LLMs are really bad with letters, and in my limited understanding that’s because they don’t see words as strings of letters, they see them as tokens. It’s all numbers by the time the LLM is processing it.