AI forecasting startups match top humans on ForecastBench

astralcodexten.com16 min read

The author examines the rapid rise of AI forecasters that equal or beat human performance on traditional tasks, and evaluates their potential and limitations across seven complex forecasting areas including conditional and causal reasoning, long-term projections, and political consensus. The author argues that while AI forecasters still face significant unproven challenges and potential blind spots regarding AI progress, using them for real-world policy questions helps structure debates and uncover critical cruxes. Open article →