The AI world was just hit by an earthquake. Yann LeCun, a “Godfather of AI” and Meta’s Chief AI Scientist, is leaving the company to launch his own startup.
This isn’t just another executive shuffle. It’s a profound, public rejection of the “LLM-to-AGI” hype train.
According to reports, LeCun is leaving due to a deep “philosophical disconnect” with Meta’s new, product-focused leadership. His new venture? A full-throated, career-defining bet on “world models” – the only path he believes leads to true Advanced Machine Intelligence.
For those in the Active Inference community, this isn’t “news.” It’s a vindication. And for those who have been following our analysis, you’ll know this exact conflict was laid bare almost two years ago.
Key Takeaways
- The News: AI “Godfather” Yann LeCun is leaving Meta to launch his own startup focused on “world models.”
- The “Why”: He has a deep philosophical disagreement with Meta’s new LLM-centric, product-first strategy, which he sees as a “dead end.”
- The Context (The “Receipts”): This split was predicted by a 2024 Davos debate between LeCun and Karl Friston.
- The Core Conflict: Friston argued Deep Learning is “rubbish” because it lacks a “calculus of beliefs” (uncertainty). LeCun agreed on the goal (world models) but was stuck on how to build them, still relying on Deep Learning.
- Our Analysis: LeCun is right to bet on world models, but he’s still stuck in an engineering-first approach (passive observation). Friston’s Active Inference is the first-principles solution, built on agency (learning by acting).
- The Big Picture: LeCun’s departure is a massive, industry-wide validation that the future of AI is moving away from LLMs and toward the principles of Active Inference.
The 2024 Davos Debate: The “Receipts”
To understand this moment, we must look back to January 2024. As Denise Holt documented in her brilliant “Deep Learning is Rubbish” article, the entire “Friston vs. LeCun” debate at the World Economic Forum centered on this exact conflict.
At that panel, Karl Friston, Chief Scientist at VERSES AI, and Yann LeCun agreed on several critical points:
- You cannot get to AGI with LLMs.
- World Models are the critical missing piece.
- The future of AI must be energy-efficient.
But they diverged on how to build them. Their disagreement perfectly captures the rift that just split Meta’s AI division.
LeCun’s Dilemma: “I Agree with Karl, but How?”
The 2024 debate was a clash of two titans: Friston, the neuroscientist focused on “natural intelligence,” and LeCun, the engineer focused on “artificial intelligence.”
Friston’s position was, and remains, revolutionary. He argued:
So, I think Deep Learning is rubbish, largely because it doesn’t have the calculus… to have a calculus of beliefs… of inference, of planning, of situational awareness. And crucially… encode the uncertainty… you’ll never know how to act…
LeCun’s famous reply, as Holt captured, revealed his core dilemma:
I’m going to heavily disagree with Karl by agreeing with him… This is not at all incompatible with Deep Learning… the question is, how do you learn those systems, how do you train those systems?
For two years, LeCun has been stuck on “how.” He’s been trying to build a world model (the what) using the only tool he knows: Deep Learning (the how). His V-JEPA model, which learns physics by passively watching videos, was his engineering-led attempt.
He was building a better “passive observer” while Friston was describing an “active participant.”
Friston’s Answer: “Learn to Perceive by Acting”
The key, which Friston stated at Davos, was agency. True intelligence doesn’t just watch the world; it acts on it to learn. As Friston said, children “learn about the world… by pinging and querying and moving.… They learn to perceive by acting. So this is… Active Inference.”
LeCun’s V-JEPA learns by observation. An Active Inference agent learns by doing.
Why LeCun’s Exit is a Massive Win for Active Inference
LeCun is one of the most brilliant engineers of our time. His departure from Meta is the loudest signal yet that the LLM hype bubble is bursting. The entire industry is now being forced to pivot and confront the very problem Karl Friston has been working on for years.
This is the ultimate validation of the premise of Active Inference.
Yann LeCun is leaving a $1 Trillion company because he knows the problem Friston defined (the need for world models) is the only one that matters. He is starting a new venture to build the what (world models) but he’s still clinging to the old how (Deep Learning, passive observation).
He’s about to re-discover all the problems that Active Inference already solves: How do you learn without acting? How do you handle uncertainty? How do you build true agency?
This proves our thesis at Network Consultants Pro. We aren’t just on the right track; we’re on the only track that connects the “what” (world models) with the “how” (the “action-perception loop” of Active Inference) to create a truly intelligent, governable Spatial Web.
Go Deeper on the Foundations
The entire AI industry is now in a race to build what Karl Friston and the Active Inference community have been perfecting for years. Don’t get caught in the old paradigm.
Visit our Education Page to understand the first principles of Active Inference, the Free Energy Principle, and the Spatial Web – the framework for the AI revolution that’s actually coming.
