A Resignation That Changed Everything
In November 2025, Yann LeCun walked into Mark Zuckerberg’s office and did something that stunned the tech world — he quit.
For twelve years, LeCun had been the beating heart of Meta’s artificial intelligence ambitions. He built FAIR — Meta’s Fundamental AI Research lab — into one of the most respected AI institutions on the planet. He won the Turing Award, the Nobel Prize equivalent of computing. He trained a generation of AI researchers. And quietly, steadily, he became the industry’s most inconvenient voice — the man who kept insisting, publicly and loudly, that everyone was building AI the wrong way.
So when he left, people didn’t just take notice. They started writing cheques.
Four months after walking out of Meta, LeCun announced that his new venture — AMI Labs — had raised $1.03 billion in its very first funding round. No product. No revenue. No customers. Just a bold idea, a legendary name, and a bet that the current path of AI is heading in the wrong direction.
What Does “AMI” Even Mean?
The name is quietly poetic. In French, ami means friend. For a company building AI that’s supposed to work alongside humans — not just generate text at them — it’s a deliberate choice.
Headquartered in Paris, with outposts in New York, Singapore, and Montreal, AMI Labs isn’t positioning itself as just another AI startup. Its founding philosophy is captured in a single line that they’ve essentially turned into a manifesto:
“Real intelligence does not start in language. It starts in the world.”
That one sentence tells you everything about why AMI exists — and why LeCun believes the billion-dollar chatbot industry has fundamentally misunderstood what intelligence actually is.
Why LeCun Walked Away — And Why It Matters
To understand AMI, you have to understand LeCun’s frustration.
For years, while the rest of the world was celebrating ChatGPT, Gemini, and Claude as revolutionary leaps in intelligence, LeCun was the person in the back of the room raising his hand and saying, “Wait — this isn’t real intelligence.”
His argument isn’t that large language models are useless. It’s that they’re statistical illusions dressed up as understanding. These models work by predicting the next most probable word in a sequence. They don’t know that fire is hot. They don’t understand that dropping a glass will break it. They can’t plan a route through a building because they’ve never experienced space. They’ve simply read so much text about the world that they’ve learned to sound convincingly like they understand it.
And in LeCun’s view, that’s a ceiling — not a foundation.
He believed the only way to build truly intelligent machines was to start over. To build AI that experiences the world the way animals do — through observation, interaction, cause and effect — rather than AI that just reads about it.
The problem? That kind of research is harder, slower, and far less commercially exciting in the short term. Inside a company focused on deploying products, it’s a difficult case to make.
Outside? Turns out, it’s worth a billion dollars.
The Funding Round — Numbers That Broke Records
Let’s talk about the money, because the scale here is genuinely extraordinary.
- Total raised: €890 million (~$1.03 billion USD)
- Valuation: $3.5 billion — before a single product has shipped
- Round type: Seed — making this widely believed to be the largest seed round ever raised by a European startup
- Timeline: Raised in roughly four months from founding
The round was co-led by five major investment firms: Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions — the last one being Jeff Bezos’s personal investment vehicle.
But the investor list reads less like a cap table and more like a guest list for a dinner you’d never get invited to:
- Jeff Bezos — Amazon founder
- Eric Schmidt — Former CEO of Google
- Tim Berners-Lee — The man who invented the World Wide Web
- Mark Cuban — Entrepreneur and investor
- Xavier Niel — French billionaire and telecom mogul
- Jim Breyer — One of Silicon Valley’s most respected VCs
And on the corporate side: Toyota, Nvidia, and Samsung — three companies with a deeply vested interest in the kind of physical-world AI that AMI is building.
LeCun later revealed that he initially set out to raise around €500 million. Demand was so overwhelming that he ended up being selective about which investors to accept and still walked away with nearly double his original target.
The Team Behind the Mission
LeCun may be the face of AMI, but this is far from a one-man operation. The founding team is stacked — and almost entirely drawn from the world’s top AI research institutions, many of them having worked directly under LeCun at Meta.
Yann LeCun — Executive Chairman Turing Award winner. NYU professor. The godfather of modern deep learning. He’s not the day-to-day CEO — he’s the intellectual engine and the public face of AMI’s research philosophy.
Alexandre LeBrun — CEO LeBrun is the operator of the group. He previously co-founded Nabla, a medical AI startup, and before that founded Wit.ai, which Facebook acquired in 2015. He knows what it takes to build a company from scratch — and he knows firsthand the dangers of AI that hallucinates in high-stakes environments like hospitals.
Laurent Solly — COO Former Vice President at Meta for Europe, Solly brings the kind of institutional and regional knowledge that a Paris-headquartered global startup desperately needs.
Saining Xie — Chief Science Officer One of the most respected computer vision researchers in the world, Xie spent years at Meta’s FAIR lab and brings deep technical credibility to AMI’s research roadmap.
Pascale Fung — Chief Research and Innovation Officer A leading figure in AI and natural language processing, Fung brings a rare combination of academic rigour and real-world AI deployment experience.
Michael Rabbat — VP of World Models The title says it all. Rabbat is specifically tasked with leading the development of the core technology that AMI is built around.
So, What Actually Is a “World Model”?
This is the heart of the story — and once you understand it, the whole AMI bet makes a lot more sense.
A world model is an AI system that doesn’t just process language — it builds an internal simulation of how the physical world works. It understands that objects fall when dropped. It knows that a car turning left on a wet road at speed is likely to slide. It can predict what happens next in a physical situation — not because it read a description of it, but because it has learned the underlying rules of reality.
The technology AMI is developing is called JEPA — Joint Embedding Predictive Architecture. LeCun first proposed it in 2022, and here’s why it’s fundamentally different from what OpenAI or Google are building:
- Most AI systems today — the LLMs — are generative. They try to predict and produce every tiny detail of an output. That’s why they hallucinate — when they don’t know something, they fill in the gap with something that sounds plausible.
- JEPA is not generative. Instead of trying to predict every pixel or every word, it learns to understand representations — high-level, compressed understandings of what something is, not just what it looks like or sounds like.
- Think of it like the difference between a person who memorised a map of a city versus someone who actually lived there. The map-memoriser can describe the streets. The resident knows that a certain road floods in monsoon, that a particular turn feels different in fog, that the shortcut only works at certain times of day.
AMI wants to build the resident, not the map-memoriser.
Where Will This Technology Actually Be Used?
AMI isn’t building a chatbot. Its ambitions are in sectors where real-world understanding is not a nice-to-have — it’s a matter of life and safety.
1. Healthcare AMI’s first confirmed partner is Nabla — LeBrun’s own former company — which will get privileged early access to AMI’s world models. The use case is direct: AI that can assist in clinical settings without hallucinating. In a context where a fabricated diagnosis or a wrong medication suggestion could cost someone their life, this matters enormously.
2. Robotics A robot navigating a factory floor, a surgical assistant guiding a procedure, a drone delivering medicine in a remote area — all of these require an AI that genuinely understands physical space and can predict outcomes in real time. That’s exactly what world models are designed for.
3. Autonomous Vehicles and Driving Self-driving cars have stalled partly because current AI struggles to reliably handle the unpredictability of the real world. AMI’s approach — building AI that actually simulates physical dynamics — could be the missing piece.
4. Industrial and Hardware Design LeCun has spoken about helping companies analyse aircraft components, optimise industrial designs, and identify failure points before they happen — applications where a hallucination isn’t just embarrassing, it’s catastrophic.
5. Wearables and Consumer Intelligence Longer term, AMI envisions AI embedded in everyday devices — glasses, watches, personal assistants — that understand your environment and context in a genuinely intelligent way, not just through voice commands and calendar access.
The Open-Source Promise
In an industry increasingly defined by secrecy — where the biggest labs treat their models like nuclear secrets — AMI is taking the opposite approach.
LeBrun has stated clearly that AMI will publish its research openly and release significant portions of its code as open source. The reasoning is strategic as much as idealistic: open ecosystems attract talent, build communities, and accelerate progress faster than any single closed team ever could.
It also signals confidence. You only open-source your work when you believe your edge isn’t just in what you’ve built — it’s in what you’re capable of building next.
What Happens From Here?
AMI has been refreshingly clear about its timeline — and refreshingly honest that this is a long game.
- Short term (now to 12 months): Hire a core team of 20–30 researchers and engineers. Begin internal R&D. Publish early research papers. Release first preliminary models.
- Medium term (1–2 years): Begin formal conversations with corporate partners. Start deploying technology in controlled environments, particularly in healthcare.
- Long term (3–5 years): Build what LeCun calls “fairly universal intelligent systems” — AI capable of being deployed across any domain that requires machine intelligence interacting with the real world.
And in a twist that perhaps no one expected — LeCun has hinted that Meta itself could become one of AMI’s first clients. “The work we are doing is not in direct competition,” he said. “Our focus on world models for the physical world is very different from their focus on generative AI and LLMs.”
The student may have left the building. But the teacher is happy to sell them a better tool.
The Bigger Picture — Why This Matters
Here’s the thing about AMI that goes beyond the funding number: it represents a genuine philosophical split in the AI world.
On one side, you have the LLM camp — OpenAI, Google DeepMind, Anthropic, Meta’s Llama — all doubling down on making generative AI bigger, faster, and more capable. The bet is that scale solves everything. More data, more compute, more parameters — and intelligence will emerge.
On the other side, you now have LeCun and AMI — arguing that scale isn’t the answer if you’re building on the wrong foundation. That true intelligence requires understanding the world, not just describing it. That until AI can reason, plan, and operate in physical reality with genuine comprehension, we’re building increasingly sophisticated autocomplete — not artificial general intelligence.
AMI’s CEO put it with a wry smile when he noted that “world models” would probably become the next AI buzzword, with every startup suddenly claiming to build them just to attract investment.
“The difference,” he said, “is that we actually mean it.”
Over a billion dollars says the world is willing to find out if he’s right.
