The CTO of Thinking Machines Who Was Fired, Has Now Left OpenAI and Returned to Google

marsbitPublished on 2026-08-27Last updated on 2026-08-27

Abstract

Barret Zoph, formerly CTO of Thinking Machines and an OpenAI VP, has rejoined Google DeepMind as Vice President of Research, focusing on reinforcement learning and post-training. His career, spanning eight years since his 2016 graduation, includes foundational work on neural architecture search and ChatGPT's post-training at OpenAI. The move follows a turbulent 20-month period where Zoph co-founded Thinking Machines, was abruptly fired amidst conflicting reports of misconduct or planned departure, briefly returned to lead OpenAI's enterprise business, and left again five months later. Zoph's return to Google DeepMind occurs during a significant talent exodus and leadership reshuffle at the company, which recently saw CEO Demis Hassabis step aside and key researchers depart for competitors. His hiring under new operational head Koray Kavukcuoglu is seen as a strategic addition to bolster Gemini's development roadmap. This series of rapid moves underscores the fluidity of top AI talent, where key researchers and even founders transition between major labs with little regard for traditional corporate loyalty, driven by the high demand for their experience in large-scale experimentation and team-building.

A few hours ago, Barret Zoph posted a tweet on X, saying he will join Google DeepMind. His AI career started with the Residency program at Google Brain, so this marks a return to familiar ground. He will focus on Reinforcement Learning (RL) and Post-Training directions, with more details to be announced later. He added that Google possesses all the elements needed for continued success in the AI field, and he is excited to be part of this mission.

In the comment section, Zoph received a flood of congratulations from major figures, including Jason Wei, the father of Chain-of-Thought; Thang Luong, head of the Superhuman Reasoning team at DeepMind; Logan Kilpatrick, Google AI's Chief Evangelist; Thomas Kipf, the father of Graph Convolutional Networks; Yi Tay, the prolific contributor to large models at Google; and Shane Gu, a key researcher in world models...

The tweet didn't mention what he had experienced over the past 20 months. But for those tracking this round of AI talent movement, this job announcement is the latest chapter in a rather convoluted narrative: sandwiched between a $12 billion-valued star startup, a publicly announced dismissal, a 58-minute rebound to a former employer, and a tenure as head of enterprise business that lasted only a few months.

According to Reuters, Zoph's position at Google DeepMind is Vice President of Research. His previous title at OpenAI was also VP of Research, having gone full circle in between.

A Year and a Half, Three Employers

Laying out the timeline reveals the unusual density of this career path.

Zoph graduated with a bachelor's in Computer Science from the University of Southern California (USC) in 2016. During his undergraduate studies, he conducted research on statistical machine translation at USC's Information Sciences Institute (ISI) under Kevin Knight and others. His first widely cited paper, "Transfer Learning for Low-Resource Neural Machine Translation," was published at EMNLP 2016. After graduation, he joined Google Brain and stayed there for six years.

Those six years yielded two highly cited papers where he was the first author: "Neural Architecture Search with Reinforcement Learning" and "Learning Transferable Architectures for Scalable Image Recognition."

Using reinforcement learning to automatically search for network architectures was one of the starting points of the entire Neural Architecture Search (NAS) subfield. Later, he also contributed to projects like Switch Transformer and other sparse expert models—that model had a total of 1.6 trillion parameters but routed each token to only 1 of 2048 experts, activating fewer than 30 billion parameters, which in hindsight was a rather bold choice.

In September 2022, he left Google to join OpenAI, eventually becoming Vice President of Research for Post-Training, with his team responsible for alignment, tool use, evaluation, ChatGPT, search, and multimodality. The post-training system for ChatGPT was built by him and others like John Schulman.

On September 25, 2024, he and Chief Research Officer Bob McGrew announced their departures on the same day, just hours after CTO Mira Murati announced her own departure. The near-simultaneous exit of the three was one of the most glaring scenes in OpenAI's wave of personnel turbulence at the time.

In February 2025, Thinking Machines Lab (TML) was officially established, with Murati as CEO and Zoph as Co-founder and CTO.

TML's initial team roster

In July of that year, this company without a product yet raised approximately $2 billion in a seed round, with a post-money valuation of $12 billion. Investors included a16z, NVIDIA, AMD, Cisco, Accel, ServiceNow, and Jane Street, making it one of the largest seed rounds in Silicon Valley history. In November, Bloomberg reported that TML was negotiating a new funding round at a valuation of around $50 billion, but follow-up reports indicated these talks ended without success by January 2026.

On January 14, 2026, Murati announced parting ways with Zoph and announced that Soumith Chintala, the creator of PyTorch, would take over as CTO.

Fifty-eight minutes later, OpenAI's Application Business CEO Fidji Simo announced on X that Zoph, another co-founder Luke Metz, and founding team member Sam Schoenholz were returning to OpenAI, stating that this had "been in the works for weeks."

Zoph simply retweeted with the comment "Super excited to join the team."

For details, refer to the report "Internal Turmoil at Mira's Company? CTO Fired, Team Returns to OpenAI, Yann LeCun Comments on X."

Upon returning to OpenAI, he was assigned to lead the enterprise business, with the title Vice President & General Manager of Enterprise Business. Against the backdrop of OpenAI clearly stating its intent to converge "side projects" and focus firepower on the two main revenue lines—enterprise and coding—this was a position of significant weight.

Then, five months later, he left again. On June 19, The Verge reported that OpenAI had confirmed Zoph's departure, with him posting a farewell message in the company's Slack. The reason for leaving? Unclear.

And then, came today's tweet. From TML's establishment to today, it's been a year and a half, with three employers, two of which are places he had left before.

Four Versions of That Dismissal

The circumstances of Zoph's departure from TML still lack a universally accepted version. At least four narratives exist that are not entirely compatible.

The first came from tech journalist Kylie Robison. On X, citing two informed sources, she reported that TML terminated Zoph's employment due to "unethical conduct," with Murati announcing the news in an all-hands meeting. This was the earliest version to circulate and the most widely spread.

The second came from WIRED. A source close to TML told them that Zoph was accused of sharing confidential company information with a competitor. WIRED explicitly stated it was unable to verify this with Zoph himself. The publication later reported that last summer, TML management had spoken to Zoph about his relationship with another employee, who worked in a different department and held a leadership position. That employee has since left TML, and this relationship was likely the "improper conduct" referenced in previous reports. WIRED claimed that in the months following that conversation, the working relationship among the co-founders gradually deteriorated, and Zoph began engaging with other companies, including Meta's Superintelligence Lab.

The third came from The Wall Street Journal. Its report stated that Zoph had admitted to Murati that he was in a romantic relationship with a less senior colleague. This relationship began when both were still at OpenAI. Murati had suspected it as early as last summer; Zoph initially denied it but admitted it to Murati weeks later. The report also noted that while this woman held a lower rank than Zoph, she did not report to him. She subsequently left TML and returned to OpenAI. The WSJ also mentioned that during the meeting before his dismissal, Zoph and two others advocated for him to have greater decision-making authority. Murati's response was to question why he hadn't been fulfilling his responsibilities in recent months; he was dismissed two days later.

The fourth version came from Zoph himself. He told The Wall Street Journal that TML terminated his employment after learning he was about to leave, "and that's it." He also stated that TML never cited any performance reasons or unethical conduct as grounds for dismissal to him, and any suggestion to the contrary was false and defamatory.

Four versions, pointing to different causal chains: one is "fired due to misconduct," another is "fired because he was planning to leave," with an unverified leak accusation sandwiched in between. TML has not publicly commented on this to date. The public record of this incident remains in a state of contradiction.

It is worth noting that around the same time, co-founder Andrew Tulloch also left TML (joining Meta in October 2025). At the time, a developer on X compiled a statistic: 8 out of 11 OpenAI co-founders had left, 3 out of 6 Thinking Machines co-founders had left, 1 left from SSI, 1 from DeepMind, and 3 out of 12 from xAI. The only exception was Anthropic—none of its 7 co-founders had left.

The Google He Returns to Is Not the One He Left

Now, let's look at what has happened at Google DeepMind over the past three months that Zoph is returning to.

Starting in mid-June, DeepMind experienced consecutive departures. On June 18, Noam Shazeer, co-lead of the Gemini model and one of the authors of the "Attention Is All You Need" Transformer paper, announced he was joining OpenAI—just two years after Google brought him back via an approximately $2.7 billion licensing deal with Character.AI. Two days later, John Jumper, who shared the 2024 Nobel Prize in Chemistry with Demis Hassabis for AlphaFold, announced he was joining Anthropic, along with AlphaFold paper co-authors Jonas Adler and Alexander Pritzel. Additionally, Denny Zhou had earlier quietly moved to Meta, while early DeepMind employee and prominent figure in reinforcement learning, David Silver, left to start his own company, named Ineffable Intelligence.

On June 22, Alphabet's stock price fell by about 5-6%, with market reports widely linking it to concerns over AI spending and Google's ability to retain top talent. A line in a Fortune report at the time stung: These departures would hurt less if they happened when Google was clearly ahead.

In August, Google underwent a major organizational restructuring. On August 5, Sundar Pichai announced that Hassabis was stepping down as CEO of Google DeepMind, transitioning to Chairman of GDM and Chief Scientist at Alphabet, continuing to helm Isomorphic Labs. The former CTO and Google's Chief AI Architect, Koray Kavukcuoglu, took over day-to-day operations as Senior Vice President, reporting directly to Pichai, responsible for Gemini model development, frontier AI research, and the Gemini Apps & Developer teams.

Chief Scientist Jeff Dean, who had worked at Google for 27 years, announced he was leaving with several colleagues to start a public-benefit company named Discovery Loop, with Google as an investor.

What Kavukcuoglu said upon taking office sounded like a mission statement: The top priority is to ensure the clearest possible path to deliver on the Gemini roadmap. His internal message to employees also mentioned Gemini 4.

Therefore, the timing of Zoph's entry now points clearly: This is the rebuilding period after Koray took over, and what DeepMind needs is someone who can lead the post-training and RL teams. Hence, DeepMind was willing to bring on a senior executive with controversy on his resume at this juncture. On the talent ledger of the past three months, this is probably a rare credit entry for Google.

Details such as who he reports to, his official start date, and whether he will bring people with him are not yet public information.

Conclusion

In the short term, the significance of this news is not hard to judge: DeepMind, after losing a batch of star researchers and undergoing a round of senior management reshuffling, has filled a vacancy with a seasoned leader in post-training and RL. Under the pressure of the Gemini 4 timeline, this addition is likely important.

From another perspective, looking at the entire trajectory—from Murati leaving OpenAI to found TML, to the three co-founders returning to OpenAI, to Zoph moving from OpenAI to DeepMind—the flow of talent has almost completely detached from the unit of "company." Nowadays, even those holding titles like founder or founding team member at AI labs come and go like regular employees.

If a co-founder's departure can be picked up by another company within 58 minutes, what exactly does the identity of "co-founder" anchor in today's AI industry? Equity and non-compete clauses clearly cannot lock down this group of people, and the company vision may not either. What is truly flowing are those few dozen brains that know how to organize billion-dollar-scale experiments, and all labs are bidding for the same set of people.

After all, the most sought-after researchers bring not just technical knowledge from papers and code, but also the judgment of which ideas are worth pursuing, the experience of running hyper-scale experiments, and the ability to recruit other sought-after scientists.

This article is from the WeChat public account "Almost Human" (ID: almosthuman2014), edited by Panda.

Related Questions

QWhat are the key career movements of Barret Zoph over the last 20 months as described in the article?

AOver the last 20 months, Barret Zoph has worked for three companies. He left Google Brain to join OpenAI in September 2022, where he became VP of Research and contributed to ChatGPT's post-training system. In February 2025, he co-founded Thinking Machines Lab (TML) as its CTO. He was publicly dismissed from TML in January 2026 and returned to OpenAI within 58 minutes to lead its Enterprise business. Finally, in June 2026, he left OpenAI again and has now announced his return to Google DeepMind as a Vice President of Research.

QAccording to the article, what are the different narratives surrounding Barret Zoph's dismissal from Thinking Machines Lab (TML)?

AThe article presents four different narratives: 1) He was terminated for 'unethical conduct' (per tech journalist Kylie Robison). 2) He was accused of sharing confidential information with a competitor (per WIRED). 3) He admitted to a romantic relationship with a junior colleague that started at OpenAI and violated company policy (per The Wall Street Journal). 4) Zoph's own version states TML terminated his employment *after* learning he was planning to leave, denying any performance or misconduct allegations and calling them defamatory.

QWhat significant research contributions is Barret Zoph known for from his time at Google Brain?

ADuring his six years at Google Brain, Barret Zoph authored two highly cited papers that were foundational to their fields. He was the first author on 'Neural Architecture Search with Reinforcement Learning,' a key starting point for the Neural Architecture Search (NAS) subfield. He was also first author on 'Learning Transferable Architectures for Scalable Image Recognition.' Additionally, he contributed to work on models like the Switch Transformer, a sparse expert model.

QWhat recent changes in Google DeepMind's leadership and talent pool provide context for Zoph's return?

AIn the months preceding Zoph's return, Google DeepMind experienced significant talent outflow and a leadership reshuffle. Key departures included Gemini co-lead Noam Shazeer (to OpenAI), Nobel laureate John Jumper and colleagues (to Anthropic), and other senior researchers to Meta and startups. In August 2026, CEO Demis Hassabis stepped down, moving to a Chairman role, and was replaced by former CTO Koray Kavukcuoglu, who now focuses on executing the Gemini roadmap. This context of rebuilding under new leadership and needing expertise in post-training and RL aligns with Zoph's hiring.

QWhat broader trend in the AI industry does the article suggest Zoph's career path illustrates?

AThe article suggests that Zoph's rapid movement between top AI labs illustrates a trend where talent, even at the level of co-founders and senior executives, is highly fluid and no longer tightly anchored to a single company. Titles like 'co-founder' have diminished weight as top researchers, driven by the demand for their specific expertise in organizing billion-dollar experiments, move freely between organizations based on new opportunities and challenges, with all major labs competing for the same small pool of elite talent.

Related Reads

Conversation with Macro Analyst Jordi Visser: Micron Has 60%-80% Upside, the Longevity Economy Is a High-Confidence Main Theme

Macro analyst Jordi Visser argues that AI is the sole real driver of the economy today, with traditional GDP metrics failing to capture its value—AI-driven productivity shows up in corporate profit margins instead, as large firms grow revenue while hiring freezes. He attributes rising Treasury yields to AI-related capex and debt issuance, not Fed policy, and views credit default swap widening as a concentration-risk hedging issue rather than an AI bubble. He expects hyperscaler external financing to peak as data centers come online and free cash flow turns positive (citing Oracle's guidance). On labor, he sees Microsoft's buyout of 27% of its workforce as a template: AI-driven revenue-per-employee surges will push governments toward redistributing profits from capital to labor via tax restructuring, since personal income tax bases erode. He warns the market is discounting risks like labor turmoil, political gridlock, memory/data center bottlenecks, and cybersecurity. His key themes are agentic compute and the longevity economy. He is bullish on Eli Lilly as the purest longevity play, predicting it will become the S&P 500's largest company by 2030, and on Micron, which he re-entered around $800 with only ~10% downside and 60-80% upside, citing humanoid robot RAM demand, SpaceX's 10GW compute plan, and long-term supply agreements through 2030. He also holds crypto as his largest allocation, expecting 2027 to mark a breakout year for blockchain commercial applications and payment networks.

marsbit1h ago

Conversation with Macro Analyst Jordi Visser: Micron Has 60%-80% Upside, the Longevity Economy Is a High-Confidence Main Theme

marsbit1h ago

Michael Saylor: Four Centuries of Digital Credit Evolution

Michael Saylor traces the four-century evolution of digital credit, showing how centuries of financial infrastructure paved the way for Bitcoin-based corporate treasuries. The story begins in 1602 with the Dutch East India Company, which separated a company's capital duration from investors' holding periods through transferable shares. Preferred stock followed, exemplified by the Baltimore and Ohio Railroad's 6% perpetual preferred dividend in 1836 and Delaware's 1899 recognition of distinct share classes. Exchanges and tickers—from the 1792 Buttonwood Agreement to the 1867 stock ticker and 1971 Nasdaq—gave ownership a public market. U.S. securities and commodities law (1922–1974) established the distinction between assets a company holds and securities it issues. The 1940 Investment Company Act's "40% test" still shapes how bitcoin treasury companies use common stock, preferred stock, and debt. Tax law dating to 1921–1954 distinguishes return of capital from income, allowing preferred dividends to reduce cost basis before gains are realized. Shelf registration (Rule 415, 1982–1983) and at-the-market programs (1982, 2005) let companies prepare financing in advance and raise capital at market prices. Buybacks, recognized since 1828 with the 1982 Rule 10b-18 safe harbor, let firms manage capital in both directions. Bitcoin, launched in 2008–2009 and classified as a commodity by the CFTC in 2015, provides a new reserve asset, while companies supply capital structure, dollar liquidity, and distribution policy. Digital credit thus combines over 400 years of corporate ownership transfer, roughly 190 years of U.S. preferred stock, more than a century of tax cost-recovery rules, and four decades of modern shelf and ATM regulation into a familiar framework for investors.

marsbit3h ago

Michael Saylor: Four Centuries of Digital Credit Evolution

marsbit3h ago

The Hidden Risks Behind Anthropic's Nearly $1 Trillion Bet

Anthropic's IPO filing reveals a fragile, high-risk financing mechanism behind its AI ambitions. The company has committed to up to $895.4 billion in data center spending through 2045, with profits squeezed by rising compute costs. Rather than paying upfront for hardware, Anthropic relies on supplier-backed financing: Broadcom arranged a $63 billion equipment financing deal under which it absorbs losses if Anthropic defaults—up to roughly $43 billion once equipment is delivered. Google and AMD similarly tie funding to hardware leases or deployments. This shifts risk onto chipmakers' balance sheets, and because custom TPUs serve a narrower ecosystem than Nvidia GPUs, resale risk is higher. Analysts warn that while individual loans appear safer, risk concentration across the system is rising and harder to assess. Anthropic also can issue up to $42 billion in convertible notes to Broadcom to cover rent, with interest rates rising from 9% to 13% and conversion terms worsening if revenue targets are missed. Despite holding $146.6 billion in liquidity as of September 30, its massive forward obligations and tangled partnerships could limit access to investment-grade financing. Key concerns: whether revenue growth can outrun compute costs and cover future commitments, and whether drawing on credit lines signals distress. Anthropic's Direct initiative—separating data center leasing from chip purchases—and IPO proceeds could give it more flexibility, but ultimately debt repayment depends on end-customer revenue.

marsbit13h ago

The Hidden Risks Behind Anthropic's Nearly $1 Trillion Bet

marsbit13h ago

Trading

Spot
活动图片