Funny you should mention this, as I've had a 1/3-finished blog post written on the cons of this approach for a couple months.  (Or shall I say, "since 18 June 2009 14:57"?)

To make my unfinishable post very, very short: I definitely see value in the relative times, but bear in mind that it doesn't answer *all* of a user's questions.  There have been countless times on sites like Facebook that I've wanted to know how far apart 2 comments were, and there's just no way to do that when they both say "2 days ago".  So I'd ask that people take the have-your-cake-and-eat-it approach that Stackoverflow uses: relative times for the main timestamp, but with the *real* time in a title tag, so you can see it on mouseover.

Also, stay away from "yesterday", as it's not a relative time in the same sense as "1 day ago" is.  That's the hard part to explain, but I think you know what I mean.  Ah, I see you've done that on Silver Arcade.  Delightful.

I think that is a great phrase or benchmark to use when choosing to present a piece of information: "What question have we answered?"  or "Have we answered the question with this representation?"

Thinking about this: sometimes, the question the user is asking may not be the question you think they are asking.  for instance, you may be thinking the user is asking "how long ago was this posted"...but I might really be asking the question:  "At about 9:30pm yesterday, I was following an interesting discussion but then had to leave-wonder what was posted immediately after that?"  Much rather just to 9:30pm, yesterday for that use case, but that's not what you were thinking... which isn't wrong usage, nor is your expected usage wrong, you were just anticipating something else. As twitter found out after their big replies blowup, it's worth noting that people don't always use your tools in the same way that you do...