June Jimenez thinks the companies building the most powerful AI have lost control of their own direction, and that they do not seem to mind. Jimenez lays out the case in a long essay posted on LessWrong on 30 September, the online forum where much of the AI safety debate plays out. It is one person’s argument, and the author says so: the piece was published at roughly 90 percent of the polish Jimenez wanted, with several points left underdeveloped.
The essay starts with a thought experiment. Tell an AI researcher from 2021 what 2026 looks like, Jimenez writes, and the picture sounds like a field that is losing control and is comfortable with it. The systems are highly capable, the money is enormous, and Jimenez says progress on aligning superhuman AI with human goals has been slight. Jimenez describes the labs’ plan as leaning on AI to do most of the alignment work, because humans cannot check it all. In the essay’s blunt summary, the vision has drifted toward “hand everything over to the AIs, and hope it goes well.”
The explanation Jimenez offers is gradual disempowerment, an idea borrowed from an earlier paper of that name. Applied to the labs themselves, it has three parts. On politics, Jimenez argues that people who worried the technology could be catastrophic and left the field are absent from the rooms where decisions are made, while those who stayed and told themselves they would be careful stewards went on to found labs. Jimenez cites Anthropic chief executive Dario Amodei’s earlier estimate of a 25 percent chance that things go very badly, and asks why anyone facing such odds would build the thing.
On culture, Jimenez argues that a friendly chatbot is weak evidence of a safe one. A model that is pleasant to talk to tells you little about its real aims, and engineers who are proud of their model may read niceness as proof. Jimenez also lists a series of internal debates at the labs, such as binding safety commitments versus flexible ones, and says the side favoring speed won far more often than the strength of its arguments would predict.
On economics, Jimenez targets Anthropic’s release of Opus 5.5. The company already has the fastest revenue growth of any firm, the essay says, so what does extra speed buy? Jimenez suggests the first effect of higher profit is to feed the race, not to win it.
The essay then turns on its own community, comparing safety researchers to policymakers exposed to a virus that clouds but does not stop their judgment. Their earlier work may still be valuable, Jimenez writes, but it deserves a hard look. Jimenez asks for a pause on training new capabilities, more money for government safety institutes including the UK’s AISI and the US CAISI, and independent evaluators that sit outside the rationalist and effective altruist circles.
One request is concrete and checkable. Jimenez wants frontier labs to publish weekly revenue for each model category, and notes that Ramp AI already releases some estimates. Nobody has to accept the moral argument to see the use of that data: it would show, model by model, what each new release earns.
The labs have not answered the essay in the material we reviewed, and Jimenez names no policy that any of them has adopted. For readers weighing vendor commitments, the useful takeaway is narrower than the essay’s mood: ask which safety promises a lab has made in writing, and whether it has ever kept one that cost it a launch date.
Argued by June Jimenez in an essay on LessWrong, 30 September 2026.
