What I think is interesting is that they're beyond what general recommendation engines can provide; it sounds like they're down to making domain specific optimizations. I wonder how many of these techniques will be generally applicable.

I also wonder if the way Netflix framed the question is optimizing for the wrong things. They're having contestants try to determine how people would rate movies, based on the corpus's collective ratings. From a business standpoint, it might be better to focus on which movies people would like that they haven't seen.