Video Recommendations in Industry

December 26, 2025·38 min·2 clips

Discover Weekly's most-loved feature started as a bug — and when Spotify fixed it, engagement collapsed.

The question is not humans versus machines. It is how a curator and a recommender system split the work when there is too much content for either side to handle neatly on its own. Corey calls that split algatorial. For film and television recommendations, that means algorithmic sorting paired with editorial judgment, with humans still involved where context changes the answer. Machines are good at organizing abundance. A curator might pull together a wide set of romantic comedies, then let algorithms tailor that set to each user. The human work is mostly cleanup, boundaries, and context. Corey keeps coming back to metadata, especially bad or inaccurate metadata, as a quiet source of weak recommendations. If the tags do not match the content, the system can make a perfectly logical recommendation that still feels wrong. He also makes room for different levels of precision. Some playlists should be broad. Narrower collections, like romantic horror comedies, need tighter edges and more care. When the host asks for a plain definition, Corey separates the work he wants automated from the judgment calls where he wants to trade signals back and forth with machine learning systems. That back and forth is the point. Human-led recommendations do not replace algorithms. They shape the material algorithms use and catch cases where the data sends the system sideways. Near the end, Corey points to Notebook LM as a useful way to work through papers on recommendation systems from a bounded source set. He likes that limit because it keeps the tool focused and lets it admit when the answer is not in the material. The episode treats curation as cleanup, context, and data work, not taste-making magic.

As heard by us

A grounded look at why human curation still matters inside industrial recommender systems.

Data Skeptic uses Corey Zechman's work at Sling TV and his long running music blog to test a stubborn question in recommendation systems: where does human taste still matter? The episode is strongest on "algatorial" work, the mix of algorithmic sorting and editorial judgment.

Read the full review in PlayNext →

Why you'd press play

If you want a calm, practical case for human curation in recommender systems, start here.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Data Skeptic