I will try to post a little about this once I’m back from vacation. Jeff may want to comment too. For my purposes, I’ve found SpamAssassin pretty effective. The recent versions do have a Bayesian module, and if that’s trained properly, the percentage of misclassified mails goes down even further.
The adaptive / Bayesian filters tend to work best for individual use; Once you have over a certain number of users sharing a wordlist or whatever, the effectiveness drops somewhat. They also need to be trained actively on an ongoing basis, and doing this (easily) is difficult without integration with the mail client itself.
SA’s resource consumption is one reason that we haven’t encouraged its use or made it an “officially supported” program.
(ps - the person who wrote that article should learn to use a spellchecker)