Google DeepMind Engineer Quits, Saying His Chip Work Could Speed Up AI

The work Robert O’Callahan once considered relatively harmless now looks to him like a way to speed up AI. The Google DeepMind engineer has resigned after concluding that his chip-design tools could help make AI cheaper, faster and more capable. His decision turns on a difficult judgment: how much certainty about future harm should an engineer need before stepping away?

The work behind the models

O’Callahan worked on tools for designing hardware chips, not on building AI models. For a time, he told himself that distinction made his job relatively harmless. He later concluded that the tools could accelerate the design of AI chips, a contribution he could no longer justify.

His objection is partly about pace. In a resignation email shared on his blog, O’Callahan called AI progress “far too rapid.” He said his team’s goal of making AI much cheaper and quicker to use was not good for people right now.

Leaving carried a personal cost. O’Callahan described the job as an “amazing” chance to be well paid for solving interesting problems. He said a transfer was impractical because Google had no other engineering teams in New Zealand, where he was based.

aiming for ASI in the near future is inherently irresponsible.

Robert O’Callahan, as quoted by The Decoder

A warning without a certainty claim

ASI means artificial superintelligence. O’Callahan calls its risk “real but uncertain”; he does not say catastrophe is inevitable. He linked to the book If Anyone Builds It without endorsing its claim that superintelligent AI must bring catastrophe. His case for leaving rests on taking a possible danger seriously before its outcome is known.

His concerns extend beyond that worst-case scenario. O’Callahan warned about people surrendering their thinking to AI, AI-related psychosis and loneliness, concentrated power, economic disruption, cybersecurity threats and weak accountability.

O’Callahan said many colleagues at DeepMind share his concerns but rarely speak publicly. That is his account of their views, rather than a public position from those colleagues. He also said regulation is “desperately needed.”

Leaving DeepMind, not leaving AI

O’Callahan does not plan to abandon AI-related work. He intends to keep maintaining the debugging tools rr and Pernosco and to study how AI agents debug code. He says he wants his next project to be “unambiguously pro-human.” For him, the line is not between touching AI and avoiding it altogether, but between kinds of work he can and cannot defend.

He expects his resignation to have a very small effect, though he argues it will not have none because some of his skills are rare. Whether others who build tools that could speed AI development would make the same choice remains an open question.