A few thoughts on what you wrote:
1. I\’m quite sure a straight binary classifier is not the answer. Interestingness is different from the spam/ham problem because the answer is fuzzy rather than black & white. My naive guess is that the best user interface will be one that subtly marks to the user that an article is worth reading (say, a brighter red used for the entry header) or not worth reading (a duller grey for the background). Ordering (higher noise-to-signal stuff down near the bottom) is also a possibility. I find it helpful, when dealing with AI judgements that are not too accurate, to let the user interface be as vague as possible.

2. I definitely agree with you about the advances since 2003–the mechanical turk has been great! I\’ve been thinking of different ways to use digg or del data, but I\’m not sure if Joshua or Kevin would be too happy with me pounding their server every time a new feed item comes in.

I know I haven’t – and I’m up to my 6th private attempt or so at different arrangements with Bayes in particular. I think a different form of filtering is what’s needed. Maybe LSI, maybe some other form of valued scoring that doesn’t result in a flat spam/ham answer. (ie. “interestingness”, as you say.)

But, in the time since 2003, what’s really shown promise are more and more varied ways of soliciting and exploiting human intelligence in the course of finding and filtering news and feed items. See: del.icio.us, digg, etal. The best news lately is pre-scanned and filtered by human domain experts.