[Spivak]

Spotify's algo for discover weekly can be incredible but you have to give it a fuckton of data to work with. My liked songs is >8k I follow probably 1000 artists (the UI doesn't give me a count).

It will surface bands with <20k listeners and songs with <1k plays out from nowhere tailored to exactly what makes the brain go burr.

For small sample sizes Pandora is still king.

[mrguyorama]

Except Spotify's discover weekly is stupendously dumb and desperate. I've put 10 years of my music listening into Spotify and it still has no damn clue what I enjoy, and anytime I listen to a song more than a few times, Spotify spends the next 4 weeks trying to cram songs with the same "tags" down my throat.

I had to spend three weeks clicking the "Don't like this" button on EVERY song in my discover weekly before the algorithm figured out "hey, maybe this guy from rural america doesn't actually like songs that are exclusively japanese meme songs"

Like everything in ML, Spotify recommendation works great if you have average or unspecial tastes, because it's just regurgitating the average of your cohort back at you.