Again, I’ve done similar things, and I’m just some schmoe who likes to drive. I have to pay to do it, no one is paying me. Yes, it does help that I’ve been in that situation before, but I haven’t really practiced it for thousands of hours. I don’t want the AI driver to be better than Mark Higgins, but I’d prefer it to be at least as good as I am when paying attention.
Ehh, its instinct, but thought is also involved. Someone who was working only on instinct would most likely over correct and still end up off the road.
Another example of humans reacting in sub-second time frames would be at the drag strip. The quickest you can react to the yellow lights going out and get the car moving is .400 seconds. The timing of the amber light being lit is randomized, so you can’t really anticipate it, you have to wait for it to go out. Even my worst reaction times have been below .600 seconds.
Nothing “complex” is going on for those sub-second decisions, even in the best-case human scenario. The self-driving systems are already better at humans at emergency braking, and for something like scabpicker’s example, dumb traction control systems are already better than a human could ever do, because it has individual wheel control.
You don’t need an LLM to vastly exceed human capabilities here. You might need them for some high-level decision making, but that can take seconds.
And that’s an example where all of your attention is focused on one simple thing, which is to hit the gas pedal as soon as the light turns yellow. The absolute best-case scenario, with no thought required, is still 400 milliseconds.
That is not something you need an LLM for. Quick, dumb reactions don’t need that level of analysis.
Yes, which is why drivers are completely awful at this kind of thing and often end up off the road when they encounter something like a deer.
Practice is training your instinct so that you don’t do the dumb thing when it takes over. Which it will, when events are unfolding rapidly.
Adam from Mythbusters brings this up all the time on his channel. I forget the details, but recently he described an incident on the show where he did the right thing–but didn’t even remember making the decision afterwards, because it was pure reaction. It’s just that he had trained himself in advance to do the right thing. If he’d gone in blind, there wouldn’t have been nearly enough time to figure out the right thing and do that.
I’m actually surprised to see you argue this, since I know you’ve written before about System 1 vs. System 2 thinking. Absolutely no System 2 thinking can go on in less than a second, and yet that’s all you would ever need an LLM for (or some variation). Simpler AI architectures do fine with System 1 thinking, since it’s an automatic stimulus->response action.
I don’t even fully buy into that distinction (it’s true at some level, but too binary), but regardless, nothing except the most basic actions take under a second while driving. And even that is only under very optimistic conditions.
Someone on Reddit posted this picture, asking whether we’ll need AGI for Level 5 self driving:
It’s a legit question. A series of signs like that confuse the current Tesla FSD software. But even a human will take a few seconds to parse the signs and figure out what the speed limit is. I plugged this prompt into ChatGPT and it gave me a perfect response:
I see these signs along a highway:
Speed Limit 70 Minimum 45
Buses speed limit 65
Speed limit 60 Trucks over 4 tons Motor Homes Campers Trailers
I’m driving a Tesla Model 3 and am not towing anything. What is the speed limit for me? Please explain your reasoning.
So yes, I think LLMs may well play a role, but only in a high-level advisory sense, in the same way that humans will engage System 2 thinking to work out what they need to do.
I imagine that for established road signs on well-traveled roads, the AI will be programmed with any special restrictions, so it won’t need to read confusing road signs.
Ideally. Unfortunately, speed limit databases are often wrong, and the car will still have to interpret situations like construction zones.
While databases might prove useful for making better decisions under uncertainty (what if there’s been no sign for a while?), I think that ultimately they’ll need to do at least a decent job at road sign interpretation. Might be a while before they solve this, though:
On the other hand, how many humans are able to work out the logic?
I imagine for temporary circumstances like construction zones, there could be a process to allow the Department of Transportation or the police inform the self-driving car manufacturer of the changes.
But there still is thought and perception happening. It’s more than a reflex. Also, you’re not jamming your foot to the floor, you hit the accelerator pretty hard and modulate it. In all but the slowest cars you’ve got to be feeling for whether the back end is getting loose and react to it if it is. If you do that on anything but a sub-second time scale, you’re probably just going to do a burnout. The 400 milliseconds is accounting for how long it takes the car to react to your inputs, not just the time of the human reaction. Some people are even faster. Bike riders often get under 200 milliseconds at places that will allow you to run a .000 light, mostly because the bike has less inertia to overcome.
For what it’s worth, the quickest reaction time for a NHRA pro is .039 seconds. That’s in a top fuel dragster, and since the whole race is over in 4 seconds, there’s not a ton of plans that get changed. But they surely sense, think and react during those four seconds.
Well, again, these awful drivers haven’t learned what to do in an adverse situation. Similar to the AI, they’d probably be winging it. Training doesn’t necessarily make your reactions a reflex, but you have at least gone through the motions, know how to identify the situation and have some idea what to do.
People can do all sorts of things automatically. I can touch type fairly rapidly, making complicated finger motions several times a second, translating thoughts into characters on the screen, but at no point am I thinking about the process. Put me in front of a keyboard with a different layout and I’d be lucky to get one character per second.
Ok, I can see from the wiki article on the Christmas tree (really I should know this since my dad drag raced) that the amber-green timing is fixed (400-500 ms). That’s not reaction time at work–that’s trying to nail a specific timing target. Totally different. I’m sure that highly practiced people (musicians, say) can get under 10 ms in those conditions.
When it comes to driving, we’re talking about events that aren’t anticipated. Like the car in front of you slamming on their brakes. Whether automatic or not, you have to decide what to do. Was it just a short blip, or is the car in front doing an emergency stop?
If you put people in one of several possible scenarios, without telling them in advance when they will happen (any time in the next several hours), or what kind of scenario it is (emergency stop? ladder in road? deer? etc.), I would be shocked if even 1% of people could handle that appropriately in under a second.
The timing between when the amber light goes off and the green turns on is fixed, that’s why you’re reacting to when the amber light goes off. But when the amber light goes on after you’ve staged and how long it’s lit before going off isn’t fixed. It was up to the guy running the tree in the past, but it was later randomized due to some folks figuring out the timing of certain christmas tree operators.
Now, there is certainly some level of timing involved. It’s randomized, but you can still kind of anticipate it if you really like risk. There’s No Prep Drag racing, which has the yellow lights removed, but the same timing set up. You are either going on green, or anticipating the light. Most people seem to try to go on green, because anticipating the light is a pretty big gamble. If you go frame-by-frame in this clip, Cali Chris’ car is moving three frames before the light goes green, but he doesn’t trigger the staging lights going off for about three frames after it does. The Firebrid doesn’t move at all until after the light is green, but it’s nose is lifting seven frames later, and the first staging light goes off about five frames after that (and he seems to have staged shallowly since the second staging light flickers). In this case, even the guy who’s waiting for the green has a reaction time far faster than a second.
Heh, but then I’m nothing special, I walk the earth, and I’d generally know what to do.
Interesting. Well, there’s certainly some strategy there when it comes to anticipating the light.
In any case, my point certainly wasn’t that humans are incapable of reacting in under a second. I’ve done reaction tests before; in fact, you can play around with one here: https://humanbenchmark.com/tests/reactiontime
I can usually get a bit below 200 ms. I seem to recall usually getting more like 150 ms in the past, which probably makes sense since reaction time decreases as we age. But that test is a pure visual stimulus → finger response. No interpretation has to take place. And even if you try not to anticipate it, to get a decent time you’ll probably have your finger already tensed and almost ready to click.
That is still not a real-world driving situation, where, even if you see some brake lights come on, you have to commit to some action. That is not necessarily slamming your own brakes on. There is some finite time, multiple hundreds of milliseconds at least, where you have to choose between several actions and then perform the required physical movement. You are still not thinking about it, but your brain has to process the information.
And that is still a best-case scenario, with an experienced, alert, undistracted driver. Which isn’t even true on average, let alone the majority of the time, for drivers on the road.
And even that was not the situation I was responding to Sam_Stone about. Within a span of a second, you might be able to perform one basic action, like applying the brakes or swerving. And let’s hope that your instincts are correct. But an LLM or other more general intelligence is not needed for these things. It may be needed for things like interpreting road signs, or properly responding to emergency vehicles, or even silly stuff like asking it to drive to the nearest burger joint that’s still open this late. But that can happen over multiple seconds, as it already does for humans.
The barrier i think they may never pass, is when power has gone out and a human is directing traffic. I’ve seen footage of a PO just standing in front of a driverless car keeping it stopped in front of a traffic tie-up. If the car could see him waving it into the next lane it would have been a 30 second problem. But as it was he was discussing with firefighters how to disable the car. A really awful situation.
(Apologies if this has been discussed. I went back 100 posts, but i don’t have it in me to read 1100.)
According to several of those tests, I am a sorry example of humanity.
Didn’t the Xbox have gesture control more than a decade ago? The problems with self-driving cars seem myriad but this doesn’t seem like the biggest to me.
It isn’t just recognizing and understanding gestures, which is difficult enough. It also has to recognize the difference between someone it should pay attention to and obey versus some yahoo who wants to mess with it.
I’d say rather that we don’t want it to be much less yahoo-resistant than it already is.
And we especially want to avoid creating single points of failure where one yahoo can create widespread mischief. Which is extra hard to prevent when all the machines are connected to the internet.
At some point, we can leave it in the hands of law enforcement. Our cars don’t bolt themselves to the pavement when parked because some yahoo can load it onto a flatbed and drive off. We don’t make windshields out of 4-inch laminated glass because yahoos can throw rocks off overpasses.
It’s true that we don’t want yahoo amplification–i.e., messing with millions of cars at once. But for messing with things locally, like someone pretending to be police and directing traffic? That’s something the cops can already deal with.
Yeah, but the biggest deterrent to yahooism is the human in the other vehicle who just might get out and lay a whipping on you, or call the cops on you, or be a cop, or be carrying a gun, or fly into road rage and chase you if you cut them off.
A person frustrated by a self-driving car may not feel that way. If you want the lane next to you, you can just start moving over, and the other driver will honk at you but probably let you in. But they might not, or they might become enraged. So you don’t do it.
But if a self driving car will 100% move out of your way if you start lane changing into it, and there’s no one inside to judge you or react to you, I could see people doing that kind of thing a lot more. Likewise, if a self-driving car will speed up when tailgated, people will start tailgating them. If you tailgate a human they migt brake check you, but an AI will just comply.
We haven’t even begun to experience what happens when some threshold is reached where self-driving cars begin to react to each other. Then it will get even more complex. I expect to see some interest feedback loops and other unknown unknowns cropping up.
On the other hand, self-driving cars are plastered with cameras and record everything. If we’re just talking about people squeezing their way into another lane, big deal. Humans should get less worked up about that sort of thing. If aggressive drivers are causing accidents or forcing self-driving cars into dangerous situations, they’ll be caught in short order with captured footage.
I give aggressive drivers a wide berth because I think they’re the ones most likely to fly into a fit of rage when they don’t get their way. I often spot them a quarter-mile behind me and change lanes to avoid them.
I’d say a self-driving car with no passengers might behave that way. But if there are riders on board, those folks are just as liable to be road ragers or armed or whatever as if they were actively steering when the aggressive human-driven car tried to bully them.
Privately owned self-driving cars should rarely be driving around empty. Hire cars a la Uber / Waymo certainly will do some, but the economics of the fleet demands it be a small fraction of the total driving.
In all, the bullyable uninhabited self-driving car seems to be a bit of a straw man. Not completely, but largely.
OTOH, your points about them interacting with one another in odd unexpected ways is spot-on. Kind of like two toddlers playing together when they don’t really get “cooperating” yet, and what little they do get is at a different level between them. I can see some “interesting” interactions arising from that.
What if auto-Teslas prove to be more aggressive than, say, auto-BMWs? Or one is prone to tailgating and the other to lane pushing? Maybe like iPhone and Android, they play well together, but are not as nice to others.
