Curating massive music libraries. I've been using a small embedding model to organise my music for DJing, and being able to generate a t-sne plot clustered on perceptual similarity has been wonderfully useful.

I've also found CLIP models useful for searching videos, just embed a screenshot every couple of min of footage and query with a description of the scene.

And as bad as generated subtitles can be, when the only other option is nothing at all they are pretty nice to have.