Teaching LLMs to Speak Spotify — Yves Raimond & Jacqueline Wood, Spotify
Spotify rebuilt its recommendation system as LLM-centric 'generative personalization' at 760M user scale
“moving from personalization as guesswork...to personalization as reasoning, which is able to analyze those results and really understand whether they are relevant to that user in a given context”
Spotify's ML leadership describes a fundamental architectural shift from traditional ranking algorithms to LLM-centric 'generative personalization' across 760M monthly active users and 100M+ content items. The new system moves from static ranked lists to dynamic, reasoning-based experiences that interactively adapt per user. This represents one of the largest-scale production deployments of LLMs as a recommendation backbone, replacing multi-stage ranking pipelines with transparent, LLM-driven personalization.