Google DeepMind Chip Engineer Resigns Over AI Pace, Not Safety Role

Most of September's AI-safety resignations came from people whose job was literally to evaluate AI risk. Jacob Coxon, Joe Benton, and Josh Engels all left safety or evaluation teams at Anthropic and Google DeepMind, each citing AI risk directly. Robert O'Callahan's resignation, announced September 24, 2026, is a different shape of story: he wasn't on a safety team at all. He was building tools to make the next generation of AI chips faster and cheaper — and quit because he decided that work itself was the problem.

O'Callahan is best known outside Google for rr, an open-source record-and-replay debugger widely used across the systems-programming world, and Pernosco, a commercial "omniscient debugging" platform built on the same idea. He'd previously worked at Mozilla before joining Google DeepMind's chip-design tooling group.

TL;DR

What he actually said

In his post "Goodbye Google", O'Callahan is specific about what he's objecting to — not a particular model's behavior, but the rate of change itself:

"AI progress is currently far too rapid (and I have doubts about the destination too)."

His stated reasoning centers less on a single catastrophic scenario and more on humanity's adaptive capacity: decisions people make about university, career, and long-term planning assume a degree of stability that a fast-moving AI economy no longer guarantees. He listed a cluster of concerns rather than one headline risk — reward hacking and misalignment in advanced systems, cognitive effects from AI-mediated work and relationships, concentration of economic power, and accountability gaps in how labs govern themselves.

He also addressed his own complicity directly: chip-design tools that make AI training and inference faster and cheaper are, in his framing, unambiguously accelerationist, regardless of the specific application. He said he searched for a role inside Google that wouldn't contribute to that dynamic and didn't find one.

Not the same story as the safety-researcher exits

The distinction matters for reading this correctly. Earlier in September, two separate, on-the-record resignations drove a narrative about an "AI safety exodus" inside the frontier labs:

O'Callahan was explicit in his post that his decision "has nothing to do" with that earlier wave. It's a meaningful clarification: Coxon and Benton/Engels resigned from roles whose entire purpose was assessing AI risk, which reads as an insider warning about the field's own governance. O'Callahan resigned from a hardware-adjacent engineering role with no formal safety mandate — his objection is closer to "I don't want to build the thing that makes this go faster" than "I evaluated this and it's more dangerous than my employer admits." Both are legitimate data points, but they're not the same claim, and conflating them overstates how organized or coordinated the departures actually are.

Bloomberg's coverage of the period noted O'Callahan's post was part of a broader cluster that also included Bilal Chughtai, who left DeepMind for METR — the same independent evaluator Benton and Engels joined. Even so, motive and role differ enough between these departures that treating them as one movement risks flattening genuinely different objections into a single headline.

What people are asking

Isn't it contradictory to quit over AI pace and then build AI debugging tools?

This was the most common pushback on Hacker News. O'Callahan's own answer is a distinction between accelerating capability (what he was doing at Google) and making existing AI more reliable and controllable (what he says he wants to do next). Critics find that line thin — debugging tools that make AI agents more effective at writing code plausibly still speed up the broader field. Whether that distinction holds up is unresolved and worth being skeptical of; it's the same tension present in Anthropic's own embedded-evaluator program, where safety-adjacent work still runs on frontier models.

Does he think AI is going to "kill us all"?

No — that framing belongs to a separate story, Google DeepMind researchers' more dramatic exit language from mid-September. O'Callahan's stated concerns are broader and less apocalyptic: pace outstripping adaptive capacity, concentration of power, and accountability gaps, rather than a specific existential claim.

What does "unambiguously pro-human work" actually mean for a chip engineer?

In practice, he's pointing at his existing open-source and commercial debugging tools — rr and Pernosco — and a stated interest in how AI agents themselves debug code, rather than any new AI-capability project. It's a narrower, more concrete answer than the phrase alone suggests: less "I'll work on AI safety" and more "I'll keep doing the developer-tooling work I was already known for, and see if it's useful applied to AI-assisted debugging specifically."

How should builders read this?

Not as a signal to stop building. O'Callahan's objection is about the margin of AI chip and infrastructure work specifically — not a claim that any particular model or product is unsafe to use today. If you're evaluating whether to trust a frontier lab's pace claims generally, Anthropic's own R&D automation index and the broader "pace the frontier" reaction thread are more directly useful reads than any single resignation.

What his existing projects actually do

Understanding rr and Pernosco helps explain why O'Callahan's "pro-human work" framing isn't just a vague gesture. rr is an open-source record-and-replay debugger for Linux: it captures a deterministic recording of a program's execution — every system call, every scheduling decision — so a developer can replay that exact run forward and backward, set watchpoints on when a variable last changed, and reproduce a bug that only shows up once in a thousand runs. It's used widely in systems-level and compiler engineering specifically because those bugs — race conditions, memory corruption, timing-dependent failures — are notoriously hard to reproduce with a normal debugger.

Pernosco is the commercial, hosted evolution of that idea: an "omniscient debugging" service built on rr's recording technology, aimed at teams debugging complex, hard-to-reproduce failures in production-scale codebases. Both projects predate his time at Google DeepMind by years — he's not pivoting into a new field so much as returning full-time to work he was already known for before joining big-lab AI infrastructure.

That context matters for his stated next step: applying the same "let you see exactly what happened and why" philosophy to how AI coding agents debug their own generated code. If an agent can be given the same omniscient-replay visibility a human engineer gets from rr, the argument goes, its debugging loop becomes more reliable and auditable — a claim that's plausible on its face but, like the broader "pro-human work" framing, hasn't been demonstrated publicly yet.

The takeaway

O'Callahan's resignation adds a genuinely new data point to September's cluster of AI-related departures — not a repeat of the safety-researcher story, but a hardware engineer deciding that infrastructure work itself crosses a line he's not willing to hold anymore. That's a harder position to dismiss as performative, since it costs him a well-compensated role over an argument about second-order effects rather than a specific safety incident. Whether it changes anything at Google, or anywhere else, is separate from whether it's a coherent position — and reactions on X split almost exactly along that line, between people who respect the follow-through and people who think a single engineer's exit changes nothing about the industry's trajectory.

Related reading

• Anthropic Researcher Jacob Coxon Resigns Over AI Safety
• Two More Researchers Quit Anthropic and Google Over AI Safety Fears
• Musk, Altman, and Hassabis React: The "Pace the Frontier" Reaction
• Anthropic's R&D Automation Index: Measuring AI Pace
• What Is an Embedded Evaluator? AI Safety, Explained
• Dario Amodei: "Pace the Frontier" and Embedded Evaluators

Sources

• Robert O'Callahan — "Goodbye Google" (official resignation post, September 24, 2026)
• Hacker News discussion
• @rocallahan on X (announcement thread, September 24, 2026)
• Bloomberg — "Google DeepMind Staffer Says AI May 'Kill Us All' in Exit Post" (related mid-September departures)

Details reflect Robert O'Callahan's September 24, 2026 resignation post and subsequent reaction as of September 25, 2026. Quoted concerns and plans are his own stated reasoning, not independently verified predictions about AI risk.