I like to gripe about our current conference/journal system, and how ungainly a method it is for publishing and disseminating quality work. Not surprisingly, other areas of computer science appear to have similar issues as well.
Hal Daume, my colleage at the U. of U., has sparked off a rather interesting discussion with his post on NLPers about the need for a new open-access journal for NLP research. The original motivation for his post is partly NLP-specific, but the evolving discussion over at his wiki provides a closeup view of the process of designing a new publication model that sits somewhere between traditional journals and traditional conference.
It's too early to tell what will come of this, but if their model gains some traction, it might provide a good exemplar for future reform in other disciplines in computer science.