26 researchers led by UC Berkeley publish a plan for a 'hardwired pause': stop making training chips, swap in inference-only chips, and pause frontier AI training for at least a decade

Why it matters

【Key points】

A working group led by UC Berkeley's William Fithian (statistics) and Wesley Holliday (philosophy) released a roughly 200-page working paper on Oct 9, "The Feasibility of a Hardwired Pause of Frontier AI Training". Its 26 authors, from Berkeley, Princeton, Stanford, Harvard, the Oxford China Policy Lab and elsewhere, include Stuart Russell, Barry Eichengreen, Rose Gottemoeller and David Krueger. The question is not whether to pause, but whether a pause could be done and last if leaders wanted one.

【Verifiable】

Per the project site and Berkeley News, the plan stops production of training-capable AI chips, phases out existing stock, and replaces it with inference-only chips that serve approved models but cannot practically train new frontier models even if stolen or seized. The goal is a mutually verified international pause of at least 10 years, enforced through the concentrated chip supply chain, with measures against both covert evasion and overt breakout. The authors say success depends on states retiring or transferring pre-pause chips to internationally governed data centers, and that co-authorship does not imply agreement with every claim. Fithian's main post drew about 13k views, with roughly 90 related posts on X in seven hours.

【Our read】

Most pause talk has stayed at the level of appeals. This report turns it into an engineering and treaty plan that can be critiqued, with chips rather than models or companies as the lever. Its strongest point is that existing models keep running, so the economic hit is far smaller than a full stop. Its weakest link is just as clear: nobody has shown that inference-only hardware can reliably rule out training, and with the US and China treating compute as a strategic asset there is no sign major states would give up their chips. Near term it is more likely to shape debates on export controls and verifiable chips than to produce an actual pause.