The Economics of Recursive Self-Improvement (metr.org)

A new paper titled “The Economics of Recursive Self-Improvement” by Parker, Tom, and seven other economists explores how AI could catalyze its own research and development through Recursive Self-Improvement (RSI). The authors emphasize that understanding RSI is essential for forecasting AI capabilities and assessing risks associated with advanced AI systems. They clarify that while RSI typically refers to feedback from AI model capabilities to improvements, there is ambiguity around its definitions and implications—some see it leading to exponential or fully autonomous growth, while others suggest a broader definition. The paper avoids technical nomenclature to focus instead on quantifying the strength of feedback effects that may enable self-sustaining acceleration in AI progress. The study reveals that the extent of potential capability acceleration depends significantly on feedback strength, with implications that cannot be ignored. Although various factors could limit this acceleration—such as constraints on data, training resources, or specific algorithmic advancements—there is insufficient evidence to dismiss the possibility of rapid capability growth. Additionally, the authors highlight the need for more comprehensive data releases from AI labs to inform their models better. Their ongoing work aims to refine quantitative estimates and explore factors that could both accelerate and decelerate future AI development, contributing to a richer understanding of AI's evolving landscape.