The problem with using our friends' activities as predictors of interest is that Facebook has no way to distinguish between strong and weak friendships[1]. There is a marked difference in my life between Facebook friends and meatspace friends.
[1] - Sure, they can determine which of my friends I spend the most time talking with on Facebook. But that says more about our relative Facebook activity than the quality of our friendship.
The real issue is that cases like those of anthuswilliams above you are unique and difficult to replicate. That's one of the noted problems with music recommendation services like Pandora too: What matters isn't what users tend towards but why, and why is a question that can't be effectively computed. It's more freeform.
I think the real solution would be a site that actively helps a user be responsible for himself. Something that doesn't simply indicate what features are available, but why users might want to use them to sort out their friends. Contrary to what the OP says in his blog post, that is possible today, but it requires responsibility on both the parts of the reader (who must sort his inputs neatly) and the publisher (who must figure out where to publish what). And web sites presently don't feel it's worth their time to help you with that task.
Now, interesting is different than "like", which is different than passion or love, which are all different than studying for school or work, or an obligation.
But anyway, a very good question, you gave me inspiration to delve on this topic and come up with lots of different approaches to determine what people will thing it's interesting.
Furthermore, what algorithm would be best? A type of Bayesian filtering is really effective at separating spam from non-spam, would it be effective at filtering interesting from non-interesting?
Applied to a social graph, it would (hypothetically) find content that is both recommended by friends and probably new to you.