To add to the last bit: It's trivial to get ChatGPT to evaluate actual code it can't possibly have seen before and correctly produce the expected output. Only limits on the size of context and output stops it from being turing complete.