A PDF can include complex mathematics, which most papers I read do. This is generally done badly or not at all in other media.
This can be handled with some effort with epub and Mobi. In HTML it requires either reference to generated images or some javascript (or browser specific support).
Oh yes, there are ways of working around the limitation. But they are all just workarounds that create difficulties for either the writer or the reader, e.g. MathML is not practically writable by a human being. MathJax looks like the nearest thing to an actual solution for HTML (even if under the hood it's made of workarounds around workarounds), but I've never encountered a site that uses it other than the MathJax site itself.
In our discussion of academic papers, Lukeprog argued that lots of smart people preferred to read ideas in academic paper format. Based on my observations, I mostly disagree. But that's just anecdotal evidence. Let's use Science!
Suppose someone at the Singularity Institute thought up a cool new idea: it could be about rationality, Friendly AI, decision theory, making money, or any of the other topics we discuss here on LW. Explaining it takes about ten pages, and it's nontechnical enough that it can be explained to a general audience of non-mathematicians. Which of the following explanations would you be most likely to actually sit down and read through?
EDIT: To state the obvious, this poll will be biased in favor of blog postings, since it's on a blog. However, I still think it'll provide data that's much better than anecdotal guessing. I've emailed a few rationalist mailing lists to try and counteract this effect.