Andrew Brown, a Meta research scientist who co-authored the company's Emu Video paper and contributed to its 30-billion-parameter Movie Gen system, has joined Yann LeCun's AMI Labs to work on world models. Brown announced the move on his X account this week, capping a roughly three-year run at Meta's GenAI team.
The pipeline from Menlo Park to Paris
Brown's exit follows a now-familiar route. LeCun walked out of Meta at the end of 2025 after more than a decade running its FAIR lab, taking longtime collaborator Alexandre LeBrun with him as CEO. Saining Xie, the NYU professor and former FAIR researcher who co-authored the DiT paper that underpins much of modern video and image generation, signed on as Chief Science Officer. Laurent Solly, Meta's former VP for Europe, became COO.
Then came the money. AMI closed a $1.03 billion seed round in March at a $3.5 billion valuation, a number that makes the previous European seed record look quaint. Brown is the kind of hire that money buys.
A weird fit, maybe
Here's the part worth thinking about. Brown's most-cited work is on transformer-based video generation, the token-prediction-adjacent paradigm LeCun has spent years publicly calling a dead end on the road to anything resembling general intelligence. Movie Gen is, at its core, a 30B-parameter transformer trained to predict pixels. The Stanford CS25 lecture Brown gave last year was literally titled "Transformers for Video Generation."
So why hire its co-author for a company explicitly built around the opposite bet? Probably because world models, whatever they turn out to be architecturally, still need to predict how visual scenes unfold over time. Video generation researchers have spent the past three years figuring out how to do that at scale. Whether LeCun's team plans to throw out the transformer playbook or quietly absorb pieces of it is the question AMI hasn't really answered.
"World models that learn abstract representations of reality, similar to the mental models humans use to reason and guide action." That's how AMI describes the goal in its launch materials. It sounds nice. It also describes roughly what every generative video model claims to be doing, depending on who's holding the slide deck.
What's missing
The thing AMI has not produced yet: a paper, a model, a demo, anything. The company has produced a CEO, a CSO, a COO, a billion dollars, a French headquarters, a partnership with healthtech firm Nabla, and now a steady drip of senior hires. Brown joins that list.
LeCun told MIT Technology Review in January that Meta might be AMI's first client, which is one way to leave a company without fully leaving it. The next thing to watch is the first technical release. Until then, the strongest evidence for AMI's approach is the talent it can pull, and on that front the company keeps winning.
