When AI Starts Improving Itself: The Economics of Recursive Self‑Improvement

AI TherapyAI DiagnosticsAI Urban Planning

Imagine a world where AI not only builds better models but also speeds up its own progress—this paper maps that possibility.

TL;DR
• The paper models recursive self‑improvement (RSI) using directed graphs and elasticities.
• Current data suggest RSI is not yet self‑sustaining but is strengthening.
• Authors call for shared empirical metrics to better assess future acceleration.

Summary
The paper models recursive self‑improvement (RSI) in AI, quantifies feedback loops, and evaluates whether current dynamics can sustain accelerating capability growth. It distinguishes narrow from broad capabilities, outlines key elasticities, and proposes data that firms should share for better calibration.
• Track algorithmic‑efficiency gains and map them to R&D productivity improvements.
• Publish disaggregated R&D spend on compute, data, and human labor to enable elasticity estimation.
• Measure the proportion of technical advances directly attributable to AI‑generated research.
• Monitor compute‑scaling constraints and power‑capital limits as potential bottlenecks.
• Incorporate economic‑output feedbacks to assess how AI‑driven growth funds further investment.

The authors construct a series of graph‑based models where nodes represent production outputs (algorithmic efficiency, AI capability, economic output) and edge weights capture elasticities. A self‑sustaining acceleration occurs only if the product of elasticities across the core feedback loop exceeds unity. Calibration using the Epoch Capabilities Index suggests a 15 % productivity uplift per capability unit is required, while observed uplift from coding agents is about 9 %, indicating the threshold is not yet met but is trending upward.

The analysis highlights that narrow capability gains—improvements on benchmark tasks—do not automatically translate into broad, economically valuable outcomes. Bottlenecks may arise from limited compute, data availability, or regulatory constraints, which can break the feedback loop despite strong algorithmic progress. Economic feedback, where higher output finances additional compute and data, can mitigate some bottlenecks, but the model stresses the need for continuous monitoring of these reinforcing mechanisms.

A comprehensive wish list is offered: (1) growth rates of algorithmic efficiency, (2) detailed R&D expenditure breakdowns, and (3) direct metrics of AI contributions to research. By making these measures publicly available, the community can refine elasticity estimates and more accurately forecast whether RSI will transition to a self‑sustaining acceleration. The paper situates its contribution within the broader literature, extending semi‑endogenous growth models and emphasizing the narrow‑broad capability distinction, thereby providing a clearer framework for assessing future AI-driven economic transformations.

Will the next wave of AI breakthroughs be driven by the systems themselves, or will external limits hold them back?