People act like transformer based embeddings are better than semantic embeddings because they can use context. But, at the same time, at least one can understand semantics when they break. We will struggle with understanding broken ghosts that are beyond our powers of language to describe.

I'm not sure what this is exactly, but it certainly feels true to me.

The real reason that things are so crazy right now is because recommendation systems can only handle 2^64 cultural constructs inside of a social recommendation system framework and so it's like a big game of ethno-cultural musical chairs.

Should there be polls on Bluesky? Asking for a friend.