I am mad at this proof. Someone that did not understand what was going on (by his own public admission in his post) got a swarm of mechanical monkeys to try random things until they worked, and now we have an unreadable proof that the community needs to work to understand while getting no credit for doing the dirty and boring work
post
Man, mathematic proofs sure have changed since I was in college.
1.3GB seems a lot
Or the AI stole the credits from this guy: https://news.ycombinator.com/item?id=50579568 "It is all human, with no AI"
From Terry Tao's blog, wow, they also claim to have shown singularity in unforced Navier Stokes. The earlier announcement used the forcing term in the Clay problem. Don't worry too much if you don't understand that. Basically there was a million dollar prize offered to solved a certain problem, but the problem as stated in the prize offer had an extra term added to make it easier, since the people who wrote it thought it would be impossible without the extra term. They solved the "easy" version several months ago and now have apparently done the hard version.
Theft theft theft
“Anthropic mathematician” 🤣🤣🤣
What's more likely, that the Riemann hypothesis was proven, or that there's some bug in the Lean prover they used that this 1.3GB of code manages to run into? I'm placing my bets on the latter...
If a coding agent produced 1.3GB of code and claimed it solved my problem, it would sure have bugs.
There are multiple kernel implementations of Lean. The proof will likely be checked against every kernel implementation available. It’s unlikely that it’s exploiting soundness bugs in all kernels at the same time.
I wonder if the Lean proof could somehow be automatically morphed into a Rocq or HOL proof and checked by those. (Rocq and HOL are proof checking systems unrelated to Lean, but with the same general idea).
Interestingly, one of HOL's software self-tests involves a large cardinal hypothesis. Explaining what that means would be a bit messy, but if you already understand it, it's sort of a lolwut moment.
The Lean experts are running the Lean code in different Lean kernel versions to avoid this problem. Some of the kernel versions are verified themselves - if I remember correctly by earlier proof checkers.
If they succeed in running it, it’s as trustworthy as humanly checkeable
The proof is probably valid but as other people have mentioned, it's being run through a bunch of different Lean implementations to make sure.
Lean is more of a proof checker than a proof finder. Checking proofs is much easier than finding them, so people generally aren't too worried about Lean. An Anthropic AI model found this RH proof and it's checked out according to at least one Lean installation so far, but they're repeating the checks as extra precaution.
The details of the proof aren't very comprehensible to humans right now, but the overall strategy apparently has been examined and makes sense.
Amazing—they got AI to successfully recreate the quintessential mathematical crank.
Press "X" to doubt.
It'll be another hallucinated solution, most likely. With very little press given when some actual mathematician debunks it a month later.
This guy: say you solved this theorem
AI: I solved this theorem
This guy:

Are we ignoring that the proof is posted? Doesn't mean it doesn't have issues, but so did Andrew Wiles' proof.
What a coincidence: https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/
Wiles proof was a human readable white paper that went through the peer review process.
The anthropic proof is 1.3 GB of lean. It will take some time to go through and at the end of the day its code. There might be bug in the interpreter. The anthropic engineer admits he doesn't understand it.
Its not really a good comparison. But the anthropic proof might end up being right all the same.
I can't wait to hear about this guy smearing his shit on the walls of an IHop bathroom while he screams about how right he was.
I'm sure that the AI is confidently incorrect about it, as well.
"Did I give you the wrong answer that took up 2 weeks of you poring through gibberish to find the mistake? That's on me. I meant (insert another confidently wrong answer)".
We really need to put a 1.44MB limit on lean proofs. If you can’t prove something interesting in that; then you don’t understand the problem yet.
The number of theorems is infinite and the number of theorems provable in any finite number of symbols is itself finite, so no matter how high you set your limit, almost all theorems won't fit.
So far we've been limited to proving theorems which have proofs that can fit in the human brain, but that's our limitation, not a truth about mathematics.
So far we’ve been limited to proving theorems which have proofs that can fit in the human brain, but that’s our limitation, not a truth about mathematics.
I wouldn't even say that. There's probably no human who understands the classification of finite simple groups aka "the enormous theorem". Its proof is around 10,000 pages spread across 100's of papers over 20 years of math journals. It has a lot of separate components and for each component, there are people who understand that component. And the general structure is understood. But probably no individual person understands how every component works well enough to recreate it.
Good, we don’t need theorems that are not understood and not understandable.
Much like knowing the fact that a rock will roll down a hill is useless without the understanding of gravity and even Newton physics
Numberphile published a quick reaction video: https://www.youtube.com/watch?v=sMZMpsHeFKc
The singularity has arrived, for the mathematicians.
Edit: it's funny to see people still dismiss the AI's accomplishments even at this point.
Though seeing, they do not see; though hearing, they do not hear or understand.
I know nothing of math; but I am cynically assuming most of the work is stolen from an unpublished preprint by someone else.
i know nothing about math
i knew quantum computing was already putting pressure on cryptography, but does this proof ultimately mean that cryptography using prime numbers will become a thing of the past?
I know a modest amount about math (happy to be corrected), and I’m pretty sure the answer is no, because people have already built and run better-optimized routines for factoring primes by assuming the hypothesis is true.
Now we know that those routines are (allegedly) giving valid answers, which was unproven before, but the gains are likely marginal. The proof doesn’t let us do anything new we couldn’t do before.
all 41 comments