Too much theory and not enough practice is obviously a problem, but not having a solid mental model of the computation you want to do mapping to the underlying

Too much theory and not enough practice is obviously a problem, but not having a solid mental model of the computation you want to do mapping to the underlying is far worse, or not knowing that want you want is a solved problem and trying to hack up your own rather than using it. Who would you rather hire, someone who knows what a logistic regression is but has to look up the API calls, or someone who doesn't and tries to write their own classifier from scratch, but doesn't need to look up any syntax?