Bottom line: Real, reliable, "full self driving" automobiles are unlikely to be achieved using current approaches. About the best that can be expected is a marketing gimmick that is sure to cause lots of accidents.
You might get a few false trips if a trash bag blows in the road and the rear sensor says nothing will hit you if you stop, but who cares? If we really prioritize safety, we'll just accept that, and find other ways to save time.
A self driving car would essentially give a person extra hours in the day, and allow people with disabilities to go outside of major cities unassisted.
I also have tesla’s fsd beta on my car, and it does a marvelous job not hitting people. If anything it’s way too cautious around pedestrians.
I recently had a situation where a large truck off to the right in front of me had a blowout and a large chunk of tire flew over into my lane. There were other cars traveling fast on both sides and behind me.
What to do? I chose to maintain my speed and lane and just hit the tire. It caused some damage to the front of my car but in my split second judgment, this was the least bad option.
What would your AI do?
It knows stuff is happening to the sides and swerving is bad.
Unlike a human, it can slow down to hit the object at minimum speed, then speed up again just in time to not get hit even if the driver behind does nothing, at an optimized profile that gives the person behind time to not overreact and cause a pileup.
Self driving also has the advantage that people might stay away. Unlike regular tailgating you have machinephobia that might actually make people not tail you.
Yes, this is undoubtedly the human drivers fault for following too closely but ... is it being precipitated by unexpected AI braking? In any case, the AI has demonstrated very limited ability to avoid these sort of collisions.
In any case, some of those cases were probably due to lack of rear sensors, or lack of any sensors at all besides cameras, and lack of any deterministic code able to override the AI.
But I've got no idea how practically usable such an approach would be.
*Something* has to interpret the data from the sensors. This is where logic and reasoning comes into play.
If a bird flies across the road in front of you as you're traveling at high speed on the interstate, slamming on the brakes at that particular moment might not be the most appropriate response.
A self driving car would essentially give a person extra hours in the day
No one is disputing the utility --- only the difficulty of actually achieving it.
We already have self driving cars in public streets that work. Arguing about the possibilities with self driving is like arguing about how good chess computers might some day become, it’s already been done.
Hitting the right price point, actually mapping the environment at scale, working out reliability, and bugs etc is real engineering. We are in that stage right between the prototype works and it’s ready for production which always takes longer than people expect but can it be done is no longer in question.
"but can it be done is no longer in question."
You are right - it doesn't seem like it can be done at all, not in the way that most people imagine "self driving" - where you hop in your car no matter the weather, tell it where to go and it just takes you there while you sleep, and it works in every country in the world on every road like any human can drive.That doesn't seem possible this century, if ever.
Waymo, Cruise, and a few other companies are sending out empty cars to pick people up. They have off site people to remotely take over if the cars get confused, but such handoffs are in no way fast enough to deal with a safety issue and can easily scale to have thousands of cars per backup driver based on disengagement statistics.
This is the same way automation has always occurred. You get the machine to deal with 90%, 99%, then 99.9% etc of cases until people move from doing job X, to being backup, to troubleshooting problems.
>>This is the same way automation has always occurred. You get the machine to deal with 90%, 99%, then 99.9% etc of cases until people move from doing job X, to being backup, to troubleshooting problems.
Of course. We are now at maybe 20-30% coverage though. Once we are at 90%, maybe in 30-40 years, then we can start saying things like "we have self driving cars".
The point of these trials is to demonstrate safety not simply test limits. Waymo cars work fine in conditions and areas outside of their current tests.
So if you took an existing Waymo car and dropped it in a middle of a previously unmapped city, it would be "fine"?
PS: I am ignoring sanity checks that would need to be removed. They are setup to be geofenced so it’s going to reject any destination outside of those until someone updates a confit file. That’s just good engineering not some abstract limitation.
Neither of these is capable of being placed into a random environment and actually "driving" from environmental feedback with minimal directions (turn left, go right, stop) --- which most any human driver can easily do.
Really it’s little different than many modern drivers being dependent on the GPS for long trips.
Once you map all the roads and account for very imaginable situation and scenario --- the unexpected and unanticipated will still occur.
Mapping isn’t about unexpected situations beyond the loss of GPS.
"Just like people do" is only true if you see 7 year olds driving around, I imagine they would have similar capacity to make decisions on the road as Tesla FSD does. Except I imagine even a 7 year old wouldn't just let go of the wheel completely.
The use case of self driving taxi already exist as in you tell an app where you want to go and an empty car with AI driver shows up. The app doesn’t always summon an AI driver, but when humans and machines are doing the same job don’t expect the humans to win indefinitely.
Extract from [1]:
The basic navigation problem has been "solved" for decades. Back in Carnegie Mellon University in 1990, I watched a NavLab van rolling down a trail in Schenley Park behind my graduate student office. Since then, every few years we've been hearing announcements about how self-driving cars were just around the corner.
But between demos and actual products lies a huge chasm. Today, not only do self-driving cars still not exist after billions of dollars and decades of investment, but the experts admit that they are at least a decade away.
And no, you don't get to nit-pick what "self-driving car" means. It still means that I can get in, tell it to go to a destination that I would normally drive to, and then I can go to sleep. That's what's known in the trade as "Level 5" autonomy. Anything less than that is not the life-changing promise that attracts all this investment and press.
If your car can operate only on a few roads, then it's no better than one of those people-mover trains at airports. Those are driverless, too.
[1] https://www.robotsinplainenglish.com/e/2021-08-29-humanoid.h...
Coverage area doesn’t somehow require intelligence and these systems operate on far more than a few roads. The same hardware and software could operate on every road in the US at which it would ‘suddenly be intelligent’ even though nothing changed.
Look people always say whatever AI systems can do isn’t intelligent as soon as they can do it. Chess used to be considered a bastion of human supremacy over machines, then go, now the last gasps of self driving.
In 20 years people will say self driving cars aren’t intelligent because they don’t work in North Korea, even as young slowly people stop learning how to drive the same way humanity largely forgot how to drive a horse and buggy.
Coverage area is required for practical self-driving solutions, and today's ML based efforts are not close to producing it.
If you measure intelligence in this context by the ability to drive on any roads anywhere, then humans are beating the crap out of "AI".
So, I guess by your metric they have human level intelligence. Or more likely, people will soon think self require minimal intelligence and we will go to the next bastion of human supremacy.
Again, you’re mixing up “could drive” with actual ability to reliably take me where I want to go. Yeah, NavLab drove from Pittsburgh to San Diego in 1995. That doesn’t mean we had self-driving cars.
And no, Waymo’s limitation is not just legal. Where are you getting that impression from? Sure, they don’t have permits to do other routes, but there’s a good reason for that: they cannot stand behind their product except within their geofence.
I am not saying intelligence equals self driving. I am saying that even if you define it that way, we are far, far away still.
That’s got nothing to do with if the technology exists. Supersonic aircraft for example exist but you can’t book a supersonic flight NYC to LA. What you’re arguing is affordable mass market self driving cars don’t exist which is a strawman that has little to do with if self driving cars exist.
> Waymo’s limitation is not just legal.
Waymo is currently testing driverless cars in every single state (CA, NV, FL) it’s legal for them to test in. Clearly that does suggest the limitation to add any more states is legal.
The problem is the insane amount of training needed. The current approach of running repeated learning iterations to iron out corner cases soon runs up against the long tail of thousands and thousands of tiny, unlikely-but-possible scenarios requiring more and more data. Every time you expand the scope of the self-driving car (different roadside conditions, behaviors of pedestrians and other drivers, new roads, weather, etc.), you discover more conditions for the model to train on.
You can't just say, "my model is intelligent. Now it's just an exercise for the reader to train the model on all possible scenarios, and voila! You have a self-driving system!". You have to actually train the model to produce a self-driving car.
This is a never-ending battle. Let me point out how utterly different it is from how human drivers deal with new conditions.
No, self-driving cars don't exist. Not yet. And it's not for legal reasons.
(One technical solution to the self-driving problem would be to also modify all the roads and surroundings to reduce the number of unlikely scenarios that our models have to learn. But this kind of solution is challenging for non-software reasons, plus it would not really be "intelligent" in terms of AGI. It would be more like a tramway. It wouldn't be sexy.)
The self driving car you want to buy doesn’t exist, but they are already providing a valuable service. Saying well VC’s want to replace all Taxi drivers and all Long haul truckers and all … just means that development will continue but that’s true of most products. We hardly say the CPU doesn’t exist because Intel and AMD are still spending billions on R&D. All existing self driving car investments are really just peanuts compared to what companies will spend as the market matures and profits roll in.
These companies are avoiding bad weather for the same reason human drivers do, they simply have zero reason to take risks.
The question that remains is whether it *can be done* in an economical manner that is at least as reliable as a human driver. Anything less is a failure in terms of money and lives impacted or lost.
Two options for the former are: brake or swerve. While for the latter a 3rd option, accelerate, can also be in play.
Which do you pick when? When you brake or swerve or accelerate, what other environmental factors (e.g. wet road, gravel road, construction workers present), should be taken into account?
What about an axe that falls off the landscaping truck in front of you? Or a mattress? Or a harmless styrofoam cooler? The self-driving car does not know what those things are, but suddenly there is <something> in the air that will likely collide with you. The computer is going to be unable to predict how the mattress or the axe or the styrofoam will fly through the air, it can only decide to take abrupt, evasive action. It is also unaware that the person behind the car has been periodically looking down at their phone instead of watching the road, so an abrupt swerve or stop may still cause an accident. A sensible human driver would realize they are taking on additional risk by following such a truck and/or remaining in front of the distracted driver, and maybe decide to change lanes safely. The car's pretend AI has no idea of any of these things until something falls off and it has to react. We'll praise it when it gets it right--ooooh how ingenious! And the apologists will claim "there's no way it could have made a perfect decision--look how many other times it gets it right!" when it fails. And the realists will conclude "ha, stupid computer, told ya so".
They might make really bad decisions in edge cases(Which will get less and less common as more cars get smart), but they might make up for it with perfect behavior in ordinary circumstances.
They will never prioritize convenience or speed or avoiding angering other people over safety. They'll do the safe thing even if no human driver could maintain that level of paranoia at all times.
Defense driving is all about making these sort of judgments --- and choosing the least bad option when there are no good ones.
It could be argued that there are already ML systems that do what you describe (i.e. make instant effective judgements/decisions about something it hasn't been exposed to during training), at least for some contexts. For more information: https://en.wikipedia.org/wiki/Zero-shot_learning
Current methods are far from perfect, but not for the reasons Marcus believes. Consider that we /should/ be able to derive hypotheses for better system behavior from a given critique, test those proposed system improvements, and then (if the hypothesis is correct) see some improvement. Marcus' crew hasn't produced any useful improvements; this is either because they are not technically competent or their hypotheses are wrong. Given that people have been hammering at these systems for ~10 years now, I doubt the failure to build better logic-based systems (or hybrid systems) is for lack of trying.
Meanwhile, the non-Marcus groups have been improving systems by leaps and bounds, through exactly the kind of iterative, hypothesis-driven improvement I described above. Certainly nothing is perfect, but we can see real progress happening.
(Finally; Tesla FSD is a strawman. It's Waymo you need to argue with, and they're doing great, AFAICT.)
Let's disperse a few thousand of them at random locations across the country for everyday use and see what happens.
AFAICT, no one has actually done this yet.
There's a more interesting conversation to be had here than the one you're offering, in any case. And you haven't addressed the total failure of the Marcus crew to actually produce working solutions...