Mapping the semantic void: Strange goings-on in GPT embedding spaces — LessWrong