If his proof is 250,000 pages, then I am extremely dubious. Just today, I was working with about 1000 lines of code and asked opus 5.4 to make a change. It introduced an obvious bug, that actually ended up doing the exact opposite of what I asked to do( print out less lines of logging, instead of more).. even if he has access to some kind of super LLM, I am extremely skeptical that it can output a 250,000 page document that is coherent and bug free, no matter what this guy says.
I would assume they do not produce the proof at once, and Lean is different from programming language (as far as I understand) in the sense that you can check if a program is correct or not within the Lean model, so it seems ideally suited for AI as they can experiment and extend the proof without ever producing a bug. I assume the real difficulty is checking that what you have proven is actually what you wanted to express but I would assume this is easier.
Makes me wonder if general programming languages are too expressive for AI making it too easy to express conflicting or incorrect ideas, we might need something like Lean but for general purpose programming. I'm aware there are some formal systems for that but I think more along the lines of general purpose programming.
Did you deliberately pick Opus 5.4 over Opus 5.5? That was the biggest thing that stood out to me in your comment. Opus 5.5 has been an all time banger for me.
You don't have access to the secret model the labs keep for themselves. Neither can you spin up thousands of subagents and burn billions of tokens a day.
Proof can be total bunk but underestimating the labs is silly.
Its a formally verified proof. Its the same thing as saying your code compiled successfully (for a compiled language). If opus made a change and your code still compiled, you would certainly still say, my code compiled. The statement being asserted in lean is analogous to my code compiled.
Lean works based on this: https://en.wikipedia.org/wiki/Curry%E2%80%93Howard_correspon...