No, they're absolutely garbage at it. I don't even understand the thought to use LLMs in the first place. And even if they weren't garbage the whole point of a music recommendation algorithm is surface music that wouldn't be in the training set so you need a way to recall likely matches at which point you've built a recommen engine.

> I don't even understand the thought to use LLMs in the first place.

You know how people believe whatever they read, hear, and watch even though it might not be true? Well an LLM is something people read and to get over the hurdle of whether something might be true or good, you simply embrace it and ignore that it could ever be wrong. I don’t get it either as I get upset when I find out a source is mostly wrong.

> Are LLMs actually good at music recommendation?

As far as I can tell, the only thing actually any good at music recommendation is (some) humans.

The Apple Music recommendations have been pretty good for me, particularly when autoplaying based on a song/playlist. Which came as a surprise because I always found the Spotify ones to be pretty useless.

Especially with the autoplay functionality on Spotify, after a while it would always throw in some songs that just did not fit in at all with the rest, and it made for a jarring experience. I've never really had that happen with Apple Music.

I’ve tried apples , agree it does less egregious stuff than Spotify but doesn’t really meet the bar for “good” in my experience. This stuff is probably really dependent on what you are looking for I imagine .

Spotify's algorithm has worked very well for me in the past.

I'm pretty sure it's not LLM based though, but rather domain specific, or possibly just a simple recommendation engine ("people who like x also often enjoy y").

Not my experience at all, fwiw.

Yeah but didn’t Spotify kill any community features long ago?

YouTubes algorithm (especially before they nerfed it) is often outstanding at it, pulling undiscovered gems that completely match your tastes.