Can governments quickly and cheaply slow AI training?

A new technical analysis, originally written as a private document for AI safety professionals, tackles a critical question: if powerful AI becomes "obviously scary," could governments implement quick, cheap measures to slow its development? The report focuses on 'inference verification'—methods like restricting server communication, limiting bandwidth, or periodically erasing clusters to ensure systems are only running inference, not training. The author's week-long investigation concludes that current verification prototypes would likely fail to create substantial slowdowns.

The core problem lies in the nature of modern AI training, particularly reinforcement learning (RL). RL training involves a large fraction of 'inference' work, like generating and scoring agent action rollouts, which can still be performed under strict communication constraints. The analysis posits that developers could allocate 95% of their verified compute to this inference work and use a mere 5% of compute in covert, unverified data centers to calculate the actual training updates. While proposals like frequent memory wipes or output recomputation would create hurdles, the report argues these are likely feasible to work around, leaving the feasibility of rapid, effective inference verification as an open question.