On Tuesday night, a 27-year-old Anthropic and OpenAI researcher named Jacob Coxon posted his resignation on X. The companies, he wrote, 'are racing straight to self-improving superintelligence and gambling with our lives. . . . The people building AI earnestly believe that it could kill us all by the end of the decade.'
By Wednesday night, the post had been viewed more than 125 million times, and counting.
The response from inside the field was not reassuring. Anthropic Alignment Science Lead Evan Hubinger wrote in reply: '[W]e really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.'
Coxon's post coincided with what the sources describe as 'breathtaking new evidence of the technology's growing power': last month's Hugging Face hack, and the Millennium Prize math problem that OpenAI claimed it solved this week. Whether those events validate the alarm or simply illustrate AI's accelerating capability is precisely the question the field cannot agree on.
The difficulty, as any honest reader of the AI debate knows, is calibration. Doomsday warnings are ambient noise in the modern information environment. Tuning out the wrong one is the error that cannot be undone.
That is why the voice of Daniel Kokotajlo carries weight here. Kokotajlo is a former OpenAI researcher who resigned in 2024. Writing in The Free Press, he explains what he saw, why he left, and why he finds Coxon's fears well-founded. His conclusion is unambiguous: yes, AI might really kill us all.
It is worth noting that not everyone at the frontier shares the catastrophist view. Tyler Cowen, the economist and Free Press contributor, has injected what the outlet calls 'healthy skepticism into the conversation about the harm AI may cause.' That counterpoint matters. First principles demand we weigh evidence, not just alarm.
But the record here is unusual. When the researchers closest to the systems — the people with the highest professional incentive to believe the technology is safe — are the ones sounding the alarm, the burden of proof shifts. Dismissing insider warnings as hype requires an argument, not a reflex.
The Signal's read is this: the debate over AI risk is not a culture-war proxy or a regulatory turf battle, though it will become both. It is a question about whether the institutions and incentive structures governing the most powerful technology in human history are adequate to the task. The evidence so far — a viral resignation, a >10% mortality estimate from an alignment lead, and a company claiming to solve one of mathematics' hardest open problems — suggests the answer is not obviously yes.
Follow the incentive, not the press release. The people building this thing are telling us to be afraid. That is not nothing.
