Netflix's Big Bet: One model to rule recommendations: Yesu Feng, Netflix

Netflix centralizes personalization efforts by transitioning from numerous specialized models to a unified foundation model based on transformer architecture. This shift addresses inefficiencies in label and feature engineering while enabling scalable, high-leverage innovation across diverse content types, including games and live streaming. By applying scaling laws similar to those in large language models, the system improves user representation through multi-token prediction and multi-layer supervision. The foundation model serves downstream applications via subgraph integration, embedding lookups, and fine-tuning, significantly increasing development velocity. Future directions include universal representations for heterogeneous entities, generative retrieval for collection recommendations, and prompt tuning for faster model adaptation. This approach optimizes infrastructure while maintaining strict latency requirements, ensuring consistent performance across Netflix’s expanding content landscape.