Twelve employees. Three months old. One billion dollars in the bank.

That’s AMI Labs — the startup founded by Yann LeCun, the Turing Award winner who spent a decade as Meta’s chief AI scientist before walking away to bet his entire legacy on one idea: every major AI company is building on the wrong foundation.

The $1.03 billion seed round values AMI Labs at $3.5 billion pre-money. That’s roughly $292 million per employee. And they haven’t shipped a single product.

Welcome to the world models era.

The Case Against Language Models

LeCun’s argument is devastatingly simple. ChatGPT, Claude, Gemini — they predict the next word in a sequence. They’re stunningly good at it. But predicting text isn’t understanding reality.

An LLM can write a flawless physics essay but can’t model what happens when you drop a ball. It can describe surgery in clinical detail but has zero grasp of the biology involved. It manipulates symbols without knowing what those symbols mean.

“Current AI approaches based on predicting the next word or pixel will not produce broadly capable intelligent agents by themselves,” LeCun told Reuters.

His alternative: world models. AI systems trained on reality itself, not language about reality.

Think of it as the difference between reading about how to ride a bicycle and actually learning to ride one. One gives you words. The other gives you understanding.

JEPA: The Architecture That Changes Everything

AMI Labs is built on JEPA — the Joint Embedding Predictive Architecture LeCun proposed in 2022 while still at Meta. The key insight is elegant: instead of predicting every pixel of what happens next, JEPA learns abstract representations and predicts at a conceptual level.

It handles uncertainty. It ignores irrelevant details. It focuses on what actually matters.

Sound familiar? That’s because it’s how humans think.

Meta demonstrated early versions with I-JEPA for images and V-JEPA for video. The V-JEPA model could watch a video with portions masked out and conceptually understand what was happening in the hidden parts — not by hallucinating pixels, but by building genuine scene understanding.

AMI Labs wants to take this from research demo to commercial-grade technology. Hence the billion dollars.

The Roster That Makes This Credible

Money is table stakes. The team is what makes AMI Labs terrifying to competitors.

Alexandre LeBrun as CEO brings something LeCun historically lacks: a track record of shipping products. He founded Wit.ai (acquired by Facebook in 2015) and led Nabla, a digital health AI company. Laurent Solly, Meta’s former VP for Europe, runs operations. Saining Xie, known for neural architecture design breakthroughs, is chief science officer.

The investor list reads like a who’s who of tech power: Jeff Bezos, Mark Cuban, Eric Schmidt, NVIDIA, Samsung, Toyota Ventures, and Temasek. When this many heavy hitters write checks for a pre-revenue company with a dozen employees, they’re not buying revenue projections. They’re buying a paradigm shift.

Where World Models Actually Matter

LeBrun was refreshingly honest about timelines: “AMI Labs is a very ambitious project, because it starts with fundamental research. It’s not your typical applied AI startup that can release a product in three months.”

But the target applications reveal why investors are salivating:

  • Healthcare: Partner Nabla is already working toward FDA-certifiable AI where hallucinations aren’t annoying — they’re potentially lethal
  • Manufacturing and aerospace: Complex physical systems need AI that understands how things work, not how to describe them
  • Automotive: Toyota Ventures’ participation signals autonomous driving and robotics applications
  • Consumer hardware: LeCun is in talks with Meta about deploying AMI tech in Ray-Ban smart glasses

The long-term prize? Domestic robots. Your Roomba can vacuum a floor. Ask any current AI to navigate your actual messy kitchen, and it’s lost. World models could bridge that gap — AI that has genuine common sense about how physical spaces work.

Competition Is Coming Fast

AMI Labs isn’t alone. Fei-Fei Li’s World Labs secured $1 billion last month. SpAItial raised a $13 million seed. Google DeepMind has internal world model programs.

LeBrun sees the wave coming: “My prediction is that ‘world models’ will be the next buzzword. In six months, every company will call itself a world model to raise funding.”

But there’s a difference between a pitch deck buzzword and years of peer-reviewed JEPA research. AMI Labs has committed to open-sourcing code and publishing papers — a stance that’s “increasingly rare” in today’s AI landscape and mirrors LeCun’s open-science philosophy from his Meta FAIR days.

The Real Question

Is a $3.5 billion valuation for a three-month-old company insane? By every conventional metric, yes.

But the bet isn’t on next quarter’s revenue. It’s on whether the entire LLM paradigm has a ceiling. If LeCun is right — if predicting text will never produce genuine understanding — then world models aren’t just another approach. They’re the approach. And AMI Labs has a multi-year head start on the foundational research.

If he’s wrong, a lot of very smart people just set a billion dollars on fire.

Either way, one of AI’s founding figures is staking his legacy on this. That alone should make everyone pay attention. The world models era has officially begun — and the race to build AI that actually understands reality is on.