Increased Compute Efficiency and the Diffusion of AI Capabilities

Training advanced AI models requires large investments in computational resources, or compute. Yet, as hardware innovation reduces the price of compute and algorithmic advances make its use more efficient, the cost of training an AI model to a given performance falls over time. To analyze this phenomenon, we introduce compute (investment) efficiency, which relates training compute investment to the resulting AI model performance. We then present a conceptual model of increases in compute efficiency and assess the social and governance implications. We find that while an access effect increases the number of actors who can train models to a given performance over time, a performance effect simultaneously increases the performance available to every actor - potentially enabling large compute investors to pioneer new capabilities and maintain a performance advantage even as capabilities diffuse. The market effects are multifaceted: while a relative performance advantage might grant outsized benefits in zero-sum competition, performance ceilings might reduce leaders' advantage. Nonetheless, we find that if the most severe risks arise from the most advanced models, large compute investors warrant particular scrutiny since they discover potentially dangerous capabilities first. Consequently, governments should require large compute investors to warn them about dangerous capabilities, thereby enabling timely preparation and potentially using their superior model performance and compute access for defensive measures. In cases of extreme risks, especially offense-dominant capabilities, the government might need to actively restrict the proliferation entirely.