The difference is the amount and nature of data that is available for training models, which go programmers > lawyers > doctors. Especially for programming, training can even be done in an autonomous, self-supervised manner that includes generation of data. This is hard to do in most other fields.
Especially in medicine, the amount of data is ridiculously small and noisy. Maybe creating foundational models in mice and rats and fine-tuning them on humans is something that will be tried.
This is true if you think of programming as chunking out "code". But great authors are not great because they can reproduce coherent sentences fast. The same goes for programmers. Actually most of the hard problems don't really involve a lot of programming at all, it's about finding the right problem to solve. And on this topic the data is noisy as well for programming.
