upvote
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.
reply
> ...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.

reply
Out of interest, which maths needed to be reconsidered as a result of attempted lean proofs failing?
reply
Our legal systems across the planet desperately need this.
reply
It's going to be fun when legal systems are analyzed by LLMs. The chances that hundreds of thousands (millions?) of laws and regulations written by legislators, judges and bureaucrats over hundreds of years are self consistent and contain no internal contradictions is very, very, very close to zero.
reply
Laws aren't supposed to be strict logical statements that contain no internal contradictions. Laws are intended to be interpreted by judges for a reason.
reply
Not at all. What an anti pattern that would be.
reply
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.
reply
I mapped the 175k files by imports into 2,647 clusters in 63 themes: https://riemannhypothesissolved.com

The main theorem's cluster directly imports only three others: a zero-free region, primes in progressions and Mertens-type zero sums. Most of the code sits 3–5 import steps below those, and no theme is more than about 4% of the code.

(I made the map. Cluster names are written by Claude from code samples, so they can be wrong.)

reply
Anyone who has taken a coffee break during vibe coding will be familiar with the perils of unsupervised universe reinvention.
reply
I don’t know about anyone else, but I’m yet to come back from a coffee break and find my agent has solved a millennium math problem. I guess there’s always P vs NP?
reply
> I don’t know about anyone else, but I’m yet to come back from a coffee break and find my agent has solved a millennium math problem.

Have you tried asking it to?

reply
Unfortunately, it is not. The same level of slop nuke occurred with Anthropic's auto-formalization of Fermat's Last Theorem (~13 million lines of code). What Anthropic's FLT did do was undercut the human-led efforts to get idiomatic formalization of the requisite mathematics surrounding Fermat's Last Theorem into Mathlib.

This result is not quite as harmful as Anthropic's FLT was because it is indeed telling us something new -- that the Riemann Hypothesis is true, with full confidence -- but it is not reusable in any sort of way, beyond proof mining to figure out what in the world the AI model did. It is still in fact harmful to the greater project of mathematics formalization (irrelevant to foundations), because it undercuts the ability to fund such efforts for reusable proofs.

reply
deleted
reply