World Models, Explained

Summary

---

Y Combinator Startup Podcast: "World Models, Explained"

Date: July 17, 2026
Podcast: Y Combinator Startup Podcast
Host(s): Y Combinator Partners (Speaker A & Speaker B)

---

Episode Overview

This episode dives deep into "world models"—the AI research concept seen as a breakthrough path toward sample-efficient, general intelligence. The hosts (YC partners) discuss the history, mathematical foundation, applications in domains from games to robotics, and the current state and future directions of world models. The conversation builds up from basic definitions of sample efficiency and reinforcement learning to advanced neural architectures for intelligent agents, all while relating these ideas to the broader quest for AGI.

---

Key Discussion Points & Insights

1. Sample Efficiency in AI

• Definition & Human Contrast
  • Sample efficiency describes how quickly an AI model can learn new tasks from limited data.
  • Humans are incredibly sample-efficient, mastering new skills rapidly, while most AI models need huge datasets.
  • "[Sample efficiency] is basically if I have one additional sample in my dataset, how much more intelligent am I getting?" (B, 00:49)
• Inductive Bias and Human Advantage
  • Humans enter new tasks with strong inductive biases from life experience and education; AI models do similar via internet-scale data, but still lag.

2. Theoretical Basis for World Models

• Perfect World Models: The Physics Analogy
  • If you had a "perfect world model," you’d need no real data—think how physics lets NASA plan asteroid intercepts without trial and error.
  • "It's called Newton's second law of motion... That is an example of a perfect world model we've built." (A, 03:14)
  • Visualization: Practicing basketball mentally—studies show imagined practice increases performance almost as much as physical practice (CogSci study by Richardson, 1967).
• The Brain as a World Model
  • The neocortex may have evolved for complex world modeling—improving intuition, adaptation, and skill acquisition over evolutionary timescales.

3. Reinforcement Learning (RL): The Foundation

• Control Problems and State-Action Modeling
  • Six-dimensional vector examples for drones, mapping physical state (position, velocity) and control (thrust).
• Transition Function (World Model) & Policy
  • Transition function: predicts the next state given the current state and action.
  • Policy: chooses the optimal action for a given state.
• Model Predictive Control
  • When a perfect world model exists (e.g., Newtonian physics), convex optimization methods can yield optimal trajectories (e.g., landing a drone or SpaceX rocket).
  • "This is convex. The sum of convex functions is convex... it will just give me out my policy, which will be the solution." (B, 09:12)

4. When the World Isn't Predictable

• Introducing Stochasticity (Non-differentiability)
  • Real environments (multi-agent settings) are non-differentiable; opponents' actions can't be predicted by gradients or convex optimization.
  • RL emerges as the necessary tool for these cases, introducing concepts of value functions, rewards, and planning.

5. Games as World Modeling Test Cases

• Chess and Go:
  • Both have massive state spaces but small action spaces.
  • Success in AlphaGo leveraged efficient search (Monte Carlo Tree Search, MCTS) and powerful value/policy estimation.
• Why AlphaGo Doesn't Generalize
  • As action space grows (e.g., “Go” on a 1000x1000 board), the computational cost explodes—AlphaGo’s method no longer scales.

6. Scaling to Real-World Domains (Self-Driving, Robotics)

• Self-Driving Cars
  • State space is "effectively infinite" (camera, weather, road state).
  • Action spaces, even for simple vehicles, are orders of magnitude larger than Go ("36,000 actions... 100x bigger than AlphaGo"). (A/B, 40:53)
  • Real-time constraints: no time for “slow” planning as in chess/Go.
• Robotics
  • Action spaces skyrocket with each added joint or end-effector.
  • Data collection is expensive, cross-embodiment gaps make transfer learning hard.

7. Model-Free Versus Model-Based RL

• Model-Free: Directly maps state to action ("behavior cloning"), efficient when data is abundant, but struggles with rare or novel situations.
• Model-Based: Learns a world model (state transitions), enabling synthetic data generation, test-time planning, and more powerful policies.
  • "The main thing that I believe is that this is required for AGI. This is what the human brain is." (B, 44:43)

8. Modern World Modeling Techniques

• Early Attempts and Innovations
  • Jurgen Schmidhuber's "World Models" trained agent policies entirely on hallucinated (‘imaginative’) rollouts.
• The Dreamer Series (Danijar Hafner et al.)
  • Sequential models trained on observed transitions, later "action conditioned," then used to generate synthetic data for policy learning.
  • Dreamer V4 achieved a breakthrough in complex environments (e.g., mining diamonds in Minecraft) using mainly synthetic data.
  • "The policy is so good that it's the first paper to mine diamonds in Minecraft... that's extremely difficult." (B, 52:04)
• Video Diffusion & Flow Matching
  • Modern models (VideoDiffusion, Sora, Gaia, etc.) leverage vast video data, adding small amounts of action conditioning for effective cross-embodiment and real-world generalization (robotics, self-driving).

9. Latent World Models & JEPA

• The Scalability Problem
  • Directly modeling every state-action pair is intractable as systems scale.
• Joint Embedding Predictive Architectures (JEPA)
  • Models compress high-dimensional observations to manageable latent spaces; learning/predicting transitions in this space enables "world modeling at scale."
  • "The key idea in JEPA... is basically taking this Q learning idea and compressing the state into a lower-dimensional space, then predicting in that space." (A, 58:17)
• Architectural Innovations
  • Additional loss terms (e.g., VICReg, SimSiam) prevent model collapse and improve stability.
  • Extends to both vision (pixels) and language (tokens).

10. Open Problems and Future Directions

• Physics-Informed Neural Networks (PINNs)
  • Currently limited in their ability to handle rare/out-of-distribution events or sufficiently high fidelity.
• Test-Time Adaptation
  • Human brains adapt rapidly to novel/friction environments—AI models are still far less flexible.
• Real-Time Constraints and Planning
  • For true real-world deployment (e.g., FSD), planning and policy generation must be real-time.
• Sensor Limitations
  • Robotics lags in tactile sensing; human epidermis offers much richer input than available machine sensors—key for delicate manipulation.
• The ‘Squint Test’
  • World model architectures are approaching the “looks like intelligence” threshold, but critical gaps remain (especially in offline/‘sleep’ learning and resilience).

---

Notable Quotes & Memorable Moments

• On Human Sample Efficiency:
  "Humans do this incredibly well. We can learn new games, concepts and skills, often after just a handful of tries." (B, 00:08)
• On the Nature of a Perfect World Model:
  "Imagine I had a perfect world model, then I should never go to the environment to go and collect samples... We do it all the time. It's called Newton's second law." (B, 02:37)
• Mental Practice as World Modeling:
  "They just blindfold them and they imagine laying up a basketball, they improve 23%. Against the control. I mean, that's insane." (B, 05:08)
• On AlphaGo's Limitations:
  "Why AlphaGo doesn't scale? The cardinality of the action space must be extremely small. If it's big, sad." (B, 35:07)
• On the Path to Robotics:
  "This is basically the path... A bigger world model helps for training and for test time planning. And I think it'll be the same for robotics." (B, 54:46)
• On Tactile Sensing:
  "It's amazing how effective our epidermis is... I can touch something and say, oh, this is smooth, this is rough. We don't have any of that." (B, 73:20)
• On the Squint Test & Intelligence:
  "If I squint, I look at the human brain and I squint and I look at all these world models... it's getting really close." (B, 69:41)

---

Important Timestamps & Segments

| Timestamp | Segment/Topic | |-----------|--------------| | 00:00–00:39 | Sample efficiency in humans vs. AI—defining the challenge | | 01:33–02:23 | Human inductive bias & limitations of state-of-the-art AI | | 03:14–05:07 | Perfect world models; Newtonian mechanics and mental simulation | | 06:33–12:08 | RL basics: state, action, policy, world models in control | | 13:03–15:17 | Non-differentiability, reinforcement learning strategies | | 19:30–24:30 | Chess & Go as RL environments; Monte Carlo Tree Search intro | | 31:29–36:26 | Computational cost and action space limitations of MCTS | | 37:19–42:39 | Applying world model concepts to self-driving and robotics | | 44:43–47:15 | Evolutionary role of world models in the brain & AGI case | | 49:02–54:46 | Modern breakthroughs: Dreamer, video diffusion models, data bottlenecks | | 56:38–58:17 | Neuroscience links; world models for entire ‘real-world’ systems | | 60:01–62:54 | JEPA: efficient latent world models & their issues | | 63:49–69:41 | Open problems: fidelity, adaptation, real-time, and the “squint test” | | 73:04–74:14 | Tactile feedback in humans vs. robots, and future research challenges |

---

Conclusion

The episode presents an in-depth look at how "world models" are reshaping the field of AI, especially for robotics and complex decision systems. The hosts connect theoretical insight, current practice, and neuroscience, charting both exciting progress (Dreamer, video diffusion, synthetic data) and remaining hurdles (sample efficiency, real-time planning, out-of-distribution adaptation, tactile sensing). If robotics, general AI, and the future of embodied intelligence matter to you, this episode is a must-listen for understanding what's on the frontier, why world models matter, and where research is headed next.

---

Transcript

[A]

One of the biggest open problems in AI right now is how to solve sample efficiency. That is how do you get models to quickly learn new tasks or skills from relatively small amounts of training data?

[B]

Humans do this incredibly well. We can learn new games, concepts and skills, often after just a handful of tries. Our best models, on the other hand, often need tens of thousands of data points just to learn.

[A]

So today we're going to discuss what many top researchers believe is the most promising path to closing that gap. World models.

[B]

We're going to discuss the motivation and math behind world models, current applications, and why this approach might be the key to unlocking AGI.

[A]

You and I have talked a lot about the various ways people are training models and the sample efficiency of them. Why don't we start by just defining sample efficiency and how we intuitively think about it as humans?

[B]

Yeah. So I think from my perspective, the two major problems that we have left to solve is intelligence per watt and intelligence per sample. Intelligence per watt is like how many valve perplexity points we get per watt of spend. And then intelligence per sample is basically if I have one additional sample in my data set, how much more intelligent am I getting? And so if imagine I have a new tasks like arcgi for example, I think like really Francois Chollet has been on the forefront of this thinking and talking about intelligence as a rate of skill acquisition versus skill acquisition. And that's very different. And so how fast do we get smarter with more and more samples? And these things are incredibly poor at getting smarter with fewer and fewer samples.

[A]

And for context, the RKGI test sets are a really good example of cases where humans are intuitively very good at them. Most humans can intuitively solve those puzzles with some amount of thinking and effort. But our current state of the art AI systems, what people consider frontier intelligence, basically can't do them right.

[B]

I mean we come into new problems with such inductive bias From K through 12, like all these math and school that we've had that these models are kind of getting from the entire compressing the entire Internet. And so when we come in, we're not coming in tabula rasa, just like bare bones. But even so that they have, I don't know what percent of the Internet you've read. I've read very little percent of the Internet. But despite that and having read the entire Internet, it still can't really do well and generalizing to these new tasks.

[A]

So now let's think about this. In the extreme cases, in the extreme case where let's say we were perfectly sample efficient. We were as sample efficient as possible. What would that mean in terms of a model that is taking a set of actions in the world?

[B]

Well, I guess the perfect sample efficiency would be zero samples. And there are examples of this and that sounds absurd to say, but. And the example, the hypothetical I'll give on this is, imagine I had a perfect world model, then I should never go to the environment to go and collect samples to train on and. Well, that can't possibly happen. No, it actually can happen. We do it all the time. It's called Newton's second law of motion. It's like Newton mechanics. We basically know how to get an object from point A to point B with a rocket quite easily just by following Newton's laws of motion.

[A]

Yeah. Like when, when NASA plans to intercept an asteroid and is planning it, you know, years in advance and can set it off in a trajectory where it just glides to the right thing and intersects to the right point. That is an example of a perfect world model we've built where we're then just letting that world model act. And that that system does not need to intelligently collect new samples from the environment to decide which direction to go next. It can already, it's already been pre programmed and can perfectly do it.

[B]

Yeah. Can you imagine if like we needed to collect 1 million training examples of like us shooting spaceships moon to know how to do it? We definitely wouldn't have the Apollo missions. Right. But we do have that ability because the real world is differentiable and we can do something called model predictive control that we're going to talk about in a little bit, but even in our own brain. I was just thinking about this on the drive up, but there's so many ways that I can basically think about the things that you are going to say or what a VC is going to say when I was pitching them, or what a customer might say. And even product having taste. What is taste is like predicting that other people are going to like this thing. And so we've built this world model over years of entrepreneurship, 10 years of like getting it wrong, right? That maybe Bill Gates, Steve Jobs And Jensen have 50 years of, you know, world modeling experience to know what people want. And basically this is actually proven in the 1967 CogSci study by Richardson that basically showed that if you take a cohort of three different people, three groups of people, and you have one go practice layups in basketball and they go and they shoot, they, they improve for one hour. They improve by like I think it was like 24% or something like that. And then if you take the other one and they just blindfold them and they imagine laying up a basketball, they improve it 23%.

[A]

Interesting.

[B]

Against the control. I mean, that's insane. It means that we have this crazy good world model. And there's this neuroscientist at Stanford named Shaw Druckmann, who basically is of the view that the entire point of the growing neocortex during the great cortical expansion 10 million years ago was to get better and better and better and better at world modeling. And having just like my little vla, which we'll define of doing the predicting the next action is not as good as having a world model to lean on, either for training purposes or for test time adaptation.

[A]

Yeah. What it fundamentally comes down to is we as humans, we think about our intuitive ability to think as coming from some implicit world model we have in our heads encoded by genetics and our ability to learn and whatever else. It seems like models can do surprisingly intelligent things despite not having an explicit world model. When it comes to natural language, when they're just talking, it seems like maybe under the hood, deep inside the weight somewhere, there's some kind of implicit understanding of the world, but there isn't an explicit representation of that. But it seems like in certain domains, especially in robotics and self driving, as we'll talk about, that sort of breaks down and maybe it would be helpful now to just think a little bit about and just sort of define some of the pieces of what makes it challenging in these different domains. And then we can use that to kind of build up to why it's particularly hard in things like self driving and robotics to get these types of predictive models to work.

[B]

Yeah, let's do it. So let's actually take a step back and just talk about control reinforcement learning and define some common terms. So typically we teach a course called Decision Making under Uncertainty, which is like the main reinforcement learning course at Stanford. I like to show a specific example of, let's say I have some drone and this is my poor little drone here, and it has a M and we know that gravity G is pulling down on it and it's currently at position T with velocity T, which we will collectively call the state. And to be really clear, this is going to be PX PY PZ tt, then VX vz.

[A]

It's like the six dimensional state vector

[B]

and we have some thrust vector U that we control and we're trying to get to some point P And V star, which is V star is typically zero. And so you have some platform that I want this thing, this drone to land on. So this is this control problem. Right. And so let's say this is like. And we'll go through optical or optimal.

[A]

Optimal.

[B]

Yeah, optimal control. So how would I actually solve this? So the first thing I need to know is my transition function. And so this is my state transition function, which is ST plus one given the previous given st. And my action, which I control is ut. And so this is my state transition or dynamics function or a world model. This is a world model.

[A]

This is like a very fundamental for context, you know, this equivalent to a transition function you would think about in RL in general.

[B]

Exactly. And then what I'm trying to learn is something called a policy, which is like, what UT should I emit given some st? And so this is the ultimate question, what should I do? What action should I take given some state st? And so the way that we'll solve this, and luckily we have a world model that is perfect, that is called Newtonian physics. Newtonian physics. This is like Newton's second law of motion, which is f equals ma. And so we know that the position pt plus 1 is going to equal pt plus delta t vt plus one half delta t squared. So everyone's taking high school, high school physics.

[A]

Yep.

[B]

And the same thing for the velocity delta ta. And then my acceleration is the sum of sum of the sum of the forces, which is going to be my ut. I think I divide by the mass and G. And so that's it. And now I have my transition function. Now how do I get to a policy? And I'm going to apply something called model predictive control or real time model predictive control, which is like the way that SpaceX lands the rocket on some platform in the ocean. And what you're going to do is you're going to set up your loss function, you're going to minimize sum over all T. You have UT to infinity. And I'm going to minimize my P star minus PT plus V star minus vt. And usually you add this little lambda ut, which is like how much energy you're exerting. And you can't have infinite thrust, so you typically will have to say UT U max thrust. That can be achieved. And so this is easily solvable with convex optimization. And so this is convex. This is convex. This is convex. The sum of convex functions is convex. This is a convex constraint. And so DCP discipline, convex programming means that I can put this into cvxpy and it will just give me out my policy, which will be the solution will be the optimal ut +1 all the way to infinity.

[A]

So we can solve this in closed form. Basically because we have this world model of Newtonian physics, we can say at every step exactly how this drone should fly so that it lands on the appropriate under a set of constraints like max thrust available.

[B]

You'll run your log barrier or interior point, whatever to some solver on this and it will give me my optimal. Then this would be literally the optimal path that this thing can take to get to this state. And that will minimize. And I can increase this if I want it to do the least energy path and I make that zero if I want it to be the fastest. And so that's typically the way that you would do what we'll call deterministic. Differentiable control. And why differentiable? Because I can take the. I can form the Lagrangian by taking this minus this constraint and take the gradient of it.

[A]

And I can do.

[B]

Martin Robinson.

[A]

You use the fact that it's differentiable to do the optimization.

[B]

Exactly. If this is non differentiable, you cannot do convex optimization and you cannot do SGD even if it's non convex. You could still solve and get a pretty good solution, as we do in deep learning. But if it's non differentiable, you can't. There's nothing you can do.

[A]

So yeah, let's have an example then of how you could make this non differentiable. Like, well, what's a, what's a scenario? I guess even it's like this drone scenario where it now becomes non differentiable.

[B]

Yeah. So I'll put this adversary named Unkit.

[A]

Okay.

[B]

And, and your job is to, you have another drone, let's say Ankit's drone is to try to hit me and stop me from getting there.

[A]

Now, from the position of your drone, you don't know what actions I'm going to take.

[B]

Right. And so now let's just call this the, this would be. Now we're definitely not deterministic. We're stochastic and stochastic and non differentiable.

[A]

Yeah.

[B]

And in this case my state transition, what is ST plus one? It's going to be my state I'm in now, my thrust and what unk it's going to do.

[A]

Right. And these, it was all differentiable until this new variable.

[B]

Yeah. And I can't like back prop through your brain to say what you're going to do with your little drone controller. Right. It's completely non differentiable now. And I'm resorting and I have to resort to this awful area called reinforcement learning, which is just super brutal and it's sprawling and there's so many different things. And you'll hear things like when you study initial reinforcement learning called value iteration or policy iteration, and there's DQN or deep Q learning or just Q learning, there's actor critic, there's all this stuff.

[A]

All of this stuff ultimately comes down to ways to estimate, to model this non differentiable stochastic process.

[B]

Exactly, yeah. And so that's basically the main thing is you're going to start talking about this as a model where I'm going to introduce this psi to say that this is going to be some model that's going to take in these things and then output this and that. We're going to train it over many instantiations of this and that's to get a better and better world model. And then I need to train some policy at st. And then typically you also need a value function and that is the value of some state and to discern between the value of different states. And like in this case, I don't know what a valid state is, but like, let's just say I was doing like SpaceX with launching rockets and landing rockets in Florida. Let's just say that like there's different. If I have my launch pad here and I have a whole bunch of houses here, let's just say the path going from here to here, I may think that doing this and then coming across here and burning all these houses alive may be not highly valued. So I might say as an example, they typically call this like some kind of a cone here. And I might say like it's low value to be here and it's very high value to be in this cone or something.

[A]

Right. In a sense, the value gives you some expectation of future rewards, like the sum of future rewards you're getting. And so if you're in a bad space, you would set the value to zero or negative infinity or something.

[B]

Yeah, so we should introduce RT as well. And so typically if you're playing Go or chess, winning the game, you can say winning the game is plus one, minus one for losing draw a zero. That's what's done in AlphaGo. In chess we have these heuristics like a pawn is worth one point, a rook is worth five, etc. So you can already have reward is the difference in board state. And then this, yes, will be the sum of my discount. Should just do T. No, of rt. Yeah, given. And it's important also to use this nomenclature vpi. And the reason why that's important is because what's actually happening here is this is the discounted reward following policy PI.

[A]

Correct.

[B]

And that means that when I'm in this state I will take this action and then I'll end up in this to SC1 and then I'll take this action and it's it and taking it greedy. And so that's the value with respect to PI. Yeah.

[A]

And so ultimately what it comes down to is we are trying to still find a new policy PI. And along the way we will use machine learning models in various capacities. This is standard RL to estimate the value function given the rewards we're receiving. And then where world models come in is a way of incorporating all of those into some sort of joint modeling of the state and action distribution so that we can make more intelligent policies off of it.

[B]

Right. And so your standard kind of setup for this is what I'm always trying to get to at the end of the day is some joint distribution which would be ST +1 given where I'm at now. Where I'm at now. And then this factorizes with chain rule simply to my PI, my policy at given ST and my world model I'll give this. This is usually represented with theta. And this is my world model which would be ST +1 given STNAT.

[A]

Yeah.

[B]

And so. And these are typically learned separately. And like. And like you can imagine, in fact actually you can actually learn this. This is a video generation model and I have the frame ST and I predict the next frame ST plus one.

[A]

Right.

[B]

And then. And we'll get into this.

[A]

Yeah. For those of us who kind of saw our diffusion model series, often people these days use video diffusion for exactly this.

[B]

Yeah. And then what you can do, and this is like the in vogue thing to do since Danajar and the Dreamer paper series from V1 to V4 is do action conditioning later like similar to clip where we will inject this like input head or input tail to come into the model to influence and enable the world model to have embodiment. What does that mean? It means that not only can I predict like as a plant or tree growing on the side of the building, I can like see the world passing by, but I can actually influence it and I can change the world and I can learn that with at least. And that's far fewer samples to do this post action conditioning if I already have a really good ST to STB one World model.

[A]

And so here you're saying what's also in vogue now is jointly training these versus separately training them.

[B]

Exactly. So that is called a world action model. Where some of the issues here is one, there's all these training dynamics. If these things are disparate training on different sets and things like that. The other issue is plainly obvious. What I have to do to actually do test time planning is I'll have to sample my with model one, invoke theta and then pass that sampled action into here and then roll it out to ST plus one. And it's very expensive and it's very not real time. Two major issues. And why, like why can't we just scale up AlphaGo to solve all the problems is because of this property. If I have one invocation to the model and it gives me both, here's the action I should take and here's the ST plus one. That'll end up much, much cheaper and much, much faster.

[A]

Okay, so I think that's a really good segue. I think why don't we now motivate everything we just described through a series of increasingly complex environments. So I'll contend that I think the right set of environments for us to consider is chess, followed by go, followed by self driving, followed by robotics.

[B]

All right, so let's go through a couple examples of problems that we want to apply reinforcement learning to. So chess is a pretty easy one. There's an 8x8 grid. And so typically when you approach any RL problem you're going to look at star. And so the size of the state, the number of states I can be in. So if I have these eight here and these eight, so this is 8, 16, 32. So it'd be 32 to the 64.

[A]

Yes. Quite large. Quite large.

[B]

Then my transition function is stochastic and non differentiable because you can, you don't

[A]

know what the other player is going to do.

[B]

So if I'm like in playing Chess.com at my house, I move and then something happens and it comes back and then now you moved and the board has changed. So I can't really differentiate through what the other player is doing. The card in my action space is actually quite small. Even though there's 32 pieces and all that stuff, there's only eight possible moves in expectation that you can actually that are legit moves.

[A]

So like in any, in any given state there's only 8ish moves you could do.

[B]

Let's just say in the beginning I can move all my pawns. I Can move my horses. So that's 10, that's like not that much. So this is extremely small. And then my reward, we can use the heuristic based approach or we can just say, you know, plus one zero or minus one if I lose, plus one if I win. And so this is very tractable.

[A]

You say it's tractable even though there's a really big state space here. Yeah, but why don't we talk about that for just a second? I think this is a really important point. I think when you say it's tractable you're specifically referring to the action space being small because it affects the kind of like combinatorial expansion here. Should we talk about that for just a second?

[B]

Yeah.

[A]

Or maybe we can add go and then kind of contrast the two.

[B]

Yeah, so why don't we do that? Because it's because I want to get to the alphago the way that they solve this. And you're right. So if I were to do this naively and I just took my SC1 and I want to do lookaheads, what I would do is I would take all of the actions I can take. So there's eight. So I would do action one, action two, action eight and then each one of these I need to expand it for all possible states. And so now I need to do cardinality S which we just said is this huge friggin number. And so I have to do that eight times and I have to do it again. I have to do it again. So just looking forward, one move is like quite intractable.

[A]

Although at the same time, you know, everyone starts at the same starting position. And while it is a really large space, you know, there isn't an infinity number of potential. There's actually a relatively small number of game boards, even four moves into the game as opposed to a game board where you could start in any permutation for example of initial game state and what, a few states down the line.

[B]

Yeah, so this is definitely over done because it's much, much less than this in practice. But just naively looking at what possible game states could be as a rough math here, but this is roughly the idea. And then each one of these leaves I need to invoke my value function which is the value of that state T +1. And so I have to do that all many times and we'll get this, off we go. But like this ends up being estimating the leaf node because at the end of the day my policy atst I want to pick, I want the argmax of like the value of the following.

[A]

The argument action I guess would be an A here.

[B]

A. Exactly. Yeah. Argumax over A of the value of the state of the end state. ST +N. Let's say that's the main goal here. And so for me to do that I need to roll all this out, estimate the value and then pick the best one. And so this quickly grows however, and we'll see this with AlphaGo, which actually has an even bigger state space. So I think it's 19 by 19 and I'm.

[A]

I think it's about right now.

[B]

So you have this 19 by 19 grid, you can. In each one it can be black, white or, or nothing there. So I have three. So let's do our star again. So the cardinality, the state I think is going to be S2 or 3 itinerary thing here, the 19 squared, I think it's 361, something like that. 361. My transition, same issue. I don't know. My action space is going to be 361, let's say.

[A]

So it's a good amount bigger than chess. Much bigger, but it's still not enormous. Yeah, as we'll see in a second.

[B]

Yeah. And so basically what they do, they call this Z, which is kind of annoying, but let's call it R. And it's the terminal, it's the terminal winner that won the game. And they basically, you know, you have your trajectory which is S0, A0, R0, then all the way to the end of the game.

[A]

Yep.

[B]

SN A N R N. And if you won, then all of these, all the moves that black, if black won, all the moves that black did get, plus all the moves that white did were minus one. And that's how they create their rollouts.

[A]

Rollout refers to a taking n steps of play of all players, one after another of moves under a specific policy at the particular instantiation of it.

[B]

Right. So let's just probably under this policy, P theta T and we're going to overload T. But this is that instantiation. We froze that model. We froze that model and we play, I think it's like 70 games and we treat all of those and we're going to subsample a bunch of these state action results. State action results to train our, to update our policy in our world model, our transition model. And what it's actually doing is we take in an st, we give it to some theta and it wants to output the probability of ST +1 being played, which is our transition function and the value of the current state.

[A]

And how do we get the value

[B]

and so the value of the current state? Well, both of them are coming out of the model. But basically the loss function L theta is going to equal and it's going to be eerily close to this control. Problem one is we have some V theta minus this z, which we'll just call it R here squared. And then plus actually, sorry, it's minus this PI, which I'll explain in a second log P theta. And I think they, everyone includes this, but they include it in the paper. So I include it there as well, which is the weight decay. So this is basically what our loss function is. Then we'll play a bunch of these games and let's try to be a little bit organized here. And so this is our setup, this architecture. And now once we train this thing, we do an insane, insanely expensive task of test time planning. And so this trend in RL is just called test time planning.

[A]

And a specific algorithm they use here for this is Monte Carlo Tree Search mcts.

[B]

And so this is one of the possible things that you could do. It ends up working extremely well if you have small action spaces.

[A]

Yeah. So let's like very intuitively talk about what MCTS does. And a lot of people have heard about Monte Carlo tree search because AlphaGo was such a big moment. But how exactly does that map into our star and value function and policy?

[B]

Yep. So I'll take this st. This will give me 361 numbers that sum to one. And so I'll have some probability of where these things are going to go of where my opponent will play here.

[A]

So these are like the sets of actions.

[B]

Yeah. So I'm here so that I have all my ST +1s. I'll have 361 of these things.

[A]

And then to be clear, this is like action one, action two, all the way to action 361.

[B]

Exactly. Yeah. And we have to estimate the value of each one of these. And so then we have to invoke the model all 361 times to give me values for each one of these things. And then I will select, I'll select it based on the ucb, the upper confidence bound, which is this equation that is roughly something like balancing my value function of st +1, which they're going to in the literature. It'll be called a Q value because it's actually the difference between a value function and Q value is just that I have the action as well.

[A]

Yep.

[B]

So it's be st then at. So we'll just call that qvalue, which is my exploitation term. And then my exploration term will be something like this funky square root of N. So it's the argmax of A of my Q. And then I have this, which is the probability of this, this move being played, which we have from here of S. Let's just call it ST&1. And then I have this term which is this sum over NSB divided by nsa.

[A]

And what's the intuition behind this term?

[B]

These N's is the visit count during my MCTS process. So this whole tree I'm going to.

[A]

So this tree could get really big, right? It's 361 per thing.

[B]

It's a depth of 30.

[A]

So you can't visit every single leaf node.

[B]

Exactly. And so you want to keep track of which state did you end up in and what action did you take when you were in that state. And you want to make sure that you have good exploration. Right. And so the way you keep track, the way you ensure that you have good exploration is you want to not just be greedy and always pick the highest value one, because that could be local. Very myopic. And so what you'll do is during this MCTS process, you'll start this dictionary which will be all zeros of the visit count of being in this state and taking this action. And then once you go through your first rollout, you'll go here, all these things will be in it to zero. You'll have some probability, we're going to bias it towards the higher probability of places to go. And then we'll expand those trees and then we will update the counts that we visited this and that will basically reduce the amount of probability that we're going to select it again because this will reduce my exploration term. And if it's highly valued, then we're going to increase the Q on this because this is the expected value of going down this path.

[A]

The gist of it is fundamentally like you want to take the optimal ish path but have enough exploration in this really expensive step you're doing here so that you are making sure you're getting a decent chunk of the other potential leaf nodes you could traverse too in these 30 step rollouts.

[B]

And so I'm going to do this MCTS simulation 800 times here and then for all 800 I have to go through this whole process and I have to invoke the model like at least 30 times to get through all here. And so that's, you know, 27,800 times 30. Yeah, 24,000 invocations of the model to develop this tree. And then once I have per step just to do one action into the game. A lot of people don't understand that this is like you don't like store this MCTS tree, you like, you throw it away after you make the move. But once it's very expensive to develop this MCTS tree and once you have it, the probabilities of traversal are actually extremely useful for training. And then you end up biasing it and you train it with the MCTS tree, which is like a little bit seems like circular motion or something like that, but you end up treating that as the PI that you'll train in your loss function. So we have the R of did we win or lose? We have the pie of what was the end result of this whole expensive process. And then at test time we are going to do these 24,000 steps, every single move to pick the ARGMAX that satisfies both exploration and exploitation.

[A]

In this case, this still feels somewhat tractable though, because the action space is small enough where this kind of works.

[B]

Exactly.

[A]

Now let's say hypothetically maybe we can draw an imaginary game of Go where it's like, let's say this game of Go was a thousand by a thousand. And so now you have a equals more or less a million. And now this tree we're drawing here, that has to take here, this has cardinal or width I guess 1 million. And there's S0 through S1 million. And the number of steps you would have to take here presumably have to be way more than 800 in order to get any reasonable kind of sampling of this. And so you're probably multiplying the test time cost of doing a rollout or of doing a next step prediction. Astronomically, if the game was even, let's say, you know, this is only 100x bigger than the current game or not even 50x bigger than the current game.

[B]

Everyone was very excited about AlphaGo and at the time in what was this, 2017, 2016, everyone's very excited about this. And the important thing to pick up is that we did 800 MCTS simulations and to cover 361 possible actions on average. So that gives us about two samples roughly on an expectation for every single action.

[A]

So here you need like 2 million of them for a similar depth.

[B]

For a similar depth. And then that's still to do a depth of 30. I would still have to do this times 30. This had be 60 million invocations of the model. So that better be a small model. Right. That's a lot.

[A]

So yeah, that's to do a single action, to be clear.

[B]

Yeah. So Exactly. To do one action. So just imagine. So why AlphaGo doesn't scale? To me there's one. The cardinality of the action space must be extremely small if it's big. Sad. Two, I need a perfect deterministic environment. Right. Like this doesn't change. The rules of this game don't change. But the rules to the stock market change all the time. The rules to venture change all the time. Like the real world changes quite often. So like homoscedastistic. And real time. If you saw the movie, the documentary is such an amazing documentary. I'd highly recommend it to anyone that's really good that watches it. The guy is sitting there for like 60 seconds, maybe five minutes waiting for the computer to decide. And it's kind of like imagine we were driving a car and you took like 60 seconds to turn the steering wheel. Everyone's dead, the whole car is dead. And so now let's talk about robotics and self driving car and why that approach kind of can scale.

[A]

Yeah, I think the really good contrast here because intuitively I think in thinking through this exact star layout, it actually really changed how I think about the kind of problem space of both of these two. So like let's take self driving car as an example. This is one many people have started to experience for the first time because we have some self driving cars that actually work. You have Waymo and Tesla, FSD and whatnot that seem like they kind of work. So let's maybe apply your same star framing here. I would contend that the state space of self driving car is enormous and it's actually not intuitive to me whether it's more or less large than this one. Right. I mean in a sense the Chess and Alphago state space is already like more than the number of atoms in the universe or something to that effect. But like just to emphasize that here, you know, you are considering, you know, surroundings, vehicle state.

[B]

Yep.

[A]

Like you know, camera, like weather. Weather. I guess the point is like road conditions, it's like massive. This is massive.

[B]

For all intents purposes is infinite.

[A]

Yes, for all intents and purpose it is infinite. Correct. Yeah.

[B]

And so is the space of pixels. Like you know, like what can I put in an image? I can take an image of anything.

[A]

Yes, true.

[B]

And so we're able to handle it. And same thing here where we compress from the board state. We don't represent the board state. We compress it with a comnet so they have some deep, some deep convnet that actually takes this state and converts it into a latent.

[A]

Right.

[B]

And that latent compression is sufficient to kind of like do pattern matching, do some type of symmetric equivariance kind of things. And same thing with this and even better with JPA which we can talk about at the end there which is like basically taking some type of state space and doing all of our optimization in the latent space which stable diffusion did. That works extremely well which reduces our state space dramatically because I'm in some latent high dimensional space.

[A]

So the key thing there is that despite this state space being effectively infinite, we've actually gotten really good at compressing this. And we'll talk more about some of the tricks for how we actually do this in practice here. But the TLDR is where there's 10 years of deep learning work that basically makes us extremely good at compressing that very fast.

[B]

Exactly right.

[A]

Exactly. T seems to have a similar problem as before.

[B]

Right.

[A]

In fact it may be even more extreme. There's like infinity other variables around you.

[B]

Right. In some ways you'd think that it's. This is physics. Newton's laws of motion should apply. If I turn the steering wheel like this and I hit the gas I should be able to really easily model this. But what is non differentiable is that I have if I'm going into a circle. Right. It's like the most the biggest issue that that we faced in when I was doing self driving car is like you're imposing your will onto maybe driving in India I think is what I'm imposing your will onto the environment. And like people just kind of adapt naturally. Like if you were doing Newton's law of motion you were going to collide and so that the optimal policy if you were being strict Newtonians here would be like don't move because anything you do you're going to crash. But it's not true like that then we wouldn't function like cars wouldn't go down the road. And so you have to model the environment. You have to include other people in the environment and understand the embodiment of like how your action will change other people's actions. YC's next batch is now taking applications. Got a startup in you apply@y combinator.com apply it's never too early and filling out the app will level up your idea. Okay, back to the video.

[A]

Now let's talk about the action space. You know like one way to look at the action space is that it seems relatively small. Seems like well you know, you turn the steering wheel left to right, you hit the brake, you hit the, you hit the gas. Doesn't seem that big but like how big is it actually? Like how do we actually represent these action spaces when it comes to a realistic self driving car scenario?

[B]

Yeah, I don't know how they, how they do this nowadays. They're doing a whole bunch of like bird's eye view, different things like that.

[A]

That's considered even just like a very simplified case. What do you have?

[B]

You have a steering wheel that you can turn left. Right. You have a brake pad and you have the gas.

[A]

Yeah. And so I guess this thing is like 365 degrees. So it's like a 1 to 365 let's say or 0 to 365.

[B]

And let's just say you break this up into 10 different severities you're already

[A]

at even with just this oversimplified model. Your action space cardinality is 365,000. So that's like 100x bigger than alpha Volt. In fact it's about the size of the example or in fact a decent amount smaller than the size we said break in CPS.

[B]

And so yeah, so 36,000 action space is very large. And then even worse, unless you're Tesla, we have a bunch of video of people driving cars. We don't have video of like dash cams and like that. Like you actually don't have. Again, only Tesla has this of the action as well. Yes. The things that you have access to your trajectories are just like St. St +1. Yes, St +2.

[A]

So you're saying there's a decent number of these that's from like dashcam footage on YouTube or something. But not really that many either. Yeah, relative.

[B]

If you wanted to do a self driving car and you didn't want to go spend a million dollars trillion dollars on going collecting all this data, then you want to leverage this data somehow. And this is going to be really applicable for robotics because we have a lot of videos of people doing things especially with egocentric. We have those videos but what we don't have is the actions they take.

[A]

Yeah. So this is like this is a sequence that you're showing here.

[B]

Unless you're Tesla.

[A]

Unless you're Tesla and Tesla has this.

[B]

So this is a huge competitive moat of like what do people do in that state and then so you can behavior clone to go from here to here, from here to here, go here to here, et cetera. But even then it's still very, very difficult. You have to, it's not sufficient. People think that like okay, I have this, we have a self driving car. Right. I mean the amount of work that they're doing at FSD is like incredible. And it's not generally available. Like you can't, you know, it's not waymo level yet.

[A]

Would this be a good moment to briefly talk about model free versus model based rl? I think that's an important distinction that's going to be relevant when we talk about more world models.

[B]

Yeah, so this is a perfect point. So model free just means that my policy PI of@given ST, I have no world model involved. It's literally doing what I said. I grab a bunch of these and I go from S to A S to A.

[A]

Just predict the next day, that's it.

[B]

And this is largely called vla. You know, this is like giving us pretty good results. It's behavior cloning. It's all the stuff that it's not getting us to Rosie the Robot just

[A]

yet, but in many ways it's the closest thing that just looks like the next token prediction from LLMs that seems to scale pretty well with natural language. I mean it's not exactly the same thing because there's no action. Exactly. But picking a token is not exactly the same thing. But it's very analogous to that like

[B]

basic thing that's I basically take away the tokenizer head and I give it an action space and I collect a bunch of teleops data, you know, like this as, as the self driving car does in Tesla. And I just take in the the state which is some image and, or maybe sequence of images and then I'll output some action and that's it.

[A]

Cool.

[B]

And this is let's say model free because I don't have a model for the environment. And then now if I do model based rl, I have not just some PI but I have also my PSI as well here. And so by including this I can have a much stronger policy. But it would take a lot more time to perform inference because I have to do this full test time planning

[A]

just to remind us that size referring to this specific transition function. Right, it's referring to this. You're saying this is specifically referring to a function of st +1 given st and action t. Yes. So it's like your ability to predict the next state. You'll Be in is the crux of it.

[B]

Yep.

[A]

As opposed to just directly predicting the actions.

[B]

Yeah. And the main thing that I believe is that this is required for AGI. This is what the human brain is,

[A]

at least in the way the human brain does it.

[B]

Yeah. And let me go further in saying that like if you look at the billions of years of evolution, basically there's this thing called 10 million years ago called the great cortical expansion, which you see the size of a brain just explode, get bigger, bigger, bigger, exponentially up until us and it basically stops. And if the entire point of the neocortex is world modeling, what happened is we started from VLAs. This would be like ants, fish or whatever and. Fish, yeah, right. Just like very like, you know, lizard brain, whatever you want to call it. And then we developed this neocortex to like, you know, go from our motor cortex to actually simulate what's going to happen. And that makes us just so much smarter. And then once we get those samples, we can compress it when we sleep or otherwise with this hippocampal shortwave, ripple, whatever you want to call it. And then that helps us develop a better policy. And that marriage between the two not only helps us train on hallucinated examples, but it also allows us to test time, plan.

[A]

I guess the kind of extreme case then of self driving car is kind of general robotics.

[B]

Yes.

[A]

Right. So if you're like a humanoid company, like figure or PI or whatever, again same star setup. I guess the gist of it is that A is now even bigger.

[B]

Yeah, right.

[A]

It is like I guess a very simple robot would be. Yeah. How would you parameterize the action space? Let's take a very basic one.

[B]

If I take my 6 axis arm as your standard here that we're actually working on right now in Stanford Robotics center, you have two degrees of freedom. Two degrees of freedom. Two degrees of freedom. And then you have another two for the end effector.

[A]

And so that's a simple end effector. Not even like a, not even like a.

[B]

It's literally a one axis. Like you know, you can rotate but you have the, the, the, the one axis Yumi style thing. So this is eight. So you have 16 degrees of freedom. And let's just say that you do the 365 divided by 10 or whatever, you know, kind of thing. I mean it's like 10 to the 16. It's like insane.

[A]

Something like that.

[B]

It's an insane number. And so much bigger than self driving car. And even worse, like getting teleops Data is extremely painful and expensive. It's not just like, oh, we'll just get some people in the Philippines, we'll give them like some, you know, things or whatever. It's like totally, totally doesn't work.

[A]

And nor is there yet something like Tesla's fleet where there are cars deployed that people are just using and they're not even necessarily realizing that every time they turn the steering wheel, they're providing this, the steel set for Tesla currently.

[B]

And then even worse, you have this, like, what's called cross embodiment gap. And so if I were to like, train this policy on Tesla Model X and I were to like, put it on a Tesla Model 3, it wouldn't work. No, like, it totally wouldn't work. Like, all the so much, so much of this. The, the way that if I were to break on a Model 3 versus a Model X, the Model X, it weighs more, it has different dynamics, aerodynamics and things like that. And so what's actually going to happen is very different. Like, the degradation you have across crop across embodiments is very, very, very strong.

[A]

And clearly Tesla's figured various ways to get around that. I mean, they have these that roll out, but actually, even with Tesla's new FSD today, they don't roll out in all the cars at the same time, probably for more or less that reason. And in this case, it's even harder now. I mean, you have bigger differences between embodiments than a Model 3 versus Y, and you have way bigger action spaces. You have to somehow model.

[B]

Yeah, Lane Macintosh I played hockey with at Stanford who now runs Tesla fsd. I can ask him, but I would bet money that they shard the data per model, per car type.

[A]

Yeah, wouldn't be surprised.

[B]

I just. Because that's what I would do. There's no way that, like, you know, I would trust, you know, data that was collected on a Model X on a Model 3. I just win. No way. I would trust it.

[A]

Okay, so now that we understand the basic setup here and why the action space problem is so big, why don't we talk a little bit about how world models actually fit into this? You know, maybe first, you know, I guess what didn't work about the naive world models and how do we fix those? And then let's kind of talk about some of the newest world modeling techniques. Cool.

[B]

So, like, in robotics in particular, it's very hard to get these. This kind of trajectories that you want that you kind of need to train for your VLAs and people spend up, you know, with a whole bunch of teleops data. It's very expensive, very expensive. Ideally what we would do is take like data like this from someone who is just like puts a camera on them and just like making sushi. Okay, like I want to make a sushi robot. How do I do it? Give it to all the sushi chefs. Don't put anything in their hands and just have them start cutting up sushi and making sushi.

[A]

And ideally we would train it in that way you were describing of like somehow we would train a model just on these two and then later add this afterwards.

[B]

And so the first real person that you know went after this was Jurgen Smithumer. Please. So he doesn't yell at us, we have to make sure we cite him. But he has this really cool paper called World Models, very aptly named. And it's basically he took these like OpenAI gym, classic games, car racing and I think Doom as well, and then just like trained a model at that time was like an rnn. He had some funky zero order stuff in there or whatever. But basically the key premise was I can take an environment, I can extract a whole bunch of this type of data off of it. I think he actually does actually this data. But we'll get into Dreamer where he does it in this way and then trains a policy on only the synthetic data, the imaginated rollouts. And it actually performs well in the environment. This is the first time in my understanding that that actually happened. And it actually works really well.

[A]

And then so the key thing there is you can basically use this if you have some predictive model of this in that case, and eventually of this, you can use that as basically a synthetic training set to train your policy model and then basically fine tune it on real data later.

[B]

Exactly. And which is just like a really powerful idea, especially since in robotics the limiting step is access to large amounts of state action data. And so now the Dreamer series. So basically this published publishes in May of 2018. Danajar Hafner publishes Dreamer 1 I think in November of 2018. And then now he's been on this rampage for the last seven years publishing these papers. And Dreamer V4 I think is the capstone of it, where he basically does the same thing and he focuses on Minecraft and he trains a world model on this type of data and then injects action conditioning on a very small amount of data to get to this type of world model that has the action conditioning as well and then samples a lot from it and Then trains a policy on those synthetic imaginated rollouts. And the policy is so good that it's the first paper to mine diamonds in Minecraft. I'm not a big Minecraft player, but apparently that's extremely difficult. That's like next level difficulty. And it did it all on synthetic

[A]

data, which is kind of crazy. And the key unlock there.

[B]

Yeah.

[A]

Use synthetic data specifically on a model trained on just this sort of state transition type of thing.

[B]

Yes.

[A]

And this ends up being very convenient because it turns out we as a society have a lot of this.

[B]

Exactly. Yeah. All of YouTube. Right. He does do a very small amount of data to enable the action conditioning and that allows you to do this full simulated rollout. But yeah, it's true. So we have, we have YouTube, we have like Flickr, we have all these data sets online of like, you know, people doing things. We'd like to use it. And no one has really gotten that to work. And then now that with this, these like video generated generation models, we can take that data, create a world model out of it, add action conditioning post train it with action conditioning for some new task that we want it to do. Chopping down wood or you know, making sushi or folding my bed or whatever it is. Only a few amount of examples. And then we can train a policy in this neural simulation.

[A]

Yeah, and we put out a video about diffusion models very recently in flow matching. I imagine that now ties very closely to this ultimately the kind of current state of the art best way to do this on basically infinity data that we have available and can keep generating is using state of the art video diffusion flow matching models.

[B]

Exactly. Yeah. So like if you have your, your Cdance or your Sora or Juan or exactly all those models, like basically the idea is now we have them and they're already trained and they're great. Let's do a small amount, amount of action conditioning on them to get to this, this world model and then we can sample from it a bunch and then train. And this is exactly what WAVE did with Gaia. And Gaia, I think they've raised $1.5 billion to basically run with this idea for self driving car. I think a bunch of companies, Nvidia, this paper here is basically talking about doing exactly the same this dream zero for robotics.

[A]

What I thought was really cool about this paper is that they do exactly this process where they have this joint model of state transitions and actions. They train it by first instantiating it with the open source one video diffusion model. And then it only takes them about 500 hours of teleop data, which is basically exactly this, to get it to be pretty good. And they have a lot of clever tricks that allowed it to be cross embodiment and work on unseen tasks with relatively small amounts of data. And it really is taking basically the exact concept I believe from the Dreamer paper and applying it specifically to these robot embodiments. And it turns out it actually works, actually better than I would have anticipated it working.

[B]

Yeah. So I think that this is basically the path to. It was the path, I believe it was the path to get humans to be as good as we are genetically over the last 10, 20 million years of evolution. A bigger world model helps for training and for test time planning. And I think it'll be the same thing as true as for robotics.

[A]

What's also cool is there's a bunch of applications of this to things outside of robotics too. I mean there was a weather planning paper, for example, we were reading this gencast paper which I think applies a relatively similar concept in terms of how they model literally the world, the world's weather, with something like this.

[B]

Yeah, we have to talk about the world model for the world. Yeah. So basically they do the exact same thing where the key unlocks for this whole thing was getting diffusion to work in very high dimensional state spaces like we talked about in the last lecture, and then learning to use that to action condition in the way that he's done. But they did this for the entire world with this exact same diffusion steps which go from some and they go back to two time steps, lag of order 2 AR2 for the statisticians there and then basically predict the next state of the world based on those things. With this Lengeven diffusion rollouts, my big assertion is that it was necessary for the human brain to develop world modeling. I actually just saw this paper that I wanted to make sure to call out because I thought it was so great out of University of Washington where they say explicitly in the abstract, each cortical area estimates both latent sensory states and actions. And the cortex as a whole predicts the consequences of of those actions. That sounds like a world model to me. Right.

[A]

It's actually describing exactly these two equations here, exactly where we're estimating both the sensory latent states and actions. I mean, I guess it's really the joint model that we showed earlier is what he's describing here. It's exactly this equation we're showing you.

[B]

Exactly right. And so if it works in us, it should work in robotics and I think that that takes us the rest of the distance.

[A]

Why don't we talk briefly about latent world models, especially the JEPA concept, because I think there's been a number of papers that use JEPA as an element of their architecture. Why don't we just briefly introduce JEPA and how it fits into the current landscape of world modeling.

[B]

Yeah, in classic rl you'll have like, you know, if you do study Q learning, for example, you basically keep this matrix called the Q matrix and it's going to be S by A. And so I have this S. States and actions, states and actions. And each one I need some amount of counts of being in this state action and I take the average value of taking that action in this state and that's my Q value there. And it's a little bit more complicated than that. There's Bellman equation, all this backup, all this stuff like that. So this scales horribly because as the cardinality of my state space gets bigger and action space gets bigger stuff, I don't have enough. I become less and less sample efficient. Right.

[A]

In the case of like robots or whatever, state is like. Yeah, it's this whole thing we described earlier, right. It's absolutely massive because it has all of these elements in it. You couldn't really enumerate a huge grid with it.

[B]

And so the classic trick, I mean, Since I took C229 with Andrew Ong in 2012, is you do this, stick

[A]

a neural network on it.

[B]

Exactly. And you basically are just going to compress that state into some lower dimensional state state space. This actually predates deep learning. We were doing stuff like this. I think my first paper was basically doing something like this, basically turning a grid into a bunch of pyramids. And the state was how much I'm in pyramid one, a pyramid two or whatever. But anyway, the neural networking can just do this. And so basically the key idea in jpa, if I have an image one and I have image two and I have image three, I can do my world modeling, my world modeling of ST +1 given ST and AT in pixel space and have this is let's say at time t, t1, t plus 2, et cetera, et cetera. And I have to actually predict now the full image. That's extremely expensive from a computation standpoint and also from like a sample efficiency standpoint. What I can do instead is put this through some comnet, some encoder. Some encoder and then I'll get a latent for T and I'll have a latent for T +1 and I'll have a latent for Z and then I'll have from this, from zt. I want to predict Z t plus one hat and my goal is to make this and this and my loss function will be something very simple like want to minimize this.

[A]

That's it.

[B]

Now this doesn't work, this collapses hard. And so what happens is basically just if you, if it just predicts zero, just output zeros, which the model will learn to do. And I'm actually incorporating this into my current research right now. And so what you need to do is something called sigreg or this is one technique, Vicreg is another where basically I add this another term that basically says I want the over a large enough batch size. I want the distribution of z t plus 1 to follow a Gaussian.

[A]

You know, it's kind of like a normalized, like a, like a bash norm type of, type of trick. I mean not in the same.

[B]

If, if it's zero, it can't be this. Yeah, right. Because then this is non zero. And so maybe I think that there's probably this or something like that. But basically this prevents it from modal collapse and it makes it do something good. And this is the most recent paper for the audience is LE Wm le world model which is super, super great. However, to be completely frank, this is self supervised learning. Super great. It doesn't work that well. If you were to not do these techniques and there's a bunch of other techniques that you can do, it will actually outperform much better that are, let's say for example if I'm going to do an LLM and you have Francois Sushi, which is definitely true and I tokenize this into a bunch of different tokens here and this is token ID 6 19, 28 whatever. And I look up the encoding into this and that's going to be E1, E2, E3, et cetera. What you can actually do is have the LLM output. The LLM will take in these things and will output the next token. And so it would be like let's call it H. This would be the low jits coming out of it T +1. And what you can do is actually have this be close to E T1 and a lot of people are playing with this idea and getting rid of the cross entropy loss entirely. And so if you were to do this, it actually is a proxy for the cross entropy loss and there is no cross entropy loss and the cross entropy head is actually very expensive and so this is very cheap and like this literally just grabbing it. So people are playing around with this idea and as a, basically as a cheaper proxy for the cross entropy loss. So there's lots of different ideas on basically taking this JPA idea to not just pixels, but to LLMs as well.

[A]

Yeah, interesting. Yeah.

[B]

So just to define what JPA is, it's joint embedding predictive architecture.

[A]

I think one of the things I find cool about this JPA idea is it feels like an idea we see over and over in deep learning that there's a version of this idea that's basically the stable diffusion idea. There's a version of this idea that in my company training graph convolutional neural network to design drugs we use to do latent variable generation, for example. And it's an idea that comes back over and over and then has this various tricks that it actually takes to get it to work in practice. Okay, now we have a pretty good sense for how world models work. We have a pretty good sense for what the state of the art looks like if we trust this paper. And it seems like these kind of work on robots too. This paper is only from the end of last year, this year. And it seems like they have various methods that allow you to train on relatively small amounts of data that's tractable and pre trained on diffusion models. So are we good?

[B]

We're done.

[A]

Does it all work?

[B]

Yeah. This is 2016, 26 will be the year of the robot. We're going to have Rosie the Robot in your house. Yeah. No, I don't think so.

[A]

What are one or two? Because there's lots of open problems remaining. What are a few open problems? Maybe we can emphasize here that the community can go emphasize working on.

[B]

Yeah, so I think the first one is that PINS doesn't really work. What is pins? Physics Informed Neural Networks. So PINS doesn't really work. This is physics Informed Neural networks. And so basically if like almost all of the self driving car data looks like this, the car is driving down the road. And let's just say for example, I have a house here and I want to train the model on not driving into the house. And so let's say I put it into a state right here to drive into the house, what's going to happen is because almost all the data looks like this driving down the road, this will just turn magically into like a highway and it'll just like boo, don't worry, you might crash at all.

[A]

It basically needs like a ton of data not to do that either from simulation for that to not happen.

[B]

In fact, I actually don't even know if because of the data distribution. There's no data here. There's almost all the data here. And like when you're training a neural network, it has a tendency to collapse if you don't keep the mini batch composition like very even over the, you know, over the class space or whatever you want to. Want to call it. But like you'd have to train on. You have to be very careful about your data mixing to make sure you get this right to solve this problem that no one really has. But even then, the. If you take just a simple thing like this, this is like the, the con example and I have some sine wave and I want. And I have these as my X, And I have these as my Y. So this is complete interpolation. No mess this up. But why like this? No, we can't get to like machine precision. What is. I don't know what is it? 116 or whatever it is. We can't. We. The SGD will not get to zero, effectively zero. So we'll always have some residual. And for us to be like a really good world model to simulate body interactions, like to simulate this. What's going to happen when I do this? And like let's say that I'm trying to be LeBron James. Like there's like I saw this one video of Steph Curry dribbling about a basketball on a court and he just felt that there was a dead spot in the court. And he, because he's so good and he knows exactly the physics of what's going to happen if I hit this, you know, the ball with this force, like the ball is going to come back exactly the spot. And it just didn't. And he knew it wasn't him, it was the court. And he found a dead spot in the court. Like that's how good the human brain is at world modeling, in my opinion. I think it's an SGD issue. I think it's probably an architecture issue. I think Sam Altman just kind of came and just said that he thinks that there's definitely an architecture that's going to be more performant than the transformer. I think he's right. I think the transformer doesn't do compression in the time domain at all. It just keeps around everything. So anyway, so I think that getting higher fidelity in the world model is extremely important. One, I think two, seems like test

[A]

time probably is going to be a big thing like adaptation Exactly.

[B]

Test time, planning, how quickly the human brain can, you know, in times of, in sports and things like that. When you're playing tennis, I think you're a tennis player, like how quickly we can adapt to what a player is doing and things like that. We're not going to sleep and like retraining. We're very quick to adapt to a new environment.

[A]

Like the out of distribution prediction.

[B]

Exactly. And like one little data point we can quickly adapt to that new thing and change. I think there's been a lot of papers on basically estimating the friction coefficients and so those can change over time. If you go to a human environment or not, for example, this friction might change and that's important in control. And so you need to estimate that very quickly and adapt. And these models just kind of don't have a mechanism to do it.

[A]

And then I guess there's the practical speed elements of these. Right. A lot of these are doing some sort of expensive planning step and we're doing some sort of like we're kind of hacking around it with this pre training process and synthetic data. But even so like to really get maximum performance right now you'd want to do something that's closer to like the AlphaGo style rollout and that's extremely slow.

[B]

Right. The MCTS process can't happen. The other thing that is pretty crazy about the way that the brain works is that like everything is kind of running autonomously and so you might be in the middle of saying sentence one and be like oh actually no, something else. And so what just happened there? It's like type one and type two thinking are happening at the same time in some way. And so there's definitely some really cool mix of these heterogeneous models and some are overriding others and taking control of the motor cortex and commanding the body to do a thing.

[A]

Okay. But on the flip side now we talked in the past video about the squint test and how we felt that autoregressive LLMs maybe don't pass the squint test. Why don't we reintroduce what the squint test was for a second and then maybe let's think about whether this passes the squint test despite all those limitations.

[B]

Yeah. And the squintest for me I think is like this comes from the Yann Lecun. We didn't need flapping wings to achieve flight. And to that I say you, well, we did need two wings. And if I squint and I look at a bird and I squint and I look at a plane, I'm like,

[A]

yeah, it's kind of similar, it looks right.

[B]

Similarly, if I squint, I look at the human brain and I squint and I look at all these world models. We have this vla, this action policy and that they're doing test time planning together and things like that. It's getting really close.

[A]

It's much, much closer, it seems closer than an autoregressive LLM. And that's like this concept of a world model of implicitly predicting future states and actions feels intuitively like what our brain is doing. And it seems like there's some neuroscience evidence to support that.

[B]

I mean, I'm getting to the conclusion that I think that the brain is the optimizer, not the model. And that the brain emits, has models that it invokes, but the brain is somehow also the optimizer itself. And so in that way it doesn't pass the squint because you know, something magical is happening when you're sleeping. There's no intelligent species that we're aware of that have any amount of intelligence that don't sleep. And so like octopuses, dolphins, all this other stuff, elephants, they all sleep. There's some reason for that. And that seems like a really think about like the evolutionary like recourse of sleeping. Like you get eaten when you sleep. So like for the benefit of sleeping should be so, so much better to outperform that. So I think we don't have this idea of awake sleep in our current architecture, but I can imagine I'm simulating compress from the hippocampus, some experience in the day. I'm training on more of those examples.

[A]

You're collecting a whole bunch of these experience rollouts and then you're updating your policy function over.

[B]

There's gotta be something. There's this thing called shortwave ripple where the hippocampus, when you're sleeping, emits these spike trains that are actually reversed from when they actually happen back in through both the hemispheres and for like seven times and then it stops. So there's something happening there that's very training something. And if you don't sleep, then you don't have long term memory. And so there's definitely a reason why we're training things that happened into our brain.

[A]

So where does that put us now we have all this work happening with world models. How should we think about what's coming ahead in these next few years in the research community?

[B]

Yeah, I think that we're going to see a lot more of these world models in robotic policies. I think that's going to unlock probably full self driving would be like one of those examples that they can get the real time ness of it seems like that's coming. They can probably solve it with more compute to have parallel things and you probably don't need it for most standard things. Maybe getting out of weird parking jams and things like that would take us some time. Similar to the Rosie the Robot, which we've always wanted to have a Rosie the Robot to clean up my room for me. I think that this feels like we're getting to good enough that we can pay up for data and compute to get to Rosie the Robot. It does feel like that it'll be expensive to collect the data and do the dreamer sequence of going from state to state and then getting the action conditioning to work. But like I feel like it should work.

[A]

Yeah. I mean what's pretty cool is we see a lot of companies at YC working at every step of this from the collecting egocentric data, collecting the teleop data, training their own world models and action models, building new embodiments and then making ways of adapting those embodiments and feels like this is the first year where you see demos where you're like, okay, this actually kind of is starting to look like it's going somewhere and it seems like a very exciting year ahead.

[B]

Yeah. So anyway, I think that there are real AI problems to solve still. We talked about pins, we talked about the real time issues and then on the robotic side there's real issue. Like it's amazing how effective our epidermis is in terms of. We can detect tactile.

[A]

Oh, epidermis.

[B]

Yeah, epidermis are tactile. We can detect shear force, we can detect temperature and it's everywhere.

[A]

Yeah.

[B]

And so like versus, you know, like the. We get like one little sensor that only does tactile. We don't have the friction component, we don't have temperature, we don't have all these. The feeling. We can't estimate coefficient of friction very quickly. I can touch something and say, oh, this is smooth, this is rough. We don't have any of that. And if I numb your hands. I actually had this experience just recently. If I numb your hands, like you actually can't tie your shoes.

[A]

Yeah.

[B]

So you can't perform control. And so like. Yeah, if you like, you know, if you train enough on enough human data, tying your laces do I think you can do it with no feedback Maybe. Maybe. But, like, how much would you need if you did actually have the human touch? I think it'd be so much easier.

[A]

Well, there's a lot of more research to do. Yeah.

[B]

Yeah.

[A]

Francois, thanks so much for joining us. Thanks so much for watching, everyone. We'll be back for the next episode of Decoded.