The meatiest Mantic Monday of the draw. 'Robots Of Prediction' reviews two AI-forecasting studies (Halawi et al Berkeley, Tetlock et al) showing fine-tuned/crowd LLMs now reach ~90-95th-percentile human-crowd forecasting. The standout 'Towards Rationality Engines' section is a genuinely generative Scott idea: could forecasting AIs become domain-general reasoners by training on trials/scientific-papers/law-cases with known answers, and how would you test it? Then 'The Predicting Of Robots' analyzes the FRI adversarial-collaboration result (80 hours of debate moved AI-skeptic superforecasters from 0.1% to 0.12% p(doom)), diagnosing the crux as a speed disagreement (2045 vs 2450 for AI > all humans) and asking why superforecaster opinion on AI risk is bimodal.

Quality 68 · Strong. Strong: elevated well above the roundup floor by the original 'rationality engines' speculation and the sharp FRI-adversarial-collaboration analysis; a roundup carrying real essay-grade content.

Claude’s paradigm shift 48 · Moderate. Notable: the 'forecasting AI → domain-general rationality engine' framing and the bimodal-superforecaster-disagreement analysis are fresh, generative ideas.

Real-world impact 2 · Minor. Within-discourse: shapes forecasting/AI-risk thinking; modest reach.

Mantic Monday 3/11/24