For the past six months, several tech CEOs, AI safety researchers, activists, policy experts and lawmakers have been advocating for an international pause on frontier AI training, in the light of recent autonomous AI agents-led cybersecurity incidents.

But formulating common-ground regulations and agreeing to an international pause on frontier AI training is easier said than done. An ongoing AI dominance race between US and China, emergence of middle powers, a blooming South East Asia chip market, current global unrest and other geo-political issues add complexities to the debate.

However, a new working paper by a team of 26 AI safety researchers tries to offer a solution to this AI quagmire: A Hardwired Pause of Frontier AI Training by controlling the AI chips that power them.

According to the paper titled “The Feasibility of a Hardwired Pause of Frontier AI Training”, an international agreement to pause the training of advanced artificial intelligence (AI) models could remain in place for at least a decade if countries could control the production of AI chips, account for existing computing hardware and develop chips exclusively to run approved models.

The paper led by 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.

According to the researchers such a pause could be implemented now or in the near future if world leaders agreed to pause it. However, its long term success would depend on international cooperation, effective monitoring of AI hardware and the pace of improvements in training efficiency.

The paper explicitly does not assess whether a pause is desirable or whether governments are likely to implement one. Its assessment is conditional on countries being willing to cooperate.

The proposed “hardwired pause” would require participating countries to prohibit the training of new frontier AI models, establish a list of approved models and halt or significantly reduce the production of chips capable of training more powerful AI systems.

“There is increasing interest in an international pause of frontier AI training that is mutually verifiable by participating states. This report studies the feasibility of a hardwired pause, a type of pause in which a prohibition on frontier training is mutually verified through monitoring of and constraints on AI hardware,” read an excerpt from the report.

Under the proposal, countries would also conduct a global census of existing training capable AI chips and gradually replace them with inference-only chips. Inference only AI chips are specialized processors built exclusively to run pre-trained artificial intelligence models and generate real-time predictions or outputs, rather than training models from scratch. According to researchers, the chips would run approved models but would not be practically capable of training new frontier models, even if stolen or seized.

The researchers said the approach could allow countries to continue using AI for commercial services, scientific research and other applications without continuing the race to develop increasingly powerful models.

During the transition, existing training-capable chips could continue providing inference services, subject to verification measures preventing their use for frontier AI training.

The report identifies specialized, model specific integrated circuits (MSICs) as one possible technological foundation for the proposal. These chips can be designed to run particular AI models efficiently, with their architecture and model weights built into the manufacturing process.

The report has argued that monitoring AI hardware could make a coordinated pause more enforceable because frontier-scale AI training requires large amounts of computing power and the AI chip supply chain has several concentrated points of control.

AI FrontPage reached out to David Krueger, one of the co-authors of the working study, who believes a hardwired pause is possible if there is mutual trust between countries when it comes to harms of unregulated frontier AI training.

“Countries need to be confident that their rivals aren’t building more powerful AI in secret, and secure in case a new government decides to rip up the agreement. If nobody has the means to continue frontier AI training, then these worries go away—this is how you get an immediate, indefinite, international moratorium on frontier AI development,” said Krueger, Executive Director, Evitable.

The researchers said computing budgets for frontier AI training had increased by an order of magnitude approximately every 18 months over the preceding five years. They estimated that, without an unforeseen methodological breakthrough, developing a model that significantly advances the frontier after a pause would require a substantial share of global computing capacity for months or years.

This could make covert violations easier to detect than attempts to train smaller models using limited computing resources. However the report distinguishes between covert violations and an overt withdrawal from the agreement. A country could openly abandon the pause and attempt to restart the AI race if it had accumulated enough training-capable chips.

To reduce the risk, participating countries would need to identify existing chip stocks and eventually retire or transfer most of their training-capable hardware outside their territories. The report has identified undeclared existing chips as a major risk for covert violations and commercially retained training capable chips as a key risk for a renewed AI race.

How to Achieve Hardwired Pause on Frontier AI Training?

The researchers identify four conditions for a hardwired pause to remain viable for at least a decade.

First, participating countries must coordinate controls on the manufacture of AI chips capable of training frontier models, preventing new training capacity from entering the system.

Second, they must share information to identify existing training-capable chips and limit the amount of undeclared computing power that could be used to evade the agreement.

Third, manufacturers must develop and produce inference-only chips at sufficient scale to replace training-capable hardware without disrupting access to approved AI services.

Finally, improvements in training efficiency must not advance so quickly that increasingly powerful models can be trained with progressively smaller amounts of computing power, undermining the restrictions.

What about US-China’s Race for AI Dominance?

The report has also focused particularly on the US and China, the two countries central to the current frontier AI race. It has argued that the durability of a pause would depend significantly on whether both the countries could account for existing AI chips and agree to retire or transfer most of their training capable hardware.

“A hardwired pause could also bring economic benefits by incentivizing companies to further reduce inference costs, leading to savings that are at least partially passed on to consumers,” the report said.

The research team includes professors from UC Berkeley and faculty from other top institutions, including Princeton, Stanford, Harvard and Oxford, combining expertise in computer science, economics, history, international relations, and Chinese and American politics with practical experience in diplomacy and government.

Through the report, the researchers’ assessment remains conditional: a decade-long pause would depend on international cooperation, effective accounting of existing hardware, timely production of inference-only chips and limits on how quickly training becomes more efficient. It remains to be seen whether “hardwired regimes” would agree to hardwiring AI chips and a global pause on frontier AI.

Also Read: OpenAI Pauses Frontier RL Training for Models Including Astra Over Safety Risks