Netflix disclosed how it integrates its Foundation Model for personalization into production systems, describing three approaches: embeddings, subgraph, and fine-tuning1. The company pre-trains the model from scratch monthly and fine-tunes it daily on the latest data. Netflix built an Embedding Store for serving profile and item embeddings and is developing near-real-time embedding inference to reduce staleness. A fine-tuning framework allows application teams to create custom versions of the model. The company is also working on a distilled version to cut subgraph inference latency.
Netflix Details Three Integration Paths for Its Foundation Model
Netflix disclosed three approaches for integrating its personalization Foundation Model into production: embeddings, subgraph, and fine-tuning, with…