Compare across CLIP checkpoints

Same data, different size models
- how does it impact attribute diversity wrt identitites? expect:
- attribute-level variance to go down
- identitites that are overfitted in smaller models to function better in larger models
same topology, different datasets
- different datasets will have different identitites (and distributions over the ones that are shared)
- how do learned biases correspond to biases in the data? anything unexpected? (beware vast search effect)

dmarx / whats-in-a-name