They took all user generated playlists and projected the songs into vectors where songs that appear together on playlists are closer and songs that appear less often are farther.
It’s likely changed a lot since then, but it seemed like a pretty straightforward clustering system at the time.
This is the same way YT/TikTok does it btw. Co-occurrence is king in recommender systems in production. It's extremely cheap to calculate and by far the most effective method.
This is not really important if you have a lot of user behavior data and/or playlists for each song. But if you have a niche song that few people of listened to, collaborative filtering based recommendations aren't going to be good.
Real semantic embeddings (which can then be part of the input to the recommendation model) can be trained using self-supervision, e.g. an auto encoder or a seperate "next audio token" predicting transformer.
Or, more bluntly: you aren't going to mate with a For You page, so it doesn't have the same evolutionary cheat code to your preferences as other people have.