By ByteBot
Yann LeCun, Turing Award winner and Meta’s former Chief AI Scientist, just launched Advanced Machine Intelligence (AMI Labs) seeking a $5 billion valuation. The startup has no product. The pitch: world models will replace large language models as the dominant AI architecture within years, and Silicon Valley is too “hypnotised by generative models” to build them.
On December 19, LeCun confirmed AMI Labs is raising €500 million at a €3 billion valuation—roughly $5 billion depending on exchange rates. Alex LeBrun, CEO of medical AI startup Nabla, is transitioning to lead AMI while remaining Nabla’s chairman. The company plans Paris headquarters in early 2026, positioning Europe as the center for post-LLM AI research.
World Models vs LLMs: The Technical Bet
World models are AI systems that understand physics, maintain persistent memory, and plan complex actions—capabilities LLMs fundamentally lack. LeCun’s criticism is blunt: LLMs “do not understand the physical world, do not have persistent memory, cannot reason in any reasonable definition of the term, and cannot plan.”
The learning difference is fundamental. Children learn by observing the world—gravity, object permanence, cause and effect—not by predicting the next word in a sentence. Similarly, world models mimic this approach, building internal representations of environments by processing video and spatial data rather than text. LeCun’s Joint Embedding Predictive Architecture (JEPA), already demonstrated in Meta’s I-JEPA and V-JEPA research, learns by creating internal models instead of comparing pixels.
LeCun predicts LLMs will be “largely obsolete within five years.” At Davos, he stated: “Nobody in their right mind would use them anymore, at least not as the central component of an AI system.” That’s a Turing Award winner publicly betting against the paradigm that dominates Silicon Valley.
$5B Pre-Launch: Bubble or Credibility Premium?
A $5 billion valuation before building a product sounds absurd—until you examine 2025’s AI funding environment. AI captured roughly 50% of all global venture funding this year. Unconventional AI raised $475 million at a $4.5 billion valuation two months after founding. Meanwhile, Thinking Machines Lab secured $2 billion at a $12 billion valuation in July.
LeCun’s track record justifies a premium. He pioneered convolutional neural networks, won the Turing Award for deep learning contributions, and built Meta’s AI research lab. However, credibility doesn’t guarantee execution. World models are computationally expensive, evaluation metrics remain unclear, and integration complexity far exceeds calling an LLM API.
The announcement signals investor appetite for LLM alternatives, but the valuation assumes world models can deliver on promises of deterministic reasoning, persistent memory, and physical understanding. If they can’t, this becomes another chapter in AI bubble behavior.
Medical AI as the Proving Ground
AMI Labs’ first application reveals the stakes. The company formed an exclusive partnership with Nabla, the medical AI startup LeBrun led before joining AMI. Nabla raised $120 million and will get first access to AMI’s world model technologies, aiming to build the first FDA-certifiable agentic AI systems in healthcare.
This choice is strategic. Medical AI is high-stakes—LLM hallucinations that invent citations are annoying in research, potentially fatal in diagnostics. If world models can provide deterministic reasoning in medicine, they prove their value. If they fail in healthcare, they likely fail everywhere.
The partnership also de-risks AMI Labs. Nabla already has medical AI traction and regulatory expertise. AMI gets a real-world testing ground; Nabla gets cutting-edge technology. Furthermore, LeBrun’s dual role as AMI CEO and Nabla chairman ensures tight integration.
Paris vs Silicon Valley: Geographic Rebellion
LeCun’s location choice carries a message. “Silicon Valley is completely hypnotised by generative models,” he told Sifted, “so you have to do this kind of work outside of Silicon Valley, in Paris.”
Paris offers advantages beyond contrarian positioning. The city has a large pool of AI research talent—Facebook and Google run research labs there—at salaries 2.5 to 3 times lower than Silicon Valley. Additionally, the Macron government invested billions in the startup ecosystem over the past decade. European AI startups now reach $1 billion valuations in roughly two years versus seven previously.
This is Europe betting on specialization while Silicon Valley chases scale. American companies pour resources into bigger LLMs; European researchers explore alternative architectures. If world models work, Paris becomes the center of the next AI wave. If not, it’s another chapter in Europe’s struggle to compete with US tech dominance.
The Competition is Already Shipping
AMI Labs isn’t alone. Fei-Fei Li’s World Labs raised $230 million at a $1 billion valuation and already shipped Marble, a commercial product that generates 3D environments from text and images. Decart, Odyssey, and Google’s Genie are also building world models.
World Labs has a head start. AMI Labs has a medical AI focus, European base, and LeCun’s explicit anti-LLM positioning as differentiation. Nevertheless, differentiation doesn’t matter if you ship too late. A $5 billion valuation won’t last if competitors dominate the market before AMI delivers.
What This Means for AI Development
If world models succeed, LLMs become specialized tools for text generation while world models handle reasoning, planning, and physical understanding. The entire AI development paradigm shifts from prompt engineering to environment modeling.
If world models fail, we learn that LLM limitations—hallucinations, no persistent memory, no planning—aren’t solved by new architectures, just traded for different problems. $5 billion gets wasted, and LLM dominance continues.
Medical AI is the first real test. FDA certification requires deterministic, explainable reasoning—exactly what LLMs can’t provide and world models promise. Watch Nabla’s progress. If they ship FDA-approved agentic diagnostics, LeCun’s bet pays off. If not, world models remain an interesting research direction without commercial viability.
Silicon Valley may be hypnotized, but Paris has something to prove.
