Lol, I built this exact thing as a 120% project at Twitter in 2011.
It was called “Flocks”. It even had transient/temporal flocks based on geo check-in and sound-fingerprint flocks for movies, tv shows, and events. We implemented the Shazam audio signature algo from a research paper we found.
The thesis was: it’s more natural for people to have conversations when there’s a shared context.
It became a widely known project within the company back then, but ultimately, like most things at Twitter, never shipped.