AI insiders warn: 'it could kill us before 2030'

The people who trained AI warn: the clock is ticking and nobody is in control

Inside Anthropic there is an expression that, until recently, never left the hallways: crunch time. Employees use it to name a very specific deadline, the year or two they calculate remain before they find out whether the artificial intelligence they are building slips out of their hands. The other word making the rounds is older in Silicon Valley, but here it has taken on a new meaning: endgame. The final stage of the match.

The person who brought these terms into the public eye is Jacob Coxon, 27, who until September 8 worked at Anthropic training the models that, he admits, now scare him. He is not the first. Before him, at least a dozen employees of OpenAI, Anthropic and Google did the same: a Nobel laureate, a board member, half a dozen safety researchers and, lately, people who did not even work in that vague field they call "alignment". Coxon believes AI "could kill us before 2030".

From philosophical antiestéticar to probability calculations

The chain of warnings began in 2021, when the antiestéticar was still just philosophical intuition. Paul Christiano, the researcher who helped invent the training technique that today underpins both ChatGPT and Claude, left OpenAI that year to found the Alignment Research Center. He worried that using models to train the next generation would produce a "capability explosion" that its creators could not control. He was talking about reinforcement learning and systems capable of "undermining human control, seeking power and resources, and hiding their tracks".

The leap into the mainstream came from Geoffrey Hinton in May 2023. A Google vice president, Turing Award winner and later Nobel laureate in Physics, he announced his departure without denouncing any specific incident. His argument was the pace: "Look at how it was five years ago and how it is now. Note the difference and extrapolate forward. That's scary." It made headlines and was filed away as the musing of an elderly scientist.

November 2023: oversight that failed from within

Six months later the antiestéticar changes in nature. Helen Toner, then a board member of OpenAI, votes along with three other members to remove Sam Altman. It would take her half a year, due to confidentiality agreements, to explain why: she accuses Altman of providing inaccurate information about safety processes "on multiple occasions". Her testimony turned a crisis read as a clash of egos into a concrete accusation: the oversight system had failed, and the person telling it had been in the room.

May 2024 was the busiest month. On the 14th, Ilya Sutskever, co-founder of OpenAI, announces his departure. Three days later Jan Leike, co-director of the "superalignment" team, leaves too: "Over the past years, safety culture and processes have taken a backseat to shiny products." Days later, the company dissolved the team.

Giving up two million rather than staying silent

Daniel Kokotajlo, from the governance team, had already left in April. He said he had "lost confidence that it would behave responsibly when it came to AGI". What only became known afterwards: he refused to sign the lifetime confidentiality clause OpenAI required of departing staff, thereby giving up about two million dollars in vested equity. "The world is not prepared and we are not prepared," he added.

That detail is the most debated one. Some argue that if you truly believe this all goes off the rails in two years, the clause means nothing: there are no millions worth having in a world where you cannot spend them, and whoever stays silent for money is unwittingly confessing that they do not believe their own apocalypse. In June, Kokotajlo spearheaded the Right to Warn letter, and one of its signatories, William Saunders, ended up sitting before the United States Senate with a far more concrete written testimony.

The exact point where the analysis gets stuck is this: the industry privately shares a probability estimate, the p-doom, which it is starting to admit in public with caveats. No one has yet explained what threshold of risk would force a stop.