How to Pace the US Frontier

This post was submitted by a SAIRC member either as a recommended read or student-created post. All credit remains with the original author.

Written with Brendan Halstead, Romeo Dean, Thomas Larsen, and one other author, this piece proposes a framework for domestically regulating frontier AI to moderate when highly capable systems arrive. It lays out four escalating options — a temporary pause on capability improvements, compute allocation minimums (70% external inference, 25% safety research), restrictions on using frontier models for AI R&D, and risk-assessment-based oversight — and argues that domestic pacing is worth doing even absent international coordination, since it slows competitors' access while preserving US technical leadership during the preparation window.