upvote
Finally finished compiling on my PC (well, actually an HPC machine at my company, but it's the weekend, and I hope nobody is using it) - it took 2.4 hours on a 480-core Xeon and produced 89 GB of compilation output. I had to make some modifications because, for some reason, Lake bottlenecks and stalls verification. I worked around this by compiling the files directly instead of using `lake build`. I also had to increase `vm.max_map_count` for some later modules.

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 :)

reply
175,187 modules... a lot of them are repetitive.

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.

reply
It is probably not that compressible. Anthropic FLT proof also has a lot of redundancies but my projection is without true mathematical refactoring it will be hard to get to 50% reduction in line count. One problem is that Lean, or type theory really, is truly verbose. In these parallely developed proofs if two agents use slightly incompatible definitions there are a lot of translations back and forth. Also FLT is following routes well established by mathematicians and it still got to 13 million lines. So I am actually surprised that GRH is not much bigger, as it reportedly ranges over a wide swath of fields.
reply
Anthropic claims to have a 1/7 the size version somewhere. https://claude.dev/blog/safely-scaling-multi-agent-teams-wit...
reply
Yes it would be interesting to see that version. If compaction comes from proof architectural improvement (what I called "true mathematical refactoring"), it could improve readability. On the other hand if it is tactical, for example using tall rw stacks that are incomprehensible to humans, shortness doesn't necessarily imply better accessibility. I suppose it has to be a lot of the first kind to get that much compression.
reply
Even with all the resources he presumably have… interesting
reply
Hence why it seems rushed. Backed by Anthropic he certainly can clean it, model access or $$$ is probably not a limiting factor; only time.

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.

reply
i would imagine the IPO has everything to do with it
reply
Apparently a mostly human/human only unfinished proof on the hypothesis was rushed to publication upon hearing this news. It is around 1000 pages long. I don't know what the approximate ratio should be but fully formal propfs generally are much longer than natural language proofs. So probably this size is ij 5he ballpark of normal. If anyone sho writes lean proofs or works in formalized mathematics can comment that would be nice.
reply
I am very skeptical of that human-only proof.
reply
It more than 250 000 pages.
reply
I am a hobbyist home-server-builder and would love to know more about you and your company. I love hearing cases where you as an employee get direct access to the hardware. can you share any more details? I'm kind of spazzing out right now with how casually you said some wildly awesome stuff.
reply
I probably shouldn't say this, but I work at a company related to CPU manufacturing.

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 ;)

reply
reply
“related”
reply
deleted
reply
480 core xeon kinda hinted at that :D
reply
Thanks for sharing!
reply
Hmm, an unnamed CPU manufacturer using a ton of a specifically named processor. No idea where you could be working, I’m stumped. ;)

Well I guess you did say ‘related to’ manufacturing which widens it a little. :P

reply
That's a fun machine but there's similar (or even bigger) configs on AWS - https://aws.amazon.com/ec2/instance-types/u7i/

Expensive, but you can just try things!

reply
Roughly what would the above cost?
reply
u7i-6tb.112xlarge

$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.

reply
Jane str
reply
Any company which self hosts in a data center has employees accessing to hardware. Every company I've worked at was on-prem. Why is this treated as an outlier? Is everyone on AWS these days?
reply
No need to be snarky... having access to HPC clusters has little to do with "on-prem VS cloud" and is certainly an outlier in the prevalent SaaS world.
reply
Not snark, I was just surprised. No need to be cynical. If we are talking about HPC clusters specifically, then of course I agree.
reply
I have no idea what I'm talking about, but is it possible the lean program itself is wrong and in fact it proves something different?
reply
It's possible, but the probability of it being wrong is much lower than a peer reviewed human paper being incorrect.
reply
Most of the OpenAI proofs from the last big dump that have been looked at with some time don't actually match the lean claim now the media circus has moved on so I think the industry earned their lack of benefit of the doubt for a lot of people.
reply
Those doubts are misplaced then. The natural language proofs have problems, sure. But that has no bearing on the correctness of Lean certificates.
reply
Does Lean compilation usefully parallelize over that many cores? That’s a mini miracle by itself!
reply
Yes but only for independent parts of the process. Having all of these modules helps quite a bit with that.

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.

reply
Did you try #print axioms?
reply
Thanks for that. Could you have an AI summarize the 460k linter warnings and decide if any of them are significant? Also, how much RAM was in the machine, please.
reply
The machine itself has 1.5 TiB of memory; I did not check the free memory while it was running, sorry :(

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. ```

reply
Thank you!
reply
Whoa.

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.

reply
This comes up frequently in threads but it's actually a much less impactful issue than you might think. There are actually several entirely disjoint implementations of Lean, each with different bugs; it is incredibly unlikely an agent would be able to exercise the same bug in every single implementation.
reply
Does a malicious proof need to exercise the same bug in all the engines necessarily? Why couldn't it exercise a different bug in each engine?
reply
You trivially could through the principle of explosion.

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.

reply
"There might be some mistakes"

"No, it has been proven"

"The prover may have mistakes"

"No, the have been verified"

"The verifier may have issues"

"No..."

Ad infinitum

reply
But every round the minimum complexity of any potential exploit is dramatically raised, until the possibility of foul play can be reasonably ignored.
reply
Agrippa's trilemma never fails.
reply
Is that just an informal version of Gödel’s first incompleteness theorem?
reply
I'd argue Gödel's incompleteness theorem is more of a formale example of the same.

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.

reply
The successor function might have a bug.
reply
I’ve heard it argued that equality has a bug. In real life it’s a lossy comparision as almost nothing is truly equal (except maybe atomic and subatomic particles?)
reply
You could maybe address this through holding out the other implementations and using them only to verify?
reply
I doubt that any of them are closed-source, so the clankers probably already know all their bugs. Who says they all don't share a similar bug that is necessary for the proof to work? No one is going to review 1.3GB of anything by hand, even if they could understand it all in a single lifetime.
reply
Are there any that don’t rely so heavily on type theory?
reply
you don't want to review the proof, you want to review the verifier.
reply
> so the clankers probably already know all their bugs

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/

reply
Some of the reimplementations are actually verified, so guaranteed to be free from the sort of bugs found in the main kernel recently.
reply
All implementations built upon the same spec no?
reply
as per https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i... before ever thinking about kernel bugs, you should think of the very basic "is the lean code even stating the problem, not even to say correctly"
reply
Yeah maybe even more important point
reply
Apparently they use the Mathlib definition of Riemann, which has been reviewed by actual human mathematicians.
reply
And I understand that definition is 16 lines long.
reply
It would have to also be really sneaky about it. A proof that's trying to invoke some bug will look very different from a proof that's the real thing.
reply
I found a bug in Lean a couple of weeks ago with an LLM, in fact, but it was in its 'macro' system that expands code into statements, not the actual proof system. It wouldn't have proven anything false, but it could have produced statements that aren't what you intended.
reply
Yes, this is starting to look like the infinite monkey theorem.
reply
...if it's tasked to find proofs of False assuming that there are still bugs left.
reply
Shell tip: You do not have to download each part separately, then cat them and then untar it. This does it in a single streaming pass without the temporary files:

    ( for i in {01..14}; do curl -fSL https://alpo.ge/lean.tar.gz.part$i ; done ) | tar xzv
reply
shell tip: you do not need for loop.

curl -fSL https://alpo.ge/lean.tar.gz.part{01..14} | tar xzv

reply
I think in this format this leads to shell expansions. For 14 thats not an issue, but I try to use for loops as it mitigates the arg max limits better. I believe if you literally quote the string, curl can internally do the range expansions too if you used square brackets instead of braces and passed it literally.
reply
HN formatting has ruined your URL, try marking the block as code
reply
I didn’t know curl allows multiple URLs like that.
reply
That's the shell expansion, not curl. Works with everything.

  $ mv file.{foo,bak}
reply
Yeah I know. But shell just passes multiple URLs as arguments to curl. I didn’t know curl would accept that. In your example if there were more expansions it is up to mv to require the last one to be a directory. And many others just don’t accept multiple arguments.
reply
It does but you have to use '[1..14]'. {1..14} is a shell expansion.
reply
curl accepts multiple URL args
reply
It does. And {1..14} would be a shell expansion not a curl feature ie the shell blew it up to 14 different args before curl saw it.
reply
Your initial comment stated that to support multiple URLs it is required to use [] syntax. This is inaccurate, because curl supports multiple URL args hence shell expansion {} works just fine.
reply
Why are we nitpicking shell syntax in a page on a monumental result like this? He was thinking its a curl feature. I was just clarifying its a shell feature, I try not to depend on shell expansions due to the command buffer limits so something like a for loop or curls internal expansion is good to know of.

edit: apparently he clarified later he meant he really didn't know of curl a b c d etc working.

reply
Where do the downloaded parts live before they're piped into tar? Do they remain in memory?
reply
tar format is designed to be streamable. It writes the extracted files to disk as soon as they are received
reply
you probably want to be able to resume one part if it fails to download
reply
Trying to flex a shell tip on curling the Riemann hypothesis has got to be the funniest thing I’ve seen
reply
An update: the Lean community is now maintaining https://github.com/leanprover-community/riemann-hypothesis and is currently checking the proof against several different toolchain setups. There are instructions to get the usual compiled files.
reply
This seems to be norm in these top-level proofs. However, some comments here say that some of the modules are repetitive. In that case, can we not ask the AI to simplify the proof based on that to make it more tractable? Same as we often ask it to improve or refactor code that already works.
reply
They just want to be as fast as possible for their claim to fame I guess. OpenAIs Proof of the Existence of Yang-Mills Quantum theory has like 250000 declarations in it and 60% is unneccessary.
reply
It's possible you can't do that if this proof-producing activity has to do with the nature of language, and not the nature of intelligence.

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.

reply
I don't think anyone's questioning the shape right away, the first pass through most agent's is always known to be bloated. Another user (jycheng) literally found 204 dead or orphaned modules a little way down. Knowing it works is the real breakthrough, let's not lost sight of that.

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.

reply
I have never seen a more out of touch claim, at this size there is just as much chance there was a bug in lean, there's bug in the code which LLMs cant resolve maybe both, than assuming it 100% correct no questions asked.

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.

reply
Say you solved GRH.

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.

reply
https://math216.wordpress.com/2026/11/20/hurried-announcemen...

This is what they intended to do, and did for months.

reply
Yes my friend thats how things used to work, and should even now. Assuming all models can solve this stuff why cant we wait a few days to cleanup our work atleast and make more sense out of it.

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.

reply
To clarify, I didn't say it's 100% correct, no questions asked. I said the opposite in fact: the expensive part of reviewing this is checking that the top-level statement actually says GRH. Actually exactly in line with what you say, that to me is where the skepticism should go.

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.

reply
In a sense that is not surprising. If there were a short proof to a famous 150 year old maths problem it would likely have been found by now.

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.

reply
I don't think this means that more concise solutions aren't available. This is likely more just the easiest path for an LLM to arrive at a proof.

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.

reply
> Little bit like solving a rubrics cube.

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.

reply
'intelligently search' is doing a lot of heavy lifting here

But I do agree with you!

reply
That didn’t age well? Alessandro Maria Masullo.
reply
I think it's more that the proof requires several leaps in multiple fields/areas of mathematics, so I wouldn't be surprised if this was actually something like 100 separate proofs.
reply
In the same sense its that the descriptive machinery isn't fully developed. If we wrote all of our proofs in a Godelian representation of statements we would also come out with tomes per proof.
reply
I've had a similar thought. Looking back, will we think that Riemann Hypothesis was just an issue of scale?

It does build on all preceding author's work though. I wonder what % of cumulative researcher work would Claude make up?

reply
Here's an allegedly human proof in fewer than 1000 pages, the result of 12 years of work. Still a draft but they published it in its current state because Anthropic published theirs.

https://math216.wordpress.com/2026/11/20/hurried-announcemen...

reply
It is really good that those things happen right now IMHO, because we really have to ask the question of value in a different way. We have long been overstating the value of success. What is really the value of having proofed the GRH as we long had empirical support? (No expert here). Now is the 1.5M SLOC something of value? Could we derive a thousand times smaller human readable version much easier now? (Like having an oracle) . I hope this will lead to a time of fruitful philosophical and societal discussion, which might result in a better world (if we haven't destroyed the planet by then irreversibly)
reply
> We have long been overstating the value of success. What is really the value of having proofed the GRH as we long had empirical support? (No expert here).

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.

reply
Indeed, my comment does not appear to have aged very well
reply
More realistically, they'll turn out to be actually really bloated bug reports for the framework the proof was made in ಥ ‿ ಥ

/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

reply
Indeed, just ask LLMs to suggest products name or domains that sound good, are easy to pronounce, suggest some immediate meaning for a certain concept, are somewhat punchy, and have other qualities that make for a strong name.

LLMs are virtually useless at this. I routinely come up with much stronger names, and when I tell the LLM, they gush over it.

reply
And the human "book" is 250k pages (!)...
reply
From Alpöge's remark paper:

> 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)

reply
More concise, without bash, find, xargs

  # 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
reply
Correction:

(Accidentally stripped out "%02")

   busybox time sh -c "$(seq -f 'tnftp -4 https://alpo.ge/lean.tar.gz.part%02g &' 1 14)"
reply
Could a math proof really need 15M lines to write out? I don't buy it.
reply
Why not? Whenever you need to enumerate arbitrary many cases you can make it arbitrary long.

Few people would consider it a beautiful proof though.

reply
I like the idea that the universe isn’t necessarily beautiful and orderly, because my life has seldom been orderly. “Imagine a spherical cow” is fine? But there are no spherical cows.

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.

reply
> 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?

reply
Now did I say we shouldn’t 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.

reply
You called it a religion, and the main thrust of the response was that the fruitfulness thus far was a reason to not consider it a religion, but rather a rational course of action. It seems you got hung up on the phrase “keep going”.
reply
I don’t necessarily agree with you, bur I really enjoyed reading that.
reply
Well, I appreciate it. In your defense, I'm not sure that I entirely agree with this take either? But I think it's important to consider that we may not be clever enough to tackle all problems.

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.

reply
reply
Man, I wish my dog could take the derivative of a polynomial! All mine can do is count one treat or two treats!
reply
This is roughly the position of Charles Fort in the late 1800’s when he started collecting reports and compiled them into his "Book of the Damned", suggesting that there’s lots of your pianos happening - rains of frogs, teleportation, clairvoyance, etc - and scientists discount them as valid evidence because they would ruin nice clean simple theories. i.e. scientists pick and choose the evidence that lets them conclude a simple theory.

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".

reply
But then Occams's Razor is no longer a guiding principle, which is fine, because the Universe has no obligation to be simple.

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.

reply
Hey, I'm a fan of Occam, but.... I recognize that maybe, kind of sort of... what if we're at the limits of what can be described by Occam? Or similarly, what if like, we can make crappy models that are parsimonious and accurate models that are not. What if the universe is parsimonious when you don't look very close and extremely chaotic and weird and inconsistent when you do? That would surely track much of my lived experience?

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.

reply
There are many things I don't know, but this much I do know: There aren't any teapots orbiting the sun deep in interplanetary space.

How do I know? Very simply. There's overwhelming evidence against the idea and no evidence for it.

reply
You don’t know that. You believe it - and to be fair, I believe it too - but you don’t know 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

reply
This gets into what the word know even means. Nothing's guaranteed. I know it as much as one can know anything.
reply
It takes thousands of hours of study for a human to get to the point of understanding what the RH says and means. It isn’t hard to imagine that there are mathematical truths that require far more than this to understand them.

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.

reply
An infinite number of math proofs really need more than 15M to write out. Could that include some easy-to-state ones that we care a lot about? Yeah probably.
reply
What do you think the upper limit on the size of a math proof is?
reply
Not sure why you got downvoted but this is actually quite deep. I guess it depends on the mass of the universe. Since our universe is approximately a black hole, its mass determines its area and the number of bits it can store on the surface through the holographic principle. If the universe is not infinite there is a physical limit to math.
reply
A finite universe does not imply a finite limit to the size of proofs of true statements.

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.

reply
I guess the upper bound is the the amount of information in the universe. That seems to be about 10^80 bits.

https://www.sciencealert.com/physicists-quantified-the-amoun...

reply
tbh i was the student in undergrad maths classes using induction to prove trivialities bc nobody handed me a “book of axioms” and my intro to writing proofs was using Coq. let’s cut some slack to our autistic LLMs — I am sure in six months we will know how confident we are in this proof; no need to rush.
reply
Do you NEED 15M lines? No, it's almost certain this proof can and will be reduced, but the goal of an initial proof of a theorem like this isn't conciseness, it's to actually... prove the theorem.

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.

reply
And presumably you need hundreds of GBs of RAM to compile it. Would be nice if someone with a beefy machine could confirm?
reply
About 1 1/2 hour in, my estimate is it takes about one day on a single DGX spark. We will see.
reply
Proof verified. Less than a day on one spark.
reply
You could run it with limited RAM, just not many Lean processes in parallel. Just need some disk space for the compiled files.
reply
some randos on twitter I've seen are on it as we speak, but the real deal would be other mathematicians confirming the proof is written correctly.
reply
Yikes. I'm definitely not taking this one for granted until this has been reviewed thoroughly. That's a lot of room for shenanigans/bugs.
reply
Apparently healthy skepticism is a downvotable offense now. Not that I care about the karma. It's just interesting that we live in a time where taking anything other than the most extreme position is unacceptable to some people on here.

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.

reply
Given the level of verification the result has already attained the skepticism is verging on the pathological now.
reply
I wasn't aware that my opinion yesterday had to take into account events that hadn't happened yet, but noted.
reply
Why can't they ask AI to shorten it and make it beautiful? Is it not that intelligent?
reply
If it was intelligent it would want to shorten the proof and remove broken pieces, because that would strengthen the argument that the proof is good… assuming it had a theory of mind capable of understanding that another thing is evaluating its performance. It hasn't, isn't… something else is underway and we are observing the parameters of that.
reply
Or if it was really intelligent it could say that shortening it and simplifying it is a waste of time since it is obviously correct the way it is.
reply
Isn't there a way to reduce this? What's the reason for such a massive size? Are there "already proven" proof libraries in Lean that the LLM is not using like reinventing the libc from scratch?
reply
The only verified “already proven” library is mathlib, which gets you up to about half way through an undergraduate degree. It may be substantially more efficient for Anthropic to direct Claude at making a Lean library consisting of all of known mathematics (at least up to the end of the 20th century), and then use that as a base for their hypothesis. Instead of banging together Lean code at considerable expense until something magiced out, Claude would have been able to build on top of the works of the math community and produce a proof that was actually readable. Claude would also be able to reference known proofs while it reasons through the problem.
reply
why a lean proof is so big? it needs to be compiled and tested with a lot of cases or what may it be?
reply
Five years ago very few proofs (outside of logic) were formalized.

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.

reply
prove that (Russell's definition of 1 as a set) (Russell's definition of + in terms of set operations) (Russell's definition of 1 as a set) = (Russell's definition of 2 as a set), that is

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.

reply
For whatever reason, the LLMs currently generate very long proofs of results. For example, they often rephrase earlier arguments rather than citing them, or repeat their own arguments. (I've experienced these myself.)

It'll just take some work to simplify this.

From Alpoge himself: "My intuition insists it will be massively simplified, likely immediately.”

reply
This is true for code, even frontier models, they will often not use a library, or not modify the library but rewrite an almost identical function(s), etc. It's very interesting and needs constant vigilance
reply
Sounds like a massively parallel search of all possible paths since they have so much compute now
reply
judging by claudes review because it actually rebuilds and proves a bunch of stuff in mathlib that wasn't sufficiently formalized, but were required dependencies.
reply
I suspect it includes a lot of formalization of existing, peer-reviewed mathematics. Including these bloats the code but makes the verification more self-contained (less AI-written unverified inputs to lean).
reply
I'm absolutely positively sure that there are no sorrys and no theorems taken as axioms _anywhere_ among those Windows 11 source code-sized Lean proofs.
reply
I mean… that would be a trivial check with `ripgrep` et al. I’d be more worried about other sneakier tactics that LLM’s like to do as to move the goal posts (prove something trivial posited as grander, take arguments in a function head but not use them in the body etc)
reply
I also am absolutely positive of that, since Comparator checks for those.
reply
Yeah, this smells like BS. A "proof" that no one will understand within the next 1-2 years, and even then it's questionable that anyone could ever seriously hold 250k pages in their mind, and that no single mistake or oversight has slipped in. This is a haystack with an unknown amount of needles in it, and gaining certainty that that amount is zero, is a whole other task.
reply
It will be condensed soon. I guess he just wanted to put it out quickly
reply