Why Frontier AI Talent Keeps Leaving, Then Coming Back

Anthropic, OpenAI, Google DeepMind, Meta, and a growing ring of startups are now competing for the same small pool of people who can still move frontier models forward. The interesting part is not just that talent is moving. It is that some of the same people are leaving frontier labs for startups, then returning, or moving again to a different lab that better matches their view of what serious AI work should be. The pattern suggests a paradox: pay and prestige matter, but mission, safety culture, and governance credibility are increasingly part of the retention package. ^{1, 2, 3}

What “loyalty-safety” means in practice

The phrase is useful because it captures two different choices that often get blurred together.

“Loyalty” is not old-fashioned corporate devotion. It is whether top talent feels enough pull to stay inside a frontier lab when startups, other labs, or independent ventures offer more freedom or upside. “Safety” is not a public policy statement or a launch checklist. It is whether people inside the lab believe the organization will actually preserve room for restraint, independent judgment, and serious risk review when the pressure to ship gets intense.

That distinction helps explain Boris Cherny’s return to Anthropic after a 14-day stint at Cursor. This is best read as an illustrative case, not a general law. Cherny initially left because he thought Cursor better understood the trajectory of AI coding. He came back because product energy and team vision were not enough to replace the safety-oriented culture he wanted at Anthropic. ⁴

"Ultimately, the speaker found that product excitement and team vision were insufficient substitutes for Anthropic’s safety-oriented culture."

— 1 Minute Signal coverage of Lenny's Podcast ⁴

That is the core of the paradox. For some frontier talent, mission alignment is no longer a decorative value. It is a work requirement.

Why people leave frontier labs

The strongest structural explanation is also the least romantic: people leave when another environment offers more leverage.

nextomoro’s lab-flow dataset shows Anthropic as the only net importer among the four major frontier labs it tracks, while OpenAI, Google DeepMind, and Meta AI are net exporters. It also finds that more than half of departures from those labs are flowing to a small set of insurgent companies, including Anthropic, Recursive Superintelligence, AMI, Thinking Machines Lab, and H Company. ¹

That matters because the move is not just about compensation. In nextomoro’s framing, hyperscaler-backed labs cannot offer the same combination of autonomy, equity, and strategic independence that insurgents can. They cannot promise equity in a company that might itself become a frontier lab, they sit inside parent-company alignment constraints, and they often have a mission subordinated to a broader consumer-products business. ¹

The same dynamic shows up in reporting on Google DeepMind. USA Business Times describes senior research scientists and principal engineers leaving after the 2023 Brain-DeepMind merger pushed the organization toward product deadlines tied to Gemini. The article also says Demis Hassabis has tried to retain talent by promising more autonomous research labs, while Alphabet has considered a sizable retention fund for DeepMind’s top researchers. Even if some of those retention details remain unconfirmed, the broader signal is clear: the cost of losing research identity has become visible enough to trigger defensive measures. ⁵

The June 2026 departures described in DEV Community sharpen the point. Noam Shazeer’s move to OpenAI after a compute dispute, and the claim that Google’s internal review cycles can stretch to six months, are classic examples of why top researchers chafe at large-lab friction. This source is more polemical than the others, so it should be treated cautiously. Still, its underlying complaint fits a pattern seen elsewhere: in frontier AI, delay is no longer neutral; it can feel like career decay. ⁶

The “why now” is mostly about 2025–2026 pressure

This talent churn is not happening in a vacuum. It is happening as the frontier-model race gets more capital-intensive, more productized, and more organizationally brittle.

Dust co-founder Stan Hulu, in the Y Combinator coverage, says the fundamental nature of work for his team shifted around November 2025 as agentic usage exploded and pricing models had to change. That is a startup example, but it points to the broader environment: the pace of AI development is changing the shape of jobs themselves. ⁷

At the same time, the 2026 AI Index shows the U.S. is attracting far fewer AI researchers and developers than it used to, with a large decline since 2017 and a sharp drop in the last year alone. It also says new AI PhDs in the U.S. and Canada are increasingly choosing academia rather than industry. That does not prove frontier labs are running out of talent. But it does suggest the funnel is getting less forgiving, which raises the value of every experienced researcher who can still operate at the frontier. ⁸

The result is a market in which elite talent can move more selectively. The strongest people are not locked into one employer by inertia. They can compare mission, autonomy, safety posture, and upside across multiple credible options.

Safety culture is now part of retention

This is where the paradox becomes more than a labor-market story.

The Cloud Security Alliance’s 2026 analysis argues that the organizations building the most capable AI systems are increasingly the ones deciding whether those systems are safe. It also says the departures from frontier labs are not random attrition, but the loss of the most intellectually generative researchers — the people best positioned to build independent research programs and train the next generation. ⁹

That framing helps interpret recent OpenAI turnover. Business Insider reports Johannes Heidecke leaving OpenAI’s Safety Systems team as the company reorganizes research and safety under Mira Glaese. OpenAI’s stated reason is that safety and research decisions are interdependent. That is a defensible management argument. But the personnel churn suggests a harder reality: when safety is asked to live inside a fast-moving commercial machine, the people closest to safety often feel the pressure first. ²

"A safety culture is not a document. It is what happens in the room when the decision is hard and the competitive pressure is high."

— Studio Hyra ¹⁰

That line matters because it distinguishes branding from operating reality. Studio Hyra’s analysis of Mira Murati’s testimony argues that safety teams are often sidelined when speed and market position become the overriding priorities. That is a structural claim, not a moral one: even well-intentioned organizations tend to compress safety review when they feel they have to. ¹⁰

The broader labor effect is straightforward. If smart people conclude that safety is mostly ceremonial, they will either leave, push harder from the inside, or choose an institution whose culture feels more credible. Anthropic’s ability to attract talent from OpenAI and Google DeepMind suggests that some frontier workers now treat safety posture as part of the job description, not a side constraint. ^{1, 2}

A return to Anthropic is not a one-off

Cherny is not the only useful example of re-entry or lateral return.

nextomoro’s dataset also shows a more complex arc in which John Schulman moved from OpenAI to Anthropic and then to Thinking Machines Lab within months. That is not a simple “leaving and coming back” story, but it is strong evidence that frontier talent is increasingly treating labs as a sequence of mission-aligned stations rather than a lifelong home. ¹

Andrej Karpathy’s reported move to Anthropic is another sign of the same pattern. He is not returning to Anthropic, but the move matters because it reinforces the idea that even highly established researchers are still willing to reposition themselves around where they believe frontier work is happening. ¹¹

The point is not that people are fickle. It is that the frontier is reorganizing around preference clusters: some people want startup velocity, some want research depth, some want a safety-first culture, and some are willing to move again if the organization stops matching their definition of meaningful work.

Governance and mission are now retention variables

A recurring mistake in commentary about these moves is to separate “technical work” from “governance” as if one were soft and the other hard. The evidence here does not support that split.

Musk’s comments, as summarized in 1 Minute Signal coverage of The Economist, frame OpenAI’s shift from nonprofit mission to a closed-source for-profit structure as a betrayal of its original mandate and a reason ethical teams left for rivals like Anthropic. Another 1MS summary of his remarks describes an industry-led safety model built around peer coordination rather than state regulation. Whether or not that proposal is workable, the talent signal is real: people are reacting to governance choices as if they are career-defining product decisions. ^{12, 13}

That is why the nextomoro data matters so much. If Anthropic is absorbing more talent than the other major labs, it may not simply be because it has the best model or the biggest checks. It may be because it has assembled a more credible compact between frontier capability and safety seriousness. ¹

This also helps explain the movement from labs to startups. Frontierbeat’s reporting on researcher-led startups says large labs have increasingly prioritized commercial benchmarks and rapid release cycles, while newer companies chase the research areas those labs have deprioritized, from reinforcement learning to autonomous labs and real-world safety. That is not a purely ideological split. It is a labor-market split between organizations that want to productize the present and organizations that want to keep the research horizon open. ³

Why this matters beyond the labs themselves

The practical consequence is that talent migration is becoming a governance problem.

Stanford HAI’s 2026 AI Index says the U.S. is attracting fewer AI researchers and developers than it once did, and that many new AI PhDs are choosing academia rather than industry. The Cloud Security Alliance and TEXXR both warn that if the most capable researchers concentrate inside private labs, universities lose the faculty needed to train students and preserve independent evaluation capacity. ^{8, 9, 14}

TEXXR puts the consequence plainly: if the pattern persists, private frontier labs, not universities, will increasingly determine which AI questions get resources, compute, and publication access. That is an epistemic shift, not just a staffing trend. ¹⁴

For builders and investors, the implication is uncomfortable but useful. The best people are not only choosing between employers. They are choosing between institutional compacts. If a lab cannot credibly explain why its safety posture will hold under pressure, it will lose people to a competitor that can. If a startup cannot offer enough mission or leverage, it will lose people back to the lab. And if both fail, people will keep moving until they find an environment that matches the way they want to work.

What builders should take from this

If you are building in AI, the answer is not simply to pay more. The sources point to a different set of levers: autonomy, mission clarity, research identity, and whether safety is actually embedded in decision-making. ^{1, 10, 15}

A few practical takeaways follow:

• Mission is now a retention lever. If your lab or startup cannot explain why its safety posture is structurally credible, top talent will assume it is branding. ^{2, 10}
• Bureaucracy has a talent cost. Large-lab friction, long review cycles, and fragmented org charts can be enough to push people out. ^{5, 6}
• Return moves are real, but they are selective. Cherny’s return to Anthropic is illustrative of a broader pattern, not proof that every startup-to-lab move is safety-driven. ⁴
• The best people are choosing leverage over comfort. Some move to startups for shipping speed and ownership; others move to frontier labs for mission and research seriousness. ^{1, 15, 16}
• Safety teams need authority, not just language. If they are always downstream of launch pressure, they will keep losing credibility with both employees and recruits. ^{9, 10}

The deepest lesson is that frontier AI talent is being sorted by more than compensation. It is being sorted by whether the institution’s mission feels real enough to work for, and safe enough to stay inside. That is why some people leave. It is why some come back. And it is why mission is becoming a retention strategy, not a slogan, at frontier AI labs.