The paper examines recursive self-improvement through an economic lens, modeling how AI systems that enhance their own capabilities could trigger an intelligence explosion. It argues that traditional growth models fail to capture the feedback loops where improved intelligence reduces the cost of further improvements.

Key variables include compute costs, algorithmic efficiency, and the elasticity of intelligence gains to investment. The analysis suggests that under plausible parameters, the takeoff speed could range from slow (years) to fast (months), depending on whether hardware or software progress dominates.

The author introduces a "recursive improvement function" that maps current capability to the rate of future capability growth. This function incorporates diminishing returns from scaling laws but accelerating returns from architectural innovations. Empirical estimates from recent large language model trajectories are used to calibrate the model.

Policy implications include the difficulty of regulating a process where the regulator's own understanding lags behind the system's evolution. The paper concludes that alignment investments must scale superlinearly with capability to maintain safety.