Determine the Continuation and Limits of Algorithmic and Hardware Trends Reducing Compute Requirements

Ascertain whether long-run trends that reduce compute requirements and cost—namely algorithmic progress that improves compute efficiency, hardware price–performance improvements, and communication-efficient training that enables more decentralized training—will continue, and characterize their practical limits and implications for the efficacy of compute governance.

Background

The paper highlights that algorithmic advances can lower the computational operations required to reach a given AI performance, while hardware advances improve price–performance, and communication-efficient training could enable more decentralized training. Together, these trends could erode the effectiveness of compute-based governance over time.

Understanding if and how these trends persist—and identifying their limits—matters for anticipating how governable future AI development will be and for designing robust governance mechanisms.