Apple bets on human DJs. I'd bet on the skip button
Apple Music launched with Beats 1 and human curators. Three weeks later, Spotify ships Discover Weekly. Two theories of music discovery.
Apple Music launched at the end of June with a big idea about discovery: people. Beats 1 is a live radio station with real DJs, led by Zane Lowe, and the “For You” section leans on playlists made by editors and music experts. Apple’s message is that algorithms can’t understand music the way humans do.
This week Spotify started rolling out Discover Weekly, a playlist of 30 songs generated for each user every Monday. No human picks them. My first one showed up on Monday and it’s surprisingly good. I saved about half of it.
My understanding of how Discover Weekly works, from what Spotify has said publicly and from the Echo Nest people they acquired last year, is that it mostly doesn’t look at the music. It looks at playlists. Hundreds of millions of user-made playlists are a huge record of people saying “these songs go together.” If lots of people who put song A on a playlist also put song B on it, the two are related, whatever genre labels say. You represent every song and every user as a vector learned from those co-occurrences, and recommend songs whose vectors are close to yours that you haven’t played yet.
What Apple’s curators know about music is real, but it’s one kind of knowledge. What millions of listeners do is another kind, and there’s far more of it. The strongest signal in all of this might be the least glamorous one: the skip. If you skip a song after five seconds, that tells the system something very honest, without you having to explain anything. You didn’t want it, at least not right now.
Human curation also has a scaling problem. A great editor can make a great playlist for a mood or a genre, but it’s the same playlist for everyone who clicks on it. Discover Weekly is a different playlist for each person, and the system learns from every skip and save it gets. Humans are good at the starting point and the framing. Behavior data is good at the long tail and at knowing you in particular.
I don’t think it’s either-or. The best version probably uses editors to set the taste and the context, like “songs for late-night driving,” and behavior to decide which of those songs to play for you. But if I had to pick one to bet on for the next ten years, I’d bet on the data. There’s a feedback loop there, and a human editor doesn’t get one.