I'm not listening to my Spotify top 100 songs of 2021 right now because if I were, I wouldn't be writing, I'd be dancing. After all, my top moods are energy and bold.*

Despite my top tracks' effects on newsletter productivity, I'm still a big fan of the streaming service and its annual Spotify Wrapped content package. I absolutely love seeing my friends' tastes and what I may have missed over the past year. Evidently the whole thing was developed by an intern, whose design process and impact is recapped here in Refinery29.

The other great part of Wrapped day: seeing how Spotify's machine learning has evolved over time to more accurately represent the music and podcasts I enjoy and not the "Three-hour White Noise for Baby Sleep" track that technically occupies more minutes of listening than any other song.

In terms of content engineering and personalization, Wrapped is extremely impressive. Like Netflix's early algorithms, it demonstrates what companies can actually do in terms of content personalization and recommendations. The amount of shares and general joy that accompanies every release of Wrapped reminds me why I like the internet in the first place.

Some notes on this year's Wrapped:

  • As always, the records released earliest in the year occupy the highest positions in my Wrapped list. My top track ("Ain't Nice" by the Viagra Boys, aka a grown man singing an Oscar the Grouch anthem) was on an album released on January 8, 2021. Number two, "Vendetta" by Iceage, was released in February. In the deep dark of a pandemic Minnesota winter, I like repeat streams of angry yet dancey Scando bands, I guess.

    Meanwhile none of the songs from Halsey's August album, which I've been mid-level obsessed with since its release, didn't make it into the top 5 despite recent heavy rotation. In a dream world, Spotify would weight release date somehow so the early birds aren't as prevalent.
  • The aura feature was a nice touch this year, although the machine learning produced hilariously mixed results. I'm curious to understand how Spotify's music-to-mo0d correlation dataset was developed, and who produced the initial set of descriptors to begin with.
  • The genre notes are similarly a little awkward, particularly when we have no access to our genre datasets to begin with. Evidently I listened to 141 genres, but my top five makes my taste look a bit one-note.
These are not five different genres.


My partner Will's top five demonstrates far more variety in his taste, but the way the ML produced the graphic makes me laugh. How can you tell a human designer didn't produce this graphic? Squish squish.

Notice any weird machine learning tics in your Wrapped package? I'd be stoked to see 'em.

*Before you get to critiquing Spotify's shitty practices for artists, and how can I support independent music and like Spotify at the same time, I also purchase records and attend concerts regularly (in non-pandemic years). I listen to Bandcamp too, admittedly not as much as I should. But the Wrapped playlist annually reminds me of the records I loved and don't have in my collection yet, so I've a list prepped for December Record Store Treats.

Also, FFS, let people enjoy the fact that they like and want to share art and culture.


How to diversify your Spotify mixes | The Content Technologist
Do you want to listen to different kinds of artists on Spotify and simultaneously raise a middle finger to systemic racism? Here’s how.
How do keyword research tools generate their data?
How do Google’s keyword planner and other keyword research tools generate their data? Here’s what I’ve learned in the past decade of SEO research.
Content personalization & personal taste | The Content Technologist
When actually engineering and optimizing content for personalization, two major challenges arise: the Inventory problem and the Interpol problem.