Teaching Gemini to Speak YouTube: Adapting LLMs for Video Recommendations to 2B+DAU
summary
YouTube rebuilds recommendations by adapting Gemini into a domain-specific checkpoint called LRM that learns a new tokenization for videos — SemanticID — via multimodal embeddings (title, transcript, frames sampled from high-engagement moments) quantized into a few tokens per video. LRM runs generative retrieval in production and experiments on ranking, handles freshness over a 20 billion video corpus, requires continuous pretraining on the order of days/hours, and after engineering cuts achieves 95%+ cost savings at scale.
AI-written summary · about
chapters (16)
- 0:00The role of LLMs in search
- 0:53YouTube recommendations
- 1:47Rethinking the system
- 2:47Large recommender models
- 3:43Tokenizing video content
- 4:34Semantic ID creation
- 5:54Continued pre-training
- 7:20Reasoning across data
- 8:15Generative retrieval
- 8:58Improving recommendations
- 9:49Serving cost challenges
- 10:29Offline inference strategy
- 11:33Challenges in video AI
- 13:37LLM recipe summary
- 15:07Future of recommendation
- 16:53Q&A session
related talks
51mGemini, Google's Large Language Model
19mGemini 2.0: Beyond the Basics
54mGemini, Google's Large Language Model
51mGemini, Google's Large Language Model
26mHow Does LLM and Generative AI Really Work?
56mBuilding Multimodal Apps with Gemini 2 0 and Gemini SDK
01h 01mBuilding Multimodal Apps with Gemini 2.0 and Gemini SDK
19mLLM-Enhanced Multimodal AI: Revolutionizing Audio/Video Interaction