Netflix replaced its famous algorithm with an LLM
Netflix spent years building its recommendation algorithm into one of the most celebrated engineering achievements in tech. The system was a marvel of feature engineering: thousands of hand-crafted signals about users, titles, and their interactions, all tuned to predict what you would watch next.
It turns out the algorithm may not have been the point.
The GenRec approach
Netflix quietly built GenRec, an in-house language model that does something radically simpler. Instead of encoding user behavior as dense numerical vectors, GenRec converts watch history into plain text. Plays, watch durations, thumbs up or down, list additions -- they become a kind of conversation between the user and the system.
A fine-tuned open-weight language model reads that text and scores all matching titles in a single pass. The model picks up on genre preferences and shifting interests on its own, without a single engineer having to spell them out as a feature.
How it works
Netflix trains GenRec in two stages. First, an unnamed open-weight language model gets fine-tuned on Netflix data so it understands the catalog and user behavior. Then a second, more frequently updated round of training turns that base model into a recommendation ranker that accounts for new titles and shifting preferences.
The key insight: a full text version of every user interaction would blow past any model's context window. So Netflix filters aggressively. High-signal events like long watch sessions stay in full detail. Brief taps or quick scrolls get dropped. Binge sessions get summarized rather than replayed event by event.
The test
Netflix pitted GenRec against its existing recommendation engine in both offline tests and a live A/B experiment with real users. GenRec won on both. The improvement was measurable, and it needed a fraction of the labeled training data the old system required.
Netflix sees this as more than just a better recommendation engine. The company's tech team described it as part of "a broader shift toward general-purpose language models and away from custom-built architectures." For a company whose entire product is recommendations, that statement carries weight.
What this means
The story here is not that an LLM outperformed a bespoke system in a benchmark. That happens regularly now. The story is what it reveals about where value actually lives in AI systems.
Netflix spent years and millions of dollars engineering thousands of recommendation features. GenRec replaced most of them with a single fine-tuning pass on a model that already understood language. The moat was never the algorithm. The moat was the years of watch history, the millions of daily thumbs-up clicks, the granular behavioral data that no competitor has.
That data advantage does not disappear when you stop hand-crafting features. If anything, a language model that reads behavior as text can extract more value from it, because it finds patterns the human engineers never thought to encode.
The skeptical take
GenRec is still early. Netflix itself calls it "an early but promising step." Off-the-shelf language models hallucinate titles, over-index on popular content, and ignore business rules -- GenRec had to be specifically designed to handle those failure modes. The transition away from hand-built architectures will take years, and Netflix is careful not to overclaim.
But the direction is clear. The company that built one of the most sophisticated recommendation systems in existence is now betting that a general-purpose language model, fed the right data in the right format, can do it better. That is not a headline about a benchmark. It is a signal about where the entire industry is heading.