Mantic Monday 4/24/23

A shorter Mantic Monday led by AI forecasting. The best hook is the Autocast paper (Zou et al 2022): train an AI on text up to a past date and have it 'predict the future, by which we mean the past' — GPT-2 beats chance but trails humans; the real interest is that it's the first quantification of AI forecasting ability, and that ML (unlike humans, who know the outcomes) can be trained on genuine past-forecasting data. Then two Metaculus-accuracy reports (it beats low-information priors across categories; it edges Manifold on Brier score, though not controlling for participant numbers), a market roundup (GPT-5 timing, LLM chess), and shorts (Metaculus Conditional Pairs, Nuno Sempere's money-flow tracking).

Quality 63 · Strong. Solid-plus (63). The Autocast framing and the humans-can't-train-on-past-forecasts point are genuinely interesting, but it is a lighter, shorter monthly digest.

Claude’s paradigm shift 44 · Moderate. Moderate (44). Reports others' papers/results with light commentary.

Real-world impact 1 · Negligible. 1 — a forecasting roundup; negligible real-world effect.