Yann LeCun could have coasted. He has a Turing Award, the field’s highest honour, shared for inventing much of modern deep learning. He spent years as Meta’s chief AI scientist. He is, by any measure, one of the handful of people who built the technology everyone else is now getting rich on. And he has spent a good chunk of the last few years telling anyone who will listen that the large language models powering that gold rush are a dead end on the road to real intelligence. In 2026 he put his reputation where his mouth is: he left Meta and launched AMI Labs to build the thing he thinks will actually work.
This is the most intellectually confrontational entry in our AI Innovation series, because AMI Labs is not just another world-model startup. It is a bet, by a founder with everything to lose, that the entire industry is climbing the wrong mountain.
The founder: the sceptic who helped build the thing he doubts
LeCun’s credibility is the whole story, so it is worth being precise about it. In the 1980s and 90s he pioneered convolutional neural networks, the architecture that made computer vision work and that still underpins huge swathes of AI. In 2018 he shared the Turing Award with Geoffrey Hinton and Yoshua Bengio for the deep-learning breakthroughs that started the current era. Then, as chief AI scientist at Meta, he became the most prominent insider willing to say the unfashionable thing: that scaling up language models, however impressive the results, will not by itself produce machines that understand and reason about the world.
At AMI Labs he is Executive Chairman rather than day-to-day CEO. That job goes to Alex LeBrun, who previously built and ran Nabla, a medical AI company, and who brings the operator discipline a research legend does not necessarily supply. The name itself is the manifesto: AMI stands for Advanced Machine Intelligence, the research programme LeCun had been pursuing with colleagues at Meta and New York University, now spun out to be chased without a giant corporation’s priorities attached.
What they are actually building
AMI Labs builds world models: AI that learns how the physical world works by training on raw sensory data, video, images, sensor streams, rather than on text scraped from the internet. The pitch is that a system fed enough real-world observation can learn cause and effect, spatial logic, and the basic physics of how things move and interact, and can then reason and plan the way a human or an animal does, rather than pattern-matching the next word.
LeCun’s long-standing technical argument is that a language model has no grounding in reality. It has read that fire is hot, but it has never been near a flame; it manipulates symbols about the world without a model of the world underneath. His alternative, developed over years of papers, is architectures that predict what happens next in an abstract representation of a scene rather than in pixels or words. AMI Labs is the commercial vehicle for finally building that at scale.
| AMI Labs at a glance | Detail |
|---|---|
| Founders | Yann LeCun (Executive Chairman), Alex LeBrun (CEO, ex-Nabla) |
| Launched | March 2026 |
| Seed funding | $1.03bn |
| Key backers | Nvidia, Samsung, Sea, Temasek, Toyota Ventures, French investors |
| Approach | World models from sensory data; open-source code and published papers |
| The thesis | LLMs cannot reach human-level reasoning; world models can |
Why this is genuinely exciting
Two things make AMI Labs stand out even in a crowded field. First, a billion-dollar seed round for a company with no product is rare even now, and the backer list, Nvidia, Samsung, Temasek, Toyota Ventures, is a mix of chipmaker, hardware giants and industrial money, exactly the partners you would want if your world models are meant to end up in robots, cars and factories rather than in a chat window.
Second, and refreshingly, AMI Labs plans to publish its papers and open-source its code as it goes. In an industry that has grown steadily more secretive, a frontier lab committing to work in the open is both a philosophical statement and a practical recruiting magnet. If the science is real, everyone gets to build on it, which is how LeCun has always believed the field should work.
The friendly sceptic’s corner
Now, warmly, the other side. LeCun has been predicting the limits of language models for years, and for years the language models have kept getting more capable and more useful, repeatedly clearing bars he suggested they would struggle with. Being early and being wrong can look identical for a long time, and the market has, so far, kept paying the LLM crowd. A thesis that has been “about to be vindicated” for several years deserves a raised eyebrow, even from admirers.
There is also the awkward fact that world models are, at this stage, more promising research direction than proven product. The idea that grounding AI in sensory reality unlocks reasoning is elegant and may well be right, but nobody has yet shipped a world model that does something economically valuable that an LLM cannot. AMI Labs has a brilliant founder and a billion dollars; it does not yet have the demonstration that turns the argument from philosophy into fact.
Finally, the open-source commitment, admirable as it is, cuts against easy monetisation. If you publish the science and release the code, your advantage is execution and talent, not secrecy. That is a noble way to build, and a harder way to build a durable business. LeCun clearly cares more about being right than about being proprietary, which is exactly why the science may flourish and the company may still struggle to capture the value.
How to think about it: bull, bear, neutral
Private company, so this is about reading the bet, not making one. (None of this is investment advice.)
Bull: a Turing laureate with a specific, deeply-considered technical thesis, a billion dollars, industrial backers who want exactly this technology, and an open approach that could make AMI Labs the intellectual centre of the whole world-models movement. If LeCun is right that LLMs plateau, he is the best-placed person alive to own what comes next.
Bear: a famous contrarian who has underestimated LLMs before, running a company whose core thesis is still unproven, with a business model that gives the science away. Being intellectually correct and building a valuable company are different achievements, and AMI Labs so far has only promised the first.
Neutral: watch the papers, not the press. AMI Labs will publish, so the evidence will be public: the moment a world model does something concrete and useful that a language model cannot, the thesis stops being a debate. Until then, treat it as the field’s most credible research bet rather than a proven direction.
What this means
AMI Labs is the closest thing AI has to a genuine philosophical fork in the road, led by the person most qualified to argue it. Almost everyone else is scaling language models harder; LeCun took a billion dollars to insist the answer lies elsewhere and to build it in the open. He may be years early, as he arguably has been before. He may also be about to be proven spectacularly right. Either way, the industry needs someone with his credibility taking the other side of the consensus, and this is the best-funded, most serious version of that argument anyone has mounted.
Related on Top Tool Stack: World Labs: spatial intelligence · Accelerated Understanding: physics AI
Did you know: LeCun shared the 2018 Turing Award for the deep-learning breakthroughs that made today’s language models possible. Now he has raised a billion dollars on the argument that those same models are a dead end.