There is a particular pleasure in dialoguing with someone who has spent more than a decade building a potential game-changer and is now on the cusp of success. That was the backdrop of my chat with Yann LeCun onstage at INSEAD’s recent AI Forum Europe, shortly after he left Meta to found Advanced Machine Intelligence (AMI Labs), a new venture headquartered in Paris.

LeCun needs no introduction to the audience overflowing the auditorium at Station F, the world’s largest startup incubator, also located in the French capital. A Turing Award laureate and long-time chief AI scientist at Meta, he has been a prominent sceptic of the idea that scaling up large language models will eventually produce human-level intelligence. AMI Labs, which has raised US$1 billion, is his attempt to prove that the path forward in the AI age runs through “world models” rather than ever-larger LLMs. And the implication, if he succeeds, will ripple far beyond the realm of technology. 

The Moravec paradox, updated

What underpins LeCun’s ambition is a decades-old conundrum. In 1988, roboticist Hans Moravec asked why computers could already play championship-level chess while being unable to perform the physical tasks a child easily performs. Nearly 40 years on, LLMs work alongside humans in business, healthcare, legal and other realms, yet nobody has built a robot that can learn to drive as reliably as humans, or one that fills a glass of water. 

This is because current AI systems are trained on language and mathematics, domains where the symbols themselves carry the reasoning, explained LeCun. The physical world, with all its noise and messiness, is not as predictable. 

“LLMs are essentially rote learners. They have some reasoning abilities, but…they only apply to things like maths and code, where language is connected with reasoning,” LeCun told the crowd gathered at Station F at the launch of the two-day AI Forum Europe in June, which INSEAD organised in collaboration The Wharton School. Other featured speakers include Philippe Aghion, the INSEAD Professor in Economics and Nobel laureate, Wharton’s Dean Erika James, and Nicolai Tangen, CEO of Norges Bank Investment Management. 

What a world model does instead, LeCun explained, is learn to understand how the real world works. "World model is what allows you to anticipate what's gonna happen in the world, particularly to anticipate what's gonna happen as a consequence of your own actions." 

Such a model can, in principle, be far smaller and far cheaper to run. “That's where AMI Labs is positioning itself. We think there’re literally millions of applications of this type of methodology to industry in lots of different areas,” said LeCun.

Sovereignty through openness

AMI Labs’ billion dollars has come roughly 40% from Europe, a third from the United States, and the rest largely from Asia. That geographic spread reflects LeCun's conviction that AI shouldn't be dominated by any single region's models.

Instead, he envisions a completely free and open foundation model that can produce AI assistants that speak every language in the world.

“You can fine-tune the system to speak any language, understand every value system, every political bias, every centre of interest… which is incredibly important because we need that diversity for the same reason we need diversity of the press. Without that, there's no democracy, there's no culture.”

To that end, LeCun recently joined Project Tapestry, an effort launched by the non-profit AI Alliance to federate compute and data contributions from governments, universities and companies worldwide into an open foundation model, without anyone having to hand over their underlying data. 

Jobs, control and the bubble question

LeCun is far more sanguine on AI’s impact on labour markets and jobs. “Certainly, our relationship to AI systems will transform professions. But only some tasks can be automated,” he said. 

“I think our relationship with AI systems is that we'll just be their boss. All of us will be kind of walking around with a staff of AI assistants that will kind of execute stuff for us, but they'll do our bidding.” 

He blames the widespread fear of losing control to AI on three reasons. The first is simply science fiction: The “robot takeover” trope has conditioned how we picture artificial intelligence. The second is projection: We assume that any sufficiently intelligent entity must share human drives, including the desire to dominate or the desire for prestige. That’s simply not true, said LeCun. Even within humanity, some of the sharpest minds there is – like academics – want nothing to do with society, much less to dominate others. They would rather be left alone with their work. 

The third reason for the fear of AI domination, and the one he considers most consequential, is that people believe AI to be smart and uncontrollable. LeCun said it’s true that LLMs are not controllable: ask one to do something and it may do something else entirely, because it has no goal of its own. It isn't predicting the consequences of its output at all. But LLMs aren’t especially dangerous, he said, since they aren't smart enough yet to do much damage. 

The architecture he is building, which he calls objective-driven AI, works such that AI systems can only act to fulfil a given goal as predicted by its world model. The model can still be wrong, which is why guardrails matter, just as human need to be governed by laws. An AI system built this way could be made structurally incapable of breaking the equivalent rule, rather than deterred from it – unlike humans who can and do break laws despite the threat of penalties. 

By the sound of it, the next big thing in AI will be an improvement over LLMs. And that’s something humankind can all celebrate. 

The next edition of the INSEAD AI Forum series will be held at INSEAD’s San Francisco Hub for Business Innovation on 18-19 September. Get your tickets here.