Yann LeCun leaves Meta: LLMs won't reach human-level intelligence

TL;DR

• Yann LeCun (Turing Award, CNN inventor) leaves Meta to start AMI Labs ($3.5B)
• His thesis: “We won’t reach human-level intelligence just by scaling LLMs”
• Drama at Meta: Llama 4 benchmarks manipulated, 28-year-old new boss, researcher exodus
• The bet: World Models that understand physics, not just language. First versions in 2027.

---

The news

Yann LeCun, Meta’s Chief AI Scientist for 12 years, is leaving. Not for another big tech. To start his own company: AMI Labs, valued at $3.5 billion before opening.

Why does it matter? Because LeCun isn’t just anyone:

• Turing Award 2018 (the “Nobel” of computing)
• Invented convolutional neural networks (CNNs)
• Founded FAIR, Meta’s AI lab
• One of the three “godfathers of AI” alongside Hinton and Bengio

And he’s leaving saying that LLMs—the technology dominating everything right now—won’t reach human-level intelligence.

---

What he said

“We’re not going to get human-level intelligence just by scaling LLMs.”

“A house cat has more common sense than GPT-4.”

“We need models that understand the physical world, not just predict the next word.”

In his last podcast, LeCun was more direct:

“Before we urgently worry about how to control AI systems much smarter than us, we need to have the beginning of a hint of a design for a system smarter than a house cat.”

---

The drama inside Meta

The departure wasn’t friendly. According to the Financial Times:

The Llama 4 fiasco:

• Meta launched Llama 4 in April 2025
• The benchmarks were manipulated (“the results were inflated a bit,” LeCun admitted)
• They used different models for each benchmark to pump up numbers
• Zuckerberg “lost confidence in everyone involved”

The new boss:

• Meta hired Alexandr Wang (28, Scale AI CEO) for $14 billion
• Wang technically became LeCun’s boss
• LeCun: “You don’t tell a researcher what to do. You certainly don’t tell me what to do.”

The exodus:

“A lot of people have left, a lot of people who haven’t left yet will leave.”

---

The bet: World Models

LeCun isn’t retiring. He thinks he has the solution. They’re called World Models.

The idea: instead of predicting the next word (like ChatGPT), predict the next state of the physical world.

LLM	World Model
Learns from text	Learns from video
Predicts words	Predicts physical states
Understands language	Understands physics
”What word comes next?"	"What happens if I drop this ball?”

If you want to understand what World Models are and why LeCun believes they’re the future, I explain it in detail in World Models: AI that predicts the physical world.

---

Why Paris

LeCun is French-American. His startup will be based in Paris.

“You have to do this kind of work outside the Valley, in Paris.”

Macron texted him when the news was announced (LeCun won’t say what it said).

Paris is the #3 AI hub globally according to Dealroom. Europe is betting on smaller, more sustainable models instead of American megamodels.

---

My take

It’s easy to dismiss this as “old scientist who doesn’t get LLMs.”

But LeCun invented CNNs. Without him there would be no image recognition, no self-driving cars, none of half the things we take for granted.

And 76% of AI researchers surveyed by AAAI agree: scaling LLMs alone won’t reach AGI.

Is he right? I don’t know. But when one of the three godfathers of AI says “this isn’t going to work” and bets $3.5 billion on something else, it deserves attention.

---

What’s next

AMI Labs launches in January 2026. First “baby” versions in a year. Full versions in “a few more years.”

“Maybe there’s an obstacle we don’t see yet, but at least there’s hope.”

If World Models work, we could see:

• Robots that understand physics without programming rules
• Cars that predict pedestrian behavior
• AI that plans complex actions in the real world

If they don’t work, LeCun will have lost a very expensive bet.

But that’s what real scientists do: bet on ideas that might be wrong.