Yeah I’m wondering how on Earth a single conjecture can require 17 million lines of code to prove, but if it involved building up an entire ecosystem of theorems that had not yet been formalized then I suppose it makes sense.
The proof has also been built in parallel, so it's possible to mechanically shrink it a lot. Some equivalent lemmas have been proven more than a hundred times independently because different agents needed them.
[dead]
Anyone who has taken a coffee break during vibe coding will be familiar with the perils of unsupervised universe reinvention.
> ...if it involved building up an entire ecosystem of theorems that had not yet been formalized then I suppose it makes sense.
This is pretty much it. Pre LLMs very little maths had been formally verified and a lot of it relied on proofs from humans convincing other humans, and issues did arise which would occasionally mean chunks of maths we previously thought were true had to be reconsidered.
Proving something in lean probably involves writing a lot of proofs for new areas, and the formal proofs are usually significantly more verbose than a human proof where a lot is often left for the reader.