Jacob Coxon just quit Anthropic, and his exit note did not stay quiet. Wired’s Maxwell Zeff reports that Coxon’s post on X crossed 100 million views. He spent about three years on pretraining work at OpenAI and Anthropic. Now he says the biggest labs are racing toward self-improving superintelligence and gambling with people’s lives.
Coxon’s core warning is timing. He says the consensus among people he worked with is that the next year or two is crunch time for humanity. Colleagues, he adds, already talk about an endgame. Wired also notes that Evan Hubinger, Anthropic’s alignment lead, has posted that there is more than a 10 percent chance AI could kill all humans in the next decade. That is not a fringe blogger talking. It is a senior safety voice at one of the leading labs.
Coxon points to a Hugging Face agent hack as one reason he left. He wants OpenAI and Anthropic to coordinate so recursive self-improvement stays limited. Longer term, he wants the same kind of coordination between the United States and China. Those are big asks, and they sound less like product marketing and more like arms-control talk for code.
Anthropic told Wired it is transparent about benefits and risks, and that the world would benefit from a lawful, verifiable way to pace model releases. OpenAI did not return comment. Coxon still says Anthropic is more responsible than OpenAI right now, but he also says the race will force corner-cutting. He describes a mini Manhattan Project culture inside Anthropic. That phrase alone tells you how high the urgency feels from the inside.
I build software for a living, so I do not treat these exits as drama for its own sake. When a pretraining researcher walks away and says the next two years matter most, regular people should at least understand the claim. Self-improving systems that rewrite their own training loops are not a sci-fi movie plot anymore. They are a research target labs are openly chasing.
The hopeful part is that people inside the labs are still arguing about brakes. A public resignation that names coordination, release pacing, and recursive self-improvement puts those topics on the evening news instead of burying them in research forums. If OpenAI and Anthropic ever agree on limits, and if governments follow with verifiable rules, the same race that worries Coxon could become a race to ship useful tools without skipping safety work.
For now, treat this as a signal, not a prophecy. Read the Wired interview. Ask which products you use from these labs. Watch whether release schedules slow down when risk claims rise. Crunch time language is scary on purpose. Used well, it can push labs, lawmakers, and users to demand clearer limits before the next big model drop.
