
METR have updated their analysis of AI’s accelerating capability to handle longer tasks across domains. The graphs show doubling times for completable task length across common benchmarks. It indicates around 2.5 to 6 months for intellectual work like coding and maths. More physical tasks like self-driving improve more slowly at 17 months. Frontier models now tackle 100+ minute tasks reliably, approaching half-day professional work. The exponential growth suggests we’re months, not years, from AI managing day-long projects.
