A researcher at AI company Anthropic said Friday that he had used the firm's Claude chatbot to prove the Riemann hypothesis, a problem that has eluded mathematicians for more than 160 years.
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.
It’s hard to understand what it means for the field, especially from the outside.
Can this help the mathematics community gain knowledge, insight, new method, and be a stepping stone for others developing new proof either on their own?
It sounds like an automated system wrote a proof, another automated system verified the proof, and it’s not clear what’s the level of mathematician involvement. 17 millions lines of code is huge, is that typical for lean proofs? Is it possible for anyone to understand it, or is the goal to pass it to checker and claim a prize?
From what I’ve read, this proof confirms what a lot math has assumed to be true for over a century. Once people can go through and figure out how it was solved, that might open up new ways of approaching other problems.
As someone very anti-LLM, even I have to admit that being able to upgrade concepts from “assumed” to “confirmed” is a good thing!
It will be interesting to see what comes from understanding the proof, or how much it can be simplified. Entire branches of mathematics can spring from attempting to prove something, and I hope that process continues.
17 million line means it is completely impossible for a small team of human to understand it.
For reference, we did slop proofs before, a proof assistant expret with full knowledge of the original pen and paper proof (that means he knows as much about the proof as the person who wrote it), can clean up through around 2K lines of slop formal proof on a good day, without doing any other work.
A very naive extrapolation means it’d take decades for someone to understand a 17 million-lines proof. It’d be interesting to see how much smaller the proof could be if some LLM had the ability to simplify and clean it up, and whether the current proof is clean enough to publish.
The current one is certainly not even close to publishing level.
And we observed that AI are terrible at cleaning up its own mess (not surprising, because it wrote said mess).
So this lean proof really is just a certificate for correctness and nothing more: it will have a hard time conveying human knowledge, and it will not be reusable to support other proofs.
The only way to comb through it is likely paying anthropic/openAI even more money to let AI work with experts to find the crucial pieces, and rebuild it from scratch.
IMO, correctness is arguably the most overrated attribute of a mathematical proof.
Simply having a lean certificate that is not reusable only adds very little value to the research in the long run…
It’s hard to understand what it means for the field, especially from the outside.
Can this help the mathematics community gain knowledge, insight, new method, and be a stepping stone for others developing new proof either on their own?
It sounds like an automated system wrote a proof, another automated system verified the proof, and it’s not clear what’s the level of mathematician involvement. 17 millions lines of code is huge, is that typical for lean proofs? Is it possible for anyone to understand it, or is the goal to pass it to checker and claim a prize?
From what I’ve read, this proof confirms what a lot math has assumed to be true for over a century. Once people can go through and figure out how it was solved, that might open up new ways of approaching other problems.
As someone very anti-LLM, even I have to admit that being able to upgrade concepts from “assumed” to “confirmed” is a good thing!
It will be interesting to see what comes from understanding the proof, or how much it can be simplified. Entire branches of mathematics can spring from attempting to prove something, and I hope that process continues.
17 million line means it is completely impossible for a small team of human to understand it.
For reference, we did slop proofs before, a proof assistant expret with full knowledge of the original pen and paper proof (that means he knows as much about the proof as the person who wrote it), can clean up through around 2K lines of slop formal proof on a good day, without doing any other work.
A very naive extrapolation means it’d take decades for someone to understand a 17 million-lines proof. It’d be interesting to see how much smaller the proof could be if some LLM had the ability to simplify and clean it up, and whether the current proof is clean enough to publish.
The current one is certainly not even close to publishing level. And we observed that AI are terrible at cleaning up its own mess (not surprising, because it wrote said mess).
So this lean proof really is just a certificate for correctness and nothing more: it will have a hard time conveying human knowledge, and it will not be reusable to support other proofs. The only way to comb through it is likely paying anthropic/openAI even more money to let AI work with experts to find the crucial pieces, and rebuild it from scratch.
IMO, correctness is arguably the most overrated attribute of a mathematical proof. Simply having a lean certificate that is not reusable only adds very little value to the research in the long run…