Testing the RL hypothesis with METR data: AI speedup started ~4 months before Mythos
austinc3301 · x · 2026-11-10
Independent re-analysis of METR's Nov 2026 AECI data testing Cohen's RL hypothesis: the larger pretrain (Mythos) only arrived in Feb 2026, so if it drove the capability speedup, acceleration should start then. But a fit allowing a one-time jump puts the start at Oct 2025 — 4 months earlier (85% likely within METR's error bars) — suggesting the bigger pretrain alone doesn't explain the speedup. Chart shows both fits, with Opus 4.6 landing just after the detected acceleration onset.
Related event: Independent Test: AI Acceleration Predates Mythos by Four Months(2 posts)→
More from AGI Musings
- Ex-AI Security Lead: AI Luminaries' Cyber Takes Are Wrong, Real Expertise Will Reclaim Authority — joshua_saxe · 2026-11-10
- SlopOps: mistaking agentic coding capability for competitive advantage is playing the slop lottery — generativist · 2026-11-10
- Eric Schwitzgebel's 22-minute interview on the possibility of AI consciousness — eschwitz · 2026-11-10
- Yudkowsky: AI fiction lags code badly — 'we may only be alive because gradient descent generalizes policies so poorly' — sebkrier · 2026-11-10
- S&P 500 firm lifespan fell from 61 to 15 years — competitive advantage matters more than ever — RogerLMartin · 2026-11-10
- Unconstrained intelligence kills intelligibility, just as unconstrained distribution killed attention — fkasummer · 2026-11-10