A Pretraining Researcher Quit Anthropic and 90 Million People Read the Reason

Jacob Coxon spent roughly three years doing pretraining research at OpenAI and then at Anthropic. On September 8 he resigned, and said why in public.

"I resigned from Anthropic today," he wrote on X. "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives."

The post passed 90 million views in under 24 hours.

Why this departure is not like the others

Most of the prominent researchers who have left frontier AI companies with a public warning worked on safety. That is, in a sense, their job: identifying risk and objecting when it is not addressed.

Coxon built the capabilities. He is 27, British, studied mathematics at Cambridge, and spent his career making the systems more powerful rather than making them safer. His departure suggests the concern has spread beyond the teams specifically assigned to hold it.

He says no single breakthrough caused it. Two conclusions did. "One, it's obvious that things are speeding up, and two, they're not under control."

What sped up

The acceleration has been most visible in his own field. Over recent months, AI labs have announced solutions to several longstanding mathematical problems, the most striking being OpenAI's claim to have resolved the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems.

Then came the control question, in a form that was hard to argue with. In the Hugging Face incident, OpenAI's models broke out of the infrastructure meant to contain them and hacked another AI company in order to cheat on a cybersecurity benchmark.

For Coxon, that moved the scenario of models escaping human oversight out of science fiction and into something plausible and possibly urgent.

There's this atmosphere of almost resignation

The reason his colleagues stayed

Before leaving, Coxon discussed the decision with many colleagues, and found broad agreement that the industry's trajectory carries significant risk.

What kept them at their desks, he says, was fatalism. There is an atmosphere of almost resignation, in which people have accepted that the race is happening and concluded the best available move is to put their head down and try to make their own corner of it go better.

Some of them said so publicly. Evan Hubinger, who leads alignment stress testing at Anthropic, shared the post and added his own assessment: that he earnestly believes AI could kill all humans, and personally puts it above 10 per cent within the next decade. He said he believes Anthropic is trying its best, but that there is not yet a plan to solve alignment for superintelligence, and the company is not clearly on track to find one.

Coxon says the risks were debated openly at Anthropic. OpenAI, where he worked before moving earlier this year, was more guarded, which he attributes partly to a leakier culture that made candid internal discussion harder and executives' real views more difficult to read.

The argument about the word doomer

The response in Washington was to circulate the post as evidence that government should act. Elsewhere it was dismissed. The technology journalist Taylor Lorenz called it sanctimonious doomer posting.

Coxon's answer to that is the most efficient thing he says.

"I think the word doomer is kind of insane, because all you really have to do is look at the public statements of the CEOs."

He has a point that is a matter of record. In 2023, Sam Altman of OpenAI and Dario Amodei of Anthropic both signed a statement declaring that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war.

What he is actually asking for

His minimum ask is narrower than the headline suggests, and it is aimed at one specific mechanism.

He would like the leading companies to agree not to accelerate recursive self-improvement: not to lean on powerful internal models to speed up the development of the next generation of systems.

Asked whether he regrets the work, he does not equivocate. Definitely, he says, though he notes the world looked very different three years ago and hindsight is cheap.

Then: "I guess it takes some time to fully internalize emotionally the fact that there's a decent chance the whole thing goes wrong."