ML//AI progress//recursive self-improvement

Recursive self-improvement is a feedback loop in which an AI system does part of the research that produces its successor, so that each improvement can speed up the next, and it is the mechanism behind forecasts of a rapid AI takeoff. The loop is reinforcing, and its gain is the return to research: the sign only says that it amplifies, while whether it accelerates, holds steady or fades depends on the size of the gain, set by numbers that can be estimated.

Recursive self-improvement is a feedback loop in which an AI system does part of the research that produces its successor, so that each improvement can speed up the next, and it is the mechanism behind forecasts of a rapid AI takeoff. The loop is reinforcing, and its gain is the return to research: the sign only says that it amplifies, while whether it accelerates, holds steady or fades depends on the size of the gain, set by numbers that can be estimated.

The returns to research: in the idea production function, the loop accelerates when the doublings of software per doubling of research effort, rrr, exceed 1. Published estimates straddle 1.

The experiments: AI researchers run on the same chips their experiments need, so the elasticity of substitution between thinking and compute decides whether more AI researchers help (compute bottleneck).

The clock: each round of train, evaluate and deploy takes time, and a floor on that generation time rules out a singularity.

Measured effects are still small and disputed. Independent studies of developers using AI found changes from about 20% slower (2025) to about 18% faster (2026, with a wide interval), while self-reports inside labs are several times larger; nobody outside the labs has measured the speed-up of research itself.