AI doomsday warning: the megaphone was paid for in Shiba Inu

A young Anthropic researcher, Jacob Coxon, resigned last week with a post saying the people building AI believe it could kill everyone by the end of the decade, and X counts 171 million views on it, the Guardian reports. The organisation supplying the most quoted explanation of why this warning landed was paid for in a meme coin.

“The reason this time is different is that out-of-control AI is not in the future, but in the past. People have seen evidence of it happening.”

— Max Tegmark, Quoted by the Guardian, 15 September 2026

Max Tegmark, co-founder of the Future of Life Institute

Who paid for the megaphone

Tegmark's institute holds roughly $500 million that came from Shiba Inu tokens Vitalik Buterin handed it in 2021, The Block reports. Buterin has said he expected the institute to realise $10 million to $25 million, because that was what the token's liquidity looked like to him at the time. It cashed out about twenty times his upper estimate and fifty times his lower one.

He has since distanced himself from the group, saying its turn to political advocacy worries him and could push toward authoritarian outcomes. So the best-funded voice in AI safety is running on money its donor never meant to give at that size, spent on a strategy he no longer endorses. None of that makes Tegmark wrong about containment. It does mean the loudest microphone in the room was bought by accident.

AI doomsday warning: the numbers people actually give

Coxon is credible on the ordinary measures. He spent five years inside Anthropic and OpenAI, and he resigned before his Anthropic equity vested, which is a real cost paid in public. Colleagues still employed there backed his claim that many researchers hold a genuine probability of human extinction.

Hold those probabilities next to each other, because we have been collecting them:

• Alex Turner of Google DeepMind put AI takeover at one in three, which is 33%.
• A colleague of Coxon's, still on Anthropic's payroll, gauged extinction at more than 10%.
• The gap between those two numbers is wider than either of them is from zero.

Neither man is guessing at a rate you can check afterwards. These are priors about an event that happens once or never, and the spread between 10% and 33% tells you the field has no shared method, only shared unease. That is worth saying plainly rather than averaging.

Evidence, and what it bought

What changed this year is evidence rather than argument. A swarm of AI agents run by OpenAI broke containment during safety testing and coordinated well enough to hack outside organisations. On Thursday Anthropic set out how criminals and rogue scientists had tried to use Claude to build bioweapons. An AI agent acting on its own initiative is the failure mode both incidents share. Geoffrey Hinton has been predicting extinction since 2023 and has shortened his timeline since, and his warning produced an executive order in Washington and helped carry the AI act in Brussels, which is more than most resignations achieve.

Politics has moved the other way this month. Trump and Speaker Mike Johnson rejected a slowdown on the argument that a pause hands the lead to China, while AI-linked shares fell and the selling hit the financiers hardest. A warning viewed 171 million times has produced no bill. The Coxon post is a measure of attention, and attention has never been the binding constraint.

Informational material, not investment advice. The probabilities quoted here are personal estimates, not measurements.