LLMs should operate on the code's AST when possible, and should be trained on code's AST as well