Machine learning is an interesting case, in that the gap between theory and praxis is quite wide.
You can basically create, run and evaluate a model within 10 minutes (e.g. with scikit-learn or Tensorflow). It would even be an acceptable model that might solve your company's problem.
But you have to plan in a few years of deliberate practice to work through statistical learning theory and its prerequisites to get an overview and to be able contribute to the scientific or open source efforts in that field.
And in my experience the opposite applies as well. Having a PhD in a field providing the necessary background in theory doesn't necessarily imply the skills and experience required to produce useful models.
That's a really good question. The best people I've seen so far have had a willingness to rigorously align their models with reality, plus the knowledge and/or experience to know what to check. Verifying assumptions, picking evaluation metrics appropriate for the problem, checking for model interpretability, checking that model decisions are sane, and so on.
I disagree wrt contributing to open source. There's still a lot of general software engineering that goes into these systems. I've found d it pretty easy to contribute implementations since the core of many ideas are actually quite simple. Of course they're only simple in hindsight, but implementation is just that.
You can basically create, run and evaluate a model within 10 minutes (e.g. with scikit-learn or Tensorflow). It would even be an acceptable model that might solve your company's problem.
But you have to plan in a few years of deliberate practice to work through statistical learning theory and its prerequisites to get an overview and to be able contribute to the scientific or open source efforts in that field.