SAM's Co-Author Took His Next Job at Thinking Machines

Piotr Dollár joined Thinking Machines Lab this spring, ending an 11-year tenure at Meta FAIR that closed with Segment Anything Model on his CV.

His last public credit at Meta was SAM2, an August 2024 paper extending prompt-based segmentation to video. By April 24, when TechCrunch first documented the movement, Dollár already appeared on TML's technical staff; Meta has not publicly announced a departure date. Two others crossed in adjacent months: Soumith Chintala, PyTorch's co-creator, became TML's CTO in January; Weiyao Wang, who spent eight years on Meta's perception systems, departed April 17.

Thinking Machines Lab raised a $2 billion seed round at a $12 billion valuation and in April signed a multibillion-dollar Google Cloud agreement for Nvidia GB300 chips. On May 11, TML opened a research preview of TML-Interaction-Small, a full-duplex model that responds in 0.40 seconds, roughly the pace of natural conversation.

The flow ran both ways. Meta recruited five TML founding members: Andrew Tulloch at a reported $1.5 billion package over six years, and Joshua Gross, who had built TML's core API product. PyTorch's stewardship passed to a team Chintala named in his November 2025 departure post; he wrote that "the project didn't need me anymore."

The five-to-three headcount favors Meta, but the exchange implies a mismatch in what each company acquired. Meta's TML recruits were product builders: Tulloch built TML's API layer; Gross ran its engineering team. Chintala, Dollár, and Wang are architects of the infrastructure the broader ML industry runs on, and their design-level knowledge of where those systems break is the harder kind to source.

TML opens the interaction model to selected research partners in the coming months. That evaluation is the first test of whether the architecture Dollár and Chintala helped design outperforms the open-source baselines they left at Meta.