Ramez Naam charts OpenAI internal horizons: p80 may be thousands of times shorter than p50, a negative RSI update
ramez · x · 2026-11-11
Ramez Naam continues his negative update on fully automated RSI timelines, posting OpenAI internal task-horizon charts (Jan–Jul 2026, fitted by monthly success thresholds):
- Limited empirical data does not support research horizons growing exponentially — the p80 (80% success) fits are sublinear or show no good fit.
- P99 horizons may be thousands of times shorter than p50; real-world p80/p50 ratios from OpenAI and GDM show a much faster drop-off for higher success rates.
davidmanheim responds that the claims are 'a bit overstated but roughly correct,' highlighting the key caveat: real research horizons could be 100× shorter than METR today while growing at exactly the same exponential rate — level discrepancies don't imply growth-rate discrepancies.
More from AGI Musings
- teortaxesTex: RL may give coding models 'full fluency' with no superhuman headroom — teortaxesTex · 2026-11-11
- Frontier AI Still Lacks 'Heresy': Breakthroughs Come from Defying Consensus — JeremyNguyenPhD · 2026-11-11
- Dev argues optimism is the only trait that matters as intelligence gets commoditized — zack_overflow · 2026-11-11
- AI risk community splits over 'intelligence explosion next year' claim as credibility war erupts — connoraxiotes · 2026-11-11
- AI lets robotics research thrive in labs with no robotics experts — 4310sy · 2026-11-11
- Dictating notes to an AI agent feels as mundane as putting on shoes — _FelixSimon_ · 2026-11-11