Ethan Mollick: AI's work capacity doubled 8 times in 30 months, METR benchmark maxed out
One Useful Thing (Ethan Mollick) · rss · 2026-10-20
Ethan Mollick publishes his new book Co-Existence, 30 months after Co-Intelligence, and charts the exponential AI progress in between.
- The charts were produced by Opus 5.5 orchestrator agents in minutes, fixing his own math mistakes along the way
- Since GPT-4's lone reign, 20 different models have held the top spot; 180 models now beat GPT-4-at-release, many open weights, 19 runnable on smartphones
- METR's Long Horizon benchmark ran out: AI went from 4-minute tasks in 2024 to 17 hours by April 2026, doubling 8 times; other studies suggest 17+ weeks
- Cost per correct GPQA answer fell from $0.30 to 0.04 cents
- New third work mode "the conjurer": summon agents, assign work, evaluate results — powerful but with new loss-of-control risks
- His classic "Otter on an Airplane Using Wifi" test fell: Opus 5.5 autonomously produced the full video
More from AGI Musings
- David Manheim: ASI existential risk likely arrives before AI's massive economic impact — davidmanheim · 2026-10-21
- Bombadil: a foundation model for plant electrophysiology trained on 4M hours of signals — vboykis · 2026-10-21
- As LLMs Outpace Robotics, Physical-World Jobs May Hold Value Better Than Desk Jobs — venturetwins · 2026-10-21
- What would AI alignment look like if developed in a more collectivist culture? — mimi10v3 · 2026-10-21
- Researchers Debate Embodied AI Stack: Steerable S0 Controllers and Closed-Loop S2 by 2027 — chris_j_paxton · 2026-10-21
- Learning moves up a layer: AI-era devs should think in systems, not code — MoonL88537 · 2026-10-21