gwern comments on Mapping the semantic void: Strange goings-on in GPT embedding spaces

[gwern]

Let’s say we look at a 1000d gaussian centered around some point away from the origin:

Why would it not be centered at the origin?