Efficiency vs Capability in the Intelligence Explosion

Abstract

AI progress can feed back into AI research in two ways. Capability improvements allow AIs to perform a larger share of AI research tasks, while efficiency improvements raise AI research output per unit of compute on tasks already automated. In a task automation model of AI research, I show that a software-driven intelligence explosion - a self-sustaining acceleration in AI progress with fixed compute and human labor - requires both margins to improve sufficiently fast, and neither margin can substitute for the other. To see whether these conditions hold, I calibrate both margins with data on recent AI progress. I estimate that capability progress is likely fast enough to support an SIE, while efficiency progress is on the knife-edge. In Monte Carlo simulations, an SIE occurs in 49% of draws, indicating that an SIE is plausible but uncertain based on historical trends in AI progress. The single largest source of uncertainty is the rate of efficiency progress, whose uncertainty is responsible for 45% of the uncertainty about whether an SIE occurs, more than all other parameters combined. Thus, better data on efficiency progress in frontier models is essential for understanding the likelihood of an SIE.