LLMs are getting better at character-level text manipulation