On why discovery ACTUALLY sucks
"Spotify's algorithm is broken." This is intellectually lazy. I am just as guilty of saying this as the next person, but never once have I sat to unpack the why.
One of the most shocking things I learned working in Silicon Valley is that all companies use the same recommendation ML models: Instagram, Twitter, Tiktok, Spotify, Netflix etc. In fact Twitter/X's algorithm has been open-sourced since 2022. Everyone's take on why Spotify's algorithm is so bad is that "Spotify is just incentivized to keep users on the platform". Ok... and isn't that the same incentive as the other social media apps out there? Yet why do they not get the same hate as spotify's algorithm. Another famous reason you've probably heard is that "100,000 songs are uploaded unto Spotify daily!!!1!" Sounds plausible until you compare against the app with the best algorithm, Tiktok, reports about 34 million uploads daily. Wtf. Either Spotify's engineers are so bad that they fucked up even an open-source algorithm or the real issue isn't actually the algorithm. No comment on the former, but let's dig deeper into the latter.
The Preview Asymmetry
- Images: 50–100ms visual processing → instant preference signal
- Video: 3–5 second hook → attention retention signal
- Music: 30–90 seconds minimum → creates 600–1800x time cost differential
It's an information bandwidth problem. Our brains process visual information 6000x faster than audio. Music requires sequential processing. I'll explain. You jump to the middle of a song, it doesn't tell you that much about it. You jump to the middle of an image, and there it is. Music requires sequential processing. Images activate parallel processing. This forces us into conservative selection strategies (artist name recognition, trusted playlists) creating a closed loop where novelty is low. To this point, even though the ML algorithms are the same, because the data skews familiar, the discovery algorithms are optimizing these existing behaviors rather than creating them.
Simply put, the algorithm is bad because the human listening signal is bad, and the human signal is bad because of the cost to preview. This is actually a core physics constraint and we need to work around it. And no, listening on 2x doesn't count. We have this same issue in Roblox discovery algorithms. We keep staffing amazing ML teams to fix game discovery with little to no returns because fundamentally the preview mechanism for games is arguably even worse than it is for music. Side note on the human behavior signal misleading the algo: this is a similar reason why NYC girlies on dating apps complain about their matches being sub-par, meanwhile they tend to message lower-tier men a lot when they get "bored" or to stroke the ego a little. Paradoxically teaching the algorithm to show them those types of profiles. Brutal.
A common pushback to this is the volume. We mentioned that briefly above, and even in our dating app example, do you notice how I specifically mentioned NYC girlies? It's not a dig at them specifically. In fact, this is just standard human behavior. But one must ask, why is it that we don't hear about the small town folk not being able to find their partners on dating apps like we do the big city ones? Why is it that nobody complained about the Youtube algorithm or even the Spotify algorithm years ago? They didn't have better algorithms then, but what they had was a smaller exploration space. If you live in a small town, you cycle through all eligible prospects entirely or at least what you believe to be a representative set. If you believe you have a general idea of the entire space then discovery is no longer an issue. As a music listener the pain is in knowing that there is more out there, but not knowing how to get there. I bet if you roughly knew what all the songs on spotify sounded like, you wouldn't care much about finding music.
Here's a line from Donald Passman's All You Need to Know About the Music Business. "Even though we have more artists each year, fewer are making it into the billboard top 100." Amazing quote, Donald, though I must admit I put this book down when I decided to build a music tech startup so I don't become hyper focused on current paradigms. I am here to disrupt it after all. Back to the quote. If it isn't physically bending your mind then your brain is already cooked. The rewards we give music (Billboard, Grammys, Top 40s) have turned this expansive space that is music into a tiny box.
In universe 25, John Calhoun built a "mouse utopia" with unlimited food, water, and an enclosed space. Population exploded at first, then interestingly: even with abundant resources, the rats' social structure collapsed. The population had grown so large that earlier rats eventually claimed all parts of the territory, such that the generation of younger rats who came later had no space and survival became zero sum even with infinite resources. What was the solution? Expand the space so that the young can stake their claim on new territories and thrive there. Similar to if you expand the space of the NYC fuckbois and spread them around america then you would no longer have a discovery problem, you would simply just... go to the city that has your type. Music awards/recognition has created this universe 25 experience which then enforces the "rich-get-richer" phenomenon, reinforced by the algorithm; leading to a discovery death spiral. The game is cooked frankly, and there's no point trying to fix it. Instead let's play a new one: humans become the new algorithm.
Our goal is to match users together using taste embeddings from their listening history and let users be the discovery vectors to one another. We plan to change music listening from passive to active by moving away from the concept of algorithm feeding you recommendations to actively allows users navigate the sound space and see songs that listeners who listen like them like. We are starting with our wedge of DJs given that DJs are the power users of discovery and have skin in the game and actively need to curate to create their sets for micro niches.