Personalization in the Era of LLMs
summary
Spotify claims personalization at scale by projecting 750 million users into user embeddings, compressing a 100M+ catalog into semantic IDs as six tokens, and fine-tuning open‑weight LLMs so the model autoregressively predicts the next song or episode as the next token. A soft tokenization layer maps user vectors into the LLM token space to produce a user‑specific token, and podcast next‑episode recommendations run in production on this stack.
AI-written summary · about
chapters (6)
- 0:00Personalization at Spotify
- 1:59User modeling explained
- 8:42Traditional vs transformer models
- 11:32Semantic IDs for catalog
- 16:24Steerable recommender system
- 19:33Q&A
related talks
30mLLM-Based Recommendation Systems
27mPlay
40mPowering Spotify's Audio Personalization Platform
27mWhich Is To Be Master? Understanding LLM Tokenization
31mComputers That “Get It”: Using LLMs to Give Software Semantic Understanding
37mExpanding the Capabilities of LLMs Through Community Collaboration- 37mExpanding the Capabilities of LLMs Through Community Collaboration
30mWhere there's a will, there's a way: The mad genius of LLMs as a classifiers