This assumes that ML and AI research will continue to be silo'd outside of domain specific research. But it's not the case in academia and also increasingly in industry. You have computational neuroscience, bioinformatics, and many other traditional disciplines which have not only incorporated ML/AI methods but also pushed the fundamental methods research forward. We're increasingly seeing interdisciplinary methods and research becoming the norm in academia. In undergrad, nearly all the social science classes and all the hard science classes had some sort programming and quantitative methods requirement. Even in industry we're seeing interesting multi-disciplinary work.Many interesting innovations in time series ML methods have come from the algorithmic trading firms and medical research community has made contributions to computer vision and unsupervised learning approaches.

I had a colleague who during her PhD in particle physics wrote from high performance parallel computation frameworks from the ground up in C which was better than Hadoop and Spark in performance. And at my last enterprise AI startup, our CTO had come from a computational neuroscience background. Whether these folks end up in creating startups is a different question, but the talent definitely exists.

The more difficult problem is how to evaluate multi-disciplinary startups and businesses. There usually isn't good empirical evidence unless they follow a more established business model.