Traditional physically-based hydrologic models’ performance has stalled in the last decade without a clear pathway toward improvements in the mechanisms. In addition, these models have grown to be complicated and expensive to run, and were not evolving toward universal access to information for the world. Deep learning models have come into the arena and have smashed the records in most if not all hydrologic prediction tasks. Yet many hesitated to join the foray because of the fear of losing some great features of the process-based models: process clarity, mass conservation, mathematical elegance, hypothesis testing, causal analysis and cross-domain interactions. More recently, there are new developments on the machine learning front – differentiable models can now seamlessly couple the strengths of machine learning and process-based descriptions and retain all of the above features. We provide an introduction to differentiable modeling in hydrology (or simply, differentiable hydrology, DH) and show its strengths in learning processes. We investigate grand questions that can be answered by DH where it can make groundbreaking contributions. Importantly, one could be a hitckhiker but not a nihilist – the techniques and experiences from the deep learning studies will provide important guidance for future developments. We demonstrate multiple examples in DH where knowledge as well as performance are gained from big data.