The principles are really easy though. At its core, neural nets are just a bunch of big matrix multiplication operations. Training is still fundamentally gradient descent, which while it is a fairly new concept in the grand scheme of things, isn't super hard to understand.

The progress in recent years is primarily due to better hardware and optimizations at the low levels that don't directly have anything to do with machine learning.

We've also gotten a lot better at combining those fundamentals in creative ways to do stuff like GANs.