Decoded

Ankit & Francois on Why AGI Needs a World Model, Not More Data

Ankit & Francois· Hosts of Y Combinator's Decoded at Y Combinator
·~74 min·English·Y Combinator
ReasoningTrainingRoboticsAI Infrastructure
TL;DR

World models — learning to predict the next state of the world — may be the route to sample-efficient AI, especially in robotics where brute-force reinforcement learning is intractable.

01The Bottleneck

Sample Efficiency Is the Real Wall

The frontier problem is not raw knowledge but how few examples it takes to learn something new — humans need a handful, models need tens of thousands.

One of the biggest open problems in AI right now is how to solve sample efficiency.

— Ankit & Francois, Y Combinator
Key Insight
The hosts lean on François Chollet's definition of intelligence as a rate of skill acquisition, which quietly changes the goal: scale maximizes total knowledge, but sample efficiency measures how fast you convert one new example into skill. A model that has read most of the internet yet fails an ARC-AGI puzzle a child solves is losing on the second metric, not the first.
02The Aha

A Perfect World Model Needs Zero Samples

A perfect world model is the zero-sample limit of learning — the more exact your model of what happens next, the fewer real-world trials you need, all the way down to none.

Can you imagine if like we needed to collect 1 million training examples of like us shooting spaceships to the moon to like know how to do it

— Ankit & Francois, Y Combinator
Key Insight
Rocketry makes the limit concrete: it is confidence in the model, not a mountain of practice launches, that carries a probe to a target years away. That reframes robotics' appetite for data as a measure of model error — the gap between the world we can already predict and the world we still have to rehearse.
03The Definition

A World Model Is Just the Transition Function

Strip away the mystique and a world model is just the map from a state and an action to the next state — and once you have that map, steering the system becomes a control problem you compute rather than learn.

So we can solve this in closed form basically like you know we can 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 thing under a set of constraints like max thrust available gravity

— Ankit & Francois, Y Combinator
Key Insight
"World model" is a grand name for the humble RL transition function — the map from (state, action) to the next state. That reframing is the whole game: once you actually have an accurate map, steering the system stops being trial-and-error learning and becomes an optimization you can solve directly.
04Why RL Is Brutal

The Dividing Line Is Whether You Can Take a Gradient

Add one other agent whose next move you can't differentiate through and closed-form control collapses into the messy, expensive world of reinforcement learning.

I have to resort to this awful area called reinforcement learning

— Ankit & Francois, Y Combinator
Key Insight
The boundary between tractable and intractable is not problem size — it is whether you can take a gradient. One drone under known physics is differentiable and solves in closed form; the moment another agent enters and you can no longer backprop through its choices, the transition turns stochastic and you are forced into sampling-based RL, where cost scales with how much of the future you must simulate.
05The Scaling Wall

AlphaGo Doesn't Scale to the Real World

Tree search is affordable only while the action space stays tiny — it goes from 24,000 model calls per Go move to a hopeless number once the world is as wide as a robot arm.

So that be 60 million uh invocations of the model. So that better be a small model, right?

— Ankit & Francois, Y Combinator
Key Insight
AlphaGo was never general intelligence; it was a bet that the action space stays small. Its budget of tens of thousands of model calls per move is affordable at 361 actions and ruinous at 10^16, and search-based methods don't degrade gracefully — they fall off a cliff the moment the world gets wide, which is exactly what self-driving and robotics do.
06The LLM Bridge

Behavior Cloning Is Next-Token Prediction, With Actions

Take the tokenizer head off a language model, bolt on an action space, and copying a human driver becomes the same next-token game that made LLMs work — if you can find the actions.

in many ways, it's the closest thing that just looks like the next token prediction from LLM that seems to scale pretty well with natural language

— Ankit & Francois, Y Combinator
Key Insight
The bridge to language models is exact: model-free behavior cloning is next-token prediction over actions instead of words. That is also why the data bottleneck bites so hard — the internet is full of what happened (video) but almost empty of what action produced it, and only a fleet owner like Tesla holds both halves at once.
07The Unlock

Learn the World From Video, Then Dream the Training Data

The breakthrough is to decouple the two things we have in unequal supply: learn the world from abundant video, spend your scarce action labels only on conditioning, then train the policy on synthetic rollouts.

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 data, which is kind of crazy

— Ankit & Francois, Y Combinator
Key Insight
The real unlock is not a bigger model but splitting the problem along the axis where our data is lopsided: video is abundant, action labels are scarce, so learn the world from video and spend the few labels only on the conditioning step. Dreamer solving Minecraft on purely synthetic rollouts is the proof that a policy can be trained without ever touching the real environment — the same move now aimed at self-driving (GAIA) and robots (DreamZero).
08The Frontier

JEPA Predicts in Latent Space — If It Doesn't Collapse First

The current frontier predicts a compressed latent of the next state instead of every pixel — far cheaper, but it invites a sharp failure where the model cheats by predicting nothing.

now this doesn't work this collapses hard and so what happens is basically just if you if you just predict zero

— Ankit & Francois, Y Combinator
Key Insight
JEPA's move — predict in latent space rather than pixels — is the same compression trick that made stable diffusion practical, now applied to control. But it exposes a brutal failure mode: a model rewarded for matching its own predictions can win by predicting nothing and collapsing to zeros, so most of the frontier work (VICReg, LeJEPA) is really about the anti-collapse regularizer, not the prediction itself.