Good advice, but I think it should talk more about the distinction between the training objective and the true objective. For classic machine learning problems, like speech recognition or face detection, these were so close that we didn't even notice there was a difference. However, now ML models are being trained to predict clicks or or other proxies of "engagement" and these can be wildly divergent from the humane objectives we want in our products. In these cases it's really important to understand the gap between what you really want and what you can encode into an objective function.
True, but I think any form of supervised ML can and often does suffer the same fate. Once you define any loss function, you narrow the signal to a subset that presumes the signal and its distribution are understood and reliable.
That's why I'm thinking the long view of ML will ultimately evolve away from supervised (short term) methods, like GOFAI's focus on curated narrow problem solvers, and toward broader more unsupervised methods in which the objective is less precisely specified and more experimentally refined, over a longer time frame (and a larger amount of data and sources).
Funnily enough us, who did ML before the explosion of Deep Learning and the use cases you've mentioned, were quite clear that the challenge doesn't lie in having the greatest algorithm, but having it's product usable for the intended purpose by an end user that would be willing to use it at some expected periodicity (use-case specific).
