While probably computationally too expensive, I would like some system where up/downvoting isn't about objective quality, but only about personal preference. Essentially the system would "cluster" up/downvote behavior from users like youtube clusters like/dislike of videos and then recommend posts which people who like the same content as you like and people who dislike content you like dislike. I am not sure how many clusters/dimensions you would need though and I guess individualized sorting would quickly become computationally prohibitive as you would have to do a scalar multiplication of the post-dimensions with the user-dimensions for each post and then sort the stuff.