The Bunkers Are Coming

Some of the people building artificial intelligence are preparing for it to go wrong. “There’s a fair amount of [disaster] prepping and bio-bunker purchasing going on,” Edouard Harris, chief technology officer of the AI-risk firm Gladstone AI, told TIME, describing researchers inside the top AI companies. He did not name them, TIME did not either, and no lab has confirmed it.

What has them preparing is a forecast. In a report published Wednesday, Gladstone says most of the insiders it spoke with believe AI will begin improving itself “within the next 3 to 12 months, and possibly sooner, assuming there’s no U.S. government intervention or slowdown.” Researchers call that point recursive self-improvement: AI doing the work of building the next AI.

The report, shared exclusively with TIME before publication, says the labs are “actively working on” automating the jobs human researchers do now: generating the ideas, writing the code, running the experiments and interpreting the results. Gladstone says it worked with researchers at Anthropic, OpenAI and Google DeepMind, quoted only anonymously. TIME describes the sourcing as conversations with half a dozen researchers inside the top companies.

One brake remains: computing power. Labs try new ideas on small amounts of compute and scale up only the ones that keep working, so if sorting good AI-generated ideas from bad ones costs too much compute, it may still make sense to rely on human researchers. The companies with the most compute can test the most ideas, and may find gains that only show up at larger scale.

Gladstone’s CEO, Jeremie Harris, Edouard’s brother, says the point is to give lawmakers something concrete. “Our goal here was to define recursive self-improvement well enough that people can have a sane conversation about it,” he told TIME. Right now, he said, legislators have “essentially zero visibility” into whether AI is helping find the breakthroughs.

The NewsCenter Take

This week NewsCenter reported on AI that could copy itself across cheap devices and learn as it spreads and on reports that OpenAI’s next model may do more of its reasoning where no one can read it. Put the three together and the shape is plain: AI may soon be building AI, and the thinking behind it is getting harder to see. The bunkers are one man’s account. The timeline comes from inside the labs.