One builds systems out of humans, ML, and other computation. Ideally integrating the strengths of each, and doing fine-grained mutual compensation for the weaknesses of each. For instance, automated task management which delegates easy cases to ML, and orchestrates humans for harder cases.
When a post is rated, a system ideally learns about the post, the poster, the rater, and the rating question. The candy machine which bribed undergrads to redundantly grade CS exams, the responses giving insight into the grader, the gradee, the exam, and the instruction.
People are multifaceted. I'm interested in someone's apps, but not their cat. In when they are writing about their research focus, not on their gone-full-nutter . It would be nice to be able to stream one's own diverse interests, without burdening people who wish to follow only one of them.
Sites, group moderation, and thumb ratings, are units of very very very course grain selection. Imagine tiktoc or netflix with everyone sees the same thing, one size fits all, and if you want something else, go somewhere else. I don't even want the same thing from hn comments at different times of day. Our capabilities are oddly impoverished - even last-century usenet had "I don't want to see any more comments from this person". Skimming hn comments for ones which match my preferences of the moment, when feasible to delegate to personalized automation, the system design could facilitate that.
User customization could be open, transparent, and collaborative. Not "you looked at this once weeks ago, so here's more, and you can't stop me". Composably collaborative - not a swamp of manually-viewed isolated playlists, but accessibly searchable/queryable - intersect playlists which include this video. Beyond "this person does good videos", to "this person does good video ratings", to "these people are good raters", to "these people do good ratings of raters"... for your personalized interests of the moment.
It's late, but hopefully those are clarified highlights.