2:25 PM·Track 3 · Room 2003
Most AI-powered search has a synthesis problem: it uses LLMs to summarize sources, which by design obfuscates the very thing users came for — the creator, the source, the human voice. At YouTube, we re-architected Search around a different bet: the best AI search doesn't replace the creator source, it amplifies it. This is the story of building YouTube's AI Search — a video-native, conversational experience that stitches the best creator moments together with LLM-generated narrative, instead of flattening videos into text summaries. I'll share: — Why video-native AI search was inevitable, and what that means for builders outside YouTube — The "sensory gap" of text-only LLMs, and why closing it changed our retrieval, ranking, and UX — The contrarian product principles behind the architecture: bet on the model over rules, focus on intent over facts, present collective intelligence over a single answer — How we evaluate helpfulness against traditional Search at YouTube scale, and the failure modes we're still wrestling with — What I'd ask differently if I were starting over For anyone building search, RAG, or any system that has to honor its sources while still feeling magical.
Search & Retrievalintermediatetalk
