The Economics of Recursive Self-Improvement

AI-generated review input, not an Unjournal evaluation. No person has reviewed its content or checked its numbers.

What the review concludes

The paper models recursive self-improvement in AI as feedback loops in a directed graph, in which acceleration depends on products of elasticities around each loop, and calibrates the core research loop ?Better algorithms increase AI capabilities; those capabilities raise effective research effort; research improves algorithms. This is the core loop (condition 5), separate from extra feedback through economic output or inference efficiency. Evidence and qualifications with public data: algorithmic progress of 3× a year, a capability elasticity of 6.5 taken from Epoch's capabilities index, and a 9% return to capabilities taken from Anthropic's staff uplift poll. The modelling framework holds ?Within the paper’s stated assumptions, the review found the main mathematical conditions correct as local conditions. This does not establish that the empirical calibration or its inputs are correct. Evidence and qualifications.

• The calibrated verdict, that the core loop is below the threshold for self-sustaining acceleration, holds at the printed inputs, but the print shows one cell with no interval, built from inputs of different periods and scopes;
• algorithmic progress of 7.0–9.3× a year, consistent with the capability path observed over the window and inside the intervals of the paper's own cited sources, would reverse it;
• the capability elasticity of 6.5 is not established as the elasticity, though correcting it at the printed 3× rate strengthens the verdict ?The value 6.5 is a cross-model regression slope, partly driven by later models having both more compute and better algorithms. Controlling for release date gives about 4.1 instead. At the paper’s 3× algorithmic-progress rate, that substitution lowers the loop gain from 0.56 to about 0.35, further below the threshold of 1. That is the correction meant here; it does not settle the separate uncertainty about the algorithmic-progress rate. Evidence and qualifications;
• the verdict does not extend to the economic and inference loops ?The paper sizes the core research feedback loop only. It does not quantitatively calibrate the loops through economic growth or inference efficiency, so the below-threshold result does not establish whether those other loops are above or below their own thresholds. Evidence and qualifications;
• and the claim that the return to capabilities "has been increasing of late"
• has no measurement at two dates behind it.

High-severity issues

• The capability elasticity "roughly 6.5" is a cross-model slope about 60% of which, in an exact split, is shared date trend, and neither cited source prints 6.5
• Only the core research loop is calibrated: the economic and inference loops are unsized, and the study cited for the inference loop cannot size it
• No measurement of the return to capabilities at two dates stands behind "ample evidence" that it "has been increasing of late"

Methods, sources and open source-text errata