OpenAI shared 80% accuracy time horizon doubles in two months

An analysis of publicly shared data on OpenAI's internal AI research and development time horizons reveals that the 80% accuracy time horizon is relatively low at around 15 minutes, but is doubling very quickly at roughly a two-month rate. The author discusses different statistical distribution fits, such as logistic and Weibull curves, and explores potential explanations for why these time horizons grow so much faster than baseline benchmarks like METR while remaining low in absolute level. Open article →