there are some other nice properties: – we can get good performance with very little data
– we recover embeddings that "look" like the original dataset
– when some percentage of the original data is available, we generally select that data first

→ View original post on X — @jxmnop