> Can you broadly explain where the gap in AI pedagogy is and how students can fill it?

Machine learning can be made much more effective and efficient, by first modeling the problem, and then optimizing that tailored representation. This is an alternative to throwing a bunch of layers of neurons, or copy pasting an architecture, and hoping something works out.

One of the most successful applications of ML is convolutional neural networks and is based on this principle. Image processing algorithms come from an optical theory which can be modeled with convolution - what if we used optimization to find those convolution kernels.

With four parameters I can fit an elephant, and with five I can make him wiggle his trunk.

Also you need to know when a problem is NOT optimization - for example solving equations via the bisection method.