New report says a global hard pause on frontier AI training is feasible

26 researchers from UC Berkeley and other institutions released a 200-page report discussing a feasible plan for a global pause of at least 10 years on frontier AI training. The researchers recommend halting the production of training chips and gradually decommissioning most or all existing training chips, replacing them with dedicated inference-only chips. The report states that such chips can efficiently run already-approved existing models, but due to hardware constraints, even if stolen or seized, they would be difficult to actually use for training new frontier models. The report also proposes international verification measures to monitor the use of residual training chips, and to leverage the concentrated nature of the chip supply chain to enforce production restrictions.