I’m no mathematician, and I’m genuinely curious: how do these things usually work? Are other mathematicians now expected to sit and read through literally a thousand pages of condensed math to review this? And somehow have enough focus to catch subtle mistakes?
Yes. Sometimes it takes other mathematician years to fully comprehend and understand 12-page proof, we're not talking about 250k pages and 900 pages.
Yes, the main objective is to review the proof, but the side objective is to understand it and publish papers that are followups, or perhaps that share a few of the initial ideas and then go to another route. In some cases, nobody believe that the proof is worthy and it just sit in a drawer forever.
Before AI, mathematicians would publish a proof (not in Lean) and other mathematicians would check it over months or years. Both steps would be manual. For example, László Babai published a proof of quasipolynomial time for the graph isomorphism problem in 2015. In 2017 Helfgott spotted a subtle error in the proof, which was then amended (as far as I know, the new proof holds).
Now we just feed it to Claude and ask it if it's correct.
Until last year, the answer was yes, other mathematicians would spend months to understand your proof.
But since this year, we now have a quicker answer with AI. See https://news.ycombinator.com/item?id=50580084