Yann LeCun recently raised €500 million to prove the entire AI industry is going in the wrong direction. His new startup AMI Labs (french startup) is built on one thesis: large language models will never produce general intelligence. You need “world models,” systems that actually understand the physical world, not just predict text about it.
A few weeks ago, researchers at UCSD published a comment (and this point is important) in Nature pushing back on exactly this. Their argument: ask an LLM what happens when you drop a glass on tile, it predicts correctly. Ask it physics olympiad problems, it solves them. If that’s not a world model, what is?
Honestly, both sides have a point and the boundary is blurrier than either camp wants to admit. Is an LLM predicting that glass breaks on tile “understanding”? Or is it pattern matching on millions of sentences written by people who understand? I don’t think anyone can cleanly answer that right now, and anyone who claims they can is maybe oversimplifying.
Where I do think LeCun has identified something real is the frozen knowledge problem. An LLM doesn’t learn from what it sees after training. It’s stuck in time. For anything that requires interacting with physical reality, updating your understanding in real time, adapting to what’s happening right now, that’s a genuine limitation that no amount of scaling has solved yet.
Whether world models are the answer to that, I genuinely don’t know. Maybe LLMs integrate that capability as they scale. Maybe you need a dedicated architecture like LeCun’s JEPA, maybe it’s both working together. It’s too early to be confident about any of those outcomes.
What I am more skeptical about is the €500 million pré-produit. LeCun is a Turing Award winner and his track record is real, but track record doesn’t guarantee execution. AMI Labs has no product, no revenue, and a thesis that hasn’t been validated outside of research papers. That said, I don’t think this is a bubble symptom specifically. The stakes in AI right now are high enough that betting €500 million on a contrarian thesis from one of the most credible people in the field isn’t irrational. But it’s just very expensive for a seed / preseed stage lol.
If AMI delivers, the implications for robotics, autonomous systems, and anything involving physical reasoning could be massive but that “if” is doing a lot of heavy lifting. The research is promising, the vision makes sense, and the rest depends entirely on whether they can turn it into something that works not just inside a lab but in the real world.