There were some interesting algorithms that came out of the contest. There was a matrix factorization algorithm that worked pretty well. Another "guy in a garage" tried to incorporate the anchoring effect into their models.

I don't think Netflix used any of the algorithms. I suspect they have more data on the user and movies than was presented in the contest.

I agree, if all of the content is garbage, then you don't need a model, you could simply pick something at random.