There have been warnings about catastrophic risk from uncontrollable artificial intelligence for the better part of a year now, but the resignation of Anthropic researcher Jacob Coxon — claiming that AI could “kill us all by the end of the decade” — unleashed a flood of built-up concern. Days later, the CEO of Anthropic published an essay warning that, within six months, swarms of AI agents could take over the internet. He called for a slowdown and passage of safety regulations. OpenAI’s CEO publicly agreed with this idea, as did leaders at Google and other AI labs.

The main worry is that, with powerful AI now building the next generation of itself, they may be on the verge of something called “recursive self-improvement.” Multiple AI researchers I’ve spoken with have seen the writing on the wall for months: without a slowdown, the growth in power in these systems will be so dramatic that there is a good chance we will be unable to control them.