Comment by apw

I thought the Owhadi et al. paper was about model mispecification; i.e., the true model is not in the hypothesis space. That's pretty fundamentally different--and far less of a problem--than gradient descent's "sensitivity to initial conditions".