Europe 2026
Personalization in the Era of LLMs - Shivam Verma, Spotify
Overview
This talk explores Spotify's approach to personalization in the age of Large Language Models (LLMs), moving beyond traditional recommendation systems. It details how Spotify builds foundational user and content models, integrating them into a steerable, personalized generative system. The core thesis is that by leveraging LLMs and advanced modeling techniques, Spotify can create more relevant and user-controlled recommendations.
Who should watch
- AI engineers working on recommendation systems
- Product Managers seeking to understand LLM applications in personalization
- Builders interested in user modeling and content understanding for large-scale platforms
- Developers exploring the integration of LLMs into existing product stacks
Key takeaways
- Spotify is transitioning from traditional multi-step recommender pipelines to a unified, LLM-backed generative model for personalization.
- Foundational user modeling involves creating vector embeddings that represent a user's taste across their entire history on the platform.
- Content understanding is achieved by fine-tuning open-weight LLMs and embedding Spotify's catalog data using a technique called semantic IDs, which tokenizes content vectors.
- Semantic IDs compress high-dimensional content vectors into a few tokens, enabling LLMs to auto-regressively predict the next item a user might enjoy.
- The system combines user embeddings and catalog understanding through a soft tokenization approach, projecting user representations into the LLM's token space for personalized recommendations.
- Features like the AI DJ and prompted playlists demonstrate the move towards more interactive and user-steerable recommendation experiences.
- The "Taste Profile" feature allows users to directly influence their recommendations by providing feedback, further enhancing personalization.
- This new generative approach aims to provide more steerable, personalized, and explainable recommendations compared to older methods.
Notable quotes
*We're kind of moving away from that siloed model of working towards this single unified model which is similar to how LLMs work.*
*The idea is that we want to expose what we know about you. And then we want to kind of let you choose which part of that you want us to kind of keep.*
Unofficial community note. Prefer the recording for nuance.