68
Anthropic mathematician claims to prove Riemann hypothesis, most famous problem in all of math
(www.scientificamerican.com)
A community to post scientific articles, news, and civil discussion.
dart board;; science bs
rule #1: be kind
Hank green has a good video about this topic
The point is AI companies use mathematical proofs as a way to evaluate and improve their models. Its not about the proofs themselves, they are a side effect. They are also completely useless for the math community because nobody can understand them. The value of a proof is in the math you used to get there and usually a proof like this involves new math that can then be used to solve other problems. The AI proofs provide none of this. Mathematicians now know something is true or not but can not use that knowledge for anything.
look, i hate the slop machine built on stolen knowledge of humanity as much as the other person, but if you can be sure something is true or not, you can definitely use that knowledge, even if you do not understand the proof.
I think they mean that mathematicians cannot themselves use this proof to prove other things, or gain new insights.
The Riemann hypothesis itself isn't very useful. It says something about the distribution of primes. But we've known for some time that all known primes fit the hypothesis, out to over a trillion of them.
Yea but it is still FAR less valuable than a proof where we actually understand the math behind it and know how to apply it to other problems. For many proofs just knowing they are true isn't all that useful.
there is no argument there
In this case, it was something that for all practical purposes was already presumed to be true, but not formally proven.
If it found an unexpected result, that could be useful knowledge, but this case is of no practical use compared to before the proof.
Looking at it, it COULD be deconstructed to an actual, usable proof. If you had nothing but time on your hands. And that it held absolutely no bugs (1.3gb of code? Absofuckinglutely unlikely)
The bigger problem is that its confusing as all hell to the uneducated public. Its a proof in the sense that it technically proves the theorem. Its not a proof in that it must be able to be peer-reviewed. AI modeled proofs are intended for the AI. They cannot be peer-reviewed, thusly they are useless.
It is, however, fun to entertain the idea that a proof does exist, we just have more work to do to get there.
My son explained it to me. Basically those files are automatic proof files, and they are syntactically correct. But if you ever wrote a program, you'll know that getting it through the compiler without error messages does only get you so far.
According to him, they already found serious issues with the actual logic.
When AI writes code, especially large code bases, it tends to be quite disorganized and unintelligible. It takes you all but 5 minutes to pick out 10 things that piss you off like, “why the fuck would it do that?” It’s complicated code, much more so than the domain is complex in its own right.
The “proof files” is just the codebase containing the code which in theory proves the theorem.
Nobody has the time to proof read that gargantuan piece of shit.
But the bigger point is: If a human had done it, then other humans would have been able to use the result. At least other field experts would find the text legible.
I think it is just as likely that it exploits the Lean theorem prover. IIRC, its evaluation model itself is known to have logic holes for convenience reasons.
It appears the future of math might be reviewing AI proofs. It's one of the last fields that managed to keep automation at bay. That's over now.
Not the last field:
