Bayesian filters work lovely
I have been running Bayespam, an implementation of Paul Graham's Bayesian mail filter idea, for a couple of days now, and it works really great. Of course, I had a huge corpus of spam and non-spam mails to train it on. No false positives, and only one or two false negatives so far. The only bad thing about insanely effective filters is that not having to deal with spam highlights how little ‘ham’ mail I actually receive.