You make an argument like you have a hammer in your hand and you are looking at a row of nails. I'm suggesting it is short-sighted to look at the problem from only that lens.

For instance, with a more holistic lens you might explore the right place for humans drivers vs autonomous vehicles. You might simulate entire cities from the ground up, specifically optimized for transportation. What if the real breakthrough in transportation systems actually comes from the ability to quickly construct / deconstruct roads? Or dynamic city zoning? Or any number of non-vehicle innovations? What if we were actually optimizing for safety in transportation and not just trying to compare machine ability to human ability in limited contexts?

The worry is this, the cars will not fail like a human will. It may even be dangerous The car may not even be able to move in certain situations (like another car blocking them in front either purposefully or accidentally). Or it may not be able to even see certain conditions a human could.

You're operating on a completely archaic understanding of AI. Modern systems can and do adapt to situations on-the-fly. What's more, they can adapt on aggregate - the experiences gathered by a single car can be shared to all cars. Very quickly, every car in the fleet will have billions of hours of cumulative driving experience.

The overwhelming majority of motor vehicle accidents aren't weird and unpredictable edge cases. They're tragic but mundane events that result from a handful of root causes - inattention, excess speed, poor judgement and unnecessary risk-taking. "Driver/rider failed to look properly" is the key contributory factor in nearly half of road traffic accidents. Computers utterly dominate humans in this respect. For every accident caused by some bizarre and unpredictable set of circumstances, there are thousands caused by someone doing something obviously stupid.

A computer can be programmed to be ultra-cautious in difficult situations. A computer can maintain 100% vigilance 100% of the time. Humans can't. Self-driving cars will undoubtedly fail in new and unexpected ways, but it's abundantly clear that they'll fail much less often than humans.

That's sort of my point. Accidents are rarely caused by edge cases, but by predictable human failures. All the information needed to avoid the typical accident was plainly available, but the driver just didn't see it, process it or react to it appropriately. Driving is generally very predictable, but human attention and perception is hugely fallible. An AI that gets the basics consistently right but occasionally freaks out in an unpredictable situation would be a huge improvement over human drivers.

In the same way, if the world is modelled accurately and succintly, can a car predict the various scenarios and adapt, even if the current sitation was not seen before (extremely high speeds or unusual number of cars, etc.)

From a human's perspective it's a stupid mistake. It seems stupid because computers think differently than people do. On the other hand a computer would probably think a human were stupid if it saw a human make a mistake when calculating the logarithms of every number between one and a million.

It's an improvement as long as those are few percent of the unsurprising and human ways human drivers fail.

It's more like humans can think, why should we trust a computer with a task where unforeseen circumstances can arise?

Further, if I got the current state correctly, no self-driving car system so far works in rainy conditions.

https://www.youtube.com/watch?v=bAxoo6JLmgg

or this

https://www.youtube.com/watch?v=JoGgE55qNBA

and this is just a L2 system...come on brah come harder

* Lack of software bugs due to careless programming.

* Millions of years of brain evolution allows adapting to unforeseen circumstances.

* Vastly decreased likelihood of being tricked by vandals exploiting flaws in recognition algorithms.

* Inability to be hacked en masse and driven into each other/houses/pedestrians.

* Much harder to back-door.

* Can be judged in court if they swerve to avoid a pigeon and hit a crowd of pedestrians.

I could go on, but this should drive the point home pretty clearly, IMHO.

They sell these features as "parking assist" or "smart cruise control" or similar, but don't kid yourself into thinking that there's not a very sophisticated software system that could, if hacked, take over control. Has this ever happened? Have there been catastrophic bugs documented?

More advanced vehicles (i.e. Tesla) have enough hardware to be be nearly fully autonomous today. (i.e. they could rend control from a driver and continue to pilot the vehicle to an alternate destination if hacked)

This is very much not true of humans.

> Vastly decreased likelihood of being tricked by vandals exploiting flaws in recognition algorithms

This is not particularly true of humans. The human eye is very easily tricked. It's true that there are things that would not fool us and would fool a computer, but the inverse is also true.

What in the last few millions of years caused humans to evolve to steer metal objects at 60mph?

For extra credit: describe what the reaction time is for Waymo's driverless cars are during various points of the garbage collection cycle.

People here have a tendency to imagine an idealized version of driverless cars and assume that actual driverless cars are the idealized versions of those cars, and then compare the idealized cars to very non-idealized humans.

I think that the real world is going to be more complicated than that.

It is easy to write code that takes a non-trivial amount of time to process to the point of output on any given hardware. I imagine that all of us have done that at various times. We can certainly imagine that reaction time could be a virtue of driverless cars -- even 99th percentile reaction time. Given sufficient hardware (and not just hardware on the core CPU/ram chips -- this is hardware out to the sensors and the subprocessors associated with each of them).

Similarly, we could imagine putting enough sensors on the vehicle that it truly has 360 degree awareness and visibility. But we do see real cars in the real world that have surprising blindspots.

There are cost tradeoffs to all of these things, and complexity tradeoffs. Is there a happy medium? I think that there very likely is, even without any fundamental advances in the state of the art. Are we within a year or two of that happy medium? Maybe.

>For extra credit: describe what the reaction time is for Waymo's driverless cars are during various points of the garbage collection cycle.

Aight, I'm not familiar with Waymo's cars specifically, but with <10K in hardware (maybe <5K now) you can design and build an autonomous vehicle kit that has a reaction loop that runs at a soft floor of 15 adjustments per second. I'm sure that with better hardware you can increase that to 60+, and 15 per second is already about 5x better than human reaction time, and fast enough that it can react to changes every 6 feet at highway speeds, and every foot at neighborhood speeds. A 60 aps loop means that its recalculating everything every 2 feet your car travels on the highway.

And yes, this isn't incredibly difficult.

For reference, in my experience, running 2 cameras and a low power lidar (along with associated outputs to hardware actuators and such) off of a 2013 (or maybe a 2011) macbook pro, the limiting factor was always the lidar's 15fps framerate.

you can say the same thing for rockets. There is entire field dedicated to this, its called Realtime Systems.

I don't think it's controversial to say that machines can easily beat those human reaction times.

Most or all autonomous vehicle software is written in C++, so garbage collection shouldn't be an issue. Of course, these programs need to be tested carefully and extremely before we trust them.

Not sure I see the point you are making... all cars have blind spots with drivers now. Human drivers are always blind to the spots they are not looking at the moment anyway. I can imagine cars being much safer than they are now.

But when you look at how you imagine driverless cars to be, you might want to temper your imagination with how driverless cars actually behave, such as the Tesla running into a trailer because it was too high off the ground for its ultrasound sensors to detect.

http://www.theverge.com/2016/5/11/11656496/tesla-model-s-aut...

(Note that this is not the famous case of the Tesla being unable to detect a white semi against a bright sky, it's a low speed no-injuries collision).

Obviously, Tesla is not the be-all and end-all of driverless cars. But when you're dealing with sensors combined with AI where each is in many ways less capable than the human eye, with different failure modes, it gets expensive and difficult to design really good fields of vision.