8 1/2 x 11 paper can be oriented in two directions, portrait and landscape. As best as I can tell, all papers in all journals are published in the portrait orientation. This made sense in 2000, when the PDF was just a brief waystation between the journal's website and the printer.
But now, I read papers on the computer most of the time. So do you, I bet, web-savvy blog reader. Most computer monitors are in the landscape direction. Don't you want to see papers oriented 90 degrees offset?
I still do read printouts, so papers should come published in both directions. Preview and (ugh) Adobe Reader, and the other viewing programs should bring up a paper, as a default, in landscape mode and then, if you hit "print" ask if you want it in portrait.
I doubt this can be done without doing all the fancy-pants type-setting stuff twice. You'd have to rearrange the text into 3 or 4 columns, move the figures around, probably resize them a little bit. All that hassle will need to be done twice. It may double the size of the document. It's worth it.
While I am on the topic, it's time to get rid of numbered citations forever. It's time to replace them with real in-text citations. It's seriously ungainly to shift back and forth between page 2 and page 17 to find out that  means Smith et al. 2004. I often literally ⌘C-⌘V a paper I am reading and open both copies, one to where I am reading and one to the ref list. If logic were king, you should be able to hover the mouse over the citation and your PDF reader will bring up a box with the full citation, including full article title. I can think of how to do this with a very little intelligence in the PDF reader and two bytes per citation, so it's cheap, space-wise.
When you download a PDF, the name should be "Neuropolarbear_etal_JON_2010.pdf." This naming scheme makes them obvious for someone looking at the title, and also compatible with the search engine on my computer. Instead papers are usually called "science.pdf", "1-s2.0-S0006322310010103-main.pdf", and "3726.full.pdf" to name 3 papers currently on my desktop. I realize commenters will point out the beauty of Mendeley right here, and I agree, but there is no reason to not also give your PDF a reasonable name. This can be solved by the journal in less than a minute per paper!
Authors submitting a paper for review should be pasting the text for each section (Methods, Results, Discussion, etc) into a box on the website and at no point should upload a word document. Then the website would create a PDF. Think of the numerous advantages:
(1) Reviewers could request a single- or a double-spaced version of the paper, depending on their preference.
(2) No need to for authors count words in each section. The server does it for you. No need to reviewers to estimate words when authors don't provide this info. Why does Frontiers ask its reviewers whether the paper violates word length policies? Computers can do that.
(3) When the server creates the PDF, it inserts line numbers. As a reviewer, I love line numbers. Of course, if for some reason, you dislike them, you can request the server to make version without line numbers.
(4) No need to submitters as PNAS to insert weird markups into their paper anymore. In fact, doing it this way would make it much easier for editors to guess at page length at all journals. And I take it that page length is something editors care a lot about.
(5) Figures could be placed in-line. Much easier to read that way.
(6) Reviewer wouldn't have to pedantically point out that Neuron has a Summary instead of an Abstract and an Experimental Procedures instead of Materials and Methods. The program would take care of it.
(7) The program could provide rudimentary typesetting to make the review copy look a little nicer.
(8) Spell checking!
(9) No more realizing that track changes was on, but not visible on the screen, and you accidentally inserted comments into your review manuscript saying "reviewer 2 is a fucking asshole. change this so that you cite their latest stupid paper."