If you were a really good academic, you could come up with a theory that predicts the performance of a model you haven't actually tested. Physics is full of this - theories that predict something novel followed by other people doing expensive experiments to test them. The guy who published the theory gets the credit. That's probably what we want from academic AI researchers - theories that we can eventually use to design models, rather than just haphazardly building something and wondering how good it will be.