There is an anecdote I half remember, that I would like to find the original source of, because it's illustrative when explaining the issues. It goes something

There is an anecdote I half remember, that I would like to find the original source of, because it's illustrative when explaining the issues. It goes something like: In the 70s/80s a research group was working on a CAD system. After adding all sorts of features, the professor assigned a task to one of his students to make it respond to voice commands. The professor expected this to take only a few months. It is difficult, even for experts, to have a good intuitive idea of where a good ML implementation is going to be easy or hard. You can be making good progress and then hit the "diminishing returns" part of the curve unexpectedly.