I've worked in other domains my whole career, so I was astonished this week when we put a million 768-len embeddings into a vector db and it was only a few GB. Napkin math said ~25 GB and intuition said a long list of widely distributed floats would be fairly uncompressable. HNSW is pretty cool.