works
Tom Davidson and Houlden. Tom How quick and big would a software intelligence explosion be? online The deployment of AI systems capable of fully automating AI research and development (ASARA) may precipitate a software intelligence explosion, where recursive improvements in algorithms and efficiency drive progress independently of increases in physical compute. Semi-endogenous growth modeling suggests that the trajectory of this explosion is governed by three primary factors: the initial research speed-up afforded by automation, the returns to subsequent software R&D, and the proximity to fundamental limits of computational efficiency. Probabilistic analysis indicates a 60% likelihood that such an explosion will compress more than three years of current AI progress into a single year, and a 20% likelihood that over ten years of progress will occur within the same period. This acceleration is expected to coincide with AI systems surpassing human expertise in broad scientific and engineering domains. While the magnitude of this transition is sensitive to assumptions regarding diminishing returns to parallel labor and the total distance to effective software limits, the intelligence explosion likely represents a substantial but bounded intensification of technological progress. This perspective provides a quantitative intermediate between extreme takeoff skepticism and theories of indefinite, near-instantaneous growth. – AI-generated abstract.

How quick and big would a software intelligence explosion be?

Tom Davidson and Houlden. Tom

Forethought, August 4, 2025

Abstract

The deployment of AI systems capable of fully automating AI research and development (ASARA) may precipitate a software intelligence explosion, where recursive improvements in algorithms and efficiency drive progress independently of increases in physical compute. Semi-endogenous growth modeling suggests that the trajectory of this explosion is governed by three primary factors: the initial research speed-up afforded by automation, the returns to subsequent software R&D, and the proximity to fundamental limits of computational efficiency. Probabilistic analysis indicates a 60% likelihood that such an explosion will compress more than three years of current AI progress into a single year, and a 20% likelihood that over ten years of progress will occur within the same period. This acceleration is expected to coincide with AI systems surpassing human expertise in broad scientific and engineering domains. While the magnitude of this transition is sensitive to assumptions regarding diminishing returns to parallel labor and the total distance to effective software limits, the intelligence explosion likely represents a substantial but bounded intensification of technological progress. This perspective provides a quantitative intermediate between extreme takeoff skepticism and theories of indefinite, near-instantaneous growth. – AI-generated abstract.

PDF

First page of PDF