ML//AI progress//intelligence explosion

An intelligence explosion is a hypothesised episode in which recursive self-improvement makes AI capability grow faster than exponentially, compressing years of progress into months, and it is the outcome that forecasts of AI takeoff argue about. Recursive self-improvement is the mechanism; the explosion is one possible outcome of it, the one where the loop's gain stays above one long enough. I. J. Good stated it in 1965: a machine that could design better machines would set off an explosion of intelligence.

An intelligence explosion is a hypothesised episode in which recursive self-improvement makes AI capability grow faster than exponentially, compressing years of progress into months, and it is the outcome that forecasts of AI takeoff argue about. Recursive self-improvement is the mechanism; the explosion is one possible outcome of it, the one where the loop's gain stays above one long enough. I. J. Good stated it in 1965: a machine that could design better machines would set off an explosion of intelligence.

Forecasts now put numbers on it. One widely cited estimate (Davidson and Houlden, 2025) gives about 60% to compressing more than three years of software progress into one, and about 20% to compressing more than ten, but only under two conditions: that an AI fully automating AI research is deployed, and that compute stays fixed from then on.

An explosion in the strict mathematical sense is a finite-time singularity, which needs the duration of each round to shrink towards zero. Rising growth rates alone give superexponential growth, fast but finite.

Whether progress looks explosive can depend on the metric (coordinate singularity).

The arguments against rest on physical inputs: experiments that need compute, packaging and memory sold years ahead, power plants that take years to connect, and a slowest loop that sets the pace (rate-limiting step).