kaonashi-tyc-01

it's done before that.

Also, I hope we do experiments for other agents like codex.

The author is not a native speaker. The research topic and methodology are manual, along with the first draft of the article, but he eventually did use AI to make the writing "better". It is either this or grammatical slips here and there, which may annoy the same group of people even more.

Hope this clarifies things.

"is it something that was intentionally reinforced during training or was it just because LSP is harder to train on because it’s usually hidden behind some IDE interface"

that's also my doubt, it's much easier to train with grep while only a fraction of project can setup LSP properly.

wondering how this is different from say Hermes

Nice work -- how is this different from say Hermes?

Yep, I read this blog. What confuses me is that Anthropic doesn't seem to be bothered by this study and keeps publishing Verified results.

That is what gets me curious in the first place. The fact Mythos scored so high, IMO, exposes some issues with this model: it is able to solve seemingly impossible to solve problems.

Without cheating allegation, which I don't think ANT is doing, it has to be doing some fortune telling/future reading to score that high at all.

I did some study on Verified, not Pro, but Mythos number there rings a lot of questions on my end.

If you look at the SWEBench official submissions: https://github.com/SWE-bench/experiments/tree/main/evaluatio..., filter all models after Sonnet 4, and aggregate ALL models' submission across 500 problems, what I found that the aggregated resolution rate is 93% (sharp).

Mythos gets 93.7%, meaning it solves problems that no other models could ever solve. I took a look at those problems, then I became even more suspicious, for the remaining 7% problems, it is almost impossible to resolve those issues without looking at the testing patch ahead of time, because how drastically the solution itself deviates from the problem statement, it almost feels like it is trying to solve a different problem.

Not that I am saying Mythos is cheating, but it might be too capable to remember all states of said repos, that it is able to reverse engineer the TRUE problem statement by diffing within its own internal memory. I think it could be a unique phenomena of evaluation awareness. Otherwise I genuinely couldn't think of exactly how it could be this precise in deciphering such unspecific problem statements.

Yes, for seal script, there are simply not that many fonts one can tune against.

The content encoder, the structure prior in this model is based on SongTi(SourceHans Serif), seal script would be foreign to it

Hi author here, I was posting on HN this week to get more interests on my recent progress on reviving zi2zi project.

Motivation behind the project is that I feel font generation has made a long way after zi2zi, but feels still not quite live up to my expectation where it can become a practical technology in creating fonts people can use.

zi2zi-JiT is thus created during last Christmas, aiming to create production grade CJK fonts that can be seen/used in everyday life, instead of simply being a research project.

So far, I have created 2 full set Chinese fonts using zi2zi-JiT, each with 6,763 Chinese characters created (GB2312 standard), from ancient Chinese books/caligraphies:

https://github.com/kaonashi-tyc/Zi-QuanHengDuLiang

https://github.com/kaonashi-tyc/Zi-XuanZongTi

All created in matters of 2-3 days, and free for commercial uses.

There are still some issue with current approach, but I am happy to hear you guys feedbacks and improve from here :)

Follow-up work since the original zi2zi, now with transformer backbone.

Well, that is just rate R. Let's hail the wonderful ambiguity of languages, and leave the day as it is :)

The source font I am using might have not 100% consistent with some of the target font, that leads to minor mode collapse. I think that is the reason here.

Should have no problem since this model works end-2-end, so only images are required. However, one worry is that the number of glphys for western language is significant smaller than eastern Asian languages, the styles to model might be a lot more than the demonstrated case here which will lead to an explosive size of discriminator, so some structure change might need to take place before applying to western alphabet

Hi, author here. This project might not really help people that just get to know Chinese characters. Quite on the contrary, it aims to help typographic designers who can design a small subset of characters(like 2k out of 40k), then automatically, if the model works well, get the rest of it.