| ▲ | djoldman 5 hours ago |
| These proofs really are quite huge. The proof is 1.3GB, ~15M lines of code: mkdir -p riemann_proof && cd riemann_proof
# fetch parts 01–14
for i in {01..14}; do
curl -fSL -o lean.tar.gz.part$i https://alpo.ge/lean.tar.gz.part$i || { echo "failed: part$i"; break; }
done
# join and decompress (creates ./lean)
cat lean.tar.gz.part{01..14} > lean.tar.gz
tar xzf lean.tar.gz
# size, file count, line counts
du -sh lean | cut -f1 | xargs printf '%s on disk (lean folder)\n'
find lean -name '*.lean' | wc -l | xargs printf '%s .lean files\n'
find lean -name '*.lean' -print0 | xargs -0 cat | wc -l | xargs printf '%s lines total\n'
find lean -name '*.lean' -print0 | xargs -0 cat | grep -cv '^[[:space:]]*$' | xargs printf '%s non-blank lines\n'
|
|
| ▲ | hafeoz 2 hours ago | parent | next [-] |
| 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 :) |
| |
| ▲ | CrimsonCape an hour ago | parent [-] | | 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. | | |
|
|
| ▲ | causal 5 hours ago | parent | prev | next [-] |
| 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. |
| |
| ▲ | striking 4 hours ago | parent | next [-] | | 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. | | |
| ▲ | javra an hour ago | parent | next [-] | | Some of the reimplementations are actually verified, so guaranteed to be free from the sort of bugs found in the main kernel recently. | |
| ▲ | tomjakubowski 3 hours ago | parent | prev [-] | | 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? | | |
| ▲ | foota 3 hours ago | parent [-] | | You could maybe address this through holding out the other implementations and using them only to verify? | | |
| ▲ | ok_dad 2 hours ago | parent [-] | | 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. | | |
| ▲ | gus_massa 2 hours ago | parent [-] | | > 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/ |
|
|
|
| |
| ▲ | NooneAtAll5 33 minutes ago | parent | prev | next [-] | | 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" | |
| ▲ | empath75 4 hours ago | parent | prev | next [-] | | 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. | |
| ▲ | red75prime 4 hours ago | parent | prev [-] | | ...if it's tasked to find proofs of False assuming that there are still bugs left. |
|
|
| ▲ | pyentropy 5 hours ago | parent | prev | next [-] |
| And the human "book" is 250k pages (!)... |
|
| ▲ | steeleduncan 4 hours ago | parent | prev | next [-] |
| 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. |
| |
| ▲ | cogman10 4 hours ago | parent | next [-] | | 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. | | |
| ▲ | William_BB 3 hours ago | parent [-] | | > 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. | | |
| ▲ | hazbot 7 minutes ago | parent [-] | | 'intelligently search' is doing a lot of heavy lifting here But I do agree with you! |
|
| |
| ▲ | rhdunn 29 minutes ago | parent | prev | next [-] | | 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. | |
| ▲ | mirrir 4 hours ago | parent | prev | next [-] | | 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. | |
| ▲ | xpct 4 hours ago | parent | prev | next [-] | | 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? | |
| ▲ | ffsm8 4 hours ago | parent | prev | next [-] | | 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 | |
| ▲ | rocho 2 hours ago | parent | prev [-] | | 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... | | |
|
|
| ▲ | BoingBoomTschak 4 hours ago | parent | prev | next [-] |
| 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? |
|
| ▲ | smurf9852 4 hours ago | parent | prev | next [-] |
| Could a math proof really need 15M lines to write out?
I don't buy it. |
| |
| ▲ | guinep 2 minutes ago | parent | next [-] | | 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. | |
| ▲ | ks2048 2 hours ago | parent | prev | next [-] | | 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. | |
| ▲ | criddell an hour ago | parent | prev | next [-] | | What do you think the upper limit on the size of a math proof is? | |
| ▲ | lixtra 3 hours ago | parent | prev | next [-] | | 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. | | |
| ▲ | piloto_ciego 42 minutes ago | parent [-] | | 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. |
| |
| ▲ | runeblaze 2 hours ago | parent | prev | next [-] | | 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. | |
| ▲ | Maxatar 3 hours ago | parent | prev [-] | | 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. |
|
|
| ▲ | stabbles 5 hours ago | parent | prev | next [-] |
| And presumably you need hundreds of GBs of RAM to compile it. Would be nice if someone with a beefy machine could confirm? |
| |
| ▲ | pama 5 hours ago | parent | next [-] | | About 1 1/2 hour in, my estimate is it takes about one day on a single DGX spark. We will see. | |
| ▲ | sva_ 5 hours ago | parent | prev | next [-] | | You could run it with limited RAM, just not many Lean processes in parallel. Just need some disk space for the compiled files. | |
| ▲ | igleria 5 hours ago | parent | prev [-] | | 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. |
|
|
| ▲ | mtlmtlmtlmtl 2 hours ago | parent | prev | next [-] |
| 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. |
|
| ▲ | itsthecourier 5 hours ago | parent | prev | next [-] |
| why a lean proof is so big? it needs to be compiled and tested with a lot of cases or what may it be? |
| |
| ▲ | killerstorm 4 hours ago | parent | next [-] | | 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. | |
| ▲ | pretzellogician 3 hours ago | parent | prev | next [-] | | 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.” | | |
| ▲ | VirusNewbie an hour ago | parent | next [-] | | 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 | |
| ▲ | coffeebeqn 2 hours ago | parent | prev [-] | | Sounds like a massively parallel search of all possible paths since they have so much compute now |
| |
| ▲ | wasabi991011 4 hours ago | parent | prev | next [-] | | 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). | |
| ▲ | ht-syseng 5 hours ago | parent | prev [-] | | 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. |
|
|
| ▲ | newsy-combi 4 hours ago | parent | prev [-] |
| 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. |