This is an interesting critique of modern AI: almost no problems have closed-form mathematical descriptions. The "Ng Reply" would be that the objective fn can

This is an interesting critique of modern AI: almost no problems have closed-form mathematical descriptions. The "Ng Reply" would be that the objective fn can be approximated by a finite composition of simpler functions, tuned by data (ie., relu.relu.relu...) However if you think about it for a second, that's clearly absurd in almost all cases. We know that we cannot even in principle formulate a closed form expression for (here, eg., people management) -- the idea that there exists some meaningful approximation feels provably silly.