I used to think that getting just to human level of 9s is enough, but I've changed my mind recently.

I think we can get to human level, by pure fact that we computers don't get distracted (they only crash ;-) ) and can also pickup signals, we cannot, due to 360° sensors. But computers are still bad at dealing with cases that are obvious to humans, due to lack of context and awareness.

I don't think that public will accept technology like that. I think there'll be a lot of public backslash once videos showing crashes in cases that are extremely obvious to humans. Videos from avoided crashes, that wouldn't be that obvious to humans won't be enough to change the sentiment.

Distribution of headlines that say "car crashed" vs "car avoided a crash" will be the same as "someone does a bad thing" vs "someone does a good thing".

Statistics can tell the difference.

It can, but when was last time that average joe understood and cared about statistic, compared to flashy headline?

The average american has about five nines on a per mile basis in the US. That's in all weather conditions and including all types of accidents / road rage issues / car maintenance levels. Under good conditions, I wouldn't be surprised if another few nines could be added to that.

What are 9s?

I'm not 100% sure, but I think they're referring to the nines of reliability. Something, such as web hosting or server time, is typically deemed as "reliable" when it is 99.999% available as intended. I think what they're trying to say is that a video doesn't need to be reliable at all. As long as you can execute something once or fake it realistically, the video is good enough. For real-world application, though, you need at least 99.9% reliability and, arguably, you'd need 99.999+% for people to accept it.

I might be totally off, though, since I've never heard of the nines being used in the context of autonomy but it does make sense in context of the conversation being had.

I understood 9s in the parent context to mine as referring to accidents per mile. Humans have 5 9s reliability in the sense that the probability of having an accident in any given mile is <0.001%

That's not how it works. Autonomous cars would need to be a couple of orders of magnitudes safer than humans. Human killing human is one thing and machine killing humans is quite another. But more importantly, the blame currently gets distributed over many people's insurance policies. For autonomous cars, all blame would end up at car manufacturer. Imagine you getting blamed for every single accident that happens ever. Even if you have achieved human-level safe driving and chance of an accident is 0.001%, multiplying that with millions of vehicles is a huge number. A dominant autonomous car manufacturer will be blamed for 4 human deaths each hour even if they have achieved human-level safety.

I do agree in principle but there will be nations like china who dont care and just adopt it. After some years the benefits will be obvious and the rest will follow.

Humans have higher expectations for technology. Better than humans is not enough.

>Better than humans is not enough.

By that logic, perfection is not enough because perfect is better than humans.

Saying “x condition is not sufficient” does not mean “x is always insufficient”.