This cynical point of view is shared by a number of engineers I know. Another version of it is 'why is it worth learning the calculus of machine learning when that is mostly abstracted away by Tensorflow/PyTorch/JAX?'

To a software engineer accustomed to operating on layers of abstraction far removed from the hardware, this may seem a reasonable point. Why is it worth learning that pesky math, anyway?

I would argue that the machine learning engineers of today are more like electrical engineers than programmers, however. When something goes wrong, you don't have nice warning messages or error catching available to you. Like an electrical engineer with a voltmeter, one must begin probing inputs and outputs each step of the way. Good luck doing that if you do not understand how the components are supposed to work.

YMMV by copying and tweaking others code, but I believe we are still far off from hands free 'autoML'. Just ask anyone who has sent a model to deployment whether AWS autoML was sufficient for them. And whether they needed someone who understands backprop at some point during the model training process.