LLMs literally cannot see letters; all the words have been covered to vectors in concept-space. It may be able to do some reasoning on concepts as part of its next token prediction, but it's hard to reason about information that's inaccessible to you