It sounds like your reasoning goes like this: my interactions with things which are supposedly AI are bad. Self-driving cars are also some form of AI, therefore self-driving cars must be bad and or far away too.
Did I understood you right?
The problem with that logic is that there is nothing in common in implementation/architecture/incentives between the things you mention and self-driving technology.
Nobody, well nearly nobody, tries to implement self-driving cars in a blackbox “AI” fashion. What I mean is that we don’t just throw sensor data at a neural network, then squint and say “i recon it is going to drive well now”. That would be madness.
Most approaches break down the problem into sub-problems covered by sub-systems. The sub-systems are fed information with known error properties and engineered to the specifications. The failure modes are painstakingly traced through and documented. Then in turn assemblies of these subsystems and the whole are reasoned similarly. Fault trees are drawn, the operational domain is considered. The reasoning why the engineers think the system is safe and have the right redundancies in place is more complicated than the code itself.
Some of these sub-systems are implemented using what one would call “AI”. Particularly in the object recognition domain that seems to be the state of the art. But the failure modes and shortcomings of these systems are considered and reasoned about the same way you would do the same with a good old-fashioned kalman-filter based sub-system. It is known that they are going to fail in various situations in various ways. The trick is to engineer the whole system such that it still remains safe despite these sub-systems having these characteristics.
I’m not saying that we will have safe self-driving soon. What i’m saying is that you can’t reason about self-driving cars by saying “commercial entity X is spamming me with bad marketing crap. People talk about AI behind said marketing crap. Therefore self driving cars are far away.”
My personal definition of what constitutes "AI" includes emergent behavior as a prerequisite. To lack emergence is to be a programmable logic controller.
Do these systems you're talking about exhibit emergence?
If so, how can you possibly reason about their failure modes?
What is your definition of emergence? I found this one on wikipedia:
“In philosophy, systems theory, science, and art, emergence occurs when an entity is observed to have properties its parts do not have on their own, properties or behaviors which emerge only when the parts interact in a wider whole.”
If i go by this then yes such a system exhibits emergence. The whole has the property of driving the car while the parts alone don’t. But so does an elevator: neither the winch, the cable, or the shaft lifts people up alone, but the whole of it together does.
> If so, how can you possibly reason about their failure modes?
Same way you do it about an elevator? Its “just” much more complex because the operational domain is much more complex.
Purported AI is not.
The self driving car won't be human like - merely a set of algorithms using trained networks to solve problems.
Like GitHub Co-Pilot reurgitating John Carmack's source code. Also, stochastic language models have reached the point where they seem like magic, e.g. automated YouTube subtitles and translations, or text to speech. But that all works by memorizing word fragments (CTC) and/or memorizing sentence fragments (n-Gram).
Yep. I marvel at the "Machine Learning AI ad algorithms" that try to sell me junk. I get 3 (three!) emails a day from aliexpress about stuff i already bought ... from aliexpress. Same thing with Amazon. I bought a, lets say "toaster" from amazon, and for the next 6 months every weird amazon ad on the wide internet and every time i hit the amazon home page it's trying to sell me toasters.
I don't even see it like "it's like a child, it finds something interesting and repeats it" because even a child gives up after a week. maybe a lot of people don't see "AI" as "old tech" - even though some of the first real "advances" - and i say that with a wry smile - were documented 40 years ago or more.
It's complete and utter snake oil. There are some cute toys that AI/ML can do, i can clone certain voices passably, for instance. There's "deepfakes" where a screenshot is enough to maybe fool someone not wearing their glasses at 3 meters, on a smart phone screen. There's obviously the AI in video games, etc.
But we're a good 25 years out from a truly autonomous, road-worthy, self-driving human transport.
Someone will always have to "look over the shoulder" until then.
The day i can hire a private self-driving taxi/limo to pick me up at my house and drive me to an arbitrary location, at any time of day or year, is the day i start to tinfoil my house to prevent the grey goo from coming inside.
edit: i should note that i would ab-so-loutely love a car service that was cheaper than hiring a limo for a week, if you can even do that. If the car is driving by itself, i can sleep while it gets from interesting place in the US to interesting place, and spend most of my time awake documenting and photographing and interviewing people, rather than hating interstate 10, 20, and 40; which is how most of my "road trips" for the past 20 years have gone. I've traveled this country "a lot", and i rarely have cute stories about places i've been to because i've either been completely wiped and just slept there somewhere, or i was in a hurry to beat the sun/state police/whatever through the state. I'd love to spend a few weeks in wyoming and montana and wisconsin, but if i have to drive it, the only thing i'll publish is a couple of shots of mountains, the kids posing in front of some tourist trap, and that's about it.
I passively consume, yet nearly universally ignore and eschew all "AI" news.