A 97-page survey maps self-improving agents — reality is humbler than the vision
maier_ak · x · 2026-09-22
Andreas Maier's full read of 'Self-Improvements in Modern Agentic Systems: A Survey' (arXiv, July 2026), 97 pages, 670 references, by 12 authors from Jilin University, KAUST, University of Alberta, and IDSIA with Schmidhuber as last author.
- Framing: from I. J. Good's 1966 'last invention' to the question of whether agents can persistently improve themselves
- Core formalism: agent = (θ,Σ); self-improvement updates θ (slow loop) or Σ (fast loop)
- Unifies scattered literatures — meta-learning, self-training, prompt optimization, tool creation, evolutionary program synthesis — under one vocabulary and evaluation protocol
- Reality check: most work only tweaks the scaffold; weight updates are rare and risk collapse; judge metrics may over-optimize to biased evaluators
- Verdict: a good map, but of a smaller territory than the epigraph promises
More from AGI Musings
- AI Futures Q2.5 update: reality running 70-90% as fast as AI 2027 predicted — ben_j_todd · 2026-09-22
- Judgment is becoming the scarce capital: why co-ownership may beat one-person unicorn firms — curious_vii · 2026-09-22
- Podcast: Epoch AI researcher on RSI, robotics, China model gap, and open vs closed safety — natolambert · 2026-09-22
- Epoch AI's JS Denain Debates RSI, US-China Gap, and AGI Timelines with Nathan Lambert — Interconnects (Nathan Lambert) · 2026-09-22
- François Fleuret: math and CS modularity explains AI's IMO success — francoisfleuret · 2026-09-22
- Gary Marcus: near-term AI risk is deepfaked disinformation, not extinction — GaryMarcus · 2026-09-22