# One Model to Rule Netflix Recommendations

Yesu Feng, Netflix | AI Engineer World's Fair 2025 | 22:28

Source: https://www.youtube.com/watch?v=AbZ4IYGbfpQ
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/one-model-to-rule-netflix-recommendations
Published: 2025-07-16
Tags: distillation, embeddings, fine-tuning, latency

## TL;DR
- Netflix is consolidating user representation and content representation in a transformer-based foundation model that supports many recommendation applications.
- Recommendation models need rich event tokenization, semantic content embeddings, and multiple training targets because user actions contain more information than a language token.
- The foundation model has improved application experiments and reduced duplicated infrastructure, while Netflix continues to explore semantic IDs, generative retrieval, and prompt tuning.

## Summary
Yesu Feng describes Netflix's plan to use one transformer-based foundation model across its recommendation systems. Netflix has many surfaces, including the main homepage, search, the kids homepage, and mobile feeds. It also recommends different content types, such as movies, shows, games, and live streams. Previously, separate models duplicated feature and label engineering. The foundation model centralizes learning from user interaction histories and provides representations that downstream models can reuse. Feng explains how Netflix tokenizes events, combines ID embeddings with semantic content information, and trains against several facets of behavior. The team has scaled from models with millions of parameters to roughly one billion, while managing recommendation latency through distillation and other serving options. Feng reports gains in application A/B tests and infrastructure consolidation. Future work includes a shared representation for different entity types, generative retrieval for collections, and prompt tuning for faster adaptation.

## Key ideas
### Netflix's many recommendation surfaces created duplicated models and features
[00:03](https://www.youtube.com/watch?v=AbZ4IYGbfpQ&t=3s)
Netflix recommendations vary across rows, content types, and pages. Rows can focus on genres, new releases, or Netflix-only titles. The content set includes movies, TV shows, games, and live streams. The product also has different surfaces, such as the main homepage, search, the kids homepage, and a linear mobile feed. These needs led to many independently built models for ranking videos, ranking rows, finding unseen shows, and ranking titles a member already engages with. The models often reused the same interaction facts through slightly different counts, time windows, slices, and sequence features. Feng says this duplication made the system hard to maintain and slowed new model development.

### A shared foundation model can centralize user representation learning
[03:18](https://www.youtube.com/watch?v=AbZ4IYGbfpQ&t=198s)
Around four years before the talk, Netflix asked whether it could centralize the learning of user representation. Feng's hypothesis was that a transformer-based foundation model could improve personalization through scale and semi-supervised learning. He also expected a shared model to improve downstream systems together, since applications would reuse the same learned representations instead of training full models from scratch. The proposal borrows from language-model development, including the idea that scaling data and model size can improve quality. The foundation model is intended to provide a common layer beneath application-specific models, while leaving room for fine-tuning and distillation.

### Recommendation tokens need to preserve the many facets of a user event
[05:25](https://www.youtube.com/watch?v=AbZ4IYGbfpQ&t=325s)
A language token is usually one ID, while a recommendation event contains several fields. Netflix must decide how much information to keep about an interaction and how finely to split it into tokens. Feng describes event information in terms of when, where, and what. When includes time encoding. Where includes country, device, page, and row. What includes the title, entity type, interaction type, and duration. Token granularity also affects the available context window. Netflix developed abstractions that allow tokenization to differ between pre-training and application fine-tuning. Those choices can affect model quality throughout the rest of the system.

### Semantic content embeddings address the cold-start problem for new titles
[08:10](https://www.youtube.com/watch?v=AbZ4IYGbfpQ&t=490s)
Netflix combines learned ID embeddings with semantic content information. An ID embedding learned from scratch cannot handle a title the model did not see during training. Semantic information gives the system another way to understand that title at inference time. Feng calls this a common cold-start problem in recommendation, unlike the usual language-model setting. The model's user representation comes from transformer hidden states, and Netflix studies how to keep that representation stable as a member's profile and interaction history change. The team also considers how to aggregate information across sequence positions and transformer layers, and whether downstream objectives should adapt the representation further.

### The training objective can predict many parts of future behavior
[09:51](https://www.youtube.com/watch?v=AbZ4IYGbfpQ&t=591s)
Netflix's objective is richer than predicting one next token. The model can predict entity IDs as well as action type, entity metadata such as type and language, duration, device, and the time of a future play. These signals can be used in multitask learning with multiple heads or hierarchical prediction. They can also act as weights, rewards, or masks on the loss. This lets Netflix direct learning toward particular parts of user behavior. Feng also describes multi-token prediction, which encourages the model to look beyond the next action. Netflix observed a notable metrics improvement from this approach, which is intended to support long-term behavior and reduce sensitivity to the gap between training and serving.

### Scaling improved quality, but recommendation latency limits model size
[11:19](https://www.youtube.com/watch?v=AbZ4IYGbfpQ&t=679s)
Feng reports that Netflix kept seeing gains while scaling over roughly two to two and a half years. The models grew from the order of ten million profiles or a few million parameters to the order of one billion model parameters, with data scaled along with them. He says the team stopped at that point because recommendation systems have strict latency and cost requirements. Further scaling would require distilling the model back into smaller forms for serving. He does not present scaling as finished. The results suggest that the model can continue to improve with scale, while production constraints determine how much of that model can run directly online.

### Downstream applications can consume the foundation model in several ways
[13:51](https://www.youtube.com/watch?v=AbZ4IYGbfpQ&t=831s)
Netflix consolidates much of its personalization data and representation layer, especially user and content representations. Application models become thinner because they build on the foundation model instead of learning every component independently. One option is to insert the foundation model as a subgraph inside a downstream neural network. Another is to use its content and member embeddings through a centralized embedding store, with refresh frequency and representation stability as practical concerns. Teams outside direct personalization can also fetch those embeddings. A third option is to extract and fine-tune the model for a particular application, or distill it when the application's online latency requirement is strict.

### Netflix is extending the approach toward collections and faster adaptation
[17:11](https://www.youtube.com/watch?v=AbZ4IYGbfpQ&t=1031s)
Feng describes three current directions. Netflix wants a universal representation for heterogeneous entities as its content types expand, including semantic IDs. It is also investigating generative retrieval for collection recommendation. Instead of selecting only one video, a model could decode a collection while handling business rules and diversity during generation. The third direction is prompt tuning. Netflix would train soft tokens and swap them at inference time to make the foundation model behave differently for different applications. In questions, Feng adds that graph-model embeddings support cold start, reinforcement learning is relevant when user rewards are sparse, and fine-grained video or clip-level embeddings are not yet incorporated into this system.

## Notable quotes
- "Our big bet at Netflix on personalization, namely to use one foundation model to cover all the recommendation use cases." (00:03)
- "The challenge back then was, is this scalable? Obviously not." (03:31)
- "We need to have semantic content information to be complementary to those ID embeddings." (08:34)
- "We see indeed high leverage of FM to bring about both A/B test swings as well as infrastructure consolidation." (16:28)

## Tools & references mentioned
- Netflix
- Transformer architecture
- DeepSeek
- LinkedIn
- Uber
- knowledge graph
- reinforcement learning

## Who should watch
- You work on recommendation systems with separate models for different surfaces, content types, or ranking tasks and want to evaluate a shared representation layer.
- Your team is building sequential recommendation models and needs guidance on event tokenization, cold start, multi-token prediction, or long-context training.
- You care about serving tradeoffs and want concrete patterns for using a large recommendation model through subgraphs, embedding stores, fine-tuning, or distillation.

## Related talks

- [Personalization in the Era of LLMs](https://aietalks.com/talks/personalization-in-the-era-of-llms) (Shivam Verma, Spotify, 20:12)
- [360Brew: LLM-based Personalized Ranking and Recommendation](https://aietalks.com/talks/360brew-llm-based-personalized-ranking-and-recommendation) (Hamed Firooz & Maziar Sanjabi, LinkedIn AI, 22:00)
- [Improving Recommendation Systems and Search in the Age of LLMs](https://aietalks.com/talks/improving-recommendation-systems-and-search-in-the-age-of-llms) (Eugene Yan, Amazon, 20:54)
- [Building Generative Image & Video Models at Scale](https://aietalks.com/talks/building-generative-image-video-models-at-scale) (Sander Dieleman, Google DeepMind, 40:46)
- [LLM Recommendation Systems (RecSys)](https://aietalks.com/talks/llm-recommendation-systems-recsys) (, 4:06:11)