How curation quietly beats recommendation

Recommendation engines are some of the most sophisticated software ever built. Billion-parameter models, real-time inference, A/B tests on a planetary scale. And after a decade of refinement, the universal experience is: the feed is fine, but it never surprises me.

Curation, by contrast, is what one person with taste does on a Sunday afternoon. They pick ten things they like and put them in a list. That's the whole technology.

Curation wins for a few reasons. It can include weird things a model would underrate. It can change its mind based on a hunch. It can refuse to include a thing because it just doesn't fit, even if every metric says it should. And, most importantly, it has skin in the game. The curator's reputation is on the line every time they add something to the list.

A model has no reputation. It just optimizes the metric it was trained on. If that metric drifts, so does the feed.

The interesting bet for the next few years is that small curated lists, made by people with taste, are going to keep beating the giant feeds for the things that actually matter to you. And they're already winning.