Regulating access to computational power and expanding know-your-customer requirements forces compliance.

At present, the most advanced AI systems are developed through training that requires an enormous amount of computational power – ‘compute’ for short. The amount of compute used to train a general-purpose system largely correlates with its capabilities, as well as the magnitude of its risks.

Today’s most advanced models, like OpenAI’s GPT-4 or Google’s PaLM, can only be trained with thousands of specialized chips running over a period of months. While chip innovation and better algorithms will reduce the resources required in the future, training the most powerful AI systems will likely remain prohibitively expensive to all but the best-resourced players.

In practical terms, compute is more easily monitored and governed than other AI inputs, such as talent, data, or algorithms. It can be measured relatively easily and the supply chain for advanced AI systems is highly centralized, which means governments can leverage such measures in order to limit the harms of large-scale models.14

To prevent reckless training of the highest risk models, we recommend that governments make access to large amounts of specialized computational power for AI conditional upon the completion of a comprehensive risk assessment. The risk assessment should include a detailed plan for minimizing risks to individuals, communities, and society, consider downstream risks in the value chain, and ensure that the AI labs conduct diligent know-your-customer checks.

Successful implementation of this recommendation will require governments to monitor the use of compute at data centers within their respective jurisdictions.14.5 The supply chains for AI chips and other key components for high-performance computing will also need to be regulated such that chip firmware can alert regulators to unauthorized large training runs of advanced AI systems.15

Through passage of the CHIPS and Science Act of 2022, the United States has instituted licensing requirements15.5 for export of many of these components in an effort to monitor and control their global distribution. However, licensing is only required when exporting to certain destinations, limiting the capacity to monitor aggregation of equipment for unauthorized large training runs within the United States and outside the scope export restrictions. Companies within the specified destinations have also successfully skirted monitoring by training AI systems using compute leased from cloud providers.16 We recommend expansion of know-your-customer requirements to all high-volume suppliers for high-performance computing components, as well as providers that permit access to large amounts of cloud compute.