so when I choose something to listen to, I am either picking with or against what’s already playing.

AcousticBrainz has (had?) a bunch of dimensional measures of various qualities of a song. How I used it was to first define a set of maybe 4 to 8 songs to “set the mood” and then pick a list of a few hundred songs that were “closest to” those songs in all the AB dimensional measures, pre-filtering out artists and songs recently played. Then - the final step was to sort the remaining candidates by their similarity to the songs most recently played. It was still random within that list, but a weighted random with the most similar (by AB measurements) songs most likely to be queued up next.