I think it's definitely a good thing to abstract away the nuts and bolts of a complicated ML algorithm for general use. Implementing the algorithms from scratch

I think it's definitely a good thing to abstract away the nuts and bolts of a complicated ML algorithm for general use. Implementing the algorithms from scratch is really useful if you want to learn the details, but most of the time the "standard" implementation will do fine, and will probably perform more efficiently. Of course as with any tool, a conceptual understanding is essential to meaningfully interpret the results though. I completely get what you're saying about not feeling like you did anything though. Once you understand the algorithms even at a basic level they stop feeling like magic, even though they'll still feel like that for other people.