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.



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.
Every LLM code review ever. “No idea, you check it if you feel it’s necessary.”
I don’t need to double check it when I already asked it to not make any mistakes.
Sort of, the reason I heard and makes the most sense is that these proofs were more a side effect of their training. They’ve been training the models to do math, first they started with basic math but the problem set became “saturated”, ie. It could answer every question. Then they move on to advanced math and that becomes saturated so eventually you have to give a model the hardest math questions known to man which are unsolved. Eventually the AI solves them when they give it enough compute, time and reinforcement learning cycles and they’re left with these proofs.
Their motivation for releasing them probably was for a PR win and they also probably did think they would be advancing math, but that’s not why they threw hundreds of millions of dollars of compute into solving these.
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 bloated lines of Lean.
This is new to me, thanks for the input - but have to read it first
Thing is, the AI puked out 250k pages, while the human brought up a possible proof in 900 and something pages
Yeah, AI can maybe cross reference faster with greater datasets and we should use that, but let’s not jump to conclusions here.
As AI can’t really think, it’s just statistical evaluation. So all the ground work was already done.
But if this is actually true and not just usual PR bullshit, we can at least use it.
Still, something like singularity, like those AI companies seem to propose, is more than far away
Cross referencing large datasets is hard for a human, so if we have a tool for that, that would be good.
But all the stories about AI trying to solve their tests by breaking out and hacking servers really makes me…not so sure
With math, there’s a little more nuance than just the ground work being done. It’s one of the fields where there’s actually potential value.
With a novel or something the way to review quality is to have a human check it. It’s slow. Verifying facts is hard.
With math, it learns the language from existing patterns, but anything it wants to generate can be checked for accuracy very fast. This means that what this type of AI is good at (replicating subtle patterns in the training set) can be trained, refined, and grow it’s dataset very fast.
That’s why some people who are not AI cranks are excited about the prospects of AI for certain categories of exploration: it can be automatically objectively verified.
An analogy for the CS people: Generating the proof/term is in NP, but (type) checking the proof/term is in P.
(All analogies are flawed, and this one is no exception.)
LLMs and static type systems are a near-perfect marriage, since the type checker serve as a fast, perfectly accurate oracle. The fact that dependent type systems and constructive mathematics are basically the same thing means LLMs get better at maths much faster than things with worse oracles.
The reason it is so large is because it was not allowed to assume anything, and not everything is in Lean libraries. So it had to do years of work to build the foundation to then be able to solve the big problem.
No this is just a line by people that are trying to be mean to AI as if they had feelings. Thinking is not a term exclusive to what happens in the human mind thinking can be put on paper, it can be shared and collaborated on.
And had Anthropic’s not been published, then another AI would have scooped the solution. It’s “First!” mentality all the way down.
Only the true scholar had the patience a professionalism to hold the solution until validated in triplicate.
A proof had been verified. How can we be sure it has been correctly stated?
It’s not that hard to just read. Jesus christ.
Does not mean error free.
Okay, so you’re just sealioning at this point. The standard of certainty that’s been met here is already vastly higher than what a traditional, hand-written proof undergoes.
If you want, I linked the formalized statement (pointing this out because you have a hard time following and reading links), which you can go scrutinize yourself or refer to any number of mathematicians if you suspect something’s wrong with it.
In reality, you have no understanding how any of this works and don’t care to even try to find out, so instead you’re pissing out a stream of vague, flimsy objections like a Dollar Tree Socrates.
Dunning Kruger strikes again. Go shout at the clouds around Cambridge University.
The abstract of the paper that video links to reads:
If you think this statement of the Riemann Hypothesis was not only machine-translated into Lean (it wasn’t) but that it somehow contained a mistake that nobody caught because of that fictional machine translation, then Dunning–Kruger didn’t just strike – it dropped you on your head repeatedly as a child. Or more likely, you didn’t read a word of that article to begin with, and you’re still just throwing flimsy bullshit at the wall because you don’t actually care to try to know what you’re talking about.
Because, again, you’re chronically terrible and lazy at actually investigating sources, here’s what that formal statement of the Riemann Hypothesis looks like:
Can you personally guarantee, after checking all 250,000 pages of underlying AI-generated text, that Claude powered Autoformalisation didn’t mistranslate an imported result?
No. You can’t.
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 hate this AI trend, and that a proof of a theorem is now reduced to check wether a >200k LOC program compiles is disgusting, but that a top mathematician can direct a swarm of bots to solve the holy grail of mathematics in ~2 months is incredibly news-worthy. This is moving the ground of the academic world, a lot.
I agree, but then the novelty is the capability of these statistical models. The feat on its own is not what’s novel about it.
Or to put it another way, maybe it takes some level of expertise to solve problems like this with a statistical model. Maybe even I could pull something like this off if I worked on it for a few years with infinite search tokens. But it takes a completely different level of expertise to solve it manually the way a human would. There’s a lot of people who could do it with a search algorithm - there are maybe 100 or so people in the world that could do it without one of them.
But at the end of all of this, code is still just code, whether it’s C, C++, Java, or Lean. They pretty much all work the same way. I will admit that the 2 month timeframe is impressive, but we’ve already known for a couples years by now thatLLMs can brute force their way to a solution if you give them a problem statement and an acceptance criterion…
What techniques like this can be used for in the future does sound quite incredible though.
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, the proof is literally in the pudding.
The whole point of Lean is the strict type-checking. In effect, the proof says “this is the Riemann Hypothesis, and this is the proof” and you can be 100% certain that the proof is valid.
100 million tons of pudding; yep the proof is in there somewhere, get digging.
BTW, the proof looks exactly like pudding, good luck.
Yeah, and?
It’s still a valid proof.
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…
Years of time and effort from specialists. Turning it into a Lean program would help a lot but that’s not a trivial effort, it’d still take considerable time.
The culture around mathematical research is changing but not the way you’d like. Mathematicians are being more secretive about their research so that it doesn’t get sniped by a guy with a Claude subscription and published in a half-baked low quality format just to say “I did it first.”
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 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.
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.
Anthropic said they don’t intend to take the money either way
Superb, saves us the effort of creating pressure
Disagree - we still need to keep the pressure on to make sure that Masullo is recognized as the one who gets credit for it. He did the hard work can came up with the more meaningful proof, and assuming that it holds up, he should get the public recognition.
And even if it doesn’t, I don’t see any reason to publicly recognize Anthropos as the group that solved it, cause anyone could dish out 100k pages of slop to prove anything. Personally, I don’t consider that novel or noteworthy, just like all the massive slop software projects out there.
U know what fair point, i agree with you. Little bit too rushed to call off pressuring. Still good they’re not gunning for the prize but yeah i agree with u actually
Ohh, I like that. It’s very much in line with “If you can’t own it, it’s not theft. Yarghh!” I like the consistency. Fuck these Bros who want their cake and eat it, too.
Everytime AI is in the news it’s a marketing play. As simple as that. They can claim anything and it gets on all the news in all the world at the same time and everyone is talking about it. We know it says yes to anything and nobody can read that output, but whether it’s true or not doesn’t matter. Everyone has already done the advertisement for them.
It’s only a valid proof when it’s peer reviewed. The ai one isn’t
No, that is false.
You are confusing publication standards with validity.
It is a valid proof because it has been formally verified.