The Godfather of AI Just Bet Against Everything We Call AI

Yann LeCun invented the technology behind every image AI you've ever used. Now he's leaving Meta to build something he thinks current AI can't do.

That should make you pay attention.

The Split

LeCun is one of three "godfathers of AI" — he won the Turing Award in 2018 with Geoffrey Hinton and Yoshua Bengio for creating deep learning. Convolutional neural networks (CNNs), the architecture powering every visual AI from facial recognition to medical scans? That's his.

He spent a decade at Meta building FAIR, the company's AI research division. He had resources. He had influence. He walked away in November 2025.

The reason's simple: he thinks Mark Zuckerberg's betting on the wrong horse.

What He's Building Instead

AMI Labs (Advanced Machine Intelligence) isn't making another chatbot. It's building "world models" — AI that understands physics, causation, and planning. Not text prediction. Not pattern matching from scraped internet data. Systems that know what happens when you drop a glass or push a door.

LeCun's been calling LLMs a "dead end" for years. Now he's putting his career behind that claim.

His argument? LLMs are "disembodied mimics." They autocomplete on a cosmic scale, but they don't understand the world. They can't reason about cause and effect. They can't plan. They live in language, not reality.

World models would change that. They'd learn how the physical world works — persistent memory, spatial reasoning, anticipating outcomes. Think less chatbot, more robot that understands gravity.

The Other Team Disagrees

LeCun's not the only godfather with an opinion.

Geoffrey Hinton (the other Turing Award winner) thinks LLMs already contain "small-scale models of external reality" buried in their neurons. He's argued they're closer to understanding than LeCun gives them credit for.

The AI industry's racing ahead with LLMs. OpenAI, Anthropic, Google — everyone's scaling up language models, not pivoting to world models.

Meta itself is doubling down on LLMs. They're building two new models codenamed "Mango" and "Avocado" for 2026. LeCun had "almost nothing to do" with Llama after the first version shipped.

That's the schism. The people who built AI can't agree on where it goes next.

Why It Matters

When someone leaves a trillion-dollar company to start from scratch, they're not hedging. LeCun's raising €500M at a $3.5-5B valuation for AMI Labs. That's not a side project — it's a bet that the entire AI boom is heading in the wrong direction.

He might be right. LLMs are stuck in text. They hallucinate. They can't update their knowledge without retraining. They break when you ask them to reason through novel problems.

Or he might be wrong. Maybe scaling LLMs another 10x solves the problem. Maybe embodied understanding emerges from enough text.

Either way, the split's real. The people who invented deep learning are now building competing visions of what comes next.

The Race Nobody's Watching

While everyone watches ChatGPT, a quieter race is starting. DeepMind's building Genie 3, which simulates real-time 3D worlds. LeCun's AMI Labs is targeting healthcare first — partnering with Nabla, a medical AI startup, to get "privileged access" to world models for clinical care.

The bet's simple: if you want AI that works in the real world — diagnosing patients, controlling robots, planning in 3D space — language alone won't cut it.

LeCun developed JEPA (joint embedding predictive architecture) at Meta specifically for this. It's designed to help AI understand the world without relying on words.

Now he's free to build it without compromise.

What This Tells Us

The AI field's splitting. One path: keep scaling LLMs, bet on emergence. The other: build systems that understand physics and causation from the ground up.

LeCun's choosing the second path. So is a growing group of researchers who think LLMs hit a ceiling — useful tools, but not the route to real intelligence.

The trillion-dollar question: which path gets there first?

When one of the three people who built the foundation of modern AI says the current approach is broken, that's not background noise. It's a signal worth watching.

---

FAQs

What are world models?

AI systems designed to understand how the physical world works — physics, causation, spatial relationships. Unlike LLMs (which predict text), world models learn persistent memory, planning, and reasoning about real-world scenarios.

Why did LeCun leave Meta?

Fundamental disagreement with Mark Zuckerberg over AI strategy. Meta's doubling down on large language models (LLMs). LeCun thinks that's a dead end and wants to build world models instead.

Who are the "godfathers of AI"?

Yann LeCun, Geoffrey Hinton, and Yoshua Bengio — the three researchers who won the 2018 Turing Award for pioneering deep learning. LeCun invented convolutional neural networks (CNNs), the architecture behind modern image AI.

Do the other godfathers agree with LeCun?

No. Geoffrey Hinton thinks LLMs might already contain internal world models. The field's splitting over whether scaling LLMs leads to real intelligence or whether we need fundamentally different architectures.

What's JEPA?

Joint Embedding Predictive Architecture — a learning framework LeCun developed at Meta to help AI understand the world without relying on language. It's the foundation for AMI Labs' world model approach.

How much funding is AMI Labs raising?

Reportedly seeking €500M at a $3.5-5B valuation. The company launched in January 2026 with Alex LeBrun (Nabla co-founder) as CEO and LeCun as executive chairman.