First we had hypothesis-driven research. But our measurement techniques have become so high-throughput that investigating one hypothesis at a time seems poorly suited to the task.
So then we had hypothesis-free research. But without knowing the hypothesis, you can never be sure the experiment is well-suited to the discovery.
So, finally, we will have all-hypothesis research, where experimental design will explicitly consider the entire distribution of possible hypotheses, and be chosen to maximize information gain.
(I think this will also help to get rid of p-hacking, because you can't p-hack if you explicitly consider the entire distribution of possible statistical analyses. I.e., if you measure the entire distribution of p over statistical methods, rather than cherrypicking statistical methods to obtain a specific value for p.)