In November 2025, Yann LeCun walked into Mark Zuckerberg's office and told his boss he was leaving. After twelve years building Meta's AI research operation into one of the most respected in the world, the Turing Award winner had reached a conclusion that put him at odds with virtually the entire industry: large language models are a dead end.
On Tuesday, investors handed him $1.03 billion to prove it.
Advanced Machine Intelligence Labs — AMI, pronounced like the French word for "friend" — announced the largest seed round ever raised by a European startup, valuing the Paris-based company at $3.5 billion. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with participation from Nvidia, Toyota, Samsung, and Singapore's Temasek. Individual backers include Tim and Rosemary Berners-Lee, Mark Cuban, Jim Breyer, and former Google CEO Eric Schmidt.
The company has just twelve employees and no product. What it has is a thesis — one that LeCun has been refining for years and that now sits at the center of a growing counter-movement in artificial intelligence research.
LeCun's argument is straightforward. Large language models learn by predicting which word comes next in a sequence. The results have been remarkable — ChatGPT, Claude, and Gemini can generate fluent, coherent text across an extraordinary range of subjects. But prediction is not understanding. These systems have no model of physical reality, no ability to plan ahead, and no mechanism for distinguishing plausible-sounding fiction from fact. They are, in LeCun's words, a "statistical illusion."
His alternative is something called JEPA — the Joint Embedding Predictive Architecture, a framework he first proposed in 2022. Rather than predicting the future state of the world word by word or pixel by pixel, JEPA learns abstract representations of how reality works, filtering out unpredictable surface-level noise. The goal is to build systems that understand the physical world the way humans and animals do: not through language, but through embodied experience. AMI calls these "world models."
"AMI Labs is a very ambitious project, because it starts with fundamental research," said Alexandre LeBrun, the company's chief executive and a former Meta engineer who previously ran the medical AI startup Nabla. "It's not your typical applied AI startup that can release a product in three months. It could take years for world models to go from theory to commercial applications."
That timeline has not deterred investors. LeCun initially sought around €500 million; demand pushed the round to nearly double that figure, with AMI ultimately turning away interested capital. The company plans to build teams across four locations — Paris, where it is headquartered; New York, where LeCun teaches at NYU; Montreal; and Singapore — prioritizing research talent and proximity to future corporate partners in healthcare and robotics.
AMI's first external collaboration will be with Nabla, applying world models to medical contexts where the hallucination tendencies of large language models carry potentially life-threatening consequences. Within three to five years, LeCun told reporters, the goal is to produce commercially viable systems that can power everything from healthcare diagnostics to robots operating in uncontrolled environments.
The founding team reads like a roster of Meta's former AI brain trust. Michael Rabbat, Meta's former director of research science, joins as vice president of world models. Laurent Solly, Meta's former VP for Europe, becomes chief operating officer. Pascale Fung, a former senior director of AI research at Meta, takes the role of chief research and innovation officer. Saining Xie, previously at Google DeepMind, becomes chief science officer.
AMI's launch comes at a moment when the AI talent exodus from Big Tech has become impossible to ignore. On the same day AMI announced its funding, Nvidia disclosed a "significant investment" in Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, along with a multiyear deal that will see Thinking Machines deploy at least one gigawatt of Nvidia's upcoming Vera Rubin systems. The two announcements together represent billions of dollars flowing away from established AI labs and toward founders who believe the next breakthrough will come from outside the companies that built the current one.
Whether LeCun is right about large language models remains one of the most consequential open questions in technology. The LLM paradigm has attracted hundreds of billions of dollars in investment and powers products used by hundreds of millions of people. Betting against it is not for the faint-hearted.
But LeCun has been right before. His pioneering work on convolutional neural networks in the 1990s was largely ignored for over a decade before becoming the foundation of modern computer vision and, eventually, the deep learning revolution itself. He is accustomed to being early, and to being alone.
The difference this time is that he is not alone at all. He has a billion dollars, a team of world-class researchers, and a growing number of investors who suspect that the most important AI company of the next decade might not be one that exists today.
