Back-of-envelope math says AI self-acceleration is within a year: 35² ≈ 1000x effective compute
ChrisGPT · x · 2026-09-27
The author works through Anthropic's AI R&D-5 threshold—the point where AI dramatically accelerates effective scaling. Using Anthropic's 35x/year effective scaling estimate, compressing two years into one implies 35² ≈ 1000x effective compute. Citing (unverified) Opus 5.5 at 55.8% on CoBench 2.1 vs. the 85% researcher-substitution level, the author argues labs are about a year from automating AI R&D, with continuous learning and long-term memory the last barrier.
More from AGI Musings
- Predicting AI Will Be 'Literally Everywhere' by This Time Next Year — MickeySteamboat · 2026-09-27
- Keeping a cool head is an act of rebellion, says AI researcher Cameron McNerney — Kyrannio · 2026-09-27
- Claim: AI Progress Next Year Will Be 6x Faster Than the Past Four Years Combined — MickeySteamboat · 2026-09-27
- Prediction: AIs with robust autobiographical memory will truly be conscious — yeastsplainer · 2026-09-27
- Ex-OpenAI/Anthropic pretraining researcher quits, slams labs' superintelligence race — Aiden_Tech_Ai · 2026-09-27
- Worker argues AI should replace much of middle management's approval busywork — sporty_outlook · 2026-09-27