4 ms·

> Can I sleep in it? > That’s it. That’s the test. Pick a vehicle. Can you get in, pick a destination and safely go to sleep? The problem with the litmus test

> Can I sleep in it? > That’s it. That’s the test. Pick a vehicle. Can you get in, pick a destination and safely go to sleep? The problem with the litmus test is that it sweeps all of the complexity of the question into the word "safely". Self driving cars aren't like, say, an elevator, which has physical mechanisms that an engineer can inspect and show to be in good working order. There is no human being or group of beings that can inspect the quality of the training data used for full self driving--the black box cannot answer for itself. A car might be able to drive nicely on straight, sunny roads... But reality has far too many edge cases such that the models can never be trained for all of them. What would the model make of a zebra or ostrich walking down the road? What if a baseball line-drives towards the vehicle? A tennis ball lands in the road--not an issue in itself for the car, but the oblivious 3 year old chasing after it would be an issue. Thus, allow me to propose an alternative razor, Invictus' Razor: It's not a self-driving car if it uses training data. Reality has too many edge cases and it's impossible to train for all of them. I believe we will need a new algorithm to do this task, something that ends up mimicking the human neocortex. What we have now is fundamentally not good enough.

throwuxiytayq 5y ago
Huh? Human drivers rely on hundreds of hours of training to safely and reliably handle most traffic situations, and still frequently fail to act appropriately in unexpected situations. We also get brain farts or fall asleep at the wheel and flatten the occasional pedestrian. According to the National Safety Council, the chances of dying from a motor vehicle crash is 1 in 103. Human driving isn't some impossible to pass benchmark of excellency - it's an endless shitshow of human tragedy.
- Invictus0 5y ago
  > Human drivers rely on hundreds of hours of training to safely and reliably handle most traffic situations Human training is different than AI training; they are so different that the words should not be the same. The difference is that humans see things how they are, not how they appear. If a human learns to drive only during the day, and then after their training has to drive at night, they will still recognize the traffic lights and stop signs--an AI will not. Moreover, a human really doesn't need hundreds of hours to learn to drive. I was driving at 5 years old, in a toy Jeep with a top speed of maybe 10 miles and hour. I knew not to drive into the flower garden or hit the dog. In many States, you can drive legally as early as 14, and historically many farm kids started younger than that with no training to speak of. When your vehicle is a slow tractor, it's not nearly as dangerous. > We also get brain farts or fall asleep at the wheel and flatten the occasional pedestrian. Great, this is something an algorithm that mimics the human brain would be able to improve upon. > According to the National Safety Council, the chances of dying from a motor vehicle crash is 1 in 103 Over what time frame, presumably a lifetime of about 75 years? A very meaningless and alarmist statistic--better to talk about the 1.1 fatalities per 100 million miles.
  - CRConrad 5y ago
    > The difference is that humans see things how they are, not how they appear. Muahahaha!
CRConrad 5y ago
> It's not a self-driving car if it uses training data. Reality has too many edge cases and it's impossible to train for all of them. Do you have a driver's license? If so, how did you get it? My guess is, if you do, you got it by learning, and then demonstrating you had learned, to drive. And you learned by... Training, just like everyone else. If that's not good enough for an automated system, then how is it good enough for human drivers? How the heck else is the system -- or anyone, or anything -- supposed to learn?
- Invictus0 5y ago
  https://www.youtube.com/watch?v=4Qt3xmUx88Q https://www.youtube.com/watch?v=4Qt3xmUx88Q Here's a five year old kid driving a toy car. He didn't take any training course to do this. He didn't need to be told not to hit the dog or drive off the side of the hill. Is this kid ready to drive on public roads? No, obviously not. But this kid is arguably better at handling edge cases on the road than any "self driving" system trained on a billion miles of road data.

throwuxiytayq 5y ago

Huh? Human drivers rely on hundreds of hours of training to safely and reliably handle most traffic situations, and still frequently fail to act appropriately in unexpected situations. We also get brain farts or fall asleep at the wheel and flatten the occasional pedestrian. According to the National Safety Council, the chances of dying from a motor vehicle crash is 1 in 103. Human driving isn't some impossible to pass benchmark of excellency - it's an endless shitshow of human tragedy.

Invictus0 5y ago
> Human drivers rely on hundreds of hours of training to safely and reliably handle most traffic situations Human training is different than AI training; they are so different that the words should not be the same. The difference is that humans see things how they are, not how they appear. If a human learns to drive only during the day, and then after their training has to drive at night, they will still recognize the traffic lights and stop signs--an AI will not. Moreover, a human really doesn't need hundreds of hours to learn to drive. I was driving at 5 years old, in a toy Jeep with a top speed of maybe 10 miles and hour. I knew not to drive into the flower garden or hit the dog. In many States, you can drive legally as early as 14, and historically many farm kids started younger than that with no training to speak of. When your vehicle is a slow tractor, it's not nearly as dangerous. > We also get brain farts or fall asleep at the wheel and flatten the occasional pedestrian. Great, this is something an algorithm that mimics the human brain would be able to improve upon. > According to the National Safety Council, the chances of dying from a motor vehicle crash is 1 in 103 Over what time frame, presumably a lifetime of about 75 years? A very meaningless and alarmist statistic--better to talk about the 1.1 fatalities per 100 million miles.
- CRConrad 5y ago
  > The difference is that humans see things how they are, not how they appear. Muahahaha!

CRConrad 5y ago

> It's not a self-driving car if it uses training data. Reality has too many edge cases and it's impossible to train for all of them. Do you have a driver's license? If so, how did you get it? My guess is, if you do, you got it by learning, and then demonstrating you had learned, to drive. And you learned by... Training, just like everyone else. If that's not good enough for an automated system, then how is it good enough for human drivers? How the heck else is the system -- or anyone, or anything -- supposed to learn?

Invictus0 5y ago
https://www.youtube.com/watch?v=4Qt3xmUx88Q https://www.youtube.com/watch?v=4Qt3xmUx88Q Here's a five year old kid driving a toy car. He didn't take any training course to do this. He didn't need to be told not to hit the dog or drive off the side of the hill. Is this kid ready to drive on public roads? No, obviously not. But this kid is arguably better at handling edge cases on the road than any "self driving" system trained on a billion miles of road data.