So I get to thinking about bayesian filters which have recently been used to fight spam. The idea is that you build a corpus of 'bad' text (in this case spam) and a corpus of good text (in this case legitimate email), and compare every subsequent incoming email to each corpus. The one it matches the most determines whether the mail is spam or not. Apparently this works well. Each new email is added to the approrpiate corpus making the system stronger as time goes by.
So what about applying this system to web pages? Add two extra buttons to your browser interface, 'Good' and 'Bad'. At each web page you visit you click either the good button or the bad button depending on your experience there. The browser adds the page to either the good corpus or the bad corpus. Over time, the browser, by comparing the page to both sets of data, will begin to be able to tell you (reasonably accurately?) whether or not the new page you are looking at is of interest to you or not.
Extending the idea, your personal Bayesian enabled web spider could be trawling around the web finding pages it thinks (knows!) you will be interested in.
Stupid idea?