Gradient descent is not a magic wand that makes computers behave like anything you want. The difference is still quite perceptible after several years and trillions of dollars in R&D, and there’s no reason to believe it’ll get much better.

Really, there's "no reason"? For me, watching ML gradually get better at every single benchmark thrown against it is quite a good reason. At this stage, the burden of proof is clearly on those who say it'll stop improving.