All 175,187 modules seem to check out with Lean 4.33.1, with no errors or `sorry`. There are tons of linter warnings - over 460k of them - but I think they might be caused by the way I invoked it. I haven't checked the actual axiom being proved, though - I'm not good at math :)
It really seems he was rushed to publish and did not bother cleaning the code up at all. I am compiling the code and reviewing it, and there are 204 dead modules that are not referenced anywhere. And a LOT of repetitive code.
I think this should not be 17 million lines, I bet this can easily be under 1 million lines of Lean with a little bit of cleanup.
I don't know if Anthropic IPO being yesterday, or American Mathematical Society Executive Committee and Board of Trustees meeting being today has anything to do with the announcement happening today, but it's certainly interesting timing.
Part of my job is to ensure that HPC programs (largely LLM libraries these days) work on their next-generation CPUs and with new BIOS revisions. We test them using a shared pool of machines with various configurations. GPU machines are much scarcer, and we have to share them with other teams, so it's hard to reserve a time slot. I usually reserve a CPU machine around the clock so I can work whenever I want (I'm more productive at midnight), and nobody seems to mind.
I've worked here for almost a year now, and I love my company ;)
Well I guess you did say ‘related to’ manufacturing which widens it a little. :P
Expensive, but you can just try things!
$62.79/hr
https://aws.amazon.com/ec2/pricing/on-demand/
I run many high core jobs for various reasons, I often reach for many smaller machines rather than one enormous node like this. But not familiar with how lean parallelizes.
It really is quite slow, I've been having my dot and its 6 sub astras churn away at a proof for Z is Diophantine in Q, even 1/10th as many modules is an overnight affair on my 5800x3d with 128gb of ram. The first version I started on was 326 or 318 variables, but it's allegedly also found some in the mid 100s.
I've been having chat make a neat graph of the spam, it's neat to watch grow (hit the replay button) https://quartic-park.sneakyness.chatgpt.site
There's no real definitive measure of progress, all of the foundational stuff is handled and now it's working on the tricky stuff that is assumed away in the original that made me decide to let them loose trying to formalize.
LLM-generated text below: ``` The build printed about 460,000 linter warnings, and none of them affect correctness. Almost all are code-quality issues such as redundant proof steps, style, and unused variables. About 33,000 are uses of deprecated Mathlib names, which only matter if we upgrade Mathlib.
Memory per process wasn't logged during the run. The container's lifetime memory peak is 441 GiB, but that includes file cache and is only an upper bound. A typical module uses about 0.5 GB of its own memory, and the heaviest use about 8 GB. With about 280 compiles running in parallel, the build used at most about 215 GB of the machine's 1.5 TB of memory, based on spot checks. ```
One real problem with depending on Lean is that if there are any real bugs, then sufficiently-large swarms of agents are guaranteed to eventually come up with proofs that depend on the bug.
That said at least one version is itself formally verified, at the machine code level if I understood correctly. It might still make mistakes but you're starting to get to a point that those mistakes become quite unlikely.
It's still a pretty tricky endeavor but we should know soonish, with some confidence if it is correct or not.
"No, it has been proven"
"The prover may have mistakes"
"No, the have been verified"
"The verifier may have issues"
"No..."
Ad infinitum
Though it is technically different. Mathematicians were happy to pick a set of axioms, they were less happy to learn any theory capable of stating facts about basic arithmetic would necessarily be unable to prove all things it could state.
I agree. So if the AIs knows some bugs, they will be using them to prove all kind of fake stuff, not only interesting stuff. https://xkcd.com/704/
( for i in {01..14}; do curl -fSL https://alpo.ge/lean.tar.gz.part$i ; done ) | tar xzvcurl -fSL https://alpo.ge/lean.tar.gz.part{01..14} | tar xzv
$ mv file.{foo,bak}edit: apparently he clarified later he meant he really didn't know of curl a b c d etc working.
If the modules are repetitive and also signify that no intelligence has taken place and no understanding has occurred (if there was intelligent understanding in place, the first thing an intelligent being with a theory of mind would do is strengthen their claim to a coherent thought by, well, making it more coherent) then we're looking at something that makes strong claims to what language is, NOT what intelligence is.
Size can be solved next, FLT went from 13m to 2m lines in a couple of months, the same will happen here for sure. The tough (or expensive part) imo is checking that the toplevel statement actually says GRH. That's the real review effort, not solving for size just yet.
I say that with the belief that this is likely correct. They should have solved the size how much would it take 3-4 days?
Who are running defense for here, I dont get the point of defending slop just because its something of value.
Code is meant for humans to read and make sense of, same with maths these concepts are for Humans even if machines do generate them. Otherwise machine should be talking to each other in some unknown super compressed language.
Would you really wait 4 days knowing that every big AI lab on the planet is racing for this result?
Worse is better wins again.
This is what they intended to do, and did for months.
Didnt OpenAI learn and started to cleanup the work a tad bit... ofc its not perfect but this thing is a joke.
I dont and wont understand the HN mindset of winning the race at any cost mindset.
This seems like a stock pump attempt that failed miserably.
I dont see the point we can all be civil and build better stuff now. Why try to one up others at the cost of throwing out slop.
Trust me if the value of your product can even be measured in days, then thats not a product worth investing in.
I personally think AI companies deserve to fail if they cant even understand that.
What it "works" here to me is simple that it type-checks. Another user (hafeoz) even compiled all 175k modules upthread, with no errors and no sorry. Of course, a kernel bug is possible, but there are independent reimplementations of the checker, so you'd need the proof to slip past every one of them. A wrongly stated theorem is the more realistic failure, I think someone brought this up as well, and that only needs a page or two of definitions checked by humans, not 15M lines.
On readability: nobody was ever going to read 15M lines, or tbh even 1M after cleanup. That's what the kernel is for. Cleanup makes it easier to learn and build on, nobody's saying that shouldn't be done. My argument is simply dead modules are ugly, not wrong. OP himself says he expects it to be massively simplified.
As for "why not spend 3-4 days on it first": it went out today, and people were independently checking it on their own hardware within hours. That's arguably the better order, iterative progress first on successful ideas, not perfection. That's kind of what the whole YC ethos (who's page we're on, is. lol). Besides, a human proof got rushed out within a day because of this one, so the incentive to publish first was clearly real. We're on the same page here, just the 3-4 days of simplification first isn't the right way to progress imo.
Presumably as LLMs work their way through the rapidly shrinking list of Millenium Prize problems, it will be the case that none of them have concise enough solutions that they would have ever been solved by a human unaided.
That said, you are probably right that the most precise proof is too long for a human to generate unassisted.
Little bit like solving a rubrics cube. There's an infinite number of ways to solve it and you can use an infinite number of steps to reach a solution. Pros tend to use fewer steps, amateurs tend to use more.
That's actually a great analogy for AI maths in general. There's one or more paths from the start state (unsolved cube) to the end state (solved cube). You have a set of rules that are verifiable. Deploy a swarm of agents to intelligently search this space, pruning branches that lead to nowhere. With enough compute, you will find a solution, but it might not be the most elegant solution.
But I do agree with you!
It does build on all preceding author's work though. I wonder what % of cumulative researcher work would Claude make up?
https://math216.wordpress.com/2026/11/20/hurried-announcemen...
I think this has been debated many times before (I also suppose that this paraphrases Terrence Tao's take on AI), the value is in elevating the discipline. Those problems were selected because they were understood to be so out of reach that new constructs and theories would have to be developed. Those new tools would benefit mathematics at large because they might in turn unlock new, harder, or more generic problems to be solved. The value lies in the potential to unlock further discoveries, by having an increasing number of humans understanding, sharing and building upon each other and into ever more distant fields and disciplines.
The elephant in the room is that so far nothing indicates that LLM-generated proofs contribute meaningfully to building new frameworks and furthering our understanding of mathematics, therefore benefitting it. What we have is an inscrutable series of steps to prove that a single statement is true, plus the knowledge that LLMs aimed at a narrow problem can brute-force it given enough resources. The proof is technically only useful as it serves to solidify corollaries derived from the GRH by earlier humans. Will LLMs come up with interesting/important/meaningful corollaries of their own? That's a creative endeavour, not something that boils down to an optimization/path-finding problem.
/Edit: I'm not even "against" ai value. For business proposes it's basically a match made in heaven... But their ingenuity and novelty is really overstated. They're really average. If you consider their output good, you're probably incompetent. If you consider it garbage,you're probably better then average. But for business, average is good enough hence me saying being not "against" ai value
LLMs are virtually useless at this. I routinely come up with much stronger names, and when I tell the LLM, they gush over it.
> Claude cutely called it the GRH... "book" — I looked at Theorem 1 and it began by noting that because the reals are totally ordered, if σ = 1 then σ > 1 or σ < 1
I respect that he decided to just post the lean proof and not bother others with 250k pages of badly written math that nobody would be expected to read anyway and would only serve as some kind of vanity thing. It is more respectful compared to some other approaches lately, imo.
Maybe the future of mathematics is gonna be AIs spitting out thousands lines of lean and human mathematicians trying to decipher them (with AI)
# parts 01-13 17.2M each
# part 14 5.0M
busybox time sh -c "$(seq -f 'tnftp -4 https://alpo.ge/lean.tar.gz.partg &' 1 14)"
# file count
busybox time sh -c 'cat lean.tar.gz.part*|tar tvzf - lean/src|wc -l'
175187
# line count
busybox time sh -c 'cat lean.tar.gz.part*|tar Oxzf - lean/src|grep -c "[^ ]"'
15531759
# extracted size
mkdir rh
busybox time sh -c 'cat lean.tar.gz.part*|tar xzf - -C rh'
du -hcs rh
1.3G(Accidentally stripped out "%02")
busybox time sh -c "$(seq -f 'tnftp -4 https://alpo.ge/lean.tar.gz.part%02g &' 1 14)"Few people would consider it a beautiful proof though.
The assumption that the universe is orderly, has beautiful laws, and doesn’t have potentially one-off events or rules is a religion.
I have had this thought experiment for decades. “Imagine that once every 1.2 million years a grand piano spawns approximately four miles from the city center of modern day Phoenix. That’s just the rule. That’s a law of the universe.” There would be no way to measure this because if it happens to happen during our time it would just be a weird event that nobody could explain and we would have to wait another 1.2m years to see it again. Ain’t nobody got time for that? So all of our explanations for this would be wrong and would never be able to explain this or predict the long term future of the universe because hypothetically our universe would be dominated by the mass of these grand pianos eventually. Maybe beings then could piece together what had happened then? But likely not. They’d be just as baffled as we are.
Parsimony makes sense to us, because we’re simple creatures that live in finite time with finite senses. There’s no reason that nature should favor simplicity over complexity other than our own biases.
No reason? Isn't the fruitfulness of this so far itself a very good reason? We've taken this approach and successfully explained and predicted so much of the universe already. How is that not a reason to keep going?
I’m just saying that the universe is under no constraints to be understandable to humans, and our human limits might impose limitations on what we can model.
Like, my dog is pretty clever, she's a good dog too! She can roll-over and shake, and after much work I can get her to "stay" for a limited period of time. She's a good dog.
She's not smart enough to do calculus I don't think, and even if maybe she is physically smart enough to do calculus there's no way I could communicate the concepts to her. So... what's like that for us humans? What's like that for our experience of the universe? Like, I can look at a spreadsheet, and do some math on various columns in a dataset. I can do the SVD here or there.
But... I cannot really intuitively understand something in... say, 30 dimensions. The best I can really do when I feel really clever and think really hard, is maybe 4 or 5 if I'm being generous as to what I consider a "dimension." If I do something like say, "ok, picture a cube, and then visualize the heat on the cube as a color scheme, there's a burner below the cube." That's "kind of" like visualizing 4-d, and maybe I can add some roughness to cube too, to add another dimension to the data.
But that's the best I can do and the temperature is not even a spatial dimension . Meanwhile, it's not uncommon for like 10 dimensions to be useful in a dataset for prediction of some value - maybe even more. So, what are we not smart enough to understand? What are we simply too dumb to truly get? What will we never be able to understand unaugmented? And maybe parsimony and the "unusual effectiveness of mathematics" or whatever is that? But maybe... math isn't unusually effective, maybe the universe is substantially more complex than we can even comprehend? And our simple first order approximations are useful, but miss massive amounts of the underlying geometry/structure/system that makes up reality?
I guess, that's kind of the Donald Hoffman take on it right? This is the interface, but our interface kind of sucks.
This gave rise to the term 'Forteana' for unexplained happenings, and The Fortean Times magazine. I think this stuff likely felt more plausible in the 1880’s before much understanding of electric fields making hair stand on end, drafts and flickering oil lamps being common, before atomic theory and the electromagnetic spectrum and radioactivity and thermodynamics when there were still a lot of big unexplained things around, before tape recorders, point-n-shoot cameras, then smartphones, were invented.
The ubiquity of smartphones on Earth selling at a billion a year and still no convincing recordings of Bigfoot, ghosts, UFOs, force fields, teleportation, is evidence indicating absence despite the saying “absence of evidence is not evidence of absence”. If you aren’t warm, that is some evidence that there’s no fire. If we had a lot of independent witnesses and videos of your piano appearing we would still lean towards it being a magic trick, hoax, CGI, because we have agreed that you can’t create energy and can’t teleport therefore it didn’t happen as it appears.
I haven't read Carl Sagan's Demon Haunted World but I think it's a counter position, look at all the nonsense humans have believed over history and still do believe today, and about the only way we can divine the real from the nonsense is repeatable tests - "science as a candle in the dark".
Also, claiming a teapot orbits does not put the burden of proof on the claimant, because, again the idea that the one making an extraordinary claim MUST offer extraordinary evidence is just that, what we think is a nice approach to take. Truth does not care who has the burden of proof.
That's why miracles are not to be brushed off; maybe you just never saw one. Maybe they only occur under some weird circumstance the vast majority of people do not experience. Not say there is material proof of them, but mostly the objection to them is based on politics and adherence to the religion of "we have the universe all figure out" We could be living in a simulation, and "miracles" are a bypass of the simulated reality.
But who knows? I'm surprised most days that there's something rather than nothing, that I exist at all (it's quite strange to be anything at all), and that math and computers work. I'm quite glad they do? But yeah, things seem pretty damn miraculous to me most days.
How do I know? Very simply. There's overwhelming evidence against the idea and no evidence for it.
The parsimonious answer feels good to us, and I think mostly it bears fruit - but there’s no guarantee that parsimony is the only way to model things accurately
Pretty much every major result in mathematics - assuming ZFC as the background theory - requires thousands and thousands of pages to formally prove from the axioms of set theory.
There are proofs that might not be capable of being written down in the material universe, even if we assume a plank volume per symbol. They would exist nevertheless.
https://www.sciencealert.com/physicists-quantified-the-amoun...
Fermat's Last Theorem was initially proven in Lean in September and that proof required 13M lines, it has since been reduced to 2M lines and it will likely be reduced even further. Anthropic claims they can reduce it down to less than half a million lines by doing similar techniques to what software developers do... find common and reusable snippets and factor them out instead of repeatedly proving similar sub-theorems over and over again. There are also opportunities to replace lengthy sequences of explicit reasoning with automated tactics, reuse existing mathematical libraries including the Lean standard library, and reformulate arguments at a higher level of abstraction.
I guess I'm supposed to either accept the galactic proof, which likely hasn't been fully read by a single human yet, as gospel and speak breathlessly about the singularity, or I should be insisting the proof will fall apart and make accusations about deliberate academic espionage.
Proof assistants existed for 50 years - they are not new. But it's REALLY hard to formalize something which is not just logic.
Even just logic is not trivial. Russel's Principia Mathematica needed 360 pages before it can prove 1+1=2.
He did not prove one apple and one apple make two apples, he proved that his complicated definitions of arithmetic using sets were not completely broken.
It'll just take some work to simplify this.
From Alpoge himself: "My intuition insists it will be massively simplified, likely immediately.”