Here is how to see it. Business rules evolve over decades by the gradual accretion of special cases, often at the whim of a single whale customer; their implementations are usually untested (except for the narrow edge cases that surfaced them in the first place) and almost never documented. Now what is wanted is for "AI" to help with porting those edge cases, diagnosing their conflicts with each other, or composing new ones; but "AI" was trained on a corpus of other people's edge cases, definitionally unrelated to one's own, equally poorly documented, equally poorly tested. Shall that succeed?
"there is really nothing of interest to a software engineer in machine learning at all"
I disagree
Here is how to see it. Business rules evolve over decades by the gradual accretion of special cases, often at the whim of a single whale customer; their implementations are usually untested (except for the narrow edge cases that surfaced them in the first place) and almost never documented. Now what is wanted is for "AI" to help with porting those edge cases, diagnosing their conflicts with each other, or composing new ones; but "AI" was trained on a corpus of other people's edge cases, definitionally unrelated to one's own, equally poorly documented, equally poorly tested. Shall that succeed?