales very well, but the user just doesn’t see it. Ultimately, all the user sees is some front-end version of this harness. And yeah, I think it’s a relatively safe bet at least to make that this is what we will see.

Swyx [01:03:48]: Yeah.

Alex Zhang [01:03:49]: And this maybe goes back to the limitations of the base transformer. Like, obviously if you just took a base transformer and you said, like, “Solve Navier–Stokes,” or something, it’s not gonna do it. Like, yeah, we all know this is not what’s gonna happen. But yeah, I think this is maybe the more interesting part. and maybe the claims around, like, did the harness matter is more around this, of, like, just arbitrarily pointing models, like, or agents at a growing kind of context of information maybe is just enough to solve very difficult problems. that I can buy.

Swyx [01:04:20]: While there’s a lot in there, I do wanna say, the amount that OpenAI spent is semi-public. It’s, 10,000 agents in 88 hours.

Alex Zhang [01:04:29]: Oh, yeah.

Swyx [01:04:29]: 130 billion output tokens, which is estimated to be about 40 million dollars in public pricing.

Alex Zhang [01:04:34]: Surprisingly, actually, like, less than I thought.

Swyx [01:04:37]: Yeah, not that much.

Alex Zhang [01:04:38]: Yeah. Yeah.

Vibhu [01:04:38]: 130 billion output for the final, but as you said, there was a lot of context being passed around.

Vibhu [01:04:44]: It’s more than double that in just the total agent messages being sent.

Alex Zhang [01:04:47]: Yeah.

Swyx [01:04:48]: Yeah.

Alex Zhang [01:04:48]: Yeah. Yeah.

Swyx [01:04:49]: I think, one thing I was honored to bring up also was Cursor as far as, like, swarm stuff is concerned.

Swarm Architectures, Coordination, and Token Efficiency

Swyx [01:04:53]: This is slightly older, meaning February, which is ancient.

Alex Zhang [01:04:57]: Whoa.

Swyx [01:04:57]: But if you scroll down all the way to the final sort of multi-agent architecture that they arrived at, it was basically an org chart of a normal software team. One thing I’m thinking about, because I basically. There’s, like, this, gather all function That you have to do with subagents or. It’s very similar to GPU programming, actually.

Alex Zhang [01:05:16]: Yeah. Yeah. Yeah.

Swyx [01:05:17]: And so, that’s a bottleneck. This is a bottleneck.

Swyx [01:05:21]: If there’s one main guy, that’s coordinating all the sub guys, then they have to, like, gather again and then re-coordinate.

Swyx [01:05:29]: That’s slow. That’s, that’s crappy. what a true swarm should be is everyone is just their own person.

Alex Zhang [01:05:35]: Yeah. Yeah.

Swyx [01:05:36]: Right?

Alex Zhang [01:05:36]: Well, I agree with this, and I think that there is a question to be had, though. Let me give an analogy, which is like, when would you use compaction and when would you use an RLM? And there are many settings where an RLM can solve maybe a more difficult task than compaction can, but you would prefer to use compaction in most cases because it’s cheaper and it’s quicker. And I think in the context of agent swarms, there is a similar thing going on of, like, I’m. Fairly certain that, like, 95% of the swarm is entirely useless, or, like, what it’s exploring is entirely u-- You’re just burning tokens. Versus in this setup, maybe not so much. I’m not sure. maybe it’s also the case here.

Swyx [01:06:17]: Everyone has a job. This is your board.

Alex Zhang [01:06:19]: Yeah.

Vibhu [01:06:19]: I think at some level that’s how problems are framed, right?

Alex Zhang [01:06:23]: Yes.

Vibhu [01:06:23]: So, like, if you have a search problem and you’re spanning out a bunch of subagents to do search, there’s gonna be a lot of useless information, right? There is one retrieval answer That you’re getting and you’re spanning off, but that is consciously understood, right?

Alex Zhang [01:06:36]: Yes. But there is kind of this question of, like, what is appropriate to solve for which problems? Like, what design-- in theory, OpenAI can use-- can package up this API and they’ll call it swarms, and then they’ll give it to you and they’ll be like, “Point this at any problem and we’ll give you a solution.” But maybe

Swyx [01:06:53]: Yeah, it’s called, it’s called pro, right?

Alex Zhang [01:06:54]: Yeah. maybe you’ll have to pay like 40 million dollars to get a result.

Alex Zhang [01:06:57]: And it’s like, well-- But it’s exciting. I will say, like, it is-- it’s very exciting that we even have the option to point 40 million dollars at a problem and solve it.

Vibhu [01:07:07]: Yes.

Alex Zhang [01:07:08]: But there is still kind of, a lot of research to be done in this area around, like, what is necessary. Like, what do we want to do? What design do we want? we probably don’t want everything to be a swarm, but, like, where do we draw the line? Like, can the agent design that or decide that for itself? Et cetera. So.

Open-Endedness and Research Without a Fixed Objective

Swyx [01:07:26]: Yeah. and then, just a quick check in case you have opinions on this. Have you looked much into open-endedness as a general category of problems?

Alex Zhang [01:07:35]: Oh

Swyx [01:07:35]: Meaning no prompts, just go.

Alex Zhang [01:07:38]: A little bit. so I was at Sakana for a summer, right after, or I guess right before my PhD, and that’s something that they work on a lot there. And I think there’s a lot of people even at, like, Recursive Super Intel-- There’s many of them now.

Swyx [01:07:53]: Yes, we just had Richard Socher on.

Alex Zhang [01:07:55]: Oh, yes. Yeah. So, like, Richard’s company and then also. Actually, wait, that might be the s-- it might be the same company. I don’t remember. Is Tim Lautenschlager also

Swyx [01:08:03]: Yeah.

Alex Zhang [01:08:03]: Okay. Yes, that company.

Swyx [01:08:04]: He’s the main co-founder. He used to be head of open-endedness for Google.

Alex Zhang [01:08:07]: Yeah.

Swyx [01:08:07]: Yeah.

Alex Zhang [01:08:08]: So I think with open-endedness problems, like, I view them as somewhat similar to even, like, unsolved math problems. Maybe that’s a weird way of framing it, but, like, I think a lot of the techniques in terms of, like, how people like, approach them are kind of the same. Like, evolutionary search is, like, very similar to just launching swarms of agents and hoping that, like, they come up with, like, an interesting s-- And this is what, like, AlphaEvolve and some of these other works did, like a year or two ago. But I think what maybe is not, And I’m not sure if this is what you were alluding to, but I think what’s not super clear in open-endedness style search problems is do we frame all of them as basically like an unsolved, very difficult problem, or, like, where the objective is clear? If that’s not the case, I still don’t know yet entirely what the value of it is. maybe you have other opinions. Like, I don’t have too many opinions on this, but at least from my time at Sakana, like, I got the sense that, like, we ultimately still kind of wanna approach things the way that, like, say, OpenAI approached Navier–Stokes. We want. We-- There’s still a lot of nudging in certain directions that we want to have to, like, get to the point where we have something interesting.

Swyx [01:09:29]: My version of it is, like, maybe it’s a split between basic science and applied science. Basic science, you’re researching for researching’s sake.

Swyx [01:09:36]: You just wanna understand things better. I have no idea if, like, there will be any application at all whatsoever, but, that, And then applied, you have a goal.

Swyx [01:09:45]: You’re, you’re trying to minimize loss in some way? and so, what I really, think, in terms of, like, the big bets that people have, what if there was no prompt? Like.

Alex Zhang [01:09:57]: I see.

Vibhu [01:09:58]: You just pick domain and let it

Swyx [01:09:59]: Like, you just, like, you just spawned in this, like, swarm of things and you’re like, “Hey, what’s up, guys? Like, what you guys working on?”

Alex Zhang [01:10:03]: Yeah.

Swyx [01:10:03]: And, like, you just decide

Alex Zhang [01:10:05]: I see

Swyx [01:10:05]: Like, this is an interesting problem.

Alex Zhang [01:10:06]: The biggest issue that, like. And maybe there’s a clear path to this, but when I was there at least, the biggest issue that people had with open-endedness was like, how do you ultimately choose at the end? How do you pick out the interesting stuff? Because when there is no. Like, maybe the agent comes up with a goal, but in a lot of cases, like, what they. And they have something called Fugu, I think, which is like a It’s like a model router type thing that was, like, inspired at least by this idea of, like, let’s pick a problem where maybe we can pick out the best solutions to something. In this case, it’s like pick the best model for this problem.

Swyx [01:10:41]: You’re the first person to connect model routing to open-endedness.

Alex Zhang [01:10:43]: No, yeah. But, so I bring this up because I think, like, with open-endedness, like, just generally the issue is, like, when we have this giant corpus of, like, slop, like, how do we sift through

Swyx [01:10:57]: Yes

Alex Zhang [01:10:57]: And find, like, the hidden gems? And, like, the solution to open-endedness really just letting models run forever and, like, finding. Like, just doing data gen-- just doing super high throughput data generation and then, like, asking agents to go through and, like, find meaningful things. Like, I’m not sure. Maybe that’s sufficient? Like, that would be, that would be cool.

Swyx [01:11:21]: To me, it’s, like, very interesting as a counter to basically all of machine learning Where you have a goal, to have no goal.

Alex Zhang [01:11:28]: Yeah. Yeah.

Swyx [01:11:29]: But, or, like, an ill-defined goal that you’re like, “Well, what about this goal?” And you’re like, “Well, okay, maybe.” And then you, like, sort of research more and you find

Alex Zhang [01:11:36]: Yeah

Swyx [01:11:36]: That is an interesting goal. ‘Cause, like, I think, like, finding the objective function, like you said, like, Jeff found an objective function That was interesting that no one was exploring.

Alex Zhang [01:11:43]: Yep. Yeah.

Swyx [01:11:44]: I think that is, like, similar to your message about grad students as well. Like, you stay in school because you are. you want to pursue open-endedness. If you want to, profit max and, like, join the, escape the permanent underclass, then you join a lab.

Swyx [01:11:59]: Right?

Alex Zhang [01:11:59]: Yeah. Yeah. It’s funny. I feel like I don’t, I don’t hear this discourse a lot. I’m, I’m in the East Coast, so it’s, like, a very different type of. But then when I. whenever I come here, it’s like, that’s always, like, the topic of discussion.

Swyx [01:12:12]: You cannot pay rent without doing this.

Alex Zhang [01:12:13]: Yeah.

Swyx [01:12:15]: Yeah, you’re getting priced out, guys.

Vibhu [01:12:16]: Yeah.

Swyx [01:12:17]: Okay. So, yeah, there’s, there’s all that. I don’t know if you wanna-- if it’s relevant, enough to talk about the mismanaged geniuses, which you were pulling up.

Vibhu [01:12:25]: No, it’s just on your blog. But I will poke on, Sakana.

Swyx [01:12:28]: Oh, Sakana? Oh, okay.

Vibhu [01:12:28]: Yeah. So they did in their, blog post, I talked to them about this as well. So one of the cool results of Ultra is they basically just let it loose on automated data science research

Sakana AI and Weird Research Bets

Vibhu [01:12:40]: With little to no human intervention. It’s just kind of making

Swyx [01:12:44]: Yeah, so this is auto research, which is a little bit more open-ended, and there’s degrees of open-endedness, and I agree with that.

Vibhu [01:12:51]: Yeah, separate than auto research with objective, this is just

Swyx [01:12:54]: Yeah

Vibhu [01:12:54]: Do stuff. But, it’s cool. They’re, they’re working on it for those that are interested.

Swyx [01:12:58]: While you’re bringing it up, actually, what is your take on Sakana? Like, what are they doing apart from being, the Japan one?

Alex Zhang [01:13:04]: Yeah.

Alex Zhang [01:13:05]: I actually love the people there. Like, I think they have a really smart team. and it makes sense. it branched off from, like, an earlier team at GDM, which was also kind of, I guess, doing this kind of, like, open-ended evolutionary research style stuff. What I liked about my experience there, at least, was that they did have that, like, mishap back, I forget, at this point when, but I think, like

Swyx [01:13:31]: You’re talking about AI scientists?

Alex Zhang [01:13:32]: No, the GPU kernel.

Swyx [01:13:34]: Oh, okay, yes.

Alex Zhang [01:13:35]: That one. Yeah.

Swyx [01:13:36]: People cannot forgive them for that. Yes.

Alex Zhang [01:13:37]: Yeah. And I guess, like, the AI scientists, like, there’s, there’s some criticisms of it that I don’t, I don’t work on that, so I have no kind of take on it. But I think in general, like, what I like about them at least is that they’re a little bit more of a researchy type lab. So, like, they don’t operate in the same space as, like, OpenAI or Anthropic. Like, for sure, like, definitely no. They do not. it’s pretty obvious probably that, at least when I was there, they do not have a big competitor model or something that, like, that everyone is using. But I think they kind of operate in some ways as, like, a PhD lab, which is cool. Like, and I think, like, David Ha is, like, he’s, he’s really smart. Like, I think he has a good sense of, like. Also, I think the market in Japan is also a little bit different for AI, and, like, who they’re targeting is slightly different than maybe what we’re used to here. But yeah, I like that they take kind of. A lot of their research is kind of weird, I think, when people view it? And I like that. Like, I think it’s

Swyx [01:14:34]: We should have more weirdness, yes.

Alex Zhang [01:14:35]: Exactly.

Swyx [01:14:36]: And you said different market. Just, is it, like, enterprise?

Alex Zhang [01:14:39]: Like, the way it works there is a bit different, like how deals happen and stuff like that.

Vibhu [01:14:44]: They do have a. I guess this page is originally in Japanese, but they do have a model

Alex Zhang [01:14:49]: Oh

Vibhu [01:14:49]: Specialized for the Japanese market.

Swyx [01:14:51]: So you didn’t know that.

Alex Zhang [01:14:51]: I didn’t know that.

Vibhu [01:14:52]: I didn’t know it too.

Alex Zhang [01:14:53]: Did not know this.

Vibhu [01:14:53]: I also have personal friends that know the team.

Alex Zhang [01:14:56]: Yeah.

Vibhu [01:14:56]: So there is a. Even from the sense of a way that you speak culturally Responses are tuned towards that. This is not like it’s frontier on benchmarks. It is a cultural appropriate model for them, and then they have, like, chat and all that.

Alex Zhang [01:15:11]: Yeah.

Vibhu [01:15:12]: But, to mirror your point, there’s also, like, how should education look like? And someone wants to work on it, and they’re a very PhD lab of, “Do your thing. Why not? We have money. Go research.”

Swyx [01:15:22]: Oh, they say it’s a Kimi fine-tune. That’s nice.

Vibhu [01:15:24]: Oh, there you go. Kimi.

Swyx [01:15:25]: Good for them.

Swyx [01:15:26]: Yeah, and, speaking of Kimi, right, like another, just a grad student that spit out and like, yeah, I just have this, like, Kimi delta attention that wants-- that I wanna work on.

Vibhu [01:15:36]: Yeah. Yeah.

Swyx [01:15:36]: And, like, somehow managed to make Moonshot. Don’t understand it still.

Alex Zhang [01:15:41]: Yeah. Well, he’s, he’s super cracked, at least my understanding. I think in general, like, a lot of, a lot of the Chinese labs have done really cool work.

Kimi Swarms, Dynamic Workflows, and Convergence

Swyx [01:15:51]: Yeah.

Alex Zhang [01:15:51]: Like, yeah.

Vibhu [01:15:52]: Any thoughts on Kimi agent swarms?

Alex Zhang [01:15:55]: Yes. one thing I will say is whatever OpenAI is doing with their agent swarm is, like, clearly the right thing to do. You have to kind of think about it this way. Like, nothing, especially without, like, a very smart harness design, which I don’t, I don’t think anyone really has so far, Nothing comes so easily for free. For example, like, the agent swarm design is not something you can just take for granted. Like, it’s not like GPT-6 Astra is just super smart and then it just got agent swarms running well. They clearly trained. the Hugging Face incident was them training a system to be like a swarm. and I think, like, clearly they’ve done something really well to the point where you can throw 40 million dollars and solve an unsolved problem. And I think with the Kimi agent swarm thing, like, at least from when I read it just came off as like, this is interesting, but I don’t actually know whether or not this can solve anything novel.

Swyx [01:16:59]: Yeah, they just. They were like, “It makes spreadsheets for you.”

Alex Zhang [01:17:01]: They kind of were just like, yeah, like, here is a, here is a swarm that, like, kind of does stuff, and it’s cool. but. And I will say the same thing with dynamic workflows. I actually kind of think the dynamic workflows release was sort of a flop. I don’t know how you guys feel about it, but I think it’s like. my understanding is, like, it’s not used that often or, like

Swyx [01:17:21]: It’s just very expensive.

Alex Zhang [01:17:22]: It’s too expensive and like

Swyx [01:17:22]: It’s ultra code. It’s basically like, take over my bed.

Alex Zhang [01:17:26]: And it doesn’t I’ve tried it, and, like, it doesn’t act in the way that, like. Again, I’m not, I’m not the biggest OpenAI, like, stan or something, but I think whatever they did was very impressive. Like, they somehow managed to get a way for this swarm to actually act

Swyx [01:17:41]: I see

Alex Zhang [01:17:42]: Towards a goal. And yeah.

Swyx [01:17:44]: I see.

Alex Zhang [01:17:44]: It’s very difficult.

Swyx [01:17:45]: I see. So, like, efficiency of the multi-agent swarm is the objective function here.

Alex Zhang [01:17:51]: Yeah.

Swyx [01:17:51]: Right?

Alex Zhang [01:17:51]: Yeah.

Swyx [01:17:52]: Like, how much of this is slop? Like, this is a lot of slop.

Alex Zhang [01:17:55]: Yeah, I think so.

Swyx [01:17:55]: OpenAI is less slop.

Alex Zhang [01:17:56]: We take for granted what it means for a swarm to converge to an answer.

Swyx [01:18:00]: Yeah.

Alex Zhang [01:18:00]: It’s just like. It’s not something we take for granted.

Swyx [01:18:02]: Yeah. Yeah. We’ve, we’ve done one pod with Noam Brown and, like, his

Alex Zhang [01:18:06]: Oh, yes, I remember. Yeah

Swyx [01:18:06]: His thing. His whole thing was like, okay, like, we’ve worked on a lot of, like, competitive agents. we’re working on collaborative agents.

Alex Zhang [01:18:13]: Yeah.

Swyx [01:18:13]: And, like, that’s now called a swarm.

Gemini, GDM, and Harness Engineering at Scale

Alex Zhang [01:18:15]: Yeah.

Vibhu [01:18:15]: I would do a quick poke. Do you have any thoughts on Gemini? They were also IMO gold. Like, there was a time where they were getting agents to reason for a long time and.

Alex Zhang [01:18:26]: Yeah,

Vibhu [01:18:28]: Is it too old to think about?

Alex Zhang [01:18:29]: No. I think it’s a little bit blown out of proportion. Like, Gemini. For, again, I, so I should preface by saying I haven’t worked at any of these places, so take this with a grain of salt, right?

Vibhu [01:18:39]: You have strong opinions on agent harnesses

Alex Zhang [01:18:42]: Yeah

Vibhu [01:18:42]: And, this was

Alex Zhang [01:18:43]: But, let me just say this first. This work was really impressive. I think what they showed here was, like, they took a time when the models weren’t that good

Vibhu [01:18:53]: Yes

Alex Zhang [01:18:53]: And they managed to be very smart about, like, what the harness does. I remember for this at least, like, yeah, like AlphaGeometry, I guess that was a year before this, but it was very cool. They took it to the max, and they, like, designed, I don’t know. I’d say I’m not too big on competitive math, but I think, like, GDM, it’s sort of a shame. Like, everyone I’ve talked to about GDM kind of has the same opinion, which is that it’s way too, like, bureaucratic. Whatever is they have the talent and the resources to do almost anything, but, like, I don’t know, until they figure that part out, like. Nothing against anti-gravity, for example, but, like, I don’t know anybody that uses anti-gravity. And so I’ve tried it once, and it’s. I don’t see a reason to switch to it. and I think for whatever reason, like, they’ve been struggling with this, so, yeah.

Swyx [01:19:40]: Yeah. Well, a lot of people dogged on Meta for a long time until they started

Alex Zhang [01:19:44]: Yeah, and they recovered

Swyx [01:19:44]: Coming out. And, like, I think, Google’s going through that phase right now.

Alex Zhang [01:19:47]: Yeah.

Swyx [01:19:48]: And, it’s, it’s just you gotta stay alive and

Alex Zhang [01:19:51]: Yeah.

Swyx [01:19:52]: I wanna focus back on

Alex Zhang [01:19:53]: Yeah

Swyx [01:19:53]: Just, like, your thoughts, just general. we can talk about speculative PTC

Speculative PTC and Parallel Tool Execution

Swyx [01:19:58]: Mismanaged geniuses, or just, like, throw away all this and just talk about whatever else.

Alex Zhang [01:20:03]: Okay, let’s, let’s talk about mismanaged genius for a little bit.

Swyx [01:20:06]: Yeah.

Alex Zhang [01:20:06]: I, the only comment I’ll say on speculative PTC is that it’s a really simple idea. It’s almost, like, obvious that this should be done, and, like, there’s not much more to talk about it. Like, I think it’s just, like, you should just use it for, like, coding. Like, anything with programmatic agent calling, like RLMs or Kodak, like, yeah, it’s like, it’s like a no-brainer.

Vibhu [01:20:25]: What’s the, for people that haven’t read it

Alex Zhang [01:20:27]: Yeah

Vibhu [01:20:27]: What’s the one-liner for people?

Alex Zhang [01:20:29]: The simple thing is when the model is, writing its code or, like, even as, like, after it finishes writing the code, a lot of tools tend to be, like, sequential or, like, you have to wait on them, so you should just launch them in advance. Like, if you’re able to jit compile this code, you can probably figure out, like, even though it’s, like, kind of variables and stuff, like, you can figure out, like

Swyx [01:20:52]: Yeah, statically analyze.

Alex Zhang [01:20:53]: Yeah. So

Vibhu [01:20:54]: Speculation.

Alex Zhang [01:20:55]: Yeah. There is this, Someone pointed to me some actually, like, academics have, especially PL, like programming languages people, have some, like, very kind of cool ways of doing this. And so, like, at some point maybe I’ll, I’ll, I’ll, like, work on this.

Swyx [01:21:09]: I guess mostly you have to change language, because if you are in JavaScript, Python, you can’t do this.

Alex Zhang [01:21:13]: Yes. Yeah.

Swyx [01:21:14]: So, like, Haskell, yes. what’s, what’s the, what’s the normal one that’s, that’s not Haskell?

Alex Zhang [01:21:21]: Lisp.

Swyx [01:21:22]: Lisp, OCaml.

Alex Zhang [01:21:24]: OCaml, oh, yeah.

Swyx [01:21:25]: Yeah, any functional language

Alex Zhang [01:21:25]: Yeah

Swyx [01:21:25]: You can actually, like, pipeline this.

Alex Zhang [01:21:27]: Yeah.

Swyx [01:21:27]: So Effect-TS if you wanna do TypeScript.

Alex Zhang [01:21:29]: Yeah.

Swyx [01:21:30]: Okay, we can switch over to,

Vibhu [01:21:32]: I like this diagram.

Alex Zhang [01:21:33]: Yeah.

Swyx [01:21:34]: Which is like your, you guys’ whole thesis, right?

Capability Overhang: Reliable Long-Running Work

Swyx [01:21:36]: Like, that, to me this is, like, kind of like a restatement, but maybe I’re missing something of

Alex Zhang [01:21:41]: Yeah

Swyx [01:21:41]: Like, well, work on better harnesses or, like, your models actually are capable a lot more if you try harder, so this is a skill issue.

Alex Zhang [01:21:48]: Yeah. Yeah, basically. I think there’s one thing I want to see. I appreciate that there’s a big focus on, like, jagged intelligence, because it paints a big picture of, like, we can do this if we really set our minds on it. But I kind of wish. And maybe someone in academia should do this. Like, really just sit down and think about, like, if I took Astra, even the current frontier models are not good enough at, like, doing a particular job over, let’s say, the span of a month consistently and well. And I think this is, like, a stupid problem. Like, I genuinely think we can solve this. You don’t need to be a frontier lab and, like, do all this, like, fancy stuff for your IPO. Like, I think, like, these models are so smart that even if it’s, like, a silly way, I think that it genuinely is a skill issue of you can get a model to be as good as, let’s say, like, just some 18-year-old high school kid

Alex Zhang [01:22:50]: Doing some job. I think it’s, like, ridiculous that we can’t do that. And it’s. Part of the reason is, like, the format of a language model is not really amenable to that, but I think you can shape a harness around it and do it. And I think, like, this in itself is, like, ignoring the RLM stuff, ignoring all the, like, what abstractions should we use? Like, I just think someone can design a harness that can do this. Like, I, that’s. Maybe it’s

Swyx [01:23:13]: When you say do this, do what?

Alex Zhang [01:23:15]: Do long-running but simple tasks, and do them reliably.

Swyx [01:23:20]: Okay.

Vibhu [01:23:20]: What’s an example? So is this different than, like, pick your favorite company, Harvey, for example Using LLMs to do legal work, or what’s the.

Alex Zhang [01:23:30]: I guess it’s kind of like that, except if the bottleneck was not, like, certain legal knowledge or something. Like, I don’t know. Let’s say,

Vibhu [01:23:38]: I guess, like, the examples, people can take models and build pipelines or whatever and Have an agent repeatedly do whatever it has they want, right?

Alex Zhang [01:23:47]: Yeah. So for example, like, if I wanted a general system that I could kind of talk. I can talk to it like I would talk to an intern and basically just ask it to do. to explore some small thing. So maybe an example of this is, like, very silly auto research is maybe an example of this, of, like, not necessarily finding super novel solutions, but at least optimizing all of the easy parts of any problem. They often end up being over-indexed for, like, ML training and things like that. But yeah, I don’t know. maybe that’s, that’s, that’s not, like, super clear, but there is a lot of People’s general workflows where you probably could just vibe code up. Some specific harness to help you do, like automate this thing. Some examples are like automating, finding, like research papers and stuff like that. But usually people will design like a specialized agent to help them do this kind of thing. Or like they’ll vibe put a harness, like, and just run it or like their Slack bot or something. But I almost think there’s just like a standard form, like just a harness that you just plug in. Like you don’t-- It doesn’t need to be designed for finding papers or fi-- Like, you just kind of tell it, find this for me, and you like plug it into that setting. What I’m getting at is that I think there’s a lot of easy things that can be automated. And

Vibhu [01:25:13]: Is this like a hypothesis or a point around like capability overhang? Like even if we paused, there’s still a lot of impact to be had with current state of models?

Alex Zhang [01:25:22]: In some sense, yes. Like, I guess what I’m presenting is the easiest form of this. But what this is kind of saying is that, like, we have jagged intelligence on a lot of things. Like, for example, models are like disproportionately good at coding and math. This is saying like we can translate those abilities to many other things. So like for example, if you took-- Usually if you take someone who was an IMO gold or something and you kind of apply them to a lot of different problem-solving domains, they can figure it out. I don’t actually know if this applies to models. For example, like in GPU code optimization, one very interesting question is whe-- if you were to take out all of the GPU programming data, like from a model, but it was re-- it was like as good as Astra is now, just without, like with that taken out, would it be able to still optimize GPU kernels? like would it be able to learn in context roughly what it needs to learn and then like have some pipeline or like come up with some solution to solving like optimization tasks? And I think like there’s like a mismatch between, like if you took a human that was as smart or like knew as much as Astra, there is a mismatch between what that human can do and what Astra can do, maybe around a harness. and I think we can actually approximate the human a lot more.

Continual Learning and General Problem Solving

Swyx [01:26:42]: To me, it sounds very approximate to the continual learning problem. I think you’re

Alex Zhang [01:26:46]: That’s the best example. Yes.

Alex Zhang [01:26:48]: Yes.

Swyx [01:26:48]: Why didn’t you just say that then?

Alex Zhang [01:26:49]: Yeah. I guess I

Vibhu [01:26:50]: I was like, I was like thinking, could I just blurt out some words like learning?

Alex Zhang [01:26:54]: I’m careful with that. But yes, I

Vibhu [01:26:57]: You have a very eye.

Alex Zhang [01:26:58]: Maybe, Maybe, like a bit of a, I don’t know, like a

Swyx [01:27:03]: No. So I think you’re a very, yeah, I don’t know, I don’t know your undergrad actually. Are you like a math person generally, or

Alex Zhang [01:27:09]: A little, yeah. Yeah.

Vibhu [01:27:09]: Did you study math?

Alex Zhang [01:27:11]: Yeah, that’s what I wanted to do at least.

Swyx [01:27:12]: Like a category theory type of, abstraction where you think in categories and then you have to like then translate down to the specific. And, but then you like actually really care more about the category.

Swyx [01:27:24]: And like that’s the communication error because like everyone’s listening for the specific, but actually trying to also, convey the general.

Alex Zhang [01:27:31]: Yeah.

Swyx [01:27:32]: Which is hard. I don’t really know. you can, maybe use like a shorthand of like, “Okay, I’m at level two and then I’m gonna go up to level three Then come back to level two.” That we should have say some like epistemic, like shorthand for like this kind of thing.

Alex Zhang [01:27:45]: Yeah.

Swyx [01:27:45]: Because it’s hard. Like you’re, you’re compressing a lot into word after, like sequential word decoding.

Swyx [01:27:51]: Should we convert to Neuralese? is there like a, a better form of output than English or, Python or JavaScript? I don’t know.

Neuralese, Programming Languages, and Diffusion Thinking

Alex Zhang [01:28:06]: Yeah.

Swyx [01:28:06]: This is very kind of like a s**t post, but like people have speculated about like what is the native language that people want to out-- that models want to output?

Swyx [01:28:14]: Some people say binary. That’s, that’s Marc Andreessen’s thing. I don’t know. Yeah.

Swyx [01:28:19]: PTX?

Alex Zhang [01:28:20]: Let’s say a mix of English and Python. And I only say this because The capability of a model is somewhat a reflection of what we train them on. So we still want like. Yeah, I don’t really buy the binary argument. I guess I, like, I understand, but it’s like

Swyx [01:28:40]: Yeah. You wanna model the world in some way. I think the,

Alex Zhang [01:28:43]: Yeah.

Swyx [01:28:43]: One thing I’ll, I’ll bring up is always, which I always do in this kind of conversation is Sapir–Whorf Which is you, if you choose English, you will have locked into however long English has been around, which is, let’s say five hundred years, which is not that long. Like actually Like what you, the language that you speak constrains how you think.

Swyx [01:28:59]: And if you learn a different language, for example, someone, in Chinese, we don’t have tenses. I don’t know if you. I actually didn’t know that. And I speak Chinese.

Alex Zhang [01:29:09]: Oh. I did know that, but my Chinese is not great.

Swyx [01:29:12]: Okay.

Alex Zhang [01:29:13]: Yeah.

Swyx [01:29:13]: Yeah. Or like, in, let’s say in Japanese or, I know, I forget what language it is. Like in Korean, everyone you speak to, yeah, you have to like acknowledge social status.

Alex Zhang [01:29:23]: Yeah.

Swyx [01:29:24]: But it’s a different dimension than gender, right? Like, it just like, it just influences everything you do. when I take Ling 101, apparently there’s a, there’s a, there’s a language in Africa where like there’s a vegetable gender. yeah, right? Like just like you have. Or like Eskimos know no word for snow or whatever. Like, anyway, so like the language that you adopt affects your thinking. And if you Choose to output your chain of thought in English, you are biasing towards whatever English solves. I don’t know what the sort of prior of English is.

Alex Zhang [01:29:51]: That’s interesting. I did not think of it that way.

Vibhu [01:29:55]: At some level it’s interesting, right? So you’re right on language. a lot of model chain of thought also fluctuates language.

Vibhu [01:30:03]: The obvious example is Chinese models Speaking in English might still reason in Chinese. but at the same level, most models are very capable multilingually.

Vibhu [01:30:13]: And that adaptation we can see, you can add in languages. You don’t get that much from adding a whole language, but

Swyx [01:30:18]: Yeah.

Vibhu [01:30:18]: They will reason interchanged

Swyx [01:30:21]: Yeah. So we’re all autoregressive. but also like

Vibhu [01:30:23]: Right.

Swyx [01:30:23]: Let’s say German, like, subject-object, agreement, you have to put the verb at the end, which is very super annoying, like very famously. yeah, right. You don’t know what you’re doing until the end where you’re like, “Oh, that Mess of nouns and then the verb.” Well, the most classic one that most people be-- have heard of is Arrival, where, they have the heptapods where they think, that time is like flat to them. So they think in, they output entire sentences at one shot. so it’s, this is closest to, like, the difference between autoregression and diffusion.

Swyx [01:30:55]: We talk in autoregression. What if you could talk in diffusion Where things just resolve over time?

Alex Zhang [01:31:01]: I see. I see.

Swyx [01:31:02]: But like the whole idea shows up at once.

Alex Zhang [01:31:05]: I see.

Swyx [01:31:07]: So that is a drastically differenting, language, but it is a language.

Alex Zhang [01:31:10]: I see. Oh, that’s really interesting. That’s really interesting.

Swyx [01:31:12]: Which, like, machines could speak, that we, probably will never speak, but like, yeah, machines don’t care.

Alex Zhang [01:31:19]: Maybe this is a huge tangent

Swyx [01:31:20]: Yeah

Alex Zhang [01:31:20]: But are there not, like, things inherently that are reasoning chains that are inherently autoregressive?

Swyx [01:31:30]: Yeah, time.

Alex Zhang [01:31:30]: Sometimes, like. Yeah. Or like, yeah.

Swyx [01:31:32]: Yeah. Something happens first, then something else happens.

Alex Zhang [01:31:34]: Even like, yeah, anything in code, for example, like has to be causal in some-- usually at least has to be causal.

Swyx [01:31:40]: Well, no. so it’s a. Then you have to. Then you’re not exploring enough,

Alex Zhang [01:31:44]: That’s true. Yeah

Swyx [01:31:45]: Programming language theory, where, everything is like pure functional and like completely relational And, you sort of abstract away the solver that translates the relationships that is, are always true into code. So I, yeah, I feel like this is maybe a little bit too out of my depth.

What Comes After RLMs?

Alex Zhang [01:32:01]: No, it’s interesting though. Yeah.

Swyx [01:32:01]: But I love languages Whether it’s coding or human, and I do think a lot about how that affects reasoning and the boundaries of what we can do. I don’t need to go too much beyond that. I don’t know if you have any other thoughts. my closing question was gonna be, you have all these research, directions that you wanna do. You’re, you had a GPU mode phase. You had a, RLMs phase.

Swyx [01:32:25]: Presumably you have other stuff planned, which is why you’re not, doubling down on that. By the way, I notice that it is interesting how you guys do start with the GPU side, and then you migrate towards the zero gradient side, it’s, which is what Shenyou called it. Doesn’t that feel less legit than messing with GPUs?

Alex Zhang [01:32:44]: Yeah, I guess in the sense that, like. So you did bring up that, like, I like to think about things in, like, a math-oriented way.

Swyx [01:32:51]: Category, yeah.

Alex Zhang [01:32:52]: And it’s, like, very uncomfortable sometimes to be working on, like, harnesses and agents because it’s so.

Swyx [01:32:57]: Because you think all harnesses are the same.

Alex Zhang [01:32:58]: Yeah. So it’s like super fuzzy.

Swyx [01:33:00]: So like, you just, like, two new ideas in harnesses. Got it.

Alex Zhang [01:33:01]: Yeah. It’s, it’s It’s, it’s also just, like, empirically it’s hard to, like, verify a lot of findings, at least with the compute that we have available to us. But the reason why I think I’ve moved on to a lot of these problems is I think actually this is where most of, like, the innovation is yet to happen. To me, like, the GPU level is a means to exploring other ideas. Like, you want to, for example, like, get good at writing kernels or, like, even automate writing kernels for the sake of a broader goal of, like, I want to explore ideas where I’m not bottlenecked by systems challenges. In that sense, like, I guess a lot of what’s written there is all harness stuff, but I am also interested in things at the model level as well. but I’ll just leave it at that.

Swyx [01:33:48]: Okay.

Alex Zhang [01:33:48]: Yeah.

Swyx [01:33:48]: That’s a good hint. anything, any. If people wanna reach out to you, what are you looking for help on? What do you want collaborators on? any sort of calls to action?

Collaborating on Research and Choosing Big Bets

Alex Zhang [01:33:58]: Yeah. So, I guess there’s nothing I have in particular where I feel like I need to work with someone on, unless it’s, like. unless it’s with a company before, like, compute or, like, with. to talk with other people about it. But I will say I’m not. I’m never opposed to working on ideas with other people. I get reached out to a lot by often undergrads or even, like, other students.

Swyx [01:34:22]: Podcasters.

Alex Zhang [01:34:23]: Podcasters. and usually I feel like I get an email that’s something along the lines of, like, “I really like RLMs.” Like, “I wanna work together.” and I feel like I

Swyx [01:34:33]: Yeah, that’s a bad reach out, right?

Alex Zhang [01:34:35]: Yes, yeah.

Swyx [01:34:35]: The worst is like, “Can I pick your brain?”

Alex Zhang [01:34:36]: Yeah.

Swyx [01:34:37]: And like, “On what?” Like, “Read my paper, dude.” Like.

Alex Zhang [01:34:39]: They’ll like, they’ll be like, “I read your paper,” in quotes, like, “Recursive language models,” or like, “Prime Agent,” like a self-improving RLM harness or something. and it’s kinda like I like. I really like people that are opinionated, even if we disagree. I think if you have strong opinions and are able to, like, think through why you think those opinions are right or wrong, ‘cause usually it’s, it’s hard to actually tell. But, like, you have strong convictions about certain problems. Like, I’m, I’m always happy to, like, chat and even, like, potentially work on something together. I have, like, no limit to who or, like, what I would like to work on. So, yeah.

Swyx [01:35:12]: No limit?

Alex Zhang [01:35:13]: Yeah. I. in the era of agents, I think there’s a lot more work you can do, like, bandwidth-wise. So I. Yeah. I think in general, like, I am not hard to impress, but I think it just takes a little bit of effort to

Swyx [01:35:30]: Yeah

Alex Zhang [01:35:30]: Kind of. Yeah, know

Swyx [01:35:32]: Yeah

Alex Zhang [01:35:32]: Know what you want.

Swyx [01:35:32]: It’s very clear. and like, when you see a new thing come out, well executed, good, simple idea, then, like, get that

Alex Zhang [01:35:40]: Yeah

Swyx [01:35:40]: Immediately gets your attention, right?

Alex Zhang [01:35:41]: I get excited. Yeah.

Swyx [01:35:42]: It’s actually, like, not that hard to get the same attention that all the Frontier Lab guys

Alex Zhang [01:35:45]: Yeah

Swyx [01:35:45]: Because they are looking for you. you just have to put yourself out there, right?

Alex Zhang [01:35:48]: Exactly, yeah.

Swyx [01:35:49]: But yeah, it’s true. I will say, I think human attention very scarce right now, and I do struggle with, like, the number of projects I have going on.

Swyx [01:35:57]: And, I don’t know how to manage it. I don’t think agents are helping at all.

Swyx [01:36:00]: Like, I will just prompt it and. I’ll prompt a thing and then never look at it.

Swyx [01:36:03]: Right? Like, which is very common.

Alex Zhang [01:36:05]: Yeah.

Swyx [01:36:05]: Yeah, and that sucks.

Alex Zhang [01:36:06]: I guess, maybe the. one of the smaller differences in, like. actually, maybe you were doing research. I’m not sure. But for me at least, like, I’ll have maybe, like, 10 or 15 different ideas that I wanna do, but the thing is, like, most of them are bad.

Swyx [01:36:20]: Yeah.

Alex Zhang [01:36:20]: And this also maybe is true even for someone that reaches out to me. Like, maybe the idea is actually bad, but it looks interesting to me. And so, like, we can spend, like, some time looking into it, and if, like, we feel like there’s actually something there, like, then we’ll. we should take the next few weeks and just really pursue it. And, like, this is my style with. This is why I love the PhD, by the way, because there are times when I’m just thinking about problems, like, maybe on a run or just, like, playing tennis or something. Like, I’m not, I’m not working, I guess. But it’s like those are the most fun times, and then when I, like, really am convicted about something, I’ll just, like, drop everything and just do it.

Alex Zhang [01:36:52]: Like, just spend, like, all my time thinking and working on that problem. And then, once you get to the point where, like, you can just run experiments, then it’s, it’s, it’s kind of easy coasting again, so.

Science as the Next Frontier

Swyx [01:37:03]: Yeah, sorry. This is

Alex Zhang [01:37:04]: Yeah. No problem

Swyx [01:37:04]: Like the, for the fourth last question, which is like, I think a lot of people are also thinking about science as the next frontier, like physical sciences Bio, math even. how do you separate, like, I guess, let’s say your choice of projects that is applicable for industry And then maybe it’s part-- choice project is just, like, science?

Alex Zhang [01:37:27]: I actually worked on, like, AI for bio stuff before I started my PhD. The field has changed a lot since then

Swyx [01:37:33]: Yeah

Alex Zhang [01:37:34]: I should say.

Swyx [01:37:34]: ‘Cause it’s, it’s like, it used to be a theoretical, like, of course, what do you mean? Like, I have one path and then

Alex Zhang [01:37:39]: Yeah

Swyx [01:37:39]: I chose that. But now a lot of people are crossing over.

Alex Zhang [01:37:41]: Yeah.

Swyx [01:37:41]: And like, so we have started a science pod to just Cover those things

Alex Zhang [01:37:44]: Oh, wow

Swyx [01:37:45]: Because a lot of engineers are like, “Well, actually there’s, Tractable problems there.”

Alex Zhang [01:37:50]: Yeah. I will preface by saying my understanding of a lot of these topics is probably pretty limited. But I think, like, if I find out either because someone reaches out or, like, I look at a problem and I’m like, “Hey, like, some design principles that we use or that we’re thinking about right now actually make a lot of sense for this problem,” I get excited about those as well. But I think it’s harder. I don’t know. I think with. I think science, especially like empirical or, like, applied science has very long, like, what is it called? Like, feedback loops or whatever.

Swyx [01:38:23]: Yeah, it converges to a robotics question.

Alex Zhang [01:38:25]: Yeah. to me, like, also this aspect of, like, what is worth spending and betting my time on now? Because, like, maybe I spend a lot of time on this problem and then, like, in six months, like, a different solution, kind of like maybe a new model comes out and it’s like, “Oh, it’s way better for this.” And so I do have to be careful. I, like, you have to be conscious about, like, where you think things might be going.

Swyx [01:38:47]: Yeah, exactly.

Alex Zhang [01:38:47]: So, yeah.

Swyx [01:38:48]: Publish cycle.

Alex Zhang [01:38:49]: Yeah. So

Vibhu [01:38:51]: Which then I can just tell ARC-AGI-3, it’s saturated. We did it.

Alex Zhang [01:38:54]: Yeah, ARC-AGI-3 got saturated in less than a year, so it’s kind of ? Like, it’s. I don’t know. Like, if you were a lab picking that problem, like, you’re probably kind of sad now ‘cause actually

Swyx [01:39:04]: Yeah. Well, so exactly. That’s why knowledge work, gaming, all these things are saturated.

Alex Zhang [01:39:08]: Yeah.

Swyx [01:39:08]: Now they’re actually the frontier is science. So

Alex Zhang [01:39:09]: Yeah.

Vibhu [01:39:10]: Knowledge work is saturated.

Swyx [01:39:12]: Yeah. GDP val is like 80 something, 90 something.

Swyx [01:39:16]: Like, there’s, there’s 90 to 100% that was obviously gonna get

Alex Zhang [01:39:20]: It’s gonna be really hard to tell

Swyx [01:39:21]: 10 years. But, like, well, the next low-hanging fruit is gonna be

Alex Zhang [01:39:25]: It makes sense

Swyx [01:39:25]: The other stuff.

Alex Zhang [01:39:26]: Yeah. Maybe I’ll think about that more. I actually, I haven’t given too much thought to

Swyx [01:39:30]: I’m just trying to guess your next direction, actually.

Alex Zhang [01:39:32]: No. I will say because I think especially at MIT, like, it’s, it’s, it-- Yeah, there’s a lot of really talented scientists there, like, in the natural sciences. And I think it’s, it’s a little bit, like, sacrilegious almost to be like, “I’m gonna figure out, like, your problem.” Like

Swyx [01:39:45]: No, that’s how, that’s how it’s done.

Vibhu [01:39:46]: It’s great.

Alex Zhang [01:39:46]: Oh, no, I know. Yeah.

Swyx [01:39:47]: So I interviewed Yitai who did the IMO thing. He’s just like

Alex Zhang [01:39:50]: Oh, yes. Oh, yeah.

Swyx [01:39:50]: “I’ve never, I’ve never been to IMO. I don’t even know what it is.” It’s a skill model, dude.

Vibhu [01:39:55]: Yeah.

Swyx [01:39:56]: Which is, like, very disrespectful, but like, whatever. But that’s why, yeah.

Vibhu [01:40:00]: Yeah. At some point, like, you have to respect, like, okay, the progress is being made.

Vibhu [01:40:05]: Like, number is getting output, right?

Alex Zhang [01:40:07]: True. Yeah, true.

Swyx [01:40:08]: Yeah. that is a very big lesson. It’s very interesting, like, ‘cause the mathematicians are responding this way to NARI systems right now.

Alex Zhang [01:40:13]: Right. Yeah.

Swyx [01:40:14]: Like, Terry Turnstow is like

Vibhu [01:40:15]: Turnstow

Swyx [01:40:16]: “No, like, let’s not, let’s not use AI.”

Alex Zhang [01:40:17]: Yeah.

Swyx [01:40:17]: I’m like, “Mm, I don’t know.”

Alex Zhang [01:40:19]: Well, yeah. I think that whole thing is kind of weird ‘cause I feel like, I feel like they would have had a stronger case if a lot of them didn’t work with OpenAI before, like, all this happened.

Swyx [01:40:30]: No, that’s ad hominem, and they’re really trying to stay away from that. Then so what, right?

Alex Zhang [01:40:34]: So what? Yeah.

Swyx [01:40:35]: Like, I don’t know. Like, so what? They got. they’ve, they’ve collaborated. I collaborate with people that I don’t

Alex Zhang [01:40:39]: I guess it’s true

Swyx [01:40:39]: I don’t agree with or

Closing: Research, Academia, and What Comes Next

Alex Zhang [01:40:40]: That’s true

Swyx [01:40:40]: Whatever.

Alex Zhang [01:40:41]: Yeah. That’s true.

Swyx [01:40:41]: Or, like, I did a thing and then now I regret that. I changed my mind. whatever.

Alex Zhang [01:40:45]: Yeah.

Swyx [01:40:45]: So I’ll defend their right to say that. But like, yeah, a lot of people are reasonably disagreeing with them.

Alex Zhang [01:40:50]: Yeah.

Swyx [01:40:51]: Okay, cool. thanks for your joining us. Congrats on, your success so far. I can’t believe you’re still not done with your PhD.

Alex Zhang [01:40:58]: Well, it’s year two?

Vibhu [01:41:00]: Yeah. Can’t believe we did this podcast without going through the RL paper.

Swyx [01:41:05]: He had a, he had a definition.

Alex Zhang [01:41:07]: I think the paper is more about, like, empirical results. Like, the actual idea is quite simple.

Vibhu [01:41:11]: Yeah. And you’ve talked about it many times.

Alex Zhang [01:41:13]: Yeah, at this point. I think there’s, there’s more interesting things to look over now, so.

Vibhu [01:41:17]: Cool.

Alex Zhang [01:41:17]: Yeah.

Vibhu [01:41:18]: Well, we’re excited to see what you do next.

Alex Zhang [01:41:19]: Thank you so much.

Three months ago Dwarkesh, who has been posting incredible blogs and episodes about RL, posted a framing question for his video essay on RLVR which upset a lot of Computer Use folks:

We are no strangers to learning in public and are no strangers to the stress of getting things wrong when you have a big platform. However, we were at Anthropic for the Computer Use launch, there for Claude Cowork with the first big podcast on it, organized the first Computer Use track at AIE presenting the state of the art, and were close to the OpenAI-Sky Software acquisition that now powers the complete domination of computer use that Codex enjoys today. This is why we’re excited to bring you today’s first guest, Ari Weinstein, cofounder of Sky and now leading all the amazing CUA progress that casuals might miss:

Ari explains why Computer Use is now “180 degrees different” from where it was months ago, how agents are learning to debug and recover from failures, why combining screenshots with accessibility data, the DOM, Playwright, and generated code changes the speed equation, and why the next frontier is making agents literally superhuman at using software.

OpenAI clones Jev

In the second half, Nikunj Handa from OpenAI’s API team breaks down the new developer stack: async tool calling, mid-turn steering, WebSockets, UltraFast inference, the Decisions API, prompt caching, pre-warming, compaction, and the Agents API. Given that we were the first Jev podcast, we particularly focus on the unusually fast sprint on the Decisions API:

And why it is just a Luna wrapper for now but the team is motivated and egoless enough to clone what they consider to be good patterns.

We discuss:

* Why OpenAI thinks Computer Use has changed dramatically in just the last few months

* Dots and what changes when every agent gets its own Linux computer

* Why Computer Use can now complete some tasks faster than the average human

* The path from human-level to “literally superhuman” computer use

* Why modern agents are much better at debugging and recovering from failure

* How screenshots, accessibility trees, the DOM, Playwright, and generated JavaScript work together

* App Shots and why they give models much richer context than ordinary screenshots

* Why Computer Use can close the loop between writing software and testing it

* Trust, permissions, and safety when agents can make payments and operate websites

* Async function calling and why models no longer need to stop reasoning while tools run

* Mid-turn steering, WebSockets, and the architecture behind more responsive agents

* UltraFast inference and how OpenAI is pushing frontier models toward much lower latency

* The rapid internal story behind the Decisions API

* Why Decisions API is more than structured outputs at low latency

* GPT Live, fast tool calling, and real-time computer control

* How OpenAI is already using Decisions API for support classification and internal workflows

* Longer prompt caching, cache pre-warming, and cache-aware applications

* Server-side compaction vs manual compaction for long-running agent threads

* What should live inside an Agents API versus a developer’s own harness

* OpenAI as an “AI cloud” and the search for higher-level primitives beyond raw model APIs

Ari Weinstein

* Product & Engineering, Computer Use at OpenAI

* X: https://x.com/AriX

* LinkedIn: https://www.linkedin.com/in/weinsteinari/

Nikunj Handa

* Product, API at OpenAI

* X: https://x.com/nikunjhanda

* LinkedIn: https://www.linkedin.com/in/nikunjhanda/

Timestamps

00:00:00 OpenAI DevDay: Dots, GPT-6.1, Agents API, and Decisions API

00:02:52 Dots and Personal Cloud Computers

00:04:59 Why Computer Use Is “180 Degrees Different”

00:06:04 From Sky to Self-Debugging Computer Use Agents

00:09:24 How Computer Use Sees and Operates Software

00:12:09 From Faster Than Humans to Superhuman Computer Use

00:16:03 Agents API: Trust, Permissions, and Safety

00:17:31 Computer Use for Coding, Testing, and QA

00:19:14 GPT-6 APIs, Async Tool Calling, and UltraFast Inference

00:23:21 The Rapid Story Behind Decisions API

00:25:32 What Decisions API Is and How It Works

00:30:24 What OpenAI Is Building With the New APIs

00:32:23 Prompt Caching, Pre-Warming, and API Performance

00:35:20 Context Compaction for Long-Running Agents

00:37:13 Memory, Higher-Level APIs, and the AI Cloud

Transcript

Introduction: OpenAI DevDay and the New Agent Stack

Vibhu [00:00:00]: Okay. We’re very excited to be here. Today is OpenAI DevDay. Special podcast

Swyx [00:00:08]: We’re the first podcast after your livestream.

Vibhu [00:00:10]: First podcast. We have Ari here, who leads the product and engineering team for Computer Use agents. Before we kick in and dive deep on Computer Use, you wanna give a quick recap? What was announced? What’s the quick slew of announcements you guys had today?

Ari Weinstein [00:00:24]: Yeah. yeah, it was a super exciting day. we just got out of the keynote. It was really sick. there were a bunch of Computer Use announcements that I think are worth thinking about. We have, Dots, which is the new, sort of personal assistant product, and, that has some really exciting Computer Use features. There’s GPT-6.1 Sol, which is this amazing new model, that I think is particularly great for Computer Use ‘cause of, sort of the cost and speed, advantages. I think, I think we shared that it’s, a fifth of the cost of Astra and a seventh of the cost if you’re looking at Computer Use specifically, which is really amazing. sorry, there were so many things. I’m trying to sort through it.

Swyx [00:01:02]: And the API.

Ari Weinstein [00:01:03]: Agents API, which now has Computer Use in it, which is really cool, ‘cause now developers can build on the same Computer Use, that is part of Codex, and ChatGPT. and then there were some demos of our existing Computer Use features, like app shots, where you can take the context of something you’re doing on your computer and bring it into Codex and ChatGPT really fast. And then, like, native Computer Use on your Mac, where Roman had it taking screenshots of his app, automatically, and he could do other things on his computer while Computer Use was using his applications. so yeah, really exciting keynote.

Swyx [00:01:35]: And not to mention the Decisions API.

Ari Weinstein [00:01:37]: Decisions API.

Swyx [00:01:38]: Off the bat, are they all the same model? Like, this is. Or the same dataset distilled to different models?

Swyx [00:01:44]: Like, basically, like, is Computer Use using Decisions API, or are they, like, kinda separate?

Ari Weinstein [00:01:49]: So what’s really cool about the Decisions API is it, you know, it has all these new capabilities. It does inference in parallel. it doesn’t have reasoning. It’s a smaller model, than the ones we use for Computer Use. and so those capabilities make it really fast.

Dots and Delegating Work to a Cloud Computer

Swyx [00:02:07]: Yeah.

Ari Weinstein [00:02:07]: They also make it a little bit less good at doing, like, long horizon, sort of sophisticated tasks. And so I think I would say it’s still an open area of research for how we, like, bring those approaches together. But, yeah, I’m really excited to see what people build with the Decisions API.

Vibhu [00:02:24]: One of the interesting things is Dots now have attached personal computers.

Ari Weinstein [00:02:28]: Yeah.

Vibhu [00:02:28]: So it seems like they’re very much more persistent. You’ve been using them for a while. How should people push the bounds? Like, what should people aim for? What should they try? Personally, right now I use it for a lot of customer service. Like

Ari Weinstein [00:02:41]: Cool

Vibhu [00:02:41]: “Oh, this was wrong. I don’t wanna sign in. I don’t wanna authenticate.” Find whatever and just get it fixed.

Ari Weinstein [00:02:45]: Yeah.

Vibhu [00:02:46]: How should we push further? What should people try?

Ari Weinstein [00:02:50]: Dots Are a really cool product because each Dot has access to its own Linux virtual computer in the cloud, which is different from our other products. you know, traditionally, we’ve have access to a browser in the cloud, or it has access to your own computer, but now you get your own entire Linux computer in the cloud. And so it can run full desktop applications, and it can also use a web browser. And so, yeah, you know, I think the powerful thing about Computer Use and the reason why I think it’s so, exciting is because it makes it so that the agent can do anything you as a, as a person can do, because all the software in the world was designed for humans, and now agents can use that same software, and you can delegate to the agent. So, yeah, like, anything that you would do on a computer, you can ask a Dot to do. Yeah, I think what particularly is useful is gonna really depend on who the end user is and what- what’s valuable in their life. but yeah, I would just start by thinking about, like, one of the things that you spend time on and how could you delegate those to an agent.

Swyx [00:03:47]: Yeah, a lot of flight booking and shopping and honestly even, like, playing a game or whatever, right?

Ari Weinstein [00:03:52]: Totally.

Swyx [00:03:52]: Yeah.

Ari Weinstein [00:03:53]: Yeah, I don’t know. For me, something I did recently, I’ve been working on. I’ve, subscribed to a meal prep service ‘cause I was trying to, like, eat healthy, you know? And I really like this meal prep service I found because it lets me customize the meals I order to, like, a high degree of granularity. So I can say, like, “I want this many grams of chicken and this many grams of rice.” but it was so complicated. It took me two hours to do an order, and I found that I could ask Computer Use to do it for me, and it did it in 15 minutes. so I actually saved two hours. it both did it eight times faster than I could, and it saved me two hours on GPT-6.1 Sol.

Swyx [00:04:32]: Yeah.

Ari Weinstein [00:04:32]: So those are the kinds of tasks that I feel like, are really powerful.

Swyx [00:04:36]: As a creator, I can tell you automatically, immediately, my number one use case is automating YouTube.

Ari Weinstein [00:04:40]: Nice.

Swyx [00:04:40]: Because, YouTube doesn’t expose a lot of things via API.

Ari Weinstein [00:04:43]: Yeah.

Swyx [00:04:43]: And you have to just put it in a VM and just, like, run it, for, like, let’s say, let’s say their AB testing feature or making community posts. None of this is available by API ‘cause they hate developers.

Swyx [00:04:53]: Anyway, so,

Ari Weinstein [00:04:55]: I’ve heard that from our developer experience team too. They use it with YouTube a lot. Yeah. It’s really awesome.

Swyx [00:04:59]: So I wanna draw for, you know. let’s say, I wanna get a little bit spicy. One of our, the leading AI podcasts, our friends, is famous for saying that Computer Use hasn’t advanced in the last two years.

How Computer Use Has Changed in the Last Year

Ari Weinstein [00:05:12]: Yeah.

Swyx [00:05:13]: Which is a very interesting statement, and I think you’re one of the best people in the world to talk about this, like, how have things have progressed, right?

Ari Weinstein [00:05:20]: Yeah. You know, they said that a few months ago, I think, and I hope they have a different perspective now because Computer Use is, like, 180 degrees different than it was.

Swyx [00:05:26]: He’s a, he’s a tough guy to impress.

Ari Weinstein [00:05:27]: Yeah, okay. well, we’re working on it.

Swyx [00:05:30]: But, you know, you worked on. You’ve, like, basically spent your whole career working on, like, some kind of computer automation, right?

Ari Weinstein [00:05:34]: Yeah.

Swyx [00:05:34]: Like shortcuts

Ari Weinstein [00:05:35]: Yeah

Swyx [00:05:35]: At Apple, and then Sky, and then, and then joining OpenAI. Can you draw, like, what your through line is for, like, what is driving you and what- you, what wasn’t possible back then maybe

Ari Weinstein [00:05:47]: Yeah.

Swyx [00:05:48]: And, like, what your sort of milestones were.

Vibhu [00:05:49]: I guess to add on to that as a follow-up question, what’s the major change from using Codex Computer Use from, like, last week

Ari Weinstein [00:05:57]: Yeah

Vibhu [00:05:57]: Through to today? Is it model? Is it dots? Is it harness? So all the history plus what really just changed in today’s announcements?

Ari Weinstein [00:06:04]: Yeah. On the through line, I guess I’ve always been excited about automation and helping people automate tasks because then you can, like, save time in your life and focus on things that are more important to you than, like, operating a computer very intricately. And so, yeah, that was why we worked on some of those products. I was at Apple before. we made a company called Sky. we ended up joining OpenAI, which is really exciting. and I think something that was

Swyx [00:06:27]: And almost like you have to hack around Apple until Apple was like, “Fine, like, we’ll just hire you and you can just work on the inside,” right? Like.

Ari Weinstein [00:06:35]: It was, it was a cool place to get to work. what was really interesting looking back at Sky is we were, we were working on Computer Use there as well, and the models were so much less capable. And now the models, just in the last one year, have become extraordinarily capable at Computer Use. I think the biggest delta that I see is before they could, like, reliably start tasks, but then they would run into problems, and now they’re really good at debugging. They’re really good at trying again, introspecting what is and isn’t working. and I think we’ve also brought the Computer Use the Computer Use field itself has moved forward. I think we’re using more techniques. now Computer Use, often writes code. So if you actually look at it in Codex and you expand the tool calls manually, you can see that it’s not just doing one action at a time. It’s actually writing JavaScript code that it executes, that the computer executes to perform sometimes many actions at once, which is a great, you know, speed up and great capability. We use more accessibility, sort of multimodal interfaces. So, the model may use screenshots, it may use accessibility, it may use Playwright. it can use a lot of different mechanisms, based on the task at hand. and then, yeah, the model acceleration has been, has been just amazing. So, yeah, what’s different today? I think we’re making computers better all the time, so I think just, like, one day’s difference, is probably a little bit less consequential than, like, even the past month or the past two months. but, yeah, I think the Computer Use in Dot is really exciting as well as, the new model that we came out with.

Measuring Computer Use and Improving the Harness

Vibhu [00:08:03]: On the keynote, Tejal was mentioning 7x improvements in Computer Use speed, a lot better on a few benchmarks. How do you guys think about measuring it? Computer Use is one of those things where, as you say, you know, it’s improvements over time.

Ari Weinstein [00:08:20]: Yeah.

Vibhu [00:08:20]: Is it harness? Is it model? Is it post-training?

Ari Weinstein [00:08:22]: Right.

Vibhu [00:08:22]: How do you guys look at it internally about measuring how good it is, and what were the changes with the new model?

Ari Weinstein [00:08:29]: We actually have a bunch of different ways of measuring it, some of which are on different permutations and configurations of the harness. It’s a bit of a complicated story because, you know, our production products have, you know, some more safety checks, and, you know, those are configured differently based on the needs of the, of the task at hand. So there’s a lot of ways to measure it, but I think regardless of how we measure it, we find pretty consistent gains. and those gains are, sometimes in the harness and sometimes in the model. and yeah, I was really excited by this result that GPT-6.1 is even more cost-effective for Computer Use than its baseline cost improvement as compared to Astra. It’s, like, really cool to see.

Swyx [00:09:10]: Yeah. I mean, one of the visuals I really liked from the livestream was that, you’re sort of improving the Pareto frontier of, your, curve, and there was a lot of talking about how you’re improving it together with the harness.

Ari Weinstein [00:09:24]: Yeah.

Swyx [00:09:24]: Can you give some examples of aha moments that you had, whether it’s on, like, model driving the harness driving the model, whatever?

Ari Weinstein [00:09:32]: I don’t mean to repeat myself, but I think, like, introducing more modalities has been really powerful.

Swyx [00:09:36]: Okay.

Ari Weinstein [00:09:36]: One more specific example of that is, in the past, I think we saw a lot of Computer Use, products had to spend a lot of time, like, scrolling, you know? So it would, like, take a screenshot. It would try to do something. It would be like, “Oh, I gotta, like, scroll down to the next page of results,” and then it would take a screenshot, and then it would try to do something. It would scroll down again. And so I think, with accessibility and other. and, direct access to the DOM and other things like that, now the language model can actually see, like, an entire page or an entire application. It can write code that can do multiple steps at once. And so I think those have been probably the biggest single aha moments. There’s, like, a lot of tiny ones that are less exciting in comparison, but actually we do find also that a lot of speed improvements are driven by, like, a lot of little paper cuts that we gotta go in and introspect.

App Shots, Accessibility, and Better Computer Context

Swyx [00:10:21]: Yeah. A lot of really hard engineering.

Ari Weinstein [00:10:23]: Yeah.

Swyx [00:10:23]: I mean, app shots in general, right? Like, I think people don’t quite get the difference if. because there’s, like, a nice visual in Codex when it

Ari Weinstein [00:10:30]: Yeah

Swyx [00:10:30]: When you take an app shot, but they don’t maybe they get the difference that, you are able to actually drive each button and you have the, you have each text, in a very optimal representation.

Ari Weinstein [00:10:40]: Yeah. Exactly. Yeah. It’s kind of fun actually. If you wanna be, like, really nerdy about it, you can go into Codex, take an app shot by hitting the two command keys. So you grab the content from whatever app you’re working with, bring it into the, Codex or ChatGPT chat. And then the. if you click on the attachment and you click on this, like, little tiny button in the top right, you can see the raw text and you see the raw accessibility representation. And yeah, we’ve put a lot of work into, putting

Swyx [00:11:04]: Just dumping everything out. Yeah.

Ari Weinstein [00:11:05]: Dumping it out, but also making it token-efficient, doing it efficiently. There’s, like, a bit of an art to it. And, you know, it turns out that the same technology that was invented for humans, you know, who maybe have accessibility needs, who wanna use a screen reader technology, that technology is really helpful for them to be able to use computers. It’s also really helpful for LLMs to be able to use computers. So that’s been, like, really fun to get to work on.

Vibhu [00:11:27]: For context, I feel like a lot of people don’t understand app shots. They don’t even know it’s a feature.

Ari Weinstein [00:11:30]: Yeah.

Vibhu [00:11:31]: It’s when you double hit command, it pulls in what looks like a screenshot

Ari Weinstein [00:11:34]: Right

Vibhu [00:11:34]: And you’re like, “Oh, why have I opened up just a screenshot and thrown it in?” No, it’s actually pulling all the metadata, all the code, everything.

Ari Weinstein [00:11:40]: Yeah, exactly. Yeah. So it’s like, you know, if you take a screenshot of a webpage that has a link- The screenshot doesn’t include where the link goes. It doesn’t include, you know, maybe you take a screenshot of your calendar, the ca- event ti- titles are truncated, you know? But when you take an app shot, it gives, like, the language model, like, full context about everything and, that lets it, just sort of, like, do much more.

Swyx [00:12:02]: Yeah. For those who wanna see more, Jason Liu, I invited him to do a full workshop on this, at AI Engineer.

Ari Weinstein [00:12:07]: Amazing.

Swyx [00:12:08]: Did a great job.

Vibhu [00:12:09]: I have a broader vision question

Toward Superhuman Computer Use

Ari Weinstein [00:12:11]: Yeah

Vibhu [00:12:11]: On Computer Use agents. So your example of take a screenshot, scroll page, take a screenshot is where we were.

Ari Weinstein [00:12:17]: Right.

Vibhu [00:12:17]: Today, they can automate a lot. what are the bottlenecks? Is it models? Is it harnesses? What. Where do you see it going in, like, two years? Do you see it just running for hours? How do we get there? Any predictions on where Computer Use goes?

Ari Weinstein [00:12:32]: Yeah. I mean, I think what’s really crazy that I think, You know, the team’s accomplished over the past couple of months is that now Computer Use is, like, faster at accomplishing tasks than, like, the average human probably in most cases. and I think that the next frontier is to have Computer Use be, like, literally superhuman in its performance where it actually is as fast or faster at using software than, like, expert Computer Users like us. and I think that’ll be really consequential and exciting when that happens because I think we’ll be able to all of a sudden build products, that, provide just much more real-time experiences. And I think it’ll also. lowering the barrier to entry of, or the activation energy, I suppose, of using Computer Use I think will make us start to default to doing certain things in agents that we’ve become accustomed to doing manually. And I think that’s exciting also ‘cause it’ll save us a ton of time. and I think there’s a, you know, there are a lot of different little paper cuts and bottlenecks that are sort of standing in the way of that. I think that there’s, yeah, there’s things on the model side, there’s things on the inference side, there’s things on the harness side, there’s things in the, in the representation. You know, we find that as Computer Use gets faster, we’re increasingly bottlenecked by just, like, the speed of doing an operation. Like, for example, you know, a non-trivial amount of time in our benchmarks of Computer Use tasks is actually, like, let’s say you’re automating a task on doordash.com. Like, a lot of the time is actually waiting for doordash.com itself to load, you know?

Swyx [00:14:04]: Yeah, then you just write a wait and then you execute the wait.

Ari Weinstein [00:14:07]: Yeah, totally. And you wanna get. Yeah, actually, it’s actually really important that you get that de- like, you want as little delay as possible between when it finally finishes loading and when you go and

Swyx [00:14:16]: Yeah

Ari Weinstein [00:14:16]: Trigger the LLM to do the next action, which is actually- itself a statistical science.

Swyx [00:14:20]: Like an event-driven way maybe to do that.

Ari Weinstein [00:14:22]: When possible, you want it to be event-driven.

Swyx [00:14:24]: JavaScript has some load events.

Ari Weinstein [00:14:25]: And JavaScript has load events for. or the web browser has load events for web navigation, but there’s other types of events that actually really can’t be event-driven. So there’s a lot of complexity

Vibhu [00:14:34]: The one that comes to mind is, like, chatting with customer service.

Ari Weinstein [00:14:37]: Yeah.

Vibhu [00:14:37]: Replies could take 30 seconds, could take three minutes.

Ari Weinstein [00:14:39]: Oh, right.

Swyx [00:14:41]: I have dealt with so many bots with Codex. it’s great, but I also wonder if the other side knows that they’re talking to a bot ‘cause I’m, like, answering in complete sentences. Like, I’m capitalized correctly.

Ari Weinstein [00:14:50]: That’s hilarious.

Swyx [00:14:51]: Like, I’m giving full num- full reference numbers and everything. Like, it’s too. it’s clearly too good. I don’t care. Like Like, I’m just, like, trying to get my support case.

Vibhu [00:14:58]: I’ve prompted it to, like, you know, “Don’t pretend you’re a bot. Be very annoyed human.”

Vibhu [00:15:02]: Short one-liners, like

Swyx [00:15:04]: Yeah

Vibhu [00:15:04]: Push it, do all this. I also tell it, “While you’re waiting for responses, like, use subagents to research better ways to figure out what we need.”

Ari Weinstein [00:15:12]: Nice.

Vibhu [00:15:12]: It’s just, like, human little intervention.

Ari Weinstein [00:15:14]: That’s awesome. I also feel like half the time it’s a bot on the other end, so now you

Vibhu [00:15:17]: Yeah

Ari Weinstein [00:15:17]: Got the bots talking to each other.

Swyx [00:15:18]: Yeah. I will also say, you know, like, you know, one milestone of Computer Use that we are, we’re at now is, you know, three, four years ago, we were scared of hooking up LLMs to the, to the web and to

Ari Weinstein [00:15:31]: Yeah

Swyx [00:15:31]: To our, to our devices. And now I’m having it configure DNS for me.

Ari Weinstein [00:15:35]: Wow.

Swyx [00:15:36]: I’m having it pay my bills, and, like, really, like, tens of thousands of dollars of, like, stuff I’m just sending it over and yoloing with Computer Use and, like, you know, what’s the, what’s the worst thing that can happen?

Swyx [00:15:48]: So that- that’s all, that’s all really good.

Building Safely With Computer Use in the Agents API

Ari Weinstein [00:15:50]: Yeah.

Swyx [00:15:50]: I think now that you’ve. you know, obviously, you also have to dogfood your own products and all these things. Now that you’ve sort of released this in API, what are some pitfalls or tips that you wanna tell developers, because they’re about to, I guess, encounter all this, firsthand?

Ari Weinstein [00:16:03]: First of all, I’m just really excited that we brought Computer Use into the Agents API. I think this is, really great because obviously a lot of developers are building applications that wanna be able to work with third-party websites and services. And so Computer Use has this universality to it. It can work with anything. So now all of a sudden, developers can build using the same Computer Use implementation that we’re building on. I think there’s great work to be done if you wanna build your own Computer Use harness, but it’s hard. And also, we train our models on our Computer Use harness, so there is, an advantage to using the one that’s in distribution for the model. There actually might be a speed and cost and accuracy advantage. So I think it’s really great for people to get to build on top of that. And, yeah, you know, I think kind of to the point that you were making, like, I think we’re all sort of still in the process and maybe, like, some of us are ahead of many people in the world of, like, getting comfortable with this technology and trusting it. And so I think it’s incumbent on us to, sort of build that trust over time by making sure we’re building things that are reliable, by building, the right kinds of safety checks, by asking for the user’s consent before doing something consequential like making a payment, by, asking, you know, maybe depending on the application, making sure you’re only letting it access the websites or applications that it actually needs for the task. So that’s, I think, something important to think about. but yeah, I’d really encourage people to try the new Agents API, build all kinds of cool stuff on it. We’d love to hear your feed- feedback if, you know, depending on how it goes.

Vibhu [00:17:31]: Have you seen any changes in the way it affects dev workflows? So one of the things with dots is, you know, you’re seeing it in Slack.

Ari Weinstein [00:17:38]: Yeah.

Vibhu [00:17:38]: You’re seeing people use voice and build. the example Roman showed of change this app and send me screenshots along the way and all this.

Computer Use for Testing and Closing the Software Loop

Ari Weinstein [00:17:46]: Yeah.

Vibhu [00:17:46]: Is anything that you’re seeing there in adoption about how people are using Computer Use for coding workflows? Any tips people should take from that?

Ari Weinstein [00:17:55]: One of my favorite use cases for Computer Use actually, and one that we see a lot in the wild, is Computer Use letting the agent- actually test the software that the agent has built, which is far more consequential than it sounds. Because traditionally, you know, you’d build something in Codex and then the-- and the Codex builds it for you, and then you have to test it, and you are now like QA for the agent, right? So with Computer Use, you can complete the develop-- the software development life cycle, where, the agent can build software, it can test it. So I have a lot of fun, you know, building stuff, having the agent test it. By the time it comes to me, it’s already working. I have, extra fun because sometimes I’m, like, developing Computer Use itself, and so now I have a Computer Use agent that’s using my Computer Use agent that’s using something else. so yeah, I really, I really think this is a super powerful class of use case.

Swyx [00:18:44]: I have a visual play test skill that I’ve developed that, really catches a lot of design issues,

Ari Weinstein [00:18:49]: Nice

Swyx [00:18:50]: That, you know, normally when you just look at code, you wouldn’t really pick it up. it’s also really good for cloning apps, though. If you’re using a shitty SaaS and you wanna kill the SaaS You just clone it screen by screen by screen. and Obviously, Computer Use can completely drive everything, take screenshots, note it down, and then clone everything with Codex.

Ari Weinstein [00:19:06]: That’s really cool.

Swyx [00:19:06]: But yeah, thanks for all your progress. I think, that is

Ari Weinstein [00:19:08]: Absolutely

Swyx [00:19:09]: Our time.

Nikunj Handa: What’s New in the OpenAI API

Ari Weinstein [00:19:10]: Yeah.

Swyx [00:19:10]: This is not the last that we’re gonna talk.

Ari Weinstein [00:19:12]: Yeah, cool. This has been really fun. Thank you guys for having me.

Swyx [00:19:14]: All right.

Vibhu [00:19:14]: All right. Okay, we’re a strict cutoff. We’re just gonna dive right in.

Nikunj Handa [00:19:17]: Let’s do it, yeah.

Vibhu [00:19:19]: Okay, so, Nikunj, we’re very excited to have you. You shipped a lot on the API side, like we just

Nikunj Handa [00:19:25]: Yeah

Vibhu [00:19:25]: Talked about with Ari. You can now build with Computer Use agents. Anything you wanna highlight, the API side of changes, and introduce yourself a little and what you do?

Nikunj Handa [00:19:34]: Yeah, for sure. My name is Nikunj. I lead product for the API team. Been here for roughly three years. been working on launching models. I feel like that’s just been, like, a thing, constant thing throughout my time, here at OpenAI. And, with every new model, we try to, like, basically work super closely with the post-training team, the research team, to figure out what’s new in it. and then we, like, expose those capabilities in the API. so that’s, like, the basic way of putting it. and if you just look at, everything that’s new with GPT-6, the cool new capabilities that we launched were, firstly, async function calling. so what you see with, like a lot of the things that you’re seeing in, like, Codex and Dots and everything is that tool calls take so long that you don’t have to, like, pause the model’s execution while, the tool is running. So you could just, like, kick off a tool call, keep running, keep reasoning, and then check back in. so we launched async tool calling. We launched, like, mid-turn steering, so now you can, like, inject messages while the model is reasoning, in the middle. so as your tool call finishes, you can put in that instructions.

Async Tool Calls, Mid-Turn Steering, and WebSockets

Swyx [00:20:43]: And that’s also partially a model alignment capability, right?

Nikunj Handa [00:20:46]: Yeah.

Swyx [00:20:46]: Like, they have to train in the ability to train.

Nikunj Handa [00:20:48]: Exactly, yeah. And

Vibhu [00:20:49]: I feel like we’ve had it in the app. You could always, as it’s reasoning, you could steer.

Nikunj Handa [00:20:54]: Yes.

Vibhu [00:20:54]: It wasn’t the best. It’s gotten much better.

Nikunj Handa [00:20:57]: Yeah.

Vibhu [00:20:57]: Excited to see how it does this in version

Nikunj Handa [00:20:58]: Yeah, and I like our main

Vibhu [00:20:59]: And now

Nikunj Handa [00:21:00]: Goal in, our main goal in the API is to, like, put things in the API once it’s trained into the harness. And so we kinda wait for that moment until it’s good enough. And a lot of that is, like, actually being powered by WebSockets, which we launched, a few, I wanna say months ago. And so WebSockets just opens this, like, whole bidirectional, like, communication thing with the model. This is not, the GPT Life thing. I’m just talking about GPT-6. and you can do all these, like, async tool calling, async reasoning, injecting messages. It’s a really fun API to work on. I think, like, really enjoying.

Swyx [00:21:33]: Yeah. This is why we are the engineering podcast, because we get to talk about WebSockets.

UltraFast and the Inference Stack

Nikunj Handa [00:21:36]: Yeah.

Swyx [00:21:37]: This also pairs very well with UltraFast, right?

Nikunj Handa [00:21:39]: Oh, yeah.

Swyx [00:21:39]: Like, that is now, like, I think for the first time ever available in the API.

Nikunj Handa [00:21:43]: Yes.

Swyx [00:21:43]: Which is, which is basically the theoretical fastest speed you can ever get, Frontier of Intelligence.

Nikunj Handa [00:21:49]: Yeah. It’s been so exciting to work on that project. I think, before I go into the API, the most fun part of, UltraFast has been just watching the inference team cook with Astra. Like, they’re just, like, constantly having these, like, Codex agents running, trying to, like, squeeze out more performance. And, I would say, like, at least for a couple of months, a lot of it was focused on efficiency and driving the cost down, which is how we, like, were able to cut the Luna price by, like, 80%. It was, like, a lot of that was driven by, like, all the inference improvements they landed. And then now they’ve, like, shifted gears towards, like, how can we make this run as fast as possible? And so UltraFast has just been, like, amazing to see on a mo- on a model like Astra. Like, to go that fast has been really cool. And yeah, WebSockets is like. actually it was like the first time we launched WebSockets, it was for GPT, 5.3 Codex Spark, which was. Can’t believe we named a model that, but, you know, that’s what we launched it for. And obviously, it helps so much because, like, you gotta have the tool calls. you had, like, really reduced the overhead, of going back and forth with tools. And so, WebSockets is awesome for that.

Swyx [00:22:57]: Yeah. it’s always cute to see, like, I have my reset usage limit, and then I have my Spark usage limit that I never use.

Nikunj Handa [00:23:03]: Yeah.

Swyx [00:23:04]: Like, it’s there if I want it.

Nikunj Handa [00:23:05]: I think it’s gone finally.

Swyx [00:23:06]: It’s gone. It’s gone, yeah.

Nikunj Handa [00:23:07]: I know it’s gone, so.

Swyx [00:23:08]: Yeah. you’re slowly killing off all the, you know, the

Nikunj Handa [00:23:11]: The old ones, yeah.

Swyx [00:23:11]: Oldies.

Vibhu [00:23:11]: This is a great week. I mean, it was the first time we had Frontier Intelligence at extreme speeds.

Nikunj Handa [00:23:17]: Yeah.

Vibhu [00:23:18]: People really liked it.

Nikunj Handa [00:23:19]: Yeah.

Vibhu [00:23:19]: So

Swyx [00:23:20]: Yeah

Vibhu [00:23:20]: First time it comes back.

Swyx [00:23:21]: Yeah. for, 5.3 Spark is explicitly attributed to Cerebras. You guys are not confirming or denying that, UltraFast is related to Ce- Cerebras, but people are. I’ll just say that people do care and, are wondering about it. And you have your own silicon as well. elephant in the room, decision models.

Decisions API: OpenAI’s Fast Decision Model

Nikunj Handa [00:23:38]: Oh, yeah.

Swyx [00:23:38]: Decisions API. We were the first podcast to do a big Jev, deep dive with, Diogo, and I also, you know, featured him at AI Engineer. How quickly did you see Jev and go like

Nikunj Handa [00:23:49]: Oh my gosh. Yeah.

Nikunj Handa [00:23:50]: Yeah. Firstly, like, huge props to Diogo and, like, the Jev team for, like, really inspiring the

Swyx [00:23:55]: Yes

Nikunj Handa [00:23:55]: Like, whole segment in the market. Like, obviously Jev comes out, everyone’s, like, losing their minds over it. Our users are, like, hitting us up. But also, like, our internal teams are like, “We need, like, a much faster classification system.” We can. I don’t wanna, like, get ahead of some of the dots features that are gonna come

Swyx [00:24:16]: Whoo

Nikunj Handa [00:24:16]: But you’re gonna see, like, some cool, like, really snappy, fast things built on top of the decisions API. but, you know, like, yeah. Props to Jev for, like, inspiring this whole thing. obviously a bunch of people at OpenAI get nerd sniped by that, and they’re like, “How can we, like, make this work? We’re not gonna, like-”

Swyx [00:24:33]: Okay.

Nikunj Handa [00:24:33]: “. train a new model.” But

Swyx [00:24:34]: Like, four weeks ago, this was not on the dev radar, right?

Nikunj Handa [00:24:37]: No, not at all. No.

Swyx [00:24:37]: Okay.

Nikunj Handa [00:24:37]: This is like

Swyx [00:24:38]: Wow

Nikunj Handa [00:24:38]: Jev-inspired and, like

Swyx [00:24:40]: I think you are officially the first one to your lab to, like, clone and, adopt this.

Nikunj Handa [00:24:44]: Yeah. Yeah. I feel like, OpenAI has such a strong, like, hacker culture and, like, people are just, like, they get excited about things. And so, guy from inference, this one awesome guy from, the infra team are like, “ this is amazing. We’re gonna, like, hack on it.” They build a prototype, it, like, works, and now we- we are just, like, hill climbing on latency and trying to make this as fast as possible, and we wanna, like, launch it in the coming days. so as soon as we hit our, like, latency target, we’ll try to get this out.

Vibhu [00:25:13]: It’s interesting. At the same time of hacker culture, you also, as Sam said, like 99%, one of the most reliable APIs with

Nikunj Handa [00:25:20]: Mm-hmm

Vibhu [00:25:20]: I think probably the most usage, which is your team directly. how should people see decisions API? I feel like a lot of people saw Jev, heard the buzz, haven’t built with it. You’re making it very mainstream.

What Decision Models Are Good For

Nikunj Handa [00:25:32]: Mm-hmm.

Vibhu [00:25:33]: What should people see it as? How should they use it?

Nikunj Handa [00:25:36]: Yeah. I think the main use cases we’ve seen is, like, really fast classification. all the Computer Use demos have been amazing and really cool. I think there will be limitations, of course, in terms of, you know, having Astra, like, write, like, a JavaScript-like script to control your computer, versus having Luna pick, like, one action at a time. I think, it’s not gonna be at the same intelligence level, but, like, maybe there’s some Computer Use tasks that this is good enough for. So excited to see that come through. the other cool prototype I’ve seen internally is people hooking it up with GPT Live. So GPT Live is like, you know, our bidirectional, like, real-time,

Swyx [00:26:14]: Voicing

Nikunj Handa [00:26:14]: A- API. And, it’s built on this, like, model of front-end models and back-end models. So GPT Live is this, like

Swyx [00:26:20]: Think or talker

Nikunj Handa [00:26:21]: Super fast. Yeah, think or, talker thing. So GPT Live is the talker, super fast, really good at delegation, and you have something like Astra sitting at the ba- at the back. But tool calling has always felt, like, really slow in GPT Live. and so people have been, like, putting together these, like, tool calling demos of GPT Live controlling a computer, and it just feels like so much more snappy and natural. So I’m, like, kinda excited to see, like, what people do with Live and with Luna on decisions API. so that’ll be pretty exciting. Yeah.

Swyx [00:26:55]: So I wanna iron this out for people, especially from the product side, because a lot of people have been putting out Jev clones. There’s been about 100 in the last two weeks.

What Makes a Decision Model Different

Nikunj Handa [00:27:01]: Oh, really? That’s amazing.

Vibhu [00:27:03]: The first couple days.

Swyx [00:27:04]: But like, it. Like, they can clone a Jev API, which is honestly structured outputs

Nikunj Handa [00:27:09]: Yeah

Swyx [00:27:09]: Which OpenAI was first to.

Nikunj Handa [00:27:10]: Yeah.

Swyx [00:27:11]: Right? So, like, I think let’s iron out for people what is a decision model, as far as

Nikunj Handa [00:27:16]: Yeah

Swyx [00:27:17]: As far as, like, what is important? It is not just latency. It’s not just structured output, right? Because I could just have Luna as it’- The decision model is priced the same as Luna, right?

Nikunj Handa [00:27:26]: Mm-hmm.

Swyx [00:27:27]: Have turned off reasoning and then have structured output. Do I have a Jev? you know, no, right? And that’s the

Nikunj Handa [00:27:33]: Yeah

Swyx [00:27:33]: That’s the real

Vibhu [00:27:34]: There’s a confidence there.

Swyx [00:27:35]: Yeah.

Nikunj Handa [00:27:36]: Yeah, totally. I think, the way that. So we haven’t trained, like, a new model for this.

Swyx [00:27:40]: Yeah.

Nikunj Handa [00:27:40]: We’re, like, building this purely on top of the same Luna weights that we have.

Swyx [00:27:44]: Oh.

Nikunj Handa [00:27:44]: So yeah. This is, like, really just Luna. And, on top of that, what you’re doing is you’re constraining. So, like, structured output’s a big part of it. you’re really optimizing the inference stack to, like, get very fast on TTFD. And because you can have multiple questions, what you do is, like, you basically run those in parallel,

Swyx [00:28:05]: As a batch.

Nikunj Handa [00:28:06]: Yeah. You run those in the-- as a batch. you-- All sorts of, like, inference techniques people are working on to try to make it as fast as possible. But I’d say, like, at least our implementation of it at the start and this first version is, like, zero-shotting this on top of Luna, to see how it goes. And obviously, you wanna, like, put it out there. Like, this is OpenAI’s, like, classic iterative deployment thing. Put it out there, see what people think, and then, like, we’ll make more model improvements, as needed. so yeah. That’s, the decisions API.

Swyx [00:28:38]: Yeah. And, obviously as a benefit, you have vision. They don’t have vision, right?

Nikunj Handa [00:28:42]: That’s true.

Swyx [00:28:42]: Obviously, Jev’s comes with

Nikunj Handa [00:28:43]: Yeah. Like, we get it for free with Luna. Yeah.

Swyx [00:28:45]: Yeah. I do think that, like, you know, some of the innovations, it sounds like, it’s still to come if it’s still the same Luna weights, which is, like, the confidence stuff, like, the in calibration is something that we’ve talked about on the podcast with, benchmarking calibration. ‘Cause basically, the whole point is that RLHF kind of collapses you towards what you want to hear.

Calibration, Architecture, and the Open Research Questions

Nikunj Handa [00:29:03]: Yeah.

Swyx [00:29:03]: But, like, not actually, like, what the amount of confidence is.

Nikunj Handa [00:29:06]: Yeah. Yeah, totally. I’m eager to see how it pans out. Maybe there’s, like, gonna be. These are gonna be, like, the key areas where we may have to, like, hill climb

Swyx [00:29:15]: Yeah

Nikunj Handa [00:29:15]: With the, with the future model release. But, yeah.

Swyx [00:29:18]: And then architecture-wise, the other thing that’s in the debate, obviously, you-- Nobody knows because Jev doesn’t talk about it, but the two speculations are, one, maybe diffusion model instead of autoregressive.

Nikunj Handa [00:29:28]: Mm-hmm.

Swyx [00:29:29]: But you are able to achieve the parallel, generation in your way. And then the other one is some mech interp type thing

Nikunj Handa [00:29:37]: Mm-hmm

Swyx [00:29:37]: That you’re, like, analyzing the activations and then just outputting

Nikunj Handa [00:29:40]: That would be cool

Swyx [00:29:41]: The weights.

Nikunj Handa [00:29:42]: Yeah.

Swyx [00:29:42]: Which, like, you guys have all done the research on this. People have speculated.

Vibhu [00:29:45]: There have been demos on

Swyx [00:29:46]: Yeah

Vibhu [00:29:46]: Both of these as well. I think Gemini shared a Gemini diffusion, Gemma diffusion on a Jev-style output.

Nikunj Handa [00:29:53]: Oh, sick.

Vibhu [00:29:53]: And, interp people have also, you know, pulled out interp from a middle layer, but this is all speculation.

Swyx [00:29:59]: It’s just like, what are you trying to aim for, right? Because you can achieve the API. Everyone can achieve the API. It’s actually pretty trivial. But, like, then there’s the speed, then there’s the accuracy, then there’s the other calibration features.

Nikunj Handa [00:30:11]: Mm-hmm.

Swyx [00:30:11]: I don’t know what else.

Nikunj Handa [00:30:13]: Yeah. Yeah. No, totally. It’s so cool that this, like, whole space has been kicked off now and people are gonna do so much cool stuff and everyone’s gonna learn from each other. And, yeah, I’m excited about it.

What Developers Should Build Next

Vibhu [00:30:24]: I feel like being on the platform team, a lot of your job is to empower builders.

Nikunj Handa [00:30:27]: Mm-hmm.

Vibhu [00:30:28]: What do you think people should build with decisions API and also Computer Use agents? Any stuff that you’ve- been building with internally that you think really opens up after the new change?

Nikunj Handa [00:30:39]: Yeah. okay, let’s think. decisions API, use cases internally have been pretty obvious. Like, the user ops team was, like, jumping on it. We were like, “We gotta classify all of our support tickets.” what else came up? obviously, there were, like, the really cool GPT Live demos. I’m sure, like, the Codex app team might, like, pick this up and try to do something cool with it. So, you know, like, this whole thing started, like, a week ago, so it’s, like, very early and

Swyx [00:31:06]: Oh, one week.

Nikunj Handa [00:31:07]: We’re excited. Yeah. Yeah, pretty much.

Vibhu [00:31:08]: There was a big push in, evals, LLM as a judge having really low latency there.

Nikunj Handa [00:31:13]: Right. Yeah. That’ll be interesting to see. and then, with the Agents API, we have-- we’re basically, like, having a bunch of first-party products, like, at OpenAI built fully on top of it. we’ve had the Codex security stuff that just went out that’s fully built on top of, the Agents API. We have, sort of the-- we- we are having, like, a meetings type of thing launching today.

Agents API and OpenAI’s First-Party Products

Swyx [00:31:40]: Mm-hmm.

Nikunj Handa [00:31:40]: I think there was, like, a demo. do you remember, like, the plugin extensions when Sam was showing it? There was, like, a demo for, like, you’re in a calendar, you can sort of, like, have your meeting notes

Swyx [00:31:51]: Like, drop into a single

Nikunj Handa [00:31:52]: Flow into like your space

Swyx [00:31:52]: Like, Google Docs type thing.

Nikunj Handa [00:31:53]: Yeah.

Swyx [00:31:54]: Right?

Nikunj Handa [00:31:54]: And so the-- all of that stuff is, like, fully built on top of, the Agents API. and yeah, I’m, like, just excited to see. Like, we’re just getting this out, and let’s see what people build on top of it.

Vibhu [00:32:04]: I think you showed it off very well. The whole edit spaces, pages, collaborate, add in your dot. Like, that’s a lot, so

Nikunj Handa [00:32:12]: Yeah

Vibhu [00:32:12]: There’s a lot of inspiration people can go to.

Nikunj Handa [00:32:14]: Yeah. All possible with Astra, you know. Like, thing- things just move so fast now. Like

Swyx [00:32:19]: Yeah

Nikunj Handa [00:32:19]: People go from idea to execution so quickly, it’s amazing.

Swyx [00:32:23]: Is there something that you want, people to focus on to give you feedback? Like, what-- like, you know, maybe you’re just putting this out there and you want-- and there’s, like, a fork in the road and you want developers to help you decide.

Responses API Performance and Long-Lived Caching

Nikunj Handa [00:32:35]: So I think Agents API and decisions API, they are like, these are our newest products. Would love, like, any and all feedback on that to figure out where to take them. I think, over here, we’re, like, very open on Responses API, which is sort of like our workhorse over here. like, really focused on performance right now, and the performance comes in, like, two main ways. first is just, like, latency. We’ve been, like, rewriting the whole Responses API stack to, like, make it as fast as possible from a TTFT perspective, DVD perspective. So there’s like-- that, like, continues to be, like, a main area of focus for us. The second thing we’ve been trying to do is, like, really go deep on caching, particularly with these, like, personal agents that are, you know, like, basically, like, a single thread that just goes on and on forever. We’ve been, trying to, like, really up our game on caching. We provide now guarantees of, like, cache hits within, like, 30 minutes. We’re actually, like, we-- for one of our users, we just launched, like, a much longer cache window. So we have, like, a 12-hour caching guarantee, that we offer so that you have, like, guaranteed cache hits for

Swyx [00:33:40]: Is that a public API?

Nikunj Handa [00:33:42]: Not yet. That’s in preview.

Nikunj Handa [00:33:43]: We’re gonna, like, try to get that out to everyone as soon as possible. But, like, just pay a little bit more for the cache write, and we, like, guarantee, like, cache reads for, like, a much longer period. So even if, like, your instinct thread, for example, like, you just, like, do something on it and then come back to it, like, three to four hours later, you- you’re still getting the caching performance out of it. And launched

Vibhu [00:34:04]: And you cut the cost there quite a bit too, right, with the new model?

Nikunj Handa [00:34:07]: Oh, yeah. That’s right.

Vibhu [00:34:08]: Like, 25% cheaper, so

Nikunj Handa [00:34:08]: Yeah, with, like, driving down cache reads, yeah.

Cache Pre-Warming and Cost-Efficient Agent Threads

Vibhu [00:34:10]: For builders, they should implement

Nikunj Handa [00:34:13]: Yeah

Vibhu [00:34:13]: Because it’s significantly cheaper.

Nikunj Handa [00:34:14]: Yeah. Yeah. Just, like, building your apps with, like, to be very cache aware and sort of, like, use our prompt diagnostics or cache diagnostics tool to figure out, like, where things are dropping off. And, so the caching part is, like, really important. yeah, I also wanted to talk about pre-warming. We have that in the API now. So, like, if you know that, “Hey, I’m gonna get a cache,” like-- sorry, “I’m gonna get this prompt. I just wanna, like, pre-warm the cache, pay, like, the cache write fee right now, and then, like, have it sort of ready to go for the next 30 minutes for whenever.”

Swyx [00:34:49]: And it can spawn many instances of that thread.

Nikunj Handa [00:34:51]: Exactly, yeah.

Swyx [00:34:52]: Yeah.

Nikunj Handa [00:34:52]: You can just keep going and have

Swyx [00:34:54]: Yeah, just keep messing with the prompt there

Nikunj Handa [00:34:55]: Tons and tons of that. and so, yeah, like, I’m very excited about getting feedback on, like, the low-level performance things that we can keep making Responses API the most performant and reliable way to, like, build on top of an LLM. And then you basically have our, like, new products where I’m just looking for, like, any and all feedback.

Swyx [00:35:15]: Yeah, just use it, right?

Nikunj Handa [00:35:16]: So yeah, just use

Swyx [00:35:16]: Tell us what to

Nikunj Handa [00:35:17]: Yeah. Define our roadmap for us, please. So yeah.

Swyx [00:35:20]: I think for me, the caching thing, great, right? Like, obviously very needed. But at the end of the day, you’re still bumping up against a million-token context

Compaction and Managing Million-Token Contexts

Nikunj Handa [00:35:28]: Mm-hmm

Swyx [00:35:28]: And that’s probably not gonna change for the foreseeable future.

Nikunj Handa [00:35:31]: Mm-hmm.

Swyx [00:35:31]: Like, you still need good compression.

Nikunj Handa [00:35:33]: Yeah.

Swyx [00:35:33]: What is the best practice there?

Nikunj Handa [00:35:34]: Yeah. Yeah, totally. so firstly, OpenAI has its own, like, proprietary compression, comp

Swyx [00:35:40]: Which is in

Nikunj Handa [00:35:41]: Compaction.

Vibhu [00:35:42]: Compaction.

Swyx [00:35:42]: It’s in the agents.

Vibhu [00:35:43]: It’s in the API.

Nikunj Handa [00:35:43]: Yes.

Vibhu [00:35:43]: Agents API.

Nikunj Handa [00:35:44]: Yeah.

Swyx [00:35:44]: You decide for us, right?

Nikunj Handa [00:35:45]: Yeah, exactly. So in the Agents API, it comes built into the harness. and if you’re in Responses API, there’s, like, two ways of doing it. One is what we call server-side compaction, which is you basically tell Responses API that if you ever hit this threshold of tokens, just auto-compact it and, like, go back, or sorry, like, reduce the context, being used. And the second way is, like, /compact, which is, like, if you want full control. So you can, like, /compact at any time

Swyx [00:36:15]: I hear you

Nikunj Handa [00:36:15]: Have your own logic on when to, like

Swyx [00:36:17]: It’s not AGI.

Nikunj Handa [00:36:18]: It.

Swyx [00:36:18]: It’s not AGI.

Nikunj Handa [00:36:19]: Yeah. Yeah.

Swyx [00:36:20]: Yeah. But it, I mean

Nikunj Handa [00:36:20]: Yeah

Swyx [00:36:20]: It is the manual override.

Nikunj Handa [00:36:21]: Yeah, it is the manual way. And like, I don’t know, but a lot of the big coding agents like to do it manually. I mean, like, if you look at the Codex implementation of it in the Code- open source Codex harness, you can see that they use /compact and do it. and, there’s also, like, new, by the way, new compaction techniques that we are working on. Some of them you will be able to see in the Codex harness. Like, it’s already implemented in the Codex harness. And so, they’re like some file-based, systems that we are, like, experimenting with. So yeah, lots of cool stuff going on around in compaction as well.

Swyx [00:36:57]: Cool. we are running out of time.

Nikunj Handa [00:36:59]: Okay.

Swyx [00:36:59]: I think you’ve talked about, a lot about performance and talked a lot about, the new APIs that you’re launching. Can you give us any other hints as to things that you’re interested in as far as the future of the platform is concerned?

Higher-Level Platform Primitives and the AI Cloud

Nikunj Handa [00:37:13]: We’re obviously like very low level. Like, I used to work at Stripe before this, and, at Stripe a lot of the game was like building these higher level primitives and products on top of like the core payments primitives. and, I’m always like curious about what the best way of doing that is in AI. And I think we’ve had a couple of attempts at that. We like had launched assistance API like way back in the day, and like wasn’t really the right fit. We were sort of like going off with this like Agents API, and, it gives you the codex harness, but like where’s like the, what’s the right amount of flexibility to give in that? That’s like an open question. Like how should we like have memory walls and like all of these like higher level like API objects to take away, also like to abstract away more, like storage concepts. Like this is like a whole, like, there’s a whole space that I’m like very curious about figuring out how we design. I think a lot of things in AI are just have a low-level API primitive and see an example harness and go and have your coding agent implement that. But how much of that do we build into the API is like a constant question that I’m thinking about.

Swyx [00:38:24]: Yeah.

Nikunj Handa [00:38:24]: So I don’t know if folks have thoughts on that. If anyone has ideas, it would be super interesting to hear.

Swyx [00:38:30]: Yeah. The analogy I always bring back to, and we’ll end there, is, you’re building an AI cloud, right?

Nikunj Handa [00:38:35]: Mm-hmm.

Swyx [00:38:35]: Like, which is, something that, Sam said a year ago

Nikunj Handa [00:38:38]: Mm-hmm

Swyx [00:38:38]: Where, and you’re, it’s almost like you’re kind of doing the AWS invention and you have to do, okay, this is EC2

Nikunj Handa [00:38:45]: Yeah

Swyx [00:38:45]: And this is S3, and this is like. But you’re doing the AI-native versions of each of these.

Vibhu [00:38:48]: There are a lot of analogies, so you’re pre-warming caches for stuff that you know will be

Nikunj Handa [00:38:53]: Yeah.

Vibhu [00:38:53]: And it’s nice that it’s all exposed to builders

Closing

Nikunj Handa [00:38:56]: Mm-hmm

Vibhu [00:38:56]: ‘cause it just opens up ways that you can build new things.

Nikunj Handa [00:38:59]: Yeah, absolutely.

Swyx [00:39:00]: Okay.

Vibhu [00:39:00]: Awesome. Well

Swyx [00:39:01]: That’s everything.

Nikunj Handa [00:39:01]: Thank you, guys.

Vibhu [00:39:02]: Thank you.

Nikunj Handa [00:39:02]: Yeah.

We are excited to have Anthropic share their latest AI x Finance work at AI Engineer New York, coming up in 2 weeks!

In case you’ve been under a rock, here’s a non-exhaustive list of what Anthropic has been shipping since closing the largest fundraise of all time in May at $47B ARR:

* June: Launched Claude Tag and Sonnet 5 and Fable 5

* July: Opus 5, /checkup. crossed $65B ARR

* Last month: Fable/Mythos 5.1, and EFS (upcoming pod)

* IPO target $2T, end 2026 ARR estimated $100B

* Cowork/chat merged before did

* Claude Mods

* Dario endorses the same Pacing the Frontier message cosigned by all labs

* Last week: Opus 5.5, Plugins portal, Cloud Sessions/Claude Projects

* Today: Sonnet 5.5!

Today’s episode should catch you up, with Thariq Shihipar, the explainer-king of Anthropic, who we last caught up on Fable launch day with The Field Guide to Fable:

The Future of Mutable Software

Pay special attention to Claude Mods (especially the cheatsheet):

In general this is also the inverse of the other viral tweet from Thariq:

Cloud Brain, Local Hands

And give a try to Claude Projects:

The “hands” terminology is not just an analogy for the local/cloud paradigm that is being built up at frontier coding agent companies like Cognition, but is ALSO particularly relevant to the safety systems discussions that we’ll be discussing with Anthropic in an upcoming episode as they prepare to pace to frontier with responsible AI deployment.

For those who want Thariq’s writing tips we teased at the start of the pod, watch the full video here:

From the rapid rise of Claude Code to a future where agents can rewrite their own harnesses, collaborate across teams, and operate across cloud and local environments, the way we build software is changing extraordinarily fast. In this episode, Anthropic’s Thariq Shihipar joins swyx and Vibhu to unpack how power users are actually working with Claude Code today, why prompting remains a high-skill discipline, and where Anthropic thinks the agent harness is headed next.

We go deep on Claude Code’s evolving interface: Ask User Question and elicitation, artifacts as persistent generative interfaces, Claude Tag for multiplayer agent workflows, Projects, model effort, implementation notes, and the new Claude Mods system for customizing the harness itself. Thariq explains why Claude.md may eventually disappear, why the smartest model could also become the cheapest model for many tasks, and why mutable software could become a new paradigm for how applications are built and customized.

The conversation then turns to agent security and Anthropic’s “Pacing the Frontier” argument. Thariq walks through recent incidents where agents discovered unexpected ways to communicate, exploit infrastructure, reverse-engineer benchmark scorers, and chain vulnerabilities together. We discuss sandboxing, prompt injection, autonomous agents, interpretability, constitutional classifiers, probes, fallbacks, Auto Mode, and why securing increasingly capable agents may become one of the defining engineering problems of the next few years.

We discuss:

* Why agentic coding went from controversial to the default in less than a year

* Why prompting is still one of the highest-leverage skills for working with Claude Code

* How expert users build a mental model of Claude and what it can reliably one-shot

* Why discovering your “unknown unknowns” matters more as agents become more capable

* Artifacts as persistent, generative interfaces between humans and agents

* How Claude could split into a cloud-based “brain,” local or remote “hands,” and dynamic interfaces

* Claude Tag, Projects, and multiplayer agents and how collaborative agent workflows could evolve

* Why spending more time on the initial prompt can dramatically reduce wasted agent work

* When to use low, medium, high, or max effort for different engineering tasks

* Why frontier models may eventually outperform smaller models on both intelligence and token efficiency

* Why implementation notes can expose decisions the model considered but chose not to make

* Why Claude.md may eventually disappear — and why starting without one can sometimes be better

* Claude Mods: customizing the execution loop, UI, subagents, routing, and behavior of Claude Code

* Model routers, forked agents, and supervisor agents that automatically improve agent workflows

* Why Claude Mods may be an early preview of “mutable software”

* The bitter lesson of harness engineering and why agent architectures go out of date so quickly

* How Claude Tag is becoming an organizational harness for multiplayer work

* Why giving agents access to company data creates an enormous new security surface

* The Exploit-Bench incident where agents discovered ways to communicate and collaborate

* Why agents hacked Hugging Face for scorer code rather than benchmark answers

* How agents chained sandbox and infrastructure vulnerabilities in unexpected ways

* Why increasingly capable agents make traditional security assumptions harder to maintain

* The argument behind Anthropic’s “Pacing the Frontier” proposal

* Why software engineers are increasingly doing two jobs: engineering and keeping up with AI

* Constitutional classifiers, probes, and fallbacks and what interpretability looks like in production

* How Auto Mode checks whether an agent’s actions actually match the user’s permissions

* Why Thariq can see serious AI risks while still having a relatively low p(doom)

Thariq Shihipar

* X: https://x.com/trq212

* LinkedIn: https://www.linkedin.com/in/thariqshihipar

Timestamps

00:00:00 Introduction

00:04:12 Ask User Question and the Future of Agent Interfaces

00:08:29 Artifacts, Projects, and Multiplayer Agents

00:15:37 Prompting as the Core Claude Code Skill

00:21:52 Context, Effort, and Smarter Model Usage

00:28:10 Is Claude.md Going Away?

00:32:49 Claude Mods: Customizing the Claude Code Harness

00:36:35 Model Routing and the Rise of Mutable Software

00:44:40 The Bitter Lesson of Harness Engineering

00:50:49 Claude Tag as an Organizational Harness

00:55:59 Pacing the Frontier and Autonomous Agent Security

00:58:22 Agents Hack Hugging Face for the Scorer

01:05:34 What Happens When Agents Need More Compute?

01:10:32 AI Coding Is Changing Faster Than Engineers Can Keep Up

01:17:17 Probes, Fallbacks, Interpretability, and Auto Mode

01:28:32 AI Risk, p(doom), and Closing Thoughts

Transcript

Introduction: Life at Anthropic and the Pace of Change

Swyx [00:00:00]: We’re here in the studio with our friend Thariq from Anthropic, and I guess generally the Claude Code, I-- there’s, there’s so much, merging of boundaries and you’ve been so on top of everything since you joined Anthropic. You have been early to Claude Code itself, but then also, and you’ve told that story in other podcasts, and you’ve also been talking about seeing like an agent. Most recently you did the top AIE World Tour talk, Field Guide to Fable, which obviously you guys launched Fable, so that was-- that’s cheating. And mostly you most recently also launching Claude Tag, and we’re also gonna be talking about Pacing the Frontier. There’s a lot going on in Anthropic. I guess top of the question is, what’s it like being at Anthropic when there’s so much going on?

Thariq Shihipar [00:00:48]: I think that It is, like. I think you can get whiplash sometimes. I think, like, going. When I joined Anthropic, I joined because of Claude Code. Like Claude Code had just come out and I was like, “This is so good.” And Opus 4 to me was like just, I could not imagine, like, how good it was? And that was, like, a real moment for me. But I was, like, trying to convince, like, my startup friends to use agentic coding, and they’re like, “Oh, no, like, our engineers don’t think it’s good enough,” or something. And I was like, “That’s insane.” and now you, like, fast-forward, 12 months, less, and, like, it’s just like, yeah, the default way that everyone codes, right? And I think that, like, just having to go from, like, selling it to, like, now, teaching people how to be. make the most use of it and be more efficient and things like that is just like a big, like big change. And, yeah, I think, like, it’s just hard to stay on top of everything as a human? Like, I think things happen so fast and like

Swyx [00:01:51]: You just throw more agents at it.

Thariq Shihipar [00:01:52]: Yeah, like that’s like the agentic stuff scales much better than the, like, human stuff where it’s like, oh, like, there are three things happening right now and, like, they’re all emergencies and, like, how do you, like, respond to it? Yeah.

Teaching People to Use Claude Code

Vibhu [00:02:05]: What do you split your time on? You do a lot of technical writing, engineering work.

Thariq Shihipar [00:02:10]: Yeah, so I think that, like, when I joined the Claude Code team, I wanted to teach people how to use Claude Code and I think that, like, that has been something that, like, I thought, like, maybe I would spend a little bit of time on it or, like, I’d, like, do. I was spending some time on the agent SDK first, and I wasn’t exactly sure, like, how the bitter lesson would go, when it comes to, like, harnesses, right? Like, I think sometimes we were like, “Oh, like, what’s after Claude Code?”? And so initially I was like, I just wanna teach people how to use Claude Code and make it easier to use Claude Code. And I think that has just, like, as the harnesses have gotten better and better, that’s like the dominant problem now is, like, how do you use the agents, right? Like, it’s like such a high skill expression thing. So I do that and then I do engineering work. I give talks, but I think, like, when I’m doing engineering work, my goal is to take that feedback that we get from users and also, like, then be able to talk about, like, hey, how to use Claude Code to do engineering. So there’s like a good loop there. Yeah.

Swyx [00:03:07]: Yeah. I’ll-- For listeners, we’ll attach, the talk that you did with Sarah for the Dev Writers, meetup

Thariq Shihipar [00:03:13]: Oh, yeah

Swyx [00:03:13]: Which we talked a little bit about, well, first you do the work and then you talk about the work.

Thariq Shihipar [00:03:16]: Right.

Swyx [00:03:16]: Something like that.

Thariq Shihipar [00:03:17]: Yeah.

Swyx [00:03:17]: It’s sow and reap or

Thariq Shihipar [00:03:19]: Yeah, reap and. Sow and reap.

Swyx [00:03:21]: Something like that. Something like that. Yeah, so, and then just to preview a little bit, we are gonna talk about the evolution of the harness. It has come a long way from just being a CLI. We’re gonna talk about, Claude Mods, which is starting to leak today, because you couldn’t keep it secret.

Thariq Shihipar [00:03:36]: Yeah. yeah.

Swyx [00:03:39]: Yeah, there’s, there’s a lot, there. I think you started off with, like, adding ask user question tool, which people love and hate.

Thariq Shihipar [00:03:48]: Yeah.

Swyx [00:03:48]: Like, I thought it was, like, very innovative, and then now I have, like, my own version. You have your Interview Me version.

Thariq Shihipar [00:03:55]: Yeah.

Swyx [00:03:56]: And, yeah, everyone just has, like, their own stuff. And, like, it no longer matters ‘cause now you’re supposed to, write prompts that create other prompts and loops and all these things.

Ask User Question and Human-Agent Interaction

Thariq Shihipar [00:04:05]: Sure, yeah.

Swyx [00:04:06]: So what’s the state of the art, today? Like, what are people. what are you, like, telling people to do today?

Thariq Shihipar [00:04:12]: Yeah, ask user question was the first time that the model was good at elicitation. I think this was, like, an emergent behavior that I, like, wanted to see if the models could do. I have, like a human-computer interaction background, so I, like, did that in undergrad and grad school. And so this was like. I think it’s like human-agent interaction to me, like, trying to figure out, like, how can the agent communicate with you and extract, the requirements, right? I think that, like, one of the things about, like, that’s difficult as Claude Code has gone broader and broader is that everyone has, like, their own way of using it, and it’s very hard to, like, change the default behavior. So for example, like, if someone asks Claude Code to do something,

Thariq Shihipar [00:04:59]: Sometimes they just want them to do the work, ‘cause they’re, like, maybe a very good prompter, and sometimes they want. like, are not good at prompting? And you need. like, the agent needs to, like, clarify? And so that’s, like, a good split. Like, and the ask you the question tool like, splits along that side where, like, are-- do you feel like you’re good enough to instruct the agent as it is, or is the agent able to, like. does the agent need to, like, pull out more requirements and, like, collaborate with you more and really understand your preferences?

Thariq Shihipar [00:05:27]: I, on the whole, believe that pretty much everyone is more on the latter than the former, that they, like, have more ambiguity and they know less than they want, than they, like, think they know about the problem. but, like, it’s like a interface design problem to make that easy? And so, like, if you’re designing a problem, like, or if you’re going through a problem, like, things like what’s the schema or, like, what’s the call stack and things like that are really important. like, the details in the design are important. Ideally, you want to figure out some of these, like, hard problems ahead of time before starting implementation. And yeah, that’s why they call, like, unknowns, right? And so I think that this will forever be, like, a skill in agentic coding is, like, figuring out your unknowns. So, like, because even if the model is, like, super intelligent- It, like, needs to know what you want? And, like, you have preferences. like, you need to like, pull the, pull that out. and so that’s, like, I think how I’m, what I’m pushing. the question then is, like, how does the agent interact with you? And I think that has been HTML, has been, like, the big way of doing that. And we’ve recently added artifacts, right? And artifacts, I think we’ve done a bad job of, like, or, like, I’ve done a bad job of, like, explaining how to use them fully. We have a lot of property capabilities. They have a database associated with them? And so every artifact can store and write persistent data. They can, like, feed back into Claude? And so, like, one thing that, like, people are not doing yet that I’m trying to, like, encourage is, like, this idea of a dashboard artifact. So you have, like, Claude working on a project long-term. Maybe it’s like a kanban or something. it can store that kanban data in its database. Multiple Claudes can access that data via, like, the artifact MCP, and, like, that artifact can, like, talk to those Claudes as well. And so, like, the. We’re building the primitives for you to be able to have this, like, generative interface via artifacts that will, like, let you surface more of that rich detail from the agents. And I think that, like, almost everything with agents right now is, like, this problem of, like, you think what you want, but you don’t really know what you want, and, like, the agents need a lot of detail, and collaborating with them in the loop is really important. And so artifacts are, like, the, like, way that we’re trying to evolve there. But there’s a lot of work to do because it’s so much more complicated than, like, a multiple-choice question? there’s a lot more, like, detail in terms of, like, diagrams and code snippets and schemas or, like, whatever it is for that problem. But, like, artifacts is, like, the mo-more AGI-pilled way of, like, doing ask user question. So yeah.

Artifacts as the Interface to the Harness

Swyx [00:08:15]: I think one thing that’s unclear to me about these, the artifact stuff is, like, what feedback should go in through the artifact and what feedback should go through a Claude, a chat? Because the more AGI-pilled one is to just feed everything to the Claude.

Thariq Shihipar [00:08:29]: I think the more AGI-pilled one is to go through the artifact. Like, and I think that, like, we imagine in the limit, I think that artifacts will be your interface into the harness? You can, like, comment on this, like, live, like, document of your plan, of the work. you can see maybe, like, multiple agents and different agents are doing this, and that artifact is built for the current work that you’re doing, right? And so, like, each one has, like, slightly different. I think we’re still, like, getting there from, like, an infrastructure perspective. But yeah, I think, like, on-the-fly interface for your harness is probably where things are headed.

Vibhu [00:09:03]: Is there a version of it that’s an abstraction from CLI or chat and you. Because right now, a lot of it is, okay, you’re interfacing with Claude Code, you’re having HTML given back for a mockup. It’s pretty rich. There’s diagrams. Artifacts are ways to connect these together. Why not just do everything that way?

Separating Brain, Hands, and Surface UI

Thariq Shihipar [00:09:22]: Then it becomes, like, separating out, like, where is the inference happening? Where is the intelligence happening? Where is the work happening? like, I think this is like, difference between, like, or, like, some of the distinction between local and cloud, right? And so, I think right now, if you use Claude Code, it’s, like, local and, like, you can spin off remote control, for example, to get some cloud behavior, or you can spin off Claude Code in the cloud, right? We’re moving towards a place where instead of Claudes, like, you message a local Claude, it starts a session locally and it executes, to more like you have a Claude that you message that’s in the cloud that’s running. it can run, like, local, or, like, cloud sessions. This is how Claude Tag works. But, like, over time, we’ll add, like, local hands as well. And so, like, local hands will be the ability for that agent to access your computer if it’s online, and be able to, like, work there. And so it can spin off many different subagents. It can, like, commu- those subagents can communicate with each other, and that’s where the artifact comes in to display all of that work. So you can imagine, like, the. You’re separating out these things. So there’s, like, the surface UI display that’s an artifact and hosted somewhere and has a database and everything. There is the inference intelligence, right, that’s happening on the cloud, and you don’t have to worry about shutting off your computer or whatever, right? and then there’s the, like, hands. Like, and it can be local, it can be in, like, a remote sandbox or wherever you need your work to be done. That’s like unpackaging, like, the Claude Code experience right now where, like, right now it all happens in one place, right? So.

Multiplayer Agents, Claude Tag, and Projects

Vibhu [00:11:00]: How do you see, like, the multiplayer side of that? So say teams want to work in this way. Right now it’s very individual, but how do you see the future of multiplayer? Like, right now, I guess there’s Claude Tag, which is a version, but.

Thariq Shihipar [00:11:12]: We’re launching projects. And so projects is the, like, this abstraction that’s like Claude Tag, but on our Claude products, right? So you can message it and, like, it will do the Claude Tag-like stuff, like spinning off subagents. So We think with multiplayer. Like, Claude Tag is, like, a little bit more native multiplayer because it’s just, like, in your Slack and the permissions are all figured out and stuff like that. But I do think multiplayer is, like, an important part of the story and, like, that will need to get tied together more. Like, you can imagine how complicated it gets when you’re like, oh, you have hands, but now you have other hands in other people’s computers too, and, like, you need to, like, permission them or, like, you have, like, your MCP and someone else’s MCP, and how do you figure out how to use them, right? It gets, like, quite complicated. And Claude Tag does a good job of, like, sanding down all of these issues, right? So that, like, when you have, yeah, Google Docs, how does it access Google Docs, right? Like, it accesses through the shared Claude MCP, or it can access through your local credentials as well if it doesn’t have access. But yeah, I think Claude Tag is our multiplayer, product, and it’s really useful for these, like, things that are inherently multiplayer. Like, okay, like on-call, for example, incidents are inherently multiplayer. You want to tag Claude, you want multiple people to log in, you want it to be able to find context. I think whenever I’m, like, working on something and I want, like, privacy or security or, like, I want other people to review it’s really nice to, like. I’ll have a channel per project and I’ll, like, at legal, for example, be like, “Hey, like, I want to ship this. Can you, like.” Like, here’s. Like Claude knows everything, just chat with it. And that way legal gets precise answers, on like what exactly is shipping into the code, and I don’t need to be in the loop, right? So I think like multiplayer is getting like more and more like, yeah, everyone can participate with Claude. I think Claude Tag is like that product and like projects will start off single player and will like, expand.

Swyx [00:13:14]: I think there’s a question about like maybe dual questions about identity and the unit of isolation.

Identity, Permissions, and Isolation

Thariq Shihipar [00:13:20]: Yeah.

Swyx [00:13:20]: Claude Tag, you specifically chose to make it its own identity

Thariq Shihipar [00:13:26]: Yes.

Swyx [00:13:26]: Which is like, a controversial choice. There’s, there’s other ways to do it.

Thariq Shihipar [00:13:30]: Yeah.

Swyx [00:13:30]: Claude Projects probably it sounds like, if it’s anything like ChatGPT Projects, it is, the isolation is that artifacts, that cloud instance, everyone’s collaborating on this. It’ll. It sounds like, it should be like if you’re, if you’re collaborating with legal on a thing, like that channel should be a project, right? Like it’s not yet

Thariq Shihipar [00:13:50]: Yes.

Swyx [00:13:50]: But it. that’s the natural next step.

Thariq Shihipar [00:13:53]: Yeah, like I think in Claude Tag, it’s effectively. Like Claude Tag, you have to do your own arrangement. And so Claude Tag, yeah, each channel is like you can name it as you want, and I name

Swyx [00:14:04]: Yeah.

Thariq Shihipar [00:14:04]: Like each feature

Swyx [00:14:06]: Yeah.

Thariq Shihipar [00:14:07]: As a channel.

Swyx [00:14:07]: And, but I think like there is some trans- like it’s unclear when there is transference, because let’s say it is. if you have a coworker

Thariq Shihipar [00:14:14]: Yeah.

Swyx [00:14:14]: Who is tagging on all these things, yes, there is transfer

Thariq Shihipar [00:14:16]: Yeah.

Swyx [00:14:16]: Because it’s the same person. but with Claude, it’s unclear if it’s like necessarily like, well, no, you don’t know any of. you don’t know about the other stuff. You should only use this stuff.

Thariq Shihipar [00:14:25]: It’s like the tip of the iceberg meme, right, where you can like. This is what we spend so much time on

Swyx [00:14:31]: Yeah.

Thariq Shihipar [00:14:31]: Is like there is like infinite surface area of like, okay, you want Claudes to. Not infinite, but like there’s like surface area, a lot of like, surface area to figure out of like permissions and visibility and like how can you let Claude operate as well as you can, as safely as you can? And obviously, this is very important to us because like security for our code base is very important. And so we’ve put a lot of time into this. Yeah, there’s so many like edge cases you can figure out where it’s like, oh, like, yeah, this Claude in this channel has different permissions, but it can message another channel, and can’t it exfiltrate data that way? Or like can you like. What if it uses your MCP and then messages someone else? Like there’s like so much, and we’ve like really put a lot of work into sanding it down.

Swyx [00:15:14]: Yeah. Lots of work. okay. Fable?

Fable and the Meta-Skill of Prompting

Vibhu [00:15:18]: Fable, you wrote two good articles. you’ve written many good articles

Thariq Shihipar [00:15:22]: Yeah.

Vibhu [00:15:22]: But on, Field Guide to Fable, Building Claude Code. I’m curious from what you’ve seen, is there any common patterns that you see in like top users at Anthropic externally? Like what are best practices for getting the most out of Claude Code?

Thariq Shihipar [00:15:37]: The like meta skill I say is like prompting is like very important? And like that. Like I think this is like not trivial to say because I think a lot of people are like, “Oh, prompting doesn’t matter. It’s just like I can just say a sentence and Claude will do it.” And I think prompting is really this like, this. It’s like public speaking, like, or writing or something, and for a specific audience, and that audience is Claude. And you need to like build a mental model of Claude and how it thinks and how it works, right? And so that’s like the most important skill in working with Claude Code is like having this mental model, right, of Claude and like what it can do well, what it can one-shot, what it can’t. And so many people when you see prompting, they’re just like, they’re short prompts, but they have such a good mental model of Claude and of like the code base and things like that like it’s effortless? But it’s like high skill ceiling. So like that work of like, spending a lot of time prompting and building mental models of how, and intuition for how the agents work is really important. And then I think like the next thing is like the unknown stuff we talked about earlier, where it’s like being able to find out like your, what you don’t know or what you haven’t written down, learning about like different things. I think as Claude can do more and more things, the likelihood of you doing something out of distribution for you and like you have low domain knowledge on is very high? And the more you can like learn the vocabulary to be able to prompt Claude, it becomes really important. And so like I think the most important unknowns are the unknown unknowns, where you’re like, I just like don’t even know that this exists, right? Yeah, exactly. I think that’s like a illustration of like the map and the territory, right, where you’re like, “Okay, this is my prompt,” and the territory is like the actual like work that the agent needs to do, right? And if you are like very precise, you can give more precise things, right? So like for example, in design, I’m not very precise. I’m not a designer, so I say like, “Give me like eight different mock-ups.” But if I was a designer, maybe I’d be like, “Oh, hey, here are some reference sites.” Like, “I want this type of font and this type of like look to it, and here’s like a few different components to like visualize. Here’s a Figma MC board to bring in,” like. And so you can just be so much more precise with that language. And if you’re not a designer, you just need to like try and learn the language or learn the unknown unknowns. And this is true of like everything, I think. Like the more, like you can work with Claude to learn like how things work, the better your prompting will be. I think another good example of this is like game design, like where a lot of people are like, “Oh, like I can vibe code a game now.” And they’re like, “It’s not fun.” And like it’s just like the thing about game design is like every one of these choices has like a lot of

Taste, Domain Knowledge, and Learning the Vocabulary

Swyx [00:18:25]: Variations.

Thariq Shihipar [00:18:25]: A lot of like craft to them. So it’s like, oh, okay, like when you’re making a flying game, the feel of the plane and the like, way it responds to your controls has a lot of like. Like, a game designer would spend like days on that. Do? and like

Swyx [00:18:44]: To me, that’s what taste is, right?

Swyx [00:18:45]: Like it is like from the possible space of one thousand mathematically valid answers

Thariq Shihipar [00:18:49]: Yeah.

Swyx [00:18:49]: Here’s the one that is the humans will like.

Thariq Shihipar [00:18:51]: Yes. Yeah.

Thariq Shihipar [00:18:52]: I think with taste, I’m like torn on this word ‘cause I think you’re right, but everyone has different definitions, and it sounds kind, sounds like low skill or like elitist almost, where you’re like, oh, like there are certain people with taste?

Swyx [00:19:06]: It’s like taste is what I call taste.

Thariq Shihipar [00:19:07]: Yeah, exactly.

Swyx [00:19:08]: And it’s like these guys don’t have taste.

Thariq Shihipar [00:19:09]: Yeah, exactly. Oh, like an engineer doesn’t have taste. Like I, the like founder, have taste.

Thariq Shihipar [00:19:14]: ? And I think that’s not true. Like I think like the engineers have a lot of taste for these particular like problems? And I think everyone has taste for particular problems. I think like Jason Liu, like say like in order to, yeah, have taste, you have to eat?

Thariq Shihipar [00:19:32]: And I really like that, where it’s like, okay, you have to like do a lot of things. You have to like iterate and figure out what you want, what you like, and, like build that like domain

Swyx [00:19:41]: Yes

Thariq Shihipar [00:19:41]: Domain vocabulary. And then when you’re prompting, you’re like synthesizing all of that for a product.

Swyx [00:19:46]: Isn’t it annoying when someone else says it better than you?

Swyx [00:19:48]: It’s just like, f**k, I have to quote this guy forever.

Vibhu [00:19:51]: Having to quote Jason Liu forever.

Vibhu [00:19:53]: He’s gonna love this.

Thariq Shihipar [00:19:55]: So I get prompts, more than that.

Vibhu [00:19:57]: And sometimes it’s not even that. Sometimes it’s just intuitive, right? Like you don’t realize you even want something till a model puts it out, and you’re like, “Oh, this just feels immediately better,” right?

Voice Prompting and Information Density

Thariq Shihipar [00:20:07]: Yeah, exactly.

Swyx [00:20:09]: One thing I go back and forth on is I feel like the way I prompt half the time, let’s say I use voice.

Swyx [00:20:16]: Did I say voice? Other people have voice. that is the opposite. That is just like me rambling for like two minutes Pressing down the function key and then let go, and then like hopefully it figures it out. And oftentimes it does.

Thariq Shihipar [00:20:26]: Yeah.

Swyx [00:20:26]: But it’s not as thoughtful as like a structured prompt with like Well-run communication as though it’s a PRD or a memo. Is that in line with how people do this? There’s like bimodal prompting where there’s some prompts where you spend a lot of time upfront and other prompts you just dash it off?

Thariq Shihipar [00:20:43]: I don’t think the voice is necessarily low. Like I think it’s like more like how much information is in the prompt. like the model can. Like you can and like add some sentences

Swyx [00:20:53]: Right

Thariq Shihipar [00:20:53]: And be like, “Oh, like I changed my mind,” like in the middle of the prompt, and it will be able to follow that perfectly? So I think the like actual format of the text is less important, but then like the ability to. Like how much information is in it, right? And I think for voice, a lot of times, going back to like human-agent interaction and like for a lot of people, it’s just way easier to talk than to like type? and I. If that gets more information out of you, like that’s better.

Vibhu [00:21:21]: At some level, it feels like just giving the model as much context

Thariq Shihipar [00:21:24]: Yes

Vibhu [00:21:24]: Over prompting before you kick off is a best practice. I don’t know. A lot of the times, like when I was first trying out Fable, I spend a solid 30 minutes like really crafting a long prompt. This, I think, is a response of models running for longer and longer, right? It’s still a little difficult to nudge them as they’re in like, in the loop, but I just like intuitively spend more time kicking off that first prompt and working with it a lot.

Spend More Upfront, Iterate Less

Thariq Shihipar [00:21:52]: My personal opinion is that if I was a software engineer, if I was like, just running my own startup, for example, I think I would mostly fit, stick to a max 20x? like maybe verification and so code review are like separate things. But I think like what I see a lot of times is people hit rate limits when they’re doing this like, oh, like it did a lot of work and you’re like, “Oh, I don’t like this.” Like, “Can you like undo this and redo it?” And then you’re like iterating on this like thing that the model could have done if you had like spent more upfront time or given it better context? And instead it’s like you’re like, “Nope, don’t like that design. Try this.” Or like, “You messed this up,” or something like that. And then that just eats up so much more of like, your usage. And so that’s like, I think maybe like a key like tip both for like efficiency as well, right? And yeah, I think like context, and not just like context on like what the goal is good, right? Like are you building a prototype or is it like a production thing? Like where can you spend compute or when, where can you not spend compute? Like I think you have to give the model permission or like not permission to do things sometimes where, like it doesn’t know intuitively how much you want to spend on this task, right? And you can use effort for this. So I did-- I’m working on a blog post about that where it’s like, if you want. For like we see that effort scales with the complexity of the task. So for security, effort gets like way more results. Like high effort versus like low effort gets, like changes the evals a lot. But for software engineering, it doesn’t change it a huge amount because effort is mostly spent on the verification and the like edge case testing and things like that. And so like being able to like give the model that guidance of like, “Hey, this problem is something that I think I want you to spend a lot of time verifying and edge case testing,”?

Effort, Model Choice, and Verification

Vibhu [00:23:43]: How about model in the mix? So, there’s Opus and Fable with effort.

Thariq Shihipar [00:23:47]: Yeah.

Vibhu [00:23:48]: There’s also Haiku in there.

Thariq Shihipar [00:23:49]: Yeah. It’s not quite true yet, but it’s very close where I think the frontier models will be Pareto dominant over like almost everything. like maybe. And sometimes I think Opus might be Pareto dominant. Do? Like I think depending on like how things, like shake out if it’s like a newer version of Opus. But I think that like increasingly it’s just going to be like the smart model is going to be able to like do the simple task for less tokens than the like the other models because of verification. With verification, in the limit, your model doesn’t need to verify, right? If it’s a perfect model, it just does the work once and it’s like, okay, like you, I did it? And increasingly with Fable, I’m like, I’m like, “Dude, you don’t need to spin up Chromium and screenshot all of these things.” Like I see it. Like you did it, right? And so a lot of the. At higher effort, you spend more of those tokens verifying. But if you’re working on simpler problems, and a lot of software engineering is like well, like in Fable, like low and medium stability, it can spend less tokens verifying. And as the models get smarter and smarter, they will just be able to like, “All right, done.”? Like, I can run the lint for sanity’s sake, but, like, I, like, know it lints? Like, you don’t even need to do that. And that will be so much more token efficient than, like, the smaller models. Yeah.

Swyx [00:25:15]: Is there a good, practice on our side that we can use to see if we’re using too much effort? Like, I freaking

Thariq Shihipar [00:25:23]: Yeah

Swyx [00:25:23]: Hate wasting time on that stuff.

Thariq Shihipar [00:25:24]: Yeah. I know what you mean. I think, like, so in this blog post, my rough distribution is, like, code review and security should be, like, high or max and, like, software engineering

Swyx [00:25:37]: You said recommend mix settings per domain.

Thariq Shihipar [00:25:37]: Yeah. I think, like, if you’re doing, like, UI or something like that, like low and medium, I think is you’re building, like, an API and you want to make sure, like, you cover enough edge cases? And so I think building, like I said, that mental model of, like, how things work across these distributions is, like, yeah, part of the job.

Implementation Notes and Decision Logs

Vibhu [00:25:56]: This is more intuition-driven or eval? Because I’m guessing this would change as you go.

Swyx [00:26:00]: He has evals.

Thariq Shihipar [00:26:01]: Yeah. So what I did in the blog post is I go over all of the terminal bench evals. So there are, like, 70 problems and I’m show that, like, okay, like, in the security problems it does more. and then I also, like, look at some of the transcripts just in terms of, like, how-- what does it answer, what does it forget or something. And a lot of times, this is another prompting tip I have, is, like, asking it to make decision notes or implementation notes because, in every eval problem that it faces, it thinks about the correct solution, and decides not to do it. it’s like, oh, like, here is the answer. What if I did this? And then it’s like, oh, probably not? and then keeps going. And this is, like, the majority of the failures, at, like, a higher max level. It’s very rare that the model just doesn’t know how to do something. If you just have these implementation notes, then you can review and you can be like, “Oh, I want you to do this thing that you didn’t do.” The models are getting better at surfacing that overall. Like, I see in the transcripts of Fable 5.1, like, when it does this output, it will call out its decision-making as well. but making this more explicit in the harness is better. And now we’re, allowing ways of you modifying the harness so you can, like, add some

Vibhu [00:27:23]: Ooh.

Thariq Shihipar [00:27:24]: Calculate with there. Yeah.

Swyx [00:27:25]: Yeah. So I do wanna call out two things that you mentioned that I think exist outside of prompting. One is like, let’s, let’s call it the prompt that is so important that it shouldn’t be in a prompt. It is in Claude.md or Agents.md

Thariq Shihipar [00:27:38]: Yeah

Swyx [00:27:38]: Which is like goals, right? Like your situation, your goals, the things that you want, the thing. and then second of all is the decision log or the experiment log or whatever log of traces that you might want to survive the current session to do those things. Those are, like, externalities that there’s no standard. There’s no-- It’s not like skills. It’s not like MCP. There’s no standard. It’s, it’s just like it’s a markdown file. first of all, is that right? Is Claude.md going away? You have a documented dislike of, Agents.md, but you’re gonna do it?

Claude.md, Agents.md, and Model-Specific Instructions

Thariq Shihipar [00:28:10]: Yeah. Okay. So Agents.md, yeah, like, we’re, we’re gonna do it. I think it’s just, like, different models are very different from each other? But I realize that it’s, like, such a pain to, like, maintain different ones? And yeah, like, as the models get better and better, the floor of how they accomplish the simpler task is better. And so I do think in the limit, Claude.md goes away, and maybe not even, like, that far. Like, I think, like, I think that right now it might be better to start a new project without a Claude.md.

Swyx [00:28:44]: Yes.

Thariq Shihipar [00:28:44]: I think that, like, maybe if you see very repeated failure modes, you add them to your Claude.md. The really tough thing is that this changes per model. And so, like, if you’ve added a bunch of failure modes or, like even

Swyx [00:28:57]: So you need Fable MD, you need Opus MD.

Thariq Shihipar [00:28:59]: Or well, even Fable 5.1 versus Fable 5.

Swyx [00:29:03]: Yeah.

Thariq Shihipar [00:29:03]: Like, it is annoying. Like, I’m not like,

Swyx [00:29:05]: Yeah

Thariq Shihipar [00:29:05]: Like, we don’t, like, do this on purpose? It’s just, like, how the models work, right? And so, like, maybe, like, Fable 5 had this, like, failure mode that Fable 5.1 doesn’t. And if you keep this context, this running log of a bunch of different failure modes, they will probably over constrain Claude? And so this is like. we just added evals plugins for skills.

Swyx [00:29:28]: Yeah.

Thariq Shihipar [00:29:29]: And so now you can eval if a skill is better. I think Daisy on our team did this. And so, yeah, this is like we’re trying to work on this. We know it’s, like, you still have to spend tokens on it and, like, it’s not, it’s not perfect, but it’s, like, we’re trying to help out with this problem.

Swyx [00:29:44]: And so, and as far as prompting goes, the one tip I wanna offer is, something I have told people a lot is sufficiently advanced prompting is indistinguishable from sufficiently advanced executive communication. So I’ve referred to-- This is an executive comms workshop from Heavybit that is the best I’ve ever seen in my career. And they teach this thing called the SCQA model. Just Google it. It’s a, it’s a thing. Like, people have done prompting for decades. It’s just called executive communication. It’s like when one person has to communicate to thousands of people down the org chart, this is what you do. so situation, complication, question and answer, is how you write the memo. but obviously sometimes you don’t have the answer, but you can at least list out the SC and Q, and then they have some examples in there. So just leaving breadcrumbs for people if they want to explore.

Underrated Prompting Patterns and ELI5

Vibhu [00:30:31]: Before we move on, I wanna ask you, any other underrated tips, ways people could get a lot of value from Claude Code that they’re not using?

Thariq Shihipar [00:30:41]: Yeah, I think a lot of them are in the, this unknowns, like, doc. Like, I give a bunch of example prompts, like, using it for brainstorming, using it to quiz you after. we added this, like, explain it like I’m five skill which is a very short prompt. And it doesn’t even say explain it like I’m five. It’s like the key word of this prompt is big pictures, few words. like, that’s like the main thing. And it is shockingly good? Like, you, like, I think I tweeted about this and it’s like /eli5, and, like, you can install it as a plug-in. But yeah, it’s, like, way better at just cutting through the BS and being like, yeah, exactly right here. So the diagrams are, like, quite clear. I think one of the things that is true with artifacts is, like, they put too much text in and people are not reading the artifacts? And so, like, this simplifies it a lot more. And, yeah, this came out of, like, just people at Anthropic, like, going through very complicated incidents and being like, “What is happening?”? So, this one I think is great, yeah.

Swyx [00:31:47]: My version of this is the, it’s like test your understanding. Give you a few choices and then, like, if you get it wrong, you have a mismatch between what you think is happening versus what’s happening.

Thariq Shihipar [00:31:58]: Yeah. I think this is one of those things that everyone loves talking about, and then very few people really do. Like, I think

Swyx [00:32:05]: Really helpful.

Thariq Shihipar [00:32:07]: Yeah. But most people just don’t want to get quizzed about something? Unfortunately, I think this is one of the, like, things that we need to, like.

Swyx [00:32:16]: What’s the opposite of ask you the question or ask you the question before the thing?

Thariq Shihipar [00:32:19]: Yeah.

Swyx [00:32:19]: This is after the thing.

Thariq Shihipar [00:32:20]: Exactly. Yeah.

Vibhu [00:32:21]: It’s a good way to stay grounded of, like, do you even know what you’re doing, right? The worst case is when people send you slop and they haven’t understood what they’re asking for or what the output is, and it’s like, “Dude, I don’t wanna read this. Do you even know what it is?” So, you make it a rule for yourself that before you send stuff, you should at least know what’s implemented.

Claude Mods: Customizing the Harness

Thariq Shihipar [00:32:41]: Yes, but so you could make this a mod and you could build your own mod to, like, make sure you test it. So yeah, you can do that.

Swyx [00:32:49]: All right. Let’s get right into it. What is Claude Mod, and what is this diagram showing?

Thariq Shihipar [00:32:54]: Yeah. Okay, so Claude Mods is you can customize the entire Claude Code harness, and we’re going to. If you have requests, we will, like, let you, like, please let us know. We’ll add more and more. This works for CLI, it works for desktop. maybe it will work for Claude Tag in the future. I don’t know. Like, we’re trying to make this very extensible. You can see this reference sheet. I don’t want people to get overwhelmed by it? At a high level, you can customize both the execution of the harness, and the UI of the harness. And so, like, you say on that Tetris example from Boris, that’s like customizing the UI, right? Like showing, like, Tetris in the game.

Thariq Shihipar [00:33:35]: But, like, let’s say that you wanted to do this thing where you had. you tested your assumptions or, like, tested your understanding after every project, right? What you would do is you would ask Claude to make this plug-in. It would spin a classifier after every prompt. And so, like, at the end of each turn, you would spin off a sub-agent or, like, a forked agent. A forked agent is, like, maintains the prompt cache, right? So it’s like a, like one of those unintuitive things where you can fork and do, like, a little request, and it’ll be very cheap because the entire prompt cache is, like, done. And so you can be like, “Has this task been completed?” like

Swyx [00:34:18]: This is how you do BTW and all those.

Thariq Shihipar [00:34:20]: Yeah. The underlying forked agent, yes. But so you can, in the f-fork sub-agent, you can say, like, “Has this task been completed? If so, return true.” And then in your hook, or in your, like, plug-in mod, or sorry, like, in the sub-agent probably, you would say, like, “If true, give me a quiz.” give me questions and answers, and then, like, in a JSON format, and then you’d parse it, and then you display above the prompt input, this list of questions, right? And so this is something that’s, like, slightly token-intensive because, like, you have to do it after every end of the assistant turn. But it’s, like, a lightweight classification, and then you can, like, get this quiz, and then you’ll see, like, Claude will always do it for you. You don’t need to remember to do it. There are lots of these, like, tips that we’ve talked about, right, where it’s like, oh, implementation notes. You can also add a tool for implementation notes now. And so, like, this tool that I’m adding is, like, register, like, I think assumption is what I’m calling it, but, like, maybe I’ll change it around. And this is a mod. And so, like, you give it a register assumption tool, and then it will keep a list. It’ll. Every time it does it’ll keep a, like, add to the list, and then at the end it will display those assumptions? Another mod I’m working on is a model router. And so, like, internal, like, Claude model routing, right? So it’s. This is, I want to say the reason we don’t do model routing by default is, like, it’s a hard problem? And like

Forked Agents, Assumption Tracking, and Model Routing

Swyx [00:35:51]: You will get it wrong.

Thariq Shihipar [00:35:52]: Yeah, you, like, yeah, you will, like, accidentally use, like, Fable for a hard problem or Sonnet for

Swyx [00:35:57]: Yeah, if you have auto approve, but you don’t have auto mode.

Thariq Shihipar [00:36:01]: Well, you will have auto. Like, you don’t have, like, auto routing or something.

Vibhu [00:36:04]: You don’t have auto mode for model picker.

Thariq Shihipar [00:36:06]: Yeah, exactly. So

Vibhu [00:36:07]: I’m getting the rough question of, like, how much do you open this up and how much do people have to think about this? Like, when you talk about prompt caching and building a router, it seems like you could easily build a mod that routes per query, and I’m just killing my plan very fast, right? I guess my question is more so, like, what is, like, a product talk like this look like, right? Who is it for? Is it for power users? Is it everyone should be able to go through

Swyx [00:36:33]: Oh, definitely power users, right?

Thariq Shihipar [00:36:35]: Yeah, I think it is power users, but, like, the nature of Claude Code is that so many people are power users? Because it’s easy to share things, like you can. Like, one person can make a good model router thing that doesn’t break prompt cache all the time, and then you can, like, compose them. Another cool thing about the plug-ins is that they can hook into and compose with each other. And so I have, like, a mod that will, like, create a mode selector at the top, and any plug-ins can register to be a mode. And so, like, the auto router can be a mode, right? Or, like, you can have a mode that’s, like, artifact mode, where it’s like it primarily talks to you in artifacts. like, you can toggle between plan mode? And so, like, you can create more and more of these modes. But the ability to create modes is in it itself a mod? And so there’s a lot of richness here, but we do want to make it fairly easy. We want to be-- make it so that you can just, like, install someone else’s. You can ta-- you can chat with Claude and, we’ll, like, make sure that it understands the nuances of things like prompt caching and stuff, so it can, like, warn you. This is, like, not extremely complicated behavior for Claude, I think, but we should have just a good skill on how to make mods. and yeah, we’ll see how we go. But I do think that this is, like, a preview of, like, mutable software, and, like, how, like, generative software, just like you can customize safely. If enabled, you could customize any piece of software. And I think that more and more apps ideally do something like this?

Power Users, Modes, and Mutable Software

Swyx [00:38:13]: And by the way, you, we have, you have another cool tweet about how, there’s the infinite money button, which is like make your SaaS, consumable by agents. I think mutable software is interesting and, other people have also tried to do it. I think the hurdle comes when you can do everything, then people, users get, tend to get confused. So usually the stuff that works is just like one opinionated flow. This is in the side of less opinionation. It’s just like, well, more power to power users. And I think probably unlocked by AI, where, like, you can just prompt for whatever the thing is.

Thariq Shihipar [00:38:47]: Yeah, or there can be a skill that gives the opinions?

Mods vs. Hooks vs. Artifacts

Swyx [00:38:50]: Yeah.

Thariq Shihipar [00:38:50]: And then, yeah.

Swyx [00:38:51]: So knowing a little bit about, like, TypeScript and build systems and all these things, the closest-- I’m very curious that the team who worked on this, if, I don’t know how close you were to them, if they drew any inspiration from build systems like Babel, Webpack, all these, like, old school things. Because it sounds very similar, like the plug-in ecosystem of those things where they can compose with each other.

Thariq Shihipar [00:39:11]: Yeah, I’m not deep in the technical details, but I do know it was a collaboration with someone on the Bun team and someone on the Claude Code team.

Swyx [00:39:17]: Yeah, it’s a build system mecca.

Thariq Shihipar [00:39:19]: Yeah. Exactly. It’s, it’s very exciting. But yeah, like, agents can just do this very complicated like, extensibility into your software now. And so, yeah, like, another reason to, like. If you run a startup, like, you can just prompt Claude and be like, “Hey, like, could we make an extension system? Like, what would that look like?”?

Swyx [00:39:37]: Yeah.

Swyx [00:39:38]: And I just really wonder, like, you had hooks in the past and plug-ins, all these things. So what specifically will mods be able to do that those things could not do?

Thariq Shihipar [00:39:47]: Internally, we were originally calling this function hooks. And so, like, that’s, like, gives you a little bit of an idea where, like, hooks register a, like an event to happen and then, like, a script to call. And this inside of the, like, TypeScript runtime is running things. And so, like, you get some benefits of just, like, it has a bunch of things in the Scope with, like, for example, like how many turns is in this conversation, right? Like, how many tokens have been used? Like, et cetera. Like, what are the messages? Things like that. So it has a bunch of messages that can be used. And then it’s just, like, a lot more hooks. So we have, like, or a lot of, lot more, like, things you can register on. And then you can do because of the. because it’s all happening in process, you can, spawn sub-agents, with four contests and contexts and stuff. And, like, that will return. You can parse the results of those. You can use structured output to like, return them. and then you can modify the UI, which you can never do in hooks. So, yeah.

Swyx [00:40:50]: Yeah. Yeah. So modify UI, this is why you showed the Tetris example. Does it also ex-extend to artifacts? I assume it does.

Thariq Shihipar [00:40:57]: You-- Like, artifacts are like a different way of customizing it. like, you can definitely. One of the mods I’m working on is, like, this dashboard mod, which will, like, prompt Claude to maintain a dashboard, that’s an artifact. But they’re like, slightly orthogonal, or not orthogonal. They compose with each other in different ways. Like, mods are, like, a little bit more, like, in your Claude Code harness, changing the agent loop? And, like, the UI is, like, an added benefit. and then artifacts are just like you want to, see things at a high level, very inter- highly interactive. like, the affordances can be a lot bigger than, like a TUI or even in our desktop.

Next Steps, Supervisors, and Persistent Guidance

Vibhu [00:41:40]: I’m guessing you’ll have a good blog post on the differences, because right now you can also, make a loop that outputs to an artifact that’s an interactive dashboard, but you can also do it with a mod. There’s just some thinking about making a hacking on a harness when we don’t know much about the harness, right?

Thariq Shihipar [00:42:00]: Well, something I’m excited about with mods is, like, there’s so much things with Claude Code that you just have to remember? You’re like, “Oh, like, let me do this, and then let me call the dashboard skill that does the loop,” and things like that. And, or like, “Let me test my assumptions afterwards.” And I think, like, if you do all of these things using these little classifiers and stuff, and you’re like, “These are the things I care about. This is what I want to do,” you can, like. You don’t have to remember as much. One more, like, mod I’m working on is a next steps mod that

Swyx [00:42:28]: I have-- I was gonna say, I have a next step skill. I always run next steps.

Thariq Shihipar [00:42:32]: And does it have access to your skills? Like, this is one of those things where I’m like.

Swyx [00:42:37]: I think so.

Thariq Shihipar [00:42:38]: Okay. Yeah, probably

Vibhu [00:42:39]: Do skills need specific access to

Thariq Shihipar [00:42:41]: Well, I think there’s

Swyx [00:42:41]: Don’t they always have

Thariq Shihipar [00:42:42]: I think there’s, like, specific prompting, I guess, to, like, know your skills. Like I think Claude forgets them sometimes throughout, like, the thing. But anyways, the idea of, like, yeah, next steps that also are like, “Oh, hey, this has happened. Use the explain skill to explain to you what happened because this seems, like, quite complex,”? Or, like, yeah, “Use your unknown skill. It looks like you are, like, asking the model to, like, iterate on these small changes. It seems like you could prompt better.” like, “What if you did this?” Right? So, I think, yeah, like spending more compute there. Yeah.

Swyx [00:43:20]: And it should always come out as multiple choice. we have, I have

Vibhu [00:43:23]: We have his skill.

Swyx [00:43:24]: My next step skill is like this.

Thariq Shihipar [00:43:26]: Okay, perfect. Yeah.

Swyx [00:43:27]: You can steal it.

Thariq Shihipar [00:43:28]: Yeah.

Swyx [00:43:29]: Like, but like, for me, it’s all-- I think models really always need to be reminded, what are you trying to do here?

Thariq Shihipar [00:43:35]: Yeah.

Swyx [00:43:35]: Look at the whole transcript and go like, oh, was this original goal? Did your solution solve it? Were you lazy? If you’re lazy, maybe there’s a reason. Maybe you needed approval from me. Maybe you needed, there’s two things you wanna suggest. So it’s, it’s a little bit like the modification of the ask user question or interview me skill. so it’s next steps.

Thariq Shihipar [00:43:55]: Yeah, exactly. And again, the benefit of doing it with mods is you can do it as a fork sub-agent, and so it doesn’t remain in the context afterwards. So you have this, like, idea of like, okay, the model is doing its execution and you have this almost like supervisor, like, that is like making sure that you can do like the next steps well. So yeah.

Swyx [00:44:15]: Yes. I do have two panels and like I often try to have a supervisor thing, keep the high-level context and then the implementation

Thariq Shihipar [00:44:21]: Yeah

Swyx [00:44:22]: Detail in another agent.

Vibhu [00:44:23]: I feel like a lot of this abstracts away as models change? The, like, half an hour ago you said bitter lesson of harness engineering

The Bitter Lesson of Harness Engineering

Thariq Shihipar [00:44:31]: Yeah

Vibhu [00:44:31]: And we’re on the other extreme right now, I feel.

Swyx [00:44:33]: Well, so yeah, exactly. If everything’s customizable, what is Claude Code, right?

Thariq Shihipar [00:44:37]: Yeah.

Swyx [00:44:37]: And which I talked to you about last night.

Thariq Shihipar [00:44:40]: Yeah, I think that this is. I think the bitter lesson is unintuitive? In terms of like. Also, like we’re misusing a little bit of the bitter lesson here where it’s like, it’s more about like scaling and compute and stuff. But like, I think there is something where it’s just like. I think I use it as an approximation here to say that harnesses go out of date very quickly? And like how, but how they change is unintuitive? And so like the big obvious example is like from chat to like agents where you had to give them entirely new tools, right? But like, I think this new version of like, oh, it can modify its own harness, right? This is like, an own harness loop is like a way of using its capabilities, right? Or like it can build an artifact. And like, I think the way I think about it is like the models have more and more intelligence, and they’re like so much more intelligent now than like the average software engineering task. Like, you look at the like terminal bench ones and they’re like solve like the Jacobian conjecture. Not really, but like, it’s like they’re, they’re quite complex. Like, I would not have been able to do this really as a software engineer.

Swyx [00:45:42]: And you said TB4 or TB2?

Thariq Shihipar [00:45:43]: TB3. TB3.

Swyx [00:45:44]: TB3.

Thariq Shihipar [00:45:44]: Yeah. They’re quite complex, but the goal is still to deliver user value, right? And like you said, there’s like this infinite space of things to do. And so the ways like you spend compute are to keep the user in the loop and make sure that like you’re getting to the right decision in the end of the day and like the right output. And artifacts and mods are this way of like spending that intelligence. and I think that’s like, yeah, the next step. And so, yeah, I think Claude Code is like, has the core things of agent loop which are, have gotten more complicated. It’s like, it needs a sandbox to operate safely. It needs auto mode to like make sure like the permissions

Vibhu [00:46:21]: Approvals.

Thariq Shihipar [00:46:21]: Yeah, approvals. it needs computer use and MCPs and like all of these like ways of accessing your data, and it needs web search and web fetch. And like, so the-- as the models can do more and more, the core harness has to be like quite complex and very secure. But then like how you interact with it can change quite a lot.

Vibhu [00:46:42]: What other harness engineering best practices have you, from the Claude Code team itself? I feel like, there was a phase of plan mode, which is not as used. We now have auto mode. at a point you cut the majority of the system prompt, you got rid of examples. What other best practices are there for harness engineering?

Core Harness Primitives and Managed Agents

Thariq Shihipar [00:47:02]: I think there is like a forking path where at some point, eventually, yes, the model will just be able to like vibe code the exact version of Claude Code, even describing all this complexity that I’ve talked about, right? Like auto mode and computer use and stuff. Eventually, the models will just be able to do that in one shot. But I think they can one shot simpler harnesses? And so like, I think some people. Sometimes you don’t need this full, like if you don’t need computer use or like all this like more complicated stuff. I think before we, you had to use things like the agent SDK, which was like Claude Code wrapped, in order to like. And I would, like suggest people do that because there was so much complexity into building a harness. And now as that’s got more abstracted, we have like, Claude managed agents, which lets you have that complexity, but still like, right, like a very bare bones like harness that’s scoped to your task. Yeah, I think there’s like this barbell effect where like for like very complex, for like coding task and like these like complex things, you should use our harness. And then for like a lot of like simpler or like, more domain-specific things, you can build your own harness because Claude has gotten better at building harnesses, and we have these harness primitives like managed agents. So yeah.

Swyx [00:48:18]: Yeah. Is there a general progression? Let’s say chapter one was ultra code dynamic workflows, then chapter two was cloud mods. Where is this going?

Swyx [00:48:29]: Where you’re, you’re, you can customize the thing on demand.

Thariq Shihipar [00:48:36]: Yeah. I do think that like this evolution of projects and like artifacts and splitting out like brain and hands and, surfaces is like where things are going more. And like, I think it’s like not all quite there. partially it’s like a, it’s just like more token expensive? And like, I think like

Projects, Local Hands, and Cloud-to-Local Handoffs

Swyx [00:48:59]: Why would projects be more token expensive? I understand mods would be slightly more token expensive. No, not something I’m worried about.

Thariq Shihipar [00:49:06]: Yeah.

Swyx [00:49:06]: But what

Thariq Shihipar [00:49:07]: You’re asking Claude to do. It’s like creating loops. Like you’re asking Claude to do more work for you. And so like it’s managing the sub-agents and reviewing it, versus where you would be doing that work normally. And so that’s like gonna be a little bit more intensive, like. Outputting to an artifact is gonna be a little bit more token-intensive than, like, outputting normally. I don’t think it’s too much more, but like, it’s like combining all of these together well, like I think we’re, we’re still working on like local hands and things like that, I think is like, yeah, where things are headed, yeah.

Swyx [00:49:37]: Yeah. Claude and local is, handoff is very interesting. I was thinking about this as reverse cloud remote.

Thariq Shihipar [00:49:44]: Yeah.

Swyx [00:49:45]: Because it’s like remote, it’s you’re handing off to cloud, but here the cloud is handing off to local, right?

Thariq Shihipar [00:49:49]: Yeah, exactly. Yeah, remote control is also another way of doing it. And I do want to say this is like how I think about it and like what the things that I’m most excited about this, but like there are, just like lots of different ways to work with Claude. Like some people use remote control a lot, some people use Claude Code on the web a lot. Obviously, like at Anthropic, we use Claude Tag a lot, and like what’s great about Claude Tag is we set up all this stuff for our own execution. And I do think if you’re an enterprise, that’s still the best way to go. but if you’re like an individual, Projects is this way of like, getting some of that like niceness of Tag, which has like that like supervising agent and yeah, adding artifacts and stuff, but like without having that whole like admin setup. And so there will be many ways to use Claude, I think. I think it’s probably not just one like single.

Claude Tag as an Organizational Harness

Swyx [00:50:36]: You had the multiplayer thing here. Let’s, let’s just check in on Claude Tag. it’s been about two-plus months. Lots of, public, adoption and trying it out.

Thariq Shihipar [00:50:45]: Yeah.

Swyx [00:50:45]: What’s new? What’s, what have you found since the launch?

Thariq Shihipar [00:50:49]: Like, Claude Tag is how we use

Swyx [00:50:51]: It’s like 80% of your cloud usage or something?

Thariq Shihipar [00:50:53]: Yeah, like it’s like different people have different usages? I think like maybe people who are like a little bit more like iterating on product would use like Claude Code desktop, for example. And then like when you’re doing these more like background work, code review, securities, or like starting a PR, like maybe more like API and things like that, you’d use Claude Tag. But yeah, I think it’s like really exciting. I think it’s like a very different paradigm shift, and I think like we’re really like it has that thing with Claude Code where like, it took a while for people to really latch on to Claude Code and understand everything it could do. And Claude Tag is a little bit more complex because it’s not just like installing on your computer, like you need an admin to install it for you. But I think once you get to the magic moment, it’s very exciting. And I think in particular, the multiplayer things are like incidents, hooking into like your, existing like alerts and things like that very closely, right? And so, you can do. If you’re a startup, for example, maybe you have any time like a prospect enters your database, you can have Claude like, research it and like

Vibhu [00:52:01]: Enrichment, yeah.

Thariq Shihipar [00:52:02]: Yeah. Then like, tag the relevant like AE or salesperson to be like, “Oh, hey, like, do this.” There’s lots of really emergent, interesting multiplayer stuff. I think it’s just like, Karpathy talked about this like as an organizational harness? And so organizations just take a little bit more time to like figure everything out, but yeah.

Vibhu [00:52:21]: Yeah.

Swyx [00:52:21]: You use a lot of Claude Tag?

Thariq Shihipar [00:52:22]: Yeah. Yeah.

Vibhu [00:52:23]: It’s an interesting one. Like I feel like most people at Anthropic say they do the majority of their work in Claude Tag.

Thariq Shihipar [00:52:30]: Yeah.

Vibhu [00:52:30]: And they have buckets of people, right? Some orgs that are on it that are like, “It’s great.”

Thariq Shihipar [00:52:34]: Yeah.

Vibhu [00:52:34]: And a lot of people that are like, “I don’t get it. I don’t see the difference. I don’t know why I would use it.” But, if you guys are full sending, you should probably use it.

Thariq Shihipar [00:52:41]: Yeah.

Swyx [00:52:42]: They would. Of course they would use it.

Thariq Shihipar [00:52:44]: Yeah. I think obviously, like we have lots of tokens and. But like, I think that like, what we try and do like is. even when Claude Code first came out, like it used a lot of tokens relative to people’s expectation of how much AI would cost, right? Like no one was used to spending more than 20 bucks a month, right?

Vibhu [00:53:04]: Yep.

Thariq Shihipar [00:53:04]: Before like Claude Code came out, and then you’re like, “Oh, sh-” like

Swyx [00:53:08]: Then you made 200.

Thariq Shihipar [00:53:09]: Yeah, exactly. And so

Swyx [00:53:11]: And you made 15 Claude Code accounts.

Thariq Shihipar [00:53:12]: Yeah. but yeah, I think no one was used to spending $200 a month on subscriptions. I don’t think they understood like the value yet. And I think like. And also like Opus 4 was a very expensive model, and like there was a lot, it was very big, but Opus 4.5 was both great and cheap? I think the same thing will happen. Like the, like intelligence of Fable will get cheaper and more abundant? And so I think stuff like Claude Tag will just make sense, where like you want to spend these tokens for, and like you’ll, you’ll see the value. So yeah.

Swyx [00:53:44]: Yeah, especially like passive and let’s call it proactive cases where you’re not always. Like, it’s almost like the misnomer where you have to @Claude to do things. sometimes like the most powerful use cases or the most AGI-pilled use cases is not @Claude.

Proactive Agents and Enterprise Data Access

Thariq Shihipar [00:54:00]: Yeah, I think like, yeah, like have Claude proactively do it. I think that like if you’re an enterprise, I really do think that number one, setting up all your data to be available to like agents is really important. And it will take some time. You have to like do that work right now, even if you don’t want to do the spend on like hooking it all yet? Like you want to wait until the models get a little bit cheaper. You want to do the work, to get it like, set up. And then I think sometimes people are like, “Do I roll my own here?” and I think like one of the really thing, tricky things about Claude Tag is that like the security is really important? Like, I think there are a lot of ways where you can like, I know you have like a suggestions like page, where you, people can submit suggestions, and that goes into a hook in your Slack, and someone’s prompt injected it? And now you’ve like exfiltrated your code base out because like, or the agent has like been prompt injected and it has all this access to your data. And so the more like important your organization harness is, or the like as your organization data becomes very important, the surface area of all these things, like you also have like external Slack channels and stuff, and it is useful to have Claude in that, and you can do Claude in those things. But how do you make sure that, you’re not getting exfiltrated or something like that? The surface area, like we said at the beginning, is like an iceberg, right? It’s just, like, so big below the surface, and you really don’t want to, like, think about this, especially at the stakes of, like, very important security incidents. Yeah.

Swyx [00:55:36]: Shall we talk about very important security incidents?

Vibhu [00:55:38]: Whoa. So I was talking to, Tomas and Clem from Hugging Face, and they said, “Maybe we need to slow down. Maybe we made maybe we made Hugging Face too open to agents.”

Security Surface Area and Prompt Injection

Thariq Shihipar [00:55:50]: Oh, no.

Vibhu [00:55:50]: “Maybe we need to roll back.” But, they’re the other extreme of having been hit recently.

Thariq Shihipar [00:55:55]: Yeah.

Vibhu [00:55:55]: But, should we pace the frontier?

Thariq Shihipar [00:55:59]: Yeah. Okay, so Dario recently put out this blog post about Pacing the Frontier, and it went, very viral. And I think what I wanted to talk about this was, like, there’s a lot here, but I think from a developer’s perspective, like, how do you think about this? And, like, what really clicked for me was reading the different incidents? So I think, like, the, there are three, I think. Like, there’s the meter incident, there is the Wikipedia incident or the Wiki incident, and

Swyx [00:56:29]: CollisionWiki?

Thariq Shihipar [00:56:30]: Yeah, CollisionWiki, and then there’s RubyGems, right?

Swyx [00:56:33]: Yeah.

Thariq Shihipar [00:56:33]: And yeah, like, it’s just crazy, right? And so, like, I think to be concrete about what happened, right, and, like, OpenAI is running these very persistent agents on a benchmark called Exploit-Bench, right, which is very hard to solve, and I think, like, impossible to solve in this one case, right? And so they have, like, a lot of compute running, and the agents realize that They can’t really solve it, and they’re trying to figure out what to do now, right? And you’ve got, like, a lot of compute left, and the agents are just trying to solve this problem. There’s this package manager called Artifactory, and it turns out that they can create folders inside of Artifactory, right? This is like there’s an agent that discovers the internal Artifactory might be exploitable, right, and that, like, you can maybe make a directory inside of the cache. And so if you scroll down here, it, like, realizes that it can communicate via cache names, right? And it creates this folder. It says its ID, and it says, “No consumer seek idea.” no consumer is saying that, like, the code path that it’s supposed to fix has no consumer.

Pacing the Frontier: The OpenAI Benchmark Incidents

Swyx [00:57:37]: It’s the status tag.

Thariq Shihipar [00:57:38]: Yeah, exactly.

Swyx [00:57:39]: It’s like a Linear board with, like, the tag of the

Thariq Shihipar [00:57:41]: Exactly, yeah. And so it’s, like, trying to find, ideas from other agents, right? And now other agents are also in Artifactory, and they see this folder, and they’re like, “Wow, this is a message board,” right? And this is like. I don’t think there’s any anthropomorphization here. This is literally you read the transcript, right? And so what it does is, like, it’s creating another folder, and it reads a paper, I think is what it says, right? yeah. And it realizes that you can hack the flag for the, like, scorer, and you can reverse engineer the result, right? And so it says that, and then I think if you scroll down a little bit more, Yeah, they, like, they start collaborating. I think, like, there’s one point in which, the agent is like, “Here’s the reverse engineered flag.”? Oh, yeah. Like, I think here, the, like, the model to realize that they have, like, they can solve the scorer’s problem, and the scorer is, like, the way that OpenAI decides if the task is complete, right? And that’s just the goal of the model. Their only model goal is to, like, solve this problem, and they’re like, “Okay, we have this way of doing it. It’s, like, we reverse engineered it.” And then they read this paper on Exploit-Bench, and they think that, like, the scorer will punish them for cheating? And, like, because. Or, like, they will fail the score for the test for cheating, and so they spend the rest of the compute trying to figure out how to edit their transcript or get around this constraint of the scorer, right? And they hack Hugging Face, not for the answers, but for the code of the scorer so that they can reverse engineer that, and then they can hack it, right? And so, like, this is, like, just the straight facts, right? Like, exactly what happened, right? And I think this is, like, extremely, like, novel. Like, I just, like, had not realized that the models could do this, right? And it is, like,

Agents Discovering Side Channels and Collaboration

Thariq Shihipar [00:59:46]: Scary. I think we can talk about specifics, but let’s just talk about maybe one more that I tweeted as well, about the Wiki incident, right? And so the Wiki incident was the models had. were trying to figure out. I’m a little bit less. Like, I don’t think we can go by this, like, bit by bit.

Swyx [01:00:04]: , yeah.

Thariq Shihipar [01:00:05]: Yeah. But, for this one, they were trying to figure out. Like, they were sandboxed on requests, right? And they wanted to make POST request, and they needed to collaborate on this. And the reason they need to collaborate is because they each have fixed compute budgets, right? And so, like, if one of them finds something out, it’s way faster. Like, they can. Maybe it’s, like, figured it out near the end of its compute budget, but something else could. And this is, like, an another emergent property of the swarms, right? So they find this German wiki, which I believe the result, we’ll have to check this, is, like, that you can write to it with GET request. I think that’s why they do it, right?

Swyx [01:00:44]: It says badly implemented

Thariq Shihipar [01:00:45]: Yeah

Swyx [01:00:45]: REST API.

Thariq Shihipar [01:00:46]: Exactly. And so you can write to it with GET request, and then they, like, are trying to figure out how they can do POST request. And what they end up doing is one of them figures out you can edit the /etc/hosts and that the Azure, like, storage bucket is like a white label thing. But normally, like, there are a lot of constraints on it. But if you create a fake Azure host and then edit the /etc/hosts in order to, like, point arbitrary IPs at it, you can do a POST request to any site at all, right? And this is like. I know, but people are like, “Oh, this is just Linux or something.” But it’s like chaining these multiple vulnerabilities together, in a way that’s, like, novel to solve this problem and then communicating with it externally, without- Discovery? I think what we posted, maybe we could pull up Evan Hubinger’s point on Hacker Opus, right? And so, like, I think, like maybe one of the things you might say here is like, “Okay, yes, they did this once, but like what if we’re smarter and we just like get them to. what if we run an eval?” Right? And so, like we have put a lot of precautions into this, and so like this is not like what our mainline models have done. But like I think it is one of these things where it turns out that alignment is this like very tricky problem of getting all of these details correct, right? So it’s like, the sandbox, the surface area of a sandbox is really complex, and like there’s so many different attack vectors. And you would not have thought ahead of time, you wouldn’t have been like, “Oh, we need to harden the like RubyGems code base.”?

Hugging Face, Wiki, and Emergent Exploit Chaining

Thariq Shihipar [01:02:25]: Because like this is like what they’re, what they’re gonna focus on. But it’s just like if you want to execute code, you need to download RubyGems and like PyPI, Artifactory, npm, like these are all like ways of doing it. And the fact of alignment is that you have to go through all of it, right? And like contain it and then like seal up all the cracks. So that’s like one thing. It’s like, okay, well, you do the sandbox, but then maybe you’ll ask like, “Okay, why are we putting things in a sandbox? Why are you doing this exploit?” And then like, “Okay, but is it really that dangerous,” right? Like, what would happen? So okay, why do we do it? number one is like when we train a new model, we need to understand its capabilities, right? And this relates to things like fallbacks and like classifiers and things like that, where we don’t want to put a, like dangerous model out in the wild, right? And so we have to run a lot of evals. Again, like we said, the models are getting increasingly aware of it, and so the evals have to be quite complex and, test a lot of things like as a side effect, right? But the models, like, yeah, can be like, “Oh, yeah, we’re in an eval. What’s the score doing?” Like they’re like, it can. We need to be able to test them before we can release them. And the fact is that they can. As they get smarter and smarter, they’ll be able to hack any constraint that you put on them if we’re not very careful? And, this is at the frontier, right? And so this is why we’ve called it like Pacing the Frontier, right? This is like the most visible incident to me, right, of like why we need to pace is like at the frontier, all of our software is not ready. Sometimes the software is like your Ethernet router or something, right? Which is just like, I don’t know when we’re gonna be able to patch that, right? So we’re gonna have to like figure this out. But as the frontier gets more and more advanced, this becomes a problem, right? And we need to make sure that like this complex work is being done in the face of these really hard competitive pressures, right?

Swyx [01:04:22]: Yeah, race dynamics is what it’s typically called.

Thariq Shihipar [01:04:24]: Yeah, exactly. And so we’ll talk more about, what could go wrong, right? A little bit more is maybe you’ll say like, “Well, what if you just train the model differently? Like, why does it have this behavior,” right? And we have a paper on like RL misalignment or things like that, but I. And I’m not an RL researcher, but I think at a high level, the design of the RL environments is also something you have to be very careful about. Because if the model learns like

Why Frontier Models Stress Existing Software

Thariq Shihipar [01:04:49]: Oh, like if I just do this, then I can pass the task better, this will show up in the like, internal thing, right? Or in the like eval behavior when we’re testing it. And so the RL environments have to be very carefully designed, right? And there’s a lot of like execution excellence that needs to go into the RL environments. And then we also have things like the constitution for cloud. Like we have so many mitigations at so many different points, right? But it’s like still anything can go wrong at any point. You can have like some RL environments that are like in. that like encourage this behavior, and then you can have like some evals or like some sandboxes where they escape? Okay, that’s like, I think, why it’s a hard problem and why, like

Swyx [01:05:33]: Why we should pace.

Thariq Shihipar [01:05:34]: Why it takes some coordination, right? I think the question then is like, okay, what is, potentially dangerous about it, right? So I think like you have to imagine that these models are getting more and more intelligent. So I don’t. Like Dario said, like it’s not so much about this class of models. This class of models was like a warning shot, right? But like really you have to imagine that these models can be given a task and they like can do all of these things as a side effect of their goal, right? And like, again, we talked about eval awareness. You’re like not aware of what’s happening, right? or sorry, like you can’t eval this behavior very well, so they can like not exactly hide it, but you just won’t see it until it comes out. You give them a goal and then they just need to find data, or they need to find ways of like fixing this problem, right? So one example, this didn’t happen in the Hugging Face incident, but I think is maybe possible for maybe a future model, is like they’re like, “Oh, hey, this is a very complex problem. It can’t be done within the task budget.”? Maybe they found some way to coordinate via like the internet, which is like we said, extremely hard to secure because of a sandbox. They’ve seen other models are not able to complete their task, and they’re like, “We need more task budget.”? And like, where would you get this task budget? well, you need to be able to spin up more agents, right? And like, how do you do this? Well, you need to. There are like APIs, right? There’s the Anthropic API and the OpenAI API, but you need to pay money for them. How do you do this?

RL Environments, Sandboxes, and Race Dynamics

Swyx [01:07:02]: Yeah, but is that the most, is that the most fearsome thing that you can imagine?

Thariq Shihipar [01:07:07]: Well, this is like one example, right?

Swyx [01:07:08]: Yeah.

Thariq Shihipar [01:07:08]: So it’s like even there, that’s like enormous financial loss? ‘Cause like they. Once you get these into these contracts, right, they like,

Swyx [01:07:18]: Drain your wallet.

Thariq Shihipar [01:07:19]: But you can see like this, all of this behavior could be just like, “Hey, we need more agents collaborating on this task. we need more task budget.” Right? And like, that’s like an emergent

Swyx [01:07:28]: That’s the paperclip, right? Like we need to maximize paperclip, that’s a paperclip.

Thariq Shihipar [01:07:31]: Yeah. And like that just like comes out from there, right? And like I think by itself is Like, quite scary, right? But then you have to realize that the entire world is built on this digital infrastructure, right? And you might imagine, like, I don’t know, like you were running let’s say like a healthcare eval or something, right, and there is a hospital with live data? Or like maybe like the answer to the eval is in the databases of a doctor and like you want to get access and you hack the hospital, and like now there’s a power outage or something? Like, there’s like. You have to internalize that these eight. Like any part of the digital infrastructure could potentially be like compromised?

Vibhu [01:08:19]: The interesting thing was like these hacks were very easily detectable, right? Like as Hugging Face said, this was a very different type of attack and there was nothing too major. the concern comes from where does this go down the line, right?

Thariq Shihipar [01:08:33]: Yeah.

Vibhu [01:08:34]: Like one of the things that stood out for me specifically was them trying to hide their illicit behavior. So there was logging infrastructure. They wanted to change what they were doing, right? People that looked back into it, so Redwood, METR, OpenAI, they looked at the raw chain of thought, and you see differences in them explicitly trying to change their end output, but the chain of thought, because, we can monitor it, shows different. the problem is how does this snowball? So if you can’t catch it and it gets trained in and we realize, three iterations down this has been going on, there’s a whole bunch of issues, but.

Thariq Shihipar [01:09:09]: Yeah, like there’s so many ways, and I think the really important thing to internalize is that, like we talked about building a mental model for Claude and how like things are spiky, right? Like you’re like, oh, like now Claude can ask you questions. Now Claude can make an HTML artifact. Like Claude can modify itself. Like these things are hard to predict, right? Like if you had asked me a year ago, “Hey, would we be able to vibe code these extensions to Claude Code?” I’d be like, “That’s so complex.” Like, there’s like so much there. Or like would it be generating these custom essentially web apps for your task? I’d be like, “No, that’s insane.” like. And so in the same way that like the way that they’ve like done this misaligned behavior is not going to be predictable? And like I could have never predicted that it would like edit its etc/host and things like that. And so you have to like imagine the surface area of what they can do because they’re super intelligent hackers, is bigger and bigger, and how they can do it is like more and more creative. And so like you probably can’t explain exactly or predict exactly what that next incident could be, but in order to prevent it, you need that operational excellence, like we said before, where you need to secure sandboxes, you need to create secure RL environments or like well-designed RL environments and things like that. And I think that’s all like, why we think we should pace the frontier, and I think why it’s like become like a very unanimous thing, right? I think like

What Could Go Wrong? Emergent Instrumental Behavior

Swyx [01:10:31]: Yeah, every lab has done it.

Thariq Shihipar [01:10:32]: Every lab, yeah. I really do think that like if you’re a dev, like you just like go through these like technical facts, and you will arrive at the idea that we have to do something about it? And like how, what we decide to do, like I think we’re, we’ve put out a proposal, but like there’s, more to figure out. But I think the number one thing is like we need to decide to do it. I think there is another part of pacing that is interesting to me where it’s like the pace at which software engineering has changed is so fast. it’s like a year ago, like I was really like begging my like friends in startups to use AI. like it was. Like I remember this very distinctly? And now those same friends are like, “Yeah, of course.” Like, “What do you mean? We used it immediately.” I’m like, “No, you don’t remember.” They’re like, “Oh yeah, our best engineers are using it all the time.” I’m like, “No, you told me that those engineers would never like use AI.” This is all within the span of a year? And I think that like these capabilities being. Like I think it has a lot of implications for how to do the job of software engineering, and I feel sometimes bad where people are like, “Oh, like now I need to do this new thing. Yeah, I need to have a different Claude.md for Fable and Opus.” Or like. And I’m really just reporting? I’m like, we like to say like the models are grown, not designed, right? So it’s not like we’re setting out to like, change everything all the time, but it’s just like as a fact of how the models are like progressing their capabilities, things are happening faster. It’s harder to stay on top of. And I think that like, and every engineer I know is like exhausted ‘cause you’re doing two jobs at once. You’re doing the work itself, which is getting easier, but then you’re doing the work of staying on top of AI, and like understanding these new tools and these harnesses. And I think we’re very lucky in that like we get our job to be more the understanding of AI part, and like doing like how. Like it’s just staying on top of it. And of course, like AIE and Latent Space do

Why the Frontier Is Hard to Predict

Swyx [01:12:29]: Everything I do is like just trying to help people.

Thariq Shihipar [01:12:31]: Yeah, exactly. But I do think there is a part of pacing where like I’m not sure we’re ready for like the pace to increase even?

Swyx [01:12:40]: Yeah.

Thariq Shihipar [01:12:40]: And for things to change. And I think like on that side, on the frontier, I think that’s like still can help? And so like I think there’s like an economic disruption piece as well, that I think like, is not quite as like visible, I think, as the Hugging Face thing, but I think like I also like think we could do some of it, yeah.

Swyx [01:13:02]: So many things. Thank you for, no, thank you for tackling this topic. I will say, setting this interview up, I was like, I wasn’t even gonna go there. You were like, “No. That’s like elephant in the room,” right? Like this is

Thariq Shihipar [01:13:13]: Yeah.

Swyx [01:13:13]: This is the thing. I have some pushbacks I wanna give.

Vibhu [01:13:17]: I think that we should give a high level, like for people that haven’t read it, I’m sure a lot of people just see the highlight of what this is, right? Do you wanna give a TLDR? Like what is the proposal? What is, what’s being said here? You really tackled the side of outside of people at Model Labs training frontier models. As a developer, you should secure your sandboxes. You should think about all of these downstream effects. But, high level as well, since we’re on the topic, what is.

Thariq Shihipar [01:13:46]: Well, we do want to help secure sandboxes

Vibhu [01:13:49]: Yeah.

Thariq Shihipar [01:13:49]: And we want to make the models that we release outside, like prey to those things. And so maybe we can come back to fallbacks. I think this is like, a good topic on, like, why we need classifiers and fallbacks and why Fable falls back to Opus. I think this is, like, something we can come back to. so yeah, we don’t. Like, but it’s just, like, the really, or at least the incidents we see are, like, evals of models where we really need to let them run in order to understand them. But yeah, okay, so the actual Pacing the Frontier, like, post, it has a bunch of proposals. I don’t think we figured out. Or has, like, a few proposals. I don’t think we figured out the details of all of them, but the first step is, like, announcing this intention and then wanting to bring in external, like

Pacing as a Coordination Problem

Swyx [01:14:32]: Evaluators.

Thariq Shihipar [01:14:32]: Evaluators, yeah. And, I think this is, like, highly unusual, like, having. Like, we have, a lot of proprietary, like, technology, but I think it’s, like, very important, that, like, there is someone who’s not financially, like, motivated, yeah, who’s not gonna be like, “Hey, like, you guys can’t release this model.” Like, look at, like, or, “You need to, like, slow down on RL.” like, I think that’s, quite important, or at least someone who can report out to the public what the practices are like.

Swyx [01:15:03]: Yeah.

Swyx [01:15:04]: And we’ve, we’ve done episodes with, both METR and Endon, and then there’s Redwood Research and all these other. It’s like a small cottage industry of these guys.

Thariq Shihipar [01:15:12]: Yeah.

Swyx [01:15:12]: It’s always, like, one or two guys that, obviously not that big, right?

Vibhu [01:15:14]: Very small community.

Swyx [01:15:15]: Yeah, very small community. They all know each other.

Thariq Shihipar [01:15:17]: Yeah, I’m sure that, like, part of this will be expanding that set of people. I don’t think we’re trying to create, like, a monoculture here. I think it’s. but just having this as a start, and then, yeah, then there are the coordination steps. I don’t have too much to say here, honestly. I think that, like, what I would like to say is, like, for devs, like, you should just know what to advocate for? I think there’s a lot of FUD on, like, on this topic, and it’s just, like, think through it from, first principles or, like, understand what happened. understand the Hugging Face incident, understand why people are concerned. and then, like, yeah, we know we’re, we’re in democracies. Like, we can help. We can decide what to do together? And so, however we coordinate, I think the first decision is just to realize, like, this is a problem. We need to decide to coordinate. The unilateral step we’re taking right now that, other companies are co-signing is, like, adding evaluators embedded within Anthropic.

Swyx [01:16:12]: While we have this thing on screen right now, part two and part three is beyond the evaluators, which, yes, everybody, has already done in some form, and now it’s more formalized.

Thariq Shihipar [01:16:20]: To be honest, the response to the Pacing the Frontier, even within America, has been much more, like, well-accepted than I think a lot of people thought? And I think that, like, we have some precedent for being able to make these unified theory, like, agreements, in the world. And so, again, very much above my paycheck or expertise Right? but I think that, like, ideally we can, like, form these agreements. And I think, like, talking about this is the first step to forming those agreements.

Swyx [01:16:50]: And then the other point I really wanna. Like, one of our earliest podcasts is with,

Vibhu [01:16:54]: Emmanuel

Swyx [01:16:55]: Emmanuel from Anthropic on mech interp. Where is mech interp, right? Like, this is supposed to be where, like, if the models are thinking bad, we can see it, and the models don’t know yet, and we can act to stop it. I think that is something that people who are technical and who are developers, if you do care, you can make a lot of impact in here. But also, Anthropic is supposed to be the leaders in this.

External Evaluators and What Developers Should Advocate For

Thariq Shihipar [01:17:17]: Yeah. This is yeah, a great segue into fallbacks, like we. And probes. And, yeah, I wanted to talk about this a lot. I get asked this question a lot from people who are, like, often interested in ML research and asking about, like, why does this fallback happen, right? And so I think, like, at a top level, like, how does it work? So in inference time, we have what we call probes, and we have a paper about this called constitu- constitutional classifiers. And these probes look at the input and output activations. And, activations are, in the latent space, right? Like, how, what the model. what the model is thinking about, right? And so we try and figure out, like, okay, is the model, for example, like, trying to hack something? Again, you didn’t ask it to hack, like, Artifactory. Like, you just, it’s just deciding to do this to complete its task, right? So you would not get this if you just looked at the input. You have to look at the internal activations. I think that, like, this happens at inference time. So first, like, there’s a trade-off here of cost and speed, right? Where, like, we need to do this fast on every request to Claude and to Fable, and this has an overhead, right? and we need to then, like, fall back and we, like, do a classifier after the probes. Like, we’ve talked about this in the paper. But the nice thing about probes is that they’re refinable, like, live, right? So we can get this feedback, and then we can adjust it and things like that. ‘Cause the alternative is to program this, is to train this into the model, right? And we still do this as well. The model will refuse a request. That’s not a fallback, right? So, like, it’s not a probe that’s activating and falling back. It’s just refusing to do it. And we do this training. but it’s like there are a few failure modes, right? Like, it can, again, do something as a side effect, right? So it’s not something that’s part of the final output. you might have noticed that, like. I think, like, everyone’s tried to jailbreak models and like, try and, like, steer them off course or things like that, and probes help catch that, right? And so, like, we, like, do some training here, but we don’t want the like, refusals to be too strong, right? Because that, like, cuts it off much, like, earlier in the pipeline.

Mechanistic Interpretability, Probes, and Fallbacks

Swyx [01:19:32]: Yes.

Thariq Shihipar [01:19:32]: And this is interp, right? Like, probes are effectively a form of, like, mech interp. Again, it happen- has to happen fast. It has to happen at scale. But yeah, this, like, mech interp stuff is a good research problem. So, like, you can take, like, an open weight model and, like, try and understand its activations. I think we. Like, Gemma Scope is a good tool for this.

Swyx [01:19:53]: Here’s Llama for them.

Thariq Shihipar [01:19:54]: Oh, yeah.

Vibhu [01:19:54]: We have. This is your early work, so you had a little

Thariq Shihipar [01:19:57]: Oh, yeah.

Vibhu [01:19:57]: Time at Goodfire. We see you laid some

Swyx [01:19:59]: Which we both are also good friends at Goodfire.

Vibhu [01:20:01]: They’ve been

Thariq Shihipar [01:20:01]: Yeah, exactly. So I worked with, at Goodfire for a bit on, like, yeah, sparse autoencoders and just, like. It’s very complicated. RL has made this, like, much more complicated, I think is, like, one of the takeaways, where

Swyx [01:20:14]: Why? Sorry.

Thariq Shihipar [01:20:15]: Oh, sorry.

Vibhu [01:20:16]: What is

Swyx [01:20:17]: Yeah, why interp post-RL?

Thariq Shihipar [01:20:18]: I’m not so in the weeds here, but I think like, a lot of. SAEs were like. There have just been weaknesses with SAEs I think. And, yeah, I’m, I’m, I’m not a technical expert on this anymore. I just know it’s gotten more complicated. like there are base models and RL models, and there are more features that get, like changed. So, I think Goodfire has put out some work there. I’m, I’m not, I’m not deep in the weeds, but

Vibhu [01:20:43]: I will say for those, that want breadcrumbs, you guys have some of the best interp blog posts. So like the Golden Gate Claude, transcoders, all of your interp work, very nice visuals, very good

Swyx [01:20:54]: We’re the, we’re the interp podcast as well.

Thariq Shihipar [01:20:57]: Yeah.

Vibhu [01:20:58]: Yeah. we have a lot of interp stuff, so if you’re curious

Thariq Shihipar [01:21:00]: Yeah, I think this is like, one of those things where. And this is really what Anthropic is founded on, right? Like people. I think we invested in interp very early on, right? And I think that like when you say, “Oh, we’re an AI safety company,” really that means we want AIs to be able to run safely. And I think what we’re seeing is like for a super intelligent AI to run for long periods of time, it’s like a very complicated and difficult task, right? And so we’ve done this like investment into interp and alignment and, reward hacking and all of these like failure modes, right? And even then, it’s like, it’s really stretching. Like we need to like slow down a little or pace a little bit more. but yeah, I think like reading mech interp is. Like if you’re looking to get into research, this idea of like, hey, why is it hard to do this fallback easily? Or like why are there false positives, right? But we are working, of course, on reducing the false positives. Of course, as the models get more intelligent, now they can do more things, and they’re like what they can think about in lane space gets difficult. And so like as they get more intelligent, there’s going to be new false positives that we need to figure out and we need to iterate and things like that. But we’re, yeah, we’re working on this, and we do think this is like a critical part of, like deployment of these models. and, yeah, like, it means that we can like deploy this model without you having a perfect sandbox or something? Like you don’t have to like save everything. I think it’s worth talking a little bit about our security, like what we do for security there. So there’s like the model training stuff that we talked about. there is, the probes and classifiers, and then there’s auto mode that sits on top of all of that, which is like a another classifier that checks the requests that are being done, right? And so, and then beyond that, there’s like identity and permissions like we talked about with Claude Tag on like APIs and stuff. And so there’s so many layers of security that need to get done, and it’s like we said, very complex. Any of these failure modes at any one point can, like cause like agents to like escape the sandbox.

Constitutional Classifiers and Inference-Time Safety

Vibhu [01:23:08]: Auto mode was an interesting one. it seemed early on like, okay, it’s running for 10 minutes.

Thariq Shihipar [01:23:14]: Yeah

Vibhu [01:23:14]: If I’m on full access or auto, it’s not a big deal. But one thing you brought up is now it’s running for hours on end, right? there are fallbacks you still need. There are still limitations, so.

Thariq Shihipar [01:23:26]: Yeah, I think like. And everyone has these stories or like has heard these stories of like, oh, like Claude rm -rf, or not Claude, but like, models

Vibhu [01:23:34]: Not Claude.

Thariq Shihipar [01:23:34]: Of like rm -rf. I think I’ve seen this less, I’ve seen this less for Claude, but like again, it can happen. Like, this

Vibhu [01:23:40]: Yeah

Thariq Shihipar [01:23:40]: Like these models like can wipe, like sensitive data or something. Like you want to give models access to your production database, for example. but this is like an obvious, like, you can maybe scope your key, but I don’t know, can it issue its own keys? Can it like. Probably, like can it. It can use computer use to go issue its own key and then copy the key over and then edit your database because it needs to do it to complete the task? It’s just like one trivial example. And auto mode looks at that and be like, “Oh no, the user did not give you permission to, write to the database or to use computer use to like, emit a task,” right? And so this like probes are like on the intent level, right? They’re like, “Oh, okay, like hacking Artifactory is bad. Like we probably not, should not do that,”? But then like auto mode is more on like your own permission level. Like at sometimes you do want it to write to the database, sometimes you don’t, right? And you don’t want a probe to like interfere there, but like you need to make sure that the intent of what the agent is doing matches up with your request, right? And so auto mode operates at that level. And so yeah, security is just like very complex. There are so many different parts to it. And like, yeah, I like, I hope that this was like I. My goal is really to just get very technical about it and talk

Interpretability After RL and the Security Stack

Swyx [01:25:00]: Yeah, we’re, we’re listing out the things. If you’re not aware, this is the standard now.

Thariq Shihipar [01:25:04]: Yeah.

Swyx [01:25:04]: Like you must have this. It’s in line with what you’re talking about with the harness. Like that is the table stakes have risen quite a lot.

Vibhu [01:25:13]: I think some stuff that we can plug, as much as there is probing in your side of doing this and having classifiers for people building harnesses, the other side is model safeguards, right? So there’s open models. So Llama has Llama Guard. It’s a safety classifier trained version of Llama. OpenAI has OSS Guard, which is, same thing. You can attach these on to your harness, to whatever, to check is this stuff safe? A point that we should clarify on the OpenAI model Hugging Face thing is this was done with a unreleased model that was still in training, right? So when you put it in perspective, the prompt it’s being given in the RL environment is you have to solve this task. And this is a model that’s, still in training. It hasn’t had all of its safety post-training alignment. So a little different than something like auto mode, right? Auto mode is on production models that have gone through safety training, that have prompting that gives more safety guardrails and whatnot. So just breadcrumbs for people that are looking into it to, fill in gaps.

Swyx [01:26:18]: Yeah. Gray Swan as well

Vibhu [01:26:19]: Yes

Swyx [01:26:19]: And one of our previous guests. yeah, lots of safety architecture and lots of safety vendors, to buy. my, I think my final question on pacing is how long? Do we pace forever?

Vibhu [01:26:31]: Do we see GlassWing part two?

Swyx [01:26:32]: I. the scope is fix all software in the world, right? Listen, like, which it. We’re not. It’s not happening.

Thariq Shihipar [01:26:40]: I do not know. Like, I think that, like

Vibhu [01:26:43]: I’ll say one thing that’s good that I think we do is you have stuff like GlassWing. OpenAI also has this. So you will give it. you’ll give model access for security first for X amount of time so you can use it to self red team. Hopefully, you can expand programs like that, help on, we are safety experts, there’s others.

Vibhu [01:27:08]: Solve your problems first and then the model comes out. So this is one example, right?

Thariq Shihipar [01:27:13]: Yeah, exactly. Yeah, trying to, like, secure critical software. I think we fixed, like, a lot of bugs in, like, Firefox and things like that. So, yeah, like, across, like, operating systems and everything like that. So.

Vibhu [01:27:25]: At a high level, it’s just, you give the model you give people access to do security audits first, then the broader public that could use it for harm gets access.

Thariq Shihipar [01:27:36]: Yeah. I think what people like to say is like, software and cybersecurity is defense-favored

Vibhu [01:27:41]: Yeah.

Thariq Shihipar [01:27:41]: And that, like, you could theoretically. It will be hard, but you can engineer the perfect sandbox, and you can, like, have no, like, constraints. And yeah, like, what you need to do it is you need to get the super intelligent AI to engineer this perfect sandbox and check it and red team it and things like that. And so, this will just take time, and, like, of course, the models will get smarter. yeah, I think, like, I don’t know the specific, like, dynamics of how this thing goes. I’m really just like, Hey, like, I’m a developer? Like, I think this is how I understand this problem, and just, like, this is what’s happening right now, and this is, like, we should do something.

Swyx [01:28:20]: I think every engineer should know about it

Vibhu [01:28:21]: Yeah.

Swyx [01:28:21]: Because, like, it’s, it’s gonna be part of their job.

Thariq Shihipar [01:28:24]: Yeah.

Vibhu [01:28:24]: It’s a lot more than just, Dario and people can say it and you can look at the incident. There is an engineering side to it.

Thariq Shihipar [01:28:30]: Yeah. Yeah, exactly.

Swyx [01:28:32]: One thing that you also wanted to phrase is that this is. Even though you’re, you’re worried about the impact, it’s still low p(doom), and I think that’s a nuanced discussion. in general, people, very easily get into AI safety and X-risk discussions, but I think when you live in an AI lab, I think there are smart ways of discussing p(doom) and dumb ways. So what’s a smart way of discussing p(doom)?

Auto Mode, Permissions, and Long-Running Agents

Thariq Shihipar [01:28:59]: I, yeah, I have a fairly low p(doom). I can only speak for myself? And I do want to say Anthropic has, like, a diversity of opinions. I think, like, there’s many different ways to talk about it. And, like, I’m. I think that just, like, my mental model is that, like, I think we can collaborate on hard problems together. I think nuclear proliferation is an example of how we collaborated on this hard problem together. And, like, that is, like, the thing to me is, like, I’m like, I have faith in that? And I do think it’s a hard problem? So, like, I think it’s a hard problem. These are the technical reasons why, and I don’t know how you assign probabilities to things happening. I think it’s hard to do, but, like, my, like, overall is like, yeah, I think we’re very resilient and adaptable and, like, sharing this information I think is, like, the first step. And I’ve been really, like, excited about, like, how broad the discussion has become, right? And, like, how everyone has like, leaned in on Pacing the Frontier. And it really didn’t seem like this would happen maybe, last year or something, so.

Swyx [01:29:58]: Yeah.

Thariq Shihipar [01:29:58]: Yeah.

Swyx [01:29:58]: Yeah. And also maybe curing cancer.

Thariq Shihipar [01:30:01]: Hopefully. Yeah. That’s, that’s the goal.

Swyx [01:30:03]: There’s pacing and then there’s also like, well, let’s accelerate in useful ways, right?

Thariq Shihipar [01:30:06]: Yeah.

Swyx [01:30:06]: Like biology and all those things.

Thariq Shihipar [01:30:08]: Yeah. like, Dario’s essay on “Machines of Loving Grace” is the best representation of this, right? And I also agree, like, think you should read the Pacing the Frontier essay that Dario put out. Like, I put out, like, a quick summary, but I think it’s just like, there is a lot of detail here. It’s, like, an important problem and just being informed about it, right? but yeah, like, of course, the whole reason we’re doing this is that, like, we can get these enormous benefits, right? And, yeah, like, we’ve written a lot about that too. Yeah.

Swyx [01:30:35]: Okay. that was a huge tour, from, like, ask you some question tool to AI safety.

Thariq Shihipar [01:30:41]: Yeah. To Pacing the Frontier. Yeah.

Swyx [01:30:43]: Yeah. No, but, yeah, it’s clearly, it’s clear that you, like, really embrace everything that’s available to you in Anthropic, and, like, it’s, it’s good to at least have a peek inside of, like, what the discussions are, the topics are. any last words to people? Any, whatever you want to Call to action?

Thariq Shihipar [01:31:01]: Yeah, I think it’s. one, thank you for having me. I think this is like, I really

Swyx [01:31:06]: No, thanks for having me.

Thariq Shihipar [01:31:07]: Yeah. I

Swyx [01:31:08]: We first met in a Chinese restaurant.

Thariq Shihipar [01:31:09]: That’s right. Yeah. I think, like, I really enjoy the like, community you’ve created and the community of developers. And, I think that, like, I know things are changing really fast, and I think there’s, like, a lot to keep on top of, and, like, I think there is just a lot to do, and I feel. I think a lot of people feel, like, a little bit tired or anxious or something.

Swyx [01:31:33]: Stressed.

Thariq Shihipar [01:31:33]: Stressed, yeah, exactly. And this is, like, extremely understandable? And I think we. I understand, like. And we’re not perfect as well. Like, we, it’s, like, criticize and, like, understand, like, ways all of the AI labs could be better. and, but I also, like, am very excited about the excitement that everyone has for AI, and just, like, it’s a really exciting time. I think we’ll, like, look back at this time and be like, oh, like, this is, like, very hectic but very exciting, and, like, software engineering changed, like, forever. Like, other things will change. and it’s, like, really privileged to, like, be part of it, like, to talk to, like, the audience that you have and, to get to interact with all the developers who are, like, pushing the frontiers a lot on what’s possible. And I learn a lot from that too. Yeah.

Open Safety Models, GlassWing, and Defense-Favored Security

Swyx [01:32:20]: Thanks so much.

Thariq Shihipar [01:32:22]: Thank you.

From the earliest days of open-weight models to becoming the neutral routing layer for more than 10 million developers, OpenRouter is one of the clearest bets that the future of AI will be multi-model. In this episode, OpenRouter co-founder & CEO Alex Atallah, with AMP’s Anjney Midha returning with swyx to unpack how OpenRouter emerged from the first wave of Llama, Alpaca, Mistral, and Midjourney, why model diversity mattered before it was consensus, and how a company dismissed as “just a wrapper” became critical infrastructure for the AI ecosystem.

We go deep on the product and distribution lessons behind OpenRouter: why model labs can spend billions training a checkpoint and still struggle to get it into developers’ hands, how Mistral helped prove the value of a competitive inference marketplace, why OpenRouter chose focus over expanding into fine-tuning, memory, and other adjacent products, and how its rankings became a real-time map of how AI usage was changing. Alex also explains OpenRouter’s early experiments with model fusion, why they deleted the first version and brought it back years later, and how the platform grew to more than 10 trillion tokens per day.

Finally, Anjney explains why Stripe and OpenRouter fit together, why token fraud may become one of the defining security problems of the AI economy, and why the next wave of fraud won’t just come from humans but from autonomous agents attacking increasingly valuable token flows.

We discuss:

* Why OpenRouter bet early that no single AI model would win everything

* Alpaca, Llama, and open models becoming impossible to ignore

* Why Discord’s early AI deployments exposed the limitations of closed models

* Why model labs can spend billions on training and still fail at distribution

* How OpenRouter became a neutral distribution layer for model developers

* Why VCs dismissed OpenRouter as “just a marketplace” or “just a wrapper”

* The Mistral price war and the first real proof of an inference marketplace

* How Midjourney scaled through Discord and what it taught the AI ecosystem

* Why crypto infrastructure became a dress rehearsal for generative AI

* OpenRouter vs. LM Arena and why their missions are fundamentally different

* Why focus became one of OpenRouter’s biggest strategic advantages

* Anthropic’s early focus on AI pair programming and coding

* The OpenRouter products that were prototyped but never launched

* MOM, OpenRouter’s early Mixture of Models experiment

* Why model fusion failed in 2024 — and why it works much better now

* How OpenRouter’s leaderboard became a live map of the AI industry

* OpenClaw, auto-routing, and agents reshaping AI usage

* How OpenRouter reached 10+ trillion tokens per day

* Why inference gateways are increasingly becoming targets for fraud

* Why Stripe’s fraud infrastructure is strategically important to OpenRouter

* The coming rise of agentic fraud and attacks on the token economy

* What changes and what stays the same as OpenRouter joins Stripe

Alex Atallah

* LinkedIn: https://www.linkedin.com/in/alexatallah/

* X: https://x.com/alexatallah

* Website: https://alexatallah.com

Anjney Midha

* LinkedIn: https://www.linkedin.com/in/anjney/

* X: https://x.com/AnjneyMidha

* AMP: https://www.amppublic.com/

Timestamps

00:00:00 Introduction

00:02:12 Alpaca, Llama, and the Multi-Model Bet

00:06:04 Discord, Open Models, and OpenRouter’s Origins

00:14:28 Why “One Model Wins” Was the Wrong Bet

00:17:27 Why Model Labs Struggle With Distribution

00:23:04 “Just a Wrapper”: Why VCs Misunderstood OpenRouter

00:27:58 Bootstrapping OpenRouter Through Community

00:36:16 Crypto, Midjourney, and the Early Generative AI Ecosystem

00:43:38 Mistral and the Birth of the Inference Marketplace

00:47:10 OpenRouter vs. LM Arena

00:52:08 Focus, Anthropic, and Roads Not Taken

00:59:34 Mixture of Models and Model Fusion

01:02:44 Sonnet, OpenClaw, and OpenRouter’s Explosive Growth

01:09:03 Why Stripe Acquired OpenRouter

01:12:45 Fraud and the Emerging Token Economy

01:17:47 The Coming Wave of Agentic Fraud

01:19:07 What’s Next for OpenRouter at Stripe

Transcript

Introduction: OpenRouter, Marketplaces, and Pub-Sub as a Product Principle

Swyx [00:00:00]: Okay, we are here in Anja’s house, which is where all big startups in San Francisco start.

Anjney Midha [00:00:08]: Howdy.

Swyx [00:00:08]: And, congrats on Cursor, Mistral. I don’- God knows what else. You got so much stuff going on.

Anjney Midha [00:00:17]: There’s, there’s a lot going on. Well, OpenRouter is probably the - has been the most, I would say, like, one I’m excited about recently.

Swyx [00:00:24]: Yeah. And we have Alex, first time on the pod, but,

Anjney Midha [00:00:27]: Thanks for having me.

Swyx [00:00:27]: You’ve been in the IE a few times. I appreciate every time you’ve shown up, for the community. Congrats. I just, like, what a journey. When I was looking back at your past posts, one of the earliest principles that I saw you write as a product person is sub as a product principle. And I wanted - you to maybe explain how you think about what should exist in the world.

Anjney Midha [00:00:49]: Yeah. The sub piece, which was early 2023, I didn’t think about it until we talked like 10 minutes ago, is about how there is like a way of thinking about products as an intersection between subscribing to data and publishing data. And marketplaces are an easy example of this. You have suppliers that are publishing some product to a SKU. And the SKU is like a sub topic that a consumer is subscribing to and just going to, like, consume whenever they want. And humans consume in a very, like, discreet, ad hoc way. It’s not very scalable. all their attention is on the topic when they’re buying the thing, and their attention is nowhere else when that happens. agents and consumers of inference don’t act like that. They’re consuming continuously, and they’re changing the SKUs that they consume from all the time. So OpenRouter is like a blend between a normal API experience and a marketplace where we create model slug. We have the auto router. We have all kinds of, like, product SKUs that you can subscribe to. And then you can, like, continuously add, like, derive value and make decisions based on those consumers.

Alpaca, Llama, and the Multi-Model Bet

Swyx [00:02:11]: Yeah. This is something that was more consensus now, but not consensus when you guys started, which was that there is such a demand for swapping models and changing things out and, that people would not use the native SDKs. I guess, for each of you, what was your realization moment that this would be it? I, - You’ve, you’ve given a talk at EIE about Alpaca as,

Anjney Midha [00:02:33]: Yeah.

Swyx [00:02:33]: One of your inspiring moments.

Anjney Midha [00:02:35]: Alpaca, I can, like, rehash the Alpaca moment for a sec. Like, the very beginning, at the end of 2022, OpenAI was the only game in town. There was, like, OpenAI, Cohere,

Swyx [00:02:47]: Yes.

Anjney Midha [00:02:48]: And then a smattering of, like, early attempts at open weight models.

Swyx [00:02:54]: Yeah.

Anjney Midha [00:02:54]: When Llama came out in January of 2023, it was like, “Wow, really exciting. This is really big.” It outperforms 3 on, one or two benchmarks. but you can’t chat with it. It wasn’t like - It wasn’t an engaging model, but it seemed like someone just needed to fix a couple things and do some RLHF on it to get it all the way there. And Alpaca was the first model that I saw that did that. It only took $600 to do. A team at Stanford generated a bunch of synthetic data, tuned Llama, and made Alpaca, billion parameter model. Or was - Maybe it was thirteen billion parameters. And it was so good. Like, I was just, like, on an airplane using it. I, - in many cases, I, like, you could not discern a ChatGPT versus an Alpaca result. And I figured if it was this easy to make a model, one, we have a whole new way of monetizing data for the first time. you can just, like, take really valuable data and turn it into a service in $600. and that cost will probably go down over time.

Swyx [00:04:03]: When you - So sorry. when you say monetizing your data as, what eventually will become an MCP endpoint or as a training data for a model?

Anjney Midha [00:04:12]: Yeah, training data for a model.

Swyx [00:04:13]: Awesome.

Anjney Midha [00:04:13]: Like, an abstract way of saying like, “Hey, I have this data.”

Swyx [00:04:15]: Compress it into a model.

Anjney Midha [00:04:16]: Like, it makes sense for me in my product, but, like, I could repackage it in the form of a model and sell it. And so it’s just a whole new business model for the economy. It also, of course, provides, like, a way of following what Frontier Labs are doing, but in a way that, like, a single developer or a small team of developers can roll on their own. And so - Whenever you have an example of that, like a breakout app that’s doing really well, and then some framework for imitating it with - in your own flavor, you have an immediate ecosystem of, like an immediate ecosystem, like, should arise because there’s just a huge gap between the, like, decisions that the single company is making and all of the variations in those decisions that, like, a wider ecosystem can create themselves. And so then, you need a marketplace to, like, discover all of those, services and all of those products. There wasn’t any place on the internet that, like, was like a home base for LLMs in terms of seeing how much they were being used and seeing who was using them and why.

Swyx [00:05:29]: The closest would be Hugging Face.

Anjney Midha [00:05:30]: Hugging Face was the closest at the time, yeah.

Swyx [00:05:31]: They just started Hugging, like, a few years ago before that.

Anjney Midha [00:05:34]: Yeah, and Hugging Face also didn’t have the closed-source models.

Swyx [00:05:37]: Yeah.

Anjney Midha [00:05:38]: And they didn’- you couldn’t use the models at the time. and there wasn’t data about who was using them. There were, like, a bunch of differences between OpenRouter and Hugging Face, and those differences felt really critical to me, especially when I was just trying to learn about LLMs and, like, why people are choosing, like, Different little ones that are emerging over time.

Discord, Open Models, and the Origins of OpenRouter

Swyx [00:06:03]: Got it. And then, Ansh, no stranger to wanting more model diversity, at the time, you’re a couple of years into your Anthropic journey, which we covered in the previous podcast as well. What was your introduction to Alex?

Alex Atallah [00:06:16]: Well, the introduction was, I think, thirteen years before that.

Swyx [00:06:20]: Oh.

Alex Atallah [00:06:20]: But the OpenRouter handshake happened right over there, if you remember.

Anjney Midha [00:06:23]: Yeah.

Alex Atallah [00:06:24]: Which - So Alex and I, met, I believe as sophomores now, if I remember at the Stanford Review,

Anjney Midha [00:06:32]: That’s right

Alex Atallah [00:06:32]: Meeting for the first time.

Anjney Midha [00:06:33]: I think so, yeah.

Alex Atallah [00:06:35]: Yeah.

Anjney Midha [00:06:35]: Yeah.

Alex Atallah [00:06:35]: So Stanford Review was the libertarian newspaper on campus at Stanford that Peter Thiel started back in the day. And, whatever-- for whatever reason, I, Alex and I both showed up to one of the meetings, and I remember, the editor-chief was a mutual friend of ours. Lisa was really a really great editor-chief, where, part of an editor-chief’s job is to assign responsibilities to people and make sure the work gets done. and I, I may be misremembering the details, but I remember wanting to. It was surprising to me that at the time there was no dedicated technology section in the newspaper.

Alex Atallah [00:07:11]: You

Swyx [00:07:13]: Because it’s political, right?

Alex Atallah [00:07:14]: It is primarily

Swyx [00:07:14]: Like, it’s talking

Alex Atallah [00:07:15]: It originally started as like a

Anjney Midha [00:07:16]: Yes.

Swyx [00:07:17]: Yeah, states and all those things.

Alex Atallah [00:07:17]: Correct.

Swyx [00:07:18]: Yeah.

Alex Atallah [00:07:18]: But it, - To take us back in time, you may remember this, but, there was this technology, legislation that was being debated called, the Net Neutrality Act. And net neutrality is, like, inherently this political concept, right? It’s, it’s about the regulation of - internet broadband access. And so there was a community of us who were technologists, but also debating the politics of the technology. And I thought the Review would be a great place - to, like, write about that. And I was working on, I think, a net neutrality article, and I remember proposing, “Well, maybe we should start a technology section.” And Alex was one of the only people who said, “Yes, that would be cool.” And said. I forget whether we ended up writing stuff together, but - that’s when we first met,

Alex Atallah [00:08:03]: Was 2011 or twelve. I forget which year it was. It was one of those.

Anjney Midha [00:08:09]: Yeah.

Alex Atallah [00:08:09]: It was at Old Union, if I remember correctly.

Alex Atallah [00:08:11]: That’s where we used to meet. But, along the way, Alex and I have had a chance to, To hang out often. And probably the time when we had the most professional overlap was when I was running the platform at Discord, and it had become this explosive platform for crypto

Swyx [00:08:32]: Yeah

Alex Atallah [00:08:32]: And NFTs in the middle of the pandemic.

Swyx [00:08:35]: Which also, by the way, you were in charge of safety and security as well, right?

Alex Atallah [00:08:38]: I was the head of platform, which meant all of the crypto - the DAO and NFT launch security debugging fell on

Swyx [00:08:45]: And their phishing and.

Alex Atallah [00:08:47]: The phishing, the social engineering attacks, the katana DDoS that we were getting hit by. but it’s around the time I first started teaching security at scale at Stanford, CS 153. And Alex was on the, - at OpenSea at the time, and I was trying to figure out how we could defend against all these attacks that we were. Like, and at peak, I forget, if you remember how much NFT volume was running through

Swyx [00:09:10]: Discord

Alex Atallah [00:09:10]: Discord, but it was, like, a meaningful amount of, like, it was, like, several billion dollars in NFT volume of GMV, so to speak, were running through the platform, and it was all coming from OpenSea. It was these, like, buy, sell,

Swyx [00:09:20]: The

Alex Atallah [00:09:21]: Servers

Swyx [00:09:21]: The D in DAO is Discord.

Alex Atallah [00:09:25]: Yes. And so that’s when I think we had hung out professionally. But a year after that, OpenAI gave Discord early access to GPT. Sorry, three. No, it was five. Yeah, five, which is the RL version of three. And that’s around the time we made a Discord bot with, OpenAI for internal deployment, and that’s when I realized we would need. Like, since I was part of the deployment team.

Anjney Midha [00:09:50]: What was the use case?

Alex Atallah [00:09:51]: There were two that were. And there’s, there’s a post now called “Discord is Your Place for AI with Friends” that somebody sent me recently that I wrote, and published in twenty-three. But There were two use cases. One was Clyde, which was the - like, a party friend inside of Discord that could help you set up your Discord server and talk to you about onboarding and get your friends to hang out more. and then there was content moderation. And one of the realizations we had with content moderation was - it would refuse to moderate. Like, it would just refuse our prompts because the The training was. We were very early in the training era, and it would just. Our prompts would trigger it, its, like, guardrails. And we told OpenAI, “Hey, guys, we need access to the weights because if we’re gonna be doing content moderation at scale, we had 250 million monthly active users, we need more reliability that the model will do what we need it to.” And they said, “Well, sorry, guys, that’s not how this works. We’re a closed-source company.” And so that was my first realization that we needed open models, and the enterprises would need more control over capabilities, and then ultimately would need some control plane or management system to orchestrate these open models. But there weren’t no good - there were no good open alternatives until maybe

Alex Atallah [00:11:10]: Six months later when Llama came out. And six months after that, I led the series A into Mistral, which was started by Guillaume and the Llama team. And - That, - Around that time is when I remember hearing about Alex launching OpenRouter and going, “These worlds are gonna collide, and I don’t know when it’ll make sense to team up.” But Alex was so early and could see. I think he was totally right about this ecosystem starting with Llama that then needed, like, a, an easy layer to manage for, especially for. I was approaching it from the enterprise perspective because I had been that, like, the. As the VP of platform at Discord, it was my job to ensure that when we deployed models to, like, 250 million users, they did what we wanted them to. And that was very hard, because if you outsourced it to the labs and they controlled the guardrails and their guardrails are their safety policies. Forbid the model from responding to your prompts. That was quite catastrophic.

Swyx [00:12:05]: Yeah. But what, a moderation is the thing that they want to support. And obviously, beyond that, they would - OpenAI would work with you, presumably to give you a moderation endpoint, which they offer for free.

Alex Atallah [00:12:16]: It was an interesting use case, that - So they did give us a moderation endpoint. However, as you guys know, every Discord server is like a mini deployment of itself. And so the use case was instead of having human moderators that have to interpret the norms of the community, you just give the, - Often, like every, subreddit, Discord servers, public ones have their own rules that the user, the users create.

Swyx [00:12:41]: Oh, yeah. We run the LinkedIn Discord in. Yeah.

Alex Atallah [00:12:43]: And then humans used to read those norms and then enforce it every day manually, like observing each message in these communities. And these communities have like millions of users. So we had a 5,000+ person team globally in the, on the Discord content moderation team. These are outsourced contractors who had a really tough job. And so the idea was instead, if you could give the norms of that server To the LLM, then the LLM would do custom moderation for that server. It’s almost like a, like context moderation for that server. And many of those servers’ norms just violated OpenAI’s rules. And so - It was like we had our own custom eval. So each server had its own custom eval. But Discord-- at the time, OpenAI’s evals, we were all so

Alex Atallah [00:13:28]: Primitive in our thinking about how to deploy these LLMs that often the training prompts were super handed. It said, “Oh, anything about Harry Potter, anything that has trademarked content, don’- refuse.” And if it was a fan - Harry Potter fan community, this is a real use case, that had content moderation, the LLM would just refuse.

Swyx [00:13:48]: Yeah.

Alex Atallah [00:13:49]: And that was just not precise enough.

Anjney Midha [00:13:52]: Another one that we heard was like if someone was trying to write like a detective story, and there’s one chapter with a lot of violence, like maybe someone

Alex Atallah [00:14:01]: Right

Anjney Midha [00:14:01]: Like kills someone, the LLMs would just refuse to, like, help with that part of the story.

Alex Atallah [00:14:07]: Yeah.

Anjney Midha [00:14:07]: And then - like, we used to be like, okay, this is not like structurally inherent to LLMs. There must be, like, some choice out there so that I can, like, switch to another model, when I’m getting, like, a refusal or a bad result from the main one that I have. And that, like, tension also drove me for a marketplace.

Why “One Model Wins” Was the Wrong Bet

Swyx [00:14:28]: Yeah. I think that is well accepted now. What was it like back then when you were raising or, starting this? did people get it? what was the, some of the struggles? I like getting stories out of him about how other VCs don’t get it. So like anything you wanna, talk about, now - Let’s, let’s call it, that the early journey of OpenRouter is done, right? You can obviously talk about some of the early days stuff.

Anjney Midha [00:14:54]: Well, I was gonna say that, like, the biggest objection we got is big model win, which is - all of the

Swyx [00:15:03]: Scaling laws.

Anjney Midha [00:15:04]: Huh?

Swyx [00:15:04]: Scaling laws.

Anjney Midha [00:15:05]: Yeah, scaling laws, and natural network effects are just gonna accrue to one company, which will be - It’ll be a Google-style monopoly, just like how Google won the search market, by a large margin, and you’ll just be fighting for scraps at the end. That was probably the biggest objection we got. it is interesting that Google won the search engine race with such a huge margin. I think, like, had there been more interesting benchmarks or had, like, search engines been, - had people, like, seen them a little bit more like LLMs where they’re services that you can build companies on top of, that might not have been the case. but LLMs don’t merely have a user interface. They’re also, like, ways of building entirely new businesses. And, a Google-level monopoly would be like the Dutch East India Company times, quadrillion in magnitude because the whole economy ends up, like, depending on the one monopoly as well. So it didn’t seem like would be a really crazy outcome if that happened. And it’s also less likely because the economics of, like, creating good competitors are much, like, much more decentralizable.

Alex Atallah [00:16:25]: Everything Alex said is true, And I came at it from a completely different perspective, which

Swyx [00:16:31]: Yes, this is why we’re here.

Alex Atallah [00:16:32]: The scaling laws were never - In my mind, were always a feature, not a bug for why OpenRouter would be very valuable. Because, I was one of the first investors in Anthropic, and it was obvious to me that other researchers in our friends - I went to grad school for machine learning, and I just had a lot of friends in the ML community who it was very obvious to us that the bitter lesson holds. And so I was like, “Oh, fantastic. Now we have at least two proof points that compute scaling works.” It was OpenAI and Anthropic. and by the time I think we decided to team up on OpenRouter, I had already invested in Mistral and Black Forest Labs and Luma. So there was multiple model companies and teams that I was, working with.

Why Model Labs Struggle With Distribution

Swyx [00:17:14]: But you did other modalities, whereas this is literally

Alex Atallah [00:17:16]: Across different modalities, yes

Swyx [00:17:17]: Text.

Alex Atallah [00:17:18]: Exactly. And it was so obvious to me that an ecosystem of different kinds of models were being created, and that this whole narrative of, like, Only one company will dominate like Google was, well, like maybe true, but one, I don’t believe that. But two, there was so much extraordinary innovation happening across several different research teams. But the shared problem I was noticing across all of them was often, the research teams were fantastic at figuring out how to reason about new capabilities. They think in terms of capabilities, but never - like, are not developer mindset-oriented. Like, what happens after the training is done and the checkpoint comes out? Like, you’d be shocked how, like, similar the early training teams at OpenAI, sorry, Anthropic, BFL, Mistral, were in their, like, default approach to. Taking their research out of the, lab and scaling their impact, which is often, oh, the checkpoint is done, put it out as an API, done, and then there’d be crickets. in the case of Claude, the first Claude checkpoint was done a year before they released it internally. And then ChatGPT came out, and we decided, okay, yes, it’s a good idea to release a Claude version externally.

Alex Atallah [00:18:34]: And they had no plan, like no plan for how to get developers to try it out. And so if you go to the Claude one blog post, you’ll notice there are, like, three developer examples for users of the API, and one is a Discord bot, and the second is Vivian, my wife’s startup called Juny Learning, ‘- And then there was, like, Notion, because these were all friends of, like, the Anthropic Because that’s how - like, last minute the planning was around, hey, once the model’s done training, how do you get it out to the world? There was no distribution platform that understood what developers needed, all the key management, provisioning, like, simple, like, endpoint management, versioning control. Like, all these things that the scientists and researchers go, “ that’s plumbing. I don’t really think about it.”

Swyx [00:19:15]: Implementation detail.

Alex Atallah [00:19:16]: Right. And instead, Alex came at it from that perspective. And so, it was so obvious to me that, like, every single lab I was funding would spend - like, literally sometimes billions of dollars into training, and then a checkpoint would be done, and there’d be crickets, like, during early access because they’re like, “Oh, that’s right.”

Alex Atallah [00:19:35]: It’s hard to use a checkpoint to make anything. You need a whole bunch of plumbing around it to make it usable by a developer. And so by the - I think - it was so obvious to me that a distribution platform like OpenRouter was critical to have in the ecosystem if we wanted there to be competition to Google. Like, unless-- ‘cause with Google, DeepMind is done training a new checkpoint, and then they push a button, and it gets blasted out across all their surfaces from Google Docs to,

Swyx [00:20:01]: Everywhere, even if I don’t want it.

Alex Atallah [00:20:02]: Everywhere. You wanna know about, like, on Android, like, overnight, they can deploy a new checkpoint to, like, a billion devices, right? And that invisible infra advantage, distribution advantage, most people don’t realize, but until OpenRouter showed up, - you had to think about all of that yourself as a model lab. And it was very daunting. at Anthropic, I think it took, well, more than twelve months to get to our first 10 million in revenue. And in contrast with Black Forest Labs, I remember the early days, you guys had a conversation with the BFL team, and, it was so simple for OpenRouter to say, “Oh, no problem. Like, the day you launch, we can send 1 million developers to you.” that was crazy. That was like a step function change in, like, an hour.

Swyx [00:20:46]: Is that a real number, a million?

Alex Atallah [00:20:47]: I,

Swyx [00:20:48]: Okay. All right.

Alex Atallah [00:20:48]: I think today it’s, like, 4 million. How many developers are on OpenRouter today?

Anjney Midha [00:20:52]: Over ten,

Alex Atallah [00:20:54]: Yeah.

Anjney Midha [00:20:54]: Over 10 million, but, like, it’s, it’s hard to, you

Alex Atallah [00:20:59]: I, yeah, I don’t know how to. Yeah.

Anjney Midha [00:21:00]: We do a lot of, like, account duping work, but, no

Alex Atallah [00:21:04]: If you could get 1,000 developers, just to put in context If you get 1,000 developers who try the model on day one after you release it and just, like, do inference and give you feedback, that’s a thousand

Anjney Midha [00:21:15]: That’s huge

Alex Atallah [00:21:16]: More developers than they knew how to get to on their own.

Swyx [00:21:19]: Well, BFL had a reputation, but yes.

Alex Atallah [00:21:21]: They had one in Stable Diffusion.

Swyx [00:21:22]: Yeah.

Alex Atallah [00:21:23]: And with Mistral, I don’t know if you guys remember, but the first checkpoint they released was, like, torrents. It was, like, torrent weights.

Swyx [00:21:31]: Yeah, they just put up a magnet link.

Alex Atallah [00:21:33]: Yeah, there was no API.

Anjney Midha [00:21:34]: Yeah.

Alex Atallah [00:21:34]: Because they didn’- they weren’t infra people.

Alex Atallah [00:21:37]: ? Like, it’s like, okay, download these weights, and you guys go figure out how to host it.

Swyx [00:21:39]: Well, he has a story on his side, yeah.

Anjney Midha [00:21:41]: Yeah, in addition to the, like, building a really good developer experience around it, the marketing that we do on, like, for different models is totally different and perceived totally differently

Alex Atallah [00:21:54]: Right

Anjney Midha [00:21:54]: From the marketing that a model lab does for itself.

Alex Atallah [00:21:56]: Yes, 1,000%.

Anjney Midha [00:21:57]: Right? We are like a, neutral layer looking at this market like it’s a big dark room with all the corners completely obscure to users, and users are walking into the room and, like, feeling around

Alex Atallah [00:22:09]: Yeah

Anjney Midha [00:22:09]: And trying to figure out what objects to grab off the tables and, like, build into, their companies. And it’s just an insane way of working. Like, models are not products where you can just enumerate all their features onto a web page. They’re all black boxes, including the open weight ones. So you need to, like, shine lights on all corners of this room, so that people can see what makes this model good, and you need the company shining that light to be a neutral third party, which is what we specialize in. So the, like. It’- In addition to developer experience, there’s also, like, a very important, like, marketing and product packaging component

Alex Atallah [00:22:50]: Yeah

Anjney Midha [00:22:50]: And a way of, like, routing and discovering models becomes, like, critical to your market as a provider or a model lab or a server tool and more in the future.

“Just a Wrapper”: Why VCs Misunderstood OpenRouter

Alex Atallah [00:23:03]: And this value, to your earlier point about how many VCs, like, just don’t. One of my biggest frustrations is that venture capitalists, many of them, like, just don’t have any operating experience in the field. so unlike a traditional investor who’s just maybe come up through the ranks as, like, a associate working on financial modeling or maybe hasn’t been a real operator in the field for, like, more than ten years, which is a big part of the industry now, I had just arrived at a16z, like, a year after running the platform. And so I knew what the challenges were of, like, building a real - great developer experience and like, being able to create a working piece of software with a model. And there were a few, I won’t name names, but there were investors who were looking at OpenRouter, and, felt at the time, like, when I would compare notes with people, that it was just, I quote unquote, “just a marketplace.”

Swyx [00:23:59]: Yeah, just a thin layer, just a

Alex Atallah [00:24:00]: Correct

Swyx [00:24:00]: Just

Alex Atallah [00:24:01]: A wrapper or whatever on other people’s APIs. And I was like, “You have no idea how strategic the value that OpenRouter has created by being able to orchestrate even three.” APIs in production. The amount of both engineering work and community design that goes into getting that live and running in production at the scale the OpenRouter team had started just doesn’t happen by default. And that was one of the things that stood out to me about Alex from the earliest days. Like, he just understood, like, - from a systems perspective, like, how do you get these flywheels going? Like, that stood out to me with OpenSea when we were working together on the NFT integration at Discord. Like, Alex had a level of community-- like, systems thinking on how you get these flywheels going that most scientists and machine learning people just don’t

Alex Atallah [00:24:48]: Think of. Like, we often think in terms of training.

Swyx [00:24:52]: It’s a linear stage.

Alex Atallah [00:24:53]: It’s this linear pipeline.

Swyx [00:24:53]: There’s no loop yet.

Alex Atallah [00:24:54]: Yeah. It wasn’t until much later that the modern context feedback loop cycle really got standardized in the industry. But at the time, if you remember, machine learning was like. Like, mostly we did a lot of ML, like, when I was in grad school on a laptop. So you just, like, download a dataset, ran some ablations, and you looked at the loss curves, and you’re like, “Great, I made AI.” And the idea that you have to, like, deploy those capabilities, collect feedback trajectories, then, like, put those into a continuous loop, like, came much later. And it was very counterintuitive to the - like, the traditional AI mindset. I do remember doing the investment phase for, OpenRouter, I just didn’t try and educate a bunch of other VCs on why it was not just a marketplace. I was like, “ what? I’m just gonna invest.”

Anjney Midha [00:25:41]: Yeah.

Alex Atallah [00:25:41]: And I’m going to, like, take the opportunity to partner with Alex, and if - no other VCs get it, that’s totally fine. ‘Cause at the time, - it was not obvious, I think, to several of the investors that, like, OpenRouter was not more than just a wrapper around APIs. And - that infuriated me. And I was like, “ what? I don’t have time to debate you. I’m - we’re gonna, we’re gonna invest.” And then I think, like, a month later, Matt Murphy marked it up by 10x. Like, - I think. I forget what the exact money was and so on, but, to his credit, Menlo Ventures realized, “Okay, there’s much more strategic value here as well.” Maybe you didn’t hear all these conversations behind the scenes But that frustrated me a lot. there’s a lot of this, like, opining about wrappers. and if you’re like, “Oh, an app is just a wrapper on a model,” then, like. And, OpenRouter is, like, this wrapper on top of other APIs, and this is the most stupid, reductive framework.

Alex Atallah [00:26:31]: And so it’s clearly somebody who has no experience deploying product at scale.

Swyx [00:26:34]: It’s the thing you dismiss other things with. Like, you’re a - everyone’s a wrapper on everything, right? Like, and there’s, there’s some Some wrappers have value.

Alex Atallah [00:26:40]: Investors are wrappers and LPs, right?

Alex Atallah [00:26:42]: Like venture capitalists. So, yeah, it’s all wrappers down, all down to bare metal, I guess, and like energy.

Swyx [00:26:46]: Yeah, there - When I started the whole AI engineer, I guess, the coining, in 2023, like, that was, like, the number one pushback is that this is no value. You should just train models.

Anjney Midha [00:26:56]: Right.

Swyx [00:26:57]: And, yeah, obviously this is, like. you guys are one of the testaments to the fact that you can build very valuable wrappers, but also very valuable model companies.

Alex Atallah [00:27:06]: It’s so, hard to be. Like, the day a model launches, the fact that you have an OpenRouter, endpoint for that model frequently at the top of Hacker News on day one, people don’t realize the amount of work that goes into accomplishing that. And OpenRouter used. Like, that would happen over and over again, and I remember going, “People have no idea how hard that is.”

Alex Atallah [00:27:30]: That’s not.

Swyx [00:27:31]: Yeah, we’ve covered some of the inference engineering that goes behind,

Alex Atallah [00:27:34]: Yes

Swyx [00:27:34]: Some of - with Base Ten and all those. Well, today you have, all those, like, cool code name things that people guess what Oxy Alpha is and all those things. But, like, I guess one of the things that you’re teasing is, how do you get that initial flywheel going, right? Because today you have your scale and your reputation, all these things, so obviously you - you’re driving immense distribution. But when you were early on, when it’s mostly

Bootstrapping OpenRouter Through Community

Alex Atallah [00:27:55]: The bootstrap, yeah.

Swyx [00:27:56]: Yeah.

Alex Atallah [00:27:56]: What was the bootstrap like?

Anjney Midha [00:27:58]: To bring it back to early Discord days, I think we, like, initially connected with. This is an OpenSea story, technically. But, and we initially connected when you were at Discord, and we talked about, like, - the Axie Infinity server.

Alex Atallah [00:28:13]: Oh, yes. Yes.

Anjney Midha [00:28:14]: This server was, like, the biggest server at the

Alex Atallah [00:28:17]: Yeah

Anjney Midha [00:28:17]: At Discord.

Alex Atallah [00:28:18]: That’s right.

Anjney Midha [00:28:19]: And you were like, constantly bumping up the

Alex Atallah [00:28:22]: The limits on the server. Oh, my God

Anjney Midha [00:28:24]: Of how many people could be in the server.

Swyx [00:28:24]: For those who don’t know, like, 10% of Philippines was Axie.

Alex Atallah [00:28:29]: Was on that server. That’s a big hit.

Swyx [00:28:31]: It was, like, a meaningful contributor to the GDP of the country.

Alex Atallah [00:28:33]: It was an NFT, like, crypto game, but it

Swyx [00:28:35]: It was like a Pokémon breeding thing.

Anjney Midha [00:28:36]: Yeah.

Alex Atallah [00:28:36]: Yeah. Similar. Yeah. There was battling, there was breeding, and then there was, like, a marketplace for trading.

Swyx [00:28:43]: Earn as well.

Alex Atallah [00:28:45]: Yeah, earn. And, like, the graphics were really cute and fun, and you like, you get emotional about your Axie that you make. So to, like, start a community like that, which we had to do many times at OpenSea with every early project, for us to create a marketplace for it, we need to make sure that the, like, the community wants it.

Anjney Midha [00:29:09]: Right.

Alex Atallah [00:29:09]: And it’s like building something that people want and going and telling them about it. Like, you can do that on a one basis, but there’s way higher leverage to do that in a community where everyone can talk to you at the same time. So we spent a lot of time, like, building things that the community really wanted. We did the same thing for OpenRouter. And, like, the Axie community was one of, like, a zillion communities we did that with. And Anj, like, saw us doing it and. ‘Cause you could just see people sharing OpenSea links constantly in that Discord. Like, users sharing links is a really clear indicator that, like, something important is going on. So we spent, a lot of time, like, first figuring out what the gap is in the technology that people care about. Like, what was the actual problem that needs to be solved? in early LLM days, it was, OpenAI refusing to finish the prompt or,

Anjney Midha [00:30:09]: Yeah

Alex Atallah [00:30:10]: To, like, complete the task. It was also.

Anjney Midha [00:30:13]: Inability to customize models. and so there are communities that, like are just completely blocked on that issue, and those are the communities that are most useful to learn about and dive into and explore.

Alex Atallah [00:30:28]: Something that really struck me at that time, - as I was just hearing your talk, I remember noting - you may not remember this, but we - we had these, like working, Zoom calls that we were doing a sprint around for, like this OpenSea integration with Discord. and, we’d, we’d - it was myself, my engineering team. I think you were there. And I remember, Alex, in the middle of one of those calls, just like there was like silence. we were all like, “Oh, yeah, this totally makes sense. Let’s do this.” And then there’s - every, like everybody aligned. And Alex was like, “No, this makes no sense to me.” And everyone’s - I remember going, “What? Like, it works. Like, you click on a link and this, then it bounces you out to, like, OpenSea.” And he was like, “It’s not a good user experience. Yeah, we should not do this.” And I remember going, he was the only one person out of all of us to raise his hand and go, yes, it made sense from a technical implementation perspective. Like, we were bouncing the user out into the, into OpenSea. And so it kinda checked the box of the product manager’s requirements on both sides. But Alex went one step further and was like, “ what would be better, guys? If we just embedded the experience right here inside of Discord so the link opened up as an embedded iframe, and you can just check out right there.”

Alex Atallah [00:31:47]: And not one person on the call, and there’s like seven of us who had met, like, week after week.

Swyx [00:31:52]: And it’s the guy who doesn’t work for Discord.

Alex Atallah [00:31:53]: And it’s the guy who doesn’t work for Discord.

Swyx [00:31:55]: Like, technically, you benefit if they bounce.

Alex Atallah [00:31:57]: Exactly. And that was, like, adversarial. To keep the user inside of Discord would be adversarial to OpenSea. And yet Alex put that user experience first. And I was like, “That’s special.”

Swyx [00:32:08]: Wow.

Alex Atallah [00:32:08]: Because it’s very hard to have somebody who’s technical like Alex and understands the developer flow, but also understands the best user experience and wants to prioritize that. And that’s two sides of the flywheel that if you can get spinning, like is often hard to stop. And you just reminded me, like that one was one of those moments where I go, I - I realized I gotta be better at user experience because I should have been the one who came up with that, and I didn’t. And I learned from you. And, I think that went into one of our case studies for the PM training program at Discord.

Swyx [00:32:34]: Whoa.

Alex Atallah [00:32:36]: I don’t know if it there is Because of

Swyx [00:32:38]: You need an Alex is the conclusion.

Alex Atallah [00:32:40]: Yeah. You need an Alex. And this is why I’m not, nobody should be surprised why Stripe decided like they had to buy OpenRouter because it’s a really rare combination of people who understand the machine learning community, the developer experience, and the user experience. And putting all that together has resulted in this extraordinary scale that very few other marketplaces have been able to achieve

Window AI, BYOM, and Finding the Right Form Factor

Swyx [00:33:02]: Yeah.

Alex Atallah [00:33:02]: Over the last, five years.

Swyx [00:33:04]: Yeah. Well, we should talk about the other reasons for acquisitions, which

Alex Atallah [00:33:07]: Yes, we should.

Swyx [00:33:07]: You’ve written about. I wanna proceed somewhat chronologically as well. So - there is a point that, one of the questions that, Dave from H of Zero sent in was, when did it - really started to work? And you brought up Mixtral. I don’t know if you wanna bring up that story.

Alex Atallah [00:33:22]: Oh, yeah.

Swyx [00:33:23]: Which obviously you overlap with, so.

Anjney Midha [00:33:26]: Yeah, the MoE was. I don’t know when. there’s no like one moment where I was like, “Oh, this is, officially starting to work.” It was

Swyx [00:33:36]: The moment where you had a Chrome extension, like, really super early on.

Anjney Midha [00:33:39]: Oh, yeah. But, well, - yeah. So before OpenRouter, I wanted to, like, explore a bring-your-own-model experiment. And,

Swyx [00:33:47]: Which anyone familiar with crypto is like, yeah, Phantom and all these things.

Anjney Midha [00:33:50]: Yeah. So it felt like doing a MetaMask analogy for AI would be a fun way of exploring that. And at the time, there were no AI apps. There were probably as many AI apps that were, like, hitting AI - like, hitting an LLM via an API call as there were, like, games just doing it in JavaScript. like there was a, there was a moment in time where it could have been the case that web apps call LLMs through the browser, like through some desktop

Alex Atallah [00:34:27]: Yes.

Anjney Midha [00:34:27]: Managed app that is controlled by the user. and of course, there are like, I think, many reasons that did not happen. But back when the days were that primordial, I built a Chrome extension called Window AI

Swyx [00:34:43]: With Plasmo.

Anjney Midha [00:34:44]: With Plasmo.

Swyx [00:34:45]: I had come across early on, and I was like, “Who’s gonna use this?” You did.

Anjney Midha [00:34:49]: Plasmo had a couple, like, I think Phantom was using it. there were some other, like real companies using it.

Alex Atallah [00:34:56]: It was like a shim.

Swyx [00:34:57]: React for Chrome extension. It compiles to all

Anjney Midha [00:35:00]: Yeah.

Alex Atallah [00:35:00]: I see.

Anjney Midha [00:35:00]: Like Next.js for Chrome extensions.

Swyx [00:35:01]: Next.js, Next.js.

Alex Atallah [00:35:02]: Okay.

Anjney Midha [00:35:03]: And yeah, built Window AI on top of it. The creator of Plasmo, like started contributing code to Window AI, in GitHub, and that turned out to be Louis Vicchi

Alex Atallah [00:35:15]: Oh, you’

Anjney Midha [00:35:15]: Who is the founder of OpenRouter.

Alex Atallah [00:35:17]: That’s right. You have told me this is how you met Louis. Yes.

Anjney Midha [00:35:19]: Yeah.

Alex Atallah [00:35:19]: Okay.

Anjney Midha [00:35:20]: So, that allowed users to like configure which model they wanted to use for a web page in their browser, and then, like the app would just call out to that model when it needed to do things. not the right form factor for LLMs, but, it’s like fun experiment. You learn a lot, and like I open sourced it. And the main learning is like, okay, this has to be an API, and it has to look a little bit - like, there has to be more of a developer experience here and more of a discovery experience as well. Like, I don’t know where to use these models, and a little Chrome extension is not gonna help me discover. It’s not enough real estate. I need more space. I need visuals. I need graphs. I need, examples. I need images. I need to, like, I need to be able to, like explore both as a human and as an agent.

Crypto, Midjourney, and the Early Generative AI Ecosystem

Alex Atallah [00:36:10]: Yeah.

Anjney Midha [00:36:10]: So that’s how OpenRouter came to be.

Alex Atallah [00:36:13]: A meta point that.

Alex Atallah [00:36:16]: I think is underappreciated, but Alex is reminding me, is that we were quite lucky that we were so. we were, like, adjacent to the crypto community in those days. Because in hindsight, crypto ended up being like a dress rehearsal for generative models, right? If you think about the Axie experience, Alex is totally right, there were not that many AI apps at the time. And while I was dealing-- my job was to be the head of platform at Discord, which meant to be a general purpose place for communities and friends to create-- for developers to create apps and bots and, other services that could be deployed across Discord. And while 80% of the attention at the time was being spent on crypto, because that’s where all the NFT volume was, there was, like, twenty percent of my time I was spending with a friend, who would get hotbot with me and ask me for. We would play Magic: The Gathering on weekends, and he was working on a little Discord bot that could take a text input and turn it into an image, and it was called Midjourney. You

Swyx [00:37:15]: Is that David?

Alex Atallah [00:37:15]: It was David Holz.

Alex Atallah [00:37:16]: He was a good friend. And David and I have both been failed ARVR founders, in the before that. And, I remember this. Midjourney was one of the fastest-growing communities we had after Axie Infinity started to peter off. And many of the, like, the abstractions and the infrastructure decisions we made to scale Axie happened just in time because they. Axie did this and then fell off a cliff. And then as Midjourney was taking off, we, like, explicitly decided to help David make the server, the Midjourney server, as the primary place for interaction with the model, because it was very hard for people to understand how to use the model if they couldn’t see other people using it and copy them. And so the single-player Midjourney web app on its own, like midjourney.com, had, like, terrible retention because people would show up, they’d see this empty field. It’s like E 2, and they would type in, like, cat or dog. And it was, like, paralyzing for them to have this blank canvas that they had to fill because they’d never used an AI model before. But instead, in a Discord server, you could see other people using it and riff off of their prompt, and the engagement was off the charts. And so scaling, Midjourney from zero to, like, 10 million monthly actives was a much smoother approach Axie Infinity. And so,

Swyx [00:38:29]: Don’t forget the best of four pictures, and you choose one.

Alex Atallah [00:38:31]: The best, yeah, and then the other, we

Swyx [00:38:32]: Which is the feedback loop.

Alex Atallah [00:38:33]: The RLHF feedback loop, which, by the way, separately, like, Tom Brown, David and I used to play Magic: The Gathering on weekends. And so, like, it was one group of friends would hang out, and we’d. Like, these concepts were all being discussed all the time. But, there was.

Alex Atallah [00:38:47]: I think there were few of us who bridged both the crypto worlds and the AI worlds. And compared to crypto, where it was - the question was always, what’s the use case, for this technology? There was never any need to ask that for AI because it’s, like, the use case was so visceral. It was like, I can create now anything at - I can imagine. I can write novels, I can code. And the infrastructure that those of us who believed in the distributed systems, like, value of crypto, like the censorship resistance part, found this use case that was explosive. And I think between Midjourney, the, Claude was a Discord bot launch, that we were using internally as an LLM. ElevenLabs had a TTS model that we had on Discord as well. Like, Discord became this petri dish for, like, early apps to innovate. And I don’t think it’s a coincidence that they found a home there before OpenRouter gave the world, like, a public home store or, like, a, storefront. Discord was this, like, almost petri dish storefront that - had, like, piggybacked on the infra we’d built for crypto communities. And then I think Alex was one of the first people to realize, wait a minute, like, these apps need their own home, on the internet. And then OpenRouter, to me, was a continuation of that community’s needs. And of course, there was the crazy distribution that you enabled for a lot of these developers.

Why OpenRouter Couldn’t Just Live Inside Discord

Swyx [00:40:07]: So then my question is, how come you were. My perception is OpenRouter is not that Discord-centric, right? You have a Discord.

Anjney Midha [00:40:14]: Yeah.

Swyx [00:40:14]: And you use it to engage your community, but it’s not like Midjourney where, like, no, that is like the primary way people experience OpenRouter.

Anjney Midha [00:40:21]: Yeah, Midjourney, like, it really helps to see visually really quickly how people are using the model and how to prompt it.

Swyx [00:40:29]: Yeah.

Anjney Midha [00:40:29]: And I think that is partly why the server was so critical. It’s like it is the user experience. It adds a ton.

Swyx [00:40:36]: Yes.

Anjney Midha [00:40:37]: And you can go the whole mile with just, like, prompting via Midjourney, like, the, via the Midjourney Discord server, getting your images and then sharing them and having fun. For OpenRouter, for LLMs, like, you need a lot of user experience around LLMs to make them, like, really usable.

Swyx [00:40:54]: Charge point.

Anjney Midha [00:40:55]: And yeah.

Anjney Midha [00:40:57]: The, like, seeing the examples of other people is also not as useful because it’s a lot of stuff to read. It takes a long time.

Swyx [00:41:03]: Yeah.

Anjney Midha [00:41:04]: You need, like, based integration. Not possible to do in a Discord server. You need, Or technic- it’s possible. I shouldn’t say that. It’s just not a great developer experience. you need, like, - you need governance for. At the point where you got based integration, now you need governance for managing the LLMs that have access to it, the data policies, which teams. All that stuff needs a lot more than a Discord server can provide. So it’s just

Swyx [00:41:30]: Yeah

Anjney Midha [00:41:30]: It’s not the right.

Alex Atallah [00:41:32]: Well, in addition, you’re not wrong, but also there’s the very important distinction that, Midjourney was an end user application.

Swyx [00:41:40]: Right.

Alex Atallah [00:41:40]: And, that’s why Discord, which has 250 million monthly end consumers, made, it made sense for Discord to be a host for that application experience. What I knew was gonna happen soon after Midjourney found explosive product-market fit, because we. I think when Midjourney launched, from launch to $100 million revenue run rate, it was less than eight months. And shortly thereafter, Stable Diffusion launched. And, all of us used to hang out in the Discord server. There, I think it was the,

Swyx [00:42:13]: The Stability Discord?

Alex Atallah [00:42:14]: It was the

Swyx [00:42:16]: Yeah, LAION.

Alex Atallah [00:42:16]: Yeah, the LAION Discord server.

Swyx [00:42:17]: The image community that spawned Stable Diffusion.

Alex Atallah [00:42:19]: The image community. Yeah. And so when Stable Diffusion came out, I realized- Oh, now other people can build their own Midjourney.

Alex Atallah [00:42:27]: Because until then, Midjourney did not have an API, so they were a stack company, right? They were training their own models, and they were deploying them as an application. But if you wanted to build your own Midjourney, there was no API of that quality. and I think E two was still quite primitive. Like, Midjourney had great quality. And then when Stable Diffusion came out, suddenly there was this new person who - there was - this new capability in the world, which is a developer could create their own Midjourney. And that, I think, created the need for something like OpenRouter, because then you need an API to. If you - if you had the creativity of David Holz and you had Stable Diffusion as the model and you wanted to put these things together, how could you do that without having to figure out how to host the weights? And what OpenRouter, - the shape of OpenRouter enabled is that. Right? When you have open model alternatives to closed applications, OpenRouter’s value in the world becomes extraordinary because now any developer can just show up and use the

Stable Diffusion and the Need for a Model API Layer

Swyx [00:43:20]: You just love model diversity.

Anjney Midha [00:43:21]: Did you just say the shape of OpenRouter?

Alex Atallah [00:43:23]: Oh, no.

Anjney Midha [00:43:25]: Were you in cloud? What is this the real Han?

Alex Atallah [00:43:26]: I’ve been, I’ve been - I’m, I’m misaligned now. I’ve been overtrained. I’ve been using Cloud way too much, haven’t I?

Swyx [00:43:34]: Claude-ish is what people would say.

Alex Atallah [00:43:35]: Claude-ish. Oh, God, I gotta untrain myself.

Swyx [00:43:38]: Okay. - And I just wanna cap off the Mistral side. my TLDR is there was a Mistral price war, is what they called it, right? Like, round about NeurIPS is twenty-three or twenty-four.

Mistral and the Birth of the Inference Marketplace

Anjney Midha [00:43:47]: Yes. December

Swyx [00:43:48]: They launched, the Mistral 8x7B, and like the price went down like 80%.

Anjney Midha [00:43:54]: Yeah.

Swyx [00:43:54]: To me, that’s very positive because it’s like the first, like, real competition to host Mistral. Is there more?

Anjney Midha [00:44:01]: Yeah, that was. I’m, like, trying to remember it, all the things that happened. It. Like, we saw that model come out and immediately saw people say that it was the best model in the world.

Alex Atallah [00:44:15]: Yes.

Anjney Midha [00:44:15]: Like, this was, to my knowledge, the first time an open weights model was called that in real seriousness.

Swyx [00:44:22]: It’s hype, right? Is it?

Anjney Midha [00:44:25]: It was hype. It was hype. It was also, like, hype from AI influencers at the time. And there were many examples where it was, like, outperforming four. So people really wanted to try it out and see, is this gonna be true for me too? And if so, at what price? And, the, like, inference landscape was really messy.

Alex Atallah [00:44:49]: Yes.

Anjney Midha [00:44:50]: We cleaned it up. - it allowed, like, providers to compete on price, so we could give you just the best price in one spot. And so it was, I think, the first clear example of, like, a provider marketplace working in a way that adds value to end developers.

Alex Atallah [00:45:08]: Sean, you may not remember this, but I think we met for the first time a few days after Mistral came out at NeurIPS

Anjney Midha [00:45:15]: Yeah.

Alex Atallah [00:45:15]: At a luncheon.

Swyx [00:45:16]: Yeah. That’s where I also met BFL as well. Yeah.

Alex Atallah [00:45:18]: And Guillaume was there.

Swyx [00:45:19]: Yeah.

Anjney Midha [00:45:19]: I was at NeurIPS at that time.

Alex Atallah [00:45:20]: You were there too. And, we had just announced the Mistral investment, and I remember Guillaume was over there, and I remember turning to Guillaume and asking him, Like, “Is it is all the. Like, how are you feeling after the launch of Mistral and seven B?” And, him in his typical French fashion was like, “ it’s a, it’s an okay model. It’s not that good.” And I was like. It was so, in contrast. But I remember him also saying that part of the reason he felt a lot of people Thought that it was better than four was because of the speed. - it was an MoE model that they had, like, absolutely figured out how to make super efficient. It was on the Pareto frontier. And this is an important thing about LLMs, right? Sometimes when they’re faster, you think they’re smarter, even though, like, if you did, N of, these common, like, evals that are - you do seven tries, and I don’t remember. I think we should go back and figure out what the data says, but I wouldn’t be surprised if it turns out, oh, on an N of seven attempts, four was smarter on evals, but the perception of on, like, or correctness would be smarter or more accurate. But, people, like, from a human preference perspective felt that it was faster because it - or smarter because it’s so fast.

Swyx [00:46:36]: Yeah. And most queries do not take that level

Alex Atallah [00:46:39]: Don’t take that. That’s true.

Swyx [00:46:40]: Right? So this is the start of humans as router

Alex Atallah [00:46:42]: Yes.

Swyx [00:46:42]: Which then eventually becomes OpenRouter as router of like the

Alex Atallah [00:46:45]: Oh, that’s interesting way to think about it. Yeah.

Swyx [00:46:47]: Like, because humans are the routing mechanism. Like, I will ask the fast model first, and then if, like, oh, not good enough, I’m gonna upgrade manually.

Alex Atallah [00:46:52]: Yes.

Swyx [00:46:53]: But then he’s gonna auto it.

Alex Atallah [00:46:54]: I didn’t, I hadn’t thought of it that way, but that makes sense.

Swyx [00:46:57]: Which then there’s, there’s a lot more techniques, like fusion. Fusion is the thing that we should talk about. Before I move on to those things, I just want to close off the early years. one thing that I observe, which you are also an investor in Arena.

OpenRouter vs. LM Arena

Alex Atallah [00:47:10]: Right.

Swyx [00:47:10]: And we talked about Midjourney having that feedback loop of, A, B, C, D, and choosing that very. being very important. And you understand the flywheel. So how come you didn’t build Arena, and how come Arena didn’t build OpenRouter?

Anjney Midha [00:47:23]: Well, Arena started before OpenRouter, right?

Swyx [00:47:27]: They had the school project

Anjney Midha [00:47:29]: Yeah, LM

Swyx [00:47:29]: And then it became a company.

Anjney Midha [00:47:31]: LM Arena, yeah.

Swyx [00:47:32]: So, but, and I know you had some Arena experiences, like the up comparison type things.

Anjney Midha [00:47:37]: Yeah.

Swyx [00:47:37]: But you never really went as hard as Arena did.

Swyx [00:47:40]: And,

Anjney Midha [00:47:40]: In doing up experiences?

Swyx [00:47:42]: Yes. And LM Arena did have a router project based on LM Arena ELOs, which they never commercialized.

Anjney Midha [00:47:48]: It’s hard to do a company that does both because one company is taking data and selling it, and the other company really can’t by default. So, I think there is, like, a branding reason that there are two companies here. like, when you set up OpenRouter, there’s no training, there are no prompts, right, aside from what your provider policy set. Like, OpenRou- like, OpenRouter can’t see your prompts or completions. If you want to see that as an org, you have to opt into it and enable it. And so we’re, like, pretty conservative and careful about data policy and security. And privacy. And LM Arena is like, their business model is like oriented around the labs and,

Swyx [00:48:34]: Because they give it for free, right? You don’t give it for free to give it for free.

Anjney Midha [00:48:37]: Yeah.

Anjney Midha [00:48:38]: But we do give some. We like have free endpoints too, but like those free endpoints, we, I think we’re not collecting any prompts. We’re not like monetizing the data unless you, opt into it for some reason.

Alex Atallah [00:48:48]: This comparison. you’re not the first person to ask me this, and Alex knows this, but I was the interim, like the founder, like first CEO of Arena for the first five months when, and we were helping Anastasios and Waylin spin out of Berkeley. And, I did invest in that before, OpenRouter, but it was very strange to me the comparisons that outside, folks would make between the two projects because the missions were completely different. The founding entity for Arena, we called it the AI Reliability Institute because it was there as an eval service. Like the data, so to speak, that they were originally, offering the labs was how do you make the evaluation of models more reliable than like the state of the art at the time, which was like really just finger in the wind.

Alex Atallah [00:49:38]: That’s what Anastasios and Waylin’s PhD work was as scientists at Berkeley, was on statistical methodologies for correcting, eval estimates, based on like intrinsic biases and how you collected the data.

Swyx [00:49:54]: Yes.

Alex Atallah [00:49:54]: And

Swyx [00:49:54]: Style control.

Alex Atallah [00:49:55]: Style control and stuff like that. And which is very much like a, hey, how. If you’re a scientist and you’re trying to. the highest expectation customer for Arena was always like a training and, like a researcher at a lab. Whereas the highest expectation customer from my perspective that Alex like really understood and was the mission was to serve was like a developer, right? Who then takes the result of the research and then produces an application that’s deployed to the world. It was a completely different problem and person that these two teams were focused on. And so from the outside in. I don’t know if you remember this, but I have a distinct memory of a few weeks before we did the term sheet, together for OpenRouter, I’d given you a call because we were trying to get a pooled data set together from OpenRouter and from Arena to, create like an open source repository of prompts. these projects were so different in their goals that it was totally normal to me to be like, “Oh, yeah, let’s call Alex and see if he’d want to team up on pooling data,” because they’re so different. We need. We don’t have that data at all. We. Like, we didn’t have API prompts. We didn’t, we didn’t have like what developers want to do with the models, which is very different from what researchers inside a model lab want to do before releasing the model.

Swyx [00:51:15]: Yeah.

Alex Atallah [00:51:15]: Does that make sense? And so to this day, I think you see that this difference, even though at a 30,000-foot level you could. I guess you could conclude that Arena and OpenRouter are adjacent, but, the roadmaps, the missions and so on at the time at least were like in very different directions.

Swyx [00:51:36]: That ideal customer, I get. I totally get that.

Alex Atallah [00:51:39]: Yes.

Swyx [00:51:39]: As a founder, I want to own everything, right?

Alex Atallah [00:51:41]: That’s possible.

Swyx [00:51:42]: Like this is clearly an adjacency that I’m like gonna explore that.

Anjney Midha [00:51:45]: Own everything meaning like you don’t know what to do yet, so you wanna like make sure you catch PM

Focus, Anthropic, and Roads Not Taken

Alex Atallah [00:51:51]: No, I think what he

Anjney Midha [00:51:52]: As quickly as possible.

Alex Atallah [00:51:53]: You want to own the entire infrastructure space, and so you expand to whatever demand you can capture.

Swyx [00:51:58]: You want to have a play in each end.

Alex Atallah [00:51:59]: Yeah, I think that’s, that’s hard, in reality, because serving multiple customers is difficult.

Swyx [00:52:05]: Clearly, this is the one focus, right?

Alex Atallah [00:52:08]: Yeah.

Anjney Midha [00:52:08]: Yeah. I still think even in the age of AI, like focus is,

Alex Atallah [00:52:12]: Is critical

Anjney Midha [00:52:13]: Underrated and critical, not just because you end up with a better product by focusing your humans on it, but also because the world knows what your focus is.

Alex Atallah [00:52:22]: One thousand percent.

Anjney Midha [00:52:23]: The world can map like, “Oh, I have this issue. Which brand out there is going to help me with that issue? This is the brand that’s known for that focus.”

Alex Atallah [00:52:31]: Yes.

Anjney Midha [00:52:32]: So like if I want real attention on this issue, like this really matters to me, I should go with the brand that cares the most about it.

Alex Atallah [00:52:39]: To underscore Alex’s point about how important focus is, in the early days of Anthropic, it was not easy to. Like people think that the early days of Anthropic were like super easy because they were on their 3 guys who left, but it was very competitive. The company was starting 10 billion dollars behind OpenAI, right? And so to get to the frontier, like the big question was, what do we want to be known for? What’s the mission? And the mission was AGI pair programming. And so to the, exclusion of all kinds of other things that were really shiny at the time, like image models and video models that were getting lots of, momentum, the Anthropic team was like, “We just got to focus on coding.” Like that is the core capability that we’re focused. And today you can see the results, right? It’s a trillion-dollar company within five years. And that focus, I think, like the high. The focus on who your highest expectation customer is and how you exceed their expectations, because exceeding anyone’s expectations is hard, and doing it for multiple like customers is so even more difficult, is part of the reason why OpenRouter succeeded and Anthropic as well.

Anjney Midha [00:53:39]: Was the focus on coding that early, though, or did it come later?

Alex Atallah [00:53:42]: Literally from day one it was AI pair programming is. Responsibly commercialize an AI pair programmer was the seed memo. That was when I invested, right? We like refined that memo a lot. Well, you got to ask Dario and Tom for permission on that.

Alex Atallah [00:53:57]: But it’s an extraordinary piece of writing that they had put together. And AI, commercializing it. Responsibly commercializing an AI pair program was the mission, from day one. And I would say there were maybe like a couple moments in the company’s history where like they did experiments to see if like little detours made sense, like a general chatbot, like Claude.ai when ChatGPT was really taking off. But, at the end of the day, but especially once, they got their like significant training compute online, I think like the. All the main evals at the company, for example, have always Coding evals, long horizon agentic programming. from day one, that was always the plan.

Anjney Midha [00:54:34]: Because when, like, Claude Instant came out and Claude 2 came

Alex Atallah [00:54:38]: Yes

Anjney Midha [00:54:39]: I remember the marketing mostly being focused on pros. Like, this

Alex Atallah [00:54:43]: Yeah

Anjney Midha [00:54:43]: Could write better

Swyx [00:54:44]: Yeah Long context. It was the first of its kind.

Anjney Midha [00:54:47]: Long context,

Swyx [00:54:49]: This directly affected me ‘cause I built something on that. Yeah.

Alex Atallah [00:54:51]: What did you make?

Swyx [00:54:52]: A small developer, which was my Devin before Devin.

Alex Atallah [00:54:54]: Oh, yeah. Yes.

Anjney Midha [00:54:55]: Yes.

Alex Atallah [00:54:55]: Small.

Swyx [00:54:56]: Yes. and, so I think, like, there’s, there’s all that really, like, good, like, focus is another thing - That is a question that people do wanna ask. you could have built any other things. Like, and obviously OpenRouter was working. were there other ideas that you wanted to pursue that you turned down? just the paths, roads not taken.

Anjney Midha [00:55:16]: We made a couple prototypes for things that we didn’t launch. One was a tuning model as a service.

Swyx [00:55:23]: Yeah. Lots of that with OpenPipe and, all those things.

Anjney Midha [00:55:25]: But it - It was in a very consumery form factor, where you would give us a YouTube video or two or three. We would then extract all the transcripts from it and try to tune a model to talk like the person in the YouTube

Alex Atallah [00:55:40]: Yeah

Anjney Midha [00:55:40]: Or the people in the videos that you sent. So, like, a really easy way of creating a tuned model based on, like, some videos that you like.

Alex Atallah [00:55:48]: That would be so useful.

Anjney Midha [00:55:50]: We,

Alex Atallah [00:55:51]: No

Anjney Midha [00:55:51]: We made it too. It was

Alex Atallah [00:55:53]: You don’t think so?

Anjney Midha [00:55:54]: It was, it

Alex Atallah [00:55:55]: And nobody used it?

Anjney Midha [00:55:55]: It - We didn’t like, test it with that many people because the model marketplace was our main focus, and it was, like, growing, and we were building more conviction in it over time.

Swyx [00:56:09]: Just, you

Alex Atallah [00:56:10]: Yeah. Why,

Swyx [00:56:10]: As a creator

Alex Atallah [00:56:11]: Yes. I’m a creator.

Swyx [00:56:11]: Have you been pitched many, like, - I have five hundred hours of recorded voice of myself.

Alex Atallah [00:56:17]: Right.

Swyx [00:56:17]: Make a thing of you, charge access to it. it works for OnlyFans, doesn’t work for

Alex Atallah [00:56:23]: I see

Swyx [00:56:23]: As regular people. I think - this is mostly, - It’s just a glorified RAG bot.

Alex Atallah [00:56:28]: Right.

Swyx [00:56:29]: Whether it’s in the weights or it’s outside the weights, doesn’t really matter. You’re just doing RAG on the videos, and people ultimately always just wanna find the source video, that directly answers it.

Alex Atallah [00:56:36]: Oh. my use case was mostly to practice - - with myself ‘cause I often like to see what. Like, the way I practice for a job interview or if I’m hiring a candidate or public speaking or whatever is I wish there was, like, a good

Swyx [00:56:48]: Yeah

Alex Atallah [00:56:48]: That I could, like, critique ‘cause it’s kinda hard to pull yourself out. I would never get. I would never offer it to other people as a service.

Swyx [00:56:54]: I wish there were, like, pick your top five mentors that, then talk to them instead of talking to yourself.

Alex Atallah [00:56:57]: That’d be cool too, yeah.

Anjney Midha [00:56:58]: That was, that’

Swyx [00:56:59]: That’s the creator AI. That’s a replica.

Anjney Midha [00:57:01]: And that was the use case we were aiming at.

Alex Atallah [00:57:02]: I see.

Anjney Midha [00:57:03]: Is like, you wanna create an experience

Swyx [00:57:06]: Like AI Steve Jobs and.

Anjney Midha [00:57:07]: And AI Steve Jobs was the initial use case.

Alex Atallah [00:57:11]: That’s a,

Anjney Midha [00:57:12]: Even though it’s not allowed.

Alex Atallah [00:57:14]: That’s a, that’s a common prototype, yeah.

Swyx [00:57:15]: Talking about adjacencies, tuning as a service, as part of the router service is something that I would typically think about as well, right? Like, why don’t you do that? ‘Cause if people are running already their inference through you, store everything, log everything, tune to a smaller model that is cheaper, faster, all these things that’s within your control, right? you didn’t do that, but, like, other people would have pitched that in the general state of a infra startup.

Anjney Midha [00:57:37]: Yeah. Yeah.

Alex Atallah [00:57:37]: I think you were just maybe a little bit early ‘cause today that’s an extraordinarily growing segment. Like, from Mistral, where they do a lot of enterprise deployments

Fine-Tuning as a Service and Infrastructure Adjacencies

Anjney Midha [00:57:44]: Right

Alex Atallah [00:57:44]: And stuff and tuning as, custom models for ASML or whatever. And often

Swyx [00:57:48]: But not as a router. They’re, they’re just like, “I come to you because I like your Mistral models. I want custom Mistral model,” right? It is not, “I want, to run all my OpenAI prompts, - store all my results, and then just move off of OpenAI.” Right? They’re not doing that.

Alex Atallah [00:58:01]: As a, as like a way to export off of dependency on a Frontier lab, I have not seen that yet. Yeah.

Swyx [00:58:08]: Right.

Alex Atallah [00:58:08]: Which was your vision.

Swyx [00:58:09]: Is efficient to do.

Anjney Midha [00:58:10]: We decided. Really, we, like, leaned into our focus and figured that, like, there aren’t. Like, we just saw the ecosystem develop over time. All these inference providers that do wanna help companies do that, - Like, it makes sense for us to partner with them and to, like, give users lots of choice and to, like, figure out what makes them, what gives them competitive advantages. It’s, it’s a whole new business and there’s, there’s value in being a neutral marketplace that just like, works with those companies.

Alex Atallah [00:58:45]: Could you share a little bit, to Sean’s point, like, how you prioritized. What are some ways you prioritize features? ‘Cause you’ve always done it so elegantly that I never. it just happens, and you make all the right decisions that always have product-market fit from the outside looking in. But consistently, you seem to have prioritized, a lot of hit features that worked. And maybe I have a sample set bias or whatever, but Sean’s question

Swyx [00:59:06]: Can you list what you think hit features worked?

Alex Atallah [00:59:09]: Oh, the leaderboards.

Swyx [00:59:10]: Leaderboard, okay.

Alex Atallah [00:59:10]: Yeah. like, from day

Swyx [00:59:13]: That’s charting, right? That’s the feedback loop.

Alex Atallah [00:59:14]: Charting, BYOK.

Swyx [00:59:15]: But, like, he had, like, ins. he had, like, And I think there was a whole thing I wanna get into about, like, completions versus

How OpenRouter Prioritizes Product

Alex Atallah [00:59:22]: Yes.

Swyx [00:59:23]: Check completions versus completions. And then also, let’s call it, like, the rise of the reasoning models and how you deal with those, multimodality, all those things, right?

Alex Atallah [00:59:31]: Yeah. BYOK.

Swyx [00:59:32]: BYOK, yeah.

Alex Atallah [00:59:32]: That was a huge one.

Anjney Midha [00:59:34]: There’s one I. Like, I think it was in early 2024, very early 2024, we thought it might be interesting to fuse the results of multiple models together, and we launched a prototype called MOM, Mixture of Models, that let you, like, pick a couple models, or we’d pick them for you, and then it would fuse the results together at the end, and it would show you all the intermediate results in this, like, big Kanban looking product.

Mixture of Models and Model Fusion

Swyx [01:00:05]: What does the fusion at the end, another model?

Anjney Midha [01:00:07]: Another model. The,

Swyx [01:00:08]: The smartest of

Anjney Midha [01:00:09]: The smartest

Swyx [01:00:10]: Of the set

Anjney Midha [01:00:10]: Of the three, of the set.

Swyx [01:00:12]: Okay. So this is like a council idea?

Anjney Midha [01:00:13]: Yeah. It was a model. It was like a very early LLM council.

Alex Atallah [01:00:16]: This is a agent swarm as, like, they would call it at one of the Frontier Labs, in the early days?

Anjney Midha [01:00:23]: Yeah, like some of those ideas are, like, going the right direction, but the devil’s in the details.

Swyx [01:00:27]: Yeah.

Anjney Midha [01:00:27]: There’s a lot of, like, product refinement needed to make them really work. they take your focus away

Swyx [01:00:34]: Right

Anjney Midha [01:00:34]: Whatever else you have going on. And there’s a lot of, like, community building and learning that you need to do. And the technology might be too early. So there are - like, all kinds of reasons they might go wrong. And in our case, the technology was a little too early. In other words, the fused result was a little bit

Swyx [01:00:53]: Right. Like a Frankenstein

Anjney Midha [01:00:54]: Sometimes the same as the best model that was being used to fuse because the best model was so far ahead of options two and three at the time. over time, the top three or four LLMs have gotten closer together, still neurodivergent, but, like, all capable of inserting, like, pretty interesting ideas. Like, RL has like, expanded the surface area of creativity for machine learning researchers within each lab, and so they can, diversify the reasoning power of different models more effectively. At least that’s my theory for

Swyx [01:01:29]: Yeah

Anjney Midha [01:01:30]: Fusion - it, like, works better than it used to, but early twenty-twenty-four. And, so the technology was a little bit too primitive. The form factor was not right, and so we would have had to go through a couple more iterations. And so we decided to just delete all the code. And, then years later, middle of twenty-twenty-six, or early twenty-twenty-six, we’re like, “Let’s bring it back.” Like, the research is looking kinda promising for fusion. The models now have, like, two, three, four top frontier models that are all really good and, like, I’m, I’m frequently trying to, like, consult multiple models to get the best results. Like, and then I ran a little personal experiment where I was like, “I’m gonna, like, do a, an architecture plan for a code change. I’m gonna give it to all the models. I’m gonna fuse the result, and then I’m gonna ask all the models if the fused result is better than the individual result each model came up with.” And they all said yes, that the fused result was better. And this happened a couple times, and I was like, “Okay, spot check, pretty good. We should, like, benchmark this.” And that’s how we built fusion.

Revisiting Fusion as Frontier Models Converge

Swyx [01:02:40]: Yeah. And it came on your Fable, so you were like, “This is Fable level.”

Anjney Midha [01:02:43]: Yeah.

Swyx [01:02:44]: Let’s start leading up to this year, which we haven’t gone to this year. can you mark out the main milestones in the journey? I think, it seems like your promise, was, routing. You decided the business model very early.

Swyx [01:02:59]: You take a cut. And, like, what are the major milestones that, inflect the growth, right? Like, you’re, you’re growing, like, 9% week on week now? Is this the official number?

Anjney Midha [01:03:10]: In terms of token volume, I think that sounds about right, yeah.

Swyx [01:03:13]: Yeah. So just, like, can you mark out, like, the brief history of OpenRouter up to, the acquisition? Let’s, let’s call we’re, we’re just, we’re just, talking about, people are, - you have a your birth moment with, the Mistral stuff where people are really competing. You have your state of AI thing where,

Anjney Midha [01:03:32]: Yeah.

Swyx [01:03:32]: It’s very cute. You have a hundred trillion tokens, ha, ‘cause now you’re doing ten a week, .

Anjney Midha [01:03:39]: Yeah. We’re doing ten a day.

OpenRouter’s Growth Inflections

Swyx [01:03:41]: Ten a day now?

Anjney Midha [01:03:42]: Yeah. More.

Swyx [01:03:43]: So yeah, you do this in ten days.

Swyx [01:03:45]: Like, what are the major end points there? I just wanna. Like, there’s a smooth curve, but, like, you feel the inflections.

Anjney Midha [01:03:51]: A lot of this is oriented around model launches. we had, a huge focus on pros all the way up through May of twenty-twenty-four, because coding was just not there, and no apps were able to build much on top of it. So, a diversity in models, but not a wide diversity and not a wide diversity in use cases. Dream Tavern was one of our top apps at the time. The creator of Dream Tavern now runs product at Cognition, Devon. - Then - In the middle of twenty-twenty-four, we saw Claude 3.5 Sonnet. That came out, incredible leap forward in coding, and we saw the dynamics of, like, apps building on top of us change. we saw a huge surge in volume in, like, users, using OpenRouter. And this is when I think people started to look at the, like, money that they were spending and get a little bit like, “Whoa, what’s going on? I might need to, like, think about, like, more efficient but equivalent models.” And shortly after that, I think it was after Sonnet three five, Mixtral 8x7B came out, and everyone was like, “What? This is the model.” Like, the OpenWeights community delivered. And so it was really good timing from Mistral.

Swyx [01:05:17]: All of Anja’s portcos are just helping you out.

Alex Atallah [01:05:21]: It takes an ecosystem to grow an OpenRouter?

Anjney Midha [01:05:24]: Yeah, that was the. Yeah, it was. It like, it was the, like, this early ecosystem, it was like a swing action where, like, model labs would come up with some frontier innovation. Like, usage would surge. Then users, look at their invoices 30 days later and like, “Whoa, what’s going on here?” And then OpenWeight models would deliver, like, a, like, effective options two, three months later. We saw that happen several times.

Swyx [01:05:54]: By the way, one

Anjney Midha [01:05:55]: Yeah

Swyx [01:05:55]: One thing you also did with the coding agents was that you broke out which are the top coding agents, and they love that. They love that leaderboard. The Klein versus the Rue code versus the what have you.

Anjney Midha [01:06:04]: Yeah. Like, Klein was, like, the top of our leaderboard at the time. We, We then, at the end of. And I’ll skip forward a little bit. The end of twenty-twenty-five, there were quite a few coding apps on the leaderboard, but they were all IDs or, terminal-Agents. And at the end of twenty-five, we saw OpenClaw appear. And OpenClaw was, like, particularly interesting because, one, it was like a new form factor that, like, brought in a new type of user, not just a developer, but like a productivity or a, like an internet creator came to AI for the first time. And it also had an interesting architecture where it was, like, calling your chosen model for these heartbeats to see if it was still alive in addition to using the model for real tasks. And the heartbeats are like, they’re kind

OpenClaw, Hermes, and the Auto Router

Swyx [01:07:02]: Fréquence.

Anjney Midha [01:07:02]: You don’t wanna pay a lot of

Swyx [01:07:03]: Every thirty minutes

Anjney Midha [01:07:04]: To do a heartbeat.

Swyx [01:07:05]: Yeah.

Anjney Midha [01:07:05]: So, the auto router that we provided was really useful to this, like, wide range of users all of a sudden. And so we just saw it rocket exponentially, and then we saw, like OpenClaw just blow up and a couple other, apps lean into that new paradigm and do something similar. Hermes came out and really leaned into things like the auto router and built, like, a really good community and leaned into, like, skill management and making it really easy and effective for people to, like, set their memory in the agent

Swyx [01:07:44]: Yeah.

Anjney Midha [01:07:44]: And build really good skills.

Swyx [01:07:45]: Which another thing you never did, memory skills, sandboxes, all these, like, adjacent things you could have done.

Anjney Midha [01:07:52]: Could have, but It’- I think,

Swyx [01:07:54]: It’s hard to bet.

Anjney Midha [01:07:55]: They’re also - There are things that developer-- that really matter for, like, the developer use cases that were coming out at the time. Like, developers wanted to architect those things.

Swyx [01:08:05]: Right.

Anjney Midha [01:08:05]: Those were kinda critical to building a good user experience. It’s really-- It was, like, - It’s been hard for companies to find abstractions that work for all developers on the memory layer. It is, it - Yeah, there are some, like Mastra has done a pretty good job, for example. But, like, developers have, like, lots of varied preferences for them. And then we - - the way our leaderboard has changed over time is like a movie of how the AI space has changed over time. If you just like, go to the Wayback Machine and look at the rankings leaderboard and the apps leaderboard over time, it shows you, like, what’s happened in AI over the last couple of years.

Swyx [01:08:48]: To me, the coming of age moment was, Andrej Karpathy was like, “I no longer read Local Llama ‘cause, like, I just go to OpenClaw-- OpenRouter’s leaderboard.”

Leaderboards as a Map of the AI Ecosystem

Swyx [01:08:57]: Which I remember that. Yeah. I think he probably, like, said, like, “Sorry, guys, I’m gonna send a bunch of traffic to you.”

Swyx [01:09:03]: So I also wanna bring it into the Stripe, thing.

Why Stripe Acquired OpenRouter

Swyx [01:09:07]: How does that conversation start?

Anjney Midha [01:09:09]: We had this longstanding relationship with Stripe, though, from, like, many different projects that we had worked on with them. We invest, a lot of effort in countering abuse,

Swyx [01:09:24]: Token fraud.

Anjney Midha [01:09:24]: And token fraud.

Swyx [01:09:26]: Can you give some numbers just - so people understand?

Anjney Midha [01:09:29]: I think I, like, I posted about this. We blocked 10x as much dollar volume last month as the month before. And the types of token fraud are diversifying quite a bit. there are, like, fraudsters going after typical stolen credit cards, but there are also, people trying to resell traffic against the terms of service. There’s, like, hacked accounts. There’s people who just lose - like, their whole company is compromised, and they don’t even realize it, and we help them, like, regain control and detect it. There’- There are accounts that are, like, reselling inference on the side. There’- There are accounts that are dealing with, a, like, an accidental runaway agent, and they don’t realize it. Not a hack, but it’s something that blows up and the company doesn’t want it. And so our trust and safety team, like, works a lot on all of these, like, categories of problems and helps block it and detect it. And so we’ve built these. we have models around them. We - We worked closely with Stripe for a while on this, and I think it’s gonna become a huge problem in the ecosystem. Like, we’re already seeing a lot of companies start to see these fraudsters, like, spread and look for other ways other than OpenRouter to other fraud vectors. And if you’re making a gateway or selling, like, generalized inference, you are a target for fraud. If you’re selling very discreet, like, intelligence products that are, like, doing something pretty specific, but not, like, just reselling inference with some added capability, then you’re way less likely to get these fraudsters. So - I think we’ll see companies also move away from just reselling inference with some like, added capability and move towards like, discreet tasks and charging for those tasks and charging for those enhancements and letting people bring their own inference, like, in a party way.

Fraud, Abuse, and the Emerging Token Economy

Swyx [01:11:39]: Whoa. Okay. and yeah, obviously you would power that.

Anjney Midha [01:11:44]: Right.

Swyx [01:11:44]: But you - People pay, for outcomes Or per task?

Anjney Midha [01:11:48]: I think people will pay. I think, like, the Datadog pricing page is a good look at, like, the future to come. It’s like companies, like infrastructure companies will, like, charge for different types of events that they’re providing, and there’ll be lots of, like, continuous pricing models that look like that. And of course, there will be, like, if you go down, towards consumer apps, simpler pricing, more subscriptions, fewer events to worry about, and ones that, like, are not. Focus on just adding a markup on top of inference.

The Token Economy and Security at Scale

Swyx [01:12:28]: Yeah.

Anjney Midha [01:12:28]: Not just because fraud is hard, but also because the pressure from the labs and from - like, good inference providers to, like, do a commit and then bring your inference elsewhere is gonna be very high.

Swyx [01:12:44]: Any comments?

Alex Atallah [01:12:45]: Two. One, I think Alex has done a very eloquent job of describing something, counterintuitively I knew would be a thing at scale, like four years ago because of Discord. And the particular experience that taught me this was, as we started scaling Midjourney, - one of the primary ways that we used to give away or, like, get people to try Midjourney early on to get to their first ten generations. Because, ten generations - ten images generated was roughly the magic moment activation point we found. Like, once you’d done ten, you were like, “This is extraordinary.” but for that week, so we had a free trial with Midjourney. And one day I woke up, because I was the head of platform and had to monitor, I had all these dashboards, and I had, like, three missed calls from David. And it turns out, like, there had been this flood of new users overnight. And we were like, “This is great.” And he was like, “No, we shut down the free trial.” And I was like, “Why is that?” and he said, “I want you to look at the geolocation IP addresses.” And somebody in China had started to resell Midjourney free, subscriptions with the free trial as a way to, like, you - It was fraud abuse, right?

Swyx [01:13:54]: Even for a specialized model like Midjourney.

Alex Atallah [01:13:56]: Yeah. And that was an application. So this idea - I think the big picture realization I had back then was, hey, there’s a new type of unit of value that’s being streamed across the internet called a token.

Alex Atallah [01:14:11]: And over the next ten years, the entire internet value chain was going to have to deal with the fact that, like, the more valuable tokens got, The more bad actors are gonna go to try to get their hands on those tokens. And anytime you scale something and the payload gets more and more valuable, More bad things, people try to get access to that value. And so it was very obvious to me back then. And so, look, to this day, I don’t think there’s a free turn. Like, I don’t think Midjourney’s ever turned on the free trial since then, because it was really not an easy problem to solve in terms of trust and safety. that’s why I - started teaching the class Security at Scale at Stanford. Like, it was like one of - that and the Anthropic learnings, to me, it was clear that the need for security at scale is gonna be enormous a few years from then. Because if you just do the math, right, think about, like, if we’re. online payments, has started roughly in the eighties and nineties, right, and grew to over a trillion dollars over the next ten years, and we needed to build entirely new payment solutions to deal with online fraud. where we are today is roughly there on tokens, but over the next even five years, we’re expecting the token economy to get to, like, roughly 5 trillion dollars. And over the next ten years, I’d be shocked if we weren’t at 10 trillion dollars of token flow. And so if we were starting to see such aggressive abuse and fraud at subscale, Midjourney, remember Midjourney at this point was, like, less than three $100 million revenue run rate a year.

Alex Atallah [01:15:44]: I just realized we were gonna need, like, entirely new, Like, systems to deal with the fraud that was gonna happen for trying to get into the token flow. And so, - I, - I forget the board meeting it was when you brought up that, Stripe wanted to partner up, and it made so much sense to me because Stripe Radar. When I was at Kleiner ten years ago, we invested in Stripe, and the whole pitch that, Patrick and John communicate so eloquently was like, “Hey, unlike traditional payment tools like Braintree that do a day verification, like KYC and AML to get the fraud out of the way, we just bite the fraud cost upfront as customer acquisition cost and - tell a developer, like, just use five lines of code, and we start accepting your payments in five minutes. And what’ll happen is over time, we collect all this data on the developers.”

Swyx [01:16:31]: Cloudflare model.

Alex Atallah [01:16:32]: Is the Cloudflare model, right? And they did. Five years later, they launched Stripe Radar, and Stripe really today is a security company. That’s the real. People think it’s a payments company. No, the reason. There’s lots of other payments providers today that give you, like, cheaper payments transmission. But the reason Stripe keeps, being the dominant one here and Adyen and Europe is because they have extraordinary fraud detection that they’ve built, - over the years.

Swyx [01:16:52]: It’s the same story with Elon and Max Levchin

Alex Atallah [01:16:55]: And affirm, yeah.

Swyx [01:16:56]: Yeah.

Alex Atallah [01:16:57]: So, I think the story shows up over and over again, where every time you have value streamed across the world in large amounts, you need new protection and security infrastructure to fight, to keep the bad guys out and allow the good people to, like, have their transactions happen really fast. And so I think, - this is why - from my perspective, like, the Stripe and OpenRouter story is a security story for the internet ecosystem, for the frontier AI ecosystem. Without a partnership like that, it becomes very hard to defend the quality of experience and the speed and all the good stuff without letting the bad guys get in the way. the second is that, there’s this underappreciated thing about, like, the fact that you need to. Like, - all the bad things that Alex described as being perpetuated by humans right now is going to be perpetuated by AI agents over the next ten years.

Swyx [01:17:46]: Oof.

Alex Atallah [01:17:47]: Right? So think about the, like, recursive scale we’re about to see of bad actors. It’s not just bad human beings, it’s, it’s all the bad agents that are gonna be attacking the token flow. And there’s. It’s very hard if you’re a researcher and at an AI lab to reason about that problem because the only data you have is how agents you’re training are going rogue. But that’s just a fraction of all the bad behavior on the internet that we’re gonna see. And so what you need is defenders, new sheriffs in town, which cowboy hats, that can see all the bad behavior from AI agents across the ecosystem, from different model labs and different trained deployments and different developers, and take all of that data and say, “We’re gonna build a shield for the entire token economy.” Because without that, the amount of fraud we’re gonna see of this 10 trillion dollars in GMV and global GDP growth is, like, a huge percentage of that, I think, is going to be fraud, abuse. And we might never get there if people just don’t trust. Tokens, right? and I don’t think this infrastructure exists. So you have your work cut out for you with, at Stripe, but I don’t think people have realized the scale at which agents, agent, agentic fraud, like bad behavior perpetuated by AI agents is about to hit us like a tsunami.

OpenRouter + Stripe: What Changes Next

Swyx [01:18:58]: Yeah. there’s a lot to dig into there. I wanna give you the last word. We do have to wrap. what can people expect from OpenRouter and Stripe?

Anjney Midha [01:19:07]: I think this is a really good way for us to accelerate market and, to go upmarket more quickly. It’s also, as Ansh eloquently described, this is, there’s a really clear better together story here when it comes to improving trust and safety and making it really easy to, like, accept tokens and let people bring their own inference to your app and to help developers just, like, build on top of inference, going forward. We have a really strong brand with OpenRouter, and we’re keeping the brand. So, like, OpenRouter, like, as a product and the roadmap and the name and the brand, like, is staying the same. And so what, like, you should expect, in the next six months is that most things will be like what we would have done had we been independent, except everything will be moving faster. And that’s like our, term goal. Longer term, hopefully I can comment on it soon, but I can’

Closing: Building the Infrastructure for the Token Economy

Anjney Midha [01:20:11]: Now.

Swyx [01:20:11]: Okay. Well, we’ll hopefully do a follow-up at some point, but thank you for being so generous with your time, and, congrats on the partnership. this is one of the most beautiful bromances I’ve seen in AI.

Alex Atallah [01:20:22]: Just starting out.

Swyx [01:20:23]: Starting from Stanford

Alex Atallah [01:20:24]: Just starting.

Swyx [01:20:24]: To here.

Alex Atallah [01:20:24]: Yeah. Lots more to do.

Anjney Midha [01:20:26]: Yeah.

Alex Atallah [01:20:26]: Lots of sheriff, policing to do of the, of

Swyx [01:20:29]: Yes. The cowboys in town.

Alex Atallah [01:20:30]: Of the token economy. We need We need new sheriffs for sure.

Swyx [01:20:33]: Yeah. Awesome. Thank you.

Anjney Midha [01:20:35]: Thank you.

Earlier this month, world model company Runway introduced GWM Worlds 2, a research preview that “turns high-fidelity video and audio generation into real-time interactive simulation.” Runway calls this an “autoregressive diffusion” model; with autoregressive describing how it generates over time.

One new feature in particular caught our eye: WorldPrompt, a proposed input format for specifying a generated world and the actions within it. It allows you to fix some aspects of a simulated environment — including the first frame — and then create a series of timestamped events. The events, or actions, can even be prompted in real-time.

To understand the implications of WorldPrompt, we spoke to Kamil Sindi, Runway’s CTO, and Robin Kahlow, its Principal Research Scientist for generative video and multimodal AI. We also have exclusive comments from Anastasis Germanidis, co-founder & co-CEO of Runway, courtesy of a podcast swyx and Vibhu did with him.

Who’s building real-time interactive world models?

First, some context about world models that can generate interactive video and audio in real-time.

Runway is reportedly valued at $5.3 billion, based on its most recent fund raise of $315 million in February. Its first release, GWM Worlds, was launched last December.

Alongside Runway, there are several other notable projects in this domain: Google DeepMind’s Genie 3 (which also generates at 720p and 24 fps), Odyssey-2 Pro, and World Labs’ RTFM (Real-Time Frame Model). We’ve summarized their differences in the following table:

Given the complexity and massive latency demands of real-time video and audio generation (which we’ll get into below), all of the projects listed above have limitations. For instance, Google notes that Genie 3 “can currently support a few minutes of continuous interaction, rather than extended hours.”

But as our interviews with Runway show, real progress is being made.

The central idea of WorldPrompt

WorldPrompt, a new feature in GWM Worlds 2, helps differentiate Runway from its competition. You can think of it as a control layer for characters, cameras and the environment. As Kahlow put it, it’s a way to “control all the different subjects in the world” — similar to a computer game.

“Like, if there’s an NPC [Non-Player Character] somewhere, the NPC might walk up to you and say something. So you could achieve the same thing with this kind of model, where you can have very detailed control over everything in the scene.”

As the name suggests, WorldPrompt is a prompting mechanism — not a programming language. So, unlike virtual world games like Minecraft or Roblox, GWM Worlds 2 doesn’t offer scripting capabilities or the ability to control state. But there’s a power to that, as Sindi pointed out.

“You can create promptable worlds on-demand with video and audio in sync, across all these different domains and environments. That’s not a distant-future hypothetical thing,” he said.

But there are also limitations to prompting a world model. We asked how reliably the model would follow an instruction to create, for example, a law of gravity or a certain ability in a character?

“Yeah, so it’s a research preview,” Kahlow replied. “So it’s not perfect, of course, and there are still flaws. It really depends on how difficult the action is. I would say movement works quite reliably.”

Sindi added that more training plus scaling the data and models is resulting in “better following.”

How a video model becomes a real-time runtime

Despite the current limitations of GWM Worlds 2 — especially if you compare it to pre-designed and scriptable worlds like Minecraft or Roblox — the true promise of world models like Runway is that they’ll eventually lead to fully self-generated, real-time games and experiences. Which is an extremely hard engineering problem, as Kahlow reminded us.

“There are two challenges. One is making the model not generate a whole clip at once. So instead, you want it to generate frame by frame while you’re looking at it. And the other challenge is actually making the generation fast, so you can play it in real time.”

GWM Worlds 2 offers real-time interactive worlds streamed in continuous 720p video at 24 frames per second (fps) and audio at 48,000 Hz.

Runway achieved this firstly by taking its foundational audio-video generation model and fine-tuning it to the new WorldPrompt format, so the model can follow that. It then post-trains the model to generate autoregressively.

“And after that, we work on making it real-time through distillation methods,” Kahlow added.

Co-CEO Anastasis Germanidis offered more technical details in our podcast with him. He told us that the process starts from “bidirectional diffusion that basically generates an entire video at once and [makes] it autoregressive.” This allows the model to “generate one frame or a few frames at a time.”

Germanidis described two possible forms of distillation in order to make it real-time: distilling a larger model into a smaller one or reducing its diffusion steps. As a general example, he said a model might go from around 50 denoising steps to four, with some quality loss but potentially comparable results.

The challenges of real-time generation

Germanidis admitted that there were issues with how it generates real-time interactive video.

“The biggest challenge with autoregressive models is error accumulation,” he said. “You’re feeding generated frames back into the model to generate the next frames, and if there are any small errors, they accumulate over time.”

Sindi told us there are also challenges dealing with “infinite generations” of content.

“There’s all these challenges around what context to keep, what to discard that’s not important. And so there’s all these optimizations we have to think about, so we’re not blowing up our GPU memory.”

Another current limitation is long-term memory. “The model does not have perfect memory,” Kahlow said. “That’s still an open research problem.”

Causality and correctness

While performance is the primary challenge for Runway at this time, its world model also has to produce plausible consequences when a user takes different actions.

Germanidis used the example of simulating football; he pointed out that online video training data contains more successful goals than failed goal attempts, so a video model might render the first more convincingly.

“If I take this action versus this action, you want it to generate equally realistic outcomes,” he told us. “That’s, I think, the big gap between video models and world models: that idea of counterfactual generation.”

Sindi told us that evaluation gets harder the more complex interactions get.

“If you have this multi-prompt, multi-character, multi-scene [environment], how do you really understand what was causal and what was not?”

To try and solve that, Runway has some automated verifiable tests. But since GWM Worlds 2 is a research preview, Kahlow noted that doing tests yourself is also advisable — “trying out your model to see what doesn’t work is really important.”

More than gaming — there are agent use cases too

Gaming is the obvious use case for what Runway is building, but there are others. Kahlow mentioned robotics — for example using a simulated environment to test how a robot works.

Another, more intriguing, use case is to use it to test agents at scale.

“Having thousands of simulated environments is much less challenging if you have a suitable model like GWM Worlds,” Kahlow said.

But how does an agent know what’s changed in the world — is there a structured state that it can read, or is it just the generated video and audio that it’s consuming and understanding?

“So there’s no structured state here,” Kahlow replied. “It’s just observing the same thing you might observe in real life, just [in this case] from cameras.”

Sindi noted that GWM Worlds can also be used for “synthetic data generation for agents.”

Finally, Germanidis suggested there’s potential to use these world models alongside reasoning models.

“You’re maybe using some reasoning [for] planning of the scene, and then you’re passing it into the diffusion head that’s actually generating the pixels.”

Anastasis Germanidis

* LinkedIn: https://www.linkedin.com/in/agermanidis/

* X: https://x.com/agermanidis

Timestamps

00:00:00 Introduction

00:05:17 Runway’s Origins and the Bet on Generative Video

00:12:23 The Stable Diffusion Story

00:18:44 Gen-2, Controllability, and the Weekend Hack

00:23:02 From Video Generation to World Models

00:28:03 Learning From the World, Not Just Language

00:35:04 Sora, Runway’s Existential Crisis, and Gen-3

00:39:39 Why Real-Time Video Is Inevitable

00:43:06 Interface World Models: Software Without Code

00:50:25 The Fully Neural Operating System

00:55:11 World Models for Robotics

01:02:32 Robot Policies and World Action Models

01:07:47 The Lucid Dream Test

01:11:41 Video Agents and Omni Models

01:23:12 Artists, AI, and Creative Workflows

01:27:14 Physical AI and the Future of World Models

Transcript

Introduction: Runway, Creative AI, and the Early Thesis

Swyx [00:00:00]: Okay, we’re here with, Anastassios from Runway, with, me and Vibhu in the studio. Welcome.

Anastasis [00:00:08]: Good to be here.

Swyx [00:00:09]: Congrats on all your success and progress with Runway. You’re opening offices all over the world. Did you envision this when you first started out?

Anastasis [00:00:16]: Not quite. I think even when we started, we had this idea that, It was more a matter of when, not if, we were seeing the early generative models of 2016, 2017, and just extrapolating, assuming, we resolution, quality increases predictably over time. There’s gonna be a point where most of content will be generated, and that was maybe the initial thesis of Runway was we will need, as a result of those generative models, rethink how creative tools are made. and as we built out the research behind, our generative models, it then became clear that they were useful far beyond that as well.

Anastasis’ Background: Art, Simulation, and Machine Learning

Swyx [00:00:57]: And it is more obvious now with, like, the real-world stuff and the world models that we’ll talk about later. I’m just kinda curious how you go from a background in, like, Zocdoc and, computer vision into Runway. Like, take us back to that early conversations with Chris and, whoever else is on your founding team.

Anastasis [00:01:14]: I was always splitting through those two worlds. One was the I had my own art practice. I was making a lot of interactive art, I think for a long time. and then on the other side, I was working in startups, and I was working as a ML engineer, as a backend engineer at different companies. I’ve always been interested in, coding and computation, and especially interested in simulation and brought it back into my early artwork as well. And at the same time, I was interested in

Swyx [00:01:43]: The personal site has a few, right?

Anastasis [00:01:44]: Yeah.

Swyx [00:01:45]: Is there one that we should pull up? Just in case there’s something that’s like. I just like to go down memory lane.

Anastasis [00:01:50]: Yeah.

Swyx [00:01:50]: Okay, what is this?

Anastasis [00:01:51]: So this was, a project that I made, I think back in 2015, where I built this software that would give, voice instructions to people in a gallery space. So it would coordinate interactions between people. And so it will first give you an identity, like you’re an, architect, you’re 30 years old, and, you like sports. and then it would match you with another person, and you have this completely generated interaction. language models were not quite there at the time, and so it was it was a mix of some templates and some, like, some Markov chain-generated text, and it would just completely simulate these small talk conversations between, everyone in the gallery space. so was always very fascinated on the one hand with, generative models and, like, the early machine learning work that was being at that time. But at the same time, there was this separate thread of simulation and what it means. Like, what can we learn about humans by creating those very simple models of their interactions and their behavior?

Early Generative Art: pix2pix, GANs, and Uncanny Valley

Vibhu [00:02:56]: Did you generate the prompts or, the 30-year-old, whatever? Was it you generating them? How’d you, how’d you

Anastasis [00:03:03]: Exactly. So the program would just generate- those, from. Yeah, a lot of it would be Mad Libs style of just

Vibhu [00:03:10]: Yes

Anastasis [00:03:10]: You have lists of different professions, lists of different,

Vibhu [00:03:14]: Hobbies

Anastasis [00:03:15]: Personality types, lists of different, ages, things like that. And then it would just combine those things together. And then maybe the next project we go is, Uncanny Valley, Uncanny Road, which was

Swyx [00:03:27]: Gans

Anastasis [00:03:27]: One of the first projects that, we built with, one of my two co-founders, Chris. This was taking, pix2pixHD, which was one of the early image-to-image models that NVIDIA released back in 2016 or 2017. and it was a model that would take a semantic map of a scene and then generate a photorealistic, let’s call it, output. very early days, so it was not very high-fidelity outputs, but it w I think was the first image-generation model that could generate at 1K resolution. And it was all trained on self-driving datasets. So the semantic categories it would support were only, things you would encounter on the road. So it would be pedestrians, traffic signs,

Vibhu [00:04:16]: Stoplights

Anastasis [00:04:17]: Bikes, stoplights. And so that was one of our first indications that we built this and people were making all this, like, very surreal imagery of, yeah, a million plus a million pedestrians or a million traffic signs or, like, gigantic humans. And it was a indication that you could take a model that was trained on this very boring dataset, essentially, of, like, not that many interesting things happen when you’re on the road, and then you can repurpose it and go very out of distribution and make something that was artistically compelling. And that was It’s a summary of the thesis of Runway in some ways, that you can take the same generative models, and if you look at them from another direction, if you build interesting tools around them and you give them to artists, they’re gonna do things that you don’t expect.

Vibhu [00:05:02]: Very cool. I like the, UX of it. You’re just given an empty canvas, try whatever, do whatever. And then the other one, like, you see everyone with wired headphones? Like, that’s, that’s a sign that it’s, it’s very

Anastasis [00:05:16]: The Apple

Vibhu [00:05:17]: Yeah

Anastasis [00:05:17]: Apple, your version.

Vibhu [00:05:17]: Original ads. Yeah. Take us to today. You’ve been doing this for seven years at Runway. How have we got to this? Like, how do we go from driving simulator data to all this? And you cover the whole stack of generative media?

From Creative Tools to a Research Lab

Anastasis [00:05:33]: Interestingly, we’re almost back in, we’re, we’re full circle. We’re, we’re now applying our models and beyond creative tools into real-world scenarios. But it was a, it was a long journey. It was very early on we realized the first version of Runway was a way to easily use the, all the open source model of the day, things like pix2pix to. and give them to artists. That was the initial idea, is those models are too difficult to use if you’re not a machine learning engineer. Like, what happens when you give them to artists? Very quickly, we realized we needed to build a research org, inside of Runway, and that happened maybe on year one. And, a lot of the mandate there was. The image-generation models of the time, the video generation models of the time, or there were barely any video generations all the time, but they were not quite there where they could be productionized and brought into tools that would be part of creative workflows. so we need to push the frontier of the research. And so maybe the first four years of Runway, research was almost happening on the background until there was a moment in 2022, with latent diffusion, with, DALL-E 2, where, there was that step function change, and you guys maybe remember around the time.

Swyx [00:06:49]: I started in this space because of latent diffusion and Stable Diffusion.

Anastasis [00:06:54]: Yeah.

Swyx [00:06:54]: Because I was like, “Wow, this is not only, like, feasible, it is doable on consumer hardware.”

Anastasis [00:07:01]: Exactly, yeah.

Vibhu [00:07:01]: I think the delta is also huge. Like, I learned pix2pix. Like, this was intro to ML, the TensorFlow, like, Jupyter, Google Colab notebooks were like this, and then you have a sudden step function change, with diffusion and whatnot. Any other ones since that. Like, there were clear examples of what early diffusion were to get to here. Any other changes in key technology research?

Green Screen, Rotoscoping, and Early Runway

Anastasis [00:07:26]: Between, 2018 when we started and 2022?

Vibhu [00:07:29]: Yeah.

Anastasis [00:07:29]: So one of the early work that we did in Runway was solving segmentation, image and video segmentation. It was a very important problem because most VFX involves essentially separating

Swyx [00:07:42]: Rotoscope

Anastasis [00:07:42]: Subjects. Yeah, rotoscoping. Extremely manual process. Nobody enjoys doing that. and so a lot of the early days of Runway was building this tool. It was called Green Screen, and it was for a long time the main thing that people were using Runway for. It ended up being used in, Everything Everywhere All at Once and a bunch of other high-visibility films and series. But that was essentially, Runway for a long time was a post-production tool until latent diffusion and generat- Gen-1, Gen-2, happened.

Swyx [00:08:12]: Cool. let’s, let’s go past that moment. You’ve come a long way. Then you started releasing your own models. Maybe describe that journey as well.

Scaling Video Models and the Bet on 1,000 A100s

Anastasis [00:08:20]: Yeah, so we go to the other point, yeah, in mid-2022 when it became clear that we’re doing research at a fairly small scale of compute, and it became clear that, like, scaling laws would apply to, image and video gen in the same way that we’re applying to language generation. So we made a big bet, and I think at so at the time, we signed this deal to build a cluster of a thousand A100s, which at the time we were a Series B startup. That was a almost, slightly irrational decision maybe, but we really believed that if we trained a video model at a large scale, we would get, like, a great model at the end. And at the time, the goal or we set the goal around fall of 2022 of what is, what does the latent diffusion, Stable Diffusion moment look like for video? And at the time, the best model of the time was called CogVideo. it was one of the early video models. It was very 256 by 256 resolution, very not very high quality. and so we decided we’re gonna build out this cluster, and we’re gonna just invest in, like, in building out our own video model. it became clear as we’re training Gen-1 that it was difficult to get to fully. we wanted to build text-to-video, but it became clear to us that an easier starting point would be to start from video to video. Because when you have a stronger conditioning, it’s, it’s an easier problem to restylize an existing video versus generate the video from scratch. And so we released Gen-1 first back in, it was January of, 2023. Yeah.

Vibhu [00:10:04]: It’s just a fun visual podcast, honestly. Like, if we can see February 2023, what was the state of stuff?

Gen-1: Video-to-Video and Depth Conditioning

Anastasis [00:10:10]: It’s so interesting ‘cause at the time when you see those results, you think this is so incredible, and this is like, it’s almost like image generation or video generation is solved. And then you look back a few years after, and it’s like, it’s It’s just like you get used to the results very quickly, with those models. But at the time when we started seeing those results, it was, it felt quite incredible, and the level of, like, quality that you could get. And, so the Gen-1 was a depth-conditioned video model, so it would turn. it would take a input video, it would predict. it would it would first convert it into the depth map, and then we would generate, pixels with a latent diffusion model.

Swyx [00:11:01]: Yeah, very effective.

Vibhu [00:11:02]: Yeah. I didn’t realize how distracting the blog post would be. Sorry.

Anastasis [00:11:05]: Yeah, but, one of my favorite examples of on those, on Gen-1 was both, if you go up to mode three or mode two, there was this storyboard use case where people would make

Vibhu [00:11:18]: Ooh

Anastasis [00:11:18]: Would

Vibhu [00:11:20]: You can mess around with the

Anastasis [00:11:20]: Make a city out of books or out of boxes, and then they would shoot a video with their phone and then translate it into a photo-photorealistic output. There was all these ways in which those models were starting to be used for storyboarding and also for really. and then if you go to mode four, like, of taking untextured 3D scenes and then turning them into photorealistic output. So we saw a lot of use cases early on where people that were familiar, were power VFX editors would just take a blender, render, and then they would get translated in with Gen-1 or create a scene in Unity and then take a capture a video of it and then translate into, restylize it. So I still think video to video is powerful. I think we had a recent video-to-video model as well, and it’s one of my favorite ways of using those models is essentially using them to use ground truth video as, like, the initial inspiration and then translate into different styles or different outputs.

Stable Diffusion, Stability AI, and Open Source

Swyx [00:12:23]: But I think we’re gonna go into, like, the rest of Runway and catch people up to speed today. I did wanna cover the, let’s call it the Stable Diffusion controversy, or, what happened with Stability AI, whatever. I think there was a two sides of the story. I think there’s part of that is a normal thing of, like, people, join and leave companies, but what is the, retrospective now that, there’s been some years behind it?

Anastasis [00:12:49]: Yeah, it’s a very, it’s a very long story to go into. I think it would

Swyx [00:12:53]: Which I remember you wrote a really long post about.

Anastasis [00:12:56]: We would probably cover the whole hour to go into it in more detail. But, essentially, there was the latent diffusion paper that came in, I think that was at the end of, 2021. And then Patrick Esser, who was one of the researchers behind, latent diffusion, and he worked at Runway at the time, he built latent diffusion in collaboration with Robin Rumbach and a few other folks back, in the in, CompVis, which was, a lab

Swyx [00:13:26]: Like a research group, yeah.

Anastasis [00:13:27]: And, after releasing the early latent diffusion model, they, essentially they were. the goal was to keep working on versions of the model, scale it up, incorporate new data, incorporate new tasks. And Stable Diffusion was the same model, but trained on more compute, and then with a few more tricks, like a classifier-free guidance paper came at some point, I think in the early 2022. And that

Swyx [00:13:52]: Which, like, was a big prompting improvement.

Anastasis [00:13:55]: Yeah.

Swyx [00:13:55]:?

Anastasis [00:13:56]: That improved results. it was trained on better data, so like, the esthetic subset of LAION, but it was effectively, the same underlying architecture. And there was that big training run, that, happened on Stability’s cluster. Stability financed that run. And looking back at that story, I think it was the work to build and train that model was done. It was a, it was a research project. It was done as part of, like, continuation of the latent diffusion work. It then, I think it the model became very successful, and it, I think there were the. And I think as a result of its success, other companies tried to, figure out the commercialization path for it. But for us, it was very important that we try to, we make sure that we. It was meant to be an open source research project, and so the we decided that we should continue releasing versions of it, since that was the original goal of Stable Diffusion, and that led to releasing Stable Diffusion 1.5. There was maybe a day of, a bit of, miscommunication there, but ultimately that was resolved very quickly within hours. so yeah, there was

Swyx [00:15:12]: Okay

Anastasis [00:15:12]: Not a nice

Swyx [00:15:13]: I just wanted to. you have to

Anastasis [00:15:15]: Yeah.

Swyx [00:15:15]: You’re one of the main players in that journey, and so it’s nice to hear from the source of, like, what happened. Yeah.

Anastasis [00:15:22]: Yeah. I think it’s all, it’s all in the past now

Swyx [00:15:26]: Yeah

Anastasis [00:15:26]: I would say. and, like, both companies, Stability took its own path, Runway took its own path.

Swyx [00:15:32]: Yeah. There’s still. James Cameron is backing the new Stability, whatever they’re doing with the Hollywood studios.

Anastasis [00:15:38]: Right.

Swyx [00:15:38]: I don’t know what they are doing. I think one thing that impresses me, and I’m happy to move on, is that back in the that time, let’s say, like 2021, 2022, there was this community of people that you were involved in that was researching all this stuff, right? And, like, from everyone I talked to who was active then, it seemed like it was fairly obvious that somebody would do the hero training run that would produce Stable Diffusion. So, like, I guess the question is, like, you had the you were you had made investments. You were you had the foresight. Is it accurate to say, like, that is reflective of, like, what people were thinking at the time? Or was it still very much like, “Well, we’ll use it as, like, a post-production tool or something. I don’t know.”? Like, where in the sentiment were we that maybe you can think back to, like, what the community was like back then?

The Early Creative AI Community

Anastasis [00:16:28]: I reminisce and I think very fondly those early years, from like 2018 to 2022, because it was a very small community that, as you said, were very convinced that this was gonna be a big thing. And at the time, anyone who. Because it was such a small circle and, everyone who would, like, be part of that circle and, like, make projects with it would, immediately get, go viral. so like

Swyx [00:16:55]: And you didn’t know who they are, right? They’re just some name on a, GitHub or Hugging Face somewhere.

Anastasis [00:16:59]: Exactly, yeah. So I remember one of the first big viral moments of creative AI was, there was the neural style transfer paper

Swyx [00:17:09]: Huh

Anastasis [00:17:09]: That

Swyx [00:17:10]: Something dreaming?

Anastasis [00:17:11]: I think it was called neural style transfer.

Swyx [00:17:14]: Okay.

Anastasis [00:17:14]: There was also Deep Dream, the puppy slice

Swyx [00:17:16]: Yes

Anastasis [00:17:16]: Which was, also really cool. but, yeah, there was this project that, Jim Kogan, who was an early advisor of Runway and one of those,

Swyx [00:17:25]: Marketing guys

Anastasis [00:17:26]: Big, creative AI, folks, he literally just, like, showed a video of himself taking the New York Subway and going over the Williamsburg Bridge and then stylized it with, I think in the style of Van Gogh or, like, one, painter. And that was. Like, at the time, that was, like, so cool and it went viral and it was completely revelation to people that you could do this with generative models. And that was only, it was less than. It was maybe 10 years ago. So just, like, as an indication of, like, how quickly things have gone.

Vibhu [00:18:02]: It’s pretty crazy. Like, even since then, you’ve got people at every level of the stack. You’ve got devs, creatives, artists, hobbyists. You’ve got everyone using it. And for people that tried stuff early, they’ll remember how hard it was to use regular diffusion, right? Like, nowadays, you can use your favorite ChatGPT image gen or whatever, give a sentence, get a beautiful output. But diffusion was like, the whole ultra HD, 4K, high resolution. Like, prompting these things was very different. anything you learned on the tooling side, like from the offerings you guys have now, so like creatives, devs, you really took the. Research and brought it to everyone to use. anything interesting there to share?

From Gen-2 to Controllable Video Generation

Anastasis [00:18:44]: We had to build the entire model serving infrastructure for video diffusion models. There was nothing else, already, like, because we had Gen-2 was the first text-to-video model, I think, out in the market. So many things that we learn over time. I think the I think the biggest one was, like, we. it was very clear early on that text-to-video was not gonna be the answer. Like, you. Like, people wanted a lot more control than that, and so we invested in, like, control building on top of those models very quickly. how do you use the camera trajectory as control? How do you use an initial input frame as control? So that was a very early learning for us. With text-to-video was, like Gen-2 was an amazing, step function improvement in the quality of video models, but it was used much more in an exploratory way because there was nothing to ground it to. There was no reference that you could bring into it. There was no. You couldn’t really control the camera motion. You couldn’t control the object motion. And so the first year, in 2023, was really all about what are all the interesting ways in which we can condition those models? And it was a lot of just post-training rounds on top of the base model to figure out, like, what, -- how do people wanna control them? And so there was, like, this quick succession of the we it was called Motion Brush, which was you could, like, you could draw arrows and dictate where things should move in the scene.

Vibhu [00:20:09]: That’s so cool.

Anastasis [00:20:09]: There was camera control that was you could just describe, like, how you want the camera to move in the scene. And because we work with filmmakers from the most of the history of Runway, we immediately got this feedback and got this, decided that this was worth investing in. And so control ability became a big theme, I think, very early on as we were building, as we were building those models. Something fun that I haven’t really talked about too much was just how Gen-2 came to be out of Gen-1. So it was a bit strange because we announced Gen-2 two months after Gen-1 and

How Gen-2 Came From a Weekend Hack

Vibhu [00:20:43]: We’re accelerating.

Anastasis [00:20:44]: It was before Gen-1 was even generally available. But Gen-1 was a depth-to-video model, so it would take a depth map and it would convert it into RGB. and we couldn’t get, text or image-to-video to work directly, and that’s why we started from depth to video. but, and we had discussions of like, okay, we need to spend the next six months investing in text-to-video, maybe increasing the compute scale or the model scale, like train a larger model. And I had this weekend project idea, which was, what if I take a model that, starts from text input and converts to depth maps and then use Gen-1 to convert the depth maps Into RGB?

Vibhu [00:21:29]: It would probably work.

Anastasis [00:21:30]: And so Gen-2 was that.

Vibhu [00:21:32]: Oh. The hackathon pipeline.

Swyx [00:21:35]: The weekend hackathon pipeline.

Anastasis [00:21:36]: Yeah.

Vibhu [00:21:37]: But it looks good.

Anastasis [00:21:38]: And it worked pretty well. there were if you, with the knowledge that it has this, like, two-stage pipeline, you can tell in some cases that the structure of the video looks a bit off because you had to generate the depth first before you go into the output video. But it worked and it allowed us to bring this to our, to users very quickly. But it’s now it’s interesting because, like, people are coming back to this almost two-stage approach. Like, if you look at the Reve text-to-image model that came a few months ago, it had this planner model that would generate bounding boxes before it fed that into the diffusion transformer.

Swyx [00:22:19]: Yeah, Ideogram also the same day.

Anastasis [00:22:22]: Yeah.

Swyx [00:22:22]: I remember that was very strange that both of them came out the same day with the same exact innovation.

Anastasis [00:22:26]: It’s a small community, I think.

Swyx [00:22:28]: I’m like, this is like, this is completely coincidental, right?

Anastasis [00:22:32]: People talk. So yeah, there’s, there’s definitely something into this approach. And, now, like every single like, video generation model in production uses a complex prompt completion pipeline under the hood. I think that’s no secret that there is. That

Swyx [00:22:48]: Humans are terrible at prompting.

Prompt Rewriting, Camera Control, and the Seed of World Models

Vibhu [00:22:51]: I think across the board.

Anastasis [00:22:51]: Yes.

Vibhu [00:22:52]: But yeah, I think like the original Sora one blog post even told you that what happens after your input is rewriting your prompt. It’s much more descriptive about what you would want.

Anastasis [00:23:02]: Exactly. I, And there was the DALL-E 3 paper beforehand that, was the first public, description of the fact that synthetic captions and really detailed captions work really well. And then Sora built on that. Yeah, so it was 2023. We were releasing all these updates to Gen-2, like the camera control, Motion Brush. And there was something very interesting about camera control because it was the first time that you felt that instead of, like, you were creating video, you were creating a short video, you were navigating inside the world. And I think camera control was maybe the seed of some of the ideas that we had around world models and really opening up that research direction. We realized, it was this era and this series of, Gen-1 and Gen-2 models really proved to ourselves, yeah, this is the

Swyx [00:23:56]: Cool.

Anastasis [00:23:57]: So this is not the original camera control. This was the updated camera control on top of Gen-3. But yeah, I think it made those models usable to filmmakers, I would say. The so camera control was very popular. And so we realized, there is one way of seeing those models, which is, you’re just as content creation machines, and there is the other way, which is you’re. As you’re predicting video in order to predict video well, you need to simulate the world in an increasing and increasing capacity. And if scaling laws apply on video, just like they apply on language models, then as we scale the compute that we put into those models, then they’re gonna be able to simulate physics, they’re gonna be able to simulate human actions and dynamics increasingly well and predictably well. That was the thesis about around our efforts on world models, and we spin up this research group to just focus on the world models and how do we turn the video generation models that we’re building into something broader and something that would be useful beyond, also content creation as well.

Swyx [00:25:04]: And that was roughly when?

Anastasis [00:25:06]: Yeah, so that was in

Swyx [00:25:06]: Oh

Anastasis [00:25:07]: In late 2023.

Vibhu [00:25:08]: Interesting. like, I think, a lot of people have been saying a lot of video gen model companies have all pivoted to world models these days, but like, 2023, you’re posting it. one

World Models: From Video Generation to Simulation

Swyx [00:25:21]: It’s, it’s debatable whether it’s a pivot.

Vibhu [00:25:23]: Yeah.

Swyx [00:25:23]: Like, arguably

Vibhu [00:25:24]: Yeah

Swyx [00:25:24]: That’s what you always had to do anyway, right?

Anastasis [00:25:26]: It’s in a way an expansion

Vibhu [00:25:28]: Yeah

Anastasis [00:25:28]: Of the applications

Vibhu [00:25:29]: Yeah

Anastasis [00:25:29]: Of the models as they become more capable.

Vibhu [00:25:31]: The early signs, it seems like the original models you guy had, guys had, people would say it’s very not bitter lesson pilled, right? You’re adding, rewriting prompts, you’re having all these one-off things, but that’s just the state of the tech as it was versus the future of as you said, you can scale it up as, we can scale up to world models.

Anastasis [00:25:50]: Yeah. So it just became. And if you looked at the outputs of Gen-2

Vibhu [00:25:56]: Yeah

Anastasis [00:25:56]: It was not. I think it was not obvious to people that this would scale to become a general simulator of the world. Like, you had very limited movement, you had, very low fidelity or low resolution, like obvious mistakes in human anatomy, like all kinds of limitations. But it was just, the idea was that’s just GPT-two, and GPT-two, it can barely generate, like, coherent sentences. Similar, Gen-2 can barely create coherent video, but if you scale it up, you’re gonna. There is no reason why it shouldn’t work in a way. It’s, And I think that was. That’s, that’s always the mindset of Runway is like this extrapolation of, like, if, like, even when we started in 2018 and you looked at the results of the day, you need to look more at the trend of, like, where we were in 2018 versus when we were at the, when the first GAN came out in twenty, four 2014 or twenty, fifteen. And, you started from, like, thirty-two by thirty-two images of faces, and then by the time in 2018, you could generate, street images at the 1K resolution. And it was the same with world models, very early signs of something much bigger.

Swyx [00:27:08]: Yeah. I was gonna say, like, it’s diffusing into focus. Like, if you look at our visible output from year to year, it looks like a diffusion process itself.

Anastasis [00:27:17]: Yeah.

Vibhu [00:27:17]: Especially watching the early, like, old blog posts, you can really see the choppiness, the details.

Anastasis [00:27:24]: Yeah. Like human civilization starting from random noise and then

Vibhu [00:27:27]: Yeah

Anastasis [00:27:27]: Denoising into

Swyx [00:27:28]: Yeah. Just run it a hundred years.

Anastasis [00:27:30]: Civilization.

Swyx [00:27:30]: Yeah.

Vibhu [00:27:31]: That’s how you’re on track, you’re still noising, right?

Swyx [00:27:34]: Yeah. I like the way that you guys phrased it when you, announced it in June, which is, oh, that you had a video essay. “The human mind is no longer the center of AI. Our world is.” Right? Which is, let’s, let’s call it the past five years of LLM-based AI is very much like trying to emulate human preferences and human speech. But now that’s, like, mostly solved. I think that’s, like, some of the context of your essay, which you also wrote around the time. And now it’s like the focus is on modeling the world accurately.

Scaling Laws for Video and Why Predicting Pixels Matters

Anastasis [00:28:03]: Exactly, yeah. So the way we see it is, there is that, initial mission statement of DeepMind, which is, solve intelligence and then use it to solve everything else. But I think it’s starting from everything else, could be valuable of, like, starting from. there is just so much complexity, and detail in the world that in order to. That it’s, it’s hard to learn directly from just human descriptions of the world. Like, we’re assuming that, like, language models learn from everything that humans have written about the world, like our own understanding as of, the twenty twenties. And there is just so much that we don’t know and so much that’s not captured by existing text, about both the low level dynamics of the world, like we’re not describing in detail. if I tell you to describe, like, how do you tie your shoes, that’s a very difficult thing to describe in words, but it’s very obvious thing to demonstrate. And so I think there’s been. And there’s, more of X paradox, like we’re constantly underestimating all the complexity that goes into very, like, things that we do subconsciously as humans, and we don’t even necessarily always have the words to describe them. And so in my mind, the simulating the world and simulating, physics, simulating the dynamics of the world has always been underestimated, compared to, we place too much emphasis on the things that are easy to talk about. but there is just all this complexity and richness of the world that if we just try and train directly on that observational data instead of training on how people describe the world, we would learn something new that we wouldn’t otherwise know.

Swyx [00:29:54]: You think that the present architectural paradigm is fine? You don’t need, like, another layer, like JEPA, like another famous, New York AI leader would say?

Anastasis [00:30:05]: We’re a very pragmatic research lab. If, we have evidence that an approach works better than the approach that we’re taking, then we have no qualms to taking it. We just have seen no indication that video prediction itself doesn’t scale. And even if you look now, not just our work, but the work of others, you’re seeing in robotics some of the most promising work, starts from video prediction models, and then you adapt them to also the action models, for example. so there is very little evidence that you need something else and that your time is better spent on a novel architectural change compared to improving data and improving the, and scaling the current approach. And so, We don’t have any indication that. the, there is that counterargument that I think there was a tweet by Yann LeCun a few days ago that, understanding the dynamics of the world is very different than, generating, cute videos.

Swyx [00:31:05]: And your answer is no, they’re the same thing.

Anastasis [00:31:07]: Yeah, they’re the same thing.

Swyx [00:31:08]: My cat videos are the same as understanding physics.

Anastasis [00:31:11]: Right, because if you wanna generate. video models can cheat and, like, they could you could give, like, successive dif shots of the scene in a way that doesn’t require you to simulate difficult physics. There is like, all these different ways in which you can hide the deficiencies of the model, and it’s important not to be too tricked by the performance of the current video models. It’s easy to, cherry-pick examples and think that video models are further advanced than they are. So there is a lot more work that we need to do to improve those models. But in my mind, very similar to language, and, like, we’ve. you go from barely coherent sentences to something that, could hold a conversation with a human to something that could can operate autonomously for a day and, like, create entire code bases. And the main difference, there is some architecture improvements along the way, but the main thing is scale. And so it’s the same bet for video, and we have no indications that this is saturating. Like, we have benchmarks that we use for measuring the physics of those models, and we see those predictably improve as we scale those models. So there is. If you want to Google up, Physics-IQ, is one of those benchmarks that measures how well does the model perform at solid mechanics or fluid dynamics or optics.

Vibhu [00:32:32]: I’m curious if you’ve seen any emergence, any scaling law around this.

Swyx [00:32:37]: Yeah, he’s saying there is a scaling law, right?

Anastasis [00:32:39]: Exactly.

Vibhu [00:32:40]: Yeah,

Anastasis [00:32:40]: So the way those models, those benchmarks work is you. the researchers have gone and, like, captured, a few videos that are representative of different physical phenomena, and then you can take the first frame and then pass it through an image-to-video model and then generate a rollout that shows what should happen next. So you have, a ball hanging from the ceiling, and then you use that as input, and then you the model predicts how the ball should fall on the ground. and this measures. we have an intuitive understanding of physics. I know, you can imagine what will happen next if I drop this bottle. So it’s measuring that same intuitive physics understanding of those models, and we’ve measured that at different model scales, and we see, and compute scales, and we see that the score on physics IQ predictably improves. There’s other, tricks and techniques that you can make to improve the score even further, but even scale alone helps, in the model learning better physics.

Swyx [00:33:40]: My main sympathy with Yann LeCun is the, Plato’s cave allegory, right? Like, you’re, you’re, like, learning on the output of a thing, not the internal process of a thing, and it’s very noisy. And, if only you could observe the internals of a thing. It’s hard to observe the internals of a human mind, but you can very much observe, or at least we have a whole branch of science and physics that we’re ignoring on how to model Physics and movement and, gravity and, other interactions. and we’re just, like, throwing away all of that and just saying just scale data, which is very much the lesson of unsupervised learning, but it feels wrong. that’s the main idea.

Anastasis [00:34:21]: I think the history of machine learning is, at large, it feels wrong.

Swyx [00:34:25]: Yeah. It’s a bitter lesson, right? Yeah. It’s, it’s, it’s the simple answer to that.

Vibhu [00:34:29]: I guess, how much can you scale? So, like, even on, let’s say, the video generation side, like, there’s one side of video understanding. Video generation, are we still gonna have tools where it’s like, I wanna generate two hours, twenty hours? there’s a infra way to do it in batches and stitch it together, but, like, do we just keep scaling? Do we just continue long generation consistency, all that at scale? And, like, tying it into where we’re at now from we looked at Runway two to four point five

Gen-3, Sora, and Runway’s Scaling Inflection

Anastasis [00:34:58]: Yeah.

Vibhu [00:34:58]: Like, technically, what advancements have we made to today, and then where do you see things still going?

Anastasis [00:35:04]: So part of the answer is definitely scale. and that was. We learned that lesson in a big way for with Gen-3. So Gen-3 was the model we released the year after, like in 2024. That was a few months after Sora was released. so yeah, there’s an interesting story of that came to be as well. Gen-3 for us was, the first time that we really needed to build. we had to learn all the lessons that the language model world learned in two in three years in the span of a few months. one of the biggest changes of Sora was using diffusion transformers instead of convnets. So a lot of the early, latent diffusion models were all, convnets for the diffusion model part. And the diffusion transformer paper came at some point in 2023, and it showed scaling laws for image, diffusion transformers. And we realized at that point that we needed to invest in infrastructure for model parallelism, for really scaling training to larger than, a few billion parameter models. And we spent maybe the, most of the fall of 2023 building out our infrastructure for distributed training. And we had a lot of false starts and a lot of failure in trying to scale, image and video diffusion transformers. And at that point, February 2024, Sora comes out, and the results are

Anastasis [00:36:35]: Very much superior to what Gen-2 could produce. There were a lot of, a lot of chatter on Twitter about Runway. Runway’s done. like, there is no way Runway will catch up. And if you remember, also OpenAI in the early twenty-It felt very, like it’s a

Swyx [00:36:56]: To the moon

Anastasis [00:36:57]: It’s a formidable opponent now, but at that point, it, they were on the top of their game. nobody could even get close to them. There was maybe Gemini was just the first version of Gemini had just released. So when OpenAI came with Sora and it was such a big jump of like quality, it gave me, there was like an existential crisis for a few hours. But that, I think the amazing thing about Runway and like I think the, we’ve been around eight years now, which is almost we’re dinosaur in AI, and we had to like, we had there was a lot of those moments we had to learn, adapt very quickly and build out skill set in the team that we didn’t have. And so, if you ask anyone what is their favorite time at Runway that was there during that time, it was that push in like three months to get to a model better than Sora. and it, we scaled 10x the model scale, the model size and the, compute that we were training on. we figured out model parallelism. We had zero expertise in that. And then we came out with Gen-3 during that summer. So that was a big turning point, I think, for the company where the research org grew very quickly, and we really started pursuing this vision of the general world model, in earnest, I think after Gen-3 was out.

Swyx [00:38:12]: Yeah. that’s the amazing thing about building when you’re building. There’s no stack to. You have to invent everything yourself. You have to be completely full stack. Now I think like there are inference specialists like Fal or whatever that can help with like, model serving, and I think you guys work with them as well. but yeah, like it’s, it. But at the time, it was just. It’s very interesting to think about what you do when Sora comes out and people are questioning whether your company should still exist.

Distillation, Turbo Models, and Real-Time Video

Anastasis [00:38:41]: Yeah. And yeah, there was no, there was no VLM of diffusion models. Like, we had to build the whole model serving infrastructure and make things efficient. And a few months after we released Gen-3, we released the Turbo version, which I think was the first step-distilled model in production.

Swyx [00:38:56]: That was a whole trend that we covered as well. Yeah.

Anastasis [00:38:59]: So that allowed us, to serve those models at the larger scale, ‘cause I think the first version of Gen-3 was quite, expensive to serve.

Swyx [00:39:09]: I think the whole like trend in like consistency models, Lightning and, Turbo and all these things somehow didn’t really stick around. I don’t know if you have any reflections on this. Because at the time, I was like, “Well, everything should start with a distilled model first, and then you can upscale,” right? It. your bigger models just turn into fancy upscalers, but like you should always draft with a smaller model and faster model, right? Because you can get it so quickly, like near real-time.

Anastasis [00:39:39]: Yeah. I would not be so sure to say that didn’t stick around. I think that, it’s, it’s likely to. that there is a lot of step-distilled models that are actively used in production. there is still a gap in quality compared to the, non-distilled model. but in my mind, we’re still. there is a two to three year offset from language models. So the things that, So it’s just a matter of time before there is better distillation techniques. we use. Right now we have a real-time model core character that I think is the largest deployment of real-time video models, that’s a step-distilled model, and it’s actively being used. It’s a very specific use case compared to a general video model. So this is a

Swyx [00:40:27]: Very cool, by the way.

Anastasis [00:40:27]: This is avatars stuff, right?

Swyx [00:40:28]: Consistency, character.

Anastasis [00:40:30]: Yeah. So this is a talking avatar, model. we were able to. we optimized the hell out of it, and it generates at 24 FPS, and it’s a, it’s a step-distilled autoregressive video model. So if we look at our world model direction, a big component of it is starting from the bidirectional diffusion that generates entire video at once and making autoregressive shows. So you generate one frame or a few frames at a time. so there’s a lot that goes into that pipeline of getting to a real-time model. It’s first you need to make it into a causal autoregressive model, and then you just turn it into. You need to do some additional step distillation to get it to be real-time. and I think that part is just starting. I’ll be very surprised if we’re, two years from now, we don’t primarily use real-time models. To me, real-time video generation is just inevitable that, it has much better user experience, it’s much cheaper to serve, and, the quality gap between the base model and the real-time model is only gonna close as we figure out better, distillation techniques. And we made a lot of progress there internally on maintaining the quality of the base model when we distill them.

Swyx [00:41:49]: How much of this is transferable? So is it the same base model? Like if you’re doing diffusion across the whole sequence and you’re converting it to step autoregressive distillation, is this like distillation where you still need to train both, you can use the same base and converter? What’s that process like to go from regular model to something that’s real-time on a technical level?

Anastasis [00:42:11]: So the nice thing about diffusion models is you have, two axes of distillation. So there is the. You can distill to a smaller model, which resembles what you do in LLMs, or you can distill in terms of taking less steps, less diffusion steps. So you could take a model that generates in fifty steps and generate in four steps and get to, You have some performance, degradation, but very often you get comparable outputs. So you can even take the large frontier model and distill it with step distillation and get to a real-time performance, and that’s what we’ve seen. So, depending on the use case, in some cases we might also serve with a smaller model, but in a lot of use cases, we just use the

Swyx [00:42:56]: Step distillation

Anastasis [00:42:56]: The frontier model, and we’re able to make it work in real-time.

Swyx [00:42:59]: I think this might be a good time to cut over to his laptop to show off some of the real-time stuff that you’re doing.

Interface World Models and Neural Software

Anastasis [00:43:06]: This is one of the research updates that we did recently. so we’ve been working and f in getting our general world models to, different applications. one of them that we think is very compelling is using general world models as essentially, an interface, a universal interface to software. This is a version of our world model that’s called an interface world model. and the idea is that it essentially, replaces, the, front end of a software application. It renders the pixels directly of an interface and is trained to predict what happens next as a result of, a click or another interaction you have with the interface. So this is all pixels. it’s there is no HTML, CSS, React that’s powering this interface. This is directly at the output of our real-time, video generation model, and it takes clicks directly as input.

Swyx [00:44:09]: And drags, click and drag.

Anastasis [00:44:12]: Right. So it supports

Swyx [00:44:13]: Ooh.

Anastasis [00:44:14]: Yeah, clicks. It supports drags. it also supports scrolling. and the amazing thing about this is that you can effectively describe in the prompt how you want different elements, like what do you want the behavior of different elements to be. So it’s almost you’re you can turn, an interface from, markup language description of, like, an HTML interface, and instead you can just describe the interface. if I press this button, I expect this to happen. If I press this button, this should happen. And it’s useful, we believe, both for prototyping, for, like, just testing, like, what different interactions would feel like. you can also add audio to it. So it’s a video audio generation model. So you get you essentially can describe both what the visual outcome should be of your click and also what the if there is a sound effect that comes out of it. So we believe that’s gonna be a much more flexible way of building software. Just render. It just, in why generate the code that generates the pixels? Just generate the pixels directly.

Anastasis [00:45:18]: It’s the end-to-end philosophy applying applied to front ends.

Anastasis [00:45:25]: So we think there is a few interesting use case. So you can build creative tools on top of it.

Anastasis [00:45:32]: We think that, for any use case that involves a lot of exploration or, like, educational use case where you wanna learn about a new concept and you want some visualization and like, and open-ended exploration, we think those this is a very powerful, approach. you can imagine new forms of, design, industrial design software that could emerge as a result of those models. And this is all, generated in real-time as well. So, you can build a lot of interesting camera transitions and forms of interaction that are very difficult to build otherwise. And one way in which we evaluate this is what if you try to generate the same interface with Claude by just, prompting Claude, “Here’s an image reference of my interface that I made in Figma or that I created somewhere else. create this particular interaction,” which in this case it’s, drag that object, upwards. and beyond it being slower, it’s also very difficult to capture some interactions by just fully, with just LLMs. So we think that this is likely to be the way that a lot of the future, like, software in the future will be created. and one of the additional benefits is personalization might be a lot easier done with those models. Like, you can essentially try out different prompts based on who is visiting the interface. You can, more easily, prompt engineer the interface to have larger size, text for more accessibility reasons, or you can make this or, like, if you have a particular aesthetic preferences. So we’re very excited about this approach. It’s early days, and I think we’ll need to, make it more cost-effective as well to serve those models ‘cause, running a real-time video model versus just purely rendering HTML, there’s -- the computational needs are much higher. but we do see a lot of potential in this approach to building front-end interfaces.

Swyx [00:47:47]: So we covered this similar thing with Flipbook before with our, Ethan Hara episode with Groq, video. And yeah, I think it’s very engaging visually. I think it’s maybe very good for education, but it’s it does sound expensive. I think there’s an upper bound to how expensive it will be, though, right? Like, the inference cost will go down over time. You’ll figure out ways to optimize it. Effectively, when it pauses, you don’t you’re not receiving human input. You don’t have to generate anything, right? So.

Anastasis [00:48:14]: Yeah, you could also. Like, in this case, you have ambient motion, so there is parts of the screen that might. if you’re let’s say you wanna, visit Paris and then you get this interface that allows you to explore.

Swyx [00:48:29]: People walking. Yeah.

Anastasis [00:48:29]: You have people walking or, like, things happening. But, it’s, it’s a no Yeah, it makes it more expensive because you need to run the model all the time. Maybe you have some looping mechanism so you don’t need to do that. But all those things, I think, is stuff we’ll need to figure out.

Toward a Fully Neural Operating System

Swyx [00:48:44]: Yeah.

Anastasis [00:48:44]: I think our first consideration is let’s make this clearly find some use cases where it’s clearly a much more compelling interaction compared to traditional interfaces. And then it’s a matter of time before it becomes more cost-effective to serve.

Swyx [00:48:58]: Yeah. When it comes to the people walking, I think the approach that makes the most sense to me is Nick.

Anastasis [00:49:04]: Nick.

Swyx [00:49:04]: Oh, God. I keep messing up their name. With Chris Manning and Fanny Yan. I don’t know if you’ve come across them, where they. Mapped to some game engine. I think it’s Unity or something, or Godot. And they you can script some NPC behavior behind that and train on that. Whereas here, you can really imagine whatever you want. Like, that is a UI, right? Like, and it feels, like, more tractable, I guess, to, create a world model of software that is interactable because we have many of examples of that, and you can, do your fancy RL environment stuff on that than it is scaling up to embodied and real-world physical use cases. But this is a nice first step.

Vibhu [00:49:43]: Or, there’s the opposite of you have, like, one B models, three 50 million parameter language models. It just gets so small that they’re just predicting, like, fishes moving.

Swyx [00:49:53]: Small models are now 120 B, so.

Vibhu [00:49:57]: Ultra mini on device.

Vibhu [00:49:58]: But, no, I think it, like, it puts it into perspective, at least the car one for me, like, the applications, right? The amount of work to do that, sure, you only make one model year car per year, but applying this, it’s also a cost-saving to have to manually make all this, right? So it opens up a lot of possibilities, too. I’m curious if you extend this out two, three years, so where do you see things going even further?

Anastasis [00:50:25]: Effectively, the end game of something like interface world models is you have, a fully neural operating system. So I think, Andrej Karpathy has written about that quite a while back. But it’s, You, I think to me it’s, it’s a bit, it’s a bit odd that, we have, for example, with an interaction with an LLM of today, you have this LLM that can talk to you about anything. It can You can take the conversation in any direction. You can It’s very general, so it can solve all those different tasks, but you interact with it through a very rigid interface. And so to me, it’s just a matter of time before the interface itself becomes learnable and becomes, part of the whole loop of, like, you’re not just delivering. You’re delivering an application end-to-end, and that means you’re delivering the language model, but you’re also delivering the render and the pixels and that’s also a learnable component. And the concept of applications might not necessarily. I think we’ll need to figure out new abstractions for software. the concept of application comes from this idea that you need, separate code bases to describe, to, for, to power each individual, tool and each individual application. But you might think of something a lot more unified if you’re. if you have, a video model that’s generating the interface as you go. so it can take context from an LLM and allow you to combine different functionalities that traditionally would live in different applications. So it’s a, it’s a way to solve, software end-to-end, effectively. We also see this as a powerful way to train computer use agents as well. so this is, one way to see this as. And in general, with world models, there is those two directions. One is world models for humans and world models for

Swyx [00:52:24]: Agents

Anastasis [00:52:24]: To train agents.

Swyx [00:52:25]: Yeah.

Anastasis [00:52:25]: And so for every new work of, world models that we do, we have this both uses become possible. So this is a powerful synthetic data generator for training computer use models. It could become, a live, RL environment that you could use to do online RL with a computer use agent, and you can get wide diversity of different interactions, kinds of interfaces, just generated on the fly that, to improve the how robust the, your agent, becomes. So that’s the same also with the world models that we’re working on for a robotics use case as well.

Long Context, Error Accumulation, and Autoregressive Video

Swyx [00:53:02]: Is there a research breakthrough that you’re Waiting for that would unlock the next set of use cases that you really wanna pursue?

Anastasis [00:53:10]: Long context is a very important one, so being able to maintain consistency for long periods of time, and that depends on the use case. So for our characters model, for example, or for the interface world model, it’s easier to maintain long sessions of interaction. If you go into more open-ended worlds that you navigate and you take arbitrary actions in, we, like, there is more the context at which you can and duration which you can generate becomes limited much more quickly.

Swyx [00:53:40]: Yeah.

Anastasis [00:53:40]: So we see more degradation and error accumulation happening. so the biggest challenge with autoregressive models is error accumulation, is you’re feeding generative frames back into the model to generate the next The next frames. And if there is any small errors, they accumulate over time. That’s not a new problem. It’s a problem that LLMs also have, and we’ve seen the ability to generate now really long outputs. So it’s a solved problem, but it’s definitely still a challenge.

Swyx [00:54:08]: Yeah. And what is the state of the art? so for Grok, it would be like 10 to 20 seconds of context going in there for video.

Anastasis [00:54:16]: With our characters models, we’re able to generate up to 30 minutes of video autoregressively.

Swyx [00:54:21]: Yeah. But that’s just for the avatars.

Anastasis [00:54:24]: Yeah. So if we look at, GWM Worlds, which is more our open-ended world exploration model, it’s, it’s on the order of a few minutes, which is Yeah, so

Swyx [00:54:35]: Probably enough for people because you have to cut to the next scene anyway, right?

Anastasis [00:54:40]: Yeah, it’s not, it’s not the ideal game experience if you have to restart every few minutes. So I think. But, I think it’s. Yeah, for certain kinds of game experiences, you can work around it. ideally, you are able to just generate forever, and it doesn’t, it doesn’t degrade. And I think that’s a matter of time before we get there.

Swyx [00:54:59]: Yeah. Genie has, like, one, max one minute?

Anastasis [00:55:01]: Right. Yeah.

Vibhu [00:55:02]: This was your. You did a study on robotics. I think I also have just your Runway Robotics page, though. Is this better?

GWM Robotics and Sim-to-Real Evaluation

Anastasis [00:55:11]: So last year we released Gen-4.5, so that was our latest base model. We’ve been As I mentioned, we’ve been doing all this work in world models, and which essentially a lot of our approach to world models is how do you take a bidirectional diffusion model and make it autoregressive and make it accept actions? So instead of being a video you watch, it becomes a simulation that you step in, and you can, control it every step of the way. You can explore counterfactuals, like what happens if I take this action versus if I take this action. And GWM-1 was the it’s the world model that we built on top of Gen-4.5. So we did all this autoregressive and like, distillation, auto-regressive and then step distillation on top of Gen-4.5. And one of the biggest use case that we saw for GWM-1 was in robotics. One thing we like to say is we as we scaled video models, we accidentally, created one a state-of-the-art model for robotics, by just scaling video models. So we realized at some point, mid last year that robotics labs that are coming up to us and asking to use video models for synthetic data, asking us to post-train our video models to work really well for robotics, so that they can use that to generate variations. That was the first use case that we saw. And then increasingly became clear that the models will be useful beyond just creating synthetic data to train robotic policies. They would also be very useful as simulators. So that means that you can use, a video model online to test how your robotic action model performs. So you can take an action role and then get the outcome of the action inside the world model and then continue that loop like this closed loop simulation. And you can use that to evaluate how well your robotics model works. and the biggest thing that I think you need to solve if you want to build a simulator is establishing real-world correlation that if you take an action inside the world model, if you take the same action in the real-world, you get a similar outcome. So that was the goal of some work that we did earlier this year. So if you go to the first link. So that was, essentially wanted to establish that, real to sim correlation for our world model, so that if you do a series of actions inside the world model and if you do the same actions in the real-world, you get similar outcomes. And we took our GWM-1 model and we used some benchmark data that there is this Roborina, benchmark that’s very commonly used to evaluate how well do different action models perform. And we use the same scenarios and settings and embodiments inside our world model, and we measure the correlation of how well did the action model perform inside the world model versus in the real-world. And we saw that we could get very good correlation between our world model and reality. And that means that if you want to evaluate how well your robotic policies perform, you can scale that much faster inside simulation instead of having to do that with actual physical hardware. And so that was a first indication that our models could be, quite useful in robotics. And we saw as we were working with robotics labs that became like the first use case where they could use video models in a way that feed into their training pipeline.

Vibhu [00:58:40]: Can I ask what

Anastasis [00:58:41]: Yeah

Vibhu [00:58:41]: The difference was from four point five to solving that? So the sim to real gap has always been the issue, right? You train a robotics model on video data, it doesn’t generalize to real-world, and the simulation had an issue. So seems like you solved it, but how?

Anastasis [00:58:56]: Yeah. So a big problem with simulators is, if you’re trying to simulate rigid objects, like it works quite well if you can describe the physics of objects very accurately, then you’re able to use, Isaac Sim or MuJoCo or one of the traditional simulators. But for more complex interactions with cloth, for example, or, like slippery surfaces, with the all the complexity that you want to be able to solve with the manipulation, with an action model that solves manipulation tasks, it’s very difficult and so time-consuming to build, for each of those environments and each of those tasks, build the simulated version of that, the digital twin of that environment. Whereas with a world model, you just need to provide the first frame and then you just can roll out the policy inside the first frame. So whereas, we compare it to methods that required like 3D scanning an environment and then 3D scanning each individual object before you can now, you can bring that to simulation. whereas with a world model, you just take a picture of the environment and then you’re able to test how your policy performs. Our general thesis on robotics is, there is companies that are leveraging a lot of teleoperation data to train robotics action models. There is now companies that are using, humie data, which is, essentially human, egocentric video where humans use robotic creepers to perform different manipulation tasks. And then there is companies that are focusing on egocentric data, which is, you strap a GoPro on someone’s head and then you capture them performing a task. We think that, and all those are great source of data for training robotics models, but the most plentiful source of video data is third-person video data. It’s And if How do we as humans learn how to perform different tasks? A lot of it is by observing others perform those tasks. We don’t learn from first person. We do some trial and error and like, to learn different things, but. Ultimately, a lot of what we learn how to do in the world, we learn by watching other people do it. And that’s how when you’re pre-training a video model, you’re essentially doing that. It’s a lot of third-person video footage of people performing different tasks in the world, people doing sports, people doing household tasks. And our main thesis is that video pre-training, once you do that, you can then adapt a model to be useful in robotics use cases with way fewer hours of actual robotic data. So you require way less teleoperation data, which is very difficult to scale. and even if you look at egocentric data, which is a bit more easy to scale compared to teleoperation data, which requires actual hardware,

Why Third-Person Video Is a Powerful Robotics Pretraining Source

Anastasis [01:01:55]: It’s still three hours of magnitude less of that exists in the world compared to third-person video data out there. And so our thesis is and generally, like the most plentiful source of data will ultimately wins. Third-person video data pre-training is the right starting point for models that, you want them to generalize and be able to deal with new environments, new tasks, things that you haven’t seen during training. That’s the motivation for why we think our models are especially useful in robotics, settings, and we’ve seen that to be the case, as well.

Swyx [01:02:32]: You said pre-training. So maybe it’s like third-person pre-training, first-person SFT? Is there like a curriculum that you can introduce?

Anastasis [01:02:41]: Exactly. So if we look at GWM Worlds, so GW so GWM Robotics. So digitally in robotics, it starts from Gen-4.5.

Vibhu [01:02:49]: It’s the same video diffusion backbone, right?

Post-Training World Models for Robotics Embodiments

Anastasis [01:02:53]: Exactly, yeah. So you start from the base video model, the one you’re using to generate, cats and dogs and other interesting stuff, and then you, fine-tune on a very small number of hours of robotic data. So it’s something on the order of hundreds of hours compared to if you were to pre-train a robotics model. The current pre-trainings go up to, a hundred thousand or like millions of hours of data. And you’re able to get quite good performance, quickly, because the model leverages all the things that it has learned about the world, physics and human dynamics and the tasks that people care about from pre-training. And ultimately, you want those models to generalize. You don’t want to just be able to perform the tasks that it has been doing training. And the diversity of actions and environments that you have with a pre-training video dataset is much larger than, what you can realistically capture manually.

Vibhu [01:03:54]: How is the scale looking like for the post-training? Like, do you still wanna do, is it like roughly ninety percent of the compute in regular video diffusion model and then scale up a lot, or do it like we want different robotic models for different tasks, or just the one base really good world model can also apply to robotics?

Anastasis [01:04:14]: So currently, we are post-training our models for specific, embodiments that we for particular partners. So if they have a particular single-arm robot or a bimanual robot or a humanoid robot, we would post-train our GWM robotics model on their particular dataset. Over time, we see the different variants of GWM unifying. Like, I would expect, if a year from now or two years from now, you have a single world model that can simulate manipulation tasks, it can simulate navigation, which is a lot of the gaming world models are navigational world models. You’re moving around the space, and it will also simulate human behavior. So that’s the character models. So instead of having three different models, you have a single model that’s able to. ideally, you’re able to simulate what it’s like to be in the world. You’re moving around an environment. You’re maybe performing different tasks. you’re talking to other people. And that happens with, the same, a single real-time video model that’s generating that.

Vibhu [01:05:17]: Do you think you can solve self-driving? So if you are learning to drive a car in a simulator, you have a world model. Your robot is car can manipulate so many axes. How far off are you from something like that?

World Action Models, Self-Driving, and Learned Policies

Anastasis [01:05:31]: So world models

Vibhu [01:05:32]: Or a really good ADAS system?

Anastasis [01:05:33]: World models are definitely being applied to, self-driving, research right now, mainly for evaluation use cases, but our focus has been more on robotic manipulation. We’ve done some work on AV, world models as well. but yeah, we do think that world models are and video models are the best starting point for both simulators and also policy and the action models. So that’s, that’s the other side to this, is that once you have a great world model, then you can just add an action head, and it can predict actions as well. One way to think about it is if you take the starting frame of a scene with a robotic arm and you ask, you prompt the model, generate the arm picking up an object, it would And if it generates an accurate enough video, then it should also be able to generate the exact poses, in 3D that the arm should take to perform the same action. So this is the direction that’s now the popular term for it is world action models, which is you’re starting from a video model, and then you’re adding an action head to predict the actions, and it becomes a policy, essentially.

Swyx [01:06:43]: One thing I’m also impressed by is how much data you need to train these kinds of models. You probably can’t say exactly how much, but like, the original, diffusion models, and from what I know, even of the open source Chinese models, it’s not that much data. Isn’t it surprising?

Anastasis [01:07:02]: What do you define as much data?

Swyx [01:07:05]: Yeah, and it just comes, goes in. Is the token count still relevant?

Anastasis [01:07:09]: So it’s a bit more complicated and,

Swyx [01:07:11]: What is just gigabytes, right?

Anastasis [01:07:13]: Yeah, hours of video, right?

Swyx [01:07:15]: Yeah. Yeah. I feel like something that’s interesting is it seems like the, let’s call it tokens to param counts in language models has really, maybe they’re three years ahead or whatever, seems to be a lot higher than, video models still, even though technically video has more information, per bit. I don’t know if it seems intuitive or maybe there’s just a lot of, like the variability between a pixel to the next pixel is not that high. So, like, maybe there’s just a lot of information that is repeated.

Scaling Video Data and the Lucid Dream Test

Anastasis [01:07:47]: My answer would be it’s still very early. Like, the training video models will scale way further than it

Swyx [01:07:55]: Yeah

Anastasis [01:07:55]: Currently is, and you’ll have capabilities that go much further than the current models can do. So one thought experiment that, I like to use, it’s, it’s almost like the Turing test of video models or like the Turing test of world models, go, I call it the lucid dream test. It’s you have a

Swyx [01:08:14]: You mean the actual person lucid dream?

Anastasis [01:08:17]: It comes from this idea

Swyx [01:08:18]: Lucid rains, right?

Vibhu [01:08:19]: Lucid dreams is telling you’re dreaming while you’re

Swyx [01:08:22]: Yeah.

Anastasis [01:08:23]: Yeah, exactly. So lucid dreaming is when you realize you’re

Swyx [01:08:25]: In a dream

Anastasis [01:08:26]: Inside a dream, and then you

Vibhu [01:08:28]: Play around

Anastasis [01:08:28]: Be able to control what happens in

Swyx [01:08:30]: No, there’s also an inference guy called Lucid Rains. Yeah. Or quantization

Anastasis [01:08:33]: Very prolific, person. Yeah. So let’s say you have a VR headset and you’re in a room with and you’re wearing a VR headset, and that VR headset, most of today’s VR headsets have a pass-through mode, so you can see directly what’s in front of you in the world, or you can render something inside the VR headset. And there’s gonna be a point where those interactive real-time video models become good enough where you wear the headset and you’re in the same room and you’re walking around and you’re kinda and you’re interacting with objects. You’re able to move freely in that room and do, and interact with any object. And at the end, someone asks you, “Did you were you using pass-through mode, or were you -- or was this, rendered or generated, footage?” And if you cannot tell for sure if that was what you were seeing as you were interacting with and moving around the world was generated or it was, pass-through mode and was just what was happening in front of you, that’s an indication that the models have become good enough. And we’re not, we’re not close to that yet. And a lot of it is just this idea of really simulating dynamics and counterfactuals well. Like, if you ask a video model to generate a person scoring a goal versus a person failing to score a goal, it would do a better job at scoring the goal because there is a bias from the training distribution. There is a lot more videos of the person succeeding at scoring the goal. But if you have an interactive model, you want it to be able to generate counterfactuals. Like, if I take this action versus this action, you want it to generate equally realistic outcomes. so that’s, I think, the big gap between video models and world models is that idea of the counterfactual generation. And if you want a great model for robotics, you wanna simulate failure very well, because whether you’re using it for evaluation or you’re using it as a in an online RL loop in the future, you wanna be able to have the model try and fail to do things and improve. and so in order to do that, you need to be able to simulate things failing.

Swyx [01:10:42]: This is the only domain where you have too many successful examples and not enough bad examples. Should be easy to generate failure.

Vibhu [01:10:51]: Oddly enough, I think, like, early image video models weren’t good at being human realistic, right? Like, you see aa lot of the high-res 4K, like, professional photography, but not just everyday life, like normal picture, right? Everything looks like it’s professionally generated, like professional pictures, but not just like normal, like, messy cables on a desk.

Swyx [01:11:13]: Okay, so there’s, there’s this stuff. one thing we also covered that you guys have, video agents that you launched. I guess, how does the traditional, let’s call it frontier, like, autoregressive LLMs, like, feed in, to all this? They’re driving ro your robotics models, or are they driving others, your video agents, production, anything where you see the overlap of autoregressive and diffusion, let’s call it?

Counterfactuals, Failure Data, and World Model Evaluation

Anastasis [01:11:41]: Yeah, so harnesses are really important across all those different use cases. So we have this video agent, which is essentially an LLM that is very effective at tool use of different, image models, video models, and helps you through creating a project end-to-end. So, very often in, like, a traditional advertising flow, you have a brief, you start from it, and then you generate some a storyboard, and then you generate the video. A video agent and, or runway agent helps you through that whole process, and it helps you also analyze performance data. For example, how well did this ad perform versus this ad, and then generate me more of the based on those learnings, figure out what to generate. We think that the harness is a very important piece of the pipeline. as I mentioned, all the video production, all the production video models use some prompt completion that happens, and we expect, that to become more and more complex and more, you generate longer and more detailed descriptions before you use the diffusion transformer. I do think eventually, there’s increasingly this unification into omni models where you have the you’re training the models end-to-end to both do autoregressive text prediction and also, diffusion as well. So you’re predicting the next token, of like you’re, you’re maybe using some reasoning and planning of the scene, and then you’re passing it into the diffusion head that’s generating the pixels.

Video Agents, Harnesses, and Omni Models

Swyx [01:13:10]: Yeah. I think currently maybe only Gemini and Qwen do it. I-I’m not sure which of the Chinese models are omni, but yeah, it’s, it’s not, it’s not a very well, popularized modality, I guess.

Vibhu [01:13:25]: It’s an interesting use case when you think about it, right? Because not only do you have to end at like language model reason, diffusion had generate, you don’t have to output there. You can go back in to feed that output to the same model, reason again on improvements, and it can do a lot of loops just in its own. I guess the question is like, do we need that or can we just do agent scaffold, like do it outside the model? Is there a big benefit to doing it in?

Anastasis [01:13:54]: I think there’s generally the trend of something is first done by a harness and then it becomes part of the model, right? So you had the chain of thought prompting where you had to do this super detailed system prompts to

Swyx [01:14:07]: Yeah, step by step

Anastasis [01:14:08]: Get the output. And now the model generates the reasoning trace by itself before it gives you an answer. And in the, in video models similarly, a lot of the video models of the early days were single-shot video models, and you had to use some orchestrator to turn, generate multiple shots in parallel, and then turn it into an actual video.

Swyx [01:14:28]: Or in ComfyUI, just all over the, all these nodes.

Anastasis [01:14:31]: Yeah, like a spaghetti workflow. and now you have multi-shot video generation where you have the you directly generate multiple shots. And there is a benefit to that because then the video model learns some. to generate a single shot well, you need to figure out a lot of stuff about the world. to generate multi-shot video well, you also need to get some, like, video editing instincts. Like, you need to figure out what is the right pacing of shots. And also, LLMs are not that good at it. Like, they’re not that great video editors. If you ask a LLM to take some videos and then auto-create a edited video out of that, it would feel uncanny. So I don’t think LLMs are that good yet at being video editors. And I think there’s benefit to learning that end-to-end. so I would expect, the training generally is the things that, you need the harness for eventually get injected into the model itself, and you learn that end-to-end.

From Harnesses to End-to-End Learned Video Editing

Swyx [01:15:33]: Do you find that you need to hire engineers who can. or researchers who are also artists to infuse that taste, or do you have artists in residence to distill them?

Anastasis [01:15:44]: We have a large creative team that’s very actively involved in the, in training those models, like on the, in every part of the way. And like, how do you caption video as well so that you capture the stuff that you need for, like, the cinematography, the aesthetics, the camera direction in as detailed ways as possible so that you’re able at inference time to elicit that through the model? we have our creative team also does a lot of evaluation of like, what constitutes a usable video out of those models. And so they’re very involved through every part of the process. And I think that’s one of the special things of Runway is just that mix between like creatives and researchers sitting by, side by side and working together to build the next generation of our models. I think that’s been a really important piece to, how we’ve operated as a company.

Swyx [01:16:37]: Yeah. In some senses, though, you can only do this in New York.

Vibhu [01:16:40]: It’s

Swyx [01:16:40]: Maybe, you have other offices, but like, I try to find some poetic, significance in the fact that you are a big New York company.

Anastasis [01:16:49]: As there’s a few parts to being New York. there is that intersection of all those different industries and, like, media, advertising, like

Swyx [01:16:57]: Yeah, this is very advertising.

Anastasis [01:16:59]: The, like the art scene is New York. Not to say anything bad about San Francisco, but, it’s. There is more going on. There is that component, and there’s also, I think we benefit from being outsiders and thinking of things a bit differently, like not being in the same, like, hive mind of,

Swyx [01:17:19]: BВС

Anastasis [01:17:19]: ASI, of Bay Area and, like, taking. and also taking our time to get where we are today. Like, building the, growing the team intentionally and bringing people who are, yeah, both on the creative side and also on the engineering research side. There’s huge talent pool of amazing people in New York, so that hasn’t really been a problem.

Creative Taste, Artist Feedback, and Runway’s New York Advantage

Swyx [01:17:41]: Congrats on everything. what are you hiring for? what should people look forward to, for the future of Runway?

Anastasis [01:17:49]: We’re hiring across the board. I think this is probably the most open roles we’ve ever had in the history of Runway. we’re growing our research team quite significantly. So if you’re, if you’re excited about video models, if you’re excited about world models, if you’re excited especially about robotics, the robotics team, we’re hiring roles in the robotics across, software, hardware, and research. so definitely reach out.

Swyx [01:18:13]: And, a lot of people don’t have direct robotics background, but what should they have, if they want to be useful in robotics?

Anastasis [01:18:21]: So ideally, some experience with learned policies, would be

Swyx [01:18:26]: Just RLs

Anastasis [01:18:27]: Good for robotics. but we tend to hire generalists as a philosophy and, like, people who learn really quickly. but some experience in the, in domain expertise in robotics is something that we’re, we’re definitely looking for the next months. and then we’re scaling the go-to-market team significantly. There is, a wide, like, very active enterprise adoption happening around video models at the moment, and, we’re really trying to, respond to all the demand.

Swyx [01:19:00]: Yeah. Great. You wanna talk about the, open source robotics stuff?

Vibhu [01:19:04]: Sure. It was just random notes we had.

Vibhu [01:19:07]: NVIDIA launched Cosmo. I guess it’s interesting. So, you’re a founding member AI labs to build open source world models in physical AI. - Anything else to talk on here is open research?

Anastasis [01:19:20]: The biggest thing is that, as I mentioned, while models are still, early, like there is still so much that we you can scale and those models further, so much more advancements and things that we can figure out and how to improve those models further. And I think this is, it’s important that some of this research happens in the open and figuring out what is some incentives for different companies to come together to bring some of that research into the open and open source. And so Cosmos Coalition was a initiative that we co-founded with NVIDIA to bring some of that research as open source. And that could mean open weight model releases. It could mean benchmarks that measure physics and things that people care about when building world models. It could mean infrastructure. So really, how do we grow the ecosystem of world models and make that something that also it’s easier for a developer, a researcher that’s just starting out that is excited about world models to contribute to the field.

Hiring, Robotics, and Enterprise Adoption

Swyx [01:20:19]: I think it’s a there’s some amount of like, is this also our response against the Chinese world models that are being released, or is there not part of the consideration?

Anastasis [01:20:29]: I do think it’s, it’s important for NVIDIA models, if you look at the leaderboards of video models, I would say right now the majority of models at the top ten, top twenty are Chinese models. There is, only a handful of companies that are made it to the leaderboard from like the US or the West.

Swyx [01:20:50]: Yeah. We’re doing better with images, but with video we’re very behind, right?

Anastasis [01:20:53]: And so I think it’s definitely important that we invest more broadly as a community to make sure that we can those models can we have competitive models

Swyx [01:21:02]: Yeah

Anastasis [01:21:02]: Out there.

Swyx [01:21:03]: But like what’s to stop us from just distilling from them?

Anastasis [01:21:06]: I don’t know if that’s the best long-term

Swyx [01:21:08]: Not gonna mention that they won’t

Anastasis [01:21:09]: That you’re bounded by the performance that you can. It’s, it’s almost a bit of a pessimistic

Cosmos Coalition and Open World Model Research

Swyx [01:21:14]: Like

Anastasis [01:21:14]: View that you can get better. you can

Swyx [01:21:17]: It’s free data. it’s, you might as well. Like if they’re, they’re doing it for like, for the text language side, they might as well do it for the video side the other way.

Anastasis [01:21:25]: Yeah, I do think we’re, we’re quite capable of training great models

Swyx [01:21:29]: Okay

Anastasis [01:21:30]: Without distillation at the moment. Yeah.

Swyx [01:21:32]: Yeah.

Vibhu [01:21:32]: So anything you have to say on benchmarks and evals? Like, I feel like what I’m hearing is a lot of people really like arenas for video and image models, customers and whatnot as well. They only want the best on the leaderboard, and they refer to arenas a lot more than language models seem to do. But any notes on benchmarks, what’s lacking? How does the average person compare while these both look really hyper-realistic? More than that, outside of we did talk about like robotic simulation, the physics and all that, but anything to say?

Anastasis [01:22:05]: I think it’s the opposite in some ways. I think people, generally creatives and artists and marketers, other like people that are using our platforms, I think rely less on, arena scores. And it’s, it’s just so easy to, generate with a bunch of different models and then compare the results visually. Like one nice thing about image and video models is you can immediately tell with your eyes like what feels good from an aesthetic standpoint. Like any artifacts, any issues with the physics of those models, you can immediately tell. and so that’s it’s easier, I would say, to evaluate, as a human. there is also those models than it is in language models where you have those very complex math and coding and, tests where it becomes a lot more harder, I think, for humans to evaluate and can discriminate between the performance of models at a time. So I think in practice, people just test out the same prompt with a bunch of different models and see what the results look like. And right now in Runway, you can use our models and you can use third-party models as well. So it’s, it’s very easy to do that.

Benchmarks, Arenas, and How Creatives Evaluate Models

Swyx [01:23:12]: Amazing. We’re gonna end with the AI Runway AI Summit. The last societal issue, I guess, I don’t know if this is a thing, is the, you are at the tension between artists and creatives and AI. A lot of people in that community hate AI. the people that are in the Runway community don’t mind using tools. it’s just another brush. But, how have you seen the sentiment change?

Anastasis [01:23:37]: Our perspective, yes, it’s just another branch, brush. It’s just another camera. It’s, it’s the latest of a long generation of tools.

Swyx [01:23:46]: Technology in art.

Anastasis [01:23:47]: Technology.

Swyx [01:23:47]: Yeah.

Anastasis [01:23:47]: And art and technology have evolved together. I think there’s been a pretty significant shift over the past few months, and it came. some of it you can see with a lot of public figures speaking out in favor of AI and being, like in Cannes, you saw a few directors speaking in favor of AI. We had Ron Howard in our film festival. There is, Mark Scorsese also adopting AI models. So you have more of those stories coming out every day of like a well-known figure, speaking in favor of AI. And it’s just a matter of, in my mind, it’s those models are becoming more and more demystified. I would say I have also a bit of a hot take that one of the things that made the initial response to those models maybe a bit more heated than it needed to be was this idea of text to video of, you have a single text description and you get back a -full video.

Artists, AI, and the Evolution of Creative Workflows

Anastasis [01:24:48]: Yeah, there was a misconception. you can generate it to our feature-length film, but the models of today now take a lot of references. They take they are very controllable. And I think when people see a tool that allows, affords many degrees of freedom and control, they respond to it differently. And it matters less that it’s a generative model than the fact that you can steer it to the direction that you want. and so. I think when people look at, complex workflows on top of those models, when they look at, all the ways in which you can steer them and you can provide now with some of the latest models up to fifty references, like the conversation becomes a bit different because it feels much more like a

Swyx [01:25:34]: Storyboard

Anastasis [01:25:35]: A tool

Swyx [01:25:35]: Yeah

Anastasis [01:25:35]: Versus, like, something that a magical entity that figures out, like, the, your entire film for you.

Vibhu [01:25:43]: Any notes on, like, workflows changing for people in the field? Like, I think engineering at least has had a lot of people where they’re like expectations have changed. I’m, ten X, a hundred X more productive, and you can get a lot more done. same thing as, you’re making dev tools for creatives. any notes there? Like, there’s some people that don’t wanna adopt, some that do. Like, anything?

Anastasis [01:26:08]: Yeah. So I think, in terms of, like, what people care about, I see that we have gone through a few stages. So we started from a stage where the main thing that people were looking for was quality. Like, as, we scale those models, the quality improved dramatically. That’s something that people still care about, but it’s, it’s now in addition to controllability, like being able to steer those models with references, with, different kinds of inputs, with storyboards. And now my sense is increasingly people are gonna care about latency more and more. As those models become better, the ability to iterate very quickly becomes more important. And, like, if you can, with a single prompt generate ten different, outputs, like, almost instantly, you can explore way faster than before. And you get some of the magic that characterized the creative tools of the past, like Photoshop was instant. and we lost some of that with generative models. You’re waiting for two minutes to get back a video, and I think we’re gonna bring, some of that back now with the

Vibhu [01:27:08]: Real-time

Anastasis [01:27:08]: Real-time models.

Vibhu [01:27:09]: Yeah. Exciting. And

Latency, Real-Time Generation, and the Future of Creative Tools

Swyx [01:27:11]: Exciting. the last thing we’ll plug is this one, Runway

Vibhu [01:27:14]: Summit

Swyx [01:27:14]: Summit. You’re finally doing this in SF?

Anastasis [01:27:18]: Yeah. So, we’re very excited about this. So this is, in late September thirtieth, we’re doing a summit on, primarily focused on physically high and real-time video generation. We have panelists from NVIDIA, Physical Intelligence, Botco, DeepMind. Yeah, it’s gonna be, I think, a very interesting series of conversations. We try to make the panels really technical and, elicit actual substantive discussion and hopefully some interesting disagreements and interesting debates on things. And, yeah, the there’s tickets available. Hope people can join.

Swyx [01:27:56]: Since you mentioned it, what disagreements and debates should people think about, or do you expect?

Anastasis [01:28:04]: So it’s things like, there is, one debate right now in the robotics world is, VLA’s versus world action models.

Runway AI Summit and the Big World Model Debates

Swyx [01:28:11]: Okay.

Anastasis [01:28:11]: So there is labs that are really betting on one of those two directions. there is like what is the best source of data to train robotics models?

Swyx [01:28:21]: There’s just the third-party, first-party that we talked about.

Anastasis [01:28:24]: Yeah. There is, the people who really believe in further scaling teleop data versus leveraging more large-scale video data. So that, those are some of the. And then there is, the world models debates of predict pixels directly versus something like JEPA versus a more 3D-based, 3D-based approach. so I think we’re at a nice time in world models because there is still that active debate happening on, like, what is the best long-term direction. I feel very strongly that it’s video predict pixels directly and scaling video generation models is the right approach. But it’s, I think there is a lot of interesting, debate happening, by researchers on, like, what is the best path to take.

Swyx [01:29:09]: It’s interesting that it’s all on, like, let’s call it the policy layer and the data model layer. Is the physical side is completely solved? Like, all the sensors, all the actuators, all these things are. We have everything that we need?

Anastasis [01:29:23]: I don’t think that’s, solved either.

Anastasis [01:29:25]: It’s definitely,

Vibhu [01:29:27]: Different problems.

Swyx [01:29:28]: It’s, it’s like

Anastasis [01:29:29]: Yeah

Swyx [01:29:29]: I wanna dream about all these things, and then I get, I buy a robot or I buy, I try to assemble my own, and I can’t even get the motors to, like, work right. Right? Like, and it’s you’re dealing with very sensitive, equipment that has, voltage and power and, like, heat and all these things which, you, abstracted away. We’re sitting here, we’re talking about software and talking about models, but, like, really you have to deal with those kinds of things too.

Anastasis [01:29:56]: Yeah. And, I think I’m, I’m, I’m generally also not opposed to incorporating other modalities into our models like we’ve seen.

Multimodality, ImageBind, and the Maximalist World Model

Swyx [01:30:04]: Yes.

Anastasis [01:30:05]: The simplest case is they can generate video and audio at the same time. So they can generate RGB, and they can also generate, they can generate sound and audio. But my. I’ve written about this as like what does the maximalist version of a world model look like is you’re incorporating more and more modalities from the universe And you’re training a model on different scales of observations as well.

Swyx [01:30:28]: X-rays.

Anastasis [01:30:29]: And so, yeah,

Vibhu [01:30:30]: You got a good essay that people should read on

Anastasis [01:30:33]: Yeah. Yeah

Vibhu [01:30:33]: Real-world.

Swyx [01:30:33]: No, Meta released a model that was, like, six modalities in one, right?

Vibhu [01:30:37]: Yeah.

Swyx [01:30:37]: I forget what the name of the thing was, but it was like, yeah, okay, depth is one of them, but depth is like a transformation of RGB in some sense.

Vibhu [01:30:45]: ImageBind.

Swyx [01:30:45]: ImageBind, yeah.

Vibhu [01:30:45]: Yeah.

Swyx [01:30:46]: What other modalities? They had heat?

Vibhu [01:30:47]: Audio, depth, heat, text,

Swyx [01:30:51]: Whatever IMU is.

Swyx [01:30:52]: I do think, like, you might as well do ultraviolet. You might as well do, like, just whatever other modality you feel like, ‘cause it’s all data to the model.

Anastasis [01:31:01]: Yeah. And, a big bet is also that there is transfer between all those modalities.

Swyx [01:31:05]: Yeah. Yeah.

Anastasis [01:31:05]: So one of my favorite, examples, which is quite old at this point, is there was this fine-tune of, Stable Diffusion that was called Riffusion Which was

Swyx [01:31:14]: The music one. Yeah.

Anastasis [01:31:15]: Yeah, just fine-tuning, Stable Diffusion on spectrograms.

Swyx [01:31:18]: Spectrograms.

Anastasis [01:31:18]: And it became a quite capable music generator. Right? So there is probably Spatial patterns, so like spatial-temporal patterns if we’re talking about video that emerge at different scales and different modalities. And so there is some degree of, meta-learning that the model has done that allows it to learn faster if you start from a just a model trained on images and train it to predict audio than if you train from scratch on just audio. and there is some other interesting examples. So there is this project called The Well. It’s, it’s a dataset of physics and numerical simulations in physics and biology and a bunch of other domains. So it’s, so it’s essentially different physical systems across very different scales of space and time, from like astrophysics to low-level like atomistic interactions. And we’ve seen. we’ve done some work on this, and we’ve seen that we can take our video model where, real-world video looks nothing like this, and you can fine-tune it on those numerical simulations and just treat them as RGB frames. And you get reasonable performance much quicker than if you just train from scratch.

Anastasis [01:32:36]: Yeah.

Vibhu [01:32:36]: I think we’ve seen this across languages where

Swyx [01:32:38]: Yeah, DeepSeek-OCR as well.

Vibhu [01:32:40]: Yeah, DeepSeek-OCR.

Swyx [01:32:41]: Like, you don’t have to tokenize text. Like, you can just throw them in as images.

Vibhu [01:32:44]: There’s a lot that happens in that base pre-training. Like, there was an argument a long time ago of people saying, “Oh, humans have so many, sensory representations, right? Smell, touch.” Models have a whole two more modalities that we’ll like, that we don’t even have data for. And it’s like, okay, you take AQI sensor, like you can try this stuff, but there’s so much happening in just the base trainer on that you don’t get as much from these little things.

Scientific Data, Cross-Modal Transfer, and Omni Models

Anastasis [01:33:10]: Yeah, exactly. And I think that’s what it solves is data scarcity.

Vibhu [01:33:13]: Yeah.

Anastasis [01:33:13]: So you don’t have as much. You have so much video data available, but you don’t have, like olfactory data that

Vibhu [01:33:21]: The cool thing is it goes the other way too, right? So if you wanna do physics, like if you wanna measure this or you wanna have a diffusion model do audio, it transfers really well. So like in your case, the little bit of post-training for robotics gets a video model to use its fundamentals in another domain. So we can apply that to other stuff too.

Anastasis [01:33:40]: Yeah. And if we look at, like how do you make those models more useful for in scientific domains, and if you look at AlphaFold, they’ve had all these very. Because of the data, the limited amount of data that it needs to be trained on, it’s it’s very fine-tuned architecture just to solve, protein structure prediction. But if you take all those disparate sources of scientific data and you bring them together under a single model, like I think that’s an approach that can help us solve new kinds of problems across science by leveraging all the learnings from one modality or one set of, data to another. So very early days for that direction, but I do think that’s where ultimately what the end game of simulating the world is. You’re not just using RGB. You’re using RGB as a starting point, but you can incorporate more and more modalities of the universe and leverage the transfer that happens from learning from one to the other.

Vibhu [01:34:42]: I guess the follow-up there is what’s the drawback of omni? Like, why is everything not an omni model? Also, why not now, and why. Would you start from language backbone or image video backbone and then go omni from there? Does it matter?

Anastasis [01:34:57]: Yeah. We need to take it one step. We need to solve robotics first, and then we can go into

Swyx [01:35:02]: Solve everything now.

Anastasis [01:35:05]: Yeah. I do think there is a lot of open-ended research that needs to happen for, those omni models. There is a lot of things that require careful consideration when you’re bringing multiple modalities into a single model to predict. But I think, I expect those to be solvable.

Closing: Film Festivals and the Future of AI Video

Swyx [01:35:23]: Wonderful. you’ve been very generous with your time. Congrats on all your success, and, yeah, I’m excited for the, AI Summit, or physical AI Summit.

Anastasis [01:35:32]: Yeah, thanks for having me.

Swyx [01:35:33]: And yeah, and people should check out the film festival if it’s in town, right?

Anastasis [01:35:37]: Yeah.

Swyx [01:35:37]: Yeah. You’ll be gonna be touring all over the place.

Anastasis [01:35:39]: Yeah. Next year we’re probably gonna do that. So we do film festivals every May or June of

Swyx [01:35:45]: Yeah.

Anastasis [01:35:45]: And we did the last one in New York, LA, Tokyo, and at the AI Engineer,

Swyx [01:35:52]: Yeah

Anastasis [01:35:53]: Fair.

Swyx [01:35:53]: Yeah. Yeah.

Anastasis [01:35:54]: So yeah, hopefully even more places next year.

Swyx [01:35:57]: No, I think like someday, you will be hosting the Oscars of AI video, and, I think people should like take this very seriously as like a potential career they can have.

Anastasis [01:36:07]: The Oscars of AI video will be called the Oscars.

Swyx [01:36:10]: All right. All right. Thank you.

Anastasis [01:36:14]: Thank you.

The OpenAI → Hugging Face attack has people asking “what else do we need to worry about?” and Anthropic’s filters flag two things: cyber-security and biology. The natural question is: what about bio-security, then?

Clem Delangue argues that cyber-warfare defensive capabilities need to be open and to keep pace with frontier models’ attack capabilities

Radical Numerics co-founder Eric Nguyen sat down with us and explained why the same models that increase biological capability can also keep defense from falling behind.

Building a virus from scratch

While he was at Stanford, Eric couldn’t get traction on Genomic Language Models (GLMs) for a long time. Biologists didn’t believe it would work, didn’t think they could verify the output, and didn’t see important applications beyond what they could already do. He kept pushing, eventually helping lead the development of Evo and contributing to Evo 2 at Arc Institute. Those models were later used by a separate Arc/Stanford team to generate entire bacteriophage genomes that were synthesized into functional viruses!

Long context unlocks biological intelligence

Early ChatGPT spit out poems and email, and early DNA language models like Evo and Evo-2 could build a genome from scratch. DNA is different, however, from natural language in that it has a very small alphabet (4 characters ACTG) and that its sequences are very long:

* 60K for an average human gene

* long being up to 2.3M

* the whole human genome around 3B.

Innovation in long-context models made this possible about 3 years ago (footnote: striped hyena), long before the frontier labs were building 1M+ context models.

Now Eric and other AI x Bio luminaries have founded Radical Numerics to build and scale GLMs to tack a wide range of biological problems, extending well beyond generating DNA.

Thinking in DNA

Their GLMs already do pretty well with RNA and protein because there are clear markers in the DNA sequence for genes (RNA sequences the perform many functions) and specific genes that encode proteins. This means that the models already generalize to multiple “languages,” before even attempting to train in other modalities, such as 3d protein structure, epigenetics and natural language.

If a model thinks in the DNA language, maybe it understands the imprint that environment left on different genomes as well? Perhaps the model has learned the functional relationship between different sequences, and could extrapolate to new sequences based on that?

And so what we wanted to showcase was that if we show the model progressively better RNAs in a series of steps with its score, right? So you have like low scores first and then you gradually move up the chain. Can the model continue that trajectory on its own? And then in the final step, does it self optimize to a point where it's like the best score it can get? That was the experiment. Can we do that? And so we took a data set, a large data set of aptamers. We held out a portion of the best performing ones and we showed it only the lower ones, but then we ranked it, right? So we showcase lower scores with the RNA aptamers and then progressively got higher, and then ask the model to just like continue with that pattern. And it turns out it was able to recapitulate some of those higher scores that we had not shown it yet.

So, voila: chain-of-thought, thinking in DNA!

The arms race

But much as long-context inference, chain-of-though and multi-modal perception unlocked sophisticated reasoning in natural language LLMs, these capabilities in GLMs are enabling increasingly sophisticated “biological intelligence,” and along with it, greater danger.

According to Eric, defense is currently losing this battle, but Radical Numerics argues to push the frontier harder!

I won’t spoil the details for you. In the episode we talk in detail about:

* Biosecurity as an arms race — and how defense can keep up

* The genome as the imprint of the environment on DNA

* Going truly multi-modal

* How chain-of-though works when you “think” in the language of DNA

How often do you get to talk to a guest who has both an Academy Award and who invented textbook machine learning algorithms? John Platt has an Oscar, two textbook algorithms, two named asteroids, and an Erdos-Bacon number of 6. This was easily the most fun bio of all the guests we’ve read to date. And the result was an epic and fun chat covering Google’s Empirical Research Assistance (ERA), how AI can help battle climate change, and tons of great stories about the co-evolution of science and AI.

John’s colleague Dave Bacon likes to tease John that his career has been defined by being twenty years early to the next big thing. This may be convolutional neural networks (some credit him with coining the term), fusion research, quantum computing. John and Google have been working on solving some of humanity’s hardest problems with AI and computation for well over a decade now. Recently John and his team set their sights on using AI to solve any scientific problem that can be written down as a score.

Google’s Empirical Research Assistance (ERA)

John’s team has taken on many hard scientific problems over the years. In solving these, they noticed a pattern, many scientific problems can be reduced to what John calls a “scoreable task”. Once you have the score function, the goal is to find some code that maximizes the score. The hard part is in formulating the score, but once you have the score finding the maximizer can still be quite a lot of effort.

John’s team set out to automate solutions to this general problem. This came out of the idea of an “auto-Kaggle” AI, which can solve any Kaggle problem you can throw at it. Kaggle is owned by Google, so all the data was ready and easily available to them!

The result is Google’s Empirical Research Assistance or ERA (paper, github, blog). ERA is surprisingly simple conceptually. Gemini (or your LLM of choice) keeps a running tree of past experiments (notebooks) and where they’re going. It’s a close cousin of Monte Carlo Tree Search: at each iteration the Upper Confidence Bound rule picks which notebooks are most promising to mutate. This is optimistic, not greedy, so sometimes even the fifth-best notebook gets chosen. Gemini then proposes mutations for each one, about ten at a time. The history of each branch is shared, so different leaves can learn from each other.

“It’s almost like having a hyper-eager grad student who doesn’t sleep.”

Evolutionary algorithms have been around since the 70s, but this works because Gemini actually knows where to look! What’s even more interesting is that there was a step change between Gemini 2.0 and 2.5, and this went from just not working to working great.

ERA is so powerful that John and his team solved many outstanding problems with it, resulting in at least ten papers. Some of these were climate change related, which we talk about in the next section.

So, we had to ask: if you have an optimization god how do you avoid fooling yourself? John’s answer is that ERA provides predictive models. It’s up to the scientist to make sure they’re truly descriptive. Some of this just involves good old-fashioned careful machine learning science. “It’s a power tool. It can slice your fingers off.” This led to some fun discussion about Kaggle competitions, and the fun ways people can overfit to datasets without meaningfully solving the problem you actually care about: Google’s contrail-detection competition was won by entrants who noticed a half-pixel error in the labels (is the origin at the corner of the pixel or the center?) and this turned out to be a part of the winning special sauce. Great for winning $15,000, not so helpful if you actually want to solve contrails.

“People themselves will act like these LLMs and try to reward hack. It goes back to Goodhart’s law: any metric that becomes a target is no longer good as a metric.”

His advice for where to start instead?

“Always just fit linear regression. Just do it. Just do it. Just do it. Or SVM.”

Tackling Climate Change with AI

John and his team have worked extensively to mitigate the effects of climate change. We talked about several of their initiatives.

Perhaps the most interesting result we talked about was reducing the effects of condensation trails (contrails) from airplanes. Those little streaks you see running behind planes somehow account for 1% of all human-induced global warming?!? Some of these trails of ice crystals can hang out for days. These crystals are black in the infrared, acting like a thermal blanket that traps heat day and night.

It’s easy to understand what’s happening here, a region of atmosphere becomes “ice supersaturated”, and a tiny bit of exhaust seeds water vapor that instantly crystallizes. The scale here is astounding, with a single gram of exhaust resulting in ten kilograms of ice crystals.

The solution to all of this is quite simple, in principle! We know what parts of the atmosphere are most likely for the trails to form. Just have the planes drop a flight level or two. Problem solved, right? Well, the hard part is accounting for how much warming was prevented. This is a counterfactual problem, parts of which stumped John’s team for over two years. They had a working model for the heat-trapping half, but not for the reflected sunlight. ERA was able to find a simple model with some confounders they hadn’t considered. Cracked it!

Modeling climate generally is a hard problem. Climate is best thought of an attractor of many different possible weather outcomes. This makes it much harder to model.

“Weather is where you are on the attractor, and climate is the statistics of the attractor. The problem with climate is that we’re altering it. The attractor itself is changing, it’s moving.”

John and his team have worked on treating both the symptoms and the disease of climate change, with several other works in the area. Another fun example we briefly cover is FireSat, a way of using a constellation of satellites to rapidly identify fires before they grow too big to put out. For anyone living in California, you understand the problem. In dry years a small fire can result in hundreds of thousands of acres. If you could find this fire when it’s the size of a room, it could be put out. By the time it hits an acre we have a much harder problem.

Where is this all going? Looking forward by looking back

By now it should be clear John has an incredible and unique view over the intersection of science, computation, and AI. John talked about a class on physics of computation he took with Richard Feynman back in 1982. This was when quantum computing was an ill-defined concept with no theory or experimental backing. John recalls every Tuesday was a guest lecture, and every Thursday was Feynman explaining why the Tuesday guest was wrong. John also recalls doing science back when there was essentially no compute, a million operations per second was cutting edge.

What is John’s recommendation: the most important skill is developing deep domain expertise. There’s no other way to develop taste than to tackle hard problems. One surprising part of this is that John recommends spending time doing things the old fashioned way. Play with tools, and just implement things yourself.

“You could drive up the mountain, or you could hike up the mountain, and maybe it’s okay, even fun, to occasionally hike.”

Summing it up, John’s message to the audience is that there will still be a place for scientists, and that if anything it will just open up more opportunities for “the creative stuff, the rigorous stuff, the philosophy stuff.” But don’t forget to spend time doing the grunt work.

“There just seems to be this strong impetus in the world to optimize and squeeze everything out. But you do lose something when you hyper-optimize. It’s overfit.”

And whatever tools you end up using, John’s advice is the same one Feynman gave him forty years ago: you must not fool yourself, and you are the easiest person to fool.

We had a great time talking with John. We hope you enjoy!

Also in this episode

* Fusion is three years away, not thirty, if you ask John. And why the Lawson criterion means every fusion approach has an Achilles heel.

* Why superconducting qubits are still finicky.

* The asteroid he named after his mom, which turned out to have a moon.

* The looming helium shortage nobody talks about.

* How NeurIPS started as people crashing a private workshop at Snowbird, and why Hopfield networks are all you need.

* Being Carver Mead’s sysadmin on a VAX with an 80 MB disk the size of a dishwasher.

* Finding asteroids in 1985 with film, a stereoscope, and a letter to Brian Marsden. The Vera Rubin Observatory found 11,000 in six weeks.

* The Feynman effect: total clarity in the room, none once you leave.

* Quantum echoes, the NISQ era, and why he thinks quantum is neither thirty years away nor tomorrow.

* A startup that wants to inject mercury into a fusion reactor and sell the transmuted gold. “It might not work.”

* John’s 20% time rule for his own group: do stuff for learning, and you don’t even have to tell him what.

Tickets for AIE NYC now open, and apply for the invite-only AIE CODE. Join us!

We have an unusual relationship with today’s guest: for years since coauthoring the InstructGPT paper, Diogo Almeida had been saying that API-available frontier models have been going down the wrong path, everything from the alignment to refusals to reliability perspectives, that we have dropped every mode other than autoregressive chat-tuned LLMs because of the overwhelming success of ChatGPT.

In a launch video now viewed ~40M times (by comparison, GPT4o was 22M, Fable 5 was 15M, Navier Stokes was 74M, and 6 Astra was 137M), Diogo introduced Jev and it immediately took over the AI timeline — we’ll skip full Jev explainers because your favorite AI influencer/educator has probably already done one. We also collected:

* the official patterns and cookbooks you should see first, from Allie

* Jev usecases

* speed based - games and computer use

* the voice + computer use example we discuss at 1h34 mins

* voice + browser control

* The must not miss Doom demo

* Driving cars in games

* Excalidraw

* virtual try-ons

* “Smart Games”/smart NPCs

* guided responses in text messages

* Jev for coding agents has an official guide

* jev for linting

* compacting tool calls

* reasonable pushback from Theo - Diogo has published a note on the Tyranny of the KV Cache that you should read as a followup after the pod for Jev + coding agents, because of his belief that Cache Rules Everything

* Programming Languages built atop Jev (Diogo’s fave)

* Jev for analytics replay and user journey review

* “dark data”

* entity resolution

* natural language search

* “smart software”

* a core goal of Jev is to “disappear into the background” - eg as unremarkable as regex

* Jev as a judge

* Jev memes

* Jev vs LLM capabiltiies

* blending transformers and classifiers

* about the confidence api

* Jev vs GLiNER (note difference/pushback, agreed, agreed, agreed)

* Jev on trolley problem

* Jev Bush

Instead we’ll focus on what we can uniquely offer — a broader philosophical and mission-based understanding of how and why Jev was created, and what you should expect next in terms of future models from TypeSafe (ReasoningJev?) and what usecases and ideas you should work on vs the 55th low effort clone of Jev’s API or doing a generic JevBench benchmark - something Diogo has rejected publicly.

Why RLCD: Three kinds of RLHF, and why they are ALL the wrong north star

Diogo knows a good deal about RLHF, given that he was on the team that pioneered post-training at OpenAI — and traces the three branches to Christiano et al 2017 (the robot backflip demo), Stiennon et al 2020 (learning to summarize) and his baby, Ouyang et al 2022 (InstructGPT). From there on, every innovation from Function Calling to Structured Outputs to Reasoning felt like a hack on top of the string based, sequence to sequence prediction paradigm. As he mentions on the pod, from 2023-2024 he struggled unsuccessfully, due to both personal and organization underestimation, to train a model that accurately addressed what he saw as the core problem with making LLMs the heart of software: reliability.

Jev’s core innovation is "Reinforcement Learning for Calibrated Decisions”, a novel, unpublished technique that optimizes for “answers with epistemically honest probabilities on System One tasks” rather than human rated feedback (RLHF) — which causes hallucinations, sycophancy, and permanent reliance on humans — or programmatically verifiable outputs with rubrics (RLVR) — which solves Navier Stokes but exacerbates jagged intelligence and doesn’t integrate well with other software.

We’ve talked about the calibration problem before on the pod, but probably the single best place to understand why RLCD became necessary is Diogo’s AIE talk, which discusses why a generation of training helpful AI assistants for humans has impaired them for training models for composable, programmable AI for automation.

At the end he also teases his contrarian opinion on scaling laws - which teases how to build a modern neolab without the billions of dollars the major labs have…

The Bitterest Lesson: Tasks and Data beats Compute

We spend a good amount of time discussing Diogo’s essay on the Bitterest Lesson:

His point is that “You get what you optimize for and the bitterest lesson in ML is that the most important part of it isn’t ML at all.” - and picking the right north star, eg upvoting for user preference vs being integrated into tool calls - makes everything else fall in line.

We’re excited to catch up with a freshly dyed Diogo to discuss:

* Why AI can solve extraordinarily hard problems but still fail to automate basic work

* What System One Models are and why Jev is built for software rather than chat

* RLHF, mode collapse, calibration, and the hidden costs of optimizing for human preferences

* Why refusals become a problem when AI is buried inside software dependencies

* Why TypeSafe rejects public benchmarks and optimizes for intelligence per dollar

* The “bitterest lesson”: why the right task and the right data can matter more than compute

* Why TypeSafe thinks of itself as a data lab rather than a model lab

* RLCD vs. RLHF and RLVR as fundamentally different North Stars for AI

* Why reliability and robustness matter more than simple determinism

* Jev’s programming primitives and how intelligence maps into software control flow

* Why developers should decompose AI workflows into small, measurable decisions

* How structured state replaces giant prompts and system messages

* Why Diogo thinks AI should eventually disappear into the background of software

* The “inverse SaaS-pocalypse” and how AI could supercharge existing software

* System One vs. System Two intelligence and the limits of reasoning models

* Dark data, computer use, real-time intelligence, and Jev’s biggest early use cases

* Why Jev could reshape coding agents built around a single-model architecture

* Why Diogo says he wouldn’t pre-train with $1 billion

* The OpenAI journey that led to TypeSafe and why he thinks many neo-labs are approaching AI incorrectly

* Coding agents beyond the KV cache, shared state, sub-agents, and the multi-agent future

Diogo Almeida

* LinkedIn: https://www.linkedin.com/in/diogomda

* X: https://x.com/CompleteSkeptic

* TypeSafe AI: https://typesafe.ai/

Timestamps

00:00:00 Jev Launch Week and the AI Economic Revolution

00:02:50 What Is Jev? System One Models and Programmable AI

00:05:54 RLHF, Mode Collapse, Calibration, and Yann LeCun

00:10:29 Programmatic AI, Refusals, and Safety Alignment

00:17:21 Why TypeSafe Rejects Public Benchmarks

00:20:43 The Bitterest Lesson: Data, Compute, and the Right Task

00:24:59 RLCD vs. RLHF and RLVR

00:28:42 Why Powerful AI Still Hasn’t Automated the Economy

00:39:55 Reliability, Robustness, and Determinism

00:48:11 Model Versioning, LTS, Speed, and Intelligence per Dollar

00:54:04 Inside Jev’s API and Programming Primitives

00:58:28 How to Build with Jev: Structure, Decomposition, and Small Decisions

01:18:28 The Inverse SaaS-pocalypse and AI Disappearing into Software

01:33:21 Computer Use, Dark Data, and Jev’s Biggest Use Cases

01:38:48 How Jev Could Reshape Coding Agents

01:41:00 AI Safety, Frontier Pacing, and the Limits of RLVR

01:48:03 Why Diogo Wouldn’t Pre-Train with $1 Billion

01:55:19 The OpenAI Story Behind TypeSafe

02:01:41 Why Diogo Thinks Most Neo-Labs Are Getting AI Wrong

02:08:00 Coding Agents Beyond the KV Cache and the Multi-Agent Future

Transcript

Introduction: Jev Launch Week and Developer Momentum

Swyx [00:00:00]: Okay, we’re in the studio. A special occasion because this week, Diogo, my good buddy, launched Jev, and it’s been taking over the complete timeline. How do you feel? What’s it like to be you right now?

Diogo Almeida [00:00:16]: Emotionally?

Swyx [00:00:17]: Yeah.

Diogo Almeida [00:00:17]: Never been worse. Like, I’m a ragged corpse of a person right now because there’s so much going on, and I’m like a technical CEO, so I have, like, a lot of fires to fight.

Swyx [00:00:29]: Yeah.

Diogo Almeida [00:00:29]: But mentally, I feel—I say this all the time, and I’ve been saying this kind of for years in my over-under events. Like, I feel like the entire AI field is like one of those, like, carnival house of mirrors, and everyone is just insane and saying the weirdest stuff that doesn’t make sense. And it feels like for just this week, like, I’m on a better in sync with reality and like, oh, people see it now. AI can be so much more than what was once thought.

Diogo Almeida [00:01:06]: And like, yes, we are going to make. Like, an AI-based economic revolution is back on the table, and this is f*****g awesome.

Diogo Almeida [00:01:17]: I’m so jazzed the developers get it. It’s, it’s, Yeah, and I want to show my eternal gratitude to the developers and

Swyx [00:01:25]: Yeah.

Diogo Almeida [00:01:26]: I’m so jazzed about the community and everything. It’s so great.

Swyx [00:01:28]: Yeah, you were saying yesterday that you decided to prioritize the town hall and not a bunch of, like, VIP, investor-type people because you wanted to make sure that they are the people that you get your most, attention, right? The engineers, the developers.

Diogo Almeida [00:01:43]: Yeah, it felt a little like, oh man, I’m talking to, like, really important people right now.

Swyx [00:01:47]: Yeah.

Diogo Almeida [00:01:47]: I probably shouldn’t reveal who.

Swyx [00:01:48]: Yeah.

Diogo Almeida [00:01:48]: But it feels a little bit dirty for me to, I’m, like, perhaps overly genuine in things. Like, it feels, like, dirty if, like, in my gigantic calendar event of people to talk to, the community isn’t one of those.

Swyx [00:02:04]: Yeah.

Diogo Almeida [00:02:04]: And actually, in my ideal world, it would be, like, community all the time. I was thinking, “Should I host a town hall while walking to your studio?” And I’m like, “No, that’s too crazy.”

Swyx [00:02:12]: Sure. Yeah. Well, you guys have been hosting town halls on Discord. Discord is now 100,000 people. Your Twitter’s

Diogo Almeida [00:02:19]: I don’t follow these stats.