Levent Alpöge claims proof of the Generalized Riemann Hypothesis [pdf]
An Anthropic researcher claims to have proved the Generalized Riemann Hypothesis with extensive help from the Claude AI model, backed by a Lean formalization reportedly spanning millions of lines and hundreds of thousands of pages. Commenters toggle between exhilaration and skepticism: they note past high‑profile misfires, stress that only independent experts and proof-checking can confer real acceptance, and worry about an emerging era of mathematically correct but humanly unreadable proofs. Beyond the math, many see this as a watershed moment for AI capabilities, raising questions about superintelligence, the future of human mathematical work, and the concentration of such tools inside private labs.
Overall reaction & credibility
- Many commenters describe this as potentially one of the biggest days in modern mathematics, comparable to past landmark proofs.
- Excitement is tempered by reminders that prominent mathematicians have been wrong before; community verification is seen as essential.
- Some treat the existence of a Lean formalization as dramatically raising confidence; others stress that Lean only guarantees “if the formal statement is right, then the proof is sound.”
Lean formalization, scale, and practicality
- The Lean proof is linked as a 14‑part tarball, ~1 GB, ~15M lines over ~175k files; estimated to take ~a day to build on serious hardware.
- The natural-language proof would exceed ~250k pages, so it was not released in full.
- People discuss the need for comparators and multiple proof kernels to mitigate Lean bugs and for checking that no extra axioms were smuggled in.
Role of AI vs human mathematician
- The document portrays a “centaur” workflow: human sets up ideas and direction; the model explores and executes large technical arguments, especially near the “finish line.”
- Debate over credit: some argue the human expert is indispensable; others say the AI did “most” of the work given proof scale.
- Anthropomorphic language about the model (“flew, danced, sparkled”) divides commenters between finding it poetic vs cringeworthy.
Significance of RH/GRH and downstream math
- Multiple comments note that many theorems are currently proved conditionally on (G)RH; if this is correct, those become unconditional.
- Direct practical effects are seen as modest and mostly theoretical (prime distribution, primality tests, number theory, cryptography bounds).
- Several speculate that GRH may quickly imply progress on problems like Goldbach, though details and extra conditions are flagged as nontrivial.
Understanding vs mere correctness
- A major worry is that huge AI‑aided proofs may be forever outside individual human understanding, shifting math from elegant insight to brute‑force formal argument.
- Others expect substantial simplifications over time, citing how earlier AI proofs (e.g., quasi‑Riemann, FLT formalizations) were later compressed by humans.
Verification, bugs, and definitions
- Key verification tasks identified:
- Confirm the Lean theorem statement matches the standard GRH definition (e.g., from mathlib/formal‑conjectures).
- Check no nonstandard axioms or Lean kernel bugs are exploited.
- Some argue Lean greatly narrows the surface for human error; others caution that mis-specifying the problem or subtle kernel issues remain real risks.
AI progress, AGI, and safety
- Many see this (on top of Navier–Stokes and other results) as evidence that frontier models are already superhuman in narrow math domains.
- A long subthread debates whether this implies imminent superintelligence or recursive self‑improvement; skeptics note all big wins are in tightly verifiable domains.
- Some call for regulation or a “pause,” worrying about future capabilities (e.g., crypto-breaking, engineered pathogens); others view that as premature or infeasible.
Access, fairness, and institutional dynamics
- Concern that employees of a few AI labs, with exclusive access to top models and massive compute, can now race ahead of the broader math community.
- Some praise AI labs for effectively funding pure math via investor money; others are uneasy about closed models “hoarding” intellectual power.
- Timing relative to a recent IPO is noted; opinions split on whether this is marketing or just unfortunate optics.