Last month, Yann LeCun's new startup AMI Labs raised $1.03 billion to build world models. Not a better chatbot. Not a faster LLM. Something fundamentally different. LeCun, a Turing Award winner who spent twelve years leading AI research at Meta, left because he believes large language models are a dead end. His bet is on world models.
That's a strong claim from someone who knows the field as well as anyone alive. But what does it actually mean? What is a world model, and how is it different from the LLMs most people interact with every day?
What an LLM Actually Does
A large language model predicts the next token. Not word — token. A token is a chunk of text: sometimes a full word, sometimes part of one, sometimes just punctuation. The model looks at everything that came before and asks: what token is most likely to come next? It learns this by training on enormous amounts of text — billions of documents, books, conversations, code — absorbing the statistical patterns of how language works.
When you ask ChatGPT to explain quantum physics, it's not reasoning about quantum physics. It's generating text that looks like a good explanation, based on all the explanations it's seen before. The output is often excellent, but the mechanism is pattern matching over language, not understanding of the subject. Think of it like a brilliant parrot that's read every book in every library. It can answer questions, write code, translate languages — but it's operating entirely in the world of text. Tokens in, tokens out.
That said, modern LLMs aren't stuck in pure pattern-matching mode. Chain-of-thought reasoning lets a model break a problem into steps and work through them sequentially, rather than just outputting whatever token feels statistically right. Tool use takes it further: instead of guessing at a math problem, the model can write and execute code; instead of hallucinating a fact, it can search the web. These techniques don't change what the model fundamentally is — a token predictor — but they dramatically extend what it can do. The strawberry problem is a good example: the model didn't get smarter at counting letters, it learned to delegate to a tool that could.
What a World Model Actually Does
A world model predicts what happens next in a physical environment. Not the next token. The next state of the world.
Drop a ball off a table. A world model predicts the trajectory, the bounce, the roll. Not because it read about gravity in a textbook, but because it's learned the rules of how objects behave by observing thousands of examples of objects behaving.
Self-driving cars are probably the most intuitive example. The car needs to predict what every other vehicle, pedestrian, and cyclist is going to do in the next few seconds. That's a world model. It takes the current state of the environment, a 3D map of everything around the car, and simulates forward in time. If the car in the next lane is drifting left, the world model predicts a lane change and adjusts before it happens.
Video game AI is another one. When an NPC in a game navigates around obstacles, anticipates your movements, or reacts to changing terrain, it's running a simplified world model. It understands the physics and spatial rules of its environment and acts accordingly.
The key difference: world models operate on physical states — positions, velocities, forces, spatial relationships — not sequences of text.
The GPS vs. the Travel Writer
Here's an analogy that makes the distinction click.
A world model is like a GPS. It has a map of the actual terrain. It knows where roads are, how long they take, where traffic is backed up. It predicts your arrival time based on the physical reality of the route. If a road is closed, it reroutes based on the real road network.
An LLM is like a travel writer who's read every travel guide ever published. Ask them how to get from Toronto to Montreal and they'll give you a beautifully written answer with plausible directions. Most of the time, it's right. But they've never actually driven the route. They're working from descriptions of the route, not the route itself. If there's a new highway that didn't exist in any travel guide, they won't know about it. If you ask about a road that's described differently in different guides, they might blend the descriptions into something that sounds right but isn't.
The GPS understands the territory. The travel writer understands descriptions of the territory. Both are useful. They're not interchangeable.
Where LLMs Fake It
LLMs are so good at language that they create the illusion of understanding the physical world. Ask one what happens when you push a glass off a table and you'll get a perfect answer: it falls, it shatters, glass goes everywhere. But the model has just seen thousands of descriptions of glasses breaking, and it's generating the statistically most likely one. Where it breaks down is uncommon scenarios — the ones where the training data leads you to the wrong answer.
The carwash problem is a good example. "Should you drive or walk to the carwash?" The model defaults to "walk" because that's the most common answer for a short distance. It has no physical model of what a carwash does, so it misses the one constraint that makes this question different: the car needs to be there.
Where World Models Fall Short
World models are powerful but narrow. A self-driving car's world model is excellent at predicting traffic dynamics. Ask it to write a poem and it has nothing to offer. It operates on 3D point clouds and velocity vectors, not language.
Building a world model is also expensive and domain-specific. You can't train one general-purpose world model that handles everything from fluid dynamics to social interactions to orbital mechanics. Each domain needs its own model, its own data, its own physics. That's the core tradeoff: deep accuracy within a domain versus broad approximation across all of them.
Where Things Get Interesting
The frontier of AI research right now is trying to combine both. What if you had a system that could reason about language like an LLM and simulate physical reality like a world model?
This is exactly what LeCun is betting on with AMI Labs. His approach, called JEPA (Joint Embedding Predictive Architecture), learns by predicting representations of the world rather than the next token in a sequence. Think of it as learning to understand situations rather than learning to describe them. LeCun has argued for years that this is a more promising path to real machine intelligence than the autoregressive text prediction that powers ChatGPT, Claude, and Gemini.
Robotics is where the need is most obvious. A robot that can understand the instruction "pick up the red cup without knocking over the blue one" needs both: language understanding to parse the instruction, and a world model to plan the physical movements. Neither system alone is enough.
Video generation is another interesting case. Models like Sora produce videos where objects move with realistic momentum, light behaves correctly, and gravity works. Some researchers argue these models are learning implicit world models as a byproduct of learning to generate realistic video. Whether that counts as "understanding" physics or just mimicking it convincingly is an open question.
Then there's simulation. Companies like Wayve and NVIDIA are building world models that generate entire synthetic driving scenarios. Instead of logging millions of miles of real driving, you simulate them. The world model creates the scenarios, and the driving AI learns from them. World models training other AI systems.
Why the Distinction Matters
If you treat an LLM like a world model, you'll over-trust its answers about the physical world because they sound so confident. If you treat a world model like an LLM, you'll miss that it can't explain its reasoning or generalize beyond its training domain.
The practical takeaway: use the right tool for the right job. Need to generate text, answer questions, or write code? That's an LLM. Need to predict physical outcomes or plan actions in 3D space? That's a world model. Need both? That's the hard problem everyone's working on — and LeCun's bet is that the systems that crack it will be the ones that actually feel intelligent, not just sound intelligent.