Did Claude discover a proof of the Riemann Hypothesis?

  • Topic: AI 
  • Thread starter mfb
  • Start date
  • Featured
39 replies · 2K views · last reply
Messages
37,577
Reaction score
14,448
https://alpo.ge/remark.pdf
I believe I have, working with Claude, discovered a proof of the Riemann
Hypothesis. In fact, the method easily extends much further, resolving the
so-called Generalized Riemann Hypothesis, and I expect, with more time, the
Extended Riemann Hypothesis too.
By Levent Alpöge / Anthropic

Can't judge it myself, we'll see what the community will say about it.
 
  • Like
  • Wow
Likes   Reactions: sbrothy, atp_ent, Greg Bernhardt and 4 others
mfb said:
https://alpo.ge/remark.pdf

By Levent Alpöge / Anthropic

Can't judge it myself, we'll see what the community will say about it.
Big if true. Again it is easy to post that they have solved it without understanding the proof. We don't yet have a digestible proof of Navier-Stokes, Yang-Mills or this one.
 
The RH proof is supposedly so much lean code that a written out proof is estimated (by Alpoge I think) to take 250,000 pages lol. Nobody gonna read all that. But maybe they will have AI compress it down to a few hundred.
 
A separate supposedly entirely human generated proposed proof was released in a hurry today as well. It's about 1,000 pages and has been in review for some months.

Hi everyone,

I am at the American Mathematical Society Executive Committee and Board of Trustees meeting now, and about to chair an hours-long meeting, but this is urgent, so I will try to get this out quickly. 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. I hope people quickly check it, formalize, fill any small gaps if they occur. I really hope it works. I have great faith in Brian’s judgement, and Brian has put a lot of thought into checking what Alessandro has done. There is a long story to tell, but right now I want to get the file out. Here it is.

https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/
 
It's a bit wild that after the NS announcement, the partial Yang Mills, ETH, etc. announcement, they didn't immediately rush to announce/publish this.
 
But I guess if they have all the timestamps to back up the timing.. credit could be shared between humans and AI? I don't think I have enough popcorn for this scenario.
 
  • Like
Likes   Reactions: bhobba and Rodney182
Matterwave said:
It's a bit wild that after the NS announcement, the partial Yang Mills, ETH, etc. announcement, they didn't immediately rush to announce/publish this.

I believe that there is a high likelihood that Alessandro's proof could have errors or be incomplete. I don't think they've had the time needed to finalize and verify everything.

If he rushed it out before it was done, then someone using AI could rapidly absorb it and scoop him. Mathematicians in the age of AI are becoming afraid to share their work too soon.
 
atp_ent said:
I believe that there is a high likelihood that Alessandro's proof could have errors or be incomplete. I don't think they've had the time needed to finalize and verify everything.

If he rushed it out before it was done, then someone using AI could rapidly absorb it and scoop him. Mathematicians in the age of AI are becoming afraid to share their work too soon.
That's fair, but now who gets the "who solved it" credit? 😂

If either has a critical error then I guess the other would get credit. But assuming there's no critical errors..
 
Matterwave said:
The RH proof is supposedly so much lean code that a written out proof is estimated (by Alpoge I think) to take 250,000 pages lol. Nobody gonna read all that. But maybe they will have AI compress it down to a few hundred.
Correction on this, Alpoge doesn't seem to be the one to give the estimate.
 
As described in the OP link, this reads more like a collaboration than a pure AI proof. There appeared to be a lot of guidance and suggestion of intermediate goals and strategy by the human mathematician.
 
  • Informative
  • Like
Likes   Reactions: Matterwave and Greg Bernhardt
PAllen said:
As described in the OP link, this reads more like a collaboration than a pure AI proof. There appeared to be a lot of guidance and suggestion of intermediate goals and strategy by the human mathematician.
I think this is true, but given how long the proof is in sheer scale, I wonder how much of it really was guided by Alpoge. It's just purely my speculation.
 
It does look like Alpöge helped significantly to make it happen, but he also admits he doesn't understand the proof yet.

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....

I am presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics. It being obviously inappropriate to attempt to post some massive unreadable travesty of "slop"2, I decided to only present Lean.

https://alpo.ge/remark.pdf

The parts he knows about and understands, and the massive unreadable travesty of "slop", as he puts it, might not even match the lean or comprise a valid and complete proof.
 
Matterwave said:
Correction on this, Alpoge doesn't seem to be the one to give the estimate.
Actually, he is. It is in footnote 2 of his post. But in this footnote, he gives an example of how textualizing the lean formalization leads to ludicrous detail no one would ever write in a paper.
 
PAllen said:
Actually, he is. It is in footnote 2 of his post. But in this footnote, he gives an example of how textualizing the lean formalization leads to ludicrous detail no one would ever write in a paper.
Ah it does appear to be there~ I saw a YouTube video attribute it to someone else so I didn't want to give the wrong attribution.
 
I went to a talk about AI a number of years ago, where the speaker was showing trend lines and making the case that there was no sign of them slowing, and that so far, milestones arrived sooner than predicted. I asked him about Riemann really falling within 10 years, and he said “maybe, why not”. I thought this was lunacy at the time, but here we are - sooner than my proposed 10 years.
 
Matterwave said:
The RH proof is supposedly so much lean code that a written out proof is estimated (by Alpoge I think) to take 250,000 pages lol. Nobody gonna read all that. But maybe they will have AI compress it down to a few hundred.

The lean code is 17.4 million lines.

https://github.com/CocoSgt/riemann-hypothesis-lean

It's surreal that there is a proof for GRH that we can be fairy sure is correct, but nobody can understand it and probably nobody has even read it.
 
  • Like
Likes   Reactions: Rodney182 and pines-demon
atp_ent said:
A separate supposedly entirely human generated proposed proof was released in a hurry today as well. It's about 1,000 pages and has been in review for some months.



https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/
If Masullo is providing a purely human proof it will have an inconvenient and an advantage. The inconvenient is that it can be safely be archived with the thousands of human proofs of the Riemann hypothesis that ended up nowhere, but if he insist enough somebody might read it and prove him right. The advantage is that it is human made so he understand and can explain his proof before we even understand the Lean proof by Alpöge.

Masullo effort could even beat Anthropic if he can give it to an AI to do the Lean check, that way we will an understanding and a proof.
 
Masullo's work is too long to look at, but browsing through it I cannot see the new ideas. It all seems quite the standard things in the automorphic world that have been around for a while now. Of course if Brian Conrad has been looking at it for some time, it must be promising. May be he should write a short (say up to 100 pages) sumary of the ideas.
 
martinbn said:
Masullo's work is too long to look at, but browsing through it I cannot see the new ideas. It all seems quite the standard things in the automorphic world that have been around for a while now. Of course if Brian Conrad has been looking at it for some time, it must be promising. May be he should write a short (say up to 100 pages) sumary of the ideas.
I just saw this invalidation: https://philarchive.org/rec/NIEIOM

Again Masullo's proof can be almost safely archived until he gets enough people to a validate his version (or correct his errors if he has any).
 
Matterwave said:
But I guess if they have all the timestamps to back up the timing.. credit could be shared between humans and AI? I don't think I have enough popcorn for this scenario.

And then the wrong way around?

Claude et. al :-p
 
It seems to say the lean code starts only from 3 foundational axioms. Doesnt this imply that perhaps 10 million or more of the 17 million Lean lines is just building up well known math?
 
PAllen said:
It seems to say the lean code starts only from 3 foundational axioms. Doesnt this imply that perhaps 10 million or more of the 17 million Lean lines is just building up well known math?
It would be interesting to know how much percent of that is just basic math.
 
Greg Bernhardt said:
I think it would be an interesting thread to start maybe in the Topology forum detailing at least some of the details or background of the proposed proof. In general, rather than attacking the zeros of the Riemann zeta function directly, the proof targets an equivalent formulation involving the von Mangoldt sum:

##\displaystyle\psi(X) = \sum_{n \le X} \Lambda(n)##

where ##\displaystyle \Lambda(n)=\begin{cases} \log(p)\text{ if } n=p^n \\
0 \text{ otherwise}
\end{cases}
##


The Prime Number Theorem already shows that ##\psi(X)## grows roughly like ##X##. Proving the Riemann Hypothesis is mathematically equivalent to proving that the error term fluctuates within a square-root ceiling:

##\psi(X) - X = O\left(X^{1/2}(\log X)^2\right)##

and my understanding is that proving this square-root ceiling is the basis of this proof.
 
  • Informative
Likes   Reactions: Greg Bernhardt
The formalisation of Fermat's Last Theorem provides an interesting basis for estimating how much Alpöge and Claude's enormous Riemann hypothesis proof might eventually be compressed.

Consider the comparison:

Fermat's Last TheoremProposed RH/GRH proof
Human-readable proofApproximately 129 pagesOver 250,000 pages (generated draft)

Anthropic reported that its Lean formalisation of Fermat's Last Theorem followed an existing exposition of Wiles's proof by Darmon, Diamond and Taylor. The conventional published proof occupies approximately 129 pages.

This is striking: 13 million lines of Lean can correspond to a mathematical argument expressed in roughly 129 pages, when written for experts who can rely on established theorems without proving everything from first principles.

The lean code is 17.4 million lines for the RH/GRH. If we assume that the relationship between Lean code length is the same for both proofs, we obtain

$$
129\times\frac{17.4}{13}\approx173\text{ pages}.
$$

That would be an enormous reduction from the reported 250,000-page generated manuscript.

However, 173 pages is only a very rough estimate. The two proofs involve different mathematics, and we don't know how much of the Lean code is needed to establish already known results or how much involves genuinely new mathematics. So the final human-readable proof could be much longer.
 
  • Like
  • Informative
Likes   Reactions: pines-demon, Matterwave and Greg Bernhardt
PAllen said:
It seems to say the lean code starts only from 3 foundational axioms. Doesnt this imply that perhaps 10 million or more of the 17 million Lean lines is just building up well known math?
I think every valid proof should basically start from those 3 axioms. It's a standard check you didn't silently add new axioms. Note that an axiom in lean is different than e.g. an axiom of group theory which, in lean, is more considered a definition.

Such basic definitions are already in mathlib (standard lib of Lean). So it depends a bit on what you mean by "well known". The model could have:

1. Rebuilt parts of mathlib.
2. Built known definitions into mathlib which weren't in mathlib before.
3. Built brand new definitions.