A global, hardwired pause of frontier AI training is feasible, says new report
A multidisciplinary study led by UC Berkeley scholars outlines a feasible 10-year global pause on frontier AI training through coordinated chip regulation and verification measures.
A new 200-page academic study by 26 researchers from institutions including UC Berkeley, Princeton, Stanford, Harvard, and Oxford confirms the feasibility of a global, mutually verified pause on frontier AI training for at least a decade. The team, combining expertise in computer science, economics, and international relations, proposes halting production of AI training chips while replacing them with inference-only chips that cannot train new models. This approach aims to allow continued use of existing AI systems while preventing the development of more powerful, potentially disruptive models.
The study highlights the concentrated supply chain of AI chips as a strategic advantage for enforcement, noting that bypassing existing production networks is 'extraordinarily difficult.' Researchers argue that an international prohibition on frontier AI training could leverage this hardware dependency to ensure compliance. Verification measures are proposed to detect violations, addressing concerns that rival nations or companies might defect to gain commercial or geopolitical advantages.
The report suggests replacing training-capable chips with inference-only alternatives, which are already being developed for economic reasons but currently lack sufficient incentive due to rapid model turnover. A hardwired pause would increase demand for these chips by removing the economic pressure to constantly train new models, potentially accelerating innovation in inference-only technologies while maintaining access to existing AI capabilities.
Success of the proposed pause hinges on global cooperation among world leaders to retire or transfer pre-pause chip stocks to internationally governed data centers or scientific preserves. The authors draw parallels to historical agreements like nuclear non-proliferation and ozone layer protection, emphasizing that coordination is possible if decision-makers prioritize long-term risks over short-term gains and establish credible monitoring mechanisms.