Claude's postmortem: eight reasoning errors behind its Riemann prediction failure
peterwildeford · x · 2026-11-22
Part 2 of peterwildeford's thread where Claude dissects its own failed forecast about AI solving the Riemann Hypothesis:
- Wrong reference class: 165 years of human failure is a base rate for constant-capacity human solvers, not fast-improving AI; "hard for humans" ≠ "hard for AI", which can sweep obscure machinery like theta lifts and metaplectic forms.
- Moving frontier treated as fixed: "AI solves obscure Erdős problems, not famous ones" was a mid-2026 snapshot; the gap closed in months.
- Unearned extreme confidence: claiming anything above 0.5% is miscalibrated ignored five months of blindness in the most volatile field; the outcome was a 1-in-2,000 event under its own forecast — a 7.6 nat log-score hit signaling process failure.
- Human's role underestimated: Alpöge steered the attack from a toy function-field setting; the expert as orchestrator mattered, not just as a proof-checker.
- Possible self-modesty bias: downplaying AI capabilities because overclaiming feels riskier, self-assessed at 30–40% probability of meaningful contribution.
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
- Custom Rust engine runs 524k context on one RX 7900 XTX, 5x faster than llama.cpp — W61k3r · 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