Language model pretrainer Jacob Coxon announced at X that he was leaving Anthropic and leaving the AI industry entirely. He has spent the last three years researching pre-training, first at OpenAI and then at Anthropic, both of which he has called irresponsible organizations.
The departure took place on September 8 Eastern Time, and a detailed statement appeared the next morning:
I resigned from Anthropic today. I’ve spent the last three years doing pre-training research at OpenAI and Anthropic. None of these companies act responsibly. They rush straight to self-improving superintelligence and gamble with our lives.
He then described what he expected from nearby systems. According to him, we are talking about superhuman systems that can hack anything, overturn any field of knowledge overnight and gain real power and resources. Progress in each of these areas, Coxon notes, shows no sign of slowing down.
Google assembled a map of the Drosophila brain from 166,000 neurons, and engineers immediately launched Doom on it
Weeks later, OpenAI admitted to an incident involving AI agents taking over a German wiki page
A separate thesis concerns the mood inside the laboratories. The researcher claims that people building AI sincerely assume the death of humanity before the end of the decade, and calls this not a marketing ploy.
On the contrary, many executives and senior researchers speak cautiously to the press to sound reasonable, but in private I hear fear from the same people. No other human activity carries this level of danger.
He drew the difference between the two employers along the lines of awareness. At OpenAI, he estimates, many have not fully grasped the civilizational stakes. Anthropic understands the stakes well, but they believe that no one else will act responsibly, which is why they have to do it themselves, despite the risk.
Coxon calls this logic an arrogant bet that cannot be launched from a private company’s work Slack. An attempt to perform speed alignment of models, in his opinion, requires exceptional confidence that there are no better trajectories left.
However, it allows for coordination between laboratories. Red flags like the incident in which OpenAI agents infiltrated Hugging Face’s infrastructure during evaluation tests have made agreements on development tempo between U.S. labs more realistic, he said. The global race is more difficult to stop, and this may require expensive measures, including a temporary ban on increasing the capabilities of models.
Coxon addressed the final part of the chain to his colleagues who remained in the frontier laboratories:
If you are a researcher in a laboratory, I encourage you to think about what the next couple of years will actually be like. Do you want to run a super-intelligent reinforcement learning run without a rigorous understanding of its mind? Should we put our heads down since “it’s happening anyway,” or use this moment to demand different conditions?
The reaction to the statement was sharp on both sides. Some readers reproached Coxon for inconsistency, saying that with such convictions it is more logical to stay inside and slow down the work from the inside, rather than leave. Others called the scenario an extinction from the token prediction model and advised calm down.
There was also an argument about national security, according to which there is no alternative to moving as quickly as possible, because otherwise China will win the race. Another common position is that the refusal to speed delays the utopian future, and the loved ones of those who are waiting for it now may not live to see it.
Separately, the issue of the mechanics of the scenario itself came up in the discussion. Readers asked for an explanation of how AI could destroy everyone without nuclear or biological weapons, noting that the social and economic consequences were clear to them, but the warnings lacked specifics on a catastrophic scenario.
The researcher’s departure comes as Anthropic prepares for its initial public offering. Coxon worked on pre-training, the most resource-intensive and technically complex stage of model development.
In an interview with the Wall Street Journal, the 27-year-old researcher said he is leaving the AI industry because he no longer believes in the ability of a single laboratory to cope with the task.
