>Tesla is a level 2 system

Does autopilot actually meet the qualifications for a level 2 system? AFAIK since autopilot is basically lane following and adaptive cruise control and braking wouldn't that be considered level 1?

Yes, the Tesla would do that, at one point in Autopilot's history. Perhaps it still will, but for the past many months I have not had to disengage it in those types of scenarios, to the point that the car has reacted before I have, and this is with the previous generation of sensors. From a technical perspective Tesla has come quite close in practice in a three year old production vehicle to what Audi claims to be ready to ship. With the current generation of sensor package, I would not be surprised to see Tesla's vehicles achieve L3 in practice (not at a regulatory level) before the new A8 actually starts deliveries. My point being that Tesla's technology has taken quite a leap from where it was originally, even with the original sensor package. Treating what Tesla is shipping in software as a fixed point and saying "this would put Audi far ahead of competitors" (as the article does) misses what has allowed Tesla to make such rapid progress in the space: continuous iteration on the live platform based on data collected from their existing fleet across all conditions.

Do you have statistics that per kilometre it makes stupid mistakes more than a human driver?

Lots of human caused accidents were driving into obvious things too. I don't drive drunk or sleepy, so maybe I'm above average, but my only standard for safety for a self driving car is that it's safer than me driving.

Sample size is too small to tell. However, if you ignore that, Tesla's Autopilot is at least 2-3 times worse than the average human in a passenger vehicle. And it only gets worse if you restrict further to middle aged people in expensive vehicles.

Source: http://www.greencarreports.com/news/1106613_how-safe-is-tesl...

Stats are on low page 2 and high page 3.

TLDR: Tesla compared their numbers to a statistic that includes much more dangerous forms of transport(bike, 18-wheelers, whatever). The Insurance Institute for Highway Safety did a study on just driver fatalities in passenger vehicles(cars and light trucks) and came up with 1 fatality / 438 million miles driven. Versus the Tesla figure of 1 / 130 million miles.

Whole article is worth reading as it goes into more detail on statistical issues with Tesla's safety statistic claims.

I'll add that human driving ability is not uniformly distributed. Most accidents are due to particular demographic groups: people that drink and drive, teenagers. It's entirely possible for a self-driving car to be worse at driving than most humans, while still better than the average. In that scenario getting really bad drivers into self-driving cars would improve average accident numbers, but getting non-bad drivers(aka the majority of drivers) into self-driving cars would make average accident numbers worse.

It's absolutely irrelevant if it crashes "less" than a human driver. It just cannot fail to detect an obstacle in front of it to be allowed on the road, period. If there is any edge case where that can happen, then the car shouldn't be certified as safe to use. Like, I don't understand how it's not obvious. A heart surgery machine that kills the patient every once in a 1000 operations would be taken off the market faster than you can say "litigation", even if a human surgeon has a worse survival rate(say 1 in 100 operations, so lets say the machine is an order of magnitude safer). That has already happened with radiotheraphy machines, there was a model that on average definitely saved lives but it had a failure mode where it would kill a person. Would you continue using it? Of course not.

Or airplanes - autopilot in planes has contributed tremendously to improving flying safety, but every time there is a crash under autopilot both Boeing and Airbus will ground every single plane of the same type to figure out the cause, they don't just go "oh well, it's still safer than flying manually so it's good for us".

I really don't understand this notion that an autonomous car merely needs to be "better than average driver" to be allowed on the road. Absolutely disagree here.

Here's a justification for you: It's impossible to get a failure rate of 0, so there has to be a point where we get a satisfactory(for now) level of safety.

Traffic kills so many people annually, that even a slight increase in safety is going to save more lives that are lost to aircraft accidents, radiotheraphy machine failure and such.

You can't get to a failure rate of 0, but I can understand people wanting the car to drive at least as well as a human who was in, say, the top 5% of human drivers.

Who'd want an autonomous car that's merely better than the _average_ driver, when every driver knows they're above average :)

Of course, and I acknowledge that accidents will happen anyway. But we cannot have a car that is on the road with "autonomous" driving functionality when it is known that the sensors can't detect an obstacle above half-pillar height(like in that famous tesla crash), or that the lane detection can fail in intense sunlight(like dozens of videos on Youtube show). It should have no known failure states upon release, and then ones which are found should be fixed later. I'm just worried that we're in such a rush to release autonomous cars on the road that we are ignoring known issues for the sake of improving average safety on the road - literally no other industry works like this.

Isn't your argument why we have the saying that "the best is the enemy of the good"? You seem to be dismissing any automated product that is not perfect even if it is demonstrably already much safer overall than the manual human equivalent. Surely this can't be the right policy if the goal is to maximise safety.

This is not to say that current autonomous vehicles actually are much safer than the human equivalent. Indeed, everything I've seen so far suggests that they have a long way to go before crossing that threshold, so personally I'm a sceptic about claims that we will see viable fully autonomous vehicles any time in the next few years. But this doesn't seem to be your argument here.

> Isn't your argument why we have the saying that "the best is the enemy of the good"?

Regardless of the OP's position on this, I do think the reality is that in practice it's difficult to sell or legalize anything that isn't significantly better than 'the average driver'. Whether it makes sense or not, we seem to demand any 'AI-powered' system to be better than the average human. I personally wish this wasn't the case, but it appears to be so.

For example, I personally marvel at the abilities of Siri and Google Now, at the very least when it comes to understanding what I'm saying. It's quite possible, in fact, that both are often objectively better at understanding me than most people around me: more than once I said something to Siri where I immediately realized that it was mumbly enough for any real human in my vicinity to ask 'what?' (often only to right away act on what I said, indicating that they did manage to interpret what I'd said). And yet Siri or Google Now actually transcribed what I'd said.

Being as good as the people around me would not be good enough, because every time Siri fails it makes me shake my fist at the whole concept of 'personal assistants' on my iDevices, rather than sigh at the umpteenth person around me saying 'what?' and then obviously acting on what I'd said, which is much more common.

I've not studied this properly, so I might be wrong. But I get the impression that I hold my AI to a much higher standard than I would my human interactions.

How many other industries have such a high potential for saving lives by improving average safety? Traffic kills more than 80000 people every year in the EU.

Would you rather have a large chance of your car driving into something it couldn't detect and a slightly smaller chance of being seriously injured or the status quo? Keep in mind that the odds of being seriously injured in a car crash in one's lifetime are already fairly low. How many people do you know that were seriously injured in a vehicle crash?

Most of the safety of AI comes from it being implemented in vehicles that are already far safer than average.

I'd rather have the solution with fewer serious accidents, even if it introduces types of accidents that don't happen with human drivers. I know a handful of people who were seriously injured in traffic.

I think accidents in traffic will be completely 100% solved by brake assist and collision prevention systems - I think they are meant to be mandatory in all new cars sold in US by 2020 if I remember correctly? So not long now. AI solves a different category of problems and isn't really necessary to solve crashes when there is a lot of cars driving close to each other.

I'm looking forward to those things as well. But bicycles don't have brake assist. It takes a very sophisticated collision detection system to prevent right-hooking a cyclist.

Electric assist bikes/mopeds will let most cyclists safely keep up with traffic in the kinds of places where you have to worry about those kinds of collisions, bike lanes will go the way of the dodo and car traffic will no longer have to take right turns across a lane of (bike) traffic.

Having a traffic pattern that requires a right turn (or left for certain countries) across a different lane of traffic is shitty and when a better solution shows up it will probably be implemented.

What you describe sounds like an irrational fear of machines, or perhaps fear of change?

Let's take airbags as an example. Do you understand why airbags are allowed, or do you also find it obvious that airbags shouldn't be certified as safe to use? Airbags provide increased passenger safety on a statistical basis, as they mostly provide a cushion for impact. However there are plenty of edge cases where if you sit too close to an airbag, or let your small child sit next to an airbag, then that airbag will actually cause significant damage that wouldn't happen without the airbag.

Autonomous driving can be viewed from the same safety perspective as airbags. Sure there are new edge cases that can cause harm, but on a statistical basis if the AI is even just slightly safer than a human, then at the end of the day there will be fewer accidents and fewer deaths.

--

Your airplane example is a straw man argument. Nobody has claimed that we should stop advancing AI driving tech once we slightly surpass human skills. That's just the point where it makes sense to allow it on public roads.

It's not about the number of edge cases, it's about the overall result. Airbags reduce variance in outcome. They hit a lot of people who would otherwise be unharmed in the fact with a small explosion in order to save a few lives in serious accidents.

The worst case is better and the average case is better because you've eliminated some outliers. However,the median is far worse. The group is effectively being punished to reduce the effect of the worst performing members.

An AI that drives into a partial lane obstruction is only better than no AI for a very small number of cases.

My argument with your radiotherapy machine is that we're currently using the 'model that on average saves lives', and therefore anything we can replace it with that 'on average saves slightly _more_ lives' is an improvement.

Once in a while the hot water tap will blow scalding steam on your hands, but the average temperature is nice.

(Do you feel lucky?)