Jacob Coxon spent three years working on pretraining at OpenAI and then Anthropic. This week he quit both his job and the industry, telling the Wall Street Journal that the labs racing to build self-improving AI are gambling with outcomes nobody can walk back. The story broke Tuesday night and was still climbing Hacker News and X early Wednesday morning, roughly ten hours after the Journal's Amrith Ramkumar first reported it.

  • Jacob Coxon, 27, a pretraining researcher who worked at OpenAI before moving to Anthropic, resigned this week and says he is leaving AI research entirely.
  • He told the Journal: "We're on track for a lot of the most aggressive of these scenarios where by the end of next year things could be out of control already."
  • He says colleagues inside Anthropic have started using the words "crunchtime" and "endgame" in internal conversation.
  • On X, he drew a sharper distinction between the two labs: OpenAI staff, in his telling, haven't fully absorbed the stakes; Anthropic's leadership has, but presses ahead anyway because "no one else will act responsibly."

What did Coxon actually say?

Coxon didn't publish a lengthy resignation letter the way some past departures have. His public statement is a Twitter thread from the handle @hilbertspaess, paired with the Journal's reporting. The core claim is blunt: "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives." He argues that near-future systems will be capable enough at hacking, resource acquisition, and scientific work that the label "superhuman" stops being marketing copy and starts being a literal description, and that nothing in current lab behavior suggests that trajectory is slowing down.

RelatedGemini 3.8 Flash Model Card Leaks Ahead of Google Launch

He also draws out a distinction that's easy to miss in the usual "AI lab quits over safety" framing. He isn't accusing Anthropic's leadership of not understanding the risk. His claim is closer to the opposite: that they understand it clearly and race anyway, on the theory that a company more careful than everyone else still loses control of the outcome if the other companies aren't careful too. That's a coordination problem, not an ignorance problem, and it's why he's pushing for cross-lab agreements, potentially including outright bans on certain kinds of capability work, rather than asking any single company to unilaterally slow down.

The race dynamic Coxon describes Two labs, OpenAI and Anthropic, both racing toward self-improving AI systems, with Jacob Coxon exiting the race rather than the target. OPENAI staff, per Coxon: stakes not fully internalized ANTHROPIC leadership: sees risk, races since "no one else will" SELF-IMPROVING AI "out of control by end of next year," Coxon COXON EXITS THE RACE leaves AI research entirely genztech.blog
Fig 1 Coxon's own framing: two labs racing toward the same destination for different reasons, and his exit from that race entirely, not a move to a rival.

Why does this land differently than a typical safety resignation?

Safety-adjacent departures from frontier labs aren't new by September 2026. What makes Coxon's exit notable is who he is and what he isn't doing. He's not a safety-team specialist whose entire job was flagging risk, the way Mrinank Sharma was when he left Anthropic's Safeguards Research team back in February citing similar existential concerns. Coxon worked on pretraining, the unglamorous, extremely central work of actually building the models everyone else evaluates. When someone whose job was making the models bigger and better says the trajectory itself is the problem, that carries different weight than a dedicated red-teamer saying the same thing, if only because it's harder to dismiss as someone talking outside their lane.

He's also not joining a competitor, which is the usual pattern when a researcher leaves one lab "over concerns." He says he's leaving AI research altogether. That's a smaller, quieter signal than a public letter, but it's arguably a more expensive one: no severance-adjacent equity story, no new employer to soften the exit, just someone walking away from three years of specialized expertise in a field paying enormous compensation right now.

What does Coxon actually want to happen?

His proposed fix isn't "AI companies should be more careful," which is the version of this argument that changes nothing because every lab already claims to believe it. He's specifically calling for coordination between labs, up to and including agreed bans on developing certain capabilities, precisely because he thinks unilateral caution doesn't solve a race. If Anthropic slows down and OpenAI doesn't, in his framing, Anthropic just loses the race it thinks it needs to stay in to keep the eventual winner responsible. That argument has been made before in AI policy circles, but it reads differently coming from someone who just quit a pretraining job over it rather than someone writing a think piece from the outside.

RelatedMathematician Says OpenAI Pressured Him Over Navier-Stokes Credit

  1. 2026-02-09Mrinank Sharma resigns as head of Anthropic's Safeguards Research team warns "the world is in peril," cites bioweapons and interconnected risks
  2. 2026-02-11 to 02-12Wave of safety-cited departures across OpenAI, Anthropic and xAI reported by CNN and Axios as researchers "sound the alarm on their way out"
  3. 2026-09-08Jacob Coxon resigns from Anthropic, tells the WSJ he is leaving AI research cites a race toward self-improving systems neither lab can fully control
  4. nextWhether Anthropic or OpenAI respond on the record neither company had issued a public statement as of publication

What it means for the market

Anthropic isn't public and OpenAI's equity trades in a private secondary market, so there's no ticker that moves on a single researcher's exit. But sentiment is part of how both companies raise their next round, and a WSJ-sized story about a pretraining researcher calling the core strategy reckless is the kind of coverage that shows up in due diligence conversations even when it doesn't show up in a stock chart. The more concrete read for investors and enterprise buyers watching this space: it adds to a pattern, not a one-off. A second wave of safety-cited exits in seven months, after February's, is the kind of thing that makes "responsible scaling" marketing harder to sell at face value, and it's worth watching whether it affects how aggressively either lab talks about timelines in their next public update.

What to watch
  • Anthropic's response. The company has built its brand on being the safety-conscious lab; a second high-profile, safety-cited exit inside a year tests that positioning publicly.
  • Whether more researchers follow. Coxon frames this as a pattern, not an isolated call. If a third or fourth departure cites the same "coordination, not caution" argument, it becomes a story about the field, not about individuals.
  • Any concrete coordination proposal. Coxon wants cross-lab agreements on capability limits. Watch for whether any lab, regulator, or industry body puts forward something specific in response, versus more statements of concern.

Our take

The specifics matter more than the vibe here. Coxon isn't claiming a specific model crossed a specific dangerous threshold, and nothing in his statement is a disclosure of a concrete incident. It's a structural argument about incentives, made by someone with an unusually direct view of pretraining decisions at two of the three labs most people would name if asked who's closest to the frontier. That's worth taking seriously without treating it as proof of anything more than what it is: one credible insider's read that the current pace outruns anyone's ability to steer it, delivered by leaving rather than staying to fight from the inside. Whether that's the right call is a genuinely open question. That he felt it was the only honest one is the actual news.

Primary sources

Original analysis by GenZTech, based on Wall Street Journal reporting and Jacob Coxon's public statements. Read the original discussion.