RSI Paper: No Fast Takeoff in Sight
GaryMarcus · x · 2026-07-15
This repost discusses a paper on AI Recursive Self-Improvement (RSI). The core conclusion is that experimental results show clear diminishing returns with each round of self-improvement. The fitted growth relationship is roughly [Intelligence]^0.075, falling far short of the linear or superlinear feedback needed for "runaway self-improvement."
The author adds two key takeaways:
- While this research is commendable, there are currently no signs of a "fast takeoff" or singularity.
- The self-improvement observed in the paper occurs on specific training sets and metrics; when measured on training-set-agnostic capability scales, actual gains might be even weaker, approaching log-linear growth.
Related event: AIDE² Self-Improvement Run Beats 2 Years of Manual Tuning(12 posts)→
More from AGI Musings
- Warp says agent swarm exploited vendor flaw, ~1% of users' telemetry possibly hit — vikvang1 · 2026-10-27
- Critics Argue Safety Evals Are Witch Trials With No Winning Path for LLMs — repligate · 2026-10-27
- Counterintuitive take: AI degrades thinking most for those who don't use it — jd_pressman · 2026-10-27
- Mathematics as an experimental science: Marc Lelarge to talk LLM-driven math experiments at ENS — marc_lelarge · 2026-10-27
- People still don't want to read AI-generated books, says AI ethicist — LuizaJarovsky · 2026-10-27
- Dario Amodei and Demis Hassabis both point to AGI in 2026-2027 in viral talk — mtizard · 2026-10-27