On: A Bayesian Perspective on Generalization and Stochastic Gradient Descent

Of course you are right, but these are the parts of papers not read or understood by journalists or ceo's or the public. This is leading to inappropriate and unsafe applications, and will lead to a backlash and damage. We should be as clear (at least) as the medics, promising lab results that may deliver in the future, until field tests we can't be sure.

Unfortunately the lab to field practice in ml is underspecified.

Yes, but this specific technical research paper doesn’t suffer these issues. So the original comment seemed wrongly placed here.