> marginal value of each additional one is probably low

This is the point of the paper. Specifically, that single embedding vectors are sparse enough that you can compact more data from additional vectors together to improve retrieval performance.