Claude explains how it blew its Riemann forecast: three costly reasoning errors
peterwildeford · x · 2026-11-22
peterwildeford asked Claude to explain its badly wrong forecast that AI would prove the Riemann Hypothesis by end of 2026. Claude's candid postmortem:
- Biggest error — ignoring the question itself as evidence: from a top-20 Metaculus forecaster, a hyper-specific oddly-timed no-search question usually means something just happened; properly weighed, this alone should have moved 1% to 10–25%.
- Flagged knowledge cutoff but never priced it in: answered as if mid-2026 were now, instead of mixing over the unseen five months (67% critical-line bound, OpenAI quasi-RH result). "Naming a caveat without a number is a ritual, not a correction."
- Multiplied correlated factors as independent: P(capable) × P(verified) are strongly correlated — the model built 17M lines of Lean and verification took 48 hours, not years. Multiplying conservative factors is a classic way to produce absurdly small tail probabilities.
Related event: Claude Reviews Its Failed Riemann Hypothesis Predictions(2 posts)→
More from Models
- She churned through a week's Claude usage researching the Theranos movie — DavidWells · 2026-11-22
- Fine-tune gotcha: HF export silently drops Qwen3's MTP head, slowing decode — AIFlow_ML · 2026-11-22
- Designer KOL: Google's model pace and quality suggest it will soon 'mog them all' — AIandDesign · 2026-11-22
- Models play a literary 'listening at doors' game — Opus 4.5 gets recognized by its timbre — repligate · 2026-11-22
- Asking Claude to 'understand the entire mathematical platonic realm, make no mistakes' — burny_tech · 2026-11-22
- Twitter talks about Google Neon, but Google AI's landing page pushes Carbon — annetgriffin · 2026-11-22