If the world's government and AI leaders ever agree to pause frontier AI training, is an international pact really feasible? How would it be implemented, coordinated and monitored?
In a first-of-its-kind academic study released today (Friday, Oct. 9), a multidisciplinary team of 26 researchers maps out a path toward an international, mutually verified pause of frontier AI training of at least 10 years. The team - which includes faculty from UC Berkeley and other top institutions, including Princeton, Stanford, Harvard and Oxford - combines expertise in computer science, economics, history, international relations, and Chinese and American politics with practical experience in diplomacy and government.
The researchers focus primarily on a plan that halts the production of AI training chips and replaces them over time with specialized "inference-only" chips, which serve fast, efficient inference on existing AI models but are not practically useful for training new models. Such a plan would permit the continued use of currently available AI products, while preventing new AI models from overwhelming states with social or economic disruption or threats to their national security. They also suggest verification measures that would allow world leaders to assure nervous rivals that they are upholding the pause agreement.
"We consider a scenario in which world leaders want to pause if they can be confident that others are pausing or that defections would be detected quickly enough," writes co-organizer Wesley Holliday, a UC Berkeley philosophy professor, in the study.
Photo by Jen Siska
Frontier AI training refers to the development of ever more powerful AI models. In a hardwired pause of such training, societies could control how powerful these models are allowed to get. If they judge that certain models can offer economic and scientific benefits, a hardwired pause could give them a way to reap the benefits before they become powerful enough to pose catastrophic risks, thus avoiding the dangers of an uncontrolled AI race.
The rapid progress of frontier AI systems, combined with leading AI companies' recent struggles to prevent their models from escaping containment and committing cybercrimes, has prompted widespread calls for a global pause until these systems can be built safely and controllably. For strategic reasons, however, a long-term pause could be difficult to arrange even if parties agree that it would be in their interests: without effective verification measures, rival companies or nations may fear that others will defect and thereby put them at a commercial or geopolitical disadvantage.
But the Working Group on AI Pause Feasibility's detailed, 200-page report addresses such roadblocks with a framework for mutual verification.
Stopping, replacing, retiring AI training chips is key
Frontier AI training requires AI chips, which states can govern. The supply chain for AI chips is highly concentrated with several chokepoints, bolstering governments' ability to enforce rules on their production.
"Bypassing existing supply chains is extraordinarily difficult," writes study co-organizer Will Fithian, UC Berkeley associate professor of statistics. "An international prohibition on frontier AI training can take advantage of this fact about the hardware needed for AI training."
The report proposes halting the production of AI training chips, which are needed to develop more powerful AI models. Over time, most or all of the pre-pause stock of such chips would be phased out. They would be replaced by inference-only AI chips, which are capable of fast, energy-efficient AI inference on approved models but are not practical for training new frontier AI models - even if stolen or seized - due to constraints hardwired into the chips.
During the chip replacement process, other verification measures would prevent residual training-capable AI chips from being used for frontier AI training.
Fithian and Holliday add that technologies needed for inference-only AI chips already are being developed for purely economic reasons, but that their economic attractiveness is currently limited by the rapid turnover of new models. If the training of new frontier models were prohibited, a hardwired pause would increase the economic advantages of these specialized chips and spur innovation in the development of other inference-only chips.
Other highlights of the report include:
- How a hardwired pause could be made politically palatable by offering ways to pursue AI-powered scientific and medical research, retain strategic insurance against being left behind in the AI race by defecting states, and pass on the economic benefits of cheaper inference to consumers.
- The new chips' reduced resource demands, which would decrease their environmental impact and improve frictions in the U.S. between data center operators and local communities.
- How a hardwired pause safeguards against both covert evasion - when a state secretly trains new frontier models while ostensibly adhering to verification procedures - and overt breakout - when a state openly abrogates the pause in an attempt to build a much more powerful frontier model before other states could respond.
- The researchers agree that a hardwired pause's success depends on the cooperation of world leaders and their willingness to retire or transfer pre-pause stock to other jurisdictions or internationally governed data centers, such as scientific preserves.
According to the report, one of the most significant questions of the 21st century is whether humanity can unite to regulate AI's unique "dual-use dynamics" - its abilities to benefit or harm. It adds that precedents do exist: an international agreement to control nuclear energy's military uses while promoting peaceful uses, and another to protect Earth's ozone layer.
"Coordination is possible if the decision makers care enough about the future," the authors write, "not just immediate payoffs, and if monitoring is adequate to support credible consequences for defections."
