A note on the impossibility of conditional PAC-efficient reasoning in large language models