Biological evolution is an enormously high-dimensional process fraught with contingencies and nonlinearities. For this reason, evolutionary biology has traditionally been considered an exclusively retrospective science. This state of affairs, however, is beginning to change, and notions of predicting evolution are increasingly being discussed with straight faces – a cultural shift brought about by high-throughput and high-resolution biotech, ever-increasing computational capacity, the explosion of machine learning and AI, as well as the increasing need for evolutionary prediction in our rapidly changing world. Previous assessments of predictability in evolution have, wittingly or not, employed something like a frequentist approach, whereby a fitness landscape and its derivative distribution of fitness effects (DFEs) are determined beforehand through enormously expensive and time-consuming lab work, and evolution then allowed to run its course to see if predictions were accurate. Our recent and ongoing work provides an alternative approach that is Bayesian in nature and allows for DFE updating in real time; it is derived from first principles, it has realistic prospects of practical implementation in real-world settings, and it can be implemented for a small fraction of the cost compared to frequentist-like approaches. We question the utility of DNA sequence data in predicting evolution and join other researchers in pointing to an emphasis on fitness-phenotype evolution as the way forward.