The Riemann Hypothesis, proved with Claude in Lean, and a GPGME double free

14 hours ago

Levent Alpöge, working with Claude, released a proof of the Riemann Hypothesis as 17 million lines of Lean, and the Lean and Mathlib communities certified it within a day. Also from Saturday's Hacker News front page: a double free in GPGME reachable with a crafted OpenPGP key, Valkey replacing the skiplist under its sorted sets with a B+ tree, and the Studio Display XDR hub that held a 2.5 gigabit line to 600 megabits. Lean verification of the Riemann Hypothesis, a joint statement by the Lean and Mathlib communities https://leanprover-community.github.io/blog/posts/riemann-hypothesis/ Discussion: https://news.ycombinator.com/item?id=50589618 Levent Alpöge's remark announcing the proof: https://alpo.ge/remark.pdf Verification files and results: https://math.mit.edu/~drew/grh/ thejokeisonme on the fake Collatz refutation, and TwoFx's reply: https://news.ycombinator.com/item?id=50590575 math_dandy, quoted: https://news.ycombinator.com/item?id=50590809 GnuPG Security Advisory for GPGME (CVE-2026-129975, fixed in GPGME 2.3.1) https://lists.gnupg.org/pipermail/gnupg-announce/2026q4/000517.html Discussion: https://news.ycombinator.com/item?id=50587208 lazka on Arch Linux and GPGME: https://news.ycombinator.com/item?id=50588951 The Red Hat downstream patch, via uptrails: https://news.ycombinator.com/item?id=50588137 Todd Zullinger's note on oss-security: https://www.openwall.com/lists/oss-security/2026/11/21/1 From Skiplists to B+ Trees: Making Valkey Sorted Sets More Memory Efficient, by Rain Valentine and Dragos Andriciuc https://valkey.io/blog/from-skiplist-to-b+tree/ Discussion: https://news.ycombinator.com/item?id=50534387 antirez on the Redis version: https://news.ycombinator.com/item?id=50586785 My Monitor Was Cutting My Download Speed by 75%, by Oscar Molnar https://blog.tymscar.com/posts/studiodisplayethernet/ Discussion: https://news.ycombinator.com/item?id=50587875

Ask about this presentation

Chapters

  1. 0:00The Riemann Hypothesis, proved in Lean
  2. 0:22The statement Lean checked
  3. 0:49Four checks instead of a reader
  4. 1:17Next, learning why it is true
  5. 1:39GPGME double free
  6. 2:08Valkey drops the skiplist
  7. 2:35The monitor that slowed the network
Show transcript
  1. On Saturday's HN front page, Levent Alpöge, working with Claude, released a proof of the Riemann Hypothesis as a Lean library of 17 million lines, and within a day the Lean and Mathlib communities certified it as correct. A crafted key can trigger a double free in GPGME, the library behind most mail clients that speak OpenPGP. And Valkey swaps the skiplist under its sorted sets for a B+ tree.

    Slide text

    The Riemann Hypothesis, proved with Claude in Lean, and a GPGME double free

  2. The hypothesis says every nontrivial zero of the zeta function sits on the line where the real part is one half, and mathematicians have chased it for more than 150 years. Its statement already lived in Mathlib, written and vetted by experts in Lean and number theory, so the proof had a fixed target. Alpöge's library depends on Mathlib and nothing else, and the same method settles the Generalized Riemann Hypothesis for Dirichlet L-functions. In the note announcing it, Alpöge writes, “Claude far outpaced me near the finish line.”

    Slide text

    Every nontrivial zero of ζ has real part ½, and Lean accepts the proof

    def RiemannHypothesis : Prop :=
      ∀ (s : ℂ) (_ : riemannZeta s = 0)
        (_ : ¬∃ n : ℕ, s = -2 * (n + 1)) (_ : s ≠ 1),
        s.re = 1 / 2
    17 million lines of Lean depends only on Mathlib GRH for Dirichlet L-functions too
  3. Seventeen million lines is more than any person can review, so the trust comes from checks a machine can repeat. Dozens of people compiled the library and got the same raw proof export. Two separate tools confirmed that the theorem it proves is exactly Mathlib's statement of the hypothesis. Then several independent proof checkers, written in different programming languages, accepted the export, and two of them come with a formal proof that they cannot accept a false one. Every check ran on an inert export, out of reach of any malicious code in the sources.

    Slide text

    Four checks let a proof too long to read be trusted

    Builds
    Dozens of people compiled the same proof export
    Matches
    Two tools confirm the theorem is Mathlib’s statement
    Kernels
    Independent checkers in several languages accept it
    Isolation
    Every check ran on an inert export of the proof
  4. In the thread, thejokeisonme recalled a fake refutation of the Collatz conjecture that once got past one of these checkers. TwoFx answered that Lean has since shipped formally verified checkers, and this proof went through them. A professional mathematician posting as math_dandy considers the hypothesis proved, and can't wait to learn why it is true. Alpöge expects the argument to be massively simplified, and plans to write it up properly.

    Slide text

    Mathematicians accept the proof and now want to know why it works

    “I consider RH proved.”

    “But I’m salivating at the prospect of learning why RH is true!”

    — math_dandy, a professional mathematician
  5. If you ship anything built on GPGME, the library other programs use to talk to GnuPG, it needs the new release. An OpenPGP key with a very long user ID triggers a double free, which can leak data out of the application, and remote code execution can't be ruled out. Statically linked programs need a rebuild. In the thread, lazka added that Arch Linux uses GPGME to verify downloaded packages. Red Hat Enterprise Linux 9 fixed this same double free five years ago, after a Coverity scan, and the patch stayed downstream.

    Slide text

    A crafted OpenPGP key can trigger a double free in GPGME

    Affected
    Every GPGME before 2.3, shared or statically linked
    Used by
    Most mail clients, many XMPP clients, Samba, Flatpak
    Missed
    RHEL 9 patched it five years ago, downstream only
  6. Valkey 9.2 replaces the skiplist under large sorted sets with a B+ tree that packs up to 61 members into each node. Commands behave the same, and the memory spent on ordering each member drops by more than a third. Score updates with ZADD run about twice as fast. OBJECT ENCODING now answers btree instead of skiplist, so a test or a monitor that looks for that string needs updating. In the thread, antirez, who created Redis, wrote the version Redis adopted, and says it saves more memory still.

    Slide text

    Valkey sorted sets now spend up to 43% less memory keeping order

    skiplist · overhead per member
    50.3 B
    B+ tree · overhead per member
    28.5 B
    OBJECT ENCODING leaderboard  # was "skiplist"
    "btree"
  7. Oscar Molnar pays for a 2.5 gigabit line, and got 600 megabits down through a USB-C Ethernet adapter plugged into a Studio Display XDR. Plugged straight into the Mac, the same adapter ran at full speed. That Realtek adapter sends four packets per burst, then waits for the Mac's go-ahead, and the display's hub adds about 8 microseconds to every wait. So the adapter drains at about 2 gigabits while 2.5 arrive, its buffer overflows, and the dropped frames force a hard backoff. A newer adapter that sends eight packets per burst dropped zero frames.

    Slide text

    A Studio Display XDR hub held a 2.5 gigabit line to 600 megabits

    2.5 Gbps in →
    Ethernet adapter
    4 packets per burst
    → ~2 Gbps out →
    Display hub
    +7.6 µs per go-ahead
    → Mac
    → An adapter with 8-packet bursts dropped zero frames

What happened in AI today, in your inbox

One short email each evening. Read it and you’re caught up.

Built with ❤️ in New York