It all comes down to ROI of going further; if the ROI is small, people won't go much further than twiddling some knobs. I think there's a reason you see people

It all comes down to ROI of going further; if the ROI is small, people won't go much further than twiddling some knobs. I think there's a reason you see people throwing a lot of analysis at finance. Understanding what models are doing and adding new features/data transformations is standard work for ML practitioners now. If you believe the deep learning hype, architecture engineering is the new feature engineering, and to the extent that that is true, industry will essentially resemble academia to some extent. Maybe not making anything groundbreaking, but mixing ideas together.