2 ms·

>he collision avoidance system couldn't determine if the vehicle was in-lane or not because of the very small overlap. And this some small glitch? This should

>he collision avoidance system couldn't determine if the vehicle was in-lane or not because of the very small overlap. And this some small glitch? This should not be the first problem you would solve, detect big objects, then you move to the next problem detect the object speeds, then you move to determine the objects next position. It seems they did not solved the first essential problem yet. Teslas hit cement lane separators, hit firetrucks, would have hit a tram/train, hot those small metal poles, dis-engaged in many regular city conditions. This tech is far from ready this decade - what I think it needs is better sensors that can detect solid objects or maybe 100 times more GPU power and 1000 times more training data. The numbers are not public because it will show that the AI is less safe then a human, if it were close then for sure Tesla PR could have spin it in a good light for them.

mlindner 4y ago
> And this some small glitch? This should not be the first problem you would solve, detect big objects, then you move to the next problem detect the object speeds, then you move to determine the objects next position. As far as I'm aware it was a regression (though I could be wrong), as it's a very rarely seen thing in reality and there was a sudden rash of news articles about it and then little since. > It seems they did not solved the first essential problem yet. That's not how neural networks work. > The numbers are not public because it will show that the AI is less safe then a human, if it were close then for sure Tesla PR could have spin it in a good light for them. Again, as I stated, it would result in the media comparing apples and oranges as Tesla does not restrict where you're allowed to drive. I'm sure if you took a Waymo to a snow covered dirt road in the winter time it wouldn't be able to drive at all so is that infinite dis-engagements? > The numbers are not public because it will show that the AI is less safe then a human, if it were close then for sure Tesla PR could have spin it in a good light for them. The released numbers show that at least where people drive with autopilot, there are less fatal incidents.
- simion314 4y ago
  >As far as I'm aware it was a regression (though I could be wrong), as it's a very rarely seen thing in reality and there was a sudden rash of news articles about it and then little since. This can be because the "bad press" made it that the human driver also pays attention, if we would have access to all the data I am 100% sure we will find daily cases where the human prevents this frontal colission with trucks, trains and walls to happen. How can it be a regression? Any official explanation from a Tesla engineer? Did they had some bad detection and some dude decided to change some parameter values or was this blamed on the AI? I am super curious to read this, if it was on HN I missed it. >That's not how neural networks work. I know how ANN work, but good engineers always use best tool for the job and split a big problem in smaller independent problems, it is easy to developer and test. I assume the fault is with people above that rush the engineers and force them to work with impossible constraints. >Again, as I stated, it would result in the media comparing apples and oranges as Tesla does not restrict where you're allowed to drive. I'm sure if you took a Waymo to a snow covered dirt road in the winter time it wouldn't be able to drive at all so is that infinite dis-engagements? This makes no sense dude, you are trying to hard to bend logic. Tesla can show numbers limited to a city or a road. Let me think for you and Tesla PR "On the road from A to B, 25km of heavy traffic, 500k cars daily, Tesla had 3 dis-engagements daily but the same number of humans in similar cars cause in average 5 accidents on this exact road and conditions". If the data is precise and correct and shows Tesla is better then humans then it would be public, if Waymo can do better then good for them. Though I would prefer Tesla would publish data for theentire city not only one street. Btw I notice how you Tesla and Apple fans defend the giant companies with shit like "the press will say mean things" or "some bad customer will sabotage the device" or "some idiot might do idiotic things and go to the press so let's remove your rights". The press did a good job to show that the AI can't be trusted and probably saved lives. This would not have been needed if Tesla would not have used false advertising.

mlindner 4y ago

> And this some small glitch? This should not be the first problem you would solve, detect big objects, then you move to the next problem detect the object speeds, then you move to determine the objects next position. As far as I'm aware it was a regression (though I could be wrong), as it's a very rarely seen thing in reality and there was a sudden rash of news articles about it and then little since. > It seems they did not solved the first essential problem yet. That's not how neural networks work. > The numbers are not public because it will show that the AI is less safe then a human, if it were close then for sure Tesla PR could have spin it in a good light for them. Again, as I stated, it would result in the media comparing apples and oranges as Tesla does not restrict where you're allowed to drive. I'm sure if you took a Waymo to a snow covered dirt road in the winter time it wouldn't be able to drive at all so is that infinite dis-engagements? > The numbers are not public because it will show that the AI is less safe then a human, if it were close then for sure Tesla PR could have spin it in a good light for them. The released numbers show that at least where people drive with autopilot, there are less fatal incidents.

simion314 4y ago
>As far as I'm aware it was a regression (though I could be wrong), as it's a very rarely seen thing in reality and there was a sudden rash of news articles about it and then little since. This can be because the "bad press" made it that the human driver also pays attention, if we would have access to all the data I am 100% sure we will find daily cases where the human prevents this frontal colission with trucks, trains and walls to happen. How can it be a regression? Any official explanation from a Tesla engineer? Did they had some bad detection and some dude decided to change some parameter values or was this blamed on the AI? I am super curious to read this, if it was on HN I missed it. >That's not how neural networks work. I know how ANN work, but good engineers always use best tool for the job and split a big problem in smaller independent problems, it is easy to developer and test. I assume the fault is with people above that rush the engineers and force them to work with impossible constraints. >Again, as I stated, it would result in the media comparing apples and oranges as Tesla does not restrict where you're allowed to drive. I'm sure if you took a Waymo to a snow covered dirt road in the winter time it wouldn't be able to drive at all so is that infinite dis-engagements? This makes no sense dude, you are trying to hard to bend logic. Tesla can show numbers limited to a city or a road. Let me think for you and Tesla PR "On the road from A to B, 25km of heavy traffic, 500k cars daily, Tesla had 3 dis-engagements daily but the same number of humans in similar cars cause in average 5 accidents on this exact road and conditions". If the data is precise and correct and shows Tesla is better then humans then it would be public, if Waymo can do better then good for them. Though I would prefer Tesla would publish data for theentire city not only one street. Btw I notice how you Tesla and Apple fans defend the giant companies with shit like "the press will say mean things" or "some bad customer will sabotage the device" or "some idiot might do idiotic things and go to the press so let's remove your rights". The press did a good job to show that the AI can't be trusted and probably saved lives. This would not have been needed if Tesla would not have used false advertising.