Yann LeCun, AMI Labs | Evident AI Symposium 2026

I can see part of it. Yes, part of it.

Sure. So I'm actually calling you from AMI Labs. It's pronounced 'Ami' — you know, 'ami' as in friend in French. So AMI is actually a French company headquartered in Paris. We have two big offices in Paris and New York, and smaller offices in Montreal and Singapore. So we're really a global company. We started operations in March. We raised about a billion dollars in the seed round, which is the biggest seed round in history outside the US, I believe, as far as I know. And we have around 65 employees at the moment, and the whole company is based around the idea of building AI systems for the real world, for the physical world. You know, people have been calling this physical AI, if you want, and it's built around the idea of training world models. And maybe we can go into what we mean by world model, which is not necessarily the same thing as what other people mean by that. And things are progressing really well, and we're anticipating some success in various industries over the next year or so.

Right. So there is a mistake that people in AI and the periphery of AI have kept making for the last 70 years, and it's the mistake that if a machine can do a complex task, or a task identified by humans as intellectually complicated, that automatically this machine is smarter than humans, or perhaps has reached human intelligence, or perhaps superintelligence. And we know this is completely false, because we've had systems that are capable of solving equations, or computing integrals symbolically, or playing chess or Go better than humans. So we have computers that are superhuman in many different tasks, and the list keeps increasing thanks to recent progress in AI. But this does not mean that we are anywhere close to having machines that have the type of intelligence we observe in humans and animals, for that matter. And in fact, as it turns out, it's much simpler for machines in general, and for current AI approaches based on LLMs, to manipulate sequences of discrete symbols — human language, code, mathematics, and things of that type — rather than handling the real world. The real world is much more complicated. It's messy. It's not entirely predictable. It's continuous, high-dimensional, noisy. And it turns out the methods that have been so successful in the context of language, namely deep nets, actually do not work very well for the real world. And so you can see this in the fact that we have systems that can pass the bar exam and prove theorems and do all kinds of amazing feats, but we still don't have domestic robots. We don't even have level five self-driving cars. We certainly don't have self-driving cars that can learn to drive in a few hours of practice like any teenager. So there's a huge gap, and that gap has been known for a very long time. It's called the Moravec paradox — the fact that we can handle things like logical reasoning and mathematics and stuff like that with computers relatively easily, but not the real world. So we're trying to kind of bridge that gap with AMI Labs, coming up with new paradigms of AI, really, that would be capable of handling the messiness of the real world, if you will.