But it's easy to develop any such killer apps in any of the ML frameworks. And there will not be a single such killer app, but there are many. It's more a ques

But it's easy to develop any such killer apps in any of the ML frameworks. And there will not be a single such killer app, but there are many. It's more a question what most of the community uses, esp the research community uses, as that is what gets into production first at some point. If you have your nice model already in PyTorch, why recode it in TF later? So, the ML framework adoption always has a bias towards what's more friendly for the researchers, and developers.