The difference is that the training data isn't code: programming language nerds like me couldn't give it formal semantics that let us directly treat it as code, for example ("tokens", yes, but 8 bit micro BASICs had tokens as an implementation concept and they're kinda trivial)
