They seem very convinced, or at least are pretending to be convinced, that they can build FSD without human-level AI. I'm not convinced of that at all, though.
Here's a few skills of my brain not needed for driving,
- compose poetry and alternate lyrics to "baby shark"
- recognize emotions on faces
- form opinions about music, movies, video games, etc.
- have a conversation
- remember that time I embarrassed myself in sixth grade
- contemplate my own existence
- divide other people into friend/enemy groups
- etc...
I'd argue this is a required skill occasionally, at least when you find yourself in a 4-way stop deadlock.
Case in point, how does a FSD vehicle deal with extremely heavy snowfall on a unplowed road? I spent 15 minutes rocking the car back and forth, so I could slowly back my way out of a one-way road onto a main street. After that I had to drive on countless roads with no visible markings due to the snow. I don't think the way we're approaching self driving cars today will ever work for these edge cases.
Driving is a pretty well defined space. Stay in your lane, get to the destination, and don't hit anyone. Pretty much every similar problem that significant effort has been solved. Chess, Go, Starcraft, DOTA all were called unsolvable at one point and all were eventually solved.
Your edge cases are just that. Edge cases. Nobody is going to care if they can't drive in extremely heavy snow. To begin with, they probably shouldn't do that anyways. More importantly the vast benefits of self driving cars outweighs the inconvenience of not being able to drive a dozen times a year. We all still have cell phones even though you can't make calls in elevators.
Even then, if you can do it why can't a computer? Rocking the car back and forth seems doable. Following an unpaved road is much harder, but then again are humans actually good at it either?
You might be right that they will still have steering wheels for no other reason than there's little downside to including them.
That's important, since if we are using it as an analogy for driving, it's not just the win rate we're interested in, it's the ability to consistently avoid pathological states when presented with edge cases (e.g. not braking when there's a strange looking object in front). Some would call this "common sense".
Also, games like SC2 are distinct from the driving problem as far as RL goes, since you have unlimited training data which allows you to explore all edge cases given enough compute. In self-driving, the data is scarce and you can't explore all edge cases by just throwing compute at the problem.
These just need to be relatively rare and not dangerous when they happen. If my car slams on its brakes every 100,000 miles for no good reason that's not really a big deal. At least it's not big enough to stop them from being used.
>unlimited training data
Unlimited is not possible, but we can get tens of millions of miles pretty easily. Then many many more as they roll out into production.
My go-to example: you're driving at night, two-lane road, and there's a minor accident in your lane ahead. You crawl forward to the point where you can pass the accident, and then what?
Someone involved in the accident is directing traffic. In the dark, with no safety gear. On the other side the traffic is backed up, so you see headlights facing you in the lane you have to use to pass the accident.
To me, that seems like an insanely difficult problem for a computer to deal with. A human driver, that's easy, but a computer?
(Having said that, I'm historically terrible at predicting just how fast technology is going to advance, so it's safest to assume I'm wrong and this will turn out to be easy.)
The remote driver could also call the passenger's phone to double check that the sensor data is correct, and to reassure them that the AI isn't just guessing.
The problem when you have 1.4B vehicles in circulation is that your edge cases happen dozens of time per second.
> Nobody is going to care if they can't drive in extremely heavy snow.
A loooot of people will care. Snow isn't an edge case in a very large part of the world, it comes every year and stays for weeks or months.
What we're going to have is a system that's slightly better than existing "smart" systems [0] but much worse than "fully autonomous". People who will learn on these systems will suck at driving and be in complete panic mode at every single of the "edge cases" they'll face very often
People have to understand that the PCH from SF to LA and quiet Mountain View streets are the edge case to the rest of the world.
[0] ESP/ABS, collision avoidance, blind spot warnings, auto emergency braking
Depends on what you consider 'solved'. AlphaStar was never able to demonstrate the level of superhuman performance in Starcraft that AlphaGo did. It got to grandmaster level, but even that was with significant mechanical advantages over humans, when what they wanted to test was the tactical/strategic thinking, not raw ability to issue basic commands (which an AI will easily win at, obviously).
That said, AlphaStar was vastly superior in performance and well-roundedness compared to any previous Starcraft AI, so it was still very impressive. It just wasn't "beating the best in the world at strategic thinking" impressive.
Though I'd like to see AlphaStar perform with comparable mechanical limitations to humans as well. That's a complete non-issue for Go or Chess, but for Starcraft it matters a lot.
90% of the cars on my street didn’t try to deal with that this weekend, and most of them would have failed if they tried. None of them are FSD.
1. Agree on the edge cases, to do real driving we'll need AGI or some whole new way of looking at it, its not possible algorithmically. But, I would pessimisticly draw on other examples of how tech handles this problem: if the tech can't do it, it's not something people get to do. Try getting help from a chatbot, or getting help from uber or Amazon on a situation that doesnt fit the FAQ. So I expect self driving cars will be great for a range of things that young tech-savvy people want to do (and that google et al want you doing) but won't drive you to the hospital in a snowstorm and will balk when they realize they have to turn into a rutted backroad at night, because why would anyone live there?
2. More frightening, there are people (see one of the siblings) that think nobody should be doing "dangerous" things like driving in a severe storm, that are beyond the range of a self driving car, and want to force their value judgement on others. So expect there to be calls for outlawing certain activities, for our safety.
In a few years, giving the speed of sensor development and growth of computing power, the automated driving system can continuously capture and monitor 360 degree environment, at better resolution than human and across the specter of wavelengths (visual, IR, T/GHz) and react better/faster than human. Given that, i think human driven cars will be relegated in couple decades to the position of horse carriages - not allowed on highway for example where automated cars will be driving faster at tighter distances while being networked to each other in the vicinity. In some sense the driving may happen to be best suited for insect style distributed intelligence with relatively low intelligence of any given node which still possesses heightened senses. I wonder whether Google cars communicate to each other when meet at an intersection - either option has far reaching consequences and will define our near-term future.
Plus at some point insurance industry will say their word which most probably will be based on the sole metric of $/mile loss, not AI IQ score.
On the other hand, if we set it up well, each car will learn from the experiences of every car, so novel situations will (ideally) only be novel once.
R.I.P. Elaine Herzberg