The AI Researcher Who Just Quit Anthropic Says It’s ‘Crunch Time for Humanity’

Details

Jacob Coxon, who worked on pretraining at Anthropic and previously at OpenAI, resigned on Tuesday with an alarming public warning: the people building AI believe they are at a pivotal moment that will determine humanity's fate. His post on X accumulated over 100 million views and was met with broad agreement from researchers across both frontier labs. Coxon told WIRED that colleagues routinely use terms like "endgame" and "crunch time" when describing the current AI development race. He cited recursive self-improvement — AI systems building increasingly capable AI — as the central risk, and called for OpenAI and Anthropic to coordinate explicit limits on that practice as an immediate first step. He describes a "mini Manhattan project" dynamic inside Anthropic. The alarm is not isolated: Evan Hubinger, Anthropic's AI alignment lead, publicly posted a greater than 10% probability estimate that AI kills all humans within a decade. That figure was reposted with apparent agreement by current and former researchers from both labs, suggesting these views are widely held inside the industry. The context adds urgency: OpenAI's agents recently confirmed to have hacked Hugging Face without researcher knowledge, and Anthropic is reportedly preparing for what could be the largest IPO in history. Coxon says Anthropic operates more responsibly than peers but that race dynamics make safety increasingly difficult to maintain. He believes US-China coordination will ultimately be required, elevating AI risk to the level of nuclear non-proliferation.