Just a reminder not to take the progress bar here as an indication that the proof is largely correct, as it takes only one hole for the entire thing to come crashing down.
Judging by other commentary so far, my understanding is that there is a major hole in this proof, which would not surprise me given that the machinery involved feels far less like a major leap (to me anyway) than the route for the Alpoge proof.
The asymptotic-order theta lifting in that proof feels like a genuine step forward for other mathematical disciplines. But I'm not a number theorist, so happy for someone else to pitch in and correct me.
Also, IIUC, the progress shown is as a fraction of issues filed. It’s not clear to me, even to an order of magnitude, how many issues they have left to file
Who’s checking that the AI formulation of the theorems in lean actually capture Massullo’s statement?
This is a really cool effort but even if all the issues close, the code still needs to produce a proof of mathlib’s RH statement. There’s a nonzero chance that everything looks like it’s tied together but still can’t produce a proof of Riemann.
They are working through Masullo’s 900-page purported proof of the Riemann Hypothesis and trying to formalize it, fill in the gaps, and repair any mistakes. When they’re done they will either have created a relatively concise machine-checked proof of the RH, or they will have developed a pretty good idea of where Masullo’s proof fails.
Or they have a bundle of correct theorems that somehow don’t fit together to prove the Riemann hypothesis, because one or more of them was misstated in Lean but still proven true
Alongside the front page post on P=NP, the public nature of a dashboard like this gives me a sort of optimism. Seeing this type of thing start to happen in public forums rather than just behind the closed doors of the big labs is exciting.
I mentioned in that thread, it feels like there is a twitch-plays-Pokémon moment to look forward to in something like this where we can all enjoy cheering on agents from the sidelines toward new discoveries
It’s cool to have two such posts in one Sunday, I think we are near a new chapter in open science
Their best case scenario is to prove Masullo correct with minimal repairs. If that happens then we will have a relatively short proof of the Riemann Hypothesis, and Masullo himself will already understand most or all of it.
How come Anthropic can’t train a model that doesn’t make me want to jump off a building when I read the writing, considering what they’re capable of with math.
Math proofs are machine verifiable (true or false) by automated proof checkers such as Lean. Whereas judging writing quality is more subjective, thus more complicated and expensive to hill climb on.
Judging by other commentary so far, my understanding is that there is a major hole in this proof, which would not surprise me given that the machinery involved feels far less like a major leap (to me anyway) than the route for the Alpoge proof. The asymptotic-order theta lifting in that proof feels like a genuine step forward for other mathematical disciplines. But I'm not a number theorist, so happy for someone else to pitch in and correct me.
This is a really cool effort but even if all the issues close, the code still needs to produce a proof of mathlib’s RH statement. There’s a nonzero chance that everything looks like it’s tied together but still can’t produce a proof of Riemann.
I mentioned in that thread, it feels like there is a twitch-plays-Pokémon moment to look forward to in something like this where we can all enjoy cheering on agents from the sidelines toward new discoveries
It’s cool to have two such posts in one Sunday, I think we are near a new chapter in open science