23 February 2026 · Matthieu MALVACHE · 8 min
A House Cat Understands Physics Better Than Your LLM
In January 2026, Yann LeCun walked away from his role as Chief AI Scientist at Meta to co-found AMI Labs. His pitch was blunt: large language models are a dead end for general intelligence. The future belongs to world models.
LeCun co-invented convolutional neural networks and won a Turing Award for it. So when he bets his next act on an idea, I pay attention.
What LLMs actually do
Every LLM works the same way. You feed it a sequence of tokens - words, subwords, punctuation - and it predicts the next one, then the next, until it produces a full response. GPT-4, Claude, Llama, Gemini: all next-token predictors trained on text.
This works remarkably well. LLMs write code, draft contracts, summarize papers, hold conversations that feel intelligent. But LeCun argues this success masks something important.
An LLM has never watched ice melt or felt what happens when you drop something heavy. Everything it knows about the physical world comes from written descriptions - secondhand knowledge filtered through language. It gives plausible answers about physical events, but it's reciting, not reasoning. It found patterns in training data. It didn't simulate anything.
LeCun's metaphor is hard to shake: a house cat understands more about the physical world than any LLM. The cat has never read a physics textbook, but it predicts trajectories, gauges distances, plans jumps. It has a world model. The LLM has a word model.
What world models are supposed to do
A world model learns abstract representations of the physical world from sensory data - mostly video and images - and uses them to predict what happens next.
The difference that matters isn't the input modality. It's what gets predicted.
An LLM predicts: given this sequence of tokens, what token comes next?
A world model predicts: given the current state of the world and a potential action, what does the next state look like?
That second question is what makes planning possible. If you can run a mental simulation before acting, you can weigh options and reason about cause and effect. That's how biological brains work. You don't need to actually throw a ball to estimate where it'll land.
LeCun's specific proposal is called JEPA - Joint Embedding Predictive Architecture. Unlike generative models that reconstruct raw pixels (expensive and full of irrelevant detail), JEPA predicts in a compressed latent space. It learns the abstract structure of what's happening rather than reproducing every surface-level detail. Think physics of a scene, not pixel-by-pixel video generation.
Why LeCun thinks LLMs are a dead end
He's not saying LLMs are useless. His argument is more precise: LLMs will never reach human-level intelligence, no matter how much you scale them.
He points to a few gaps that scaling won't close.
LLMs learn correlations in text. "Glass" appears near "break" and "floor" in training data, so the model produces plausible sentences about dropping glasses. But it never learned that glass breaks because of impact force, material brittleness, and surface hardness. Correlation isn't causation. No amount of text bridges that.
Then there's grounding. Language is a lossy compression of reality. When I write "the cat jumped from the shelf to the table," you reconstruct the scene using your own physical experience of cats, shelves, tables, gravity. An LLM has none of that. It manipulates symbols that refer to experiences it has never had.
And LLMs can't really plan. They generate tokens one after another. They don't evaluate multiple possible futures and pick the best one. Chain-of-thought prompting and extended thinking modes help - they're System 2 patches bolted onto a System 1 architecture. Useful workarounds, but the model still can't test what happens if it takes action A versus action B in a physical environment.
LeCun frames it using Daniel Kahneman's distinction. LLMs are reactive System 1 thinkers - fast pattern matching, no deliberation. World models would give AI a System 2: the ability to reason slowly about consequences before acting.
Where the debate stands
Not everyone agrees, and the counterarguments are real.
The scaling camp (OpenAI and others) argues that LLMs with enough parameters, data, and compute will develop emergent reasoning. GPT-4 already shows flashes of something that looks like causal understanding, even if it's built on statistics. Maybe "understanding" is just really sophisticated pattern matching all the way down. I'm not convinced, but I can't rule it out either.
Gary Marcus has argued that LeCun overstates the case against LLMs while understating how hard it is to build world models. JEPA is promising on paper. Nobody has shown one working at scale in a complex environment.
Fei-Fei Li's research at Stanford goes in LeCun's direction. Her group published a framework in 2025 arguing that understanding the 3D world is essential for intelligent agents.
And then there's the boring-but-probably-right answer: the future uses both. LLMs handle language. World models handle physical reasoning. Agents pick whichever fits the task. Language understanding doesn't become less useful just because you add physics on top.
What this means if you're building with AI
If you use AI in your business today, none of this changes your immediate plans. LLMs are the right tool for text-heavy work: documents, code, support, content.
But if you're thinking about where agents go next, world models matter. The agents I build today live in digital environments: databases, APIs, documents. The physical world is out of scope. An agent that interacts with reality (warehouse robotics, autonomous vehicles, manufacturing) needs something closer to a world model. It needs to anticipate physical consequences, not autocomplete text.
LeCun is betting this transition arrives sooner than most people expect. I don't know if he's right about the timeline. But the direction feels sound: intelligence limited to language is a subset of intelligence that also handles reality.
So where does that leave us?
LeCun quitting Meta to build world models full-time tells you how seriously at least one Turing Award winner takes the idea. Whether JEPA delivers or not, the question it raises won't go away: can you really get to general intelligence through text alone?