The résumé has been through more than most in the AI industry: OpenAI, where he worked on frontier models; the founding of Thinking Machines Lab with Mira Murati; and now Google DeepMind. The Wall Street Journal reported that Barret Zoph has joined DeepMind as vice president of research, responsible for post-training and reinforcement learning, two of the most competitive areas in artificial intelligence.
The move is notable for the direction of travel. For two years, the flow of top AI talent has run from Google toward startups and rivals, and DeepMind, Google’s crown jewel, has been the biggest source of departures. Fortune reported the same day that DeepMind’s most senior researchers are still leaving at a steady clip. Zoph’s arrival is a rare instance of the current reversing, a researcher with a marquee background choosing to build inside Google rather than outside it.
Zoph’s history explains the significance. He spent years at OpenAI, where he worked on the post-training techniques that turned large models into useful assistants, the layer of engineering that determines how well a model follows instructions, reasons through problems, and behaves under pressure. He left to co-found Thinking Machines Lab with Murati, the former OpenAI chief technology officer, and the startup became one of the most closely watched AI companies of its generation, an attempt to build frontier models with a different organizational model and a different safety posture.
That chapter ended in August, when Zoph parted ways with the company’s chief executive. The details of the falling out have been reported piecemeal, and neither side has explained it fully, but the practical consequence was clear: Zoph was available, and the market for his skills is the most competitive in technology. DeepMind’s offer, a vice presidency focused on exactly the field he has spent his career developing, is the kind of role that lets a researcher keep working at the frontier without the fundraising and organizational duties of running a company.
DeepMind’s timing is strategic. The laboratory has been rebuilding its post-training organization as the field has shifted from pre-training scale to the techniques that come after: reinforcement learning, reasoning chains, and the fine-tuning that turns a general model into a reliable agent. Google’s Gemini models compete with OpenAI’s and Anthropic’s, and post-training is where much of the race is now decided. Bringing in an outsider who helped build OpenAI’s approach gives DeepMind both expertise and a signal that it is willing to pay for the best available talent.
The hire also fits a pattern of consolidation at the top of the AI field. As the number of frontier laboratories shrinks, talent is flowing back toward the largest companies, which can offer resources, compute, and stability that startups cannot match. Thinking Machines, for all its promise, faces the same pressures as every challenger: raising capital, securing compute, and retaining researchers who can command premium packages elsewhere. Zoph’s decision to join DeepMind rather than raise another round is a data point about where the power in the industry now sits.
The technical stakes are easy to understate. Post-training is the set of techniques applied after a model’s initial learning phase: reinforcement learning from human feedback, reward modeling, and the fine-tuning that determines whether a model gives a useful answer or a generic one. In the past two years it has become the field’s sharpest competitive edge, because the base architectures of the leading models have converged while the methods used to steer them have not. DeepMind has invested heavily in this layer, and Zoph’s brief, owning it across the laboratory’s flagship programs, puts him at the center of the company’s most important technical decisions.
The move also reads as a response to a pattern that has worried Google’s leadership. DeepMind’s founding researchers have drifted toward startups and independent laboratories, where equity packages and autonomy have proved more attractive than the scale of a public company. The departures have not stopped the lab from producing leading models, but they have changed its character, and recruiting from outside has become a priority. Zoph is the most senior outside hire the laboratory has made in years, and his willingness to take the job was likely helped by the resources only Google can offer: compute at a scale few startups can match, and a research organization with deep benches in every relevant discipline.
For Thinking Machines, the departure removes a co-founder at a critical moment. The company raised its last round with a prominent list of backers and has been building toward its first frontier models, a goal that requires the exact expertise Zoph represents. Its chief executive, Murati, has not commented on the split beyond brief statements, and the company has said work continues. In the broader talent market, the episode is another illustration of the industry’s new arithmetic: the number of people capable of shaping frontier models is small, the companies competing for them are few, and the packages being offered are large enough to break apart founding teams. Zoph’s decision suggests that, at least for now, the largest checks are being written by the incumbents.
For Google, the hire is a message to its own employees as much as to the market. The company has responded to the brain drain with retention packages and promotions, but it has rarely been able to attract senior outsiders in the AI field, where the culture has favored founding teams. Zoph’s arrival gives DeepMind a new senior voice in the post-training debate and a public answer to the question of whether anyone at Google is still building the future. The coming quarters will show whether the hire translates into model leadership, or whether one researcher, however accomplished, is a drop in a stream still flowing the other way.
