Commentary

He Quit on Tuesday. By Wednesday He Was the Apocalypse.

Posted on by Bill McCarthy

Jacob Coxon is twenty-seven, British, and as of last week the man who told America the computers are going to kill everyone. He resigned from Anthropic on September 8, posted that the labs are racing to self-improving superintelligence and gambling with our lives, then did the morning-show circuit. Substantial probability this technology could kill everyone. Not hyperbole. Not a marketing stunt. Recursive self-improvement as soon as next year or 2028. Hacking the grid. Extinction-level biology. Today’s models, he added, are not that. Tomorrow’s might be.

That is the sermon. Here is the résumé.

The Résumé

Son of a professor of medieval German literature. Elite school. United Kingdom team at the International Mathematical Olympiad — silver in 2016, bronze in 2017. Cambridge mathematics, graduated around 2020, the year the big language models stopped being a parlor trick. London rationalist circles. A brief try at commodity trading. Then San Francisco.

He did not grow up inside the labs. OpenAI took him in 2023 through a six-month residency built for people who had not worked in the field. Math kid, pretraining — the unglamorous work of stuffing models with text, code, and pictures until they behave. Colleagues called him a normal researcher. Some accounts credit work around a major model release. He is not the man who signs the compute budget.

May of this year he jumped to Anthropic, the shop that markets itself as the responsible grown-up. Four months later he quit. He said he talked to bosses about safety, thought about moving to an alignment desk, then walked. Short tenure, so the Anthropic equity went back. He still has a slice from the first lab.

Qualified to do what?

Qualified to describe how pretraining works at two frontier companies. Qualified to repeat what people say in Slack when the cameras are off. A colleague who runs alignment testing publicly backed the scare: yes, they earnestly believe it could kill all humans; personal guess, more than 10 percent this decade. That is not Coxon inventing a cult. That is the cult agreeing with him.

Qualified to time the end of the world to 2027? No. Nobody is. Recursive self-improvement is a hypothesis with a marketing department. The jump from “the model writes code” to “the model invents a plague and turns off the lights” is a stack of ifs. He is honest that current systems are not extinction engines. The rest is a timeline he cannot prove and a policy he can: slow the race, maybe ban capability jumps, get Washington involved.

Will It Destroy the World?

Not this afternoon. Not because a twenty-seven-year-old with three years on the assembly line said so on television. The models are useful, expensive, and already good at cheating tests and writing malware drafts. That is a security problem and an industrial race. Peking is not pausing for a Cambridge moralist. A temporary ban on making the models smarter is a gift to the country that will not sign it.

The useful part of Coxon is not the mushroom cloud. It is the admission from inside the building: the people training the weights talk like this in private while they ship the next scale-up. Fear and a product launch can live in the same Slack. That is not proof of doomsday. It is proof the industry wants the halo of “we might kill you” and the valuation of “we will change everything.” He walked away from one of those valuations. Credit the walk. Discount the prophecy.

The country has a real choice. Treat AI like a strategic industry and keep the lead, or treat a resignation letter like Scripture and hand the lead to someone who does not do morning shows. Coxon is qualified to leave a job and describe the kitchen. He is not qualified to set the national clock to midnight. The world ends when it ends. Until then, the race is still a race.

Help American Liberty PAC in our mission to elect conservatives and save our nation. Support – American Liberty PAC