A Timeline of AI Researchers Leaving OpenAI, Anthropic, and Google DeepMind

Noam Shazeer co-wrote “Attention Is All You Need,” the paper that gave the world the Transformer. He spent the last several years co-leading Gemini at Google DeepMind — a model built on the architecture he helped invent. On June 18, 2026, he announced he was leaving. For OpenAI. To help build a rival to the model he’d just been running.

That’s one departure. Lined up in order since 2020, the rest of them tell two different stories depending which half of the timeline you’re reading. The early departures are almost all about a disagreement over safety, and they created new labs. The recent ones are almost all about money and compute, and they’re mostly just moving people between labs that already exist. Here’s the sequence, in order.

2020–2021: the split that created Anthropic

Dario Amodei left OpenAI in December 2020. He’d been VP of Research, co-leading the team behind GPT-3. His sister Daniela Amodei, OpenAI’s VP of Safety and Policy, left with him. The reported reason: they’d concluded OpenAI’s leadership was underweighting alignment relative to how hard it was pushing on scale.

In January 2021 they announced Anthropic, registering the company in California that February with $124 million in seed funding, alongside fellow OpenAI alumni Tom Brown, Jared Kaplan, Sam McCandlish, and Jack Clark — most of them co-authors on OpenAI’s own GPT papers. The pitch investors reportedly found contradictory: build the most powerful AI systems in the world, and simultaneously make them safe. Five years later, Anthropic is one of the three labs everyone in this article keeps leaving each other for.

May 2024: OpenAI’s safety team resigns

Ilya Sutskever, OpenAI co-founder and chief scientist, announced he was leaving on May 14, 2024, citing a “project that is very personally meaningful” to him. Jan Leike, who co-led the Superalignment team with Sutskever, resigned the next morning. His public statement was blunter than Sutskever’s: “safety culture and processes have taken a backseat to shiny products,” and OpenAI “must become a safety-first AGI company.”

OpenAI dissolved the Superalignment team days later and folded its remaining members into other groups. Sutskever went on to found Safe Superintelligence Inc. Leike joined Anthropic.

Andrej Karpathy, an OpenAI founding member, had already left that February — his second exit from the company — to start Eureka Labs, an AI-education startup, saying at the time there was no drama behind it.

September 2024–2025: Murati leaves, and Meta starts buying

Mira Murati, OpenAI’s CTO and one of the people most publicly associated with shipping ChatGPT, announced her departure on September 25, 2024, after six and a half years. In February 2025 she launched Thinking Machines Lab as CEO, taking OpenAI veterans Barret Zoph, John Schulman, Lilian Weng, Andrew Tulloch, and Luke Metz with her.

By the summer of 2025 the pattern shifted from people leaving to start something to people leaving to join something that already existed and was writing very large checks: Meta’s newly formed Superintelligence Lab. OpenAI lost at least seven researchers to it, including Jason Wei (o1 and deep research), Zhiqing Sun, Hyung Won Chung, and Shengjia Zhao — a co-creator of ChatGPT and GPT-4 who became Meta Superintelligence Lab’s chief scientist.

2026 kept the exits coming from OpenAI’s side without a single unifying cause: Bill Peebles left in April after three years leading Sora, without announcing a next stop. Johannes Heidecke, who led safety systems, left in July. Brad Lightcap, employee since 2018, left to “start something new.” By mid-2026, coverage was counting a dozen senior OpenAI departures for the year alone.

2026: Google DeepMind’s turn

Then the exodus swung to Google. In June 2026, Shazeer left for OpenAI. In the same window, John Jumper — who shared the 2024 Nobel Prize in Chemistry for AlphaFold — announced he was leaving DeepMind for Anthropic, alongside fellow DeepMind researchers Jonas Adler and Alexander Pritzel. Denny Zhou, who’d founded Google Brain’s reasoning research team eight years earlier, had already been at Meta for months by the time this made headlines. Fortune reported that of everyone who left DeepMind in the prior twelve months, 25% went to Anthropic, 21% to Meta, and 14% to OpenAI — Anthropic, not Meta’s checkbook, had become the single most common landing spot for departing DeepMind talent.

And in May 2026, Karpathy resurfaced — not at his own company, but at Anthropic, on the pretraining team under Nick Joseph, working, per his own announcement, on “using Claude to accelerate pretraining research.” The OpenAI co-founder who left twice to build his own thing had joined the lab founded by the OpenAI VP who left because OpenAI wasn’t safety-focused enough.

What changed between the two halves

Run the two ends of this timeline next to each other and the difference is obvious. 2020 to early 2025, the marquee departures were people quitting over a disagreement about how the work should be done, and the result was a new lab each time — Anthropic, SSI, Thinking Machines. Every one of those companies still exists and matters.

From mid-2025 on, nobody’s founding anything. Wei, Zhao, Shazeer, Jumper, Karpathy — they’re moving laterally, from one already-massive lab to another, and the stated reasons (when there are any) are about where the most interesting pretraining work is happening, not about a values dispute with the old employer. That’s a different kind of signal. A researcher leaving to start a company is a bet that the field needs something that doesn’t exist yet. A researcher leaving DeepMind for Anthropic, or OpenAI for Meta, is a bet about which existing lab is going to win — and the fact that DeepMind is currently on the losing side of that bet more than anyone else is probably the most concrete data point in this whole list.

Sources