Intel / video

Teaching LLMs to Speak Spotify — Yves Raimond & Jacqueline Wood, Spotify

Source: youtube.com
Author: AI Engineer
Date: Published Sep 25, 2026

Context

Spotify presenters Yves Raimond and Jacqueline Wood describe a way to teach a language model both ordinary language and identifiers for items in Spotify's catalog, so it can connect a listener's request to music, podcasts and audiobooks. According to the talk's description, the training recipe, which Spotify calls NEO, adds catalog identifiers to an language model, grounds the new identifiers while the original model stays frozen so its language ability is kept (the description says continued erased it), and then tunes the model across many Spotify tasks.

The description presents a steerable DJ, prompted playlists and an editable taste profile as the listener-facing side of what Spotify calls generative personalization, and says one in four US Premium subscribers use a single system it calls the Large Taste Model every day. It does not say in its text how each feature uses the trained model, and its evaluation figures are the presenters' own.

Key takeaways · AI-distilled

Spotify's NEO recipe adds semantic IDs for catalog items to an open-weight LLM such as Qwen, then grounds those new tokens with the backbone frozen; the team says continued pre-training instead wiped out the model's language ability.
After , Spotify instruction-tunes the model across many Spotify tasks at once, and Jacqueline Wood reports that this multitask tuning even helped cold-start audiobook recommendations.
Spotify reports that 98% of generated semantic IDs come out valid even without , a useful data point for teams weighing constrained decoding against plain beam search.
Spotify's generative personalization spans a steerable DJ, prompted playlists, an editable taste profile and personal podcasts, and one in four US Premium subscribers now uses the system daily.

Terms in this piece · Glossary

open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
structured output — Forcing a model's response to match a schema, so downstream code can parse it instead of guessing at prose.

Read the source www.youtube.com