METR proposes apple-picking model for AI R&D diminishing returns
soumitrashukla9 · x · 2026-08-18
METR published a note introducing an "apple-picking model" to measure AI agents' contribution to frontier R&D. The model likens AI optimization to a robot picking apples: agents handle low-hanging fruit (easy optimizations) but struggle with higher branches (hard breakthroughs). It implies that while agents push the frontier, returns diminish sharply and are not additive—e.g., as of March 2026, agents provide 1 week of human-equivalent labor, but running an agent twice doesn't yield two weeks of work.
More from AGI Musings
- Prediction: Frontier models may surpass Instruct baseline by September 2026 — maksym_andr · 2026-08-18
- Polymarket prices just 9% odds a Chinese firm tops Chatbot Arena by 2026 — Polymarket · 2026-08-18
- Five Eyes warns AI advances will trigger sophisticated cyber-attack onslaught — connoraxiotes · 2026-08-18
- What Gives Human Life Meaning When AI Does Everything Better? — PeterDiamandis · 2026-08-18
- Scholar proposes using AI teams to conclusively settle colonial economics debate — paulnovosad · 2026-08-18
- OpenAI Foundation holds 26% of OpenAI; AI is about to unlock hundreds of billions in philanthropy — iskander · 2026-08-18