Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.
- 140
- 2
Jump in the discussion.
No email address required.
Notes -
DJ Claudify
I've been meaning to try this for ages, but kept being too lazy to bother. There's a website called Exportify that dumps your saved Spotify playlists into a csv. The setup is painless: approve access, no suspicious permissions, and you walk away with every track you've ever saved alongside metadata like timestamps and Spotify's own popularity scores.
I exported the lot. I don't really do dedicated playlists; my method is to encounter a song I like, save it on the spot, or, on rare occasions, find something decent on the recommended page. The result is a library in strict chronological order, where almost every track carries temporal and emotional baggage. Here's the song I was listening to during a breakup. That's the one that was playing when I first stepped off the plane in Scotland. Consistency and curation are for try-hards; if I don't like what's playing I queue something better, or jump down the list.
I fed the csv to Opus 4.7 and ChatGPT 5.5 Thinking and asked them to chart my mood and vibes over time. Both did remarkably well, well past reading tea leaves. Then I asked for recommendations, ideally things I hadn't heard before.
From Claude:
"176 tracks in your library have a Spotify popularity score of zero. 244 sit below 10. About a quarter of your library is essentially invisible to the algorithm. Names like Chikoi The Maid, Holy Fawn, Suave Punk, Caleb Bryant, Mokadelic show up there. Meanwhile your highest-popularity saves are exactly the global hits everyone agrees on (Blinding Lights, I Wanna Be Yours, Espresso, deja vu, Hey Ya, 505). That bimodal shape tells me you save the songs everyone agrees on and you save the songs almost no one knows, with a thinner middle than most people have."
Yeah Boi. I can out-hipster the hipsters, I listen to songs and genres you people have never even heard of. All while not predicating my personality around it, because I do listen to normal, human music. Sabrina Carpenter? You've heard of her.
Both delivered. Claude was a touch better, with something like a 50% hit rate by recommendation-to-save ratio. Better than any human I know. Better than Spotify's own recommendations, which is the part that surprised me. Spotify has spent over a decade and presumably a small country's GDP building a hybrid recommender that fuses collaborative filtering, raw audio analysis via convolutional nets, and NLP over reviews and metadata. They have every play, skip, save, and 30-second-bailout from half a billion users.
In other words, they try pretty hard.
And yet a model that has never heard a single waveform is beating them on my library, presumably because it's drawing on the entire written corpus of human music criticism: every RateYourMusic thread, every Pitchfork review, every breathless Reddit comment about a B-side. Collaborative filtering also has a pull toward the popular middle, which an LLM doesn't share, since it isn't trained on user-item interaction logs at all.
Easy chronology probably helps too. Spotify mostly sees a bag of vectors with timestamps; a language model can read your library as a story and notice when the vibe shifts.
You should try it. It helps to give honest, immediate feedback on what works and what doesn't. A sampler:
"Avril sounds like something they'd play in an aquarium that sprang a leak. It's definitely for someone, I'm not sure if that someone is me."
"Blink sounds like elevator music on ketamine. A maybe?"
"Loretta fucked my ear canal and got it pregnant. Instant add."
"Mona Lisa can stay in the museum."
"Vapour has a lot going on. Someone else can get it on, fucking hell, there are notes in there that are best appreciated by a cat."
"Catamaran sounds like it was recorded inside a toilet. From the adjacent stall. Miss."
"Mountains isn't bad, but it's for a very moody teen girl, or a middle aged woman going through a divorce/mid life crisis. Close but no cigar. Sigh. I'll add it anyway."
The whole experience was unreasonably pleasant, especially given that I'm not even sure Claude can ingest audio, or was meaningfully trained on music samples. The rest of you are getting mogged by a blind deaf entity that lives in a computer. I added something like 20 songs over the hour, which is a ridiculous number against my usual rate of about one a week.
I don't think they do. They certainly could but frankly their recommendation algorithms have always been utter garbage and they've gone out of their way to make it impossible for users to even manually help them. For example I can't say "Never recommend this artist for this playlist" nor can I say "Never recommend this album or song for anything (but do recommend other songs from this artist)". Their recommendations to playlists keep repeating the same small number of songs over and over and can't even distinguish multiple identical versions (so a playlist that has song X from the original album will get recommendations for song X from a compilation album).
This runs in parallel with them having a massive team to work on the Spotify client with the result that it has only degraded with time since people need to justfify their existence while the real development needs could be filled by just a couple of small teams.
I find the "discover weekly" recommendations not bad, and "release radar " sometimes has something I want to hear. Totally agree that radio is dogshit though.
Until recently the shuffle algorithm was also totally fucked and would get stuck in loops - literally play e.g. 20 songs, then go back and play those same 20 again. I don't know what the fuck kind of shuffle algorithm could result in that, but it rinsed several tracks so bad I can't listen to them anymore.
Now they are trying to push Spotify as a video platform so sometimes I open the app and see some ridiculous TDS thumbnail. I'd probably jump ship if there was another platform with good music recommendation algorithms. I don't know what the eight thousand Spotify employees are doing all day, but I'm not sure it's making my life better.
There is a new function that will go back through your whole listening history and find songs that you listened to a lot at one point, and then stopped playing. I think it just heavily weights skips as "I don't actually like this song" rather than "Not in the mood just now, but there's a reason it's on the playlist."
I just want a braindead "roll the dice" shuffle. Yeah, it might result in some tracks playing twice in a surprisingly short window, but you can do a lot worse (repeating sequences of 20 tracks) and you can't easily do a lot better in my view.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link