but why use “~float16” for something that in its own documentation the authors describe as a float16* ?