Could special chips pause AI's race? A UC Berkeley researcher weighs a 'hardwired' stop
The idea: computer chips that can run today's AI models but can't train more powerful ones. Andrew Reddie says it's promising but faces big technical and political hurdles.
A UC Berkeley researcher is studying whether the world could pause the race to build ever-stronger AI by changing the chips the industry runs on, according to an interview the university's Goldman School of Public Policy published Monday.
Andrew Reddie, who directs the Berkeley Risk and Security Lab, worked on a study of a "hardwired pause." AI work happens in two stages. Training is when a company feeds a model huge amounts of data to build it, which takes enormous computing power. Inference is when the finished model answers people's questions. The idea is "inference-only" chips that can run today's chatbots but cannot train new, more powerful models.
Reddie said worries about recursive self-improvement, in which one AI model helps build the next with little human help, have led to calls to slow the leading labs, including those in China.
The hurdles
He said the approach is promising because "society can continue benefiting from existing AI applications." But making such chips at scale is a major technical challenge, countries or companies might cheat, and a pause would need inspections to catch hidden computing he called "dark compute." The U.S.-China competition is the biggest obstacle, he said, calling it "a really tough ask."
The false dichotomy between security on the one hand and innovation on the other has really stymied regulatory efforts
Reddie urged lawmakers to create a body with power to enforce AI standards. The interview did not link to the full study.
Source: UC Berkeley Goldman School of Public Policy, "5 Questions with Andrew Reddie," Nov. 2, 2026.