Yann LeCun Left Meta to Build a Brain — The Theory Behind It Came From a 1943 Book

Key Takeaways

• •Yann LeCun left Meta in late 2025 after 12 years as Chief AI Scientist
• •AMI Labs raised over $1 billion — Europe's largest-ever AI seed round
• •LeCun argues LLMs are a dead end — real AI needs 'world models' that understand physics
• •The 'world model' concept traces to Kenneth Craik's 1943 book about internal mental models

Timeline

In 1943, Scottish psychologist Kenneth Craik published a slim volume called 'The Nature of Explanation.' Its central argument was deceptively simple: the human brain builds small-scale models of external reality and uses them to predict events. We don't just react to the world. We simulate it internally. We run mental experiments before taking physical actions.

Craik died in a cycling accident in 1945 at age 31. His book was largely forgotten outside cognitive science.

Eighty-two years later, Yann LeCun — one of the three 'godfathers of deep learning,' a Turing Award winner, and Meta's Chief AI Scientist for over a decade — left Meta to build exactly what Craik described. His new company, AMI Labs, raised over $1 billion in early 2026, making it the largest seed round in European AI history. Backers include NVIDIA, Temasek, and capital linked to Jeff Bezos.

LeCun's thesis is a direct challenge to the AI industry's dominant paradigm. Large Language Models — GPT, Claude, Gemini, Llama — are fundamentally limited, he argues, because they only predict the next token in a sequence of text. They don't understand physics. They can't predict what happens when you push a cup off a table. They have no model of the world.

LeCun's argument is stark: LLMs only predict the next word. They don't understand physics, causality, or the physical world. They are 'autoregressive parlor tricks,' he says. Building real intelligence requires world models.

LeCun calls LLMs 'autoregressive parlor tricks.' Impressive, yes. But not intelligence. Not even close.

AMI Labs is building what LeCun calls a 'world model' — an AI system that learns to predict the physical world the way a baby does. Babies don't learn language first. They learn physics. They learn that objects fall, that things hidden behind other things still exist, that actions have consequences. This intuitive physics comes before language, before reasoning, before everything else.

This is exactly Craik's argument from 1943. Intelligence is not pattern matching. It's not statistical prediction. It's building an internal model of reality that's accurate enough to simulate the future.

This is the most significant philosophical schism in AI since the symbolic-vs-connectionist debate of the 1980s. If LeCun is right, the entire LLM gold rush is building on sand.

The implications are enormous. If LeCun is right, the current LLM gold rush — hundreds of billions of dollars invested in scaling language models — is building on a foundation that can never achieve true intelligence. You can scale GPT to a trillion parameters, and it still won't understand that water flows downhill.

This is the most significant philosophical schism in AI since the 1980s, when the field split between symbolic AI (rules and logic) and connectionism (neural networks and learning). The connectionists won that debate. LeCun was one of them — his 1998 paper on convolutional neural networks (LeNet-5) proved that neural networks could read handwritten digits, launching the deep learning revolution.

Now LeCun is saying that his own revolution doesn't go far enough. Neural networks that predict text are a stepping stone, not a destination. The destination is machines that understand the world.

The AI community is divided. OpenAI and Anthropic believe that scaling language models, combined with reinforcement learning and tool use, will eventually produce general intelligence. LeCun believes this is like trying to reach the moon by climbing a very tall tree. You're going up, but you're on the wrong vehicle.

A billion dollars says LeCun is right. Or at least that his investors think he might be.

Kenneth Craik never saw a computer. He died two years before the first electronic general-purpose computer (ENIAC, 1945) was completed. But his argument — that intelligence requires an internal model of physical reality — may turn out to be the most important idea in AI, 82 years after he wrote it down.

How did this make you feel?

Recommended Gear

Keep Reading

AI for Good

Amazon Cut 16,000 Jobs Citing AI — The Luddites Fought This Same Battle in 1811

6 min read

AI for Good

AI Is Now Detecting Wildfires Within 10 Minutes of Ignition. The Root of Fire Lookouts Goes Back to the Roman Empire.

7 min read

AI for Good

AI Can Now Translate Sign Language in Real Time. The Fight for Deaf Communication Rights Is 250 Years Old.

7 min read