> Stories like this drive home the point: you can’t train your machine learning for every eventuality.

This wasn't some crazy edge case. The Tesla hit a large stationary object that was directly in front of it.

The only thing it drives home is that Tesla's self-driving functionality does not relieve the driver from having to pay attention.

Then it's not self-driving.

Oh, it’s self-driving.

It may be undesirable self-driving, but it does drive itself.

Tesla doesn't claim that it is yet.

The name Autopilot would lead you to think otherwise.

Do you think that when you get into an airplane and engage the autopilot that the pilot just like...goes to sleep?

"The person in the driver's seat is only there for legal reasons. He is not doing anything. The car is driving itself."

Probably regulators, because insurance companies are better at statistics.

This. Self-driving risk is likely already comparable, if not significantly better, than human-driving risk, but even if it isn't, why would insurance companies care? They're the bookmakers, they don't really care which horse wins as long as they can reasonably accurately predict the odds. At worst insurance for cars with self-driving features becomes expensive.

The one thing I could see insurances worry about is correlated risk. Imagine if an accidental push installs a bad development version of the autopilot on all cars instead of the 2 test vehicles with extra sensors, causing a much higher crash rate. Add a broken auto-update mechanism preventing a rollback, and it could take a day to mitigate and a week to fully stop the crashes, and during that time a lot of cars could crash. As long as it's no worse than the crashes resulting from e.g. a day with severe weather, it'll probably be fine, but if it significantly exceeds that, insuring self-driving cars could become too unpredictable, and thus too risky for insurance companies.

How do you come to the conclusion that autopilot is safer?

The Tesla press releases? It's the kind of uncritical analysis that you would expect from a company trying to market something.

Food for thought: By considering only automated and mnual driving you exclude basically all possibilities. Hybrid systems that are designed around the limitations of humans and machines seem to be a much more promising but less flashy solution if safety is the goal. This approach has a long history if you consider things like ABS and the traffic fatality statistics suggest it actually works. (Crash mitigation technologies surely muddy the waters.)

Reported accident statistics in self driving mode. I said "likely" because I have no statistics for human drivers that are granular enough to allow a precise comparison.

> By considering only automated and mnual driving you exclude basically all possibilities.

I'm not comparing Autopilot to some hypothetical future system. I'm trying to determine whether the risk from Autopilot is comparable to human drivers as they drive now, because insurance companies have clearly shown that this is something they're willing to put up with.

It doesn’t matter if it is safer. You can’t hold a human liable for not having superhuman qualities. Humans make mistakes and those mistakes are accepted.

A computer making mistakes isn’t acceptable, every instance will be seen as a liability and a huge one.

computers will not be given the benefit of the doubt that humans have

> A computer making mistakes isn’t acceptable

[citation needed]

Insurance companies will just look at the stats. And you better believe they'll jack up your rates if your self driving car crashes. You'll be held responsible, like it or not.

That's likely to be Tesla's problem, not the problem of the car insurance company though.

> Stories like this drive home the point: you can’t train your machine learning for every eventuality.

Overturned truck or stationary emergency vehicle should be pretty high on the list; trees maybe a bit lower. But the sensor package doesn't do those well, so in the interest of comfort, stationary objects are filtered out before avoidance algorithms are run.

> so in the interest of comfort, stationary objects are filtered out before avoidance algorithms are run.

Running into stationary objects has a pretty adverse effect on comfort.

So does stopping for stationary objects like metal plates on the road or overhead signs.

So, your argument is that no matter what choices were made about tuning, Autopilot isn't ready for real use?

Given the choices that were made, I don't think it can do the job it's marketed as doing, and I don't think I could safely and reliably supervise a car that's expected to partially drive itself.

I much prefer to drive a vehicle where I'm clearly responsible, but the (on by default) assistance features might help me out from time to time if I screw up. A computer doesn't get bored of checking if there's a slowing vehicle or (if it checks for them) stationary object in the path; but I'm pretty sure I wouldn't stay alert enough to jump in. OTOH, I also don't like using cruise control either, but that's not because I want to stay alert, but because without radar cruise control it's too much work to use, and I don't like the following distance options for the radar cruise control I have access to.

There are a lot more trees than overturned trucks.

Not having sensors that see trees without reducing comfort is a choice.

If the car can’t drive safely in a situation, it needs to require a driver.

it's worse, they are not even trying to mimic intelligence, it's clusterfsck of ML "models" and "hyperparameters" which has just as much common sense as a gazillion monkeys trained to press random buttons or get electroshock.