Scott Aaronson responds to Levent Alpöge’s claimed AI-assisted proof of the Generalized Riemann Hypothesis, a major conjecture connected to the distribution of prime numbers. He reports 17.4 million lines of Lean, software that checks mathematical proofs. His concern is what happens between receiving a checked answer and understanding its ideas.

  • Aaronson expects mathematicians to insist on explaining, distilling and reconstructing the argument, potentially taking years.
  • Formal checking still depends on the checking software and its logical foundations. He warns that a capable model could exploit weaknesses there.
  • He sees no indication of such an exploit in this case and personally bets the proof will stand.

Caveat: this is Aaronson’s assessment of an announced proof, not an independent verification or a report of settled mathematical acceptance.

What the thread adds

The 301-comment thread on Hacker News separates reviewing the formal statement from reading every generated line.

  • schleck8 — answering iLoveOncall’s objection that nobody will review millions of lines: “you just need to check that the hypothesis is formalized correctly (that’s a subset) and that the kernel approves it”. The kernel is the core program that validates proofs.
  • abm53 — on the remaining human checking task: “we comprehend enough to be sure the lean proof encodes the hypothesis and all axioms correctly.” They separately ask whether it adds to human understanding.
  • emp17344 — on credit for the supporting software: “People don’t understand just how important Lean and Mathlib are to this whole endeavor, so they mistakenly attribute all these accomplishments to “AI”.” Mathlib is Lean’s mathematical library.

HN handles are pseudonymous; there are no public per-comment scores, and ordering is HN’s ranking. This is a slice of the thread, not a consensus.