[flir]

> attention concentration rather than interestingness

With sufficiently rich data, I think they're equivalent.

I think collaborative filtering works much better once you have a critical mass of users/data - enough that taste clusters emerge. Otherwise it just defaults to whatever's globally popular. Google Reader managed it quite well. Librarything is fantastic at it (far better than Amazon).

I wrote a simple recommendation engine for musical taste in the early 2k's (pre-last.fm), but it had very little data available to it. The result? It recommended Radiohead to everyone. (Looking back, this was my first exposure to The Bitter Lesson).