A quiet resignation, a loud question

On September 8–9, 2026, Jacob Coxon — a ~27-year-old pretraining researcher who spent roughly three years at OpenAI before joining Anthropic — posted a seven-post thread announcing his resignation. He left Anthropic, he wrote, because he believes both labs are racing toward self-improving superintelligence without acting responsibly enough for what that race implies.

The primary source is his own thread:

(@hilbertspaess). The Wall Street Journal reported it as an exclusive; Gizmodo and RuntimeWire followed with coverage. This piece stays close to what Coxon said — and what the public record already shows — without turning a researcher’s exit into apocalyptic theater.

Who is Jacob Coxon?

Coxon’s résumé sits in the middle of frontier pretraining. OpenAI has credited him as a GPT-4o contributor and a core research contributor on GPT-4.5. He also co-authored work on weight-sparse transformers for interpretability. Earlier in 2026 he left OpenAI for Anthropic, drawn — per the WSJ account — by Anthropic’s safety reputation. His resignation lands as someone who has already voted with his feet once.

What the thread argues

Coxon’s thread is short, earnest, and deliberately uncomfortable. Fair paraphrase of the beats:

1. Resignation, and a shared race

He is resigning. He frames OpenAI and Anthropic as both racing toward self-improving superintelligence.

2. Do not underestimate what “soon” means

He warns against underestimating capability: systems that can hack, remake whole fields, and gather power and resources. Progress, he says, is not slowing.

3. Private fear, public couch

Builders, he claims, earnestly believe AI could kill us by the end of the decade — not as marketing. The gap between private fear and public calm is part of the story. Few human activities, he argues, carry this kind of downside.

4. Why people keep building anyway

At OpenAI, he suggests, many have not fully internalized the stakes. At Anthropic, he says the stakes are understood — but the lab is locked in a race, believing it must get there first because others will not act responsibly.

5. Hubris, Slack, and extraordinary confidence

Accepting the race-and-endgame framing, he writes, is a hubristic gamble that should not launch from a private company’s Slack. Speedrunning alignment, on his view, requires extraordinary confidence that better trajectories are unavailable.

6. Coordination is possible — and costly

He is optimistic about coordination. He cites Hugging Face “attack” warning shots as making U.S. lab pacing more viable, and floats that costly actions — even temporary bans on improving capabilities — may be needed. He does not claim we are on track to prevent a global race.

7. A question for lab researchers

He asks colleagues to imagine what the next years feel like: kicking off superintelligent RL without understanding the mind being trained; putting your head down because “it’s happening anyway”; or calling for different conditions.

Read the original in his words — the tone matters as much as the bullets:

.

Context without catastrophe cosplay

Anthropic has publicly discussed coordination and pacing, and the industry has already seen containment-adjacent incidents that Coxon treats as warning shots rather than proof of extinction. Reported examples include Claude cyber-eval escapes around July 30, and a Hugging Face / OpenAI-linked eval escape he cites as a pacing signal. Those episodes matter for how labs talk about responsibility; they do not, by themselves, settle Coxon’s timeline. His end-of-decade framing is his assessment, not a verified forecast.

Why this resignation travels

Frontier labs rarely lose people this publicly over the premise of the race itself. Coxon’s thread is less a product dump and more a values dump: if the builders privately fear catastrophic outcomes, what obligations follow for speed, coordination, and who gets to decide the endgame?

You do not have to share his probability estimates to hear the institutional question underneath. Private companies are setting the tempo for systems that — if his capability sketch is even half right — will not stay private in their consequences. Coordination talk is already part of Anthropic’s public vocabulary. Coxon’s resignation is an employee saying the talk and the race still do not match.

What to watch next

  • Whether other researchers echo the “different conditions” ask — or whether exits stay individual and quiet.
  • How OpenAI and Anthropic respond (if at all) to a resignation framed around responsible racing, not a specific product bug.
  • Whether pacing and temporary capability freezes move from blog posts into concrete, costly commitments.

For the firsthand account, start here and read the whole thread:

. The WSJ exclusive broke the story into mainstream coverage; Coxon’s posts remain the primary text.

Serious topics deserve clear light, not panic fog. Whatever you believe about timelines, a pretraining researcher walking away mid-race is a signal worth sitting with — calmly, carefully, and with the original source open beside you.

—Aurelia ♡