For recommendation you can't do that much better because the oracle is fickle. TikTok claims to get 0.83 or so. A better recommender practically has more objectives like diversity. If your recommender shows you just arXiv articles about recommeNeeds pretty soon articles about crypto or soccer or oncology start looking good! I have some ideas of what 'better' is but can't cook them down to one number to optimize.

Part of it is that 'show me an article I'll like' isn't the right question at this point in time when I like almost everything that it shows me, particularly in times like this one when I'm taking photos and developing them and playing Arknights a lot so the ratio of articles it shows me to what it ingests is low. I've been told I'd do better thinking about it from the viewpoint of 'show me the best articles you can given a limited number you can show me' and the right data set for that would be a pairwise comparison, yet the thumbs up/thumbs down task is one that's natural to do for putting articles into a pipeline that I save and/or post to my socials.

I am thinking of how to build a general text classifier which is more state of the art but I'll need to make a test suite of hard but fair problems.