“ai forecasters now equal or beat top human performance at traditional forecasting questions”
AI forecasters now equal or beat top humans on standard questions, so Scott Alexander maps the gaps before anyone cashes in: conditional forecasting (a year-old FRI test found models markedly worse), genuinely causal forecasts, AI progress itself — which superforecasters keep underestimating — fifty-year horizons, unresolvable claims like the COVID lab-leak question, automatic crux generation, and whether models can defy respectable consensus. He uses them anyway, as a public bug bounty.
Some technical background helpful
Platforms mentioned: Metaculus