I would love to see Spamassassin's bayesian filters merged with Mozilla: I could point at the "main text flow" (usually, a table cell) and say "this is ham", and I could point at a block of "sponsored links" or "text advertisments" and say "this is spam". The granualarity of content is usually a range of table cells or a frame. You can guess what's spam or not by page location and/or content. Get started people! :-)