Why discovery actually sucks
[Draft stub — Kai-authored framing, replace with your Pages content.]
The reason every music discovery feature ships, plateaus, and dies the same way is not the algorithm. It is the asymmetry between preview cost across media. An image takes 50 to 100 milliseconds to evaluate. A short video takes 3 to 5 seconds. A song takes 30 to 90 seconds, minimum, before you know whether you would put it on at home with someone you love. That is a 600 to 1800x time-cost differential. Every piece of music discovery infrastructure is built on top of that asymmetry without acknowledging it, which is why the systems all converge on the same failure mode: optimizing what you have already proved you tolerate, because exploration is too expensive.
I noticed this first at Roblox, in a completely different domain. The recommendation system worked beautifully on what users had already engaged with, and was structurally incapable of meaningfully widening their world. Same shape: cheap evaluation of the known, expensive evaluation of the unknown, optimization that compounds the asymmetry instead of solving for it. The people who pay attention to this in product circles call it the cold-start problem. It is not a cold-start problem. It is a perception-cost problem, and it is a hardware fact about music as a medium that takes time to unfold.
Once you see it, every existing system reads as a workaround for the asymmetry instead of a solution to it. Spotify Discover Weekly works on collaborative filtering, which sidesteps the preview cost by transferring trust from people who already paid it. Recommendations from a friend work for the same reason: they are a paid-it-already trust transfer. Algorithmic genre buckets are a preview-cost reduction by typecasting. None of these systems make new tracks cheaper to evaluate. They make recommendations cheaper to act on by making the act feel less risky. That is a different operation, and it is why none of them produce the bath-playlist moment.
The bath-playlist moment is what happens when someone you trust hands you music you have never heard, and you put it on, and three songs in you realize the curator understood something about you that you have not explained to them. The cost of that moment is paid by the curator, in advance, when they listened to those tracks one by one with their full attention. They paid the 30 to 90 seconds per track so that you do not have to. The system compresses the preview cost by absorbing it into a relationship that already has trust collateral.
The structural lesson is that the right intervention is not better recommendations. The right intervention is paying the preview cost in a different currency. There are several of these worth building. Mashup-as-discovery is one. Curator-as-trust-transfer is another. Both of them work because they convert the preview cost from your 30 to 90 seconds into somebody else's 30 to 90 seconds, where the other party already had collateral that makes the bet less expensive to take.
This essay is the diagnosis. The next one is what you do about it.