That's precisely the problem though. You cannot still read and understand an AI written proof at the current skill level of the AI being applied, because they're orders of magnitude longer than human written proofs even when they don't need to be, and spend most of that length on the parts that aren't important. This has been really thoroughly documented by expert mathematicians who are engaging with AI in public like Terence Tao and showing in detail how much work it takes working alongside AI to figure out how to understand AI generated proofs. With human generated proofs that process is forced to happen before publishing the proof because the new style of AI generated proofs validated only by formal verification is supplanting the old human peer review process that forced the burden of understanding onto the publisher and not the reader.
> You cannot still read and understand an AI written proof at the current skill level of the AI being applied, because they're orders of magnitude longer than human written proofs even when they don't need to be, and spend most of that length on the parts that aren't important.
That doesn't seem to be true. The OpenAI NS paper was 166 pages. Wiles-Taylor proof of Fermat's last theorem is 129 pages. The length is not unprecedented for a difficult unsolved problem.
To be honest, I feel like the difficulty of reading AI proofs is due to the fact that we are on the verge of being beyond human comprehension. This is a demonstrable fact as no human has figured this out despite the problem being open for almost 100 years.
> To be honest, I feel like the difficulty of reading AI proofs is due to the fact that we are on the verge of being beyond human comprehension.
I can see where that's coming from, but I really don't think it's the case. Even with Astra, the proofs you get are just off in a way that doesn't signal superhuman comprehension. As 9question1 says, a common theme is that they dwell on insignificant steps. Another one is that they'll often be full of terminology that either doesn't exist, or has this weird quality where it looks like it is trying to make some minor insight seem much greater than it is. At first glance, that'll often make it look like it knows more than you, but when it's really just doing the same thing but in a more complicated and worse fashion, that to me isn't a signal of comprehension at all. The bizarre thing is that despite all the "stochastic parrot" style nonsense you'll get in individual proof steps, they still often combine to something valid.
In either case, what all of this means is that the working mathematician still needs to go through, and generally completely rewrite, any proof output by an LLM. Otherwise you are passing the burden of unreadability onto the reader.
Yeah, that mirrors what I've seen throwing some of the leading models at a set-theory problem that's stumped me (https://mathoverflow.net/q/511601): in this case, the problem does not easily yield to the standard tools, but the LLMs do not recognize it as a major open problem they should give up on. So they seriously try it, but typically end up in a loop of inventing certain classes of simple solution or counterexample attempts, defeating them, and trumpeting each one as a major result, each time inventing some new terminology.
It's definitely quite curious that the AI labs are able to push these results through seemingly with pure brute force. Perhaps it's largely a function of how many monkeys you have attempting various constructions on top of the known results and strategies the models have memorized.
> This is a demonstrable fact as no human has figured this out despite the problem being open for almost 100 years.
That's not true. Alpoge and Buckmaster's related LLM-assisted blowup result (https://news.ycombinator.com/item?id=49605915) utilized a strategy developed recently by Cordoba and Martinez-Zoroa.
