> So he think humans are intervening once every 1-2 miles to train the Waymo
Just to make sure we're applying our rubric fairly and universally: Has anyone else been in an Uber where you wished you were able to intervene in the driving a few times, or at least apply RLHF to the driver?
(In other words: Waymo may be imperfect to the point where corrections are sometimes warranted; that doesn't mean they're not already driving at a superhuman level, for most humans. Just because there is no way for remote advisors to provide better decisions for human drivers doesn't mean that human-driven cars would not benefit from that, if it were available.).
To apply this benchmark, you'd have to believe that Waymo is paying operators to improve the quality of the ride, not to make the ride possible at all. That is, you'd have to believe that the fully autonomous car works and gets you to your destination safely and in a timely manner (at the level of a median professional human driver), but Waymo decided that's not good enough and hired operators to improve beyond that. This seems very unlikely to me, and some of the (few) examples I've seen online were about correcting significant failures, such as waiting behind a parked truck indefinitely (as if it were stopped at a red light) or looping around aimlessly in a parking lot.
You'd also have to believe that when you wished to change how your Uber driver drove, you'd actually have improved things rather than worsened them.
Let's suppose Waymo's fully automated stuff has tenfold-fewer fatal collisions than a human. There's no way to avoid the fatal accidents a human causes, and the solution to Waymos getting stuck sometimes is simple. The point is that the Waymo can actually be described as superior to a human driver, and the fact that its errors can be corrected with review is a feature and not a bug - they optimize for those kinds of errors rather than unrecoverable ones.
Just to make sure we're applying our rubric fairly and universally: Has anyone else been in an Uber where you wished you were able to intervene in the driving a few times, or at least apply RLHF to the driver?
(In other words: Waymo may be imperfect to the point where corrections are sometimes warranted; that doesn't mean they're not already driving at a superhuman level, for most humans. Just because there is no way for remote advisors to provide better decisions for human drivers doesn't mean that human-driven cars would not benefit from that, if it were available.).