The spam tuneup I performed last month is working well. I certainly feel like I'm seeing less spam. The stats are good too. 3100 of the messages I received this month were classified as spam (that's ⅔ of my mail!). Of those 3100, 2900 were caught by the Bayesian detector with 99% confidence. About 1200 were caught by Vipul's Razor and about 1200 were caught by Pyzor.
Bottom line: the Bayesian stuff really works if you train it. The collaborative things are helpful but not complete.