Levent Alpoge Claims Proof of Riemann Hypothesis
56 points by mitsuhiko
56 points by mitsuhiko
Reportedly, 250,000 pages. May as well ask an oracle for an answer.
Not really. Even his summary already reads like it has enough detail to let experts in the area start understanding the roughest sketch of the approach.
The 250,000 page result is evidently a mess, and will certainly be cut down in size, with a lot of combined human and AI effort over the coming months.
I don't know much about frontier mathematics but as I see it the complexity can be:
Which is it? Obviously a combination of all, but if there weren't much of the first kind, I'm guessing we would have found a proof already?
From page 3 of Levent's letter:
I also emphasize that I may have baked lots of strange suboptimal choices into the argument, and expect short arguments to follow fairly quickly, though I have not had time to search for such myself (yet!).
So the expectation from the author is that there is a shorter, simpler proof. How much it will shrink, I don't know, but at least right now we can expect a heavier portion to fall in your third category.
Seems very plausible it's a methodological artifact. AI models are compulsively verbose in almost anything they do, and the more passes they take over something the more wordy and repetitive it gets. This is true for prose as it for code. Even AI-generated visual media often have this distinct horror vacui aspect.
Hurried Announcement: Alessandro Masullo GRH
Alessandro Masullo has a proposed proof of GRH, and Brian Conrad has been looking closely at it for months and gave extensive feedback which is being incorporated, and it is so far looking good. It is all human, with no AI. The hope was to bring Alessandro to Stanford for a year to polish the writing and then announce it after running it by enough people. Now it seems best to put out the current draft.
1000 pages, vs 250k in the slopped up proof.
If it turns out that the reviewers used Claude for their work, and this ends up being a repeat of the Navier-Stokes situation (except even higher-profile)... this might just be the funniest timeline.
For some extra irony, Levent Alpoge is one of the people who worked on the Navier-Stokes proof that OpenAI (allegedly) stole.
edit: quoting the Acknowledgements section of Masullo's paper,
This work started off when the author was nearing the end of his first year as a Graduate Student at Stanford under Brian Conrad. It took nearly 12 years to assemble. [...]
If any message at all this work is meant to send to the mathematician who feels discouraged by the pace and tone of the current climate, it is this: we are still here. We are very much alive. We are not going anywhere
There are apparently hundreds (thousands?) of other mathematical proofs that start with "assuming the Riemann hypothesis...", so if this holds up those all become established at once.
Any particularly important? I've heard the same but I don't know which ones have the most importance.
I'm not sure if this counts as one of the most important ones, but if my memory serves me correct, there are implications for prime number tests. There is a relatively slow polynomial algorithm that show definitively whether a number is prime or not (AKS), and there's another, faster algorithm (Miller-Rabin) that, if you run it a certain number of times, shows definitively whether a number is prime or not if the riemann hypothesis is true. Which means if it's true, we can check reliably whether numbers are prime in a faster way.
The Miller-Rabin test usually refers to the probabilistic version. The so-called Miller test is the deterministic one. And indeed GRH implies that the latter algorithm runs in polynomial time, but funnily enough the quasi-RH result proved by OpenAI in October already sufficed for this implication. See discussion in this MathOverflow thread.
Miller-Rabin is an interesting case because if you run the probabilistic version about 50 times or so, the chance of the algorithm failing is less than the chance of your hardware failing. If you get the wrong answer, it is more likely to be caused by a cosmic ray flipping a bit, etc, while your program was running.
Compare with the deterministic version of Miller-Rabin, which if I understand correctly requires running 2(ln n)^2 iterations of the core algorithm (assuming RH). That surpasses 50 iterations as soon as you get to n=149. So yes, it would be "deterministic" but not meaningfully more deterministic than hardware in the grubby real world.
Still exciting stuff, though. :-)
Amazing if true (in the sense that there's no subtle hole in the proof, obviously I don't assume the person in question would try to lie).
The timing of this being released right after the IPO feels a little cynical though.
The timing of this being released right after the IPO feels a little cynical though.
they, er, elliptically address this in the document. certainly seems to me it would be weirder to hold on to it longer.
For the record, Levent Alpöge has already been in the news months ago in relation to the Navier-Stokes problem, where OpenAI announced a proof and then Tristan Buckmaster and he claimed they were already working on that and OpenAI may have (deliberately or inadvertently) worked off of their unfinished progress.
The timing of this being released right after the IPO feels a little cynical though.
They couldn’t have released it during the quiet period.
LeanProver zulip chat. Doesn't look like anyone else has finished compiling it yet (but given how large it is, this is unsurprising). Edit: someone has finished compiling, and is now running comparator.
Tristan Buckmaster (with Alpolge) also claims to have unforced finite-time blowup with Navier Stokes.
There are so many instances of experts in one field accepting the glaring problems with someone's work because they see it as intersectional. This intersectionality bias is something that interactions with language automation have been taking advantage of. There are currently no experts in understanding the claims of the outputs of automation absent of the process known to work outside of automation. This bias is why we don't have big stories about the extended exploration of proofs either automated or not, and again an output carries the weight of an output when in fact are two different things, categorically, and should not assume to have many overlapping qualities. This is obvious with proofs, but really just about any output you see has this intersectional bias. Writing about it gives this bias away since the intersection is outside of the scientific process.
There has also been this. https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/
An entirely human generated 900 page proof.
Meta but related, from "Math StackExchange" What does proving the Riemann Hypothesis accomplish? tl;dr: not much. The most important thing is that a bunch of other theorems will be automatically true as someone else already pointed out.
Funny: Claude didn't know that was claimed to be proven by Anthropic today, even though it said it searched the web.
Prompt:
what's the big deal about the riemann hypothesis? is it easy to state?
Claude:
... (explanation) ...
Recent news: it’s still unproven. Last month there was a notable related result. OpenAI claimed to settle the “quasi-Riemann hypothesis,” a weaker version of the Riemann hypothesis, which is arguably the biggest open question in math. Roughly, the weaker version says the zeros stay inside some strip around the critical line, not exactly on it.