Analysis
Barret Zoph, who co-founded Thinking Machines Lab alongside former OpenAI CTO Mira Murati, has landed at Google DeepMind as vice president of research, TechCrunch reported. A Google spokesperson said the company is glad to have Zoph 'bringing his RL and post-training expertise to Gemini' -- a specific technical mandate, not a vague executive hire.
Zoph's path here has been unusually volatile even by frontier-lab standards. He spent two years at OpenAI before leaving in October 2024 to co-found Thinking Machines with Murati, positioning himself as CTO of what was, at launch, one of the most hyped new labs in the industry, reportedly valued near $50 billion in its earliest funding conversations. In January, Zoph and fellow co-founder Luke Metz left Thinking Machines to return to OpenAI, where Zoph was assigned to lead AI enterprise sales -- a notable pivot from research leadership to a go-to-market role. He lasted roughly five months in that stint before leaving OpenAI again in June, and has now surfaced at Google.
Three employer changes in under two years, for a co-founder who was CTO of a company his own departure helped destabilize, is the kind of churn that would be disqualifying in most industries but reads as normal turbulence in the current AI labor market, where compensation packages and equity upside move fast enough to make loyalty economically irrational for in-demand researchers.
“Google, Meta and OpenAI have all been active in re-recruiting talent that cycled through smaller, newer labs over the past 18 months.”
Thinking Machines itself has lost multiple founding team members since its 2024 launch, a pattern that raises real questions about the durability of star-studded lab launches generally -- assembling a marquee founding team is not the same as retaining it once competing labs can outbid on comp or offer a more stable mission. Google, Meta and OpenAI have all been active in re-recruiting talent that cycled through smaller, newer labs over the past 18 months. Pulse has previously covered Google DeepMind's research and product moves as the lab has ramped up its own reasoning-model efforts.
For Google specifically, landing Zoph is a direct reinforcement-learning and post-training hire at a moment when RL-based fine-tuning has become the primary lever labs use to differentiate model quality post-pretraining -- the area where OpenAI's o-series and Anthropic's extended-thinking models have made their most visible gains. Google recruiting specifically for this expertise signals where it sees its own competitive gap against OpenAI and Anthropic on reasoning-model quality.
What to watch is whether Thinking Machines' remaining leadership can stabilize the founding team, or whether Zoph's exit is one more data point in a broader unwind of the company's original pitch to investors.
