Interesting. a bug of mine suggests that OpenAI's "small" embeddings are locations on a sphere? I haven't chased it down, it's ancillary to what I'm trying to accomplish, but that's a dramatically smaller semantic space than I'd intuitively have thought.