Noa Nabeshima comments on An exploration of GPT-2′s embedding weights