Andre Karpathy argues that replacing systems of imperative code with deep learning models leads to better results and that software 2.0 is a set of pervasively learned models, leading to less code writing. This approach is beneficial for those who prefer machine learning.