Clown Cast

Joey Musselman
  • 5.0 (1)
  • Technology
  • Updated Daily

Podcasts about whatever I find interesting — history, tech, weird rabbit holes. I was making these for myself anyway, so I figured I'd share. Fully AI-generated: Claude researches and writes each episode, and Piper voices it, one of them trained on my own voice. It all runs on a pipeline I built with Claude Code on an old laptop in my house. Enjoy.

  1. 3h ago

    Melted for Silver: How Formats Decide What Survives

    Seventy-five percent of silent films made before 1930 are gone. Eighty-six percent of pre-1965 recordings have vanished. Not from decay alone, but from deliberate choices: studios melted nitrate film for silver, platforms deleted archives, and only economically-viable content got migrated forward. This episode examines how recording format choice and infrastructure ownership act as invisible filters, determining which parts of human culture survive to the next century. 00:01:32 - Why studios melted movies for silver 00:05:08 - People didn't choose what got recorded 00:10:13 - Format and economics as invisible filters 00:13:38 - A solution that was built but not funded --- Sources & further reading: • David Pierce — "The Survival of American Silent Feature Films: 1912–1929" (Library of Congress, 2013): https://www.loc.gov/static/programs/national-film-preservation-board/documents/pub158.final_version_sept_2013.pdf • CLIR/Library of Congress — "The State of Recorded Sound Preservation in the United States: A National Legacy at Risk in the Digital Age" (2010): http://www.clir.org/wp-content/uploads/sites/6/pub148.pdf • Bill Klinger — "Cylinder Records: Significance, Production, and Survival" (Library of Congress National Recording Preservation Board): https://www.loc.gov/static/programs/national-recording-preservation-board/documents/klinger.pdf • Library of Congress — "Library of Congress Study Finds Majority of U.S. Recordings Not Available to the Public" (2005): https://www.loc.gov/item/prn-05-190/ • Library of Congress — "Sound Preservation" program page: https://www.loc.gov/programs/national-recording-preservation-plan/sound-preservation • UC Berkeley Library — "Project IRENE: Using optical scanning to digitize ethnographic field recordings on wax cylinders": https://exhibits.lib.berkeley.edu/spotlight/project-irene • Lawrence Berkeley National Lab — Project IRENE homepage: https://irene.lbl.gov/ • Variety — "MySpace Music Data Loss: 50 Million Songs" (2019): https://variety.com/2019/music/news/myspace-music-data-loss-50-million-songs-1203165649 • NPR — "Myspace Says It Lost Years of User-Uploaded Music" (2019): https://www.npr.org/2019/03/18/704458168/myspace-says-it-lost-years-of-user-uploaded-music • Lost Media Wiki — "Doctor Who (partially lost episodes)": https://lostmediawiki.com/Doctor_Who_(partially_lost_episodes_of_British_science_fiction_TV_series;_1963-1974 • SlashFilm — "How The BBC Lost Over 100 Episodes Of Doctor Who": https://www.slashfilm.com/675692/how-the-bbc-lost-over-100-episodes-of-doctor-who/ • Vint Cerf — "Warns of Digital Dark Age" (ACM, 2015): https://cacmb4.acm.org/news/183231-vint-cerf-warns-of-digital-dark-age/fulltext • Computerworld — "Vint Cerf worries about a 'digital dark age'": https://www.computerworld.com/article/1638666/vint-cerf-worries-about-a-digital-dark-age-and-your-data-could-be-at-risk-2.html • Andrew Potter — "Digital Preservation in 2025" (Substack): https://metaarchivist.substack.com/p/digital-preservation-in-2025 • Tom's Hardware — "Internet archival sites struggling to preserve the internet because of skyrocketing hard drive prices due to the AI boom": https://www.tomshardware.com/pc-components/storage/internet-archival-sites-struggling-to-preserve-the-internet-because-of-skyrocketing-hard-drive-prices-due-to-the-ai-boom • DSHR's Blog — "Internet Archive's Storage" (January 2026): https://blog.dshr.org/2026/01/internet-archives-storage.html • UNESCO — "Endangered languages, endangered thought" (UNESCO Courier): https://courier.unesco.org/en/articles/endangered-languages-endangered-thought • UNESCO — "3,000 languages may go extinct by end of 21st century": https://www.dailysabah.com/arts/3000-languages-may-go-extinct-by-end-of-21st-century-unesco/news • …and 7 more in the episode research notes This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.

  2. 19h ago

    Right But Late: Speed vs. Correctness

    Nine episodes into recommender systems and it's finally time to answer the phone. When a user opens the app, a single request arrives at the gateway, bounces through a BFF (Backend for Frontend), and fans out to multiple services. But none of the architecture matters if it doesn't arrive in time—and the core tension is simple: a recommender that's correct but slow loses to a popular list that shows up fast. Today's episode explores the serving layer, fallback strategies, why the user staring at a spinner doesn't care about accuracy, and what the AWS picture looks like. 00:00:00 - Starting with nine episodes of recommender systems 00:01:48 - The BFF making parallel calls on your behalf 00:05:05 - Fan-out patterns and the slowness trade-off 00:06:38 - Mise en place: prep versus the finishing station 00:08:29 - Building the feed's fallback ladder 00:12:17 - Task definitions as recipes 00:15:41 - The AWS architecture picture --- Sources & further reading: • Sam Newman, Pattern: Backends For Frontends, 2015-11-18 — — BFF origin, one per experience, ownership, aggregation, partial responses.: https://samnewman.io/patterns/architectural/bff/ • Microsoft, Backends for Frontends pattern, Azure Architecture Center, 2025-03-19 — — extra hop, duplication, when not to use, GraphQL.: https://learn.microsoft.com/en-us/azure/architecture/patterns/backends-for-frontends • Dean and Barroso, The Tail at Scale, CACM 56, 2013 — — tail latency in interactive services.: https://research.google/pubs/the-tail-at-scale/ • Jas Chhabra, Caching challenges and strategies, Amazon Builders' Library, 2019-12-05 — — cache addiction, cold start, soft/hard TTL.: https://builder.aws.com/content/3K9TeXp6strMiJ4vhakpRD7BQWO/caching-challenges-and-strategies • Amazon Builders' Library, Timeouts, retries and backoff with jitter, 2019-12-05 — — p99.9 timeouts, retry amplification, jitter.: https://builder.aws.com/content/3EumjoZascWd1oZiEgL8ORlv3qE/timeouts-retries-and-backoff-with-jitter • M. Nottingham, RFC 5861, May 2010 — — stale-if-error, stale-while-revalidate.: https://www.rfc-editor.org/rfc/rfc5861 • Martin Fowler, CircuitBreaker, 2014-03-06 — — trip, half-open, Nygard.: https://martinfowler.com/bliki/CircuitBreaker.html • AWS, What is Amazon ECS? — — ECS definition, layers, capacity options.: https://docs.aws.amazon.com/AmazonECS/latest/developerguide/Welcome.html • AWS, Architect your solution for Amazon ECS — — task definition/task/service/cluster; Fargate vs EC2 vs Managed Instances.: https://docs.aws.amazon.com/AmazonECS/latest/developerguide/launch_types.html • AWS, Architect for AWS Fargate for Amazon ECS — — Fargate isolation, Spot.: https://docs.aws.amazon.com/AmazonECS/latest/developerguide/AWS_Fargate.html • Danilo Poccia, Introducing Amazon Bedrock AgentCore, AWS News Blog, 2025-07-16 — — preview date, seven launch components.: https://aws.amazon.com/blogs/aws/introducing-amazon-bedrock-agentcore-securely-deploy-and-operate-ai-agents-at-any-scale/ • AWS What's New, Amazon Bedrock AgentCore is now generally available, 2025-10-13 — — GA date, nine Regions.: https://aws.amazon.com/about-aws/whats-new/2025/10/amazon-bedrock-agentcore-available • AWS, What is Amazon Bedrock AgentCore? — — definition, current component list.: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html • AWS, AgentCore Runtime docs — and https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-http-protocol-contract.html — microVM sessions, 8 h max, 15 min idle, port 8080, ARM64.: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-how-it-works.html • All fetched 2026-10-15. This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.

  3. 19h ago

    Four Seats: Where LLMs Actually Plug In

    When Spotify's language model faithfully narrated the wrong number of artists discovered, it revealed a deeper problem: guardrails need to sit upstream of the model, not downstream. This episode maps the four specific places where large language models actually fit inside production recommender systems—and all four share identical problems: they're expensive at scale, and their output always requires external validation. Joey walks through the architecture, constraint by constraint, with research-backed choices and an honest accounting of what he hasn't built himself. Key timestamps: 00:00:00 - Spotify Wrapped and the confidently wrong bug 00:01:30 - Joey's honesty check: five years of backend work, zero production recsys experience 00:04:35 - Cold-start retrieval: why you don't throw away IDs 00:09:25 - Generative retrieval and what the vocabulary shift means 00:12:45 - The container is right; the contents might be wrong 00:14:20 - The fix revisited from Part 2 logic --- Sources & further reading: • Okura, Tagami, Ono, Tajima, *Embedding-based News Recommendation for Millions of • Singh et al., Better Generalization with Semantic IDs, arXiv:2306.08121, 2023-06-13 • Hou et al., Large Language Models are Zero-Shot Rankers for Recommender Systems • Sun et al., Is ChatGPT Good at Search? (RankGPT), arXiv:2304.09542, 2023, EMNLP 2023 • Rajput et al., Recommender Systems with Generative Retrieval (TIGER) • Yoshnee Raveendran, "Inside the Archive: The Tech Behind Your 2025 Wrapped • Anthropic, "Structured outputs" docs • Zheng et al., Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena • All URLs fetched and quoted from on 2026-10-15. • Users*, KDD 2017 — — IDs fail on expiring news: https://research.lycorp.co.jp/en/publications/233 • content vectors; CTR +23%, duration +10%. Abstract page only. • random IDs vs content features at YouTube ranking. Abstract.: https://arxiv.org/abs/2306.08121 • arXiv:2305.08845, 2023, ECIR 2024 — — position and popularity: https://arxiv.org/abs/2305.08845 • bias, bootstrapping, 200-user budget limit. Full PDF text. • permutation, sliding window, 10 calls per 100 passages, $556: https://arxiv.org/abs/2304.09542 • 440M distillation. arXiv HTML v3; its latency figures render garbled and are not used. • arXiv:2305.05065, 2023, NeurIPS 2023 — — Semantic IDs, cold: https://arxiv.org/abs/2305.05065 • start, invalid IDs, beam-search cost. Full PDF text. • Highlights", Spotify Engineering, 2026-03-12 • batch precompute: https://engineering.atspotify.com/2026/3/inside-the-archive-2025-wrapped • stats block, distillation, LLM judges on ~165,000 reports, the timezone bug. • [primary, vendor case study] Meta AI, "Spotify personalized recommendations built with Llama" • 2024-12-18: https://ai.meta.com/blog/spotify-personalized-recommendations-built-with-llama/ • LLM explanations; the unmethoded "four times" figure. • constrained decoding: https://platform.claude.com/docs/en/build-with-claude/structured-outputs • and its stated exceptions. • arXiv:2306.05685, 2023, NeurIPS 2023 D&B — — 80%+ agreement: https://arxiv.org/abs/2306.05685 • judge biases. Abstract. • [preprint] Wang et al., Large Language Models are not Fair Evaluators, arXiv:2305.17926 • 2023-05-29 — — order hacking, 66/80, calibration. Abstract.: https://arxiv.org/abs/2305.17926 This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.

  4. 19h ago

    Opening Night vs. Dress Rehearsal

    A team from EPFL discovered something every recommender builder should fear: the model that performed best in offline testing was the worst performer live. This episode breaks down why offline metrics deceive, and covers three evaluation approaches—offline metrics, live A/B tests, and off-policy evaluation—that reveal whether your system actually works. 00:00:00 - The Paradox: Best Offline, Worst Live 00:01:47 - Starting with Offline Metrics 00:07:10 - How Offline Evaluation Fails 00:15:35 - Three Evaluation Approaches 00:17:20 - Why Procedure Matters --- Sources & further reading: • Garcin, Faltings, Donatsch, Alazzawi, Bruttin, Huber, Offline and Online Evaluation of News Recommender Systems at swissinfo.ch, RecSys 2014 — — offline/online inversion, borrowed clicks, CTR +35%.: https://www.epfl.ch/labs/lia/wp-content/uploads/2023/01/p169.pdf • Maksai, Garcin, Faltings, Predicting Online Performance of News Recommender Systems Through Richer Evaluation Metrics, RecSys 2015 — — "exact inverse" of online CTR ordering.: https://www.epfl.ch/labs/lia/wp-content/uploads/2023/01/p179.pdf • Krichene & Rendle, On Sampled Metrics for Item Recommendation (Extended Abstract), IJCAI-21, of the KDD 2020 paper — (paper page https://research.google/pubs/on-sampled-metrics-for-item-recommendation/) — sampled metrics inconsistent, collapse to AUC, AP vs NDCG decay.: https://www.ijcai.org/proceedings/2021/0651.pdf • Ji, Sun, Zhang, Li, A Critical Study on Data Leakage in Recommender System Offline Evaluation, ACM TOIS 2023 — — random splits leak the future; timeline scheme.: https://arxiv.org/abs/2010.11060 • Kohavi, Deng, Frasca, Longbotham, Walker, Xu, Trustworthy Online Controlled Experiments: Five Puzzling Outcomes Explained, KDD 2012 — (course-hosted copy) — Bing bug, OEC, sessions/user.: https://draft.epsy5221.education.uconn.edu/wp-content/uploads/sites/2358/2019/04/Kohavi-et-al-2012-Trustworthyness-of-research.pdf • Kohavi, Deng, Frasca, Walker, Xu, Pohlmann, Online Controlled Experiments at Large Scale, KDD 2013 — — alerts need statistical and practical significance.: https://exp-platform.com/Documents/2013%20controlledExperimentsAtScale.pdf • Fabijan et al., Diagnosing Sample Ratio Mismatch in Online Controlled Experiments, KDD 2019 — — the 50.2/49.8 example; telemetry-loss cause.: https://exp-platform.com/Documents/2019_KDDFabijanGupchupFuptaOmhoverVermeerDmitriev.pdf • Schultzberg, Ankargren, Frånberg, Risk-Aware Product Decisions in A/B Tests with Multiple Metrics, Spotify Engineering, 2024-03-05 — — success/guardrail/quality metrics, ship rule, preplanned analysis.: https://engineering.atspotify.com/2024/3/risk-aware-product-decisions-in-a-b-tests-with-multiple-metrics • Parks, Aurisset, Ramm, Innovating Faster on Personalization Algorithms at Netflix Using Interleaving, Netflix TechBlog, 2017-11-29 — (read via web.archive.org) — two stages, team draft, >100× fewer users, limits.: https://netflixtechblog.com/interleaving-in-online-experiments-at-netflix-a04ee392ec55 • Li, Chu, Langford, Wang, Unbiased Offline Evaluation of Contextual-bandit-based News Article Recommendation Algorithms, WSDM 2011 — — replay on a random bucket, ~40M Yahoo! events.: https://arxiv.org/abs/1003.5956 • All fetched and quoted on 2026-10-15. • [not fetched] Kohavi, Tang, Xu, Trustworthy Online Controlled Experiments, Cambridge University Press, 2020 — publisher page errored; nothing quoted. This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.

  5. 19h ago

    One True Name: How Data Contracts Keep Systems Honest

    Episode six of our recommender deep-dive leaves the algorithm's brain and explores its plumbing. When a basketball game ends, three systems fire—editorial publishes a recap, video publishes highlights, scoring publishes the final score. Each speaks a different schema. How do they become one unified event the feed can rank? Through data contracts: written promises between teams about what data flows and when. Joey explores the mechanics using Earthsea's concept of true names—and what breaks when those contracts fail. The recommender doesn't crash. It quietly learns the wrong thing. And that's worse. 00:00:00 - The Plumbing Problem: Three Systems, One Event 00:03:01 - Earthsea and True Names: The Le Guin Metaphor 00:08:10 - What Breaks in the Sports Feed 00:11:41 - Silent Failure: When the Plumbing Lies 00:13:12 - Feature Stores and the Model --- Sources & further reading: • Chad Sanderson and Adrian Kreuziger, *An Engineer's Guide to Data Contracts — Pt. 1*, Data Products Substack, 2022-10-05 — — contract definition, contracts as code, CI enforcement, FORWARD mode.: https://dataproducts.substack.com/p/an-engineers-guide-to-data-contracts • Bitol / LF AI & Data, Open Data Contract Standard README (v3.2.0 at check) — — definition, version, license, sections (also https://bitol-io.github.io/open-data-contract-standard/latest/).: https://github.com/bitol-io/open-data-contract-standard • Confluent, Schema Evolution and Compatibility — — modes, BACKWARD default and rewind rationale, upgrade order.: https://docs.confluent.io/platform/current/schema-registry/fundamentals/schema-evolution.html • Apache Kafka, Introduction — — events, partitions, per-key ordering, retention.: https://kafka.apache.org/intro • Confluent, Kafka Message Delivery Guarantees — — at-least-once default, idempotent producer, transactions.: https://docs.confluent.io/kafka/design/delivery-semantics.html • AWS, Kinesis Data Streams Terminology and concepts — — shards, partition keys, 24h default retention.: https://docs.aws.amazon.com/streams/latest/dev/key-concepts.html • AWS, Handle duplicate records — — producer vs consumer retries, embedded primary key.: https://docs.aws.amazon.com/streams/latest/dev/kinesis-record-processor-duplicates.html • Tyler Akidau, Streaming 101: The world beyond batch, O'Reilly Radar, 2015-08-05 — — event vs processing time, skew, watermarks.: https://www.oreilly.com/radar/the-world-beyond-batch-streaming-101/ • Akidau et al., The Dataflow Model..., PVLDB Vol. 8, 2015 — — late/retracted data, correctness-latency-cost trade-off (listing only; full PDF not read).: https://research.google/pubs/the-dataflow-model-a-practical-approach-to-balancing-correctness-latency-and-cost-in-massive-scale-unbounded-out-of-order-data-processing/ • Martin Zinkevich, Rules of Machine Learning, Google for Developers — — skew definition and causes, Rule #29.: https://developers.google.com/machine-learning/guides/rules-of-ml • Del Balso and Hermann, Meet Michelangelo: Uber's Machine Learning Platform, Uber Engineering, 2017-09-05 — — feature store, shared training/serving pipeline.: https://www.uber.com/blog/michelangelo-machine-learning-platform/ • Feast, Point-in-time joins — — point-in-time correctness, backfill leakage.: https://docs.feast.dev/getting-started/concepts/point-in-time-joins • All fetched on 2026-10-15. This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.

  6. 19h ago

    Reading the Ground: Signals and Cold Starts

    Recommender systems work like animal trackers—reading your behavior to guess what you want next. But a click is the noisiest signal; six clicks on a rival's drama doesn't mean betrayal. In this episode (part 5 of our recommender series), we explore the signals that feed user models, why dwell time matters more than clicks, what silence reveals, and the cold start problem that haunts every new user. 00:01:42 - Clicks: the cheap, noisy signal 00:06:30 - Dwell time: why staying matters 00:08:01 - Negative signals: mutes, hides, and what they mean 00:13:03 - Cold start: learning from nothing --- Sources & further reading: • Yifan Hu, Yehuda Koren, Chris Volinsky, Collaborative Filtering for Implicit Feedback Datasets, IEEE ICDM 2008, pp. 263–272, DOI 10.1109/ICDM.2008.22 — — no negative feedback, preference vs confidence, α = 40, 300,000 set-top boxes, CF cold-start quote (venue via api.crossref.org): http://yifanhu.net/PUB/cf.pdf • Xing Yi, Liangjie Hong, Erheng Zhong, Nathan Nan Liu, Suju Rajan, Beyond Clicks: Dwell Time for Personalization, RecSys 2014, pp. 113–120, DOI 10.1145/2645710.2645724 — — program abstract only (ACM full text returned 403): dwell as relevance proxy, normalization across devices and contexts: https://recsys.acm.org/?p=7780 • Paul Covington, Jay Adams, Emre Sargin, Deep Neural Networks for YouTube Recommendations, RecSys 2016 — — completion as positive, CTR promotes clickbait, watch-time weighting, impression features for churn, demographic priors for new users: https://research.google.com/pubs/archive/45530.pdf • Jiahui Liu, Peter Dolan, Elin Rønby Pedersen, Personalized News Recommendation Based on Click Behavior, IUI 2010, pp. 31–40 — — interest drift, local trend and sports examples, 10 virtual clicks, +30.9% CTR, +14.1% visits: https://research.google.com/pubs/archive/35599.pdf • Fangzhao Wu et al., MIND: A Large-scale Dataset for News Recommendation, ACL 2020, pp. 3597–3606 — — impression logs, 1M users and 160k+ articles, 84.5% of articles survive under two days, WikiData entity linking, entities help: https://aclanthology.org/2020.acl-main.331/ • Yueqi Wang et al., Learning from Negative User Feedback and Measuring Responsiveness for Sequential Recommenders, RecSys 2023 Industry Track, arXiv 2023-08-23 — — not-to-recommend loss, the dislike and skip live results, counterfactual responsiveness: https://arxiv.org/abs/2308.12256 • Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, Domonkos Tikk, Session-based Recommendations with Recurrent Neural Networks, arXiv 2015-11-21, ICLR 2016 — — GRU4Rec: https://arxiv.org/abs/1511.06939 • Wang-Cheng Kang, Julian McAuley, Self-Attentive Sequential Recommendation, arXiv 2018-08-20, ICDM 2018 — — SASRec, recent vs long-range attention, efficiency claim: https://arxiv.org/abs/1808.09781 • All URLs fetched on 2026-10-15. • [secondary] Martin Beck, "Yahoo Scientists: 'Dwell Time' Is A Better Metric Than Click-Through Rate", MarTech, 2014-10-07 — — the press-reported Yahoo details in Part 3, marked UNVERIFIED: https://martech.org/yahoo-scientists-measure-dwell-time-alternative-click-rate This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.

  7. 20h ago

    The Expediter: Why Your Feed Needs a Policy Layer

    A ranker scores each item individually. But you experience a page. In part four of the recommender series, Joey explores the re-ranker—the policy engine that handles diversity, freshness, deduplication, and editorial overrides. It's the difference between individual station cooks and the expediter at the pass. [00:00:00] YouTube's three failed fixes. [00:01:35] The ranker grades the exam, the re-ranker arranges the seating chart. [00:06:28] Why YouTube's best result never shipped. [00:11:25] How hand rules blocked better ML models. [00:17:27] The system becomes a curator. --- Sources & further reading: • Mark Wilhelm, Ajith Ramanathan, Alexander Bonomo, Sagar Jain, Ed H. Chi, Jennifer • Jaime Carbonell, Jade Goldstein, *The Use of MMR, Diversity-Based Reranking for Reordering • Abhinandan Das, Mayur Datar, Ashutosh Garg, Shyam Rajaram, *Google News Personalization • Fangzhao Wu et al., MIND: A Large-scale Dataset for News Recommendation, ACL 2020 • Gurmeet Singh Manku, Arvind Jain, Anish Das Sarma, *Detecting Near-Duplicates for Web • Yaron Greif, Pinterest Engineering, "Using PID controllers to diversify content types on • Robert B. Cialdini et al., Basking in Reflected Glory: Three (Football) Field Studies • All fetched and quoted on 2026-10-15. • Gillenwater, Practical Diversified Recommendations on YouTube with Determinantal Point Processes • CIKM 2018, doi:10.1145/3269206.3272018 — — pointwise-sorting: https://jgillenw.com/cikm2018.pdf • problem, last-layer DPP, windows, Table 1, +0.52% watch time, heuristics critique. *ACM PDF 403 • author copy read.* • Documents and Producing Summaries*, SIGIR 1998 • MMR: https://www.cs.cmu.edu/~jgc/publication/MMR_DiversityBased_Reranking_SIGIR_1998.pdf • definition, λ behavior, five-user pilot. • Scalable Online Collaborative Filtering*, WWW 2007 • item churn, time decay, Popular baseline and: https://www2007.cpsc.ucalgary.ca/papers/paper570.pdf • ~38% win, stories as article clusters. • dataset scale, 84.5% under two days, news cold: https://aclanthology.org/2020.acl-main.331.pdf • start. • Crawling*, WWW 2007 — — simhash, 64-bit / k=3: https://research.google.com/pubs/archive/33026.pdf • over 8B pages. • home feed," Pinterest Engineering Blog, 2020-07-10 • — hard-coded insertions, delayed models, global share targets, PID boosts, 4–12% range. • [primary, abstract only] Tien T. Nguyen, Pik-Mai Hui, F. Maxwell Harper, Loren Terveen, Joseph A. • Konstan, *Exploring the filter bubble: the effect of using recommender systems on content • diversity*, WWW 2014 — — "slightly: https://api.openalex.org/works/doi:10.1145/2566486.2568012 • narrowing" finding. Full text 403. • JPSP 34(3), 366–375, 1976: https://www2.psych.ubc.ca/~schaller/528Readings/Cialdini1976.pdf • apparel after wins; "we" 26% vs 13.5%. • [primary, abstract only] Jonathan A. Jensen et al., "Forty years of BIRGing," Journal of Sport • Management 30(2), 149–161, 2016 — — Study 1: https://repository.lsu.edu/kinesiology_pubs/364 • reproduced, Study 2 not. This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.

  8. 20h ago

    Ranking: What 'The Score' Actually Is

    What does a score of 0.83 mean? It's not relevance—it's a predicted probability of an action you might take. In episode 381 of the recommender systems series, Joey explores ranking: how systems order hundreds of candidates to show you the top few dozen. He works through published papers from Google and YouTube, learning how the third layer of recommendations actually works. Key moments: 00:00:00 - What is a score? 00:01:48 - From millions to hundreds: retrieval vs. ranking 00:05:19 - Hand-built cross features 00:06:57 - Features matter more than tuning 00:10:06 - YouTube's pointwise ranking approach 00:11:43 - Measuring a better thing 00:13:21 - Judging execution: the diving analogy 00:16:38 - The plumbing problem: architecture first 00:18:29 - Offline and online disagreement 00:20:22 - Models predict; people decide --- Sources & further reading: • Cheng, Koc, Harmsen et al. (Google), Wide & Deep Learning for Recommender Systems, arXiv:1606.07792, 2016-06-24 — — O(10) ms / O(100)-item funnel, feature families, memorization vs generalization, crosses, Tables 1 and 2 (read from the arXiv PDF).: https://arxiv.org/abs/1606.07792 • Zhao, Hong, Wei et al. (Google), Recommending What Video to Watch Next: A Multitask Ranking System, RecSys '19, September 2019 — — engagement vs satisfaction, MMoE, pointwise for serving, "manually tuned" combination, shallow tower, Tables 1 and 2. ACM PDF returned 403; text read from a mirror of the published PDF (https://daiwk.github.io/assets/youtube-multitask.pdf, DOI and front matter match); authors cross-checked at https://research.google/pubs/recommending-what-video-to-watch-next-a-multitask-ranking-system/.: https://doi.org/10.1145/3298689.3346997 • Covington, Adams, Sargin (Google), Deep Neural Networks for YouTube Recommendations, RecSys '16, September 2016 — — user-item history features, retrieval source as a feature, churn, clickbait, weighted logistic regression for watch time.: https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45530.pdf • Burges (Microsoft Research), From RankNet to LambdaRank to LambdaMART: An Overview, MSR-TR-2010-82, June 2010 — — RankNet, flat/discontinuous metrics, Figure 1, lambdas as arrows, Yahoo! LTR Challenge.: https://www.microsoft.com/en-us/research/publication/from-ranknet-to-lambdarank-to-lambdamart-an-overview/ • He, Pan, Jin et al. (Facebook), Practical Lessons from Predicting Clicks on Ads at Facebook, ADKDD 2014 — — trees + LR, historical features dominate, Normalized Entropy, calibration, downsampling re-calibration.: https://quinonero.net/Publications/predicting-clicks-facebook.pdf • Wang, Shivanna, Cheng et al. (Google), DCN V2: Improved Deep & Cross Network..., WWW '21 — — feature crosses as "the key" (abstract only).: https://arxiv.org/abs/2008.13535 • All external URLs fetched and quoted from on 2026-10-15. This podcast episode was fully generated by AI — research, script, voices, and production. Built with Claude, Piper TTS, and automated pipeline tooling.

See All (76)

About

Podcasts about whatever I find interesting — history, tech, weird rabbit holes. I was making these for myself anyway, so I figured I'd share. Fully AI-generated: Claude researches and writes each episode, and Piper voices it, one of them trained on my own voice. It all runs on a pipeline I built with Claude Code on an old laptop in my house. Enjoy.

Information

  • Creator
    Joey Musselman
  • Years Active
    2026
  • Episodes
    76
  • Rating
    Clean
  • Show Website
    Clown Cast