AMI Labs: LeCun's $3.5B bet against the AI mainstream

Yann LeCun thinks you're all doing it wrong.

The former Meta Chief AI Scientist walked away from one of the best-resourced AI labs on the planet because he believes the entire industry is building on sand. Scaling transformers, throwing more compute at next-token prediction, hoping intelligence will somehow fall out of bigger models. He's been saying it for years. Now he's put his career where his mouth is.

AMI Labs, which stands for Advanced Machine Intelligence Labs, launched about a month ago. Today the company announced a seed round north of $1 billion, putting the valuation at $3.5 billion. For a company with 12 employees. Let me say that again: twelve people, three and a half billion dollars.

Jeff Bezos wrote a check. So did Mark Cuban. The investor list reads like a who's who of people who got rich betting against conventional wisdom. Whether that pattern holds here is the question nobody can answer yet.

What LeCun actually believes

LeCun has been one of the most vocal critics of the large language model paradigm. His argument goes something like this: autoregressive text generation is a parlor trick. It produces fluent language without understanding. It can't plan. It can't reason about the physical world in the way a toddler can. And no amount of scaling will fix that.

He's been pushing an alternative vision he calls "world models," systems that build internal representations of how things work and use those representations to plan and predict. Think less "guess the next word" and more "simulate what happens next." It's closer to how we think biological intelligence operates, though comparisons to the brain are always dangerous.

At Meta, LeCun had resources but not freedom. FAIR did interesting work, but it was still tethered to a company that needed to ship products. AMI Labs is his chance to go all-in on the research direction he thinks matters. Twelve researchers, no product roadmap, a billion dollars of runway to prove or disprove a thesis about the nature of intelligence.

I find myself genuinely torn on this. The transformer skeptic in me thinks LeCun might be onto something real. We've all watched GPT-scale models plateau in certain ways that more data and more FLOPS haven't fixed. But the pragmatist in me looks at a 12-person company valued at $3.5 billion and feels a little dizzy.

The money is getting absurd

AMI Labs isn't even the wildest number from this quarter. Project Prometheus raised $6.2 billion. Humans& hit a $4 billion-plus valuation. Ricursive AI, same territory. We're in a period where AI companies are raising money at valuations that would have been considered delusional two years ago.

Some of these bets will pay off. Most won't. That's how it always works with technology waves. But the sheer volume of capital is hard to ignore. Investors are terrified of missing the next Anthropic or OpenAI, and that fear is driving check sizes that have no relationship to revenue, product, or even a plausible path to either.

Here's what I keep coming back to: the companies raising the most money right now fall into two categories. There's the "scale harder" crowd, building bigger models with more compute. And then there's the "rethink everything" crowd, which is where LeCun sits. Both camps are attracting billions. They can't both be right about the path, but they can both be right that AI matters. The question is timing and approach.

Is this a bubble? Wrong question.

I'm tired of the bubble discourse, honestly. Yes, valuations are disconnected from fundamentals. Yes, most of these companies will fail. Yes, there are echoes of 1999. But calling it a bubble doesn't tell you anything useful about what to do or what to watch.

The dot-com bubble was real. Amazon still exists. Google went public after it popped. The bubble was real AND the technology was real. Both things were true simultaneously.

What I'd rather ask: which of these companies are building something that will matter in five years? LeCun has credentials that few can match. He co-invented convolutional neural networks. He was doing deep learning before deep learning was cool. If anyone has earned the right to take a contrarian bet on AI architecture, it's probably him.

But credentials don't guarantee outcomes. Plenty of brilliant researchers have been wrong about the future of their field. And $3.5 billion is a lot of pressure on an unproven thesis, no matter who's behind it.

What I'm actually watching

Forget the valuation headlines. Here's what will tell us whether AMI Labs is for real:

Who are the 12 people? At that funding level with that team size, every hire is a statement about research direction. If LeCun is pulling top people away from DeepMind, FAIR, and Brain, that says something about how seriously the field takes his vision.

What do they publish? LeCun has always been open about his research. If AMI Labs goes quiet, that's a red flag. If they start dropping papers that make other researchers rethink their assumptions, that's the signal.

How long before the investors want results? A billion dollars buys a lot of patience, but not infinite patience. The tension between "fundamental research" and "return on investment" kills companies. LeCun knows this, which is presumably why he raised so much at once.

The real story

The real story isn't the money. It's that one of the most respected minds in AI looked at the current trajectory of the field and said "no." That's worth paying attention to regardless of whether you think the valuation makes sense.

LeCun could be wrong. Transformers might keep scaling. The next GPT might actually reason. But if he's right, and the current paradigm does hit a wall, then the company that already spent five years exploring alternatives will have a significant head start.

Twelve people. Three and a half billion dollars. One very specific bet against the consensus.

I don't know if it'll work. Nobody does. But I'll be watching.