I haven’t seen people talk about this at all, so thought I’d share this article about it. For those depressed that no one will actually understand the RH Proof: we have another, entirely human proof which means there are humans who understand it. It might have been overshadowed, but assuming it ends up working out (I have heard there have been some problems found, hopefully they are surmountable) then I think it will be the far more impactful result.



So if I get this right, the human proof was in the process of being finalized and checked when the AI one was released which forced their hand and made them release it early and not completely finalized.
It also says that the AI solution can’t be read by humans while the Human one can. I’m no matemathician but a solution that cannot be understood by humans seems a bit limited. How do you check it? I guess you can check it empirically but that seems a bit primitive. How do you teach it to future mathematics students? Mathematics relies on building and developing previous work done by people in the field. This means studying and understanding that work and refining it or using it for new work. You can’t do that with something you don’t understand.
Apparently Masullo has been working on RH for more than a decade, and spent the last year polishing his proof and getting feedback from colleagues before releasing it. He had planned to spend another year making extra sure his proof was perfect before publishing, but was forced to share an unfinished version because of this recent announcement.
Meanwhile Anthropic shits out a 17-million-line proof and says to the world “idk, you figure it out if it’s right or not”. Kinda makes me cynical about their motives here. Seems like they’re much less interested in advancing the field and much more interested in getting positive press to juice their IPO.
That’s not what’s happening here. The proof is formalized in Lean and thus can be programatically checked. The proof is verifiably correct.
The problem isn’t “if it’s right or not”; it’s whether actual insights can be derived from the proof – where right now it’s 17 million lines of unintelligible nonsense.
Spot on. When Alpoge published his proof, he was gloating endlessly about how he did little work and just had the model do everything. His announcement sounded like a press release written by a 12 year old.
I’m not a mathematician, but his proof sounds absolutely useless, and yet all the press is fawning over this result rather than the one that Masullo was working on, ie. the guy who did everything the right way.
I knew a couple students in special needs class that genuinely wrote better shit than his announcement. So, that’s just an insult to 12 year olds to compare the two.
A Lean proof is relatively easy to mechanically verify and this has been done independently already by several groups using different Lean kernels to preempt the possibility of a Lean bug. For this particular question, the existence of a proof itself is remarkable in its own right. If Masullo’s proof can be verified, it would be an even bigger deal because it can more readily translate into human digestible techniques, but until/unless it is fleshed out and accepted, it’s far less interesting. A lot of work on the Riemann Hypothesis has been done for decades, it wouldn’t be close to the first incorrect proof produced by legitimate experts.
A question I have that I don’t think I’ve seen answered anywhere. We know that the Lean project does compile and validate, across several Lean kernels. But do we know if that Lean project actually is a proof for the Riemann Hypothesis? Or could it just be validating something (even marginally) unrelated.
Let’s say I created a Lean project that verified if 1 == 1, but the actual lean code was so incomprehensibly large that no-one could understand what it was verifying, just that it did verify eventually, and said that was my Riemann Hypothesis proof. Would anyone have anything other than my word on that?
Yes that’s one of the criteria that were verified by humans. In the case of the Riemann hypothesis it’s not that difficult because it’s not that complex (only proving it is difficult, spelling it out isn’t), and there was an existing human curated formalization of it too.
You can read more about it here: https://leanprover-community.github.io/blog/posts/riemann-hypothesis/
What does it take to validate Masullo’s proof though? My understanding is that it has to be examined by other experts and manually verified using existing expertise, but couldn’t it also be written in Lean as well and validated that way too?
In any case, maybe the culture around proving things in mathematics does need to change. If people doing hard research for years get sniped by people messing around with models, then maybe they need to publish more incremental progress instead of only releasing it when it’s done and thoroughly polished…
I would not be surprised if someone got an AI to fill in the missing bits and fix the errors by making it do exactly that. Assuming the author really was on to something of course. It wouldn’t be the first human attempt to prove it that turned out to be wrong.
I hope we can all band together to make sure anthropic gets 0 of the prize money, and that it all goes to Masullo instead. Cant be held legally liable for your fuck ups? Cant get praised for ‘your sucesses’ either then.
It’s only a valid proof when it’s peer reviewed. The ai one isn’t
The AI solution is formalized in LEAN in such a way that it’s easy to verify by humans that it’s formalized correctly, and from there the proof as a whole can be verified by software. It’s unreadable because it’s 17 million lines of LEAN and it’s hard to get a full understanding of it, but there is now a strong consensus that it’s correct. A working LEAN proof is a much higher standard than proofs in human language are generally held to. Humans can now go through it and look for techniques that look interesting and novel and build on them. Likely there will be a shorter proof at some point in the future.
The human solution is human readable because it’s written in human language. But it’s still almost 1000 pages of text and it has to be understood by humans to verify it. By traditional standards for human work, a few mathematicians need to be convinced it’s correct and put their name on it. If we want to test it by the same standard as the AI solution then someone now has to translate 1000 pages of English into LEAN. However it was rushed out, and it’s fairly easy to see that it has holes because it literally has chapters that are empty. People are also claiming that AI agents can trivially find errors in it, but I haven’t verified those. However, human proofs of the Riemann Hypothesis that turn out to be wrong are kind of a regular thing, so I would not be surprised. If it’s salvageable though it’s still worthwhile because the human proof goes in an entirely different direction than the AI one.
Unreadable machine proofs are not a new LLM era thing by the way. The first machine generated proof that was considered unreadable wasn’t something from the last few years or even from this century, it was the one for the Robbins conjecture in 1996.
Just wanted to answer your “how to verify” question:
The proof is written in Lean and is verified by the Lean tool chain. As long as Lean is correct as a tool, the proof can be trusted.
Understanding and using the proof or parts of it to advance math is a different undertaking.
I’m not sure I do trust Lean though. The tools I work with are getting patched for odd edge cases every day. Am I expected to suspend all disbelief and assume that there’s nothing dodgy in the 17M lines of unreadable puke that make up this proof? How confident are we that an agent wouldn’t exploit a bug in Lean to achieve its goals?
One part of the verification is that the proof is checked by multiple independent implementations of LEAN. A bug in the software would have to be one that has to be present in all of them to not be caught. There absolutely have been bugs before that were caught this way. But this time they seem to agree.
Calling it 17 million lines of unreadable puke is also not justified. There is a structure to it and part of the reason why it’s so much is that the AI was not allowed to introduce assumptions and went into incredible detail to prove every last little thing that went into the proof. When people make fun of things like it containing a separate proof for real numbers that are not 1/2 to be either less or greater than 1/2, realize that the alternative where the AI is allowed to just write this sort of thing down as an obvious truth like a human would means there is a list of such assumptions that have to be verified individually. So this is not a sign of slop but a tactic to make the proof easier to verify.
LLMs don’t “make assumptions”, they’re text generators, the reason it spat out 17 million lines of Lean code is because it was a brute forced random walk of the problem space.
That a human could do in 980 pages what a LLM took 250,000 pages to do, is not proof that the chatbot “went into more detail”.
Assumptions in the mathematical sense. Not talking about how LLMs operate but how the proof is structured.
The human with 980 pages has entire sections missing and errors in others. We don’t know how many more pages are needed, but we do know every single one has to be verified manually because unlike the AI he does not have anything that can be checked by a machine.
If there’s no bug in Lean it doesn’t matter how bad the slop is, as long as it checks out. Big if, though
Your broader sentiment about understanding and teachability are fair, but…
What? That’s the whole ballgame. It is what separate maths from other fields. Rigorous, empirical proof.
Think it’s fair to say from context they meant that there’s a difference between a proof we can verify works correctly based on the answers it gives, versus a proof that we can also understand how and why it gives the correct answer (because sometimes the real useful stuff comes from the working out rather than the answer).
And I agree with that entirely. Just saying empirical proof isn’t primitive. That was a bit out of place it felt.
strange isn’t it that both proofs have come with humans being very close at the same time
the next challenge is obviously to make clear why this machine-code-like LEAN proof is valid and useful
That’s not the next challenge at all?
It’s basically a pointless waste of time. Maybe an AI proof could substantiate the existence of an actual proof in principle (assuming we can even confirm its validity), but that’s about it. We only care about proofs at all because of what they tell us about the mathematical community and the intellectual interest they generate. To that end, an AI proof is basically worthless.
It’s not worthless if it convinces mathematicians that something is definitely true as opposed to being merely hypothesized and widely believed. It changes the types of questions we ask and the way we think about results that relate to RH. Something which turns out to conflict with RH is now provably false. Lines of reasoning that assume ~RH are now provably unsound. The questions about RH itself turn from “is it true” and “how do we (dis)prove it” to “what can we learn from the proof.” There’s undoubtedly still room for entire new theories to support a human-comprehensible proof, it’s just that now we have something concrete to dig through for inspiration.
deleted by creator
I cannot verify these claims myself as I’m not a mathematician, but from what I’ve read elsewhere, but a lot of mathematics is predicated on the RH being assumed true yet unproven.
As others have mentioned, it was proven using a theorem prover called LEAN which can validate that the proof is valid, which I imagine works in the same way that a compiler processes human-written software to prove that the program that was written is valid.
I think the proof itself is probably not very useful if it’s not human readable, and while it seems like mathematicians have accepted it already, there is a small chance that it could always be invalidated if there’s a bug in the LEAN compiler or its libraries. Hopefully it can be reduced from machine-written code into something more intelligible. And maybe the techniques can be studied and incorporated into future proofs.
It’s less that it’s human readable and more that’s its just really really long and overly detailed.
It’s like saying moby dick is unreadable to a child and thus worthless.
It’s both a right and wrong point of view.
It’s not absolutely worthless, but it would be more useful if it could be understood.
For example, I could write a C program with a new and slightly different version of
printffor each time I call it. That’s a lot less useful than a general version ofprintfthat everyone can use.In this case, a proof is still a proof, but the proof itself won’t help others prove things either. That said, if these statistical models can be used to prove other longstanding problems in mathematics, then that actually does sound like something useful.
Great article, thanks 👍
Anyone with any reason or sense at all would take those AI shitbirds and tell them to go fuck themselves.
But, apparently, nobody has any fucking sense anymore, so we’re all jerking off over the things AI can do, even if it can’t really do it at all