I haven't read this paper, but the way deep learning is evolving is becoming more like general-purpose programming. People have even started calling it "differe

I haven't read this paper, but the way deep learning is evolving is becoming more like general-purpose programming. People have even started calling it "differentiable programming". The basic driving force is that the neural network architectures people are using are becoming more and more complex, employing state and dynamically changing structures. To express this deep learning frameworks are becoming like general-purpose languages with the constraint that everything has to be differentiable for optimisation to work. So I don't think the driving force is that "normal" programs will have machine learned components, but that machine learning is becoming more like programming.