Why are Machine Learning projects so hard to manage?

I am not sure why this article didn't generate much response - but I think we need to discuss this too.