- 135comments
- 185comments
- 33comments
- 219comments
- 58comments
- 11comments
- 45comments
- 242comments
- 71comments
- 18comments
- 237comments
- 35comments
- 17comments
- 51comments
- 9comments
- —discuss
- 168comments
- 146comments
- 24comments
- 13comments
- 3comments
- 89comments
- 174comments
- 16comments
- 72comments
- —discuss
- 15comments
- 280comments
- 148comments
- 136comments
Seems like this would have strong implications for distillation and/or smaller types of transformers!
How? I don't see it. (I'm familiar with the ML side, not the combinatorics side.)
No, this is pure graph theory, and is quite far away from anything machine learning.
Actual meat: https://arxiv.org/abs/2510.20765
Isn't it actually the bread? The meat is given, if I understand correctly.
Nice, it's pre-AI
I am not a mathematician but are most papers now accompanied by a lean proof?
Is there a central repository of lean proofs shared by mathematicians like an npm repository of JavaScript packages?
Does it all depend on a stupid is-odd package in the end?
It's new but there is actually a registry now: https://palomar-registry.org/
LLMs have gotten good at creating Lean proofs so the are much more common but not universal. And they depend on https://github.com/leanprover-community/mathlib4
No, almost none (except for in certain fields, such as HoTT) have formalized proofs.
Why not? It seems like this should sort of be the standard now? Or is it hard to make lean proofs in all fields?
i think there are a few reasons.
- lean proofs are hard, and a lot of the time there is so much mathematical machinery that folks are working on that you would need to not only prove your result, but also all of the machinery that your subfield it is built on. it would be infeasible for many authors to do all of this work (this might be a major part of multiple careers, and when there are 5 folks in your entire subfield, the payoff is not really worth it)
- human proofs are readable, and can illustrate concepts better than lean proofs. human proofs give insights into how to think about a type of problem, and this is often the most valuable part of a proof/result.
- lean proofs are often very difficult to read; while they give you a "verified" check mark, they do not necessarily improve the bounds of human understanding if that makes sense.
Wondering: if the process for the upper part of the sandwich is the complement of the process for the lower part, why was it so much more difficult? What would go wrong if you took one of the earlier lower-sandwich processes, and complemented it in a similar way? I have to assume it's something, but what?
Hilarious - a mathematical result that afaict has nothing whatsoever to do with AI, and 75% of the comments are about AI, including this one!
lol yeah we're doomed