Accidents per miles is not too informative. For example do teslas fair better than an average human driver on eg. someone running a red in front of the car? On a suddenly overturned car on the freeway? Do we have enough data to answer these questions?

I don’t understand this kind of question. Obviously you need some statistical measure to compare safety, and incidents per mile (or vehicle years, or hours driven) is an accurate way to estimate accident rates. All we care about is that on average it crashes less than human drivers. If there are specific conditions where it fares worse those would be pretty obvious to address, and naturally the odds of when crashing the situation being weirder than usual is a given if safety improves for the average case.

That aside, the answer to the first one is yes. Because of sonar and the cameras, a Tesla can see traffic two or three cars ahead and will initiate braking way earlier than a human would in the case of someone else running a red light - there a few videos of this exact situation available in YouTube. As for the second, probably not enough data, but in absolute numbers humans are responsible for nearly al of those cases so far, they happen weekly.

By the way, the overturned truck was not a fatal accident - the car triggered emergency braking and the driver came out without a scratch. The fatal one was a couple years ago when a Model S ran under a white truck making an unsafe u-turn.

Incidents per miles is fine with a big enough data set —- but I disagree that it is enough in case of a new technology that can potentially kill. Like let’s say teslas are absolutely safe on the highways, much more than human drivers, but would have a tendency to hit pedestrians much more so than humans. It is possible that it will have a better incident per mile record even though no sane person would legalize them in this case.

> If there are specific conditions where it fares worse those would be pretty obvious to address

Like, if (inSpecificSituation) { payMoreAttention(); } ? This is the actually hard part of the problem, not breaking when something is close and stay in the lane. They can’t even create adequate test environments for the many many special cases that can trivially happen in a city.

Also, how does it see traffic two or three cars ahead? It sounds like a marketing gimmick but correct me if I’m wrong.

> Incidents per miles is fine with a big enough data set

Only if you're comparing comparable conditions. If you compare autopilot in the sun to humans in the snow, infinite miles won't make it any more valid of a autopilot/human comparison.

Tesla's miles driven is self-selecting, because AP won't engage when it can't because of poor conditions.

Humans don't get that luxury.

You can't say for Tesla (deaths/miles driven on good-for-AP roads in acceptable conditions) and for humans (deaths/miles driven on all roads in all conditions) and then act as if they are comparable or prove "improved safety", because they don't.

> All we care about is that on average it crashes less than human drivers.

All we care about is that on average it crashes less than human drivers in comparable conditions. If you compare a 40 year old driving a tesla on autopilot on a sunny freeway to an 18 year old driving a 1995 beater without modern safety features, no ADAS, etc, in a snowy busy intersection, autopilot doesn't have to be that great to look better. If you compare tesla on autopilot to "average human miles" you rolling some very invalid comparisons into the averages. You can only usefully compare averages if they're averaged over comparable contexts.

A less extreme comparison, if my memory serves, is tesla rolling out ADAS with autopilot when they did a before/after comparison. They compared autopilot-available + ADAS-engaged to autopilot-unavailable + ADAS-unavailable, or something along those lines, if I remember correctly.