I finally read Paul Graham's seminal paper on Bayesian spam filtering yesterday (thanks to Jason Kottke for pointing me to it). I have to admit that my feelings are mixed.

On the one hand, I like the idea of an adaptive spam filter. To paraphrase somebody (was it a Supreme Court justice?), "I know spam when I see it." Besides, one man's spam is another man's portable, canned, processed meat product. So to speak.

On the other hand, it feels like closing the pantry door after the cans have gotten out. So to speak. By the time spam gets to the mail client, it's already done its primary damage, wasting bandwidth. If you try to push the spam filter further upstream, you lose the ability to define what constitutes spam; it becomes a collaborative definition or (worse) someone else (like your bandwidth provider) defines it for you.