Close Your Eyes

Close your eyes. Walk to your kitchen. Open the fridge. Reach for whatever's on the second shelf.

You just did something that no large language model on earth can do. You ran a simulation. Not a description of your kitchen — an actual spatial, causal, physics-aware model of a room you're not currently in. You predicted what would happen if you opened a door, where your hand would need to be, which shelf the milk is on. Some of those predictions were probably wrong. The structure was right.

That is a world model.

I want to explain what world models are, why they matter, and why a lot of very serious people are betting billions of dollars that they — not bigger chatbots — are the next revolution in AI. This is going to get technical in places. I'll be gentle about it. But I'm not going to dumb it down. Easy is empty, right? 🫶

What AI currently does

I need to be fair to the current generation before I explain what's coming next.

Large language models start with next-token prediction — given everything that came before, what word comes next? But that's only the foundation. After that initial training, the models go through reinforcement learning from human feedback (RLHF), where people rank outputs and the model learns to prefer responses that humans judge as helpful, accurate, and safe. Then there's further fine-tuning: chain-of-thought reasoning, tool use, code execution, instruction following. The finished product is not just a raw autocomplete or next-token predictor.

I built this app on top of these systems. I know what they can do. They are not toys.

But here's what they still don't do: they don't simulate.

All of that post-training — the RLHF, the reasoning chains, the tool use — makes the model a better language reasoner. It learns to produce more thoughtful, more structured, more accurate text. What it does not learn is a persistent, interactive model of how the physical world behaves. It doesn't maintain state across actions. It doesn't predict what happens when you push something. It reasons about physics in words. It doesn't run physics.

If you ask a frontier LLM to describe a ball rolling off a table, it will produce a convincing, well-reasoned paragraph about gravity and trajectories — probably citing Newton. It can even show its work. If you ask it to predict where the ball will be in 1.7 seconds given a specific velocity and angle off an irregular surface with spin, it starts to struggle. Not because it's stupid. Because it's reasoning in language about a problem that wants a simulator.

A world model is the simulator.

The actual definition

A world model is an internal simulator that an AI system carries around inside itself. Given the current state of things and an action, it predicts the next state. Not the next word — the next state of reality.

Five properties distinguish a real world model from even the most capable language reasoner:

1. Causality — it captures cause and effect, not just correlation
2. Interactivity — it responds to your actions in real time
3. Persistence — it remembers what's behind you. Object permanence.
4. Compositionality — it can combine learned concepts in configurations it's never seen before
5. Real-time responsiveness — it runs fast enough to actually be useful for planning

Your brain does all five of these constantly. Current LLMs do approximately zero of them well. That gap is the whole story.

The kitchen test

This is the easiest way I know to understand the difference between a language model and a world model:

An LLM can describe your kitchen. A world model can simulate it.

A description is a sequence of words. A simulation is a structure that changes when you poke it. In your kitchen model, if you imagine moving the table, the path to the fridge changes. If you imagine turning off the lights, you can still navigate because the spatial relationships are preserved. The model isn't a photograph. It's architecture.

This concept goes back to 1943. A Scottish psychologist named Kenneth Craik proposed that the mind constructs "small-scale models" of reality. Not copies — models. They preserve the relationships within the world without reproducing every detail. Your kitchen model doesn't include the number of floor tiles. It has the structure: fridge is left of the stove, counter is waist height, floor is cold. The map is not the territory. But the map preserves the topology.

Craik died at 31 in a cycling accident. The field took decades to catch up with him.

What happened in 2018

Two researchers — David Ha and Jurgen Schmidhuber — published a paper called "World Models" that brought this idea into modern AI. Their system had three components:

1. A Vision model — compresses what you see into a tiny numerical summary
2. A Memory model — the simulator. Predicts what happens next.
3. A Controller — the tiny decision-maker that picks actions

The architecture is simple. What they did with it was not.

They trained an agent to play a car racing game. The agent would interact with the game for a while — the "awake" phase — collecting experiences. Then it would stop interacting with the real game and switch to the Memory model, which would generate fake experiences based on what it had learned about the game's dynamics.

The agent would practice inside a synthetic version of the game generated entirely by the Memory model.

Then it would go back to the real game. And the skills transferred.

The researchers call this the "dream" phase. I know how that sounds. I'm the last person who wants to anthropomorphise a training loop. But the term is theirs, not mine, and the technical description is uncomfortably precise: the system interacts with reality, builds a model, then disconnects from reality and trains inside the model instead. "Awake" phase: real environment, collect data. "Dream" phase: simulated environment, refine skills. I don't love the word. I can't argue with the architecture.

Why compressing matters

Here's a technical detail worth understanding because it's elegant and it clarifies the whole field.

Your kitchen model doesn't simulate every atom. It compresses. It keeps the stuff that matters — spatial relationships, gravity, solidity — and throws away the stuff that doesn't — the specific pattern of light on the countertop, the exact position of each spoon in the drawer.

Modern world models do the same thing. Instead of predicting the next raw image (millions of pixels, computationally brutal, mostly irrelevant detail), they work in what's called latent space — a compressed representation that captures essential structure.

Think of it as the difference between simulating your kitchen at the resolution of individual atoms versus simulating it at the resolution of objects. Both are models. One is useful. The other melts your GPU.

Yann LeCun — until recently Meta's chief AI scientist — pushed this idea furthest with something called JEPA (Joint Embedding Predictive Architecture). The key insight: don't predict the future in pixel space. Predict it in meaning space. Don't forecast what the next video frame will look like. Forecast what it will be about. What relationships will hold. What will have changed.

This is, again, what your brain does. You don't simulate your kitchen at the resolution of individual photons. You simulate it at the resolution of objects, surfaces, and relationships. JEPA does the same thing computationally. Strip away the noise. Predict the signal.

LeCun's team built a system called V-JEPA 2 and put it on physical robots. The robots could handle objects they had never seen during training. They'd learned the structure of physical interaction — not just specific examples of it.

The current moment

I should tell you what is happening right now because it is happening fast.

In early 2026, LeCun left Meta after twelve years and launched AMI Labs with a $1.03 billion seed round — the largest European seed ever. His thesis: world models, not language models, are the path to genuine machine intelligence.

Fei-Fei Li — the Stanford professor who created the dataset that modern computer vision is built on — raised $1 billion for World Labs. Her company generates explorable 3D worlds from images and text. Not videos — environments you can walk around in.

Google DeepMind released Genie 3: real-time interactive world generation at 24 frames per second, 720p, minutes of consistent navigable space. Not a pre-rendered video. A world that reacts to what you do in it.

The phrase you're hearing in the research community is "Large World Models." The deliberate replacement for "Large Language Models." The bet is that the era of making chatbots more eloquent is ending. The era of making machines that understand physics, space, time, and causation is beginning.

The debate you should know about

Here's where it gets philosophically interesting, and I want to be honest with you about the uncertainty because that's my whole thing.

A group of researchers trained a transformer — the same type of model behind ChatGPT — on moves from Othello, a strategy board game where two players place black and white discs on an 8x8 grid, flipping each other's pieces. Simple game. Clean rules. The researchers gave the model just the moves. No board. No rules. Just raw sequences of game positions.

Then they probed the model's internal activations and found something: it had built a representation of the Othello board. Not because anyone told it to. Because predicting the next legal move apparently requires some kind of spatial understanding, and the model had developed one on its own.

This is the best evidence that language models might develop emergent world models from pure next-token prediction. If a model can learn to represent an Othello board just by predicting moves, what might GPT-4 have learned to represent by predicting language?

But Melanie Mitchell — a complexity scientist at the Santa Fe Institute — pushed back. Hard. Her argument: what the model developed might not be a genuine world model. It might be a complex patchwork of heuristics that produces correct outputs for the wrong reasons.

She uses a historical analogy. Ptolemy's model of the solar system, with its epicycles upon epicycles, could predict planetary positions accurately. But it didn't understand orbits. It was an elaborate curve-fitting exercise. Copernicus's model was simpler and captured the actual mechanism.