Comment by gwern | Hacker News Reader

Comment by gwern | Hacker News Reader

Parent Full thread

I increasingly recommend against the Arxiv HTML version. I thought it had an acceptable start and they would fix the remaining problems and rapidly become on par with the PDF, but that seems to not be happening.

The HTML version is seriously buggy; and the worst part is, a lot of those bugs take the form of silently dropping or hiding content. It's bad enough when half the paper is gone, because at least you notice that quickly, but it'll also do things like silently drop sections or figures, and you won't realize that until you hit a reference like 'as discussed in Section 3.1' and you wonder how you missed that. I filed like 25 bugs on their HTML pages, concentrating on the really big issues (minor typographic & styling issues are too legion to try to report), and AFAIK, not a single one has been fixed in a year+. Whatever resources they're devoting to it, it's apparently totally inadequate to the task.

I think development on the TeX-to-HTML compiler has slowed down at some point, and it's far from perfect yet. Some of the issues are probably HTML5 limitations, unlikely to be fixed any time soon (unless one wants formulas to become graphics).

But there is another problem: It takes too long to load on mobile and doesn't reflow. I thought mobile was one of the reasons people wanted HTML in the first place!

In PDFs on arXiv, syntax highlighted codeblocks are graphics.

I think that's essentially only true if they are that in the original source. You can check for yourself, most papers have the TeX source available on arxiv.

> Some of the issues are probably HTML5 limitations, unlikely to be fixed any time soon (unless one wants formulas to become graphics).

You can convert a lot of formulas into either Mathjax/Katex-style fonts or MathML, or even just HTML+Unicode. (I get a very long way with pure HTML+Unicode+CSS on Gwern.net, and didn't even have to write a TeX-to-HTML compiler - just a long LLM prompt: https://github.com/gwern/gwern.net/blob/master/build/latex2u... )

But that's missing the point. Who cares about all of the refinements like reflow or pretty equations, when you are routinely serving massively corrupted and silently incomplete HTML versions? I don't care how good the typography is in your book if it's missing 5% of pages at random and also doesn't have any page numbers or table of contents...

Reply on news.ycombinator.com