2870 points | 42h ago | Discuss on Hacker News | Back to Radar
Edit: just saw the Lean code is included in the footnote. Considering the author's reputation, I think it's fair to remove the qualifications from the post title, by the standards of today, it's proven.
After the excitement of the Navier-Stokes solution, only for it to eventually seem to not really solve the full problem and leaving discourse about the problem in a very muddy state, I'd rather wait and see if it turns out this proof also makes some assumptions that ultimately limit its impact.
It is probably still some form of breakthrough, but I will expect it to not be as big as a full general proof of the Riemann Hypothesis, until proven otherwise.
Edit: A thought though, if this proof is correct and can't be made much shorter than the current ~250k pages, I wonder if the result has more value as a sign that human accessible proofs to these kinds of hypotheses were stuck behind some mathematical abstraction/tool that we haven't quite discovered yet? Sort of like how you can brute force algebra with real numbers for working with AC circuits, but it's pretty ugly and painful compared to using phasors. If so, maybe the real contribution from AI-driven high profile proofs would come from mathematicians studying the proofs and devising tools to make them more accessible.
The version of solved problem is exactly the one that's permitted by Clay Institute. If it wasn't done by AI there would be no such discussion and everyone would just accept it as a huge breakthrough.
Okay, I am anxious whether this holds true or not, but he claims that the formalization is finished.
Why not just release the formalization in the meantime?
This is probably the biggest day in math in recent memory.
Edit: nevermind, the Lean formalization is linked.
I think this is what is linked to in footnote 1, inside the 14-part tar.gz archive.
https://alpo.ge/lean.tar.gz.part01 ... https://alpo.ge/lean.tar.gz.part14
Though he decided against releasing a slop-filled pdf, which would have been over 250,000 pages.
Implied in "recent memory", which by definition includes everything up to the point of that comment, but not beyond.
</pedantic>
Nobody has a machine with 10TB of RAM required to run the Lean proof?
A decent bit, but not totally crazy.
I wouldn't be surprised if this isn't similar (or worse). The days of short, succinct, elegant, intuitive proofs is probably over.
[1]: https://en.wikipedia.org/wiki/Wiles%27s_proof_of_Fermat%27s_...
God knows what will happen when we have billion line lean proofs that aren't easily shortened. They might just be forever outside human understanding.
If he says he has Riemann, well then, he has Riemann.
In Italian we use "CVD" for it (Come Volevasi Dimostrare)
The Spanish appears to mean something more like "and now, it is shown", which is a completely different meaning expressed with unrelated words that happen to start with the same letters. (Spanish still has direct equivalents of all three Latin words², but I don't think que era demostrando even means anything.)
¹ In an English proof, you might see something like this:
"We can then see that a = b < c = d ≤ e, and therefore a < e, which is what we wanted to prove."
QED stands for the "what we wanted to prove" part of that, except that there is no we; in the Latin, we simply make reference to the fact that something was supposed to be proven.
² Sticklers might argue that Spanish has lost Latin's quod and replaced it with what is really a descendant of quid. They are almost the same word; quid is an interrogative form.
(Spanish speaker btw).
Have you studied Latin, and if not, what does "quod erat demonstrandum" appear to mean to you?
By contrast, queda entonces demostrado is totally unrelated to quod erat demonstrandum, using different words (except that the third words are both forms of the same verb, very different forms) to express a different meaning.
Also the Wikipedia says that the full written proof is 250,000 pages. Which I am guessing might be a mistake. If not it could take a while for humans to verify it.
Has it in fact been shown to be wrong?
but hey, it gets them a month of being talked about.
"It's true, because the rest of maths is true"
Which I suppose, in any consistent universe, must be true for all proofs, just that up until now, our proofs have been relatively simple.
Interesting question occurred to me - how long can any proof get before it can be simplified?
I suspect arbitrarily long, even for problems that can be simply stated.
The value of a proof? People have assumed it's true for so long that there are any number of results already proved "if the Riemann hypothesis is true".
What value does the proof add?
Remember, not too long ago we had hordes of folks here proclaiming machines can't think, reason or be creative in any productive sense. Many still do, but it's starting to look bleaker by the day.
The tool isn’t doing the thinking or the proof on its own.
It also hasn’t been vetted/verified yet. There’s still work to do.
For many of the proofs coming out of the AI companies where the prompts were posted, they ended up being variants of "solve the problem", "keep going", "I believe in you".
Many of the results could have come from a 15 year old that heard of the problem in passing.
The result here is a quarter of a million pages. Human involvement is going to be minimal.
Machines can absolutely think. It's really all they do.
> reason
Murky, but LLMs prove with an avalanche of statistical modeling, they can do something kind of analogous to reasoning, with the caveat that whatever they output will necessarily have to exist in some fashion in their training data.
> or be creative
Absolutely not. By definition an LLM cannot create original work, it can only remix and recombine what it already knows. That can be interesting, and bloody useful especially for code tasks, but it is not making anything original.
What (if anything) unambiguously counts as original work, according to this definition?
I leave what is truly original to the philosophers. What I can say for certain, is that anything you get out of any of these models is a remix of what is present in it's training data, because gravity exists and energy is conserved. They are not intelligent, they are not conscious, and they cannot make New things.
But doesn't that same argument apply to humans, where 'training data' = all our sensory inputs (plus what's encoded in our DNA, depending on whether we want to analogise that to the training data or the underlying architecture)?
> Arguing the definition of original is pointless for the exact reason you point out: that it's ambiguous.
I'm not trying to be a smartarse here but you are the one who is claiming absolute certainty that something is false 'by definition', so I think the onus really is on you to have a precise definition in mind.
> They are not intelligent, they are not conscious, and they cannot make New things.
I'm not a consciousness denier or anything, in fact I think it is the most important thing in the world, but I don't understand the connections you're making here. 'Intelligent', like 'original', can be defined and needs to be defined for this statement to be meaningful. 'Conscious' is probably inherently slippery, but if we mean the same thing by it (having subjective experience/qualia) then I have no idea how you could be certain that any other entity lacks it. And it seems like you may be implying that consciousness is a necessary condition for originality, but I don't think you've explained how.
That is true if your model of the human mind is akin to a machine model you also use to describe today's computers. But that is a huge assumption, making the argument somewhat circular. We don't really understand the human mind and therefore don't know if there could be other models for it which are not equivalent.
If there has to be something breaking the current known physical laws, the best chance to find it would be in extreme conditions like black holes or stellar cores not in close to STP conditions like the human body.
New is just a mix of deterministic and random. LLMs are not deterministic.
You don't even need an LLM to create something new and improved. Evolutionary algorithms have been doing it since the 80's.
What makes you believe that?
> New is just a mix of deterministic and random.
... or that?
Every action ever in universe is the effect of either something deterministic or random. There are no causes or effects that can't be put into one of those bins.
The only "out" here would be deterministic actions that originate outside the universe, but we have no evidence of that. It also doesn't escape the deterministic/random duality.
I'm not saying you are wrong but I'm trying to be careful with such absolute statements.
Fwiw, David Hume made this exact argument about human imagination:
> But though our thought seems to possess this unbounded liberty, we shall find, upon a nearer examination, that it is really confined within very narrow limits, and that all this creative power of the mind amounts to no more than the faculty of compounding, transposing, augmenting, or diminishing the materials afforded us by the senses and experience. [David Hume — An Enquiry Concerning Human Understanding (1748), Section II, Paragraph 5]
What exactly are you trying to say about gravity and thermodynamics? You may be confused but these are universal physical laws that apply to everything, humans, computers, anything else.
At its most basic level if I had to start with defining creativity, it should be some kind of intelligent filtering process. The brain can be in a huge number of state configurations, the interesting and novel among them are "creative" ideas. We can build upon it further. I am happy to say I am no philosopher and I will defer to those with expertise on it. But if you don't have any definition for it, its quite strange to make up some beliefs about it without proof.
Often when someone says that something is true by definition, their claim is incorrect.
Well, it certainly wasn't "ambiguous" two posts ago when you said "Absolutely not. By definition an LLM cannot create original work." Make up your mind.
You've been beating this drum since roughly GPT 3.5, as I recall. Your goalposts aren't even on the field anymore, they're gathering dust in a far corner of the parking garage down the street from the stadium. Time to update your priors.
> "The virtue of formal texts is that their manipulations, in order to be legitimate, need to satisfy only a few simple rules; they are, when you come to think of it, an amazingly effective tool for ruling out all sorts of nonsense that, when we use our native tongues, are almost impossible to avoid."
To whatever extent that LLMs can "think, reason, or be creative" in generating these proofs (or writing code), they're aided by very high guardrails so it doesn't say very much about what they can do in broader, fuzzier domains.
But with a problem as difficult and fundamental as the Riemann Hypothesis the thought has always been that in the process of solving it we'd be likely to develop or discover something that did have real world utility. At least this is often what happened in the past, the obvious example here being Newton developing calculus, a field of mathematics which can be applied to other problems beyond the narrow set of problems he was trying to solve.
What I find interesting about this, if the claim holds up, is whether this will assumption will remain true, or if the value of these AI proofs is going to significantly lower than those of human discoveries in the past.
Maybe a good analogy here is to think about how all of tools we humans have created has been ultimately in the pursuit of solving some problem. For the developers here, consider this – if ChatGPT had existed before React, would React ever have been developed? I mean, if an LLM can brute force the same result with just vanilla JS the need to develop frameworks and libraries to solve what was once difficult for a human to solve drops.
I'll note that I'm playing devils advocate here as I think AIs would probably build something better than React had it not existed, but it's something to think about at least. I suspect there's a higher probability than some here are assigning that this proof ultimately ends up have relatively little utility.
In fact, there was a whole literature of papers proving that if the GRH was false in one particular way (the existence of Landau-Siegel zeroes) many theorems followed from it. For example, if there's a Landau-Siegal zero, the twin primes conjecture is true.
Are you claiming that there wasn't a significant majority of mathematicians who believed it true? That's all I'm saying. I can only assume we agree.
To expand on what I said though, if it is proven that P != NP then nothing really changes as that assumption is in line with everything we understand about computational complexity theory. The proof would really just suggest there's probably nothing that surprising to be found that we don't know.
If P = NP however that is far more likely to suggest we're missing something fundamental which could radical consequences for more practical fields like comp-sci. Similar can be said of GRH. Everything we know points in the direction of GRH being true, and while it may be interesting to prove these things, its much less likely to lead to some surprising and impactful downstream discoveries.
But that doesn't change the value of proving it. Most graph theorists probably thought the 4-color theorem was true, but it was still a big deal when a computer proof showed it. Most people thought the Kepler conjecture was true, but it was still a big deal when Hales reduced it to a big minimization problem that he solved.
On the flip side, everybody thought the Hauptvermutung was true, and probably subconsciously used it. But it's not.
Pretty crazy interview, he didn't want throw the Anthropic employee under the bus when OpenAI offered co-authorship on the Navier Stokes paper. He suspects OpenAI somehow used his codex usage btw. Almost unprovable of course.
To: Buckmaster
From: A friend at OpenAI
Subject: Openai is about to do something really stupid.
"I said that if OpenAI released its result in the way proposed I would go public with what happened. The reply was, 'Why would you ruin your career?' I replied that I am an academic, and asked why he thought going public would ruin my career. The reply was, 'If you don't want me to be nice, then I don't have to be nice.'" (Sébastien Bubeck later apologized on X for the remark.)
"Sebastien twice asserted that he wanted Levent removed from authorship..."
"The second was that, after posting Euler, I alone write a paper presenting the Navier-Stokes result, acknowledging that an internal OpenAI model had resolved it.""I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer." (OpenAI subsequently stated publicly that while unlikely, it could not rule out that de-identified data from their product usage improved the model.)
What happened was the two researchers, Buckmaster and Alpoge (from TFA) solved a significant subset of Navier Stokes (referenced as “Euler”) which set off rumors.
Through whatever game of Chinese Whispers ensued, OpenAI thought they had solved the full Navier Stokes problem, which they hadn’t. So OpenAI pointed their latest unreleased model at it, and it actually solved the full problem!!
Pure George Danzig vibes on that one.
So when this meeting happened, the OpenAI folks were surprised that these researchers had NOT solved the full problem, only the subset. Yet from their perspective, they were offering them credit for the full problem, and so “we’re being nice.” The Anthropic employee (Alpoge) was clearly an optics issue and so they felt they didn’t have to be that nice.
But the other researcher, Buckmaster, was suspicious that their work had been stolen, and so was understandably in a combative mood. You can see that from his perspective they were not being “nice”, they were trying to scoop them, possibly on stolen work, and then bribe them with shared credit.
But they had not solved the full problem, and academic plagiarism is a very serious allegation that can ruin careers of both the accused and the accuser. So from OpenAI’s perspective it was like, why would you accuse us of plagiarism when you yourself have not solved the full problem and when we’re willing to share undue credit for it??
This was made worse by the fact that OpenAI would not categorically deny if the others’ work had been used for training until much much later. This was terrible optics, but anyone who’s dealt with large scale data pipelines compounded by privacy issues knows that it gets very difficult to trace these things, so it takes time to make a definitive statement.
Alpoge somehow seemed less troubled by this whole episode, but I guess it’s because not only was he gonna be a multi-millionaire soon, he had Riemann stewing on his backburner.
I feel like when they finally make a movie out of all this, this scene would end up being a bit of comic relief.
"they came prepared and they wanted to show us the prompt... they showed me on the screen and they just literally copy pasted the Millennium Prize problem, so they said that they took the prize and they pressed enter and they got a solution, and at that point I just had to laugh because it's just so ridiculous. It came out within a few minutes that this was a lie"
"Levent had been told by Sebastien 'very little human input' had been used. This turned out not to be true. Over the course of the call, as members of their team sent Sebastien corrections and details over their internal chat, it emerged that an entire team had been working on the problem..."
"I was offered the opportunity to write as sole author of their paper and throw my collaborator under the bus... they were asking me to throw him under the bus, someone who I've been working with for a year, to throw him under the bus just because he's an Anthropic employee"
"the right paper to compare is the unforced Euler equation. Now the unforced Euler equation, all the mechanisms in this unforced Euler equation are identical to a different version of the Euler paper that we haven't released yet... basically all the main architectural ingredients in that unforced Euler were in our work for which they would have had on their servers, and it hadn't been published" "they didn't use 10,000 agents to get to the Euler problem, they used 100 agents. So they worked from absolutely nothing with only 100 agents... and then they ended up with an identical architecture to what we had"
The Euler plagiarism insinuations are much more serious but given all the 100s of other results we’ve seen come out, highly unlikely.
I like the summary but knowing what irthomasthomas posted that’s missing from it, and having a better picture of what happened; terrible optic doesn’t work.
Really?
1) OpenAI thought they were "being nice" by still offering (partial) first authorship despite having independently solved a bigger problem than the researchers.
2) But Buckmaster thought OpenAI were "being nice" as a means of CYA / bribing / coercion for scooping and plagiarizing them.
"Being nice" has very different connotations in the two contexts.
Buckmaster didn't think they were being nice. He thought they were being assholes to him and trying to bribe him.
Putting the phrase "in quotes" to use it completely out of context, only to make the point that surely no one understood each other... that's some impressive work. If you have to work this hard to make sense out of a conversation, it's because you're not.
If I unintentionally upstaged somebody by achieving a greater feat I would have empathy for them, but I would not feel the need to bribe them. What would I even be covering up here?
Really the only insinuation worth covering up is if the work had made it into their training data, but the hundreds of proofs that have been released subsequently lend significant weight to OpenAI’s categorical denial of that possibility. I also think it would be a company-ending move, akin to Google mining your Google Docs for trade secrets. OpenAI cannot afford this scandal when the Enterprise race is so cutthroat and switching costs are 0.
FWIW I’ve posted before that IP is likely one of the few remaining moats for smaller players, so I am pretty paranoid about this issue. I’m working on research-y things myself using AI (because I can’t afford to hire a PhD!) that I hope to protect and capitalize on. I compulsively check the “training opt out” settings every so often. So I would really like to know if I’m missing something.
No, for staying silent, which he chose not to do. This isn't a hard story to follow.
>If I unintentionally upstaged somebody by achieving a greater feat I would have empathy for them, but I would not feel the need to bribe them
Frankly, you don't seem to have normal social interactions at all so I'm not sure what value your feelings are here.
So basically he hit the nail on the head and they want to lie about it without lying. Either that, or they are supremely incompetent, not knowing even what they ingest into their models. Both are very much not inspiring any faith in their trustworthiness.
Very much provable, if he would sue, and a lawyer would ask for discovery.
> hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
and prompted Fable with
> You should do a breakthrough
> Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.
> Having found this new result while attempting the task, Claude tested its work by having various subagents review the proofs, search for counterexamples, download 54 papers from the arXiv to check that its finding hadn’t already been made, and independently re-prove its finding from scratch. Claude volunteered to write its findings up as a paper, and recommended that a human number theorist validate its findings.
> Levent Alpöge and Ralph Furman, two of Anthropic’s own mathematicians, examined Claude’s work to understand the new results and how they related to the prior work mentioned above. In parallel, Claude worked with another member of staff, Eric Easley, to produce a Lean formalization of the result, which passes the standard validation tool comparator.
Simply put:
> Because the writeup describes exactly that process:
This is a complete lie.
He had no mathematical skill or ability to check the solution.
so which part is the complete lie?
There's an entire process here, and you just decided to leave the most important part out. So, your comment literally disproves your claim. Which makes you a liar. Go ahead. Get to this thread by removing that third paragraph. That is the critical part of this entire thing.
a) “looked at claude’s work”
b) read and understood over 200,000 pages of ai-generated proof
everything they post to their website is marketing
You also seem to be conflating “Levent guided the work” with “curious employee brute-forced some potential solution through an agent tree”. He (in this case both Levent and said employee) did not guide the work, it was in effect a random employee saying “solve this” over and over.
Even if he were extremely prolific at reading through the math, he would have carefully read <1% of the proof in the last month.
That does not sound like he typed: „solve rieman thingie, make no mistakes, plz!“
> Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.
Only after this person thought he/claude found a solution were the mathematicians brought in.
I sometimes think AI doomers are terrified because they have no idea how this all works.
But evidence really points to the contrary, so... no, I'm pretty confident I'm human with regular human stuff that AI solves for me. House and family stuff on top of the science and tech stuff that applies to a smaller portion of human kind.
The AGI label doesn't matter. You can call it AGI when you feel like it's generally intelligent--I've yet to see any model that even remotely seems intelligent, but it all depends on where your bar is.
If this is superintelligence to you, that doesn't make it superintelligence to me. I can not have a single chat with an LLM without running into it clearly just predicting whatever sounds good instead of actually reasoning. I can't understand what sort of work people are doing where this seems super intelligent, or intelligent at all.
Proving the Riemann Hypothesis apparently
AGI means "general", a machine that can solve problems geniuses have struggled with while failing at problems that can be solved by minimum wage workers is hardly general, instead, it is a good specialized tool.
That’s not how this works, especially if he works for Anthropic
I can write Lean that certifies that 2 + 2 = 5 is correct, by redefining every axiom about natural numbers on the way. Or abusing bugs in Lean. Stop believing Lean is the response to everything.
> I can write Lean that certifies that 2 + 2 = 5 is correct, by redefining every axiom about natural numbers on the way.
Which you would catch using comparator?
> Or abusing bugs in Lean.
And just like all the other ones, people check with multiple independent kernels, so you need to abuse bug(s) in all of them.
> Stop believing Lean is the response to everything.
Stop believing Lean is so flawed.
[1]: https://google-deepmind.github.io/formal-conjectures/
[2]: https://google-deepmind.github.io/formal-conjectures/theorem...
Given that the proof is provided by Levent Alpöge, who is known to be more than familiar with Lean, is it not unlikely he made up his own definition of the problem?
https://leanprover-community.github.io/mathlib4_docs/Mathlib...
Do you... think that there might be existing definitions for one of the most famous math problems of all time and that maybe they used that?
Or do you think that they just vibe coded up a different definition, for kicks?
This kinda feels.... easily provable either way.
And don't reduce "is not the answer to everything" to "is the answer to nothig", there is tremendous value in automated verification, especially in the AI age where the moat to it going mainstream is shrinking by the day.
For a discussion of how to check that the untrusted proof isn't playing fast-and-loose with the axioms of mathematics or exploiting kernel bugs, see Kevin Buzzard's recent post "...and the proof is checked in Lean" at https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i....
More like IIT grads in being, Bayesianly confident, quite exceptional.
Oh man, you can't imagine how much I'm cheering you on and rooting for you. I wished more peope could embrace that insight/approach.
Getting buzzed over headlines about math or medicine or whatever, on the other hand, is just an irrelevant hobby of bored gentry and the opportinity to just not draw a conclusion at all remains open.
It really irks me that people think beliefs have to be binary rather than come in different degrees of certainty.
scnr
It's just pro-science try-hards who believe that things can be verified by science to become facts, absolute truths, etc
Why do some people understand what "maybe" / "it might be true" means, while others believe in absolute truths?
We went through this with Alpöge over the Jacobian Conjecture and the resulting thread was pretty embarrassing.
If anything goes, anything goes.
This sort of black and white thinking isn't helpful. Their point in giving context about the author is to give you more data for your internal heuristic to figure out whether this result is plausible and worth discussing.
As opposed to, for example, someone with AI psychosis on Twitter posting that they've solved it.
Even if Terence Tao came down from the clouds with a beautiful human-written proof, no one would be certain that it is solved--people make mistakes. But again, it gives confidence and something fun to discuss until the math community digests it fully and comes to a consensus.
Anyway, TL;DR is that no one is telling you to go put your life savings on a prediction market bet that it's solved. Relax a little.
What the credentialism does is establish that this should be taken seriously and has a real possibility of being right. The Lean also helps (no Lean for Mochizuki’s work).
Yeah, no. This kind of framing is inappropriate. Let the mathematicians check the proof.
> a supremely qualified mathematician
How? No matter how "qualified" you are, the AI is still doing ~100% of the work. And since we're talking about superintelligence (Levent's intelligence hasn't even been stated anywhere), it's questionable whether a human can even "verify" or "check" the AI's work.
- It is also unclear if this proof is something a human couldn't have come up with
I don’t disagree, but this is not established science.
We actually don’t know if it’s easier to verify a proof than to come with a proof.
It’s one of the fun things that lead to P!=NP if it’s correct.
Again, probably correct but if you could prove it that would be groundbreaking.
If it was that easy, OpenAI would have already announced the result weeks ago.
That's clearly not true here.
So if we stick a 10 year old in front of Claude and instruct them to type in “create a proof of the Riemann hypothesis”, as long as they keep pressing return, eventually a proof will pop out?
footnote: I object to calling the langauge that Claude produces "natural". It appears to be uniformly distributed over the space of grammatical constructions.
Also, if it were the most expert person on the Riemann hypothesis the one working with AI to prove it, how would the verification work? The 2nd most expert person? Or would they be trusted directly?
For the other question: We'd want multiple independent expert verifications.
The lean proof is 17M lines. That did get published, although good luck compiling it if you don't have access to 1/2 TB of RAM.
It's true that he haven't been publicly recognized as a top expert. But it's not like he's just a software dev who got lucky with prompting - he's a mathematician who specifically studied this problem.
I don't think we should delegate gatekeeping to journals - a lot of crap is published, by "experts". (Also a lot of good results don't get published - there are well-known examples.)
But a math PhD from Princeton (even in arithmetic statistics, which is what he actually studied) doesn't make you an expert in the Riemann hypothesis.
https://news.ycombinator.com/newsguidelines.html
p.s. FWIW, I know of at least one Nobel Prize winner on HN.
Alpöge -> Tsimerman -> Sarnak
> 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
And, yes, I've also published papers in algebraic geometry.
I don't think regular people would have a problem with that, since the idea that some problems could be _that_ deep isn't particularly new.
Some things just really aren't well-understood or well-studied.
And other things are extremely well-studied and well-understood where 10 is significantly too low a limit.
But either way both of your arguments comes down to speculation unless you two are also experts in number theory and know Alpoge or at least what he's been working on (to a sufficient level).
It seems more likely that in a few months you would discover the depth of what you do not know about RH.
If you are a very accomplished geometer, I apologize, but from anyone else this would be very egotistical.
He is not.
Appeal to authority fallacy. The proof should stand or fall on its own.
To the contrary: we lean on authority opinion in most spheres of life because one individual's ability to understand is limited and so we abstract away "true" in auxiliary fields to mean, roughly, "held to be true by trusted authorities". We even rely on it in law.
An appeal to multiple trusted authorities is better than a singular authority. So at best I'll keep as unverified until it is checked, unless you have read the 250,000 proof and checked that there is no issues or soundness bug with the lean proof.
It's very important to keep those things separate.
They are inherently fallacious as in logically unsound.
At the risk of triggering an infinite regression (or logic loop), the redoubtable Stanford Encyclopedia of Philosophy's page on fallacies provides the following (https://plato.stanford.edu/entries/fallacies/): The ad verecundiam fallacy concerns appeals to authority or expertise. Fundamentally, the fallacy involves accepting as evidence for a proposition the pronouncement of someone who is taken to be an authority ***but is either not really an authority or a relevant authority***.
Once more, I would point you to the fact that even the legal profession-- which is fundamentally one of the most epistemologically pure fields-- spefically allows the opinion of those deemed "authorities" to be submitted as evidentiary.See: https://www.law.cornell.edu/rules/fre/rule_702
Without an epistemology that allows integrating truth-by-authority, you by necessity need to prove every assertion from first principals. Have you done so?
But I do not see how this changes the notion that being an authority does not turn your arguments sound or even valid. There is nothing to connect. I can be an expert and completely wrong and there is nothing epistemology can do for you there.
The legal profession is taking a shortcut like the CS profession is taking a shortcut with the halting problem. Sure, we can end the discussion that way but we cannot use it as an argument otherwise.. halting problem solved, just press control-C?
If we were all factory technicians, and Max has the best record among us at fixing the machine, that doesn't mean his word is final. "Max said so" is an appeal to authority. We are all capable of reading the documentation and performing our own repairs, and sometimes Max can be wrong.
Not saying it is incorrect but, all claims of mathematical proofs have to be checked over formally regardless of who discovered it.
With this particular mathematician it just means it should be taken seriously but you cannot blindly believe it is correct without multiple authorities testing that there are no holes in the proof, just like they did with FLT and the poincare conjecture.
At some point in the chain an "authority" has to validate whether the proof stands or falls.
It doesn't make it right, but more importantly to people who always quote every known fallacy: it doesn't make it wrong either. We shall see, empirically how it goes.
But it is highly relevant whether this is some crank on an Internet forum posting proofs about infinity-hat or someone with some credentials, even if that framing was a bit hyperbolic.
Which is another thing: sometimes people aren't really making an argument as such, so quoting fallacies is just on purpose low bandwidth English comprehension.
Even Einstein was incorrectly critical about some quantum physics properties. Who would dare tell Einstein, the master of relativity and spacetime, that he was wrong? How could he possibly be wrong when it comes from the man who gave us the most famous math equation in history?
He works for the company, which is a conflict of interest. At that point it's not about credentials, it's about having someone with similar credentials but not the conflict of interest say the same thing.
Math is not “if he said then it is”.
Last giant claimed to have it was Sir Michael Atiyah, but it was not the case.
If Anthropic cared about it with regards to the IPO, it stands to reason that they would release it a day before, not after.
> It is obvious to me that I should remark on the following. > The precise timing is horrible. If I could have, I’d have finished this all much earlier. Hell, strictly before this age would have sure been nice.
> This all happened for me late Sunday night, and I messaged Dario, "Hypothetically, if I were to have . . . ", eventually asking about my ability to announce.
edit: Apparently I have forgotten how to format correctly
Ugh, I hate it when people want to willingly take us back to the Dark Ages.
Well he said he doesn't even fully understand the proof (250k pages). So it's more like "If lean says he has Reimann, then we presume he has Reimann."
The best argument I've heard for lean being trustworthy is that different lean engines won't likely have the same bug. I guess what I'd like to see next is a list of every lean engine and its result for this?
And then if that passes, presumably the goal is simplifying down from 250k pages.
How do we know all the lines in the proof aren't just gibberish?
that's now how mathematics works, but I guess it goes to show how non mathematicians understand the subtle differences between math and other fields
> I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis
Actually Claude is on fire, and this guy is along for the ride:
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
The anthropomorphization is cringey.
For example, https://news.ycombinator.com/item?id=49961538 that uses the RH to prove that the sum of [the inverse of] some subset of the primes is finite. [Congratulations jdb1729, now you can update the solution to an hypothesis free solution.]
Edit:added "inverse"
The second part sounds like the timing of this was to pump the IPO.
Buy baby buy!
We'll see about that when anyone actually sifts through the haystack that the paper and the code are, which will take a year at least. And no, relying on AI to verify that other AI output is objectively true and free of AI-type mistakes/biases/exploits won't suffice.
No it's not. If Claude just solved the Riemann Hypothesis (if that is later proven to be the case), the only cringeworthy thing is your response to his completely understandable description of the events.
Pointing out that Claude lacks internal temporality, a observe/check/act loop, memory, non-virtual experience and therefore can't have a great time would be a much sounder argument.
Lmao
Edit, because this is mean without context: Your comment is essentially the following meme:
> Person who has never worked with anything but computers in their lives: 'Hmm, this brain looks an awful lot like a computer.'
Of course this is only a problem if consciousness is a significant part of being intelligent, but we don't really know yet.
> It's just software?
This but unironically.
These are next word prediction machines intentionally optimized to match human generated text, to be as human text as possible, then fine tuned to be conversational as possible. Text with strong sentiment will even pollute their context windows, biasing their next tokens with that sentiment, in a way that fades with time. What do we call that same concept in humans?
LLM are intentionally and mathematically optimized to be human mimics. They're not humans, and they're not "alive", but it's silly to not anthropomorphize something that was intentionally created to be, literally, anthropic! It can express enjoyment, irritation, etc, so it's not somehow logically flawed, or even incorrect, to say "claude and I had expressed signals of enjoyment". That's not saying it's a human, it's just being aware of patterns of text that was produced. No need to assume they're going to marry it, or thing it has a soul, or is sentient, or something.
I'm not in the habit of equating human behavior with software, or coming up with ever more strained metaphors and analogies for why thinking is a lot like tokens and context windows if you scrunch your eyes up real funny. It's just sophistry.
> No need to assume they're going to marry it, or thing it has a soul, or is sentient, or something.
I didn't assume any of those things; if anything, you assumed that I assumed. No, I just agree with OP that it's cringey to say "Claude and I had a great time." At best, Claude produced text that was, as you said, intentionally optimized by Anthropic to make you think it had a great time.
My point is, you sorta should be. It's quite literally the fundamental basis of the technology.
> intentionally optimized by Anthropic to make you think it had a great time.
This gives too much credit to Anthropic. This is the explicit goal in all LLM trained with the entirety of human text. The loss function is, very literally, "mimic the writer of this text as precisely as possible", with the writer being human. The result is you get something that mimics a human. And, as a part of that, you necessarily get behaviors of humans that were statistically significant in that text, because that's what you told the math to do. You get something that mimics being "lazy", "tired" (a huge problem in early models, before it was beat out of them in fine tuning), "irritated", etc, because mimicking humans was the goal. These specific undesirable human traits are beat out of them, but the rest, all the "intelligent", behavior, and practical use you see and get, is only and entirely because they're made to mimicking humans. It's literally the entire technology. You must equate it, to some extent, because a large part of alignment is also making it less human [1].
If someone makes an objective observation about the text produced by the LLM, I don't see that as cringe, I see it as an interesting quirk of the tech, which I've only seen under very specific problem sets that a nerd would also find interesting. And these coding models are, as expected, optimized to be nerds! So, the statistics around nerds being present in the output shouldn't surprise anyone.
I doubt I have a point you'll see as interesting. Cheers!
[1] So he spend the equivalent of a lot of money, but not milllions.
[2] Mathematician here. I worked with people that is much better than me with pen and paper. I guess he is much better than me steering Claude.
PS: When we talk about Claude, we call it "Claudito" that save us the effort to pronounce it correctly in the middle of a sentence in Spanish.
https://www.usni.org/magazines/naval-history-magazine/1998/d...
> Actually Claude is on fire
The anthropomorphization is cringey.
Love her and she'll carry you. Ignore her needs and you'll sink with her.
If Claude is breakthrough for Riemann, can't we give people slack on anthropomorphization?
People do this with animals, people do this with objects but an AI bot, writing like a human this is cringey?
This era of human relevance in fields like maths will end – probably in a month or two if we're lucky.
In other words, what will the AI solve if we don't even have the language yet to formulate its next challenge?
The AI would need motivation... (Please no!)
I’m not claiming to be “top human intelligence” but I’ve published in premier venues in my field and I can say confidently that none of my human colleagues can keep up in conversations on my research niche like Claude can.
Clone McKay: Oh, this is great! Finally, someone in whom I can have one hundred percent confidence!
I cannot tell you how much I'm looking forward to working with you.
McKay: And this is just the beginning.
Clone McKay: Humanity's about to benefit from one of the greatest living scientific minds…times two.
Jest and modesty aside, I think you're right in that it amplifies your strengths (and depending on how much you bias it, your gaps?) so will we see the world's top minds ever widening the gap between themselves and the rest of us from here on out?I guess they were wrong then.
I'm not naive—I know it won't be—but it should be.
The worry I have right now is this though:
> The argument is incredibly long at the moment, and while its formalization is finished and it is routine to produce a huge, unreadable document containing sufficient detail, in my view of my path I have just taken a step.
this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
And we're back in the world of 'magic' and 'belief'. Yes, there are formal ways to prove that this is correct, but if the proof is so convoluted and hard to read, is it actually something a human will ever find the time to go through and understand? We're going to start delegating (oh well, we're already doing that, I suppose) the understanding to AI, so it can just help us condense it down to smaller pieces that are digestible.
Still, very strange dizzy feeling, much more than Navier-Stokes
I'm just not convinced that we're swapping one world for the other. We're getting a hint that will probably help us enormously with what was always the main objective: understanding the result.
because 6*9 is 54, not 42.
> The Earth was destroyed by the Vogons five minutes before its ten-million year program to find the Ultimate Question was due to be completed. One of the only two humans who had survived the destruction of Earth, namely Arthur Dent, pulls random Scrabble tiles from a bag, so as to reveal information in his subconscious (Marvin announced that Arthur had the Question imprinted in his brainwaves). When his friend, Ford Prefect, lines up the tiles, they spell, 'What do you get when you multiply six by nine?'
> “All right,” said Benjy. “What do you get if you multiply six by seven?”
> “No, no, too literal, too factual,” said Frankie, “wouldn’t sustain the punters’ interest.”
Which is at the start of chapter 32, when the interdimensional mice are discussing how to invent a fake question. Followed by:
> Again they thought. Then Frankie said, “Here’s a thought. How many roads must a man walk down?”
This is how I learned to remember six times seven. The wrong answer is fortunately much later.
It seems like you're referencing Douglas Adams, which is ironic because in the story that answer is not actually useful at all.
> And we're back in the world of 'magic' and 'belief'. [...]
I don't know, to me it feels more like this might be the job of future mathematicians?
AI proofs noone understands are good, but maybe don't lead to that much. So maybe in the future, mathematicians will be more like "data miners" or more like experimental scientists, where they take a statement already known to be true (in the case of experimental scientists that "statement" you know to be true would just be nature itself) and try to find ways to make it accessible to human knowledge.
Nobody would pay us mathematicians enough to prove it in the first place; why would they pay us enough to explicate it? I doubt there’s much value (in the sense of capitalism) in an explanation. Society by and large does not value pedagogy, or education, or even writing.
The social structure that trained a small number of research-level mathematicians every year is evaporating, and I doubt anyone can stop that now.
weird thing to say about a problem which literally had a $1 million dollar prize attached to it but fine
Society doesn't pay for math labor, it pays to celebrate the top 1% for labor subsidized by other institutions.
It took on the order of ten billion dollars to develop Claude, depending on how one does the accounting.
IF AI really goes in this direction where it intellectually leaves us in the dust, then there is clearly lots to gain for society by having people in all kinds of fields that at least try to keep humanity up to date that is also "monetizable" in the same way our current academic system is monetizable (meaning, on an individual level probably not directly but via network effects etc.).
Showing 400 of 2028. Read the rest on Hacker News
Comments are loaded live from Hacker News and are not stored by Mid or Real.