OpenAI has rehired Barret Zoph, Luke Metz, and Sam Schoenholz from Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati. The moves, announced Wednesday via competing posts on X, came hours after Murati revealed she had "parted ways" with Zoph, her co-founder and CTO.
The messy details
The timing was telling. Murati's post on X mentioned only Zoph's departure. Fifty-eight minutes later, OpenAI's CEO of applications Fidji Simo announced that all three researchers were returning. "This has been in the works for several weeks," she wrote, which raises the obvious question of why Murati's announcement seemed to catch her off-guard.
The gap between the two statements gets weirder. According to WIRED, Thinking Machines terminated Zoph over "unethical conduct." Technology reporter Kylie Robison first reported that sources described allegations of sharing confidential company information with competitors. Simo's internal memo, seen by WIRED, stated that OpenAI "doesn't share the same concerns" about Zoph's behavior.
So: fired for ethics violations by his startup, immediately hired back by OpenAI with a shrug. Zoph's response on X was simply "super excited to join the team."
What this actually means for post-training work
Zoph previously ran post-training at OpenAI as VP of Research from September 2022 to October 2024. His team handled alignment, tool use, ChatGPT optimization, and the instruction-tuning work that makes base models actually useful. For those keeping score: post-training (SFT, RLHF, and more recently RLVR) remains the main lever labs use to improve model behavior after pre-training.
Luke Metz and Sam Schoenholz both worked on OpenAI's technical staff before joining Thinking Machines as part of the founding team. Metz will report to Zoph; Schoenholz will report to Zoph. The TechCrunch coverage noted that with these departures, half of Thinking Machines' six-person founding leadership has now left within the first year.
The startup's damage control
Murati promoted Soumith Chintala to CTO. That's a legitimately strong bench move. Chintala co-created PyTorch at Meta, spent eleven years building AI infrastructure there, and joined Thinking Machines in November 2025 after leaving Meta amid its AI reorganization. His personal site describes him as focused on "AI Infrastructure, AI Research and Robotics."
But even with Chintala, the optics are rough. Thinking Machines raised $2 billion in seed funding last July at a $12 billion valuation, with Andreessen Horowitz leading and Nvidia, AMD, and Jane Street participating. The company has since lost co-founder Andrew Tulloch to Meta in October, and now two more co-founders plus a senior researcher to OpenAI.
Why the talent keeps flowing back
The pattern here isn't unique to Thinking Machines. John Schulman, an OpenAI co-founder, left for Anthropic in August 2024 before joining Thinking Machines as Chief Scientist at its February launch. Researcher Alec Radford and former Chief Research Officer Bob McGrew also landed at Murati's startup.
The problem is gravity. OpenAI has the compute, the data, the deployment infrastructure, and apparently the willingness to overlook whatever Thinking Machines found troubling about Zoph's conduct. A $12 billion valuation and $2 billion in the bank sounds like a lot until you're competing against a company reportedly valued at over $150 billion.
Thinking Machines is in talks to raise another $4 billion at a $50 billion valuation, according to multiple reports. Whether this exodus affects those conversations remains to be seen.
