I think you are right. I think that we first have to have great implementations and applications, and only then can hope to evolve / select one or a few APIs/libraries that can cover most use cases well. It is impossible to design a good API upfront and then hope that the implementations will come for the simple reason that there are few people in the Clojure community that know this field well. That's why you see so many abandoned toy projects. People get all warmed up for a great idea, but when they see how hard it is to go beyond the basics, especially when compared to mature libraries such as numpy/tensorflow etc. they leave for the next cool thing.