7 ms·

FSD beta is fairly good, certainly better than the standard Autopilot. It's definitely not perfect, but so far for basic driving from point A to B it seems to w

FSD beta is fairly good, certainly better than the standard Autopilot. It's definitely not perfect, but so far for basic driving from point A to B it seems to work well. I can only see the tech improving as it gathers more data and feedback. It's a ML training data problem at this point, no?

mjhay 4y ago
Real FSD, where it is not necessary to have a human ready to take the controls at any time, isn't something you can just get more training data for. It's just that current ML methods aren't up for the task. Basically, it turns out you need to have something that can think, and ML is nowhere near that.
- BoorishBears 4y ago
  That is what no one at all wants. Tesla seems to be taking the big ball of ML approach, and that relies on a model that can practically think, and that in turn means you're essentially waiting on AGI. Which is ridiculous. Instead you break down the problem that is driving. Throw pieces of ML at specific problems, like traffic light classifiers, but have a much more structured approach to the high level decisions that are made. The benefit of this is that you gain clear introspection into why your vehicle is making the decisions it makes, and a direct way to influence them other than "throw in more data and hope next time it makes the right decision!" The latter is how you kill people, and it's what Tesla has been doing so far. - Hint for other readers: Please read the comment if you want to write a reply to it. Literally nothing was said about people wanting a car you can just get into, but of course the first reply didn't make it 2 sentences in before jumping to completely orthogonal conclusions to anything that was even remotely implied in this comment. It just gets exhausting to have a conversation about a topic like this if you have to repeat every point 10 times over.
  - danaris 4y ago
    "No one at all" wants a car where you can just get in, input the address, then ignore the wheel and what's going on outside until you get there? I think many, many people would beg to differ.
    - BoorishBears 4y ago
      Like I said in my edit since I was rate-limited, this is a nonsensical response to what I said. > Basically, it turns out you need to have something that can think, and ML is nowhere near that. This is what I was replying to. The comment I replied to didn't say a word about hoping into your car and inputting an address, and my comment repeatedly mentioned ML, so it's really hard to imagine you actually read either and came away with that erroneous of a conclusion.
ProjectArcturis 4y ago
With about a million edge cases, yeah.
BoorishBears 4y ago
FSD beta shows that Tesla is not working on self-driving. They've frozen their sensor suite in place because Elon wanted to be able to pre-sell FSD. The class of mistakes it makes demonstrate that Tesla is *willfully* marching into the local maxima problem: They're taking an approach that improves their actual product, AP, while sacrificing any hope of actually solving FSD. Full-self driving and what autopilot actually does take two completely different approaches. AP needs to be able to gracefully handle control to a non-trained layman. Current L3/L4 vehicles that rely on a driver have trained operators ready to takeover while the vehicle thinks everything is fine, and they need to aggressively make decisions like no-go, while AP never no-gos. For people who are in the self-driving field and are familiar with the approaches to perception and decision making (which I hope more of which will chime in here), FSD beta makes mistakes that are kind of like someone saying their robot can now feed you steak, and the robot starts by grasping the sharp end of a knife and cutting the steak with the handle. Sure you might make some progress getting mashed steak into your mouth, but the robot is cutting off its hand to do it. It's not going to result in much steak in your belly unless the robot's approach is completely reworked.
holmesworcester 4y ago
It's shockingly good and definitely worth 20% of the value of a car, for me. Two main reasons: 1. Your body doesn't feel tired after driving, because besides the effort of a hand resting on the bottom of the wheel, your body can be completely relaxed the vast majority of the time. 2. When FSD is on, you can completely stop paying attention to navigation-related decisions. You still have to pay attention to driving and intervene sometimes (most interventions are just a tap on the gas to give the car the confidence to go, because it's biased to caution) but every decision about when and where to turn is completely taken care of for you. This frees up a lot of mental overhead when you're in totally unfamiliar territory, including overhead that, without FSD, would pull your attention away from driving, making you less safe. I didn't get the magic of #2 until I used it on an unfamiliar drive home in New England backroads with lots of turns and realized I was paying attention to the road and my conversation with my partner, and none at all to where we were going. Then we were home. The one non-ideal thing about FSD Beta right now is that it's really easy to lose it for an arbitrarily long period of time. What happens is it's fairly easy to have your hand on the wheel on a straight stretch of highway and still get the nag. And there's a small but non-zero false negative rate when you wiggle the wheel to satisfy the nag but your wiggle is insufficient. And because it becomes so habitual and because the car beeps at you for other reasons too, there's a fairly high rate (at least for me) of wiggling to satisfy the nag and failing to notice that you didn't wiggle enough. The odds are small but after months of long drives it's pretty easy to accumulate the 5 "strikes" that trigger loss of FSD Beta without violating the intent of the nag, i.e. taking your eyes off the wheel or the road.
- holmesworcester 4y ago
  I'll also say that all of this value will probably be completely unapparent to someone who tries FSD Beta for the first time or infrequently, because like ChatGPT, FSD Beta is not a human-replacing AI that always gives you the right answer. It's a tool you learn to use, by building up a mental model of it over time. The first time you use FSD Beta will be a lot more like the first time you drove a car in actual traffic, or the first time you drove with a student driver. You'll feel safe enough to keep doing it but it will be very unsettling. But over time as your model builds of how it behaves, it becomes a tool you'll want to turn on in a growing number of situations. And you'll miss it. The value will also be much less apparent to someone who does short (<1 hour) drives on familiar roads.

mjhay 4y ago

Real FSD, where it is not necessary to have a human ready to take the controls at any time, isn't something you can just get more training data for. It's just that current ML methods aren't up for the task. Basically, it turns out you need to have something that can think, and ML is nowhere near that.

BoorishBears 4y ago
That is what no one at all wants. Tesla seems to be taking the big ball of ML approach, and that relies on a model that can practically think, and that in turn means you're essentially waiting on AGI. Which is ridiculous. Instead you break down the problem that is driving. Throw pieces of ML at specific problems, like traffic light classifiers, but have a much more structured approach to the high level decisions that are made. The benefit of this is that you gain clear introspection into why your vehicle is making the decisions it makes, and a direct way to influence them other than "throw in more data and hope next time it makes the right decision!" The latter is how you kill people, and it's what Tesla has been doing so far. - Hint for other readers: Please read the comment if you want to write a reply to it. Literally nothing was said about people wanting a car you can just get into, but of course the first reply didn't make it 2 sentences in before jumping to completely orthogonal conclusions to anything that was even remotely implied in this comment. It just gets exhausting to have a conversation about a topic like this if you have to repeat every point 10 times over.
- danaris 4y ago
  "No one at all" wants a car where you can just get in, input the address, then ignore the wheel and what's going on outside until you get there? I think many, many people would beg to differ.
  - BoorishBears 4y ago
    Like I said in my edit since I was rate-limited, this is a nonsensical response to what I said. > Basically, it turns out you need to have something that can think, and ML is nowhere near that. This is what I was replying to. The comment I replied to didn't say a word about hoping into your car and inputting an address, and my comment repeatedly mentioned ML, so it's really hard to imagine you actually read either and came away with that erroneous of a conclusion.

ProjectArcturis 4y ago

With about a million edge cases, yeah.

BoorishBears 4y ago

FSD beta shows that Tesla is not working on self-driving. They've frozen their sensor suite in place because Elon wanted to be able to pre-sell FSD. The class of mistakes it makes demonstrate that Tesla is *willfully* marching into the local maxima problem: They're taking an approach that improves their actual product, AP, while sacrificing any hope of actually solving FSD. Full-self driving and what autopilot actually does take two completely different approaches. AP needs to be able to gracefully handle control to a non-trained layman. Current L3/L4 vehicles that rely on a driver have trained operators ready to takeover while the vehicle thinks everything is fine, and they need to aggressively make decisions like no-go, while AP never no-gos. For people who are in the self-driving field and are familiar with the approaches to perception and decision making (which I hope more of which will chime in here), FSD beta makes mistakes that are kind of like someone saying their robot can now feed you steak, and the robot starts by grasping the sharp end of a knife and cutting the steak with the handle. Sure you might make some progress getting mashed steak into your mouth, but the robot is cutting off its hand to do it. It's not going to result in much steak in your belly unless the robot's approach is completely reworked.

holmesworcester 4y ago

It's shockingly good and definitely worth 20% of the value of a car, for me. Two main reasons: 1. Your body doesn't feel tired after driving, because besides the effort of a hand resting on the bottom of the wheel, your body can be completely relaxed the vast majority of the time. 2. When FSD is on, you can completely stop paying attention to navigation-related decisions. You still have to pay attention to driving and intervene sometimes (most interventions are just a tap on the gas to give the car the confidence to go, because it's biased to caution) but every decision about when and where to turn is completely taken care of for you. This frees up a lot of mental overhead when you're in totally unfamiliar territory, including overhead that, without FSD, would pull your attention away from driving, making you less safe. I didn't get the magic of #2 until I used it on an unfamiliar drive home in New England backroads with lots of turns and realized I was paying attention to the road and my conversation with my partner, and none at all to where we were going. Then we were home. The one non-ideal thing about FSD Beta right now is that it's really easy to lose it for an arbitrarily long period of time. What happens is it's fairly easy to have your hand on the wheel on a straight stretch of highway and still get the nag. And there's a small but non-zero false negative rate when you wiggle the wheel to satisfy the nag but your wiggle is insufficient. And because it becomes so habitual and because the car beeps at you for other reasons too, there's a fairly high rate (at least for me) of wiggling to satisfy the nag and failing to notice that you didn't wiggle enough. The odds are small but after months of long drives it's pretty easy to accumulate the 5 "strikes" that trigger loss of FSD Beta without violating the intent of the nag, i.e. taking your eyes off the wheel or the road.

holmesworcester 4y ago
I'll also say that all of this value will probably be completely unapparent to someone who tries FSD Beta for the first time or infrequently, because like ChatGPT, FSD Beta is not a human-replacing AI that always gives you the right answer. It's a tool you learn to use, by building up a mental model of it over time. The first time you use FSD Beta will be a lot more like the first time you drove a car in actual traffic, or the first time you drove with a student driver. You'll feel safe enough to keep doing it but it will be very unsettling. But over time as your model builds of how it behaves, it becomes a tool you'll want to turn on in a growing number of situations. And you'll miss it. The value will also be much less apparent to someone who does short (<1 hour) drives on familiar roads.