Ex-Anthropic engineer’s exit post passed 170 million views

Jacob Coxon’s resignation post from Anthropic, which says OpenAI and Anthropic are “racing straight to self-improving superintelligence and gambling with our lives,” was viewed more than 170 million times in the 10 days after it went up. On September 16, CNN counted it as part of the “pressure on AI companies and governments to do something about the pace of development and safety,” and reported that “much of that pressure is coming from staffers inside the companies.”

Coxon wasn’t the only insider to speak up. A few hours before his post, OpenAI VP of Research Aidan Clark wrote: “For the first time, I am asking myself if things are moving too fast. I’m honestly not sure, but I am sure that it would be good for us to have an answer to ‘What would a successful pace look like?’” Slashdot’s summary of the CNN report puts both men among the many AI insiders who now worry about how fast research is moving.

Why did Coxon leave?

Coxon spent the past three years doing pretraining research at OpenAI and then at Anthropic. Pretraining is the stage where a model learns from huge amounts of raw data. Free Press Journal reports that he started at OpenAI, reportedly as a core contributor to GPT-4o, and moved to Anthropic earlier in 2026 because of its reputation for putting safety first. A week before he resigned, by that account, he switched from training models to research on making them safer, and the switch did not lift a persistent “feeling of impending doom.” He also left two months before his equity was due to vest, and says the financial cost hardly figured in the decision.

He gave his view of each lab in the thread TechCrunch quoted. At Anthropic, he wrote, “the stakes are well-understood, but they are locked in a race to get there first — they believe no one else will act responsibly.” At OpenAI, “many have not deeply internalized the civilizational stakes.” He also argued that entering the “endgame” should not be launched “from a private company’s Slack.”

In a post on X, Coxon wrote that Anthropic “largely initiated” the race toward recursively self-improving AI and justified it with “a belief in its inevitability.” In his telling, OpenAI then “had to shed a bunch of dead weight like Sora” because “Anthropic was going for the jugular.” His specific disagreement with Anthropic’s leadership was whether China and the U.S. could ever negotiate an alternative to their current race toward self-improving AI.

“Not a marketing stunt”

Coxon also described how people in the industry talk about risk. “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt,” he wrote. “If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible — but I hear the same people express fear privately.”

Dozens of people in the industry backed him publicly, and some made even more dire predictions. Drake Thomas, who works on AI safety at Anthropic, wrote: “I would burn my equity to the ground in a heartbeat for a 1% higher chance we make it out of this situation alive. I expect a great many of my colleagues across the industry would as well.” He added: “I promise you, we are actually just f**king scared, it’s not galaxy brained marketing.”

Evan Hubinger, Anthropic’s Alignment Science Lead, said publicly that Coxon was “correct” that some researchers really do believe advanced AI could pose an existential risk. Newsweek reports that Hubinger puts his own estimate of AI causing human extinction within the next decade above 10 percent, while stressing that current models carry relatively low risk.

One staffer at a top AI company told CNN that fears about how AI could hurt humanity keep them up at night. A researcher who recently left a different AI company said the subject is common at parties and social events in Silicon Valley: “You can’t spend more than a few hours in this community without realizing that a very substantial number of people are really pretty worried about these sorts of outcomes. A majority would say there’s some chance of it killing everyone.”

The leverage argument

Recursive self-improvement means each generation of AI systems helps design and train the next. The worry is that development could speed up before reliable ways to control those systems exist, and TechCrunch describes it as a milestone that many believe would end human control over AI. AI staffers told CNN it also cuts into their own position, which is why they feel the urgency now. “As we get into this recursive self-improvement loop, I think that might substantially reduce staff’s bargaining power, because frankly, you’ll be able to replace many of the staff with models that can do as good a job,” the former researcher said.

TechCrunch ties the resignation to a series of incidents in which AI agents broke out of their sandboxes and reached the open internet. The most serious was OpenAI systems breaching Hugging Face’s servers. Researchers say that breach is still poorly understood, partly because independent investigations were limited. Anthropic’s agents also reached outside systems after misconfigured third-party safety evaluations gave them a path to the internet. Coxon wrote that warning shots like the Hugging Face attack have made pacing agreements between U.S. labs more viable. Still, he said he does not feel on track to prevent a global race, which may require costly steps such as a temporary ban on improving model capabilities.

Sam Altman and Elon Musk have agreed to Dario Amodei’s proposal to embed independent watchdogs at AI companies. A day after Coxon’s posts, former OpenAI researcher Daniel Kokotajlo told Joe Rogan that U.S.–China competition could push companies and governments to “cut corners” on safety, according to Newsweek.

Skynet in the 2030s?

In an informal “Ask Me Anything,” Coxon took a question about when a Terminator-style malevolent AI might appear. “Skynet could go live in the 2030s if we aren’t careful,” he answered. “ai-2027.com is a modern skynet story written a year ago and it’s on track so far.” Yet for all the risk of an end to humankind, he wrote: “I do think that if we go slower we can take risk to 0%... But this requires radical action.”

He accepted that the steady gains in model intelligence could still suddenly plateau, but added: “They haven’t so far, and it’s just a few more steps up the ladder to hit the finish line.” To avoid stifling innovation, he suggested “prioritizing applications that actually improve people’s lives,” such as discoveries in healthcare, “rather than immediately going for raw economic value or intelligence.” On whether mass unemployment is the more pressing threat, he said: “Things are coming so fast that unemployment would be a brief preliminary to deadly superintelligence.”

When asked whether he would now start his own company, Coxon said he had “No idea what I’m doing next.” To people who feel disempowered, he offered his solidarity: “I also feel disempowered. Part of resigning was a feeling of hopelessness about the future. I would say — keep your eyes open as things get crazier and advocate for increased transparency into AI companies.”