Dario Amodei, a co-founder of Anthropic, among others, has warned that artificial intelligence ( AI ) poses an existential threat to humankind. Others have gone further and argued for pausing the training of new frontier AI models. The counterargument is that regulation, including self-regulation, is unlikely to be effective. And even if a pause is feasible in principle, there is little hope that it could be sustained in practice, given that the United States and China desperately fear falling behind each other.
But a new report by an interdisciplinary group of researchers, of which we are members, concludes that a global pause on frontier AI training is both feasible and sustainable. If implemented appropriately, such a pause could continue for a decade or more, depending on factors such as the pace of gains in AI training efficiency.
Frontier models require AI chips whose supply chain is highly concentrated. A “hardwired pause” would exploit this chokepoint by halting the production of training-capable AI chips. It would allow production of chips which are capable of fast AI inference on approved models but which, owing to their design, cannot be practically used for training new frontier models.
Once production of inference-only chips ramps up, the pre-pause training-capable chips would be phased out, so they could not be used to fuel a renewed AI race. This hardwired pause would allow states to reap the economic benefits of AI inference without continuing AI training that may result in overwhelming disruptions to society and national security.
The supply chain for AI chips is extremely bottlenecked. Advanced lithography ( extreme ultraviolet ) machines are available from just one supplier, the Dutch firm ASML, and ship in containers spread over 20 trucks and three cargo planes. ASML depends on mirrors from a single German high-precision optics firm, Zeiss. Refined rare earths are available from only a limited number of national sources. Other materials, such as high-purity quartz, come from a handful of locations.
As for the use of those chips, an evader using a single computer to train a small AI model may be impossible to detect. But hundreds of thousands of chips would need to be used for months or years at a time to train AI models at a scale significantly beyond that of current frontier models like OpenAI’s Astra. After commandeering such a large hardware fleet, an evader would need to deceive the world for a long period into believing that the fleet was being used for, say, lucrative and computationally intensive inference.
This suggests that a limited number of interventions—including controls on the production and sale of training-capable chips, advanced lithography machines, and rare earths – would go a long way toward supporting a pause in the training of frontier models. This would still allow production of inference-only chips, such as certain model-specific integrated circuits that hardwire the architecture and weights of an approved model into the chip. Model-specific chips already exist, but their economic attractiveness is limited by the high turnover of frontier models. Producing them would become more attractive with a pause on new frontier model training.
Some will object that pausing the production of training-capable chips will slow the advance of AI. That, of course, is the point. Still, with adequate cooperation it could be possible to construct internationally governed scientific preserves where training-capable chips could be used for declared, transparent research workloads.
But could cooperation on a pause, between the US and China and more generally, be sustained over time? States could be tempted to defect from such an agreement, despite the risks, if the potential gains from AI advances are great, and hiding defection could become less difficult as training algorithms grow more efficient.
But decision-makers should understand that they are engaged in a repeated game, that monitoring will occur, and that immediate gains may be swamped by future losses if agreement breaks down. Adequate monitoring mechanisms can limit the risk of defection. The history of nuclear weapons containment suggests that an international agreement is worth trying.
There are caveats, to be sure. Policymakers in the US and China may lack the patience needed to sustain cooperation over time. Given existing stocks of training-capable AI chips, countries might not sustain a pause unless a sufficiently comprehensive census is taken. Chip fabrication would have to be strictly monitored, and developers of frontier models could become privileged monopolists unless regulation prevents them from exploiting their market power.
Whether the development of frontier AI should be paused is for policymakers to decide. But they should recognize that a pause is feasible and, if implemented properly, that it is likely to be sustainable.
Barry Eichengreen is a professor of economics and political science at the University of California, Berkeley, and a former senior policy adviser at the International Monetary Fund; Will Fithian is an associate professor of statistics at the University of California, Berkeley; and Wesley Holliday is a professor of philosophy and the chair of the Group in Logic and the Methodology of Science at the University of California, Berkeley.
Copyright: Project Syndicate
