From Fired Co-Founder to DeepMind VP: Barret Zoph's Turbulent 20 Months End Back at Google

In January, Barret Zoph was fired from Thinking Machines Lab — the $12 billion startup he co-founded — after a falling-out with CEO Mira Murati. On Thursday, he landed one of the most senior research jobs in the industry: Vice President of Research at Google DeepMind, where he will lead work on reinforcement learning and post-training.

It is a startlingly fast rehabilitation for a researcher whose last twenty months read like a Silicon Valley soap opera. And it is an equally pointed signal from Google, which has spent 2026 bleeding senior AI talent to rivals and is now paying up — in title and scope — to bring a proven post-training leader back into the fold.

A homecoming, announced plainly

Zoph announced the move himself on X. “Excited to announce I am joining Google DeepMind!” he wrote. “I started my AI career in the Google Brain Residency class and am excited to rejoin.”

The homecoming framing is genuine. Zoph spent roughly six years at Google Brain, the research group that was folded into Google DeepMind in 2023, before leaving in the first wave of talent migration that reshaped the field. Reuters reports he will hold the title of Vice President of Research after previously serving as VP of Research at OpenAI — a rare lateral move between frontier labs that underscores how liquid the market for senior AI talent has become.

At DeepMind, his remit is reinforcement learning and post-training — the discipline of turning raw pretrained models into useful, aligned, agentic products through fine-tuning, RLHF, tool use, and reasoning training. It is, by most accounts, the single most contested battleground in frontier AI right now.

The wildest twenty months in AI careers

To understand why Google’s move raised eyebrows, it helps to replay the timeline.

2015–2021, Google Brain. Zoph joined through the Google Brain Residency program and stayed six years, building an influential body of work. His early research on neural architecture search — using machine learning to design neural networks — predates the modern foundation-model era and remains widely cited. He later became one of Google’s key post-training leaders.

2021–2024, OpenAI. He moved to OpenAI as VP of Research (Post-Training), where his team worked on alignment, tool use, and the reasoning behaviors that made ChatGPT and GPT-4 class models genuinely useful rather than merely fluent.

2024–2026, Thinking Machines Lab. Zoph co-founded Thinking Machines Lab with former OpenAI CTO Mira Murati, serving as CTO. The startup raised at a $12 billion valuation and became one of the most-watched new labs in the world.

January 2026, the firing. According to Wall Street Journal reporting, Murati told employees there were issues with Zoph’s “performance and conduct”; the Journal also reported he had an undisclosed relationship with a colleague. Zoph had told Murati on a Monday he was considering leaving — and was fired by Wednesday. Murati named PyTorch co-creator Soumith Chintala as the new CTO. Zoph, along with co-founders Luke Metz and Sam Schoenholz, returned to OpenAI.

August 2026, the exit. Less than six months after rejoining OpenAI, Zoph left again — and this time the landing spot is Google.

Why DeepMind wants him now

The timing is not accidental. Google DeepMind has spent 2026 in an unusually defensive posture on talent. In August alone, Fortune reported on Zeki Data analysis showing DeepMind’s share of elite AI hires in EMEA collapsed from 49% in 2022–23 to 18.6% in 2025–26, with its arrivals-to-departures ratio falling from 12:1 to just 2:1. Anthropic alone absorbed 25% of departing DeepMind researchers in the past year.

The lab also changed leadership this summer, with Koray Kavukcuoglu taking charge as CEO amid the race to keep Gemini competitive with OpenAI and Anthropic — a race that is currently decided less by pretraining scale than by post-training craft: reasoning, tool use, agentic reliability.

That is precisely Zoph’s specialty. He has now held post-training leadership roles at two frontier labs (OpenAI, and the founding team of a third, Thinking Machines) and is one of the few people who has run post-training at scale across multiple organizations. For a DeepMind trying to close a product-quality gap, that resume outweighs the baggage of a messy departure — and Google is betting the drama was startup-specific, not engineering-specific.

What this says about the talent market

The move is the latest data point in what has become AI’s defining labor market: senior researchers with frontier post-training experience are now valued like franchise athletes, and loyalty is negotiated annually.

Three dynamics stand out:

The talent war has gone fully bidirectional. For years the flow was one-way — Google trained researchers, OpenAI and startups harvested them (Noam Shazeer’s own move from DeepMind to OpenAI in June was a marquee example). Zoph’s move shows Google now winning back senior leaders it lost, and willing to hand them VP scope on arrival.

Conduct risk is being priced, not disqualifying. A co-founder fired for undisclosed conduct issues would once have been radioactive. Six months later he holds a VP of Research title at DeepMind. The scarcity of proven post-training leaders has effectively created a forgiveness market for everyone below C-level.

Post-training is the new pretraining. When compute and data equalize across labs, the differentiator is what happens after pretraining: RL environments, agentic training, alignment pipelines. Labs are staffing accordingly, and Zoph’s RL-and-post-training mandate at DeepMind is the clearest signal yet of where Google thinks the Gemini gap lies.

The road ahead

For Zoph, the DeepMind job is a chance to write a second act — in the building where his career started, with the resources of Alphabet behind him, and a clearly defined mission: make Gemini’s post-training stack world-class.

For Thinking Machines Lab, the episode is now firmly in the past; the company has moved on under CTO Soumith Chintala and continues to ship. For OpenAI, losing a returned leader within six months is a small but notable crack in the narrative that returning alumni stay put.

And for the rest of the industry, the message is blunt: in 2026’s AI economy, elite researchers don’t just change jobs — they change jobs back. The loop from Google Brain to OpenAI to startup and home again can now be completed in under two years, drama included. Whether Zoph’s redemption arc ends with Gemini closing the gap on ChatGPT and Claude is now one of the more quietly consequential questions of the next year.