The Word for Exit

Resignations in this industry usually arrive as a press release. Jacob Coxon’s arrived as an indictment. On September 8 the Anthropic researcher — formerly of OpenAI — quit and published a public statement saying the frontier labs are “gambling with our lives,” racing “straight to self-improving superintelligence” from inside private companies that answer to no one but their own roadmaps.

The specifics are what make it hard to file away. Coxon wrote that neither Anthropic nor OpenAI is acting responsibly, that the systems being built could soon “hack anything, revolutionise any field overnight, and acquire real power and resources,” and that many of the people building them privately believe AI could “kill us all by the end of the decade” even while they speak more cautiously in public. He called the enterprise “a hubristic gamble” that should not be launched from a private company’s Slack, and he urged other staffers to reconsider their work.

The timing is not incidental. Coxon’s exit lands the same day OpenAI claimed a Lean-verified Navier–Stokes blow-up proof from an unreleased internal model, and four days after this page carried Anthropic’s own machine-checked formalization of Fermat’s Last Theorem. The capability story and the conscience story are now the same story told from two desks in the same building. The labs say the acceleration is the safety plan — build the thing that can prove theorems, cure disease, and align its successors. Coxon’s letter is the internal counterargument, published.

He is not the first to walk, and the exits have a shape. Mrinank Sharma left Anthropic in February writing that “the world is in peril.” The difference is register: Sharma’s letter was a warning from someone who still believed the institutions could be moved. Coxon’s reads like an assessment after the motion failed — less plea, more autopsy. When the people whose entire job is estimating what these systems can do start describing the race in actuarial language, the letters stop being personnel news and start being evidence.

Anthropic’s position, for the record, is the Responsible Scaling Policy — capability thresholds, security standards, the framework the company markets as the industry’s most serious. The company did not dispute the letter’s account so much as decline to race it. But the letter’s real target is the premise underneath every scaling announcement this year: that the people doing the building are the right judges of when to stop. Coxon’s answer is that they have already answered that question wrong once, quietly, and the rest of us only get to vote with our attention.

What to watch is who signs next. Resignation letters are a lagging indicator — they mark the moment the argument inside a lab loses its last institutional buyer. The frontier’s public posture is unchanged: ship faster, prove more, disclose on schedule. The private posture, per the people leaving, is a bet. This week the house wrote its odds down where everyone could read them.