Yann LeCun just made the boldest move of his career. After 12 years at Meta—including time as the company’s Chief AI Scientist—the 65-year-old Turing Award winner has left to found AMI Labs, raising $1 billion in first-round funding at a $3.5 billion valuation. His mission: build AI that actually understands the physical world.
This isn’t just another AI startup launch. When one of the three “founding fathers of deep learning” bets his life’s work on an alternative approach to the current AI paradigm, the entire industry should pay attention.
The Bet Against Large Language Models
LeCun hasn’t been subtle about his views on LLMs. Speaking to French journalists recently, he criticized the “herd behavior” of AI giants caught “in a kind of rut” that pushes them to “all work on the same thing because they can’t afford to be late.”
The core problem, according to LeCun: Language models do not seek to understand the world. They seek to predict the next element in the sequence.
This fundamental limitation manifests in practical ways:
- LLMs hallucinate because they optimize for plausible text, not truth
- They struggle with spatial reasoning and physical causality
- They can’t reliably plan multi-step actions in the real world
- Robotics and autonomous vehicles remain stubbornly difficult
“These models do not allow for sufficient levels of intelligence for applications such as robotics or autonomous vehicles,” says Alexandre Lebrun, AMI Labs’ new CEO and LeCun’s former Meta colleague.
What Are World Models?
LeCun’s alternative vision centers on what he calls “world models”—AI systems capable of understanding how the physical world works without explicit supervision.
The classic example: a ball rolling toward the edge of a table. A world model should understand—without being told—that the ball will fall to the floor. This seems trivial to humans but represents a fundamental gap in current AI systems.
LeCun has been developing an architecture called JEPA (Joint Embedding Predictive Architecture) for nearly a decade. Unlike language models that predict the next token, JEPA learns representations of the world that enable prediction and planning.
The key differences:
| Approach | How It Works | Strength | Weakness |
|---|---|---|---|
| LLMs | Predict next token in text sequences | Language tasks, coding, reasoning | No physical world understanding |
| World Models | Learn causal relationships in physical reality | Robotics, embodied AI, planning | Still largely theoretical |
Why This Matters for Enterprise AI
The enterprise implications of this shift extend beyond robotics labs:
1. Reliability and Grounding
Current LLMs require extensive guardrails, RAG architectures, and human oversight because they lack true understanding. World models could provide a foundation for AI that reasons about consequences, not just correlations.
For enterprise applications where errors are costly—logistics, manufacturing, medical devices—this reliability gap is a blocker.
2. The Physical World Interface
As AI moves from pure software into physical operations, LLM limitations become critical. Supply chain optimization, warehouse robotics, autonomous inspection, predictive maintenance—all require AI that understands physical causality.
Companies investing heavily in AI-driven physical automation should watch AMI Labs closely.
3. Reduced Compute Requirements
LeCun has argued that world models could be far more efficient than LLMs, which scale compute exponentially. If true, this could democratize advanced AI capabilities for companies without hyperscaler budgets.
4. European AI Competitiveness
AMI Labs represents a significant win for the European AI ecosystem. LeCun, born in France’s Parisian suburbs, is building a company that could challenge American AI dominance with a fundamentally different approach.
With Laurent Solly (former Meta France executive) as COO, the company has strong European leadership even if headquartered globally.
The AMI Labs Team
The founding team combines research excellence with operational experience:
Yann LeCun (Chairman)
Turing Award winner (2018), former Meta Chief AI Scientist. Published over 200 research papers. Developed foundational work on convolutional neural networks and deep learning.
Alexandre Lebrun (CEO)
Former Meta AI executive. “I started reading his papers when I was a student. For me, he was God,” Lebrun said of LeCun. Brings operational and product experience to the research vision.
Laurent Solly (COO/Co-founder)
Former Meta France executive. Provides enterprise go-to-market experience and European market expertise.
The Contrarian Thesis
LeCun’s move requires believing several contrarian positions:
- LLMs will hit fundamental limits that can’t be fixed with more scale
- World models are buildable with current (or near-future) technology
- A startup can out-execute Big Tech on foundational AI research
- The market timing is right for a paradigm shift
The first two are scientific questions. The latter two are business questions. $3.5 billion in valuation suggests investors believe the answers are yes.
What About the Competition?
LeCun isn’t the only researcher skeptical of pure LLM approaches:
- Geoffrey Hinton (fellow Turing Award winner) has expressed concerns about LLM limitations
- Yoshua Bengio (third co-winner) remains focused on academic research
- Google DeepMind continues investing in world models and embodied AI
- Tesla’s FSD implicitly relies on world model concepts for driving
But AMI Labs is unique in having a Turing Award-winning researcher building a company specifically around this thesis, with significant capital from day one.
Implications for AI Strategy
For enterprise AI leaders, the AMI Labs launch suggests several strategic considerations:
Short-term (2026-2027)
Continue investing in LLM-based solutions—they’re the proven approach with established tooling. But monitor world model developments and avoid over-committing to architectures that may face fundamental limits.
Medium-term (2027-2028)
Evaluate world model applications as they emerge, particularly for robotics, automation, and physical world interfaces. Consider pilot projects that test reliability and planning capabilities.
Long-term (2029+)
The AI landscape may look fundamentally different if world models prove viable. Build flexibility into AI architecture choices to accommodate potential paradigm shifts.
The Risk: “His Thesis Remains to Be Proven”
Not everyone is convinced. Antoine Bordes, head of research at Helsing (and former LeCun colleague), offers a measured take: “The limitations of the language models he notes are no longer so controversial. However, his thesis on world models remains to be proven.”
This captures the situation perfectly: LeCun is right that LLMs have limitations. Whether world models are the solution—and whether AMI Labs can build them—remains uncertain.
Conclusion: The Third AI Revolution?
LeCun describes world models as “the third AI revolution.” The first was rule-based systems. The second was deep learning (which LeCun helped create). The third, he argues, will be AI that understands how the physical world works.
At 65, with $1 billion in funding and the credibility of a Turing Award, LeCun is betting everything on this vision. As Alexandre Lebrun puts it: “This is his life’s work.”
For enterprise AI leaders, the message is clear: the current LLM paradigm may not be the end state. The most important AI breakthrough of the next decade might come not from scaling existing models, but from fundamentally rethinking how AI understands the world.
AMI Labs just became the company to watch.
At Virge, we help organizations navigate the evolving AI landscape. Understanding paradigm shifts—whether they succeed or fail—is essential for building robust AI strategy. Contact us to discuss how these developments affect your enterprise AI roadmap.