AI Reports · GovAI (Oxford) · 2026-11-09 · Research paper
Economics of AI Efficiency vs Capability in the Intelligence Explosion October 2026 Pre-Print Karthik Tadepalli A software-driven intelligence explosion is plausible but uncertain.
The paper models AI progress feeding back into AI research through two channels: capability gains that automate more research tasks, and efficiency gains that raise research output per unit of compute. It finds a software-driven intelligence explosion requires both to improve fast enough, with calibration suggesting capability progress is likely sufficient but efficiency progress is on a knife-edge. In Monte Carlo simulations an intelligence explosion occurred in 49% of draws, and the rate of efficiency progress accounts for 45% of the uncertainty, so better efficiency data is essential.