Probabilistic Numerics is the notion that computation itself can be described as a form of learning, from electronically produced data. This removes the conceptual separation between empirical and computational information. One advantage of this view that will be discussed in detail in the talk is that probabilistic numerical computation allows seamless inference across dynamical systems. This can provide significant efficiency gains in ``physics-informed’’ versions of machine learning. On a more abstract level, the talk’s central argument is that one should not think of a “solver” for PDEs, ODEs or DAEs as an encapsulated piece of immutable code, but an interactive, adaptive part of the machine learning tool-chain.