ChatGPT can already run code, which allows it to overcome some limitations of tokenization (eg counting the letters in strawberry, sorting words by their second letter). Doesn't seem like adding a Prolog interpreter would be all that hard.