BTbtown·Oct 5, 2025
To add a (positive) point of comparison: The concept of multiple named "feeds" with various degrees of "algorithm-ness" to them, including ones largely controlled by the user's explicit signals, is something that Tumblr implemented in a fascinating way.
IMO Tumblr's constraints were quite unique in the social space: a userbase that one could quite naturally (and with a badge of pride) describe as "rabid" in its expectations that their historical dash experience be preserved in much the way it had been in the early 2010s (with or without a ball pit included); a legacy PHP codebase without the benefits of Meta's resources; and ownership under various companies throughout the years that (from an outsider's perspective) did not prioritize algorithmic optimization of its feeds as a C-suite-driven priority, unlike practically every other social
Read the rest
network at scale.
The result is systems like https://github.com/Automattic/stream-builder that enabled Tumblr product teams to define Following, For You, Your Tags, Popular etc. feeds with complex but deterministic and predictable specifications, where each feed would round-robin each slot from a mixture of well-defined component streams.
I'm not as familiar with Bluesky's implementation, but I imagine it has a similarly predictable algorithm for these user-defined feeds.
I would encourage anyone building any kind of discovery experience to look at this repo, at least at its readme and guide, for inspiration. You can, and should, build magical experiences without needing to make your feed algorithms inscrutable. Not everything needs to be TikTok, and I think the world can be a bit more sane if we don't have that as our base assumption as engineers.