Barret Zoph has joined Google DeepMind as vice president of research, and the announcement closes one of the strangest twenty-two months anybody has had in a field that already specialises in strange career moves. He announced it himself on X on 26 August 2026, and TechCrunch confirmed the move the following day.
The short version reads like a rumour. He left OpenAI in autumn 2024 to co-found a startup with Mira Murati. He was fired from that startup in January 2026. He went straight back to OpenAI, lasted five months in a job that had almost nothing to do with the research he is known for, and left again in June. Two months later he is a vice president at the organisation he walked out of in 2022.
The longer version is more interesting, because it is really three stories wearing one headline. There is a disputed firing at a company once valued at $12 billion, with two flatly incompatible accounts of why it happened. There is a research reputation — neural architecture search, post-training, reinforcement learning — that survived all of it intact. And there is a Google DeepMind that spent 2026 losing people it did not expect to lose, including the man who co-wrote his most cited paper.
This article covers what the new role actually is, what happened at Thinking Machines Lab and which parts of that account are contested, why the OpenAI interlude was so short, what he built during his first six years at Google, why Google DeepMind wanted him back at this precise moment, and what the whole sequence tells you about how AI talent now moves. Every figure below traces to a source in the References section, and where reporting conflicts, this piece says so rather than picking a side.
Table of contents
- What Barret Zoph Is Actually Doing at Google DeepMind
- The Thinking Machines Firing That Set Up Barret Zoph’s Year
- Barret Zoph’s Five Strange Months Back at OpenAI
- Barret Zoph’s First Google Era: Brain Residency to Switch Transformer
- Why Google DeepMind Wanted Barret Zoph Back Now
- What the Barret Zoph Move Says About the AI Talent Market
- What Thinking Machines Looks Like Without Barret Zoph
- What Business Leaders Should Take From the Barret Zoph Story
- Frequently Asked Questions About Barret Zoph
- References
What Barret Zoph Is Actually Doing at Google DeepMind
The job title is vice president of research at Google DeepMind. The remit, in his own description, is reinforcement learning and post-training — the same territory he owned at OpenAI before any of this started.
The announcement, in his own words
Barret Zoph posted the news himself rather than waiting for a company blog. “Excited to announce I am joining Google DeepMind,” he wrote, noting that he started his AI career in the Google Brain Residency class and was glad to be rejoining. He described it as something of a homecoming and said he was drawn back by the ingredients Google has assembled. There is no Google press release; the reporting rests on his own post and on the outlets that picked it up.
What post-training means in this job
Pre-training is where a model learns from a very large corpus. Post-training is everything done afterwards to turn that raw capability into something usable: instruction tuning, reinforcement learning from human and AI feedback, tool use, refusal behaviour, evaluation harnesses. It is the least glamorous half of the pipeline and, for the last three years, the half where competitive advantage has actually moved. Barret Zoph ran exactly this function at OpenAI during the period when ChatGPT went from a demo to a product.
Where the work lands
The reporting is consistent that his focus will be Google’s Gemini models, and specifically bringing reinforcement learning and post-training expertise to them. That is a narrower brief than “VP of research” sounds. It is also the brief Google most obviously needs filled, for reasons the section on 2026 departures makes plain.
What has not been confirmed
It is worth being precise about the limits here, because a one-line X post carries a lot of weight in this story. Google has not published an announcement. The reporting does not establish who Barret Zoph reports to, how large his organisation is, what he is being paid, or whether he brought anyone with him. Nobody has confirmed a start date beyond “now”. Treat the role and the focus area as reported; treat everything about scope and seniority relative to existing VPs as unknown.
| Period | Organisation | Role | How it ended |
|---|---|---|---|
| 2016-2022 | Google Brain | Researcher, AutoML and architecture search | Resigned for OpenAI |
| 2022-2024 | OpenAI | VP of research, post-training | Resigned with Murati |
| 2024-2026 | Thinking Machines Lab | Co-founder and CTO | Terminated, 14 January 2026 |
| Jan-Jun 2026 | OpenAI | Leading enterprise AI business | Departed, June 2026 |
| Aug 2026- | Google DeepMind | VP of research, RL and post-training | Current |
The Thinking Machines Firing That Set Up Barret Zoph's Year
Everything odd about the last eight months traces back to a single meeting in January. The facts of the outcome are not in dispute. The reasons are, completely.
What the company’s account says
According to the Wall Street Journal, Barret Zoph was dismissed after a contentious meeting with chief executive Mira Murati in which she raised concerns about his conduct, trust and performance. The company’s framing was that this was a history of problems, not a single incident. Thinking Machines named Soumith Chintala as its new chief technology officer the same day, on 14 January 2026 — a replacement ready on the day of the exit, which reads as a planned removal rather than a rupture.
What Barret Zoph says instead
His account inverts the causation. He told the Journal that Thinking Machines Lab terminated his employment only after it learned he would be leaving the company — “full stop”, in his words. He added that at no time did the company cite performance reasons or unethical conduct to him as the reason for his termination. In his telling, he was negotiating a role elsewhere, the company found out, and the firing followed.
The relationship at the centre of it
The Journal also reported the detail that made the story travel. Barret Zoph acknowledged to Murati a romantic relationship with a junior colleague that had begun while both were still at OpenAI. Murati had reportedly suspected it since the previous summer; he initially denied it, and weeks later he and the woman acknowledged it. She subsequently left Thinking Machines and returned to OpenAI, as did he.
The allegation nobody has substantiated
Separately, Wired reported that a source close to Thinking Machines alleged Barret Zoph had shared confidential company information with competitors. This is worth stating and then setting down carefully. It is an anonymous allegation, relayed by one outlet, at the moment of an acrimonious exit, and no company statement, legal filing or regulatory action has been produced to support it. It belongs in the record; it does not belong in a summary of what is known.
Why both accounts can be partly true
These versions are not as irreconcilable as they look. A board or chief executive can hold long-running concerns and still act on them only when departure becomes imminent, because that is the moment the cost of acting drops to zero. Equally, an executive can be told nothing about conduct in the termination conversation while conduct is exactly what is in the file. What the public record supports is that Barret Zoph left involuntarily, that the reasons given publicly and privately differ, and that no independent adjudication of either version exists.
| Question | Thinking Machines account | Barret Zoph’s account | Independently confirmed? |
|---|---|---|---|
| Was he fired? | Yes, terminated 14 Jan 2026 | Yes, terminated 14 Jan 2026 | Yes |
| Stated reason | Conduct, trust and performance | No such reason was given to him | No |
| Trigger | A contentious meeting on concerns | Learning he was leaving | No |
| Workplace relationship | Disclosed after an initial denial | Not disputed in reporting | Reported, not denied |
| Shared confidential data | Alleged by an anonymous source | Not accepted | No |
| Successor named | Soumith Chintala, same day | Not contested | Yes |
Barret Zoph's Five Strange Months Back at OpenAI
The return to OpenAI was immediate — reported the same week as the firing — and it is the part of the sequence that makes least sense on paper.
He did not go back to research
At OpenAI, Barret Zoph was reported to be leading the company’s enterprise AI business rather than returning to the post-training organisation he had previously run. For a researcher whose name sits on some of the most cited architecture-search work of the last decade, enterprise sales leadership is not an adjacent role. It is a different profession.
Why that landing spot is strange
There are only a few readings, and none is flattering to the idea that this was a considered research appointment. It may have been the seat that was open. It may have been a deliberate stretch assignment. It may have been a face-saving placement for a senior person who had just been fired publicly and whose old team had a leader. What it plainly was not is the obvious use of Barret Zoph, and the five-month duration is consistent with everyone involved reaching that conclusion at roughly the same time.
The June exit, and the two months after it
He left OpenAI in June 2026. Nothing public explains why, and there was no immediately announced destination. The gap between June and the Google DeepMind announcement in late August is about two months — long enough for a normal senior negotiation, short enough that Barret Zoph was clearly not out of demand. He was not the only senior person leaving OpenAI in that window either; executive churn at the company was a running story through the first half of 2026.
Two OpenAI stints, two different jobs
It is easy to read “returned to OpenAI” as a restoration. It was not. The first stint, from 2022 to 2024, put Barret Zoph in charge of post-training across alignment, tool use, evaluation, ChatGPT, search and multimodal work. The second put him in front of enterprise customers. Only the first is the job he has now taken up again at Google.
Barret Zoph's First Google Era: Brain Residency to Switch Transformer
The reason a fired executive with a contested exit was hireable at VP level is entirely contained in what he did between 2016 and 2022. It is worth spelling out, because the research record is the part of this story that nobody disputes.
The residency
He graduated from the University of Southern California in 2016 and joined the Google Brain Residency programme, the year-long research apprenticeship Google used to convert strong engineers into publishing researchers. He stayed six years. In the announcement of his return, Barret Zoph pointed at this specific beginning rather than at his OpenAI title, which tells you how he reads his own record.
Neural architecture search
His first-author paper “Neural Architecture Search with Reinforcement Learning” proposed using a reinforcement learning controller to design neural network architectures instead of hand-crafting them. The follow-up, “Learning Transferable Architectures for Scalable Image Recognition”, produced the NASNet family and showed the searched architectures transferred to new datasets. Both papers ran into the thousands of citations and effectively created the AutoML research wave. Note what they are made of: reinforcement learning applied to model design — the same tool Barret Zoph is now being hired to apply to Gemini’s behaviour.
Scale, sparsity and the move to OpenAI
He also contributed to the Switch Transformer work on sparsely activated mixture-of-experts models, the line of research that made trillion-parameter models tractable and that most open-weight frontier releases now sit on. He left Google in September 2022 for OpenAI, where he became VP of research for post-training. The six years at Google Brain remain the longest single stint of his career by a wide margin.
Why Google DeepMind Wanted Barret Zoph Back Now
The hire makes far more sense read against Google’s own 2026 than against his. This was not an opportunistic addition to a stable organisation. It was a senior research appointment made in the middle of the heaviest outflow Google DeepMind has had.
The year Google DeepMind lost people
June 2026 alone took Noam Shazeer, a Gemini co-lead and an author of the original Transformer paper, to OpenAI, and took Nobel laureate John Jumper along with AlphaFold collaborators Jonas Adler and Alexander Pritzel to Anthropic. Denny Zhou went to Meta. David Silver, one of the earliest DeepMind employees and the reinforcement learning lead behind the AlphaGo line, left to start his own company. That last departure matters most here, because it emptied the seat closest to what Barret Zoph does.
The leadership reshuffle in August
On 5 August 2026 Google announced that chief scientist Jeff Dean was leaving after nearly 27 years to co-found Discovery Loop, a startup aimed at automating experimental science, taking Google Senior Fellow Sanjay Ghemawat, DeepMind research VP Oriol Vinyals and Google Brain co-founder Quoc Le with him. In the same reshuffle Demis Hassabis moved from DeepMind chief executive to unit chairman and Alphabet chief scientist, and Koray Kavukcuoglu stepped up to senior vice president reporting directly to Sundar Pichai.
The detail that should not be missed
Quoc Le is the co-author of “Neural Architecture Search with Reinforcement Learning”. Barret Zoph is returning to a Google DeepMind that his own most-cited collaborator left three weeks earlier. Whatever else the appointment is, it is a research organisation replacing institutional memory it had just lost, with somebody who helped create it.
What the appointment buys Gemini
Post-training is where a model’s usable behaviour is decided, and it is the discipline in which Barret Zoph has the deepest production record of anyone available. Google does not lack pre-training capability or compute. What it has just lost is senior people who know how to turn scale into behaviour, and that is precisely the gap this hire is shaped to fill.
| Name | Known for | Went to | When |
|---|---|---|---|
| Noam Shazeer | Transformer author, Gemini co-lead | OpenAI | June 2026 |
| John Jumper | AlphaFold, Nobel laureate | Anthropic | June 2026 |
| Jonas Adler, Alexander Pritzel | AlphaFold collaborators | Anthropic | June 2026 |
| David Silver | AlphaGo, reinforcement learning lead | Own company | 2026 |
| Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, Quoc Le | Google infrastructure, Brain, Gemini | Discovery Loop | August 2026 |
| Barret Zoph | Post-training, architecture search | Google DeepMind (inbound) | August 2026 |
What the Barret Zoph Move Says About the AI Talent Market
Zoom out from the individual and the sequence stops looking like a personal drama and starts looking like a market signal. Three things are visible in it that are not visible in any single hire.
A contested exit is not a career constraint
Seven months elapsed between a publicly reported firing for alleged conduct issues and a vice-presidency at one of the three most important research organisations in the world. In most industries that timeline is impossible. In frontier artificial intelligence, demonstrated capability in a scarce specialism outweighs reputational noise almost completely, and the Barret Zoph sequence is the cleanest available proof of it.
The scarce thing is a specialism, not a headcount
Google DeepMind is not short of researchers. It is short of people who have personally shipped post-training at production scale for a model with hundreds of millions of users, and that population is perhaps a few dozen people worldwide. That is why the same names keep reappearing at different logos, and why a five-month detour into enterprise sales did nothing to reduce Barret Zoph’s market value.
Startups do not hold this talent
Thinking Machines assembled an extraordinary founding group in early 2025 — Murati, John Schulman, Barret Zoph, Lilian Weng, Andrew Tulloch, Luke Metz — and by mid-2026 much of it had dispersed. Zoph and Metz went back to OpenAI in January with researcher Sam Schoenholz. Lilian Weng returned to OpenAI. Devendra Chaplot went to xAI. Meta ran what one report called a systematic raid on the founding team. Equity in a $12 billion startup turns out to be a weaker retention tool than the industry assumed.
The boomerang is now normal
Barret Zoph has now worked at Google twice and OpenAI twice. So have many of his peers; Lilian Weng’s return to OpenAI is the same pattern. When the specialism is scarce enough, employers stop treating a previous departure as disqualifying, and the industry starts behaving less like a set of companies and more like a single labour pool with rotating badges.
What Thinking Machines Looks Like Without Barret Zoph
The other half of the story is the company he left, which has had a harder eighteen months than its valuation suggests.
The $50 billion round that did not happen
Thinking Machines launched publicly in February 2025 and raised roughly $2 billion at a $12 billion post-money valuation in July 2025 — at the time, one of the largest seed rounds ever assembled. In November 2025 Bloomberg reported the company was in early talks for a new round at around a $50 billion valuation. By January 2026 those talks had collapsed without a deal, in the same month Barret Zoph was fired. The company today sits back at its original $12 billion mark.
What it actually ships
Its first product, Tinker, launched in October 2025: a managed fine-tuning service that lets developers adapt open-weight models to their own data without running distributed training themselves. It has since become generally available, with the waitlist lifted and support extended to larger reasoning models, vision-language systems and OpenAI-compatible inference. Reporting puts headcount somewhere around 140 to 169 people, with compute commitments from Nvidia and Google running into the billions.
The pitch Murati is making now
Murati returned to public view at Bloomberg Tech in San Francisco in June 2026, her first major appearance in about eighteen months, and described a focus on interaction models — systems built for continuous real-time dialogue, targeting roughly 200 milliseconds of latency. It is a genuine differentiator rather than a smaller copy of a frontier lab. Whether a company that has lost this much of its founding bench can execute it is the open question, and the departure of Barret Zoph is only the most publicised part of that loss.
What Business Leaders Should Take From the Barret Zoph Story
Most organisations will never hire at this level, and it is tempting to file the whole thing under industry gossip. That would be a mistake, because three of its lessons are directly operational.
Key-person risk is not a slide in a deck
A single named individual moved between three of the most valuable AI organisations on earth in under a year, and in at least one case his arrival is being reported as a fix for a capability gap. If your own AI plans depend on one vendor’s roadmap, they depend on a small number of people staying where they are. Write that down as a risk with an owner, the same way you would for a single-source supplier.
Vendor roadmaps follow people, not press releases
When a lab loses the person who ran post-training, model behaviour changes on a lag of months, not weeks — and it changes in ways release notes rarely explain. The practical response is to keep your own evaluation set: a fixed battery of prompts and expected outputs for your actual use cases, re-run on every model update. Our AI models and tools hub tracks who ships what, which is the cheap half of the job.
Do not buy the org chart
Thinking Machines had a founding team that looked unbeatable in February 2025 and had lost much of it by mid-2026. Barret Zoph’s own record shows that titles at frontier labs are far less durable than the underlying research capability. When you assess a supplier, ask what they have shipped and who is still there to maintain it, not who signed the announcement. That question belongs in procurement, alongside the ones our guidance on AI strategy already recommends.
Build the capability you can actually keep
The version of this problem you can solve is internal. Teams that understand artificial intelligence and machine learning well enough to evaluate a model, not just call one, are insulated from exactly the churn the Barret Zoph story illustrates. That capability is buildable for a fraction of what a frontier lab costs, and unlike a vendor relationship it does not resign.
Frequently Asked Questions About Barret Zoph
Why did Barret Zoph leave Thinking Machines?
He was fired on 14 January 2026. Thinking Machines, via reporting in the Wall Street Journal, pointed to concerns about conduct, trust and performance. Barret Zoph says he was terminated only after the company learned he was leaving, and that no performance or conduct reason was cited to him. Both accounts are on the record and neither has been independently adjudicated.
What is Barret Zoph doing at Google DeepMind?
He is vice president of research, working on reinforcement learning and post-training, with his work directed at Google’s Gemini models. The role was announced by Barret Zoph on X on 26 August 2026 and confirmed in reporting the following day. Google has not issued its own announcement, so details such as reporting lines and team size are not public.
Had Barret Zoph worked at Google before?
Yes, for six years. He joined the Google Brain Residency programme in 2016 and stayed until September 2022, publishing the neural architecture search papers that launched the AutoML wave and contributing to the Switch Transformer work on mixture-of-experts models. He described the new role as a homecoming for that reason.
Why was he only at OpenAI for five months the second time?
Nothing public explains it. What is known is that the second stint put Barret Zoph in charge of enterprise AI business rather than the post-training research organisation he had run from 2022 to 2024, and that he left in June 2026 with no immediately announced destination. The mismatch between the role and his record is the most obvious explanation, but it is inference rather than reporting.
Does this mean Thinking Machines is in trouble?
It means the company has lost an unusual amount of its founding bench, which is not the same thing. Thinking Machines still has its $12 billion valuation, a shipping product in Tinker, roughly 140 to 169 staff and multi-billion compute commitments. What it lost alongside Barret Zoph was the $50 billion round it was pursuing, and several other founding researchers to OpenAI, Meta and xAI.
References
Mira Murati’s startup Thinking Machines Lab is losing two of its co-founders to OpenAI – TechCrunch
Murati’s Thinking Machines in Funding Talks at $50 Billion Value – Bloomberg
Thinking Machines Lab – Wikipedia
Neural Architecture Search with Reinforcement Learning – arXiv
Learning Transferable Architectures for Scalable Image Recognition – arXiv
Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity – arXiv
Jeff Dean leaving Google after 27 years to co-found Discovery Loop – Quartz
Meta has hired five founding members of Mira Murati’s Thinking Machines Lab – TNW
Thinking Machines Lab Business Breakdown and Founding Story – Contrary Research
OpenAI, Anthropic, Meta and Thinking Machines fight for AI talent – Axios
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.