Imagine Google Maps, but instead of maps, you are looking at images of PDF files. These can be annotated by lots of different people, and the annotations are moderated, so you can see the annotations of experts in the field, or trolls, or only people in your research group, etc. Annotations would have to be able to be made private.
It would be automatic for PDFs that are already available on the web, but journal suppliers might also be amenable to it because it isn't really good for printing, it's only good for online reading. But who knows.
Imagine being able to mouse over a reference, and see not only the citation, but the abstract! And maybe even have the full text one click away.