@jwz As someone that made a career out of building recommendation algorithms, this completely tracks. When I started I naïvely thought discovery was important. Even the PMs talk about excitement and discovery. But the metrics are always a time spent or something analogous. Which isn’t wrong per se, but reinforces the problem with similarity based recommendations, you play it safe and recommend the same six things over and over again.
Now I look at recommendation algorithms for what they are: checkbox driven requirement no one cares about.
Trust me. More people at Spotify care about their annoying AI drive-time DJ than making your “Daily Mix” playlist have something than same six bands that were popular in 2011.