Netflix replaces hand-crafted features with LLMs for better recommendations

netflix implemented a new recommendation system called genrec that converts user watch history into natural language prompts for a 10-billion-parameter open-weight model. unlike the existing system that relies on thousands of hand-crafted features, genrec uses this text-based approach to score titles in a single pass via vllm, preventing hallucinations by only scoring real catalog entries. offline tests showed a 1.6% gain in ranking quality, and the system achieved these results with 40 times less labeled training data than traditional methods. the success of this experiment suggests that large language models can effectively replace custom-built architectures for recommendation tasks.