There are different classes of problem out there, and not all of them require 99.99999% accuracy. Self driving is tough, because the consequence for failure is death, and we still have a ways to go to get to human-level failure rates. But other tasks don't require this much accuracy, for instance I think ChatGPT could do the job of a lot of journalists right this second.
This strikes me as an utterly pointless thing to say. It could be rewritten as "it's not useful unless it's useful"
Right, the same critique could be made of humans, but it’s not all that insightful. I wonder if this was his way at pointing out that AIs should be able to give you a confidence interval, but he also just likes to seem smart I think