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 decided the entire industry was headed in the wrong direction.
Now he has $1.03 billion to prove it.
Advanced Machine Intelligence Labs (AMI) - pronounced like the French word for “friend” - announced Europe’s largest seed round ever on March 10, with backing from Jeff Bezos, Nvidia, Eric Schmidt, and a consortium of European investors including Cathay Innovation and HV Capital. The funding values the Paris-based company at $3.5 billion before it has built a product.
LeCun’s pitch: large language models are “a statistical illusion. Impressive, yes. Intelligent, no.”
The Case Against LLMs
LeCun has been making this argument for years, but leaving Meta to stake his reputation on it signals something different. His critique centers on how LLMs work: they predict text word-by-word, learning patterns from language without understanding what the words mean.
“AI systems based solely on large language models are unlikely to produce broadly capable intelligent agents,” LeCun stated in the announcement.
The problem, in his view, is that this architecture is inherently prone to hallucination. LLMs focus on surface-level pattern matching rather than genuine understanding. They can write fluently about physics without knowing what gravity feels like.
His alternative is JEPA - the Joint Embedding Predictive Architecture he first proposed in 2022. Rather than predicting future states in pixel-perfect or word-by-word detail, JEPA learns abstract representations of how reality works. It’s designed to mirror how humans and animals actually understand the physical world: through embodied experience, not language.
The Business Model (Eventually)
AMI has no product, no revenue, and a five-year timeline. The plan:
- Year one: Pure research and development
- Years 1-2: Begin corporate partner discussions
- Years 3-5: Deploy “fairly universal intelligent systems” across industries
Target customers include manufacturers, automakers, aerospace companies, and pharmaceutical firms - industries where AI needs to understand physical processes, not just generate text.
The company’s first partnership is with Nabla, a French medical AI startup run by AMI CEO Alexandre LeBrun. The connection is strategic: LeBrun sold his previous company to Facebook and has operational credibility that LeCun, as a researcher, lacks.
There’s also discussion of integrating world models with Ray-Ban Meta smart glasses, suggesting LeCun’s departure from Meta may not have been entirely adversarial.
The Competition
AMI isn’t alone in betting on world models. Fei-Fei Li’s World Labs in San Francisco raised $1 billion in late February 2026 for similar research. The timing isn’t coincidental - both LeCun and Li are reacting to the same perceived limitations.
But the funding environment for this approach is still dwarfed by LLM investment. OpenAI raised $110 billion in February. Anthropic raised $30 billion. AMI’s billion-dollar round barely registers against those numbers.
“In six months, every company will call itself a world model company to raise funding,” LeBrun told reporters. He’s probably right. The question is whether AMI can maintain its differentiation.
The Risk
The blunt assessment from The Next Web: AMI’s success depends on whether “being right about the problem” equals “being right about the solution.”
LeCun has been publicly critical of LLMs for years while the industry has poured hundreds of billions into them. He may be correct that current approaches have fundamental limitations. But JEPA is still theoretical. No one has proven it works at scale.
The $3.5 billion valuation rests entirely on LeCun’s reputation and the belief that he sees something others don’t. That’s either visionary investing or the most expensive research grant in European history.
What This Means
If LeCun is right, the current AI paradigm hits a wall. Systems that predict language well can’t generalize to physical reasoning. The companies that survive are those building AI that understands reality, not just text patterns.
If he’s wrong, AMI becomes a cautionary tale about betting against market momentum. The investors lose their billion dollars, LeCun’s legacy gets complicated, and the LLM paradigm continues unchallenged.
Either way, this is the clearest statement yet that serious researchers see limits to the current approach. When a Turing Award winner leaves his prestigious position to build something fundamentally different, it’s worth paying attention - even if you’re not sure he’s right.