Re-analysis of METR's time horizons finds a 3-minute to 30-minute jump is far easier than 30 minutes to 5 hours
- Sources: paper
- Summary: The paper recomputes 228 tasks across 26 models using splines and item-response theory. The resulting conversion from human task time to AI difficulty is nearly flat between 2 and 30 minutes, then steepens.
- Why it matters: Equal multipliers on the widely cited time-horizon plot do not represent equal capability gains, so a doubling rate read off that curve cannot be extrapolated past the half-hour mark.