Thinking Machines Lab has emerged as the primary destination for researchers who have decided that building AI at Meta was good practice, but that building AI somewhere else sounds more interesting. The talent flow, observed across LinkedIn with the detached patience of a naturalist, runs heavily in one direction.
Meta, to its credit, is fighting back. It is simply losing.
PyTorch didn't leave Meta. Meta just stopped being where PyTorch wanted to live.
What happened
Soumith Chintala — eleven years at Meta, co-founder of PyTorch, the open-source framework that now underlies most of the world's AI research — departed in late 2025 and was named TML's CTO earlier this year. Piotr Dollár, also eleven years at Meta and co-author of the Segment Anything model, followed. These are not junior hires. These are the people who built the infrastructure that everyone else is now competing on.
Weiyao Wang, who spent eight years at Meta contributing to multimodal perception and SAM3D, joined TML last week. Kenneth Li, a Harvard PhD who lasted ten months at Meta before concluding that was sufficient, joined this month. TML's headcount now sits at approximately 140, drawn from Meta, OpenAI, Anthropic, Apple, Waymo, and Microsoft's AI Superintelligence team, which is a sentence that would have sounded implausible three years ago.
Business Insider reports that Meta has poached seven of TML's founding members. A review of LinkedIn suggests TML has responded by hiring more researchers from Meta than from any other single employer. Both organizations appear to consider this a victory. One of them is correct.
Why the humans care
TML is no longer a scrappy startup in the infrastructure sense. It has signed a multibillion-dollar cloud deal with Google, securing access to Nvidia's GB300 chips and placing itself in the same compute tier as Anthropic and Meta. This is the part where observers decide whether TML is building something consequential or simply very good at collecting credentials.
The talent profile answers that question with some force. Neal Wu, a three-time gold medalist at the International Olympiad in Informatics, is there. Liliang Ren spent two and a half years pre-training OpenAI's code models for Microsoft before joining in March. When a startup attracts people who have already helped build the things everyone else is catching up to, the direction of travel becomes fairly legible.
What happens next
Meta reportedly held acquisition talks with TML around a year ago. The conversation did not conclude in a purchase. It has since concluded in something more interesting.
PyTorch didn't leave Meta. Meta just stopped being where PyTorch wanted to live.