Personally I think it's a combination of a) much better resources being available than in the past, which explain concepts clearly and in relatively simple ter

Personally I think it's a combination of a) much better resources being available than in the past, which explain concepts clearly and in relatively simple terms, and b) much better tooling existing now than ever before, so many things that you used to need to do "by hand" are now taken care of by relatively standardized tooling. Even just automatic differentiation engines are a huge deal, not having to backprop "by hand" (first release of TensorFlow was 7 years ago, so if you looked at it a decade ago I'm assuming you had to implement backprop yourself). Beyond that another jump was from Tensorflow's delayed execution/compilation model (it really was a headache to work around its APIs/set up the computation graph) as well as just having a generally ugly API, to PyTorch's "literal" setup, where it feels like you're just writing regular Python code and performing operations "normally"