this is the thing about most of what we're doing with machine learning. It gets exponentially harder the closer you get to 100%; most of the things we use machine learning for where it works pretty well, 99% is really pretty great. 99% won't cut it for self-driving cars.

Which isn't to say that self driving cars won't get there, just that the fact that it looks like we are almost there doesn't mean as much as you'd think; we still have the really hard bits in front of us.

> That makes human drivers 16 times safer than Uber self-driving cars.

While that's a difference large enough to make wide scale deployment of those cars irresponsible, it is not large enough to make one pessimist on self-driving cars existing soon. It's the kind of difference that often enough go away with a few years of engineering.

And I expect those to be the worse stats for all the self-driving car companies around, since others seem to be doing things in a less "move fast and break things" way.

Many accidents have no claim filed

"The average driver" isn't our cutoff for allowing people to drive. Once the self-driving cars are safer than the most dangerous drivers that we allow, there's theoretically an advantage moving those people to self-driving cars.

"The average driver files a claim for collision once every 17.9 years"

That leaves open the question of how much damage that driver is at fault for on average.

Looks like there's still a lot of work left on that front. I was trying to get some context for the 15000 miles number.

Citation desprately needed.

I would theorize the opposite.