TBH, a lot of the “machine learning” projects seem to be excuses for the developers to play with machine learning. It gets shoe horned in because it’s trendy, not necessarily because it’s appropriate.
“We have tons of log data to analyze… Let’s use machine learning to get insights about it,” but then nobody knows what exactly that means or what they’re looking for, so the project struggles.
Another good chunk of the machine learning projects are just candy for the investors. So the business people or the managers don’t really care about them. They always get low priority access to common resources and are under-budgeted. Nearly all of the companies I worked for wanted to do machine learning and all of those wanted it just to tell it to the share holders or potential investors.
His argument went with “machine learning is hard.” I might make an alternate case: managing projects is hard enough without adding unknown unknowns to your code. Every choice in ML matters, and depends strongly on your use case. How you structure a webapp is not often so unique to your project.
For anyone interested in the topic of managing ML teams, I recommend Andrew Ng’s: http://www.mlyearning.org