DeepMind Engineer Quits Because His Chips Were Working Too Well

Robert O’Callahan, a Google DeepMind engineer based in New Zealand who worked on next-generation chip design tools, resigned this week in a blog post titled “Goodbye Google,” and the reason he gave is unusual mainly for how directly it targets his own job description. He didn’t object to a specific model, a specific deployment, or a specific safety failure. He objected to the fact that his team’s entire function — making the chips underneath AI models faster and cheaper — necessarily accelerates the thing he’s worried about, regardless of which lab, which model, or which use case eventually runs on top of it. You cannot make the hardware safer by making it better at its job, in his framing, because its job is the acceleration.

“I firmly believe AI progress is currently far too rapid,” he wrote, adding that he has “doubts about the destination too.” He cited the standard menu of frontier-safety concerns — misalignment, deceptive models, reward hacking, what he calls “psychotic swarms” — as things that were once hypothetical warnings and are now, in his account, observed phenomena. He was careful to frame this as calibrated concern rather than certainty of doom, noting he doesn’t believe catastrophe is guaranteed, only that the risk is serious enough to warrant precaution nobody currently seems to be exercising. And he was explicit that the decision was, for him, a matter of religious conviction: an elder in his church, he wrote that continuing to build tools he believes accelerate reckless AI development “would not be a Jesus-following thing to do.”

The timing invites a comparison he tried, in the post itself, to head off. O’Callahan’s resignation lands roughly three weeks after Anthropic researcher Jacob Coxon’s own departure post went from a few thousand views to 169 million, turning “I quit over AI safety” into a genre with its own momentum. O’Callahan pre-empted the obvious read by attributing part of his own timing to a previously scheduled backpacking trip rather than to Coxon’s virality — a disclaimer that is either refreshingly honest or the exact thing someone says right before joining a trend anyway. He’s not the only DeepMind departure in this vein this year, either; a colleague, Bilal Chughtai, left over similar concerns earlier.

O’Callahan’s specific position is at least structurally interesting in a way Coxon’s wasn’t: he didn’t work on models, alignment, or anything resembling the sharp end of AI capability. He worked on chip tooling — infrastructure so far upstream of an actual deployed model that his resignation is closer to a foundry worker quitting because the roads near the plant are too well paved. He acknowledged as much himself, noting a transfer was impractical because Google has no other engineering team in New Zealand, and that his own departure will have only a “small — but nonzero” effect, given how specialized his skills are. That is an unusually honest thing for a resignation post to admit: that the gesture matters more as a gesture than as an intervention.

Not everyone is treating the moment as it’s being packaged. One outlet covering the resignation noted, fairly, that a handful of departures out of the thousands of people employed across frontier labs doesn’t obviously constitute an industry trend, and pointed out that engineers leaving well-compensated roles at companies whose equity has made many of them independently wealthy invites its own line of questioning about motive — not because the concern isn’t genuine, but because “principled departure” and “already vested” aren’t mutually exclusive. O’Callahan didn’t dispute this in his post; he simply didn’t address it, which is its own kind of answer.

What happens now is presumably nothing, structurally speaking — Google will hire someone else to work on the chip tools, the chips will still ship, and Colossus-scale data centers will keep getting built regardless of who signs the New Zealand time sheet. What changes is smaller and more diffuse: one more public document, from one more credentialed insider, arguing that the industry’s own foundations are moving faster than anyone building them can vouch for. O’Callahan said he expects many of his former colleagues privately agree with him. He also expects most of them to keep showing up to work on Monday. Both things, per his own account, can be true.

Sources: