Netflix’s New Recommender AI, Explained Simply

Netflix just published a research paper on a new way of doing recommendations. They took one open-source AI model, Qwen3, fine-tuned it on their own data, and it can recommend a show before anyone has watched it. In their tests it did 77% better on brand new titles than the model Netflix had before. In this video: → How recommendations evolved, from "bought together" to embeddings and two-tower models → The cold start problem, and why new shows start way behind → How Netflix gives every show a 4-part code, like a library call number → How the model reads your viewing history like a sentence and writes the next show → The experiment: writing the code one part at a time vs all at once → The results: 77% better on new titles, 22% better at answering questions, 73% cheaper to train → Where recommendations are going next Source: Netflix's paper "Personalized LLM Recommendation with Semantic IDs: Autoregressive Beam Search vs. Multi-Token Prediction at Industry Scale" (arXiv, October 2026). Chapters 0:00 Netflix's new AI recommender 1:00 What Netflix built 1:21 How recommendations evolved 2:35 The cold start problem 3:01 How Netflix does it, step by step 4:42 One part at a time vs all at once 5:30 How they tested it and the results 6:01 Where this is going 6:46 Wrap-up Links Netflix paper: https://arxiv.org/abs/2610.15864 Open-source project, semantic IDs on a smaller catalog: https://github.com/eugeneyan/semantic-ids-llm All opinions are my own and do not belong to my employer. #Netflix #AI #RecommendationSystems