Building Generative Content Understanding Systems
summary
A deployed generative recommender called Tubi FM generative recommender uses LLM base models, direct token injections, semantic IDs via RQ-VAE, and nightly incremental retraining (about a day on H100s) to score the whole inventory end to end. The uploader claims 12× impact versus a median experiment and search improvements that beat three years of prior experiments, with faster latency, instant few‑shot behavior from recent user actions, and measurable gains across retrieval, ranking, and short‑query search.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:20LLMs recommend Tubi content
- 3:45Generative recommenders language
- 6:56Multimodal user behavior modeling
- 9:25User story tokens
- 11:31Training with user histories
- 13:37Semantic IDs for recommendations
- 15:40Curated video inventories
- 17:25Results and impact
- 19:51Few-shot user adaptation
- 21:57Query intent prediction
- 24:50Multilingual search improvements
- 25:54Multitask learning expansion
- 27:39Ensembling directions
- 28:41Acknowledgments
related talks
23mML-Powered Search & Recommendation 101
25mBuilding a recommendation system for EPAM systems
21mRecsys Keynote: Improving Recommendation Systems & Search in the Age of LLMs
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
38mEmbeddings! Embeddings everywhere!
43mRecommendation Systems in Market Research
18mBuilding an Incrementally Trained, Local Taste Aware, Global Deep Learned Recommender System Model
27mEnhancing Personalization Using Recommendation Systems & Multi-Armed Bandits