How to Pace the US Frontier (blog.aifutures.org)

Last week, over 1,000 employees from the frontier AI sector signed an open letter advocating for U.S. government support in developing tools to "pace the frontier" of automated AI capabilities. This initiative aims to moderate the timeline for the development of advanced AI technologies, primarily to mitigate existential risks, while also ensuring that U.S. companies maintain competitive advantages over international counterparts, particularly China. The letter suggests that establishing domestic regulatory measures could be a precursor to broader international coordination, effectively delaying the rapid advancements that could lead to potential dangers. The proposal outlines a tiered approach for enforcing pace in AI development, starting with a temporary pause on enhancing frontier model capabilities. Future options include requiring companies to allocate a set percentage of their computing resources for external inference rather than training new models, as well as mandating transparent safety compute allocations for safety-related research. A key aspect of the plan is to implement robust measurement systems to evaluate AI capabilities and adjust regulations as necessary. Ultimately, the vision is to create a balanced development environment that prioritizes safety while allowing for continued innovation, with the potential for these domestic regulations to evolve into international frameworks.