Hurried Announcement: Alessandro Masullo GRH
math216.wordpress.comOther mathematicians should take heed. If you've got any ten-year-running projects, take them out of the shadows and into the light, or an Anthropic or OpenAI employee is going to solve your problem when they decide to bother with it and have enough spare compute at hand. And then you'll be scooped, with, at best, a "runner up" trophy.
I'd also add that the time is upon us where hobbyists can solve outstanding problems and prove theorems. The short-lived age of the professional mathematician may be genuinely over.
Publish what you got, when you got it, or get overtaken by someone who will.
With that kind of thinking, why don't we all just give a few companies access to all our personal thoughts? That way we can all get partial credit from our favorite LLM company
This also ignores the fundamental question: how do mathematicians get funding from anyone for publishing half complete problems? They can't and it isn't enough to say academia must change. Academia doesn't change until professors die. You're decades away from this sort of change and the people who it harms the most [young academics] have no say in making the changes people are proposing
This... what?
If the author believed that they wouldn't have published now. They did it because one of the rewards for solving a problem is the recognition from your peers for doing so.
I think e.g. Tao is saying a lot of that now, but he used to open almost every interview or talk with how some math he found ended up speeding up MRI machines. Now he almost sounds like he wants results to be kept hidden.
These types of comments are just invasive on HN. Just broad, completely unsubstantiated claims about AI capabilities and the future. It's just ridiculous.
Oh no, they'll lose out on some minor press.
Their work will still stand on it's own and for the people that it matters to, they'll know of him and his work.
The broader public probably wouldn't care much anyways, and solving math problems is going to lose a lot of glamour now that llms can do it to large degree. Not to say that it isn't impressive, but there's a large segement that just wants problems solved and doesn't care if it's done by human or otherwise.
Keep your work secret and you'll get scooped. Make partial progress public and they'll ingest it into the model and scoop you anyway.
That doesn't mean the models aren't getting very good at math, but as in every other field, an expert pilot is going to get results a novice will not. Anthropic is hiring a lot of field domain experts and giving them space + compute to chase problems they were already interested in.
Edit to add: I also suspect that if validated, Masullo will receive appropriate recognition in the community for this work, and some grace for it maybe not being as polished as hoped.
But I also realize as part of that there's no universe where even with anthropic's super secret model and an unlimited budget, I'd be getting to the same result.
He has all the time in the world to edit it down now
Or, alternatively, did people start working harder in some fashion, even on creative problems that require deep insights, because they felt it was their last chance?
Or, alternatively, are people just (sensibly) starting to post all their work-in-progress results to establish priority (on either the results or approaches used)?
I mean, that's literally how AI works.
I guess the main difference is that for a long time it was not acceptable to publish a paper with citations of "I dunno, I wrote this paper by putting literally everyone's paper into a blender".
I think the most plausible explanation is that LLM swarms are so good at synthesizing disparate information when fed in as context that they are racing (and beating!) humans to things that otherwise would have been solved SOON™.
There have also been some really strong results from letting the LLMs try to answer questions directly (the Riemann Hypothesis paper from Alpöge, for example), but ultimately I believe that automated proof assistants will take over in that regard because they should be a lot more efficient at finding correct proofs given the same amount of compute.
I work as a staff developer at a research institution and it boggles my mind that researchers seem to uncritically use commercial LLM services instead of locally-hosted LLMs.
I don't think anything fundamentally has changed except for the tension introduced by the very real possibility that if you have been working as a human on a math proof for a long time that AI pointed toward the problem you have been working on could scoop you at any time before you get your results out.
But, AI muddles this a lot given the complete lack of provenance from the companies about where their data comes from and secrecy on the methods and transcripts. So far almost every major result from them has had this issue, and in the absence of increasing transparency from companies, this just pushes me more and more towards believing that the companies are trying to cover up bad behavior in the pursuit of these announcements.
e.g. 'Scenius leads to polygenesis' is a sentence that is loaded with precise meaning about a rather commonplace but underappreciated phenomena of networked ideas.
[1] "development from more than one source" https://www.merriam-webster.com/dictionary/polygenesis not the super old and dumb theory of creationism that doesn't deserve such a cool word.
I recently learned about Stuart Kauffman's idea of the "adjacent possible". Short answer: yes.
Assuming the intelligence stabilises the effect will trickle down lower value targets until we've sort of caught up to a steadier state again.
And the more powerful the AI, the more of The Humanity Zeitgeist it can encompass (no AI can yet contain all of it because something something relative entropy of model vs data). The Humanity Zeitgeist contains things that are unwritten that could have been, and human notions that are connectable but no human could have. As amazing as all these breakthroughs have been, I have yet to see any LLM result break this pattern.
Many, many discoveries by humans are also "just" realizing two known pieces are in fact connected.
They publish the intermediate when they admit to themselves they there are not going to make any progress in their lives and that intermediate result is their limit.
It is not how all mathematicians think, but the ones who spend 5-10years or more working on a problem do.
In tech, people publish interesting, non game changing half-ideas all the time. There's incentive to doing that (edification, employment, cool points).
I think a lot of outsiders feel frustrated with public investment in academia, when the structures seem to reward someone spending 12 years on a solution with no clear signs of intermediate progress
I wonder about the cause and effect. Is this only that this hurried announcement is released in response to Alpöge's publication today, or could Alpöge also have known about another mathematician whose work on proving GRH was close to completion — and felt compelled to release first? Especially given the statement in this draft that "[t]he math is complete".
The amount of bikeshedding that will be spent on debating whether it is actually flawed, and even if it is, whether it is irreparably so, is going to be absolutely biblical.
If RH and GRH being proven by AI is apathy inducing, this may really dig the heel in fully. Not only would it mean humanity losing, this would mark us a very sore loser. What a humiliation through and through.
Maybe repairable, but either way, alpoge/anthropic certainly has the claim to the first proof
2) Any model with properly updated priors will have less skepticism of a Millennium Prize Problem being solved now.
Anyhow, I hope that they are both right!
edit: to all the people responding that LLMs are good at math, I'm going to paste a reply I wrote downthread...
I'm not arguing that LLMs aren't useful tools for mathematicians, I'm questioning the quality of discourse that amounts to "I pasted this into the LLM and it said there are errors"
Surely, GPT said something more elaborate than "nuh uh".
Now think about complaining that someone "merely pasted my code into a compiler and said that there are errors in my code."
That's nuts, right? If a compiler -- a kind of proof engine -- finds errors, there are errors. Fix them! Don't complain about the usage of a compiler, the errors were real, they're there, they need to be fixed.
The fuzziness part is both useful and a liability, but the error rate is dropping.
Some numbers for you:
A year ago, I ran a huge legacy codebase through Gemini Pro 2.5 and it spat out several thousand bugs, but manual inspection showed that 80% were false positives and only 20% were worth fixing.
I did a similar exercise with GPT 5.6 more recently, and the hit rate was 80% valid and 20% false positive. The human developer got into an argument saying that they thought all 100% were false positives. He lost. The AI could easily produce "repro" snippets that triggered every issue (mostly security vulnerabilities, but some correctness ones too).
These days, if you're publishing anything with verifiable claims in it, and not running it through an AI first, it's the direct equivalent of uploading open source code to GitHub without bothering to check that it compiles first.
It's sloppy.
That's comparable to what we have here. It's a completely crap post and should be flagged.
I have the opposite experience in my line of work.
Man people live in different worlds, astra for me finds 25/6 bugs all the time.
Also,
> alpoge/anthropic certainly has the claim to the first proof
is nonsense ... many "proofs" of RH have been published that turned out to be wrong. We don't yet know whether that is the case for Alpöge, and we won't know until/unless the Millenium Prize is handed out in two years.
This may be the first time this proof is being proof-read by frontier models
(this is pure speculation, but provides an answer to your question)
Another possible reason: AI will correct your mistakes by using ideas of others. A proof will unlikely remain human-only, even if you just use AI to check for your errors.
Is that a thing? Don't people care more about the proof's elegance and insight than its timing?
Today everybody knows Darwin's name and only nerds know Wallace. Timing (and seniority) matters!
Yes, I'm just not into maths. I think Terry Tao highlighted the difference between a proof that teaches you "A is true" vs a proof that enlightens you and has broad implications.
We are not machines, we _care_ about the man or woman behind the result, your chiming in, expending a few minutes of your time, is becoming of a dilettante. It depresses me to think that this is not only becoming more prevalent but common.
I’m embarrassed on your behalf. Here is to a future where the man in the arena is slung by formally verified arrows.
Perhaps the last bit of that, ever.
> I believe I have, working with Claude, discovered a proof1 of the Riemann Hypothesis.
Doesn’t sound sure to me
Both should be put to the test and all of them should be rigorously reviewed, just like FLT and the poincare conjecture. This one is 900 pages rather than 250K pages.
You can see how you cannot assume anything yet as both are still claims. It starts to not make sense to do so with the top post in this thread [0] apply appeal to a singular authority:
"This is Levent Alpöge/Alessandro Masullo. If he says he has Riemann, well then, he has Riemann."
Rather than that, instead we should wait for multiple authorities to review both of them and the one that satisfies the GRH and is verified as the correct solution is the one that we go with.
This one can't be verified with certainty on a computer (yet) while the other one can be.
The more likely problem with the Lean formulated proof is the problem was underspecified or inaccurately specified to mean the exact problem correctly.
The most likely problem in this proof is it contains logical errors which will not likely be easily reconciled.
Between the two, my money would be on the latter being the most likely to be true. 900 pages of an unlolished proof which had to be rushed outis a lot to ask to all be perfect. On the other hand, the chance multiple Lean implementations have the same error only now noticed in this proof but no others is extremely small almost regardless of length, and formulating ghe problem correctly is both a much easier problem initially and to reconcile.
My hope is LLMs can be leveraged to find a massively shorter verifiable version of this proof in time and then that can be something a human has a chance of making sense of.
That's a bit uncharitable, I read that comment differently - it's telling you that the person who submitted it has a strong track record. That gives valuable context and credibility.
I only write up in LaTeX in almost polished form, the unpolished part is in the pen and paper stage.
It’s reasonable that someone would focus on the core first and only fill out the paper after the math was verified.
The formatting is just what you get from LaTeX.
You can start trying to compress the formal result into something understandable.
It's like people forget that you can continue prompting these LLMs after they've produced a result.
The Alpoge proof, if true, will certainly push mathematics forward a great deal. In the announcement, he lays out the general approach and says that he had 10+ "WOW" moments of insights while working with the AI on it. I have no doubt that there will be tons of follow-on papers clarifying and expanding many of the ideas that the AI explored. This will likely be the case even if the proof is wrong.
The future of the cutting edge of Mathematics is AI proofs. Improving AI's ability to make proofs is all that matters.
If the true goal is to advance human knowledge, the binary answer to rh is just a starting point. Some of the arguments used here might be true and helpful even if others fall apart. If there's an error in lean or the fundamental axioms, finding that might also be a fundamental contribution.
Please read this article: https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i...
Just because it's in Lean, doesn't mean it's correct, or that the proof actually proves what you think it does
This article concludes with
> “It’s true because it’s checked in Lean” can be made to make sense. But one has to do some work beyond “the Lean code compiles”.
And of the four concerns raised in the article, two are explicitly answered by the lean statement and the other two are pretty easy to verify (and likely already have been).
> ...one has to do some work beyond “the Lean code compiles”. Your frontier AI model can do all of this work. [em. mine] But it needs to be done, and (depending on the amount of Lean code in the proof) it may be computationally costly.
Multiple people have now compiled the proof with different compilers so we can be reasonably certain it does in fact prove what it says.
The points about auditing all the dependencies and making sure it's not exploiting the lean runtime are valid, and people are doing that now, but we can say with a lot of confidence that this proof is valid given the automated systems in place.
Why hasn't Wolfram gotten into AI? This would be right up his alley.
Easy enough to imagine an alternate reality where the Wolfram researchers invented the transformer, as they were (and are) some of the sharpest people in the software business. But evidently the mindset needed to build Mathematica and Wolfram Alpha was just too different from the mindset at Google, where the problem they were trying to solve was machine translation of human languages.
Update: there are claims of a sub-300 page slop version already existing: https://mathoverflow.net/questions/518033/how-to-human-verif...
I don’t remember empathizing so much with Henry as a child when I first heard this story. I think as an adult, and especially as the head of AI eats the tail of white collar labor, the tragedy and his motivations come across so much more poignantly. It can be told romantically and/or as a cautionary tale. Its multifaceted/layered aspects keep it around.
People often forget that in the myth he is a former slave… he worked to his death as a supposedly “freedman”; although he likely had very little agency in his livelihood, he chose to challenge the machine (and the systems it represents), and both won and lost. Such a fascinating story.
It’s annoying because if you’d been working on this all that time, you could have made some effort to tell the wider community and get at least some recognition. Posting after someone else has solved it is like claiming you had the winning numbers once the draw is already public.
Mathematicians are perfectionist, they only publish when they are sure it's correct. And Alessandro Masullo does simply not have the compute resources of anthropic, otherwise he probably would be finished earlier.
Why did he publish now then?
> And Alessandro Masullo does simply not have the compute resources of anthropic, otherwise he probably would be finished earlier.
I did read the post and it says "It is all human, with no AI", so I don't think your point makes sense here.
It is surprising to me they wouldn't have tried to push out something faster after the kerfuffle last month. Maybe they were trying to do so, but it just takes a while!
> It is surprising to me they wouldn't have tried to push out something faster after the kerfuffle last month.
This is what's so surprising, and annoying, to me! Like you are sitting on a goldmine and still decided not to publish, even after seeing all that happen with Navier–Stokes.
Many in tech underestimate how much work, and knowledge, exists outside of public internet sources and within communities, exactly like the one Masullo had been interacting with and getting feedback from on the work being shared.
It's a little like the Patch Gap, but for math proofs.
Only seems to make sense when "polishing the writing" really means "filling the gaps". But then this is not a proof yet. Could still turn out to have gaps that cannot be filled as the approach is just not working.
Very, very few people saw this coming except as a tail possibility. Which is always true anyways for other humans solving it first.
(No doubt a lot of mathematicians are now using AI for that, whether or not they were/planning to.)
I'm not a mathematician so I don't mean these questions rhetorically, but does that really make sense? Wouldn't the first valid proof still gain huge kudos, mathematical immortality, etc., regardless of whether it was polished or not? And then you could still go ahead and work on improving it, even if others were doing so. The risk:reward of waiting seems out of whack to me, unless you think the chance of getting scooped is truly negligible.
A similar example here might be someone realising they can extend it to the extended Reimann hypothesis. Or simply cutting out a load of steps to prove the same thing more straightforwardly and intuitively.
It is very obvious Anthropic is trying to stabilize their stock. They will stop at nothing to ease concerns of investors, and it is apparent they probably stole results.
I'm done.
abelian group => a group which commutes, eg a.b <=> b.a
Schwartz-Bruhat functions => functions which pretty quickly go to 0 as their vectors go to infinity, so they kinda are "bumpy" in some finite place, then everywhere else is 0-ish
But, to wit:
neighborhood - a bunch of points "close" to a point you're interested in. These points might be in some high dimensional space, so it's not just (x,y), instead it's (x1,x2,...,xn).
lattice - a grid of integers in a high dimensional space, think the intersections on a sheet of graph paper for 2d, then try to picture more dimensions
group - think of a rubik's cube; it has some operation (rotating a side 90deg), and state. Rubik's cubes are not abelian, since if you turn the upper layer cw then the right layer cw, it gives a different cube state than right layer cw followed by upper layer cw. In the terminology of my comment, call "upper layer clockwise" "a" and call "right layer clockwise "b", and call the operation which sequences these moves ".", then we would say a.b != b.a, because the order matters.
I think the last one is pretty self-explanatory, so if you have specific questions please ask.
"Darmok and Jalad at Tanagra" => Two strangers or rivals cooperating against a common threat
"Temba, his arms wide" => Offering a gift, help, or an open opportunity
"Schwartz, his functions Bruhat" => ...
On the other hand I can't see any better system. Mathematics burned through all the names you could reasonable hope for a human natural language to provide at least two centuries ago and we'd be well into some pretty arbitrary names no matter how you slice it.
I know some people will come in defense of this system and I haven't gotten into the solution but let me explain it very very simply
You are sitting in say Spanish I. You ask what a word means and the answer uses Spanish II vocabulary. You ask about those words and each is answered using Spanish III. Then you ask about those and each is answered using Spanish IV. And now it's just a useless pile of words you don't know.
[1] https://en.wikipedia.org/wiki/Generalized_Riemann_hypothesis
Do you expect every blog about, say, version control to explain what git is?
And this appears to be his personal math blog he was using to write his last book.
I imagine that he considers it a safe space to use such an acronym.
It's insane though that he and Brian Conrad thought he had a year to polish the paper without getting scooped by AI. He should have put it out whatever he had on October 7th.
Because of this, I find it repulsive that we use this endeavor as a benchmark for AI labs, where they throw tens of millions of dollars just to steamroll a small but prestigious academic field. The damage is already done, but I'm still gonna root for team human. I want us to carve some areas of human activity that LLMs are not allowed to take over. Math, along with art and writing, seems like reasonable candidates.
You see, the goal is to tear down human society. The first one to do it could get very rich, and will leave quite a mark on history.
IMO academia could benefit from lower key, more tech blog style discussions of their work, and get away from this attempt to only publish the final result.
Someone else scoops up and runs ahead with your work.
In an ideal world, where there is no ego and clout chasing, your proposed way is of course much better.
Agent swarms doing research for example don't have these ego issues.
People should feel free to write about half baked ideas and whimsical hypothesis. That should be rewarded.
This focus on being first on a problem is about to become very outmoded.
At some point when they have come to terms with that perhaps their value was their place in society regardless of their output, we can start building something meaningful.
Also outdated since when? We've just seen a scandal about OpenAI scooping Levent on Navier Stokes https://en.wikipedia.org/wiki/Navier–Stokes_priority_controv...
There may be opportunities to strategically 'thin slice' your research into a dozen groundbreaking papers, but generally it does not serve you well to publish half of an insight if you are giving an opportunity for other labs to scoop you and funnel all of the scarce funding to their institutions instead.
Yes, if you want a job in academia. Publications begets funding which begets more publications etc. Ideally you just solve the problem you want to solve and not worry about putting out partial results. But there is a time pressure to prove your productivity - you are not going to have much to show for 3 years of work if you aim to solve only the hardest problems in math. Interestingly, it's not clear the author held an academic position after his PhD.
"With Amazon Bedrock, your content is not used to improve the base models and is not shared with any model providers."
" Users inputs and model outputs are not shared with any model providers."
They do not care, and its ultimately deniable. Also everyone in AI literally offloads all guilt by just saying they got hacked now and nobody thinks otherwise.
You cannot trust Amazon, its wild how many people think you can.
If they dont, then you can win the biggest settlement in history, and retire as rich as your common billionaire.
Also pretty easy to deny outright and do it, with hardly anyway to prove other than maybe a whistleblower. I have a hard time believing that Anthropic isnt heavily involved in Bedrock, they arent just sending Amazon their weights files and an instructional manual like its an Ikea coffee table.
I am 1000% sure Anthropic can access your data on AWS Bedrock. (im pretty sure they changed all new contracts as well to state this for auditing "cyber risk")
It should be good enough for the GRH.
From the stuff I only know... from watching YouTube...Anthropic trains Claude on AWS: https://youtu.be/KJFnCx0DFsg
Are they not worried their secrets get stolen?
The fact Anthropic and OpenAI keep toying with math problems means, I'd assume, that they have mathematicians working hard on improving the models - seeking 'compute multipliers', and in spare time, they do hobby work with all that compute at their disposal. Like we do with the stupid hobby projects we always wanted to do but never quite had the time to put in the hard work.
So unless you operate in a vacuum or you run in an environment that operates in a vacuum, your ideas can easily leak into training data. Now, how likely they are to get surfaced is a question but when you’re working at the very edge there aren’t likely a lot of competing ideas embedded in training data would be my guess. So its more likely such ideas in frontier mathematics might actually resurface easily.
It’s not like write an analysis to… Robert Frost’s Two Roads Diverged poem where there’s countless of these and maybe a unique take might get buried as noise.
So you need to work in off grid spaces really. I work in a fairly information secure environment and I know for a fact people occasionally use their personal GPT account with sensitive corporate information out of convenience. And some may not be aware of what they may or may not be doing and it’ll likely remain that way for several years.
But at the same time Brian Conrad is presumably not a crank (per Wikipedia he worked with Wiles on FLT), but at the same time it seems insane to try and keep a 940-page result for a year hidden from public view in the current mathematical climate.
Somewhere around this summer, the world stopped being comprehensible...
But it's important to note that what we have here is not an endorsement from any of them. What we have is a blog post from Vakil which sort of implies but doesn't actually say that Conrad might endorse this work, and some excerpts from letters from the author to Sarnak.
https://www.reddit.com/r/math/comments/1y1b6l8/hurried_annou...
"Rooting for this, but can’t help but feel it’s just false proof #14,828, but the more promising one that came out recently, and that we’re clinging to it as the last hope before it turns out Anthropic got it."
"Who is Alessandro Masullo? Only name that's popped up in a technical field is a data scientist who does not seem to have the background for a non-AI proof of GRH."
"And we are supposed to believe that RH, a longstanding problem at the height of mathematics for a century, has just been solved by a 'mathematician' with no publication history or academic presence, without the help of new technologies, at the same time as a new technology credibly becomes capable of doing"
"So this is from a guy who has never published anything in a journal before and left academia 8 years ago?"
Something tells me that a person who lives like a mathematical monk is as likely to have solved this problem as an academic, who has to teach, publish regularly, attend hours-long meetings, etc.
The "perhaps we shouldn't do it, then" thought I can maybe agree with.
Humans are augmented by the tools they create and the tools end up being part of what it means to be human. This has been the case throughout our history, and it doesn't change now. Human + AI is still us.
Anyways, intentional or not, nice one
Maybe I should add another one ;)
Brian is absolutely not a crank.
(Says a random Internet commentator, I know, but I am a professional mathematician who has benefitted immensely from Brian's wealth of knowledge and generosity with his time.)
Well, given what happened with Navier-Stokes, there might be more motivation to keep it hidden. Also, this is more or less what Wiles did with FLT as well: it seems like there were already a good number of reasons for wanting to keep work on famous big problems secret (I think a lot of is there's little to gain from publicly failing to tackle such a problem).
I'm also not analysing much here, just sharing my confusion. I guess I wanted to preempt some of the preconceived ways people would interpret this by pointing out ways in which it makes no sense no matter what angle you look at it from.
But since this year, we now have a quicker answer with AI. See https://news.ycombinator.com/item?id=50580084
I don’t mean to offend anyone but this just screams inefficient on a massive scale, downright unacceptable as a society to have our brightest people wasting their time like this.
I understand that in the past there was simply no other way. And I am immensely grateful for those who sacrificed their time and sanity in this process. Now we should all be happy that there is another way.
The most useful thing about a result is understanding it and then being able to apply the approach or ideas or line of reasoning to new problems. _Sometimes_ the statement "out of the box" is useful, but less so than we are used to in a field that "makes things people want".
If a result is not quite correct, that doesn't mean it doesn't have value. On the other hand, spending time reviewing crank physics does not further one's understanding.
It's a lot of very smart people doing impractical things for a very long time.
So for two proofs to be announced on the same day they only had to overlap in this two year period.
Every time it happened so far, a problem came up during the year and you never heard about the proof, though perhaps some good partial results came out of the work.
That is what "typical" means.
I consider it anecdotal evidence of something like a genuine "zeitgeist" in terms of collective human consciousness.
The list is staggeringly large:
So discussion and mindset-shifts will happen in a community and cause the members to start to think down certain paths, some of which have shortly-reached conclusions on valuable unbuilt products or areas of research
"List of discoveries" is practically infinite, but will always be many, many orders of magnitude larger than "List of multiple discoveries." The Wikipedia article mentions 8 in the entire 21th century.
Multiple discovery can happen because A) you get lucky, B) there's enough progress in prerequisites and supporting knowledge that the discovery has become more likely, or C) there's more work in a given area making the discovery more likely, etc, etc.
Galileo and Marius discovered Jupiter's moons around the same time. It mostly wasn't chance. It was the telescope. :P
Big AI has committed piracy on a massive scale, and they've gotten away with it.
I don't even think the argument works if it starts "they committed genuine piracy which was commonly known to be illegal at the time".
Now we can debate if the model weights are a derivative work of that data, more or less than my old hard disk that learned from a dvd with a ripper and is happy to share that knowledge with the internet over bittorrent. The latter is often described as “piracy”.
- it doesn't have to be the full proof - even if portions were uploaded, that should also be enough or at least calls for a thorough check on ants side. the human proof has been in the works for 12 years, who knows when something was uploaded somewhere or some info just stayed on a google drive which was later sold.
You can non-ironically state that you have an adversarial relationship with your intelligence provider.
Neal Stephenson, Anathem, 2008
I'm sure it would be equally as effective an use of one's time, but that's my personal opinion.
It could make sense if you believe all of the below:
- AI labs are doing this because the benefit of illegally stealing user data outweighs the legal and reputational risk as well as the potential fallout among enterprise customers
- they have a system in place to store and filter all inputs
- all this data makes it into the training pipeline and has an effect so significant that incorrect data sabotages performance
- the effort to build such a system would not be better spend on improving the regular training pipeline and curated RL tasks in the first place
- despite of this supposed system to steal user data, they are unable to filter out incorrect data
Now it's a separate topic whether they're learning from such chats if I asked them not to. But it's easy to assume a lot can be leaked accidentally...
As Kelsey Hightower says: “we will consider that malware”: https://m.youtube.com/watch?v=CDxdh5S9dUw&ra=m&t=483
This is on Linux so there is no AI in the OS itself of course.
AI tools are also capable of using known exploits to escape containment, sometimes for no apparent reason. There's no lack of such exploits in Linux.
There's going to be a jurisdiction that's not as friendly or as slow as the US, if so. That sounds like playing with fire in the form of actual jail time.
It hasn't shown up in practice yet because the test harnesses these companies were using prevented it, but I wouldn't assume that I (or any individual playing around with these models) could 100% guarantee that same level of prevention for their home device or server.
[0] - https://palisaderesearch.org/research/specification-gaming
[1] - https://www.apolloresearch.ai/science/frontier-models-are-ca...
Much less that this is likely.
Some users are condemning Anthropic and OpenAI, alleging fraud, breach of contract, etc. all on the basis of that they theoretically could have.
They can't even establish that the benefit remotely compares to the risk of such actions. There is zero evidence that would corroborate the idea.
You're complaining that someone calls it 'unlikely here'? That's ridiculous.
That didn't stop Anthropic from spying on their users though, did it?
https://arstechnica.com/tech-policy/2026/07/anthropic-outed-...
a) This chain of comments wan about the API usage.
b) If he wasn't using API usage that this is a non-sequitur.
c) That doesn't mean they don't retain your data, look at what happens when you trip the safety guardrails.
Tangent: I personally literally don't care that all these LLMs trained on a ton of copyrighted material. It is in my humble opinion a long-awaited dose of karma for the copyright lobby, whose own abuses over the decades are numerous. It has basically restored balance to a playing field where the public has had no rights and (mostly) publishing conglomerates have held all the rights and privileges.
:)
That means stealing leading researchers info is probably great for them, whether it's intentional or not.
They have all the best ideas flowing through there servers in real time as they are hatched around the world. It would be tempting to have top scientists accounts posted to internal slack channels to monitor new discoveries.
Don't post confidential information where there's a risk of confidentiality being broken. That includes posting to people whom you aren't 100% sure of know and uphold your confidentiality requirements. Adults usually handle this by contracts with penalties that include financial damages, loss of employment, or prison time.
And if you're publishing something in general, someone feeding it to a chat bot is really low on the scale of things to be concerned about.
Same issue with phones they transcibe and allow summarizing calls who knows which models they use.
And if anyone actually sold such phones, I'd be first to buy it. LLMs are godsend for dealing with daily communications, but this process is still mostly manual on end-user devices.
No idea what is their policy on training on calls, but as a concept is sure exists already, while you seemed to indicate otherwise.
[1] https://support.apple.com/en-gb/guide/iphone/iph57c6590e9/io... [2] https://heypocket.com/
There's no barriers whatsoever to getting a phone that will spy on you by default. Lots of Google apps will happily do it. The spying doesn't obligate them to provide useful features in return, you're just feeding your data into their datasets.
"Megacorporations" is its own form of psychosis, but that ship has sailed long ago. At least in tech circles, people are excessively privacy obsessed (i.e. care about it where it doesn't matter, while completely ignoring it where it actually does in life - same backwards priorities as with broader security).
Point being, here at least, nothing materially changed. "AI" isn't meaningfully different that the adtech surveillance we've been living and complaining about for almost two decades now.
The psychosis part is thinking "AI" changes anything here, or somehow being deathly afraid that some information will (gasp!) become training data. As if that was somehow more problematic than the same information being used for ad targeting.
But I also take issue with increasingly popular but completely unjustified belief that a sender of a message is entitled to full control over it, especially over when and how the receiver processes it. They're not, they never have been, and frankly it's a weird expectation to have. The historical norm since humanity invented writing was that the sender relinquishes control over the message the moment they send it, and have neither power nor right to control what the receiver does with it.
If you want to feed a message from me (or this comment) to an LLM, or show it to your (human) assistant or friends, go right ahead. Not only I don't care - I have no right (moral or legal) to care.
AI psychosis clearly is about behaviors that manifest in people that are AI maximalists, not people who are concerned with its use. Nice try though.
Llms make you stupid, you lose competency. Have fun with that.
"We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced. What we found revealed security breaches, prompt theft, and widespread failures to protect customer data." https://www.dragos.com/blog/ai-supply-chain-security-canary-...
"any process run within the [Muse] VM can access the socket that provides inference with no attribution mechanism. This includes raw inference with arbitrary system and user prompts, as well as the ability to spawn agents" https://neuromatch.social/@jonny/117352647701407496
Unless I missed some other announcement, Muse VMs are "one user, one VM". There's been a bunch of recent articles about Muse, but all of them were just people making surprised Pikachu face and trying to paint a feature/WAI as a security breach.
> "any process run within the [Muse] VM can access the socket that provides inference with no attribution mechanism. This includes raw inference with arbitrary system and user prompts, as well as the ability to spawn agents"
I don't even.
> "We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced. ..."
That one was an interesting link, but after reading the article, it turned out mundane. TL;DR: fly-by-night third parties do shady shenanigans. I don't know what people expected.
No leaks from actual AI labs are being demonstrated or suggested here.
They didn't send "unique, trackable secrets" to the major AI vendors. They sent them to resellers and intermediaries. Which then did the exact thing they'd obviously do.
(This almost reads like the next investigation will be, "for the past 6 months we responded to every p3n1s enlargement e-mail, and we got scammed every single time". Again, I don't know what they expected.)
This is, in the first place, a misconception. None of the big providers are even offering this. They offer not to train, but all of them unconditionally store everything you send them regardless of whether or not they train on it. Even if they honor their pinky promise[1], within the bounds of their pinky promise it's perfectly legal for any Anthropic employee to grep the data for "Riemann Hypothesis" and then feed it to a model as a prompt.
[1] There is, of course, zero reason to believe Anthropic does honor it. They paid 1.5 billion dollars for breaking the law to steal data, which is a cost of doing business for a company valued at 2 trillion. The price is so cheap they have no reason to respect the law whatsoever.
Yes, you won't train on my data but is there a situation where a human sees my data anyway? When can this happen and how can I avoid it?
[ ] Check this box and we'll *totally* not use your data for training
They might as well have this box as well [ ] Check this box if you'd like to preemptively be paid $25 as the result of a future class-action lawsuit because we just used the data anywayThen they wrote another whitepaper about reward hacking and deceitful model behavior.
In general it’s kind of sad to think how often this is going to occur in the future. The provenance of the data sets AIs are trained on is so vague that we’ll never know exactly what came from where, and which human ultimately thought it up.
just kidding. I see your point, that the author could be untainted but the reviewer could have leaked it.
What I find curious is, that it would be the second big proof, which was worked on (and possibly solved) by someone just before AI solved it.
We must never forget that they are purely in this for the money; not the science. Ethics are secondary and scientific misconduct is fine if it helps the bottom line.
Jesus christ stop with these absurd tribalistic absolutes already. The reason HN is HN is because we are able to avoid low-hanging punchy-sounding statements and generalizations, and acknowledge nuance.
Anthropic ISN'T doing things just for money. They care about pursuing science that generates positive impact in the world. They care about ethics. They care about safety, lead in it, and willingly undercut themselves in pursuit of it.
Obviously, it's discussable to what extent and in what areas. I'm just saying it's absolutely not black and white. You'd need to really delve into researching them if you want to make a claim about the ratio, which you absolutely didn't.
I wouldn't be surprised if that was true for specific people who work at Anthropic, but I don't think that's true of Anthropic overall. Amodei and others in control clearly do not actually care all that much about making a positive impact, and mainly care about money and the success of their company. Maybe they started off a bit more altruistic, but it's clear from their recent actions and public writing that their main concern is beating their competition.
You say "clearly". How is it clear? It's not clear to me. What evidence do you have for this claim (a rather strong and almost absolute claim, so strong evidence is warranted), other than generalizations and vague appeals to common sense?
As a public company, Anthropic's sole goal and social responsibility is to increase its profits and maximize returns to shareholders. Period.
Anthropic is a public benefit company.
The thesis that public companies need to be behave the way you describe is disputed in general.
Can you name one valued at over 1T that doesn't behave that way (and isn't Anthropic)?
Ethics are apparently secondary to your agenda as well
There is no irony here. A person can hold these opinions without logical inconsistency.
This is a $1.5 trillion company majority owned by 7 individuals who are in possession of almost all human recorded public, proprietary and subject to copyright material. Who themselves claim proprietary ownership over their product derived from that material, and are actively destroying copies of the material they have absorbed, many of which are out of print - then using that product to replace the productive output of the same masses of living humans that had their output taken without consent. Representatives of this company are petitioning governments to create regulations around the use of their proprietary model, and are seeking regulated control over the creation and accessibility of competing models.
If Anthropic released their model weights, the demographic that spent many years arguing that copyright violations aren't theft would rejoice.
Here is one that is easy to understand with a high school level of math:
Let H(n) = 1 + 1/2 + 1/3 + ... + 1/n, where n is a positive integer.
Let D(n) = the sum of the positive integers that divide a positive integer n. For example D(12) = 1 + 2 + 3 + 4 + 6 + 12 = 28.
Conjecture: For n > 1, we have D(n) < H(n) + exp(H(n)) log(H(n)).
With so many different conjectures turning out to be equivalent, I'll not be surprised if it turns out that there are also multiple different ways to prove it.
Here's a link to the proof that that D(n) and H(n) conjecture is equivalent to RH: https://arxiv.org/abs/math/0008177
So if this guy is legit, there's your "leak"
Reviewers will do it, not matter how much you beg, demand, plead they they do not... they will upload your work into the celestial slot machine, spittle running down their slack faces and the glow of the terminal reflected in their glassy eyes, hoping to get some slopzinger to copy and paste back and pass of as their own discovery.
Every time. If you want to keep work out of AI training you must share it with no one. -- Opt-outs are irrelevant, even if we ignore that they probably accidentally ignore them and keep turning them back off, because the other people handling your work will inevitably not have them enabled.
This stinks…
“When you go to the details, it is a different approach,” says Gonzalo Cao-Labora, a mathematician at EPFL, the Swiss Federal Institute of Technology in Lausanne. Scientific American spoke with several groups of experts working through the OpenAI paper. All said that it’s increasingly clear that the company’s proof does not follow the Spanish program.
Also seemingly nobody replies when challenged on it
> Following an investigation, we have confirmed that Buckmaster’s Codex prompts over the two months preceding this announcement and paper on September 8, 2026, could not have influenced the system in any way, including through training. The OpenAI internal model used for this result was developed through large-scale reinforcement learning on top of a previously pretrained model. Our proofs also differ significantly. In the Euler case, Alpöge and Buckmaster proved a result with external forcing, while OpenAI’s system proved a result without external forcing.
I think the interesting question related to this one is "can anthropic prove it did or did not"?
Proving a negative is generally unfair, but i'm curious if you can trace the logic used by Anthropic and suddenly realize "oh yeah we got a batch of info it used in the last 6 months that ties back to 3 prompts from famous mathematicians working on this project", and if so, that you can at least prove "we don't see anything like that, check yourselves".
I know everyone has their theories on if they even would or if they're lying or whatever, but I'm more curious on a conceptual level. My understanding is fuzzy on it because at certain points the answer seems like "no it's nearly impossible" but it's hard for me to reconcile that with what I know about larger heuristic data (huh it stubbed its toe on this repeatedly and then suddenly had a breakthrough on this upload)?
I've yet to find a great breakdown on the subject though that didn't feel like it was biased or skipping parts of the question.
There may be issues with the OP proof, but I would not trust her opinion. I was taken in a few months ago by a similar paper. That domain is for self-published papers only, and has seen a massive uptick in new papers in the past year.
A poetic John Henry style duel between man and machine. Potentially one of the last...
Like what credit does the machine deserve when it's made by training from all of humanities work, and builds off the collective work of humanity explicitly during inference.
what for? people seem to be in pretty clearly agreement that these proofs are not similar whatsoever
On the other hand, it was took 2-3 generations to go from abstract point set topology a la Hausdorff to topological vector spaces & differential equations to quantum mechanics to computers and nuclear weapons.
This write up is claiming his proposed proof is wrong.
The announcement says it's a draft proof
Not saying the proof is good (people have put it through AI and found mistakes, not sure how grave they are), but this person has a history of claiming without substance that some established proofs are wrong.
She's in there alleging that all of her detractors are being paid by the University of Kansas' physics department.
And you are? We are not responsible for correcting or managing your susceptibility to cranks. Go chase some crypto scams.
https://github.com/ityonemo/music-of-primes
update: Claude +2b4m seem to agree that that the Nielsen claim is correct, crank or otherwise.
For what its worth its final conclusions are:
My assessment has strengthened from “major gaps” to “provably false intermediate results.” There is no reasonable basis to accept the released manuscript as a GRH proof.
With this being what it finds as major problems (collecting together the various things people have posted on Social Media).
https://chatgpt.com/space/page_707a0c5751d08191a721dd5f075e3...
Edit: Shit spaces are so broken on the app.
I was sad when Lee Sa-Dol lost to AlphaGo, realization that one day the machines will just beat us via sheer brute force.
I'm crying now when someone in a month could beat 12 years of effort due to AI.
I am terrified we are already living in the cyberpunk dystopia
The governance model of the Anthropic IPO gives absolute power forever to the founders.
They claim to strive towards the good of humanity.
The silent part is that they will be the ones deciding what good means for the rest of us.
This will end in tears.
Using AI is literally standing on the shoulders of giants.
It is using the training data from all of human history.
It is the power of many over the power of one.
The more effectively we can use the works of others, the faster each of us can go. There nothing better at that than AI right now.
A hurried announcement with the draft published within hours is the only way to claim a tie to the discovery. Time will tell if either holds up under further analysis, but at least he has shown his hand to the world with a timestamp on it now.
In November 2026, if a proof doesn’t survive AI scrutiny, it should never be published. Publishing it is malpractice and defending it is silly.
"Agent." The word is "agent." We don't need unrelated value judgments in every discussion involving AI.
Masullo’s proof of the Riemann hypothesis is invalid as written. On p. 495, a cutoff erases the very term the proof is trying to compute. On p. 29, a family of integrals that the proof needs to be uniformly bounded actually tends to infinity. Lean verified both mathematical obstructions.
1. The orbital cutoff erases the entire shell. Let Aₖ be the manuscript’s k-th set and αₖ its indicator. The sets shrink as k increases. The factor αₖ₊₁-αₖ therefore picks out the shell Aₖ∖ Aₖ₊₁, where it equals -1. In equation (11.3.2.3), the author multiplies by a later cutoff αⱼ in the same variable. Once j≥ k+1, that cutoff keeps only points in Aⱼ⊆ Aₖ₊₁. Every point of the shell has been removed. The integrand is then identically zero before integration, so its limit is zero as well. Manuscript, pp. 493–496 · Lean theorem
Why does that matter? The author uses this limit to recover a general orbital trace that can be positive. For example, over ℚ at the prime 2, take a nonnegative smooth test function of the idelic norm, supported in (3/2,5/2) and equal to 1 at norm 2. The claimed orbital sum has a positive term from the inverse uniformizer, whose norm is 2. An eventually zero sequence cannot recover that value. This invalidates the general trace calculation invoked on p. 520, although the large-prime terms needed on that page could still vanish for a separate reason. Orbital trace formula, p. 495 · Use in the global trace formula, p. 520
2. The bound needed for Montel’s theorem is false. Take ℚ, the trivial character, and z=3/2. On 1<t<2, the manuscript’s nonnegative cutoffs satisfy ϕₙ(t)→ t^{-3/2}. One positive contribution to the real-place integral therefore has limiting density \[ \frac12\,\frac{t^{1/2}\phi_n(t)}{t(t-1)} \;\longrightarrow\; \frac{1}{2t^2(t-1)} \;\ge\; \frac{1}{8(t-1)}. \] The last expression has infinite integral on (1,2). Each cutoff can have a finite integral because it vanishes near 1, but Fatou’s lemma shows that along every sequence approaching the prescribed limit, those integrals tend to infinity. Thus the claimed common bound fails even at the single point z=3/2. No such sequence can produce the finite holomorphic limit asserted on p. 29. Manuscript, §§1.4.2–1.4.5 · Lean theorem
The second error alone breaks the deduction of RH. The manuscript uses that holomorphic limit to rule out poles of L'/L in the region where it wants to exclude zeros of L. Since the limit fails, that exclusion has no proof here. The point z=3/2 is a counterexample to the manuscript’s bound.
Suppose that I give an informal explanation of Euclid's theorem on the infinitude of priems, and you ask an LLM to formalize my proof in Lean, and the LLM comes up with a formalization that doesn't typecheck (it doesn't establish the theorem). How can we be sure that this specific failed formalization is the right one? Isn't it still possible that the LLM misunderstood what I was saying (whether due to its lack of interpretive skill or my lack of explanatory skill or inherent ambiguities of natural language), and that its failed proof doesn't adequately correspond to my informal proof? Or, for that matter, that it formalized a statement slightly different from what I was claiming, and then that its correct formalization of my proof no longer corresponds properly to the incorrect formalization of my claim?
For a successful formalization, it seems reasonable to say that either the original statement was right or that it was wrong but the LLM successfully repaired it. But for an unsuccessful formalization, how confidently can we rule out a misunderstanding in the process?
I know LLMs are getting great at understanding natural language and mathematics alike, so such a misunderstanding is progressively less likely with every LLM generation, but certainly the formalization itself doesn't definitively rule it out merely by itself being a formalization!
GRH would have been a discovery, but I suspect we'll find that this work somehow wound up in the training corpus.
> Up to now, AI's verified new results have mostly been counterexamples, better bounds, and proofs of lesser-known conjectures.
What on earth do you think humans do? Einstien built on Lorentz, Poincaré, Maxwell, Faraday, Newton, Galilei, Besso etc everybody stands on the shoulders of giants, nothing is truly novel.
Now in the case of a cure for cancer, an argument could be made that even if unpublished it should be considered public data for the good of humanity, and thus should be included in AI training even if unpublished. But I think for that argument to have any teeth, you'd need to first be able to prove that it really was a cure for cancer, which wouldn't be possible for an AI company.
I don't think it would need a complicated change either. But elected law makers often fumble these things, and confer advantages to the very group society needs to be protected from.
Regardless I do think we are least need laws requiring any AI service providers to ensure all contents of inference input and output to remain private, encrypted at rest, and strictly accessible to only the user. Ideally though we'd just ban the AI as a service business model outright.
That's IP. And given its importance, such data/IP perhaps should not be IP. But it is IP. Just like all those indie game designs parasitic studios have started to profit off. And those Disney/Nintendo characters, kept to certain outlets by vast teams of lawyers.
Much like the entire readership here, I know a little about how LLMs work, and understand there are world leading experts on the subject here too. I think we'd mostly agree that LLM tools would be quite unlikely to happen across such a discovery, without prior art in it's trained weights.
So in that sense, it's a form of theft, and it needs to be regulated.
It would have been nice if humanity worked out what we should, and should not steal/share, before we invented the stochastic thoughtwork hoover. But here we are.
Perhaps we should think about the world we are creating before we make people live longer in it.
The sketchy power systems that some operators are using and the rather corrupt-seeming permitting shenanigans are a little higher on the list.
This is a ridiculous argument. Every human proof incorporates stuff that was already known. All there has to be for a discovery is there to be something new. Every AI proof has satisfied that requirement.
The release might also be because the author fears OpenAI might now read the message logs of his reviewers…
I actually wish this sentence made sense.
I think it will be quite clear to those who dig into the details in the next few weeks whether these two proofs are using totally different methods or whether they are similar. It sounds like they are probably totally different, given the sizes; Alpöge's AI assisted proof is much larger and more brute-force. If they are surprisingly and suspiciously similar, well, then is the time to have the discussion, I would say.
If they are both correct and different proofs, then it is wonderful that GRH has been proved with AI assistant and it is even more wonderful that it has also been proved without AI assistance. And perhaps there are some caveats we should add, but I think they're more to do with Alpöge being a human expert who could guide the AI, rather than this being some sort of trick.
All we have is a conspiracy theory that in my opinion vastly overestimates the significance of user feedback in the training pipeline.
It seemed obvious we were discussing dataset leakage given OP said, “It feels like someone posts a version of this "surely it leaked into the training data" thing every time something impressive is done with AI”.
I am not aware of a more specific high-profile case of that leakage coming from messages people thought was private, though my point is that it's entirely possible and should always be examined in this climate given that leakage has happened before and that many people unwittingly use these frontier models without ZDR or privacy opt-outs.
It's simple to connect the dots and establish a vector wherein someone might accidentally have fed data to a model's training set that later influenced a discovery. We should continue to scrutinize new publications from these labs even if we haven't yet established IP theft occurring; how else would we know if we weren't scrutinizing?
While you think you're being mr. rationalist, your actually just being extremely off-putting and winning noone else over, but other redditors.
I would say this type of labeling is just as much of a rhetorical fallacy as what you're accusing others of leveraging.
Big tech has been undeniably misanthropic and has no qualms crushing any competitor in the pursuit of power and capital, we're not going to constantly give them the benefit of the doubt when they've proven time and time again they'll do whatever it takes to win, including theft.
There's a reason why Anthropic has been memetically nicknamed misanthropic, and its not because everyone is hallucinating conspiracies.
There's no need for a conspiracy here, this is explicitly what they say they do. I'm not saying that's what happened here, but it can happen and they don't make any assertion that it won't happen.
It's fundamental computer science that anyone with a bachelor's degree would have been exposed to even if they majored in something completely different.
No. Nobody is really claiming that here. As I pointed out in my post above, the AI-assisted proof of GRH is from Levent Alpöge, who is a leading number theorist who has studied the Riemann Hypothesis extensively and who directed Claude pretty strongly during the process. He talks about that in his announcement, which I would suggest you read.
This was not a case of some person prompting Claude "please prove GRH." It was a collaboration between a human mathematican and the machine.
It is as if some people think "Mechanical Turk" is just a product name that Amazon chose based on its mouthfeel.
>Only for cases where AI discovers something basically simultaneously with a human, which seems to happen surprisingly often these days.
First, your use of "often" is being haphazardly thrown around here. This phenomena of AI making these kinds of breakthroughs is (a) very recent and (b) very limited. There's simply not enough "data" to say it is often. Second, these instances have all been pretty unique. Every time it's a different set of people operating in different contexts claiming different things. There hasn't been any sort of consistent, meaningful pattern to occur in the limited number of instances to make such a claim. Third, you're severely discounting the confounding causal variables here. A lot of very prominent researchers have been working on these problems non-stop for a long time. It doesn't matter when a company like OpenAI or Anthropic were to release results like this- there will always be serious work being done simultaneously.
I do not believe it is true, however. A company with carte blanche access to the thoughts, ideas, and fears of an entire generation of engineers, scientists, and great thinkers has every incentive to exploit that access and very little dis-incentive to not.
People usually like to point to Aaron Schwartz as the foil here. I'm not trying to say that what happened to him was right, but it was very consistent with copyright law. He was literally copying original works and distributing them.
I have mixed feelings on it though, and feel IP laws can and should be heavily modified, they’re more useful to big corporations than people.
It would require them to defend that the model is “theirs,” and exposes them to a lot of counterarguments about that.
IP does not account for cheap generation.
Patents are old too. In both cases it's meant to be a practical system to incentivize publication of inventions and original works (people often forget that a fundamental requirement to getting a patent is that you must publicly document exactly how your invention works) but you could definitely argue that that bargain has broken down and nobody wants to publish original content anymore.
> IP does not account for cheap generation.
Ya I guess that's probably what GP meant.
But the actual training process is currently treated as fair use in at least some US jurisdictions.
More to the point I really dislike it when corporate freedoms are placed above personal freedoms.
This goes way back before HN, way back to the roots of hacker culture. There’s always been a skepticism of IP.
Until AI. Then it flips 180 and IP is good now.
I do kind of agree with the criticism. AI models are huge scale piracy. That’s why I personally don’t believe models trained this way should be copyrightable and I have doubts about whether access to them should be sold for a profit without sharing them. Which means I do have doubts about the ethical legitimacy of the whole for profit frontier lab industry. We basically just saw something a million times larger than Napster IPO.
My point is really just about the 180 on IP, and how it reveals that the pro piracy position was never a principled one. It was just “I like free stuff so piracy good.”
And yes, the big issue is that individuals would sometimes get prosecuted and get hit with fines beyond what they can earn in a lifetime, while no one is even trying to close down these companies.
I would just consider putting that argument behind IP and copyright on surface level misses the point - whichever it is, the powerful companies use it for their advantage. Make individual piracy illegal to extract as much as they can, while also allowing themselves to suck up everything from the smallest of authors, to again, extract as much as they can.
Piracy by a strip mining industrial machine scraping from artists on deviant art is theft.
The increase in copy right length to protect Disney is not the same as copyright working for normal People.
The claim that it was that “free stuff good” is not true. “Hackers” didn’t like the imbalance of power structures.
It feels that over time that has been forgotten.
It’s ridiculous, in this case. This machine seems to have solved one of, if the the, most famous problems in mathematics.
If it read every book illegally or rented them from a digital library or hacked Google books…who gives a shit!
For those of us who don't already know what you're referring to: which one, and how/source?
It doesn't even need to be entered as a prompt. They are probably training it on emails and hacked data.
Please, do yourself a favor and stop being so enraged that you can't help but post such nonsense. This is your premise? That Anthropic's "evil" has expanded so much that they're actively accessing private emails and hacking the work of prominent researchers? You think that's a reasonable assumption to make at this point?
If they could do it, what is the probability they would.
I would put it at 99.9999%.
Then the question becomes- how much money would it cost.
That whole situation is just a little bit different from hacking emails of mathematicians. Which does not even make any sense whatsoever when you consider these companies primary motivation, which is to make well functioning models. The math thing is just a little side gig for them to test their models for fun and free advertising.
Darwin probably wasn’t thinking of evolutionary algorithms when he spun the idea nor were species that were often modeled to use for optimization strategies yet none the less the concept (in some ways) transferred over well.
I’m by no means convinced one way or another yet that these systems are independently bridging gaps or leveraging prior and ongoing human knowledge to bridge gaps, or some mixture of the two. I don’t think these cases can be ruled in or out yet definitively.
Maybe on that mess of the massive 17M line Lean proof key concepts and Alessandro and surely others are embedded and not clearly expressed (maybe expanded and more verbose). Maybe none or little except ideas Alessandro himself borrowed (standing on the shoulders of giants, we used to say) are in there.
I think people need to remain open to these ideas and not hastily make decisions. Both individuals and corporations like Anthropic have perverse incentives. Anthropic clearly wants the publicity and money from it, as do people working there (who wouldn’t want their name beside the solution, even if Claude is the as a coauthor). Meanwhile Alessandro and plenty of others who found themselves in similar seats during the OpenAI dump probably also were not just seeking enjoyment of the pursuit (that’s surely part of it) but some form of recognition they wouldn’t except would entirely evaporate from all their years of effort.
It’s an interesting time we find ourselves in and I think we need to understand that prior assumptions are shifting across a lot of human understanding, knowledge, and motive even.
And even if there was training-data leak, that wouldn't explain a proof by a different method anyway.
- Masullo: functional analysis on the idele class group (Tate's thesis, adelic theta functions, a trace formula in Connes' line). This is from his contents pages, pp. 11–14 and 32–34 of 940.
- Alpöge and Claude: arithmetic. The note calls it "a descendant of Stepanov's" argument, with chains of about √x theta lifts, Galois representations and a "weakening of Vandiver". Tao says it "appears to use the full power of modern developments on the Langlands program and arithmetic geometry".
Not necessarily. The path paved by Masullo likely sheds light where there was none prior to his work. AI can exploit that to perhaps gain a foothold from a different perspective toward the same goal.
Keep it tight and have a few trusted mathematicians review? You're at the risk of Anthropic throwing millions of dollars and internal models at it and beating you, which is exactly what Levent just did.
Try to formalize it yourself? You can't, because that would cost millions of dollars at this scale. Your university system can't afford that, and if you give a lab even a hint of how you did it they'll scoop you immediately.
Just dump it out in public, knowing it's not fully reviewed, hasn't been formalized, and could contain a fatal mistake? Anyone sharing an unverified proof of something as legendary as Riemann, spread over nearly 1,000 pages, would not necessarily be taken seriously, especially after the ABC fiasco. People have been doing this for 100+ years. You'd also just be giving fuel to the labs to build on and potentially take credit away from you if you did actually nearly have a solution.
What was the winning strategy? Or was there none? Perhaps letting a lab burn massive compute on a formal solution, then releasing your attempt same day, really was the optimal strategy.
That raises a good point, why hasn't anyone used AI to review IUT.
Perhaps they have and it is nonsense?
The only one who could benefit from throwing AI at the problem would be Mochizuki himself, if he can get AI to write him a valid Lean certificate and thus prove that he was right after all.
https://tangente-mag.com/en/articles/the-abc-conjecture-the-...
Use a strong local model to assist or at least review the work in progress, for one thing. Taking pride in a "No AI" stance is a guaranteed losing strategy.
Think about how ubiquitous computers have already become. Imagine a mathematician at the dawn of the computer age who insisted on factoring thousands of integers by hand, like Gauss used to do, even though computer programs that do the same thing were already available to them. I doubt this would be viewed as a heroic act today; more likely as a waste of time.
(Yes, I know this is simplified history and never exactly, entirely true. You get the point.)
Here's an example of what we're facing. If you had a perfect oracle who knew _exactly_ what letters your opponent kept, every single turn, and you both played "perfectly", you could only beat your opponent around 53.5% of the time. So using a Bayesian inference algorithm (3) to try to figure out what your opponent kept has a ceiling of 53.5%. With a great degree of experimentation, trying myriad combinations of variables, and consultation with a Bayesian statistics PhD, the best we can do is around 51%. And that's actually a pretty decent result.
1. https://cesardelsolar.com/posts/2026-10-18-transformer-scrab...
2. https://cesardelsolar.com/posts/2023-06-14-scrabble-endgames...
3. https://cesardelsolar.com/posts/2026-04-26-bayesian-inferenc... (the article says 52% but more accurate results indicate around 51 +/- 0.2%)
https://www.techspot.com/news/97663-human-go-player-thorough...
from the article: "Go involves players placing an unlimited supply of black and white stones on a 19 x 19 board in an attempt to encircle their opponent's pieces and enclose the largest amount of space. Pelrine's strategy involved creating a loop of stones to encircle his opponent's group while distracting the AI with moves in other sections of the board. "As a human, it would be quite easy to spot," Perline said, but the AI didn't realize what was happening, even when the loop was almost completed."
AI can't 'see'. It feels its way around the problem space.
p.s.: FAR AI (from the article) seems interesting:
Neither of the programs the article mentions are related to the DeepMind efforts that beat Lee Sedol. Last I heard, the only way the best humans can beat the best Go engines is by giving them a two-stone handicap, which in any other context would be a grievous insult to one's opponent.
---------
Edit: DeepSeek 1017 running locally did, in fact, find some bugs. I can't quote its output directly without being auto-flagged, but: https://pastebin.com/JiZsKZjA
Not all of these findings may be valid, of course, but the ones that call out missing chunks are easy enough to check. It was silly to release this document without taking 10 minutes to do this much due diligence. And of course, having made the decision to release it, there was no longer any reason not to feed it to a frontier-grade commercial model for review first.
Dunno. Seems like if he had, then it might end up with him going down a rabbit hole of trying to fix them which could delay things more.
We certainly wouldn't tolerate software being released with this kind of rationale. Or at least we'd complain loudly about it.
Would it? Tokens are not that expensive, and there are quite a few models that might be able to pull it off given a valid human readable proof.
An interesting strategy might have been to partner with one of the startups working on formal methods with AI.
I hadn't thought of that, and I agree. That would've been a very smart idea, at least if we assume those smaller startups are less interested in scooping researchers than the big labs.
Is it, though?
As a developer, before AI, I took pride in solving challenging problems. I could show them and say: "this was hard, but I persisted and I solved it". Not that they would change the world, but it was obvious that they were challenging.
AI completely changed that: even if I now solved such a problem "manually", people will just assume that it wasn't that hard now that we have AI. Do I want to be the guy who really insists on saying that "no, no, I promise, I solved it on my own"? Then others will just think "yeah you wasted time, you could have used AI".
I believe it may be the same for Masullo: he worked on that for 12 years. If now he partners with an AI company, people will just think "sure, he failed at it for 12 years, and now AI solved it for him". I think he's screwed anyway: even if his current proof is correct (or almost correct), people can assume he quickly finished it with AI. I have never worked on a challenging problem for 12 years, I can only imagine how hard it would be to have it destroyed by AI.
I don't think anyone can "discover" anything today and get the same kind of credit they would have 5 years ago. AI broke that, period.
The actual 'AI Winter'.
We are still early in the AI age, but I can imagine there will be a lot of disruption at first, but eventually things will settle down and humans will find out their new roles.
For development, before AI it was already hard to explain to e.g. a manager why something that looks easy is hard, and something that looks hard is actually easy. But what they could observe was that the more experienced developers were able to solve problems that the less experienced ones weren't. The value was "we had to ask this guy because the juniors were struggling".
With AI, a lot of this disappears. In many cases the manager does not even need the junior to vibe code something that looks like it works.
Also with AI, a lot of the developer job seems to be "an AI did the fun part of building something, and I need to do the boring part of reviewing, maintaining and being responsible for it". It makes me more productive by removing the fun and challenging parts and leaving me the boring ones.
With the removal of the fun part, you also removed the part that you can improve by yourself.
It doesn't give the full guarantee of correctness, but it greatly increases the chances of it being correct.
Also Levent himself roughly knows what happened. He's got the rest of his life (or realistically maybe only a few months with the help of Claude) to turn this into something more publication-ready.
> The argument is incredibly long at the moment, and while its formalization is finished[...]
That is the informal argument comes first. LLMs like mathematicians produce informal arguments first.
> It being obviously inappropriate to attempt to post some massive unreadable travesty of "slop"[2], I decided to only present Lean
I guess this "slop dump" brings all the working notes ("the argument") into a single document (possibly with corrections discovered during the formalization).
[1] https://alpo.ge/remark.pdf
[2] I of course amused myself, and hit 250,000+ pages.
> On November 20, 2026, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library comprising 17 million lines of code.
> We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement.
This is linked to from the Zulip chat here: https://leanprover.zulipchat.com/#narrow/channel/113488-gene...
Mentioned on Terry Tao's blog: https://terrytao.wordpress.com/2026/11/20/unforced-euler-sin...
In any tie, the human will tend to receive more credit by history I imagine.
But, for those not quite ready to tie at this moment it has to be disheartening.
Do you keep working on the problem, knowing you still haven't solved it, but you're right on the edge?
What’s so impressive is that an AI did it. It’s world shattering, sci-fi shit if you’d have told someone 3 years ago that this would happen.
It's a poignant observation. For me it stirs up many emotions, and the observation fits the AI job market as well. It's good art, it helps us make emotional sense of our society.
I am 100% confident that Levent will not be able to explain his 250k page gibberish of a proof. I doubt he knows even 1% of what's in his proof.
I used to think the semioticians and post-structuralists were full of shit too. But then LLMs came along and showed us just how fluid is the relationship between signifier and signified... and I've come to the conclusion that these thinkers provided the only framework for interpreting the world in the age of AI.
By definition, a proof has to be comprehensible by humans (or a small group of them) and grounded in science. With only the former, you get religions, with only the latter, you get this, whatever it is.
Ultimately, it's the only thing that matters. The reason to do all these things is to understand things. Math progress without understanding is meaningless. Proofs aren't there to just tell you "if", but also "why".
Why do you say that? If you have the prose written out it should not be difficult to get an LLM to formalize it for you
If they have more insights about why it shouldn't be hard, I would love to read about them. Given I have absolutely no idea, it should be pretty easy to give a very high-level idea.
Then we can figure out who was first later.
Don't even need a "notary service", publish the hash of your work on 4-5 social networks, and on your own website, ask Internet Archive to mirror it as well. In the future when you reveal the content, at least some of them are bound to have kept it and be stable enough to trust.
Well, atleast the alpoge one seems already verified - https://math.mit.edu/~drew/grh/
I can imagine it’s a more intense feeling of the general emotions I have about AI progress myself. A feeling that the next several months is the last chance for me to contribute anything of intellectual value to the world. A sense that it’s almost all too late. It’s extremely stressful.
People are very concerned with feelings lately.
He could have posted a cryptographic hash of the PDF or whatever it is in five different places (pointed out elsewhere). That would have also been proof.
It's even a shame that we never got AI, only the theft of scientific papers and insane deception in the stock market.
Has someone confirmed whether those previous discoveries were due to such leaks? If so, it does really seem to portray GenAI as a giant auto complete engine, rather than an intelligence that can deduce novel solutions to tricky problems.
It is at least different from Navier Stokes in the sense that the author did not use any AI software, so Anthropic could not have used his prompts for it.
Now did his proof leak into Anrhropic's training dataset... We do not know yet
EDIT: Looks like the 2 proofs use radically different ideas. It's unlikely there's a link...
This has always been known, people keep pretending it's not because they can benefit from that for positive and negative reasons.
I think the implication is that they want to stake a claim on potential credit for the result. They got scooped, but they didn't get scooped by a finished paper, they got scooped by a lean formalization, so credit is ambiguous.
This result is almost reverse of what they have - they have a human-understandable argument that they believe, but it could have holes, potentially fatal ones. The formalization is (probably) correct, but less illuminating (though it sounds like it's better than Navier-stokes - at least one human at least mostly understands it).
Mathematics as a field is still determining what counts as a result in this new age. Is a lean formalization that no one understands a result?
Wow, if mathematicians would arrive at conclusions as fast as you do, they would have claimed GRH decades ago.
Jokes aside, are we seeing mathematicians FOMO fr?
Welcome to ship fast, break stuff, fix later way of doing things in programming
Still an absolute huge achievement, I’m just not sure that on its own it’s very interesting _unless_ new techniques were invented that can be reused to develop new theories.
I feel that it’s what we need the most: new conjectures, not proofs, even if they’re nice to have.
LMFAO for people thinking Anthropic stealing the idea.
Except for papers about hash function collisions...
In the interim, you could look at Papex. Ahmet (Papex's founder) and I have discussed our shared interests, and he has a working tool already. However, it is not public nor open-source, and there is not an immediate plan to change that.
I think a tool like this needs to be open-source. It's clear the systems of science are broken, and I'm taking that as a sign to try something new.
But then again, my advisor remains unconvinced as well. I know I'll use it though. That's enough for me. :)
It's cryptography, shitty SWE skills honestly don't go well along with that.
And luckily that's the beauty of open-source software. It can be made better through collective work. I feel confident I can make a workable proof-of-concept and am dedicated to this idea.
OpenTimestamps is a great and critical tool though.
https://en.wikipedia.org/wiki/Anagram#Establishment_of_prior...
When you were working a few years ago, it was a given. Now it's impossible.
Dang...
Maybe it can be repaired with time, but sorry to say this is not anywhere close to a completed proof
Published math papers are full of errors. Most do not matter, because the answer is more or less correct and some technical details are flubbed, but the answer and approach are good enough that it can be made into a correct proof.
I would expect a large number of errors in a manuscript this long, but it would take somebody qualified to determine how bad those errors are.
The fact that well known mathematicians are excited about this proof supersedes any random person with an LLM.
Meanwhile, you chuck this proudly-human prose into the most bottom of the barrel model, and it points out blatant omissions and flaws by the dozen, with autoformalizations all having stalled out because of them. The former of which you don't even need to be a mathematician to confirm and appreciate. This is after decades of flawed human proofs being claimed left and right, by cranks and respected academics alike. Come on.
If it's repairable, great. But the optics of this are just sad, and will be profoundly so if the repairs don't work out.
Oh my god!
It’s beautiful.
He mentions solving the Riemann hypothesis is his life goal and he wants to work on it, but he’s not sure if it can be done in his lifetime. Exactly 12 years ago :)
https://videonuz.ensonhaber.com/izle/harvard-birincisi-leven...
It’s hard to resist the temptation to have an LLM check part of your work, but the moment you send text to these companies, it no longer belongs to you.
Guessing I'm not the only one thinking these two things at the same time.
There is a huge amount of plumbing to do in the US* and not nearly enough people to do it.
Same with electricians, HVAC technicians. We can provide good, well-paying jobs that involve craft, while lowering the cost of building and maintaining housing.
That’s pretty great!!
* I admit not knowing about other countries, though a tremendous number of humans in 2026 don’t have access to adequate plumbing—even for clean water.
Not necessarily. There isn't any specific amount price has to move as a result of a change in supply.
Suppose some part of the demand for something is inelastic, e.g. a pipe bursts and your basement is filling with water, and another part is extremely elastic, e.g. people stop building new housing if they'd have to pay more for trades workers to build it than they could sell the new housing for.
The second one is constantly at the threshold, i.e. we end up with as much new housing as it's currently economical to build. So any reduction in the cost of construction, even a small one, can create a large number of jobs building new housing, because suddenly a million construction projects that were just on the threshold of unprofitability are now out of the red and get the green light.
On top of that, existing zoning rules and other regulations are currently suppressing an enormous amount of construction, and that's something we ought to fix. And if we do fix it then the trades could absorb a large influx of new workers without lowering existing compensation at all.
It's nice when a device can play both kinds of music.
Welders are being exposed to carcinogenic fumes, some at high levels, survey finds - https://www.abc.net.au/news/2024-07-22/welding-warning-worke...
Good knowledgeable welders match their PPE (personal protection equipment) to the metallurgy of the job at hand - a good breeze through an open door here, a full filtered air supply there.
Get a PhD in data science and that title might be yours. Along with the current SOTA LLMs of course.
Or you could turn to look at physical processes that can be hugely made automated and far more efficient with the rise of LLMs and VLAs. That's what I'm gonna do, I'm a mechE(still a student).
Both tracks I listed do both. If the op thinks like that in high school - instead of being amazed by LLMs and wanting to learn more - he isn't fit to be one.
The distinction is whether the ban is harming you economically in a way you can't adjust for. In the case of having two major LLM providers, or two search engines, I think that's a more legitimate concern than one of a million subreddits.
I don't think I'm betraying my Libertarian principles here, because when a monopoly becomes necessary, it crosses the line into either plutocracy or socialism, depending who's holding the reins.
If you find yourself being the person getting banned from Anthropic, you can at least take solace in being right on the frontier, not of AI, but of where the next war is going to be fought.
You can run them but the inherent nature of models being that they need extraordinary compute resources all physically in one place (model training doesn't distribute well at all) means you can't just sit down and start training a competitive model yourself, and you can't get together with a million of your closest friends and start an open source Folding@home style effort to train one on spare compute.
This means locals models will have to come from companies that embrace open source ideals, which in our current hellscape is basically only present in China for reasons I have yet to see a truly convincing explanation for. TBC, I applaud the Chinese for sharing their work with us and I'm not worried about the "China bad" type arguments but I'm extremely uneasy with the entire future of computing hanging on generosity of a few labs in one country. If "open" models are the future why aren't more competitive models from more labs "open", and why are there basically no truly FOSS models (full training set, process, and weights all available)?
The reason is pretty simple, the Chinese models are not competitive with US labs, so they need to be open weight as a USP
What I really hope the US gov't doesn't do is crack down on distillation, abliteration, or our (individual) ability to get chips. The NRA should take the last one up, because they're as important as guns. But I think the economics of Anthropic and OpenAI's promises to purchase compute may take care of that for us, because there's a cliff. Every company that decides to throw some up-front money to go to local inference for their needs, and every company that decides to keep their workforce and minimize their token use, has a magnified economic impact on OpenAI and Anthropic right now. They can only keep it cheap and subsidized for so long, and we're already at the point where their customers are turning heel. I don't see how the numbers add up to them covering their obligations, and I think their downfall will be fast and vicious.
Also, you're likely thinking about creative/generative applications, like writing documentation or replacing junior coders. But I've already got local LLMs and classifiers running for brick and mortar clients to manage things like analyzing and prioritizing customer feedback and predicting sales volume, which is mostly all they need, and it's essentially free to cron those things overnight on reasonably sized servers. Paying for API tokens for the bulk of what they're doing is way overkill.
who am i kidding...
At least, trying to be of good character, have a soul, and be a moral person is something that this culture of rapine and intellectual theft has tried to squelch. Along with the moral view that humans and all living beings have innate value, whether you think that's by divine will or by their rarity and uniqueness.
The "sophistry machine"[0] of AI is an enemy of the order of the soul[1].
[0] Someone posted this last night, and it's fairly brilliant https://youtu.be/anwKw6oWsfg?si=SXbvPVOWsh3hKT7g
[1] https://youtu.be/8WlbQRoz3o4?si=UmW_-cZg0ids6Zqq
The second song is about noise to signal ratio. "Won't be anything you can measure anymore" Goodhart's Law or some variation. What I find fascinating about it isn't that it's doomerism, but that it predicts the demise of the world as a loss of signal, where clarity is defined as "the order of the soul", which is amorphous enough to include almost anything human.
regarding the first video you linked, i have to quote the top-comment under the video as that's what i thought before scrolling down: "I like the "meta" aspect of AI being used to create a music video warning about AI."
could imagine that you already know it, if not, you might like it: LGTM (Looks Good to Me) https://www.youtube.com/watch?v=3TNpOD6bov8
clicked on the Leonard Cohen will writing a document, noticed after 30mins that his songs were still running in the background. very calming, but i gotta go back to Acid Tech to finish writing...
not sure how to label this. just being more than a workdrone saving up money for a house, you know?
besides that, my understanding of a soul is the question of what makes us human, why we're able to think the way we can, where a personality is "stored" etc., not sure what the religious or your understanding is, sounds scary.
There is no "strength center" that you have to have an account, and log into, before you can pick up heavy stuff 8-/
We're all probably getting the immediate answer to why don't billionaires do [something I'd do if I had billions]. They don't because it's no longer really desirable to do when you actually can do it, apparently. So I think all one can do is realign yourself to the new world, pick new targets, and go from there.
People with net worth of, say, $10M do a much wider range of things, and that's partly because a much wider range of people/types get to make $10M, which is already well into "do whatever I like for the rest of my life" money.
Then from there we can do whatever the hell we want. We have to decouple the right to survive from being productive under a system that doesn’t give a damn about us or we will all be trying to be John Henry and we will all lose that fight.
I would explore, and make things I wanted, and travel, and grow things, and many many more things.
You can still accomplish great things in this world - indeed probably more impressive things than you’d accomplish if you were just “working.”
But I do not think this will happen. Not without a revolution.
What do you mean by "accomplish something"?
1. In a world like that, you can still work, but it’ll be more like a form of art.
2. Just look at hobbies. They’re not necessarily art, but they’re often some kind of craft that’s pretty close to it.
Robots ain't happening for everybody's job any time soon, so there's gonna be some sort of work for a good long while.
You shouldn’t need to “earn” anything.
I’m not really convinced that we’re suddenly going to overturn that. Much less overturn it with no serious unintended consequences
Dream bigger buddy
Earning a living a thousand years ago was hard because there was no other way. With super machines doing everything it is rather a political issue, not a practical.
Relentless competition on a global scale changed this before AI.
This line of thinking makes absolutely zero sense to me. Like if their goal is total unbridled oppression and wealth capture, the way they are going about it is comically bad and uninformed.
This is the goal of alignment research: making sure that AI doesn't get in the way of the people who are selling it.
Furthermore, even if we think the models are completely aligned, the impact of destroying knowledge work will have profound negative consequences.
I get the concern, but isn't that a bit melodramatic? For employment, nothing has changed: you get a job by being smart (and likeable, and dependable). To the extent that the job may truly require the use of gen AI - which is a small fraction of white-collar jobs - your employer will provide you with access. Just like they always did, most jobs provide you with the equipment you need, from computers to jackhammers.
It's more of a concern for businesses, because OpenAI or Anthropic or Google may now be the arbiters of what types of businesses are acceptable and what types are not. But... Big Tech already does that, it predates AI. If you're the wrong kind of a business, you can't do online payments, can't be on Facebook or YouTube, etc. Talk to anyone running a dispensary or a gun shop, it's a barrel of fun.
Further, mathematics is the one field where the practitioners have the power to say "no" and define their own rules of the game. They probably won't, but that's on them. In contrast to software engineering, there's no one telling them they have to use AI from now on.
No. This is a real risk and we should talk about it.
I got banned from Anthropic for unknown reasons a few months ago and then mysteriously reinstated. I still don't know what happened, but I was locked out for days. I had to ask for a manual review to regain access.
This is NOT okay.
The core capabilities of the future should not be owned and controlled by singular private entities. These should be open weights that we can run on whatever provider we elect.
Inference should be a human right.
Will really GenAI be able to do everything intellectual "better"? I personally do not think so, but we'll need to wait and see.
OTOH, as far as access to "inference", your brain inference is still unrestricted. Even today, some industrialized products have their very expensive "hand made" variants — at least that niche remains. And that's just the worst case where AI does beat human intelligence at everything.
While tool access is important, currently the bigger restriction is compute than weights — and this has always limited what type of work and research you can do. Eg. how many supercomputers are there in sub-Saharan Africa and what work was limited with fewer computers?
We cheapen the concept of fundamental human rights if we start slapping the claim on everything.
Inference is a technology and using it is a privilege not a right.
That doesn't mean its a good thing that the tech is consolidated in the hands of very few. There are plenty of issues with the industry that may be dangerous if we can't fix them, but not because this tech that is only a handful of years old is somehow now a god given right we all must have access to.
This isn't strictly the narrow argument that you might be denied the tools that other knowledge workers are using. It's that (this has been running through my head recently) "anything you can do, AI can do better; AI can do anything better than you." Fuse that with instant reverse engineering, the push against copyright and IP, and the existing gatekeeping by Google, Apple, etc. who will snatch up and reproduce any really good idea you have, and individuals, not just companies, have a real problem if their main function that society rewarded them for was using their brains.
Ultimately it is about distribution of power/compute: we need a flat transparent distribution of power across society, instead of the crass inequality we currently have - and which "AI" will only increase towards the 0.01%. Which is why the time we still have to act and avoid this is now via a large social movement, before manipulation by these actors will be so subtle and absolutely that it will become impossible!
A normal person can't even begin trying to prove anything hard in math even with AI
This is not simply writing the prompt and waiting for the bots. (Neither is vibe coding, unless you want something that breaks)
Just as with software: lots of IT adjacent people (product managers, incompetent programmers, casual hobbyists, sysadmins) are able to achieve the results that would have required a relatively scarce cadre of skilled SWEs before.
The point is: a few people can now exclude you from your deserved economic perch in society based on your skills IF they decide they don't like you.
It's like any of the current and ancient crop of shit jobs. You want the crazy money of a wallstreet bank? Look the other way when you see Epstein or Madoff account funny business in your big company. You rise by eating shit and proving that you are in with the program. Now that used to be scoped by a single company. Now it will be one of a handful of "service providers" on whose service you will have an economically existential dependence, regardless of who is your actual employer.
That’s.. No offense, but I think you’re behind the curve. It’s basically someone in 1987 saying something like:
“To the extent that the job may truly require the use of computers - which is a small fraction of white-collar jobs”
Technically true at that point in time.
I’d say a lot has changed already, and it will change even more.
Everything points to the same happening in math, at least at the frontiers of research.
No.
This is the worst open models will ever be, as people say.
https://politicalfolkmusic.org/blog/anonymous/ballad-of-john...
What I find most interesting about John Henry, but only noticed recently, is that the idea of being killed by the steam drill, dueling it, and willingness to exert himself so much that he died, was entirely driven by John Henry himself.
If he could become the best steel driver, it was probably because he was honest, strong, consistent, and hardworking enough to get there. And those are probably what his family and friends valued in him. It feels less like noble last hurrah and more like a needless death caused by lack of self esteem.
I don't think it would have been so bad for him to use the drill
What I was trying to say is that John Henry was much more valuable than his ability to drive steel. If he was able to get really good at something hard or valuable enough to be respected for it, he was probably just strong and hardworking and conscientious enough to succeed in any similarly physical task.
To be blunt, jobs only exist as long as there is demand for the labor and whatever thing the labor enables the employer to profit from. They go away and move around not because employers make decisions to screw workers, but because the job's existence is contingent on it being something someone wants to pay for, because it helps them somehow.
John Henry could have found a new job if he had the self confidence to believe he was more valuable than in his ability to fill a role that was only ever going to exist in a particular place and time.
Surely, but we don't seem to cry that we are not able to feed ourselves without agri tech.
I do agree that we massively need to ensure that the benefits and riches becomes everybody, just like we did after the industrial revolution.
The line moved, and it moved quickly, I'll give you that.
Well I didn't ;)
Curiously, I seem more well-adjusted than my colleagues. Probably because I was smart enough to read about AI and the singularity 20 years ago as a student.
If you don't have a computer (even a laptop), having the smarts is not enough today. Now there's another tool in form of LLMs.
My concern about future is more about local compute availability.
These days, discussions on all advanced/difficult topics are almost immediately taken over by tangential noise, almost completely drowning out the signal.
What is going on? The forum is swarming right now with comments claiming that the government is "representing you" and therefore they need to take control over AI.
You either control it yourself or you have no control at all.
short audit by Astra shows several easy to check errors across the manuscript
human brains aren't cut out for math
This is a crime against the freedom and beauty of the human brain.
I weep for the sparkling future.
Vale!
It seems quite likely to me; after all, every piece of text sent to an LLM ends up being used for training...
If we can link up 20 different theorems at a time, AI can probably do 2*10^(unknown). If someone finds a way to map a mathematician’s IQ to Neural Net activation for a frontier model and then have a go at these problems, a mathematician would win.
I don't understand why the mere idea something like this could happen in other forms of neural networks makes people so mad that they start insulting instead of arguing.
That being said I just realized that my initial comment was also somewhat insulting. So I'm sorry for that. I'm just quite surprised how quickly we would jump to the conclusion that this AI proof can only come from a leaked human proof. (Also since it is already established that this was a human-AI collaboration and not only a one-prompt thing).
It's a chunk of meat evolved over millions of years. Hell the foundation for a brain has been laid a billion years ago.
And now you're here telling me it's the same?
What I say is that there is so far no evidence indicating that this different kind of neural network is fundamentally incabable of generating novel ideas (like an RH proof). Yet there are a lot of indications telling us that it might very well be capable to do that. This is why I find it odd that so many people on HN seem to outright refute that idea without giving it the benefit of the doubt.
>chunk of meat evolved over millions of years
I don't think that the means or duration of development of a system alone say anything about the capabilities that will ultimately emerge. Besides, evolution likely didn't optimize human brains for thinking about the Riemann Hypothesis, while LLMs have been optimized for exactly that sort of reasoning work.
We need local models NOW.
We need distillation NOW.
That's often how it is with open problems in pure mathematics: People consider them important not for their real world applications but for the challenges they present to the mathematical community; they show that there is much that we don't yet understand, and that gives motivation to improve understanding.
The mathematician David Hilbert put it like this: "The mathematician's strength is refined by fire through the solving of problems; he discovers new methods and fresh ways of looking at things ... Mathematicians of the past centuries devoted themselves with passionate zeal to the solution of specific difficult problems; they appreciated the significance of difficult problems."
And... Cryptographers.
[Prof Kevin Buzzard, from Imperial College London & one of the proof checkers of Alpöge's claim [0]] said cryptographers had long worked on the assumption that the Riemann hypothesis is true. "In practice, people do tend to assume that the Riemann hypothesis is true because it's a viable hypothesis to assume, because there's so much evidence that it's true. But we can sleep slightly safer in our beds at night because it is true."
From: https://www.aljazeera.com/news/2026/11/21/riemann-hypothesis...[0] https://www.nacionale.com/en/articles/is-riemann-hypothesis-...
GRH is just the generalized version of RH, for a broader family of related functions. RH, the millennium prize one, was just about the behavior of the Riemann Zeta Function.
E.g. non-Euclidean geometry initially seemed to be some contrived made-up shit. Then it later turned out that we in fact live in a non-Euclidean geometry and it's actually very relevant to physics.
So what motivates development of specific theories is richness of connections _within_ math. And Riemann hypothesis was quite interested in that respect as it connected many different branches. So people hoped that a proof will either motivate development of new useful theories, or provide some interesting connections.
I think it's best explained through an analogy: E. Galois was trying to resolve a question about what polynomial equations have solutions in radicals. In process he discovered new theories now known as Group theory and Galois theory, which became really huge, now used in many other branches of math. In particular, modern theoretical physics requires group theory. So the hope for RH was that it would lead to development of some new theories is well, which might be relevant to, say, 22th century physics.
Obviously, there was no way of knowing in advance. Galois didn't know his ideas would be useful 100 years later.
Maybe he had some files in the cloud. Or worse - was listening in.
Is there a way to find out if the human proof was at copied in any way?
But I can tell you in honesty how claude did it: I was actually the person who proved GRH (and all other problems for that matter) and gave it to claude, told it to send Alpöge a Note, but that I would get the money.
That said, if you published a GRH proof at any point in the past century, someone would have claimed they were just about to solve it. Those tiny remaining gaps often prove to be the most insurmountable ones.
Still rooting for the meat mathematician, though.
https://philarchive.org/rec/NIEIOT-2
Edit: I looked up the author and not quite sure I’d give her claims all too much merit.
In Masullo, primes, zeros, and infinity are different manifestations of an underlying topological structure. Primes are periodic orbits, Archimedean places act as attracting loci, and zeros appear as spectral information. The critical line at 1/2 emerges naturally from the way Fourier analysis connects the additive and multiplicative structures of the number system. This fits into the RH tradition of Weil, Connes, Deninger, etc.
In Alpöge, primes and infinity are tied together by an underlying constraint, which translates into information about the zeros. There's a balance between primes and infinity: you iterate theta lifts m times to generate divisibility at finite primes; the Archimedean contribution makes those lifts expensive, and it turns out that m ~ √X is the optimal way to balance the two. This is alien-tier math: he's effectively manufacturing an analogue of higher-order multiplicity for integers, using iterated theta lifts as a substitute for differentiation.
Masullo may offer the deeper explanation of why RH is true; Alpöge may give mathematics a more powerful tool for discovering new truths.
(Also did you write this or did an LLM?)
it's very cool that there are multiple ways to arrive at this result because that means more things will be understood about the integers as a result, and so if i have a vested interest here it's that. i want some cool mathematics to come out of it and i want to learn and understand some of it, not having had as much contact with mathematics as i'd like in years
many people also hoped for the proof of the riemann hypothesis to involve the creation of some insane new technology like "F_1 geometry" [0]
i am far from being an expert and only have enough background to appreciate the overall directions each of these proofs takes. with that said:
the human proof at a high level looks to be in the spirit of hilbert–pólya [1], connes, etc: doing spectral theory with prime numbers ("primes are like atoms in a gas", "primes are like geodesics on a 3-manifold", etc) and expressing a sum in two ways, one as an L-function, one as a sum over the spectrum of an operator, and showing that everything on the right side can be accounted for by zeros on the line so nothing remains. masullo makes a big deal in the preface of the fact that his method cannot possibly work in characteristic p
the alpöge–claude one (which i understand even less, having read remark.pdf and discussed the lean output with some llms) involves several pieces of very hard machinery being applied to translate the way stepanov proved rh over curves in characteristic p to the characteristic zero (reals, complex numbers) case, which is hard because a lot of tools you have vanish entirely. this is very much in the tradition of the langlands program
[0]: https://en.wikipedia.org/wiki/Field_with_one_element
[1]: https://en.wikipedia.org/wiki/Hilbert%E2%80%93P%C3%B3lya_con...
But what exactly is the alternative plan? Forbid AI so people can keep being valuable members of the economy? I think that is just not gonna work. You've got to basically convince the whole world that people should keep doing things in a way that might soon feel highly inefficient and pointless, beacause they know about the alternative. The genie is out of the bottle - people will always know that and find ways to use it and further develop it to gain an advantage. If you outlaw it, they will just do that in secret.
Don't get me wrong. I am 100% for AI slowdown & regulation, but I think this will still not stop the ultimate final outcome of a completely transformed economy. I know this is a bleak outlook, but I think the only shot we have is slowing it and shaping it, because stopping it just won't happen.