Executing programs inside transformers with exponentially faster inference