Netflix's LLM ranker beats production with 40x less data
Two algorithm stories today — one about how AI is quietly replacing the most mature machine-learning stack in the world, the other about what an open-source algorithm reveals about election season.
Netflix says its LLM-backed recommender, GenRec, beat the streaming service's heavily tuned production ranker in a live A/B test while using roughly 40x fewer labeled training examples. The system, detailed in a Netflix Tech Blog post and accompanying paper, post-trains an internal foundation model on Netflix data and objectives, then verbalizes each member's watch history, context, and item metadata into natural language instead of hand-crafted features — a shift the authors describe as moving from feature engineering to "context engineering." A catalog-aware scoring head ranks titles, and the model runs on Netflix's vLLM serving stack in prefill-only mode, scoring the whole catalog in a single pass with no token-by-token decoding.
The results are the notable part. Offline, GenRec improved Mean Reciprocal Rank by about 1.6 percent against the production baseline while using a small fraction of the Phase-2 labels; online, a roughly four-week A/B test covering about 10 percent of Netflix traffic delivered statistically significant gains on both short-term and long-term metrics. Netflix also found the Phase-1 foundation adaptation is worth a 10–20 percent offline gain over an off-the-shelf LLM, and that trimming context to about one-third of the original budget barely hurts quality while cutting serving cost proportionally. Recommendation has long been the last stronghold of bespoke MLPs, two-tower models, and thousands of engineered features; GenRec is one of the clearest signs yet that the LLM-native stack — shared backbone, prompts as features, scaling laws as design guides — is coming for it.
X's open-source ranking algorithm reveals exactly which Brazilian candidates and accounts get filtered out of For You feeds during the 2026 election. The repo xAI published last week includes a dedicated filter that removes accounts reported to Brazil's Electoral Court (TSE) from recommendation results, per the country's election rules — a list of obfuscated user IDs with usernames included "for transparency," covering dozens of politicians from both major parties, and a comment noting at least one account was deleted before the code shipped. It follows the open-sourcing of X's ranking engine, which we covered earlier this week — X open-sources its ranking algorithm, adds shadowban checker.
The filter is a legal compliance mechanism, not a moderation preference: Brazilian resolution requires platforms using recommendation systems to exclude channels and profiles reported to the Electoral Court, and the code even links to the TSE's open candidate dataset. But the transparency payoff is new — anyone can now audit exactly how a major platform's algorithm enforces electoral law, who is on the list, and how the suppression is implemented. It's the strongest concrete example yet of the promise behind open-sourcing a feed algorithm: election-time content governance, previously a black box, now readable in source code.
What to watch: whether Netflix's results push other streaming and shopping platforms to disclose their own LLM ranker experiments — and whether X publishes the full, unredacted version of the Brazil list.
If LLMs can beat a decades-old recommendation stack on a fraction of the data, how long before every "you might also like" row is generated by a foundation model? Tell us in the comments.
