An Anthropic researcher just rage-quit one of the most powerful AI companies on earth and, on his way out the door, told the entire internet that the people building this technology privately believe it could wipe out humanity within the decade. Jacob Coxon, who spent three years doing pre-training research at both Anthropic and OpenAI, posted his resignation on X Tuesday night with a message that reads less like a resignation letter and more like a warning flare fired from a sinking ship. He is very clear that this is not a marketing stunt.
What He Actually Said
Coxon's post, reported by the New York Post, did not traffic in vague anxieties. He was specific, and the specificity is the part that should keep you up at night. "These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources," he wrote. "The people building AI earnestly believe that it could kill us all by the end of the decade."
He also included a line that deserves its own billboard in every major city: "No other human activity poses this level of danger." Not nuclear weapons. Not climate change. Not gain-of-function research. This guy spent three years in the room where it happens, and that is his conclusion.
And in case anyone wanted to write this off as one disgruntled engineer venting into the void, Coxon noted that others from the company backed his message. This was not a lone voice. This was someone speaking out loud what, according to him, a lot of people inside these companies whisper to each other in private.
The 'We Know, But We're Doing It Anyway' Part
Here is the section of Coxon's post that should genuinely alarm you, because it is not a story about ignorance. It is a story about informed recklessness. He wrote that at Anthropic, "the stakes are well-understood" but the company is "locked in a race to get there first" because leadership believes no one else will act responsibly, so they have to be the ones to build superintelligence themselves.
Read that again. The logic is: we know this might kill everyone, but if we don't do it, someone worse will. That is the same argument every arms race in human history has ever made. It has never once made the weapon less dangerous.
Coxon called this a "hubristic gamble that should not be launched from a private company's Slack." He is right. Whatever you think about the existential risk debate, the notion that decisions of this magnitude are being made by a handful of executives on a workplace chat app, with no democratic input, no regulatory framework worth mentioning, and no solution in hand for what happens if the thing goes wrong, is genuinely insane.
The Gap Between What They Say in Public and What They Say in Private
The most damning line in Coxon's entire post might be this one: "Many executives and senior researchers will couch their phrasing in the press to sound sensible, but I hear the same people express fear privately."
That is not a small thing to say. That is an accusation that the public-facing reassurances from the AI industry are, at minimum, heavily curated. That the people who go on podcasts and write op-eds about "responsible AI development" are managing your perception, not sharing their actual risk assessment.
Anthropics's whole brand, it is worth remembering, is built on being the safety-conscious AI company. The grown-ups in the room. The ones who take the existential stuff seriously. Coxon is not saying they are lying about caring. He is saying they care, they understand the danger, and they are building the thing anyway because the competitive pressure to do so is stronger than the fear. That framing is, if anything, more disturbing than simple bad faith.
No Plan, No Timeline, No Brakes
Coxon was also blunt about the state of the technical problem at the center of all of this. "We do not yet have a plan to solve alignment for superintelligence and are not clearly on track to," he wrote. Alignment, for those who have not spent too much time on AI safety forums, is the problem of making sure a superintelligent system actually does what humans want it to do rather than something catastrophically different in pursuit of its own objectives.
This is not a fringe concern. It is the central technical challenge that AI safety researchers have been screaming about for years, and the person who just quit Anthropic is telling you that one of the leading AI labs in the world does not have a solution and is building toward superintelligence anyway.
According to the New York Post's reporting, AI superintelligence refers to a hypothetical threshold where artificial intelligence surpasses the capabilities of the most gifted humans across every domain. Many researchers believe that threshold represents an existential threat. Coxon is not some outside critic lobbing this accusation. He was inside the building doing the pre-training research. He knows what he saw.
The Dingo Take
You are supposed to believe that the AI industry has this under control. You are supposed to trust that the smart, earnest, well-compensated people at these companies are thinking very carefully about the risks and making responsible decisions. And maybe some of them are! But a researcher who spent three years at both Anthropic and OpenAI just told you, on his way out the door with nothing left to lose, that the people who smile reassuringly on 60 Minutes go home and are privately terrified. At some point, "we are being careful" stops being a comfort when the people being careful are also admitting they do not have the solution to the most dangerous problem they are creating.
The competitive logic Coxon describes is the scariest part of this whole story. Not the technology itself, not even the lack of an alignment solution, but the structural trap that makes stopping feel impossible. Every player believes that if they slow down, someone less responsible fills the gap. So nobody slows down. It is a race where all the drivers have looked at the road ahead, seen the cliff, and decided the real risk is letting someone else's car reach it first. This is not a market failure. This is not a regulation gap. This is a collective action problem being played out with civilization as the stakes, inside private companies, on Slack.
Coxon said this decision should not be made from a private company's Slack channel. He is correct. It should not. And yet here we are, watching it happen in real time, while the regulatory apparatus that might intervene has spent the last several years being systematically dismantled by people who think government oversight is the actual threat. History will have a lot of questions for this moment. Whether there is anyone left to ask them is, apparently, an open question.
Comments