Gaming the algorithm
But this time, for the fans.
I lamented the state of internet music discovery yesterday, but I often wonder if there are ways to force “the algorithm” to work for me on these music apps. Can I treat the passive recommendation algorithm as an active search tool? Here’s one experiment I’ve been running.
Spotify makes every customer “Daily Mixes.” There are 5 of them. They are not always easy to find. If you click on “Search” and then click the box “Made For You” and then scroll down to the bottom, you’ll probably find them there. Spotify doesn’t tell you what these mixes are nor what they intended them to be. I’m sure it was the project of some recommendation engineer and will probably be deprecated. But I’ve enjoyed them in the car when I just wanted a stream of a familiar mood.
As far as I can tell, these are mostly songs I’ve already listened to, and then a sprinkling in of some other tracks that must match in some fancy machine learning model (is it cosine difference on LLAMA MAMA embeddings?). The new songs they add are occasionally great, but more skippable than not. But isn’t that always true when you are searching for new music? The songs that grab you are rare. They should be rare. I wondered if I could use these mixes to distill out some actually new directed suggestions. Could I force Spotify to help me explore a genre?
I had been listening to a lot of ambient music on Spotify, not only because I love ambient music, but because it’s some of my favorite music for working. Just as Brian Eno intended it. One day, I noticed a Daily Mix that was full of ambient and tried an experiment. I made a playlist and copied the Daily Mix into it. The next day, there was another ambient Daily Mix, I added this to my custom playlist too.
I repeated this process for the next several days. Find the ambient Daily Mix. Copy into my custom playlist. Remove the duplicates. Delete any songs that weren’t ambient enough. Repeat.
The list didn’t grow very quickly. There were always a lot of duplicates because these mixes were (I think) intentionally mostly comprised of songs I’d already listened to. I did this process yesterday. The Daily Mix had 50 songs. Of those songs, only nine were not already in the custom list. Of the nine new ones, five were on albums already in the playlist. Four were by new artists. But I was getting new suggestions every day. A couple of new ambient artists a day isn’t bad, right?
I ended up with 8 hours of AI Generated Ambient:
Pick a random starting song and get blissed out.
This is a decent playlist featuring some stellar artists. I’m a huge fan of Tomasz Bednarczyk who prolifically creates sleepy synth drones. The playlist pulls several tracks from Ametsub’s 2009 album “Nothings of the North,” a favorite of mine. The mix features ambient greats like Fennesz, Federico Durand, Tim Hecker, and Willamette. And it features a bunch of great artists I’ve never heard of before. Yesterday it led me to Patricia Wolf’s 2022 album “See Through.” It’s pretty good!
Content creators have been scheming to game “the algorithm” for a decade. I’m never sure if they are hallucinating conspiracy theories or finding legitimate means of juking engagement. But as a content consumer, maybe I can game the algorithm too. Which other ways can I take a recommendation architecture designed for brain-dead revealed preferences and force it to help me explore?
Thanks for reading arg min substack! Subscribe now and we’ll game the algorithm together.