Self Driving Cars, a Cautionary AI Tale

[Physics Hunter]

I like Elon Musk as much as the next ULTRA MAGA dude, but he has been consistently wildly optimistic about progress in AI.

I will open with: I worked a lot in AI, writing algorithms and engines, and creating complete systems.  I also have a Grad level sheepskin that includes the words "Artificial Intelligence".

Let's start with Self Driving Cars...

Back in about 2013, Elon was very optimistic about Tesla's with Self Driving, saying it was a couple years off.  As detailed in this retrospective from Jalopnik (an automotive press site), we are 11 years out from those optimistic predictions, and Self Drive is nowhere to be found unless it is slow and non-lethal (food delivery drones).

https://jalopnik.com/elon-musk-tesla-sel...1850432357

I think that Jalopnik is a little too hard on Musk.  AI, and real science and engineering are difficult subjects.  You can use 'proven" "optimal" theorems and laws all day, but if you step a pinkie toe outside the limits of the derivation of those equations, approaches, or maxims you are gonna lose your entire foot.

The problem with Self Driving Car is that the set of challenges that the system must face is not closed.  (Does this sound like an unsolvable NP-Complete problem?  That is too geeky if you are not a scientist)  Just suffice it to say that weird stuff happens on the road, frequently.  Do I hit the skunk/deer/moose, shovel, refrigerator, or do I ditch onto the shoulder; the only thing that informs this decision is common sense.    Artificial Common Sense is even more rare than the real thing, which if I had a 12 pack of I could change the world for the better.  

Hitting a skunk, or even running over a fridge is gonna suck.  If it is a possum, raccoon, or empty box, it is a different story.  You brake as much as you can, look at the other lane, and assess if the shoulder if flat or a sheer drop-off and pick the best option.

And you can teach your AI (and it is an AI system using tricks that we developed over the past 50+ years) self drive system all of that.  Next time it will be a kid chasing a dog, then what do you do?

We expect a lot from AI systems, too much.  We all use them every day.  If you want a list, feel free to ask.  If you use voice recognition to reply to this, don't ask.

[Billy_Bob Patriot Master]

Modeling a chaotic and unpredictable system is not possible. Just ask the climate modeling people.. They have been at it for over 50 years and still everything fails without question.

[old dog]

AI learns through errors.  America's highways are not the place to conduct that learning process.

[Sunsettommy]

Quoting Physics Hunter,

I like Elon Musk as much as the next ULTRA MAGA dude, but he has been consistently wildly optimistic about progress in AI.

His big mistake was not doing exhaustive product testing to see how it performs under duress, should have taken his time in bringing out a brand-new concept with no prior history of effectiveness as the potential failure is catastrophic to flesh and blood riding in those cars that drives for you.

Because of crashes and lives lost trust in them has evaporated, why would anyone still buy proven unreliable and dangerous protypes anymore when your survival rate in them drops greatly in comparison to standard human based driving.

Meanwhile I would never let go of the steering wheel as I like to drive which keeps me active mentally.

[Matt]

I don't trust self driving cars and I'm not sure why we would want them. If I'm in a vehicle going at high rates of speeds I want to be fully in control of it. I don't want to trust my life to a computer program that may or may not function as intended. Also, I happen to like driving.

[Billy_Bob Patriot Master]

AI learns through errors.  America's highways are not the place to conduct that learning process.

I would prefer not to be "one of those errors"!

[Physics Hunter]

I don't trust self driving cars and I'm not sure why we would want them. If I'm in a vehicle going at high rates of speeds I want to be fully in control of it. I don't want to trust my life to a computer program that may or may not function as intended. Also, I happen to like driving.

Seems to be a lot of us here that like to drive!   

I own two muscle cars, that  I don't really need, and they are both manual transmission.  I just like them...

---

AI learns through errors.  America's highways are not the place to conduct that learning process.

Actually most current AI systems do not learn in the wild.  They are trained, limited, and tested in the factory and fixed in place until there is an update.  AI systems that are allowed to learn without specific guidance tend to go off the rails rapidly, reference no common sense.  This gets the lawyers out of sorts...