Neel Nanda comments on Linear encoding of character-level information in GPT-J token embeddings