In November 2025, Yann LeCun walked into Mark Zuckerberg's office and told his boss he was leaving.
He had spent twelve years at Meta, founding Facebook AI Research in 2013 and later serving as chief AI scientist. He had become one of the field's most decorated figures — a 2018 Turing Award winner, a professor at NYU, a co-pioneer of the deep learning revolution. He had also spent three years arguing publicly that the entire industry was heading in the wrong direction.
Large language models, he had insisted, were not the path to real intelligence. They were a statistical trick — systems that predict the next word so fluently they can pass for understanding, without actually understanding anything.
Meta had made its choice. In June 2025, Zuckerberg invested $14.3 billion in Scale AI and installed its CEO, Alexandr Wang, to lead a new LLM-focused division called Meta Superintelligence Labs. FAIR, the fundamental research lab LeCun had built, was deprioritized. The robotics group was dissolved. By November 2025, the distance between LeCun's worldview and Meta's direction had become irreconcilable.
On March 10, 2026, four months after leaving, LeCun's new company announced a $1.03 billion seed round. Advanced Machine Intelligence Labs — AMI, pronounced like the French word for "friend" — had raised what is believed to be the largest seed round ever completed by a European startup, at a pre-money valuation of $3.5 billion. The company has no product, no revenue, and twelve employees.
Investors include Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions as co-leads, with Nvidia, Samsung, Toyota Ventures, and Publicis Groupe as strategic backers, and individuals including Tim Berners-Lee, Mark Cuban, Jim Breyer, and former Google CEO Eric Schmidt.
They are not just funding a startup. They are funding a thesis most of the AI industry rejects.
What LeCun Is Actually Building
AMI Labs is building "world models" — AI systems that learn to understand the physical world from video, audio, and sensor data rather than predicting text. The core technical approach is JEPA, the Joint Embedding Predictive Architecture, which LeCun developed during his years at Meta.
The distinction from a standard LLM is fundamental. A language model predicts the next word in a sequence. A diffusion model predicts every pixel of a future image. Both require modeling everything — including all the random, unpredictable noise that has nothing to do with what is actually happening. The result is a model that is impressive at generation but carries no compact understanding of the causal structure of the world.
JEPA takes a different path. Rather than predicting the future in full sensory detail, it predicts the future in an abstract representational space, learning the underlying physical rules while ignoring irrelevant noise. LeCun's own framing: "The world is unpredictable. If you try to build a generative model that predicts every detail of the future, it will fail. JEPA learns to represent videos really well."
The intuition comes from how humans develop. An infant learning about gravity does not reconstruct a pixel-by-pixel image of a falling object. It builds an abstract model of cause and effect. JEPA trains on video, audio, and sensor streams to do something analogous — extracting physical rules in a form that supports reasoning and planning.
Meta published V-JEPA 2 before LeCun departed, demonstrating zero-shot robot control in environments the system had never seen. AMI plans to build on that foundation with its own model, AMI Video.
The Team and the Plan
AMI is headquartered in Paris, with offices planned in New York, Montreal, and Singapore. LeCun serves as executive chairman while keeping his NYU professorship. Day-to-day leadership falls to Alexandre LeBrun, formerly CEO of Nabla, the French medical AI startup, who becomes AMI's CEO. The founding team is stacked with Meta alumni: Michael Rabbat (VP of World Models), Laurent Solly (COO), and Pascale Fung (Chief Research and Innovation Officer) all came from FAIR. Chief Science Officer Saining Xie arrived from Google DeepMind.
AMI is not chasing chatbot customers. Its near-term commercial focus is on organizations running complex physical systems — manufacturers, automakers, aerospace firms, and biomedical companies where LLM-based AI has shown persistent limitations. Nabla, whose CEO now runs AMI, is the first official partner, getting early access to world model technology for healthcare documentation.
The consumer vision is further out: domestic robots, wearable AI assistants, and possibly Meta's Ray-Ban smart glasses, which LeCun described this week as "probably one of the shorter-term potential applications." Despite having left Meta, he suggested his former employer could become one of AMI's first customers — the two companies have a partnership agreement in place.
The Case For and Against
The skeptical view is straightforward. World models have shown promise in constrained robotics settings, but nothing approaching the generality that LLMs have achieved for language. AMI's first year will be pure R&D, with no commercial product. LeBrun himself predicted that "in six months, every company will call itself a world model to raise money" — an honest acknowledgment that the category is becoming a fundraising label.
The harder question is timing. LLM-based agentic AI is continuing to expand into physical-world applications. If multimodal training and improved action-prediction methods deliver sufficient commercial performance in manufacturing and robotics before AMI has a viable product, the window for a separate world model paradigm narrows considerably.
LeCun's counter is consistent: the easy path has been fully explored. The remaining frontier — genuine physical reasoning, persistent memory, real-world planning — requires different data, different architecture, and a different theory of intelligence. At Davos in January, he put it plainly: "The breakthroughs are not going to come from scaling up LLMs."
The $3.5 billion valuation for a pre-revenue, twelve-person company prices in the possibility that he is right. The investors backing him are not betting that LLMs will fail. They are betting they will not be sufficient — and that when the frontier shifts, LeCun's work will define what replaces them.
Frequently Asked Questions
Who is Yann LeCun, and why does his departure from Meta matter?
LeCun is one of the three founders of modern deep learning, sharing the 2018 Turing Award with Geoffrey Hinton and Yoshua Bengio. He founded Meta's AI research division in 2013 and served as chief AI scientist for twelve years. His departure signals a genuine rupture between the long-range fundamental research tradition he built at FAIR and the LLM commercialization direction Meta has now committed to under Alexandr Wang.
What is AMI Labs building?
AMI is building world models: AI systems that learn from video, audio, and sensor data rather than text, using LeCun's JEPA architecture. Rather than predicting the next word or pixel, JEPA predicts abstract physical representations — learning how the world works rather than how language describes it. Initial commercial targets are manufacturing, automotive, aerospace, and biomedical applications.
What was the funding round?
AMI raised $1.03 billion at a $3.5 billion pre-money valuation — believed to be the largest seed round ever by a European startup. Co-leads were Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Strategic investors include Nvidia, Samsung, Toyota Ventures, Temasek, and Publicis Groupe. The company has no product and no revenue.
What is JEPA and how does it differ from ChatGPT?
LLMs predict the next word in text sequences. JEPA predicts future states in an abstract representational space, trained on video and sensor data rather than text. It learns the underlying physical rules of the world — causal relationships, spatial dynamics — while ignoring unpredictable noise that generative models are forced to model. Meta's V-JEPA 2 demonstrated zero-shot robot control as a proof of concept. AMI will develop this into its own AMI Video model.
Is AMI competing with OpenAI or Anthropic?
No, at least not directly. LeCun has been explicit that AMI is not building a chatbot. The company targets physical-world AI in settings where LLM limitations are most acute. He has suggested Meta could become one of AMI's first customers, and the two companies have a formal partnership. The discussion of deploying AMI technology in Meta's Ray-Ban glasses points toward collaboration rather than competition with his former employer.
Who runs AMI Labs?
LeCun is executive chairman, remaining a professor at NYU. Alexandre LeBrun (former CEO of Nabla) is CEO. Key leaders include Michael Rabbat as VP of World Models, Laurent Solly as COO, Pascale Fung as Chief Research and Innovation Officer — all from Meta — and Saining Xie from Google DeepMind as Chief Science Officer. Headquarters are in Paris, with planned offices in New York, Montreal, and Singapore.
Related Articles
Dani
Dani
