[TeMPOraL]

All of them sucked, though. Turns out, running regression on some features picked as much to help the user as to satisfy business objectives, doesn't lead to a system that can capture one's preferences well.

LLMs have the benefit of understanding (or some values of understanding) how people feel about things in general, in a global sense. They may not have a database of people's choices, but they have a "database" of connotations for every word, how ideas and emotions relate, how interests connect, etc. Instead of relying on a relatively tiny historical record of choices in few, specific (and ill-defined) categories, they can just place user's history in their 10 000 dimensional latent space and use that as a direction to explore, effectively guessing whatever the user's actual preferences are likely to be, without being able to name them or fit into explicit categories.

[MarcelOlsz]

The best recommendation system I encountered was the "similar artists" cloud on what.cd on an artists page. Has never failed me.

https://github.com/WhatCD/Gazelle/blob/master/classes/artist...