This has a ton of seemingly random assumptions, why can't we compress multiple latent space representations into one? Even in simple tokenizers token "and" has no right being the same size as "scientist".