The article and some of the people quoted in it, like Hinton, seem to think MDs need so much help for diagnosis. Real life isn't House MD. The answer is very often obvious from the history/physical exam and the most basic labs.
Anyways, I don't like the fact that Hinton and others seem perfectly OK with not knowing how the machine is diagnosing. It's machine clinical gestalt.
I also think this diagnosis by machine would be very frustrating. Imagine, as a patient, asking the diagnostic robot: "why do you think this happened to me?" or "Why do you think the diagnosis is x?"
And then not having an answer other than "the imaging and tests are consistent with imaging and tests of previous patients with x disease" - this sounds like a bad answer. That wouldn't be good enough for me, but maybe it would be for plenty of other non-curious parties. Maybe there is this huge group of people who want healthcare from robots with robotic bedside manners. But I doubt it. Hinton is wrong, doctors are going to be augmented by helpful diagnostic applications. We will still have to learn to diagnose on our own but we will have help too. Maybe a robot to help triage cases into "serious/less serious" categories (and with working initial diagnosis) with good accuracy.