Jacob Coxon quit Anthropic last week and took to X with a warning that the big labs are racing toward self-improving systems that could wipe out humanity. Ten percent chance of extinction by 2035, he said. OpenAI and Anthropic are both “gambling with our lives.” Dario Amodei, Anthropic’s CEO, piled on days later with his own call to slow the whole thing down. Congress suddenly wants hearings and a select committee. The timing is convenient.
Coxon is 27. He spent time at OpenAI before moving to Anthropic. Four months in and he walked. The media coverage has treated him like a lone truth-teller who finally said what the suits won’t. That framing ignores how little leverage a mid-level researcher actually has inside these places and how quickly the story moved from one guy’s posts to coordinated regulatory chatter. Low-level employees do not normally drive global policy narratives this fast unless someone upstairs finds the noise useful.
Anthropic is prepping for an IPO that analysts have floated as high as two trillion dollars. The company spent years selling itself as the safety-first alternative to OpenAI’s cowboy culture. Now the safety narrative is working overtime right as the S-1 papers are in motion. Every headline about “existential risk” keeps the valuation conversation alive and gives the company a ready-made reason to demand favorable rules before it goes public. The danger talk functions as both marketing and moat-building.
Regulation always sounds noble until you watch who writes the rules. Compliance costs, licensing regimes, and mandatory safety reviews crush small teams and open-source efforts while the giants with legal departments the size of mid-sized cities sail through. Anthropic and its peers already sit on the compute, the data, and the political access. New barriers simply lock the door behind them. The same pattern played out in finance, pharma, and telecom. The incumbents end up writing the fine print.
Donald Trump has been blunt about it. He called the sudden alarmism a hoax, said whoever wins AI wins, and brushed off calls for new guardrails. He has no interest in slowing American labs while China keeps moving. That stance puts him at odds with the Amodei-Coxon chorus and the lawmakers now echoing them.
Matt Walsh and Andrew Torba have both pushed back hard on the same story. Walsh has treated the extinction talk as another elite overreach dressed up as concern for humanity. Torba has pointed out that the people demanding oversight are the same ones who spent years pushing content controls and deplatforming on their own platforms. Both see the regulatory push as less about stopping Skynet and more about concentrating power where it already sits.
The underlying technology does not support the apocalyptic claims anyway. Current large language models are sophisticated statistical engines. They predict the next token based on patterns in training data. They do not reason, maintain persistent goals, or improve themselves without constant human intervention and massive new compute. They remain tools. Impressive ones, but tools. The jump from today’s systems to genuine artificial general intelligence with agency and self-modification is not a matter of scaling the same architecture a few more orders of magnitude. The gap is qualitative, not quantitative. Treating autocomplete on steroids as an extinction-level event flattens the actual technical picture into something that serves the funding and regulatory goals of the people already at the top.
The sudden alignment between a departing researcher, a CEO preparing to cash out, and politicians looking for a new issue to own is worth watching more closely than the headlines suggest. The danger is real, but it is not the one they are selling.
