Scaling Foundation Models for 300M+ Users at Netflix скачать в хорошем качестве

Scaling Foundation Models for 300M+ Users at Netflix

Can one model do it all? Netflix Machine Learning Manager Moumita Bhattacharya reveals how her team moved away from fragmented, "bespoke" ML pipelines to a unified architecture that serves 300M+ users. In this InfoQ video, Moumita introduces UniCoRn (Unified Contextual Ranker) and Netflix’s proprietary User Foundation Model. Learn how they leveraged Transformer architectures - similar to GPT-4 - to move beyond language and master user behavior trajectories. If you are an engineering leader or architect looking to reduce tech debt while increasing personalization lift, this session is a masterclass in ML system design. ⏱️ Video Timestamps (For Navigation) 0:00 - The Challenge: Scaling for 300M+ Users 2:15 - The Two-Stage Ranking Framework 4:45 - Introducing UniCoRn: One Model, Four Use Cases 8:30 - System Considerations: Latency, Throughput, and SLAs 11:10 - Building a User Foundation Model (The "Harry Potter" of ML) 15:45 - Tokenizing User History: Titles vs. Words 19:20 - Results: The Impact of Personalization Magic 23:10 - Addressing Over-Personalization & Filter Bubbles 26:45 - Q&A: Fine-tuning, Cold Starts, and Multi-modal Data #Netflix #MachineLearning #SoftwareArchitecture #LLM #SystemDesign #InfoQ