Hacker News Reader

> knowing how the sausage is made, seeing the compromises that are made in order to "hit the ship date", the bugs that are swept under the rug, the boundary conditions that are assumed away as "impossible", etc., I won't be trusting any self-driving car anytime soon.

As a general member of the public who is present on the roads, you have no choice but to trust any Alpha- or Beta-level self-driving technology that certain companies thrust out on the roads today.

I think we're at least one CS breakthrough away, and how do you put a timeline on that?

Kinda like how speech recognition was just almost there for decades until modern ML made it actually as good as a human being.

> ML made it actually as good as a human being

How ML made it usable in some cases, would be more accurate. I still can't ask Alexa to play specific albums or tracks most of the time.

I don't want that level of accuracy driving cars on the road.

Google seems to have no trouble playing music from Spotify for me. It can even manage it while I yell my song requests from the shower.

The only thing I have ever wanted from phone assistants is to tell me whether a coffee shop was open or not, or what the closing time is. I still can't get this info, nor is this question understood. I still have to stop, launch a browser, navigate to the site, and dig around their website for opening hours, or in the worst case find a phone number and call to ask them.

What phone assistants are you using? I’ve never had this problem.

Siri

Same here, Siri solves this problem.

Me: “Hey Siri, how late is x open?”

Siri: One possibility is on x street, does that sound good?

Me: Yes

Siri:

“X is open from y-z.”

I just tried it and you're right, it works. Well, it didn't used to work so I stopped using Siri. I regret the error.

I wonder if people assume Siri failed to understand the query when they hear its request for clarification. I think it should include some context, eg: “Which x do you want opening hours for?”

The one thing that I can consistently get google assistant to do is to set timers: "Ok google set timer for 5 minutes". I want more of that. Could they please just implement and document simple recipes like this? Setting a timer hands-free is already incredibly convenient, and I can only imagine making something similar for a wide host of tasks isn't /that/ hard. I really don't mind if I have to formulate it in a rather specific way, as long as it works and I can find out what that way is.

The weird names and track titles that dominate electronic music has made this quite difficult for me.

Better GPUs is what enabled modern ML, it wasn't a CS breakthrough.

I dunno. I'd say new classification strategies like K-Means* really did it. Without that, all the power in the world doesn't get you anywhere.

* https://www-cs.stanford.edu/~acoates/papers/coatesng_nntot20...

I agree.

In my various conversations on this topic around the bay area, people are generally overly optimistic, who believe it'll happen in next couple or so years, and people who are overly pessimistic.

Possibly a gross generalization but generally the people who are overly pessimistic are the ones who are or have till recently been engineers, while the people who are overly optimistic are the ones further away from engineering - with the farther they are the more optimistic they are.

Moreover, in addition to what you wrote there are a lot of other open questions, not necessarily in any order and definitely not a complete list:

1) Auto Insurance (cars below refers to fully autonomous cars):

1.1) What happens to auto insurance? Do I still need to buy it if I'm not driving it, or does it come bundled in the car price?

1.2) Does everyone pay the same auto insurance for the same car model/year.

1.3) Do cars with different autonomous models have different auto insurance?

1.4) What happens in the case of car accident between self-driving and non-self driving car?

1.5) What happens in the case of two self-driving cars' accident

1.6) What happens if a person is killed due to the Car's fault. Who is liable? The insurance company? The Car manufacturer? The self-driving module manufacturer?

2) Some Edge cases

2.1) What happens if on a 4-way stop 4 self-driving cars come to stop at the same time? Do they know how to handle that, as in which car is moving forward and which one is allowing other cars?

2.2) What happens in an emergency? Can I force my car to drive rash to the hospital or will it follow the speed limit to the T.

2.3) What happens if the Car runs into a situation which it is not trained for?

3) How are the cars tested?

3.1) As a consumer do we know that the cars are fully tested to handle ALL edge conditions?

3.1.1) What is the definition/definitive list of ALL edge conditions?

3.2) When will we know that the threshold has been passed, and the car is truly autonomous?

4) Security

4.1) Assuming the cars will need to be connected, how will the security be?

4.2) Can the cars be Pwned? I'm sure they can be.... so this is the rhetorical question

4.2.2) If the cars can be pwned, will the fully autonomous cars become the next WMD (Weapon of Mass Destruction) -- if you don't agree, close your eyes and imagine a fleet of hundreds of thousands or millions of fully autonomous cars under the control of a rogue entity hell bent on causing damage/killing people.

4.2.3) How do we protect against that?

1) Car insurance can work exactly like it currently does. The car owner's insurance is liable by default, however if the car itself is in some way at fault (performing out-of-spec), then the insurance company goes after the car company. This can already happen if something like brakes fail or accelerator gets stuck.

2 & 3) Yes there are some edge cases that you can't train for, but same with humans. What do humans do in an edge case like a torrential downpour? Simply pull over.

Even if there are edge cases, consider that humans are pretty bad drivers on average, over 1.25 million deaths per year, and 50 million injuries in the US alone.

Suppose that in a couple years, self driving cars are not perfect due to the edge cases, and still cause 250k deaths / year. Would you allow them at that point? Even with those stats we should consider only allowing self-driving cars and begin to outlaw human drivers.

Car insurance: I had made a simplified argument, but to itemize, auto insurance may work the same for auto-theft and some liability coverage. But, it definitely cannot work the same for anything remotely related to driver (or in this case self-driving system's) fault.

Edge Cases: Humans use their brain and adapt. Self driving systems have no concept of adapting, hence the result is undefined.

> over 1.25 million deaths per year, and 50 million injuries in the US alone.

In US in 2016 there were 36,750 fatalities[0][1], so much much lower than the 1.25 million number you're using.

So, no, personally I will not prefer self-driving cars if they are not perfect due to the edge cases, and still cause 250k deaths / year

And the reason is simple, and why throwing statistics is now the right approach: I'd rather have my life in my hands and my driving where I can take a corrective defensive action than rely on an AI model where it is unpredictable when the lightning will strike and a bug/untrained edge case will kill me. I will prefer to be in control.

[0] https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in... [1] https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/...

> But, it definitely cannot work the same for anything remotely related to driver (or in this case self-driving system's) fault.

Why not?

Because I'm not the one driving the car, I'm merely sitting in it.

Think of it this way, do you take auto insurance for a cab ride, no you don't, because you're not driving it.

That all depends on the contract with whatever service your using. It would be totally legal for the cab driver to make you sign a contract making you liable for any damages. I'd be very unlikely to sign for a cab ride, but I might if I had my own self driving car depending on how reliable they were.

Anything is possible.

My topmost comment said these are open questions, as in undefined.

I, as in IMHO, highly doubt it would be like you're describing, but it is equally likely that it may be like how you're describing, it's just an open question.

In fact, a couple of years ago there was a small event at Tim Draper's Draper University in San Mateo about how auto insurance would look in the world of self-driving cars and it was very clear that at least till that point there was no consensus. I haven't seen anything to suggest if that has changed since then.

Thanks for the correction, the 1.25 million number is worldwide, not US https://www.asirt.org/safe-travel/road-safety-facts/

Self driving systems however do adapt. Most companies are using neural networks which can be retrained based on their mistakes. Unlike a human, they can also be trained on each and every crash for which sensor data is available.

They cannot "adapt". Your "retraining" is not a trivial process. Attempts to accommodate edge cases are not guaranteed to be able to cleanly accommodate all previous cases without discontinuities/regression/overfitting. The consequences of which could be especially catastrophic given that every other vehicle that safely operates may be relying on a particular quirk of a particular version of the neural network's particular function which it is imitating to safely operate in it's particular environment.

That's the problem with NN's. You can't say anything about them explicitly except that they meet the specification for their training datasets. Fitness for a task outside of that is not guaranteed. Humans, on the other hand, are constantly adapting, because they never stop "learning". This has the inverse effect in that you can say that a person can be trained for a task, but at the moment may not be.

When you apply an NN on an industrial scale, you can't just throw it out there without being willing to accept the liability for the destruction you cause as a result.