Tech
Meta Trained Them. Thinking Machines Lab Hired Them.
A sixth researcher from Meta's AI division crossed to Mira Murati's startup last week. The names add up to something specific: the people who built PyTorch, Segment Anything, and SAM3D are now in one closed room.
Weiyao Wang's final day at Meta came the week of April 21. Eight years in, he had shipped multimodal perception systems and co-authored SAM3D, a November 2025 extension of Meta's Segment Anything model into three-dimensional space. His next stop: Thinking Machines Lab, the fourteen-month-old startup founded by former OpenAI CTO Mira Murati.
He is the sixth researcher from Meta's AI research division to make that move in roughly twelve months. Piotr Dollar, an 11-year FAIR veteran who served as research director and co-authored Segment Anything itself, is already on TML's technical staff. So is Soumith Chintala, who co-founded PyTorch across eleven years at Meta before becoming TML's CTO. James Sun (nearly nine years, LLM pre- and post-training), Andrea Madotto (multimodal language models, FAIR, joined December 2025), and Kenneth Li (Harvard doctoral graduate, ten months at FAIR, arrived the same week as Wang) complete the roster. According to LinkedIn data cited by TechCrunch, TML hires more researchers from Meta than from any other employer.
This is not a random sample of departing talent. The cluster landing at TML is specifically the multimodal perception and 3D vision cluster: the people who designed the open-source tools that now underpin most serious computer vision work. Chintala built the framework. Dollar co-authored the model. Wang extended it into 3D. They are now building something together that they will not open-source.
Two days before Wang's move was announced, TML signed a multibillion-dollar deal with Google Cloud for Nvidia GB300 chips, announced April 22 at Google Cloud Next. The GB300 NVL72 rack ships with 37 terabytes of memory and 130 terabits per second of internal bandwidth, a configuration built for models too large to train on any conventional cluster. Founding researcher Myle Ott told Google's press team: "Google Cloud got us running at record speed with the reliability we demand." The deal puts TML's compute on the same infrastructure tier as Anthropic and Meta.
Meta built FAIR around a publication strategy. Publish PyTorch. Publish Segment Anything. Publish LLaMA. The theory was that publishing attracted researchers, and researchers built better things internally. That theory worked, until it didn't. The researchers Meta trained and put in front of the world now work under a closed door at a startup valued at $12 billion.
TML has released one product: Tinker, an API that uses reinforcement learning to automate custom frontier model creation. The founding engineer who built it, Joshua Gross, left for Meta Superintelligence Labs in March. The company has about 140 people. A PyTorch architect, a Segment Anything co-author, a SAM3D builder, and three more FAIR-lineage researchers, plus compute infrastructure now running on GB300 NVL72 racks: that is not the profile of a lab planning to wrap another company's model. It is the profile of a lab preparing to train a world model. The compute just arrived.