Google DeepMind safety researcher quits, says AI 'could kill us all'
The pacing fight just picked up a witness from inside Google's lab, Britain's government put itself on record about listening, and someone finally ran the accounting on what the data centers have to earn back.
Google DeepMind's AGI safety team produced its first public exit note, and it landed in the middle of the industry's loudest argument. Bilal Chughtai, who worked on safety and alignment research at DeepMind until July, posted this week that he "earnestly believe[s]" AI has the potential to kill us all — and that the field is running out of time to stop it. He has since joined BlueDot Impact, the nonprofit that trains people outside the labs to work on AI safety. The two-month gap between his departure and his post is the detail that matters: this is not a grievance story, it is a contribution to the pacing debate, delivered the same week Trump put Jensen Huang on speakerphone onstage to call extinction fears a hoax and Amodei's slowdown proposal became the industry's central fight. We covered the earlier wave of safety departures here — The researchers who left say AI transparency is voluntary — Trump says he has no extinction concerns.
Britain's government left the door open to binding AI rules as OpenAI itself urged London to write them. OpenAI's European policy chief Tom Duff Gordon told the UK to "act now to set up a stronger set of safety rules," with mandatory independent testing and incident reporting for frontier developers; parliament's Joint Committee on Human Rights called this week for a new AI Bill and a statutory regulator, with chair Alex Sobel saying regulators currently cannot block a model they consider unacceptably risky. Secretary of State Louise Haigh is expected today to say ministers must "heed the warnings" from the people building the technology. It is the clearest split yet with Washington, and Labour's pre-election pledge of binding regulation — shelved in favor of voluntary agreements — is suddenly back on the table.
MIT Technology Review put the honest number on the buildout: the hyperscalers need trillions, fast. Wharton's Jessica Wachter, the SEC's former chief economist, starts from spending on AI data centers reaching nearly $1.1 trillion by 2027 and concludes that if earnings don't grow into it, the buildout "will be the largest misallocation of capital in history." Columbia's Stijn Van Nieuwerburgh prices the ask at roughly $41 billion per gigawatt: about 183 gigawatts of planned compute through 2032 means around $3.7 trillion in required annual revenue at a bare 10 percent return. MIT's David Gensler calls it a parlay bet — massive hyperscaler revenues, economy-wide productivity gains, and frontier models fending off cheaper rivals all at once — and says lose any leg and the whole ticket fails.
What to watch: whether Haigh's "heed the warnings" becomes a bill, and whether DeepMind answers the exit note with anything more than silence.
Does a resignation from inside a frontier lab carry more weight than a CEO's essay? Tell us in the comments.
Sources: Bloomberg · The Next Web · City AM · The Independent · MIT Technology Review
