AI agents' task complexity is doubling every 4 months, with time horizons growing exponentially from 30 seconds in 2022 to 14+ hours today
Hacker News · 2026年4月28日
AI要約
•Researchers at METR tested AI agents from 2019 to 2026 on about 230 tasks (mostly coding, some general reasoning) and found task length is highly correlated with agent success rate (R² = 0.83). The length of tasks agents succeed at 50% of the time is growing exponentially.
•From 2019–2025, time horizons doubled every 7 months. In 2024–2025, this accelerated to doubling every 4 months. Current projections suggest agents could reach 1 work week (40 hours) by 2028 and 1 work month (167 hours) by 2029, though the faster trend is less robust because it is based on just one year's data.
•The article notes that as AI systems improve, they could accelerate development of yet more capable AI—creating a potential flywheel where agents speed up creation of more capable agents. This could lead to superexponential growth, with effects described as potentially transformative.