YouTube-style semantic IDs tackle recommender memory walls with dual-purpose tokens
_reachsumit · x · 2026-07-29
YouTube proposes Dual-purpose Semantic IDs to ease the memory-wall bottleneck in large-scale recommender systems by using discrete tokens for both identity and content reconstruction.
- Hierarchical quantization compresses continuous embeddings into semantic IDs.
- The IDs serve two roles: collaborative identity and content reconstruction.
- A lightweight decoder reconstructs embeddings on demand instead of storing massive dense vectors.
- The approach has been deployed in production at a major video platform for ranking and retrieval.
More from Infra
- Google posts first quarter of negative free cash flow after years of growth — michalmalewicz · 2026-07-29
- South Korea’s AI-linked stocks sink 10.84% as Samsung and SK Hynix plunge — emmanuelvivier · 2026-07-29
- OpenAI launches Presence for real-time voice agents and enterprise chatbots — emmanuelvivier · 2026-07-29
- French Startup ZML Releases Free Inference Server Compatible Across AI Chips — emmanuelvivier · 2026-07-29
- AI Chip Startup SambaNova Raises $1B at $11B Valuation — emmanuelvivier · 2026-07-29
- Microsoft is building internal AI models to cut OpenAI dependence and costs by up to 89% — emmanuelvivier · 2026-07-29