Huh interesting that you say it’s as high as 50%. I would tend to agree (and it’s similar when working as a data analyst), and wonder if other people have found the same in their data science roles.

Yep, I agree with you. And I’ve read interviews with several data scientists where they emphasize “make sure you know the question you are trying to answer, and how the answer to that question will be used before you start developing an approach”.

Also, good point about autopilot vs emergency manual mode.

This reminds me of the discussions we were having years ago in astronomy when storage was getting cheaper, and data volumes per unit time were getting bigger and bigger. Easy solution? Just archive all of it, of course! But without being able to effectively describe the original observer’s intent, and store and search that, the value of the data was pretty low. So why spend a few tens of thousands of dollars a year on storing data that really didn’t have much archival value?

I think data science is similar. We really have to cautiously examine what value using a particular model will add… and really examine it in terms of current context and envisioned context. Brute force cloud ML services can’t do that. Nor would we want them to. But guaranteed, a lot of people will be doing just that.