Who is hailed as the pioneer of "boomerang job-hopping" in the AI industry?
Before you stands Barret Zoph: former Senior Research Scientist at Google Brain, former Vice President of Post-Training Research at OpenAI, former Co-founder and CTO of Thinking Machines Lab, former Head of OpenAI Codex Commercialization, and current Vice President of Research at Google DeepMind.
(This space is not enough for so many people.gif)
If you still don't recall who he is, he was the one who first demonstrated the video call capabilities of ChatGPT to the world during the GPT-4o "Her" launch event.
Due to network latency issues, he was even called "a wooden table" by ChatGPT.
From Pioneer of Neural Architecture Search to Head of Post-Training
Barret Zoph's career path has been unconventional. He doesn't even hold a Ph.D., yet he has risen to the positions of Vice President of Research and CTO.
After graduating with a bachelor's degree from USC in 2016, he joined the Google Brain Residency program. Collaborating with Quoc Le, he co-authored "Neural Architecture Search with Reinforcement Learning," effectively helping to pioneer the field of Neural Architecture Search (NAS).
His contributions extended beyond architecture search.
In January 2021, in collaboration with William Fedus and Noam Shazeer, he published the Switch Transformer paper. This was the first model to achieve stable training at the trillion-parameter scale and later became a standard reference for sparse Transformer designs.
After leaving Google to join OpenAI in 2022, he quickly ascended to Vice President of Research, focusing on reinforcement learning for post-training.
The post-training phase encompasses model alignment, instruction following, tool use, and refining user-facing capabilities, which are crucial steps in turning foundational models into usable products.
USC later listed him in official promotions as an alumnus "who paved the path for ChatGPT," alongside the two authors of the Transformer paper.
In October 2024, former OpenAI CTO Mira Murati departed. In early 2025, she co-founded Thinking Machines Lab with Zoph and four others.
Zoph served as Co-founder and CTO, but cracks appeared within less than a year.
In January 2026, Zoph, alongside another co-founder Luke Metz and early employee Sam Schoenholz, all left Thinking Machines Lab simultaneously and returned to OpenAI.
According to *The Wall Street Journal*, Murati subsequently informed employees that Zoph's employment had been terminated, citing issues related to "performance, conduct, and trust." The report also disclosed that during his tenure, Zoph had an undisclosed relationship with an employee, leading to a subsequent reduction in his responsibilities.
Zoph presented a different version of events:
He claimed that before being fired, he, Metz, and Schoenholz had personally asked Murati to grant Zoph full authority over all technical decisions. He stated that Murati announced the termination only after learning of his plans to leave and that the reduction in duties was merely a temporary arrangement.
Upon returning to OpenAI, Zoph's role shifted. He was no longer responsible for technical research but was instead tasked with promoting the use of the Codex programming product to enterprise customers.
A few months later, he left OpenAI again, this time joining Google, where he resumed responsibility for reinforcement learning and post-training research.
AI Talent Transfer Like Professional Athletes
The timing of Zoph's return to Google is somewhat intriguing.
Google DeepMind is undergoing its most significant personnel turbulence in the past three years.
In early August, co-founder Demis Hassabis stepped down from day-to-day management, transitioning to the role of Chairman of Google DeepMind and Chief Scientist at Alphabet.
On the same day, Jeff Dean, a 27-year Google veteran, announced his departure. He, along with longtime colleagues Sanjay Ghemawat and Oriol Vinyals, as well as Quoc Le (who co-pioneered NAS with Zoph), founded the AI research company Discovery Loop.
Earlier, in June, Noam Shazeer, co-lead of Gemini and co-author of the Transformer paper, left Google for OpenAI. John Jumper, who won the 2024 Nobel Prize in Chemistry for his work on AlphaFold, went to Anthropic.
Analysis by data intelligence firm Zeki Data reveals a broader trend:
The inflow-to-outflow ratio of research and senior engineering talent at DeepMind decreased from approximately 12:1 in Q2 2023 to about 2:1 in Q3 2026. This means that for every 2 people hired, one person leaves.
In contrast, Anthropic's ratio during the same period was 22:1, and OpenAI's was 5.7:1.
Among those leaving DeepMind, about 25% went to Anthropic, 21% to Meta, and 14% to OpenAI.
Meanwhile, organizational adjustments are underway. An all-hands meeting on August 6th disclosed that some teams would be moved from DeepMind into the Google corporate structure, with Kavukcuoglu, as Senior Vice President, taking over the day-to-day operations of Gemini model development and frontier research.
This AI talent market phenomenon, which "looks more like professional sports transfers," and its unconventional hiring cycles are primarily seen among a very limited number of senior research leaders who have genuinely overseen large-scale training and post-training processes for frontier models.
The factors driving this talent mobility go far beyond compensation.
According to *Fortune* interviews with several current and former DeepMind employees, reasons for departures include:
Competitors poaching with lucrative offers, dissatisfaction with Google's position in the AI race, low internal morale, and the allure of pre-IPO equity at OpenAI and Anthropic.
As Google increasingly aligns DeepMind's focus around Gemini commercialization, some researchers feel that the open-ended, long-cycle pure scientific exploration is being squeezed out. This, however, was precisely the core asset that initially attracted academic talent to DeepMind.
References:
[1]https://x.com/barret_zoph/status/2092790775524516206
[2]https://viterbischool.usc.edu/news/2023/03/attention-is-all-you-need-usc-alumni-paved-path-for-chatgpt/
This article is from the WeChat public account "QbitAI," author: Meng Chen
