Researchers outline options for slowing AI, but say no clear plan exists

In short: More AI leaders are talking about slowing down advanced AI, but researchers say there is still no agreed, reliable way to do it or enforce it.

What's going on

A new report, Pacing the Frontier, A Research Agenda, argues that “slowing down AI development remains an unsolved puzzle,” according to coauthor Raymond Douglas of the University of Toronto. The debate has intensified after recent public warnings from people inside AI labs about the risks of very powerful systems.

Several well-known AI executives, including leaders at Anthropic and OpenAI, have voiced support for some kind of slowdown or pause. One reason is that AI companies are starting to use AI tools to help build newer AI models. Some researchers worry this could create a feedback loop where progress speeds up and becomes harder for people to monitor, like a car that keeps pressing its own gas pedal.

The Wired report describes a range of possible controls. One option is stronger third-party testing, where outside groups try to find dangerous behavior in a model before release. Another idea is to track “compute,” meaning the raw computing power used to train big models, similar to tracking electricity use to spot a factory running at full capacity. More far-reaching proposals include adding tamper-resistant logging or even remote shut-off mechanisms to AI chips, and international agreements to limit the hardware used to train the biggest models.

What to watch

Even if major companies agree to slow down, enforcement is the hard part. Researchers say the next fights will be about who gets to inspect systems, what data they can see, and whether governments and other countries, including China, will cooperate or try to race ahead.

Source: Wired