a smaller embedding dimension is naturally easier for downstream small-N problems, even if the larger-sized representation is objectively better (ie in the large-N case).