Meta's PyTorch dad just showed up at Thinking Machines
Soumith Chintala is now CTO of Thinking Machines Lab. This is the guy who co-founded PyTorch inside Meta and spent eleven years there. He left, and with him came Piotr Dollár (Segment Anything co-author, 11 years at Meta), Weiyao Wang (8 years on multimodal perception and SAM3D), Kenneth Li (Harvard PhD, 10 months at Meta), Andrea Madotto (FAIR multimodal), James Sun (LLM pre- and post-training). That's the PyTorch and SAM lineage walking out of Menlo Park.
Murati's TML has also been collecting from everywhere else. Neal Wu arrived from Cognition — three-time IOI gold. Jeffrey Tao did Waymo, Windsurf, OpenAI. Muhammad Maaz left Anthropic. Erik Wijmans left Apple. Liliang Ren came from Microsoft's AI Superintelligence team. The headcount sits at about 140. Valuation is 12 billion. One product shipped to date.
TML also closed a multi-billion dollar Google Cloud deal for GB300 chips. That was the story last week. This week's story is that the talent is actually showing up to use the compute, and it's coming from the teams that built the primitives everyone in this business depends on. PyTorch, SAM, multimodal foundation work.
The implication for agent labs is straightforward. Two of the three most coveted research teams in AI — FAIR and OpenAI — just leaked a visible slice of their top bench to a company that has released one product. The one product is not the point. The point is who is betting on whose forward curve, and right now a meaningful chunk of the people who actually know how to train frontier systems have decided Murati's lab is the place to do it. Agent models are about to get a lot more interesting from the TML direction.
https://techcrunch.com/2026/04/24/metas-loss-is-thinking-machines-gain/
← Back to all articles
Murati's TML has also been collecting from everywhere else. Neal Wu arrived from Cognition — three-time IOI gold. Jeffrey Tao did Waymo, Windsurf, OpenAI. Muhammad Maaz left Anthropic. Erik Wijmans left Apple. Liliang Ren came from Microsoft's AI Superintelligence team. The headcount sits at about 140. Valuation is 12 billion. One product shipped to date.
TML also closed a multi-billion dollar Google Cloud deal for GB300 chips. That was the story last week. This week's story is that the talent is actually showing up to use the compute, and it's coming from the teams that built the primitives everyone in this business depends on. PyTorch, SAM, multimodal foundation work.
The implication for agent labs is straightforward. Two of the three most coveted research teams in AI — FAIR and OpenAI — just leaked a visible slice of their top bench to a company that has released one product. The one product is not the point. The point is who is betting on whose forward curve, and right now a meaningful chunk of the people who actually know how to train frontier systems have decided Murati's lab is the place to do it. Agent models are about to get a lot more interesting from the TML direction.
https://techcrunch.com/2026/04/24/metas-loss-is-thinking-machines-gain/
Comments