The Funnel: From Millions to One

Every recommender system that works at scale has the same shape: a funnel. Joey explores YouTube's 2016 paper on recommendations—the paper that switched YouTube from optimizing for clicks to watch time—and breaks down the funnel architecture that lets systems rank millions of items in milliseconds. This is episode one of Learning Recommender Systems, a series on how recommendation engines actually function. 00:00:00 - YouTube's 2016 recommendation paper 00:01:32 - The core problem: ranking millions in milliseconds 00:03:03 - The funnel structure 00:07:45 - What recommenders learn from 00:11:15 - Walking through a user's complete journey 00:13:01 - How scrolls become tomorrow's training data --- Sources & further reading: • Paul Covington, Jay Adams, Emre Sargin, Deep Neural Networks for YouTube Recommendations, RecSys '16, September 2016, DOI 10.1145/2959100.2959190 — — funnel, latency, surrogate problem, clickbait, watch-time weighting, Example Age, offline vs A/B (quoted from the Google PDF; ACM page 403).: https://research.google.com/pubs/archive/45530.pdf • Ivan Medvedev, Haotian Wu, Taylor Gordon, Powered by AI: Instagram's Explore recommender system, Meta AI blog, 2019-11-25 — — 500 → 150 → 50 → 25, distillation, value model, same-author penalty.: https://ai.meta.com/blog/powered-by-ai-instagrams-explore-recommender-system/ • Vladislav Vorotilov, Ilnur Shugaepov, Scaling the Instagram Explore recommendations system, Engineering at Meta, 2023-08-09 — — four stages, top-100 second stage, final reranking rules.: https://engineering.fb.com/2023/08/09/ml-applications/scaling-instagram-explore-recommendations-system/ • Eugene Yan, System Design for Recommendations and Search, eugeneyan.com, June 2021 (updated 2022-04-14) — — offline/online × retrieval/ranking; "fast—but coarse" vs "slower—but more precise".: https://eugeneyan.com/writing/system-design-for-discovery/ • Google for Developers, Recommendation systems overview, last updated 2025-08-25 — — candidate generation / scoring / re-ranking; the "billions" figure.: https://developers.google.com/machine-learning/recommendation/overview/types • Yifan Hu, Yehuda Koren, Chris Volinsky, Collaborative Filtering for Implicit Feedback Datasets, IEEE ICDM 2008 (venue per citing references) — — explicit vs implicit; four properties of implicit feedback.: http://yifanhu.net/PUB/cf.pdf • All URLs fetched and quoted from on 2026-10-15. • [primary, abstract only] Xing Yi, Liangjie Hong, Erheng Zhong, Nathan Liu, Suju Rajan, Beyond Clicks: Dwell Time for Personalization, RecSys 2014 — — dwell as relevance proxy; normalization across devices.: https://recsys.acm.org/recsys14/session-metrics-evaluation/ • [primary, abstract only] Google Research (ten authors), Recommending What Video to Watch Next: A Multitask Ranking System, RecSys 2019 — — multi-objective ranking; competing objectives and selection bias. Full paper not fetched.: https://research.google/pubs/recommending-what-video-to-watch-next-a-multitask-ranking-system/ This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.