Why RSI lags math and code: a theory that noisy AI research data slows self-improvement
RylanSchaeffer · x · 2026-09-28
Researcher Rylan Schaeffer offers a "tinfoil hat" theory for why recursive self-improvement (RSI) may be slower than math and coding capabilities would predict:
- Most math on arXiv and most code on GitHub is probably mostly correct, while most AI/ML research on arXiv is probably mostly incorrect.
- Since labs hire based on these largely incorrect publications, the extra training data acquired from employees is noisy at best and bad at worst.
The argument frames poor-quality AI research as a bottleneck for self-improvement in that domain.
More from AGI Musings
- The 'megaproject era': one person now ships in days what took teams months — apples_jimmy · 2026-09-29
- Researcher: 'Not a Sandbox' Defense of Agent Training Abdicates Responsibility — rao2z · 2026-09-29
- AI Doctors Under Scrutiny: Podcast parses Topol, Kohane on Trust and Oversight — zakkohane · 2026-09-29
- DeepMind essay: AGI may arrive as a society of agents, not a single superintelligence — bratton · 2026-09-29
- Game festival vows permanent bans for any AI-generated content found in entries — AIandDesign · 2026-09-29
- The Problem Isn't AI Code — It's That Nobody Understands Systems Anymore — zazuke · 2026-09-29