What do we learn from inverting CLIP models?