Skip to content
Vale, home

The Riemann Hypothesis Is Checked. Understanding It Is Still the Job.

11 min read

Six weeks ago, in Proved, but Not Yet Understood, I wrote that a check can justify trusting an output without creating the understanding you need to extend it, and that mathematics just happens to make that gap easier to see. I didn’t expect it to become this easy to see quite so soon. 🤯

I found out the way we find out about everything these days, mid-scroll on X, sandwiched between memes. On Friday, November 20, Levent Alpöge, a mathematician at Anthropic, posted a four-page note on his personal site that opens with “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” There’s no paper attached, just a Lean archive split into fourteen parts, and the second page explains why: “It being obviously inappropriate to attempt to post some massive unreadable travesty of ‘slop’, I decided to only present Lean.” The write-up he generated and kept to himself “hit 250,000+ pages” (which led the mathematician Thomas Bloom to post on X that he wasn’t sure what surprised him more, the claim itself “or (b) that it claims to have produced a coherent proof that, written out in human-readable format, is 250,000+ pages long”).

So this post is a sequel rather than a new argument. In October the gap was spread across 719 manuscripts of very uneven importance. This time the result is the most famous open problem in mathematics, the formal check is about as strong as formal checks get, the responsible human is a serious number theorist who has been preparing for this problem since he was a student, and the gap is still there.

what was actually checked (as of Sunday)

The Lean and Mathlib communities published a joint statement on Saturday describing the release as “a Lean library comprising 17 million lines of code” and saying plainly “We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct.” What I find most interesting is how they got there. The community wrote its own file that isolates the statement and imports only Mathlib, then compared it with two independent tools against the theorem Alpöge’s library proves (so the thing being checked is Mathlib’s long-vetted RiemannHypothesis, not whatever the 17 million lines happen to define). The exported proof then went through several kernels, including two that “come with a formal proof, relative to set theory, that they cannot accept a false proof”, and everything ran “on an inert export of the proof, out of reach of hypothetical malicious code in the Lean proof sources” (after my own agent installed malware on my server earlier this year, that’s the sentence I’d most like to frame 😅).

Andrew Sutherland at MIT keeps a running log of the checks. When I last read it (updated 16:20 UTC on November 22) all 175,187 modules compiled, the five headline theorems depended only on the three standard axioms and several independent kernels had accepted the full export, while a couple of runs that repeat checks Lean’s kernel had already passed were still going. Scott Aaronson, who titled his post “The stochastic parrots have apparently proved the Riemann Hypothesis” and then had to add an update explaining it was sarcasm, wrote after the con-leche result that this is “a much higher degree of verification than a typical human paper accepted to a math journal. The proof is correct.”

It’s worth being precise about scope, since a lot of the coverage hasn’t been. What the community certified is Mathlib’s statement of the Riemann hypothesis, the release also proves the Generalized Riemann Hypothesis for Dirichlet L-functions, and Alpöge’s broader claim that “the method easily extends much further” isn’t formalized (the Extended Riemann Hypothesis is something he expects only “with more time”). No Claude model is named anywhere. The Clay Mathematics Institute still lists the problem as unsolved, which is simply how its rules work (publication in a qualifying outlet, at least two years, general acceptance by the community). Its president, Martin Bridson, put it well when he said in a statement that “the journey from today’s announcement to a place of real understanding will be long and difficult.” And just for a sense of speed, Anthropic’s own post in August about Claude raising the share of zeta zeros proven to lie on the critical line from 41.6% to 67.2% included the sentence “We don’t expect that the techniques Claude used will lead to proving the Riemann hypothesis.” Three months!

four stages, out of order

For my metacognition literature review in Cognitive Science at Georgia Tech, I spent a few paragraphs on Wallas’s four-stage model of the creative process (preparation, incubation, illumination and verification). I described verification back then as a stage that “can be seen as an internal dialogue between the subject and the creative object”, the moment when the creator turns around and works out what they actually made, and I also noted Armbruster’s warning that the model quietly assumes creativity is linear. Alpöge’s note reads almost like an illustration of it.

The preparation is his, and it’s decades long: “It was my secret in my student days that my plans were for this problem in decades, which meant then that I tried to become familiar with all tools that I could possibly recognize in the problem when it was time.” The illumination, at least in its final stretch, he credits to Claude with real generosity (“this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line”), while being clear that he steered (“I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe”). The Lean community then did one half of verification, the external correctness check, in roughly a day.

The other half, the internal dialogue, has barely started, even for the author, who is unusually candid about it. He writes that he personally has “little mastery of automorphic forms on higher rank groups”, and that in the parts “in which Claude is the expert and not me I am sort of paraphrasing a ‘slop’ summary, which might make zero sense.” So Armbruster’s caveat turned out to be the important one, since the stages weren’t just split between a person and a machine but ran out of order, with a certified check arriving before the person who spent years preparing had his own conversation with the result. His closing promise is almost word for word Wallas’s last stage, scheduled after the fact: “I intend to do my best to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly.”

what this adds to October

In October I suggested that an agent’s PR should come with, among other things, “a named human who has read all of that and can answer why.” Alpöge is close to the best named human you could ask for (an expert in the field, public, accountable and honest about his limits), and Terence Tao made exactly this point in his post, expecting that “the digestion process will take some time, though in this case at least we have a human author to communicate with to assist with this task.” And still, in his own words, he is “presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics.”

So I want to sharpen the October rule rather than drop it. Naming a human is necessary but not sufficient, because responsibility can be assigned in a minute while understanding can’t, and once an agent “flew” on the hard part, the person who steered it is effectively one more reviewer of that part. The honest move, in mathematics and on a team, is the one Alpöge actually made in his footnotes, which is saying which parts you derived and which parts you’re paraphrasing, and I’d much rather read “I set the direction, the agent wrote this module and I haven’t fully worked through it yet” in a PR description than let the author field imply a comprehension nobody has. His other admission will sound familiar to anyone merging agent code too, since he expects he “may have baked lots of strange suboptimal choices into the argument”, with “short arguments to follow fairly quickly” (which is a fair description of most agent PRs that work on the first try and are three times longer than they need to be).

It also complicates a distinction I borrowed from Aaronson in October between verification-first and comprehension-first releases, since this one has the verification-first shape but is one researcher’s personal note, and he describes the Lean-only format as his own decision (so I don’t read it as a company switching sides). Back then my tentative explanation for these choices was the distance between the new result and what the receiving community already understands, and I’d now add that the shape of a release also depends on the particular people and the particular moment. This moment includes the announcement coming the day after Anthropic’s record stock market debut, which Alpöge addresses head-on (“There is an obvious cynical view, to do with the timing, but it is false”), while the rough timeline Claude compiled for the note ends with “11/16/26: GRH,” a few days before. I can’t judge anyone’s motives from a PDF and nothing in this post depends on them, but it’s a reminder that the norms are still mostly being set by whoever ships first, with institutions like Clay acting as the slow counterweight.

where this argument is weakest

The strongest objection is one I’m half inclined to make myself, since nobody understands all of modern mathematics either (Wiles’s proof of Fermat’s Last Theorem stands on decades of other people’s work, and the classification of finite simple groups is spread across thousands of pages and many authors). In my Cognitive Science coursework I argued that scientific research is a textbook case of Hutchins’s distributed cognition, socially, temporally and spatially, and I wrote that I completely agreed with the framework. So if understanding in mathematics was always distributed across a community and its tools, why should it matter that no single person understands this proof yet?

I think that objection is mostly right, and what it changes is the question rather than the answer. Distributed cognition works because the parts of the system talk to each other through representations the other parts can read (that’s why Hutchins’s cockpit works, since the pilots and the instruments share legible state), whereas a 17-million-line Lean library is a representation that one component, the kernel, reads fluently, while for everyone else it’s very close to an oracle.

What encourages me is that the rest of the system is visibly assembling around it. Tao points to the Lean Zulip thread as the place for efforts “to verify the proof and extract the strategy of proof”, and AI is turning up on the human side too. Hours after Alpöge’s note, Ravi Vakil posted a draft proof of GRH by the Stanford postdoc Alessandro Masullo (“It is all human, with no AI”). About eight hours later he added that several people had reported “problems in the proposed proof, identified with the assistance of different AI models.” It isn’t really a priority fight (Vakil makes “no claim as to the correctness of the proof strategy”, and the draft is being revised), and John Baez made the point I’d want to keep: Masullo’s approach is quite unlike Alpöge’s, so “even if one of these proofs turns out to be right, it won’t make the other one redundant.” Catching errors is still correctness work, but Baez’s point is about understanding, since an approach is something you can learn from whether or not it survives, and that’s the hopeful version of the objection I raised in October that the comprehension bottleneck might be temporary.

The other thing that softens my worry is the asymmetry I pointed out last time. A true theorem stays true, so mathematics can afford to bank a checked proof and spend the next decade understanding it (Clay’s two-year rule is, in a way, a comprehension buffer written into the prize). Alex Kontorovich at Rutgers put the working mathematician’s version of this to the AP better than I could: “the computer is extraordinarily picky, way pickier than I am and so if you can satisfy the computer, there’s a very, very good chance you’re going to satisfy me too,” followed by “But I have to do the work.” Software doesn’t get the two years, since someone has to change the code next quarter.

the slow part, again

I still want the agents to cook, and this weekend showed what that looks like at the very top of the range. A machine-assisted proof of the Riemann hypothesis is, as far as anyone can currently tell, correct (an astonishing sentence to type) and the people checking it did so with a rigour I’d love our industry to borrow. What I’ll be watching for over the next couple of years isn’t another certificate, though, but the first human write-up that makes a number theorist say “oh, of course”, because that’s when understanding finally catches up, and Alpöge has already said that’s the work he intends to do.

Going forward I’ll keep in mind that a named author is a promise to understand rather than proof of it, and that the honest thing is to say how much of that promise has been kept so far.

Anyhow, stay tuned, and keep on coding! ✨

things worth reading

← all posts