Yann LeCun Leaves Meta to Bet $3.5 Billion on World Models

The $3.5 Billion Gamble

When someone with LeCun's credentials raises €500 million at a €3 billion valuation before even launching, the industry pays attention. AMI Labs (Advanced Machine Intelligence) officially kicked off in January 2026 with a simple but revolutionary thesis: LLMs are a dead end for genuine intelligence.

LeCun has been saying this for years, but now he's putting his career where his mouth is. At the AI-Pulse conference in Paris, he didn't mince words: "Silicon Valley is completely hypnotized by generative models. You have to do this kind of work outside of Silicon Valley."

Why LLMs Aren't Enough

Here's the core argument, and it's surprisingly simple. LLMs predict the next token. That's it. They don't understand physics. They don't maintain persistent memory across sessions. They can't plan multi-step actions in the real world.

LeCun argues that this fundamental limitation means LLMs will never achieve the kind of contextual understanding humans take for granted. A toddler who's never seen a particular object can still predict that dropping it will make it fall. LLMs, despite being trained on the entire internet, can't reliably make that inference.

What Are World Models, Actually?

If you've been following the evolution of world models in AI video, you've seen glimpses of this technology. Runway's GWM-1 and World Labs' Marble are early attempts at building AI that understands spatial relationships and physics.

But AMI Labs is going further. LeCun's vision isn't just about better video generation, though that's certainly part of it. It's about AI systems that can:

• ✓Observe and interact with physical environments
• ✓Simulate "what if" scenarios before acting
• ✓Maintain context across complex, multi-step tasks
• ✓Transfer knowledge between different domains

Think of it as giving AI the ability to imagine. Not in the creative sense, but in the predictive sense. What happens if I push this button? What happens if I turn left instead of right? What happens if I combine these two chemicals?

The First Application: Healthcare

AMI Labs isn't starting with robotics or autonomous vehicles, though those are clearly on the roadmap. Their first deployment will be in healthcare through a partnership with Nabla, the medical transcription startup whose CEO, Alex LeBrun, now leads AMI Labs.

This is clever positioning. Healthcare has massive context-switching problems. A patient's journey involves dozens of touchpoints, each handled by different systems. If world models can maintain coherent understanding across that journey, it proves the technology works in high-stakes environments.

The Competitive Landscape

AMI Labs enters a crowded field, but with arguably the most credible founder:

What makes LeCun's approach different is his explicit rejection of the LLM scaling hypothesis. While OpenAI and Anthropic pour resources into making LLMs bigger, LeCun is betting on architectural innovation. He believes the breakthrough will come from how models learn, not how many parameters they have.

Why This Matters for AI Video

For those of us watching the AI video space, AMI Labs represents something important. The physics simulation improvements we've seen in recent models are baby steps toward world models.

Imagine generating a video and being able to say "now make the character pick up that object" and having the AI correctly simulate the physics of that interaction. That's where world models take us.

We've already seen hints of this in TurboDiffusion's real-time generation and Runway's experiments with interactive video. But those are still fundamentally diffusion models with physics sprinkled on top. True world models would flip the paradigm: physics first, appearance second.

The Paris Factor

One detail that caught my attention: LeCun is deliberately building AMI Labs outside Silicon Valley, with a strong European presence centered in Paris.

There's a pragmatic reason: European AI talent is world-class but often overlooked by American companies. But there's also a philosophical one. LeCun seems to believe that the groupthink around LLMs is so strong in the Bay Area that genuine innovation needs geographic distance.

For European AI, this is validation. One of the most decorated researchers in the field is betting that the next breakthrough will come from here, not Palo Alto.

What Comes Next

AMI Labs is just getting started, but the implications are significant. If LeCun is right, we're about to see a paradigm shift in how AI systems are built.

The biggest question is timing. World models are theoretically compelling, but can they deliver practical results fast enough to matter? LLMs may be limited, but they're useful now. OpenAI and Anthropic are building empires on that utility.

LeCun is betting that the LLM ceiling is lower than people think, and that when users hit it, they'll come looking for alternatives. AMI Labs wants to be ready.

The Bigger Picture

What I find exciting about this move isn't just the technology. It's what it represents for AI research culture. For too long, we've had a monoculture around transformer architectures and scale-is-all-you-need thinking. LeCun starting AMI Labs is a high-profile endorsement of alternative approaches.

Whether world models prove to be the path to AGI or not, having Yann LeCun fully committed to the approach means it will get a serious, well-funded attempt. And that's good for everyone who believes AI research benefits from diversity of thought.

The next few years will be fascinating to watch.