This feels like it has nowhere near enough massaging of the writing style, it’s as if they just told Claude to write a book and used defaults. You absolutely must tell models to speak in certain ways, use guidance to reduce Claudisms, and run adversarial review with other separate models that take a viewpoint of a uniformed reader. That is the only way to start getting semi-reasonable writing for something like this.
So many people who are heavy users of llms do not care at all about the end result of their usage. They don't even read it even if something to be read is the deliverable.
I can't tell if they think no one notices this. If they lack taste. Or if they just don't care.
I actually feel like it’s the people that don’t use LLMs that much that do this, as the people that do use it all day surely must have a nearly visceral reaction when reading this?
And heavy users of LLMs surely know exactly how to get it to write decently?
This took like 12 hours to build on my MacBook. What’s the point of this? It takes longer to build a Linux kernel on older machines. There’s no Linux kernel build leaderboard.
Has anyone scanned through and found any unknown insights Claude might have used? Or is the whole thing just stuff we could already do but was past the scale of a human's ability to do so?
From what I can tell, yes, there are new insights. The proof goes by a route that nobody had seriously considered. It's a fundamentally new proof technique, and may open up new frontiers. At the same time, it relies on a lot of existing math from the last 100 years, so it's not like humans accomplished nothing of relevance in that time.
I don't think any one person would have been able to link so many areas of math to come up with this proof in the ai-form. Masullo's proof is definitely more digestible and human-plausible.
Something similar was said about Andrew Wiles FLT proof - i.e. that it was stunning a single person that traversed such vast mathematical territory.
Over the past 30 years (apparently I can't testify to this personally) it's much slicker but for the most part it is still covers mostly the same range of vast mathematics.
I do wonder if eventually someone will simplify it somewhat.
Apparently, if the ABC conjecture is true FLT becomes a one pager proof.
But the only "proof" we have for ABC is basically illegable with months of work and those who have taken the effort to try do not believe it to be correct.
To attack this disagreement apparently the majority of the current effort is on using AI to translate it to familiar mathematics arguments and lean to verify.
It feels romantic to say maybe the proof is correct but we're stuck between complex maths arguments people understand or alien seeming maths only one guy seems to understand.
Multiple groups using AI have tried to formalize Mochizuki’s proof, and they all got stuck in the same place: Corollary 3.12, which is the same place the human mathematicians pointed out the flaw years ago.
All evidence now points to that Corollary 3.12 is flawed, and Mochizuki’s proof would not work.
I saw something about that pretty recently. The formalization effort hit a snag in the exact same place that Scholze and Stix did when they tried to review the ABC paper by hand. Unless a fix is forthcoming, things are as they were.
At the very top of the README.md it makes it clear that the PDFs and ePUBs can be downloaded from the release artifacts, which view is linked directly. My humble opinion is that this is the most correct or at least standard way to do this on GitHub.
Rather than making my eyes bleed with claudisms, it needs completely rewriting. For example, the unitary hypotheses node page intro paragraph is:
The Lean code ends the trunk with one existence statement. A record of forty-six fields, called HypsAlt, has an element. This page says what that record is, what the words "definite side" and "closed bound" mean in the code, and what the three large branches prove.
But wouldn't an intro be so much nicer with a little narrative leading:
This final node of the main proof shows that a 46-field record called HypsAlt has at least one element. The Lean theorem takes seven inputs, including sixty literature hypotheses on unitary groups and a "closed bound" on families of arithmetic data. Three main branches use these inputs to prove the "definite side" condition and two estimates, then build the HypsAlt element.
This feels like it has nowhere near enough massaging of the writing style, it’s as if they just told Claude to write a book and used defaults. You absolutely must tell models to speak in certain ways, use guidance to reduce Claudisms, and run adversarial review with other separate models that take a viewpoint of a uniformed reader. That is the only way to start getting semi-reasonable writing for something like this.
So many people who are heavy users of llms do not care at all about the end result of their usage. They don't even read it even if something to be read is the deliverable.
I can't tell if they think no one notices this. If they lack taste. Or if they just don't care.
I actually feel like it’s the people that don’t use LLMs that much that do this, as the people that do use it all day surely must have a nearly visceral reaction when reading this?
And heavy users of LLMs surely know exactly how to get it to write decently?
I built a tool so anyone can run the Riemann proof on their spare CPU cycles.
And there's a global leaderboard too (6 people running it now). Graphics inspired by SETI@home
https://riemann-at-home.github.io/riemann-at-home/
I’m here for all the tools to understand this Riemann Hypothesis breakthrough!
I don't really understand the point of this, you don't need a distributed system, it only takes a few hours on any desktop PC.
This took like 12 hours to build on my MacBook. What’s the point of this? It takes longer to build a Linux kernel on older machines. There’s no Linux kernel build leaderboard.
I thought that the Lean project for the RH had been built and checked already several times with different kernels?
They have.
Wait hasn't it been done multiple time over?
Has anyone scanned through and found any unknown insights Claude might have used? Or is the whole thing just stuff we could already do but was past the scale of a human's ability to do so?
From what I can tell, yes, there are new insights. The proof goes by a route that nobody had seriously considered. It's a fundamentally new proof technique, and may open up new frontiers. At the same time, it relies on a lot of existing math from the last 100 years, so it's not like humans accomplished nothing of relevance in that time.
I don't think any one person would have been able to link so many areas of math to come up with this proof in the ai-form. Masullo's proof is definitely more digestible and human-plausible.
Something similar was said about Andrew Wiles FLT proof - i.e. that it was stunning a single person that traversed such vast mathematical territory.
Over the past 30 years (apparently I can't testify to this personally) it's much slicker but for the most part it is still covers mostly the same range of vast mathematics.
I do wonder if eventually someone will simplify it somewhat.
Apparently, if the ABC conjecture is true FLT becomes a one pager proof.
But the only "proof" we have for ABC is basically illegable with months of work and those who have taken the effort to try do not believe it to be correct.
To attack this disagreement apparently the majority of the current effort is on using AI to translate it to familiar mathematics arguments and lean to verify.
It feels romantic to say maybe the proof is correct but we're stuck between complex maths arguments people understand or alien seeming maths only one guy seems to understand.
https://www.newscientist.com/article/2522687-the-secret-proj...
You’re out of date on the ABC conjecture.
Multiple groups using AI have tried to formalize Mochizuki’s proof, and they all got stuck in the same place: Corollary 3.12, which is the same place the human mathematicians pointed out the flaw years ago.
All evidence now points to that Corollary 3.12 is flawed, and Mochizuki’s proof would not work.
I saw something about that pretty recently. The formalization effort hit a snag in the exact same place that Scholze and Stix did when they tried to review the ABC paper by hand. Unless a fix is forthcoming, things are as they were.
< apparently...
I believe this is true for asymptotic FLT (ie FLT for large enough n) and Mordell but not the full FLT.
Alas, it's also incomplete, and perhaps even 'wrong'.
There's no source for that other than few randos' Claude sessions.
The 'incomplete' is from the author himself.
Obviously. The "wrong" part though isn't.
I got the first 2 definitions. I'll take a break now.
> Nobody has built the Lean project in full yet
Huh? I thought multiple people did?
Yeah, I personally compiled the Lean source myself. It takes 5 mins to ask Codex to start the build, and it builds in a few hours.
Not exactly impossible to do at home.
> in PDF and EPUB
Okay, so there should be prominent links to the PDF and EPUB.
At the very top of the README.md it makes it clear that the PDFs and ePUBs can be downloaded from the release artifacts, which view is linked directly. My humble opinion is that this is the most correct or at least standard way to do this on GitHub.
I agree. I had no difficulty finding and opening one of the PDFs just for curiosity. (Of course I understood absolutely nothing in it)
I'll wait for the human rewrite.
Rather than making my eyes bleed with claudisms, it needs completely rewriting. For example, the unitary hypotheses node page intro paragraph is:
The Lean code ends the trunk with one existence statement. A record of forty-six fields, called HypsAlt, has an element. This page says what that record is, what the words "definite side" and "closed bound" mean in the code, and what the three large branches prove.
But wouldn't an intro be so much nicer with a little narrative leading:
This final node of the main proof shows that a 46-field record called HypsAlt has at least one element. The Lean theorem takes seven inputs, including sixty literature hypotheses on unitary groups and a "closed bound" on families of arithmetic data. Three main branches use these inputs to prove the "definite side" condition and two estimates, then build the HypsAlt element.
What part of “the load bearing branches name their seams” is unclear to you?