If ~70% of AI capability gains come from better RL environments, RSI may be surprisingly easy
burny_tech · x · 2026-11-22
Author burnytech argues that roughly 70% of current AI capability drivers come down to more and better RL environments and doing more RL. If that holds, recursive self-improvement along this direction seems actually pretty easy — a notable take on the RL-scaling route to capability gains.
More from AGI Musings
- Cutting entry-level jobs destroys the talent pipeline: the case for TRIBORG teams — BubblyOption7980 · 2026-11-22
- Lean experts confirm Levent's GRH proof was genuine human-AI collaboration, not brute force — anshulkundaje · 2026-11-22
- ChatGPT and Muse are free, so B2B revenue times C-end users is a misleading AI valuation — oran_ge · 2026-11-22
- KOL compares AI doomer '1% risk' argument to COVID-era safety cult — kevinnbass · 2026-11-22
- Not just fast cars: researcher argues elite users, not models alone, drive top AI results — anshulkundaje · 2026-11-22
- Emad Mostaque mocks mathematicians for pivoting to 'math is about the struggle' as AI cracks problems — rohanpaul_ai · 2026-11-22