6 ms·

“Musk has said Autopilot is “unequivocally safer” than normal driving.” I expect that if Tesla was forced to release all camera footage from each Autopilot rel

by jmpman 4y ago

“Musk has said Autopilot is “unequivocally safer” than normal driving.” I expect that if Tesla was forced to release all camera footage from each Autopilot related crash, the public would be horrified. I expect it would show that each situation would have been handled perfectly by a 16 year old driver with a week of experience. I’d like to see people advocate for all self driving vehicles to be forced to publicly release all video of any crash involving their technology. There’s too much of an incentive for our politicians to willingly discount the lives of their constituents in order to look like a tech friendly business environment and make some political points.

oehpr 4y ago
I'm not exactly a booster for tesla or anything. But are you sure this is the standard you want to set? Would humans pass this standard? If you released dash cam footage of most human accidents and asked "could an inexperienced 16 year old avoid this situation?" What would the answer be? I don't think that's a good test.
- nso95 4y ago
  I think you missed the point
  - oehpr 4y ago
    Help me out then. Because as far as I'm aware, the baseline assertion is that fully self driving cars are safer than humans behind the wheel. to REFUTE that you don't just say "look at this one example" because the assertion was never "They will never crash and nothing bad will ever happen" It was "self driving cars are safer than humans" So: How do humans do? How do self driving cars do? And there's plenty we can talk about here. Scatter shot examples of failure are not it.
    - freejazz 4y ago
      Musk said "unequivocally". Besides, the person you responded to was making a different point. They were making the point that the crashes would likely be those easily avoided by humans. A) I think this is a distinct point from yours, but still answers it anyway as far as it being a refutation that FDS is safer, and it's not pointing to a single anecdote. the poster was calling for ALL footage to be released... B) if your point is to say that there is always a human that could have made such a mistake in the car, what is that supposed to mean? not all drivers should be driving, many are even unlicensed...
      - oehpr 4y ago
        My point was NOT say that a human could have made that mistake. Obviously a human would not have (except for all the times a distracted human would do exactly this). My point was that the standard "Release all footage of accidents and look at how many accidents humans could have avoided" is not productive. The method of illustrating this is to show how this standard would have performed with human drivers, who will generate hundreds of thousands of incidents that other humans could have easily prevented. Those incidents will have some categorical differences, obviously. But incidents they will be. If robots make 1 mistake in 10000, and monkey makes 1 mistake in 1000, robot's are unequivocally safer than monkeys. It's that simple. The question is how many mistakes robots make, and how many mistakes humans make.
        freejazz 4y ago
        If humans wouldn't have made that mistake... it's certainly not safer than humans in that regard. You are saying that FDS might be safer on the whole. Apparently, it would be because it FDS doesn't make other mistakes that humans do make, if it was safer. Not sure how that supports you saying that having footage of the mistakes that FDS does is not productive. It'd be productive even if FDS WAS actually as safe as humans (I seriously doubt it is, and no doubt the article outright disputes such a notion) because FDS could be made safer. The reality is that you are saying that FDS should not be any safer than a typical driver. While I certainly understand the rationality behind such an argument, I think we should strive for better.
    - lm28469 4y ago
      > Help me out then. It really isn't that hard: in it's current state it's too good to keep drivers engaged and not good enough to be relied on at all the times. Which end up in a lot of dumb accidents like plowing full speed in emergency vehicles, turning on train tracks, emergency braking for shadows on the road, &c.
radu_floricica 4y ago
True yet probably irrelevant, because self driving and humans are very likely to have different failure modes. It's quite possible to have an overall safer AI driver, but its fewer accidents to be ones that humans don't tend to do. It's still safer overall, and thus desirable. And a separate objection: I wouldn't be surprised if most accidents by humans were actually easily avoidable by an inexperienced 16 year old - if he was paying attention. If you watch a filmed accident you definitely pay attention, and it's easy to see what should have been done. It's not particularly relevant on how people actually drive day to day.
- jmpman 4y ago
  Most accidents by humans are due to intoxication. Could autopilot/FSD do better at paying attention to a street light than a drunk 21 year old texting with his cheating girlfriend at 3am? Yes. That’s where humans make fatal mistakes. Where do Teslas make fatal mistakes? Driving into road barriers at freeway speed. Driving into stopped emergency vehicles. Driving under semis. These are fundamentally sensor problems. Basic issues a 16 year old can handle. If we want our new drivers to be safe, they spend hours in driving education class, watching footage of car wrecks, so they understand simple things like - don’t drive around railroad crossing barriers. Why shouldn’t we have training of - don’t engage camera based full self driving when driving a curvy mountain road into the sunset with intermittent freeway dividers. There are a class of failures that FSD can’t deal with. And yet it feels like people are telling me that - cars being hit by trains are rare, why are you edicating drivers to be careful around trains? Don’t you know how much safer trains are at hauling freight around the country. You’re just making people scared of trains.. all the time, the railroad barons are employing lawyers and lobbyists advocating against putting up railroad crossings, or letting the public know that a car on a train track is a bad idea. There are likely basic failure modes that Teslas are bad at. But, as an informed consumer, I should have the right to know what those failure modes are, so I can assess if I want to engage autopilot in a school zone. Should I engage autopilot while approaching a school bus that just pulled over? Will the Tesla stop when the bus’s stop sign pops out? At what angle does that stop sign not get detected? What if the sun is behind the bus? Again, these are scenarios that a 16 year old can handle, and yet, there’s nobody on this forum who would be willing take a $1k bet that today, a Tesla would properly handle that situation in all angle, timing, lighting conditions, but are instead willing to say, those pesky kids, crossing from their school bus, are acceptable collateral damage… don’t want people to understand these corner cases, because Tesla can detect a few red pixels better than distracted drunk teen.
diffeomorphism 4y ago
I see that backfiring horribly, e.g. see reddit's /r/idiotsincars. If that is your standard it seems much more reasonable to expect that "normal driving" will require substantially higher insurance premiums or even special permits within like 20 years

by jmpman 4y ago

oehpr 4y ago

I'm not exactly a booster for tesla or anything. But are you sure this is the standard you want to set? Would humans pass this standard? If you released dash cam footage of most human accidents and asked "could an inexperienced 16 year old avoid this situation?" What would the answer be? I don't think that's a good test.

nso95 4y ago
I think you missed the point
- oehpr 4y ago
  Help me out then. Because as far as I'm aware, the baseline assertion is that fully self driving cars are safer than humans behind the wheel. to REFUTE that you don't just say "look at this one example" because the assertion was never "They will never crash and nothing bad will ever happen" It was "self driving cars are safer than humans" So: How do humans do? How do self driving cars do? And there's plenty we can talk about here. Scatter shot examples of failure are not it.
  - freejazz 4y ago
    Musk said "unequivocally". Besides, the person you responded to was making a different point. They were making the point that the crashes would likely be those easily avoided by humans. A) I think this is a distinct point from yours, but still answers it anyway as far as it being a refutation that FDS is safer, and it's not pointing to a single anecdote. the poster was calling for ALL footage to be released... B) if your point is to say that there is always a human that could have made such a mistake in the car, what is that supposed to mean? not all drivers should be driving, many are even unlicensed...
    - oehpr 4y ago
      My point was NOT say that a human could have made that mistake. Obviously a human would not have (except for all the times a distracted human would do exactly this). My point was that the standard "Release all footage of accidents and look at how many accidents humans could have avoided" is not productive. The method of illustrating this is to show how this standard would have performed with human drivers, who will generate hundreds of thousands of incidents that other humans could have easily prevented. Those incidents will have some categorical differences, obviously. But incidents they will be. If robots make 1 mistake in 10000, and monkey makes 1 mistake in 1000, robot's are unequivocally safer than monkeys. It's that simple. The question is how many mistakes robots make, and how many mistakes humans make.
      - freejazz 4y ago
        If humans wouldn't have made that mistake... it's certainly not safer than humans in that regard. You are saying that FDS might be safer on the whole. Apparently, it would be because it FDS doesn't make other mistakes that humans do make, if it was safer. Not sure how that supports you saying that having footage of the mistakes that FDS does is not productive. It'd be productive even if FDS WAS actually as safe as humans (I seriously doubt it is, and no doubt the article outright disputes such a notion) because FDS could be made safer. The reality is that you are saying that FDS should not be any safer than a typical driver. While I certainly understand the rationality behind such an argument, I think we should strive for better.
  - lm28469 4y ago
    > Help me out then. It really isn't that hard: in it's current state it's too good to keep drivers engaged and not good enough to be relied on at all the times. Which end up in a lot of dumb accidents like plowing full speed in emergency vehicles, turning on train tracks, emergency braking for shadows on the road, &c.

radu_floricica 4y ago

True yet probably irrelevant, because self driving and humans are very likely to have different failure modes. It's quite possible to have an overall safer AI driver, but its fewer accidents to be ones that humans don't tend to do. It's still safer overall, and thus desirable. And a separate objection: I wouldn't be surprised if most accidents by humans were actually easily avoidable by an inexperienced 16 year old - if he was paying attention. If you watch a filmed accident you definitely pay attention, and it's easy to see what should have been done. It's not particularly relevant on how people actually drive day to day.

jmpman 4y ago
Most accidents by humans are due to intoxication. Could autopilot/FSD do better at paying attention to a street light than a drunk 21 year old texting with his cheating girlfriend at 3am? Yes. That’s where humans make fatal mistakes. Where do Teslas make fatal mistakes? Driving into road barriers at freeway speed. Driving into stopped emergency vehicles. Driving under semis. These are fundamentally sensor problems. Basic issues a 16 year old can handle. If we want our new drivers to be safe, they spend hours in driving education class, watching footage of car wrecks, so they understand simple things like - don’t drive around railroad crossing barriers. Why shouldn’t we have training of - don’t engage camera based full self driving when driving a curvy mountain road into the sunset with intermittent freeway dividers. There are a class of failures that FSD can’t deal with. And yet it feels like people are telling me that - cars being hit by trains are rare, why are you edicating drivers to be careful around trains? Don’t you know how much safer trains are at hauling freight around the country. You’re just making people scared of trains.. all the time, the railroad barons are employing lawyers and lobbyists advocating against putting up railroad crossings, or letting the public know that a car on a train track is a bad idea. There are likely basic failure modes that Teslas are bad at. But, as an informed consumer, I should have the right to know what those failure modes are, so I can assess if I want to engage autopilot in a school zone. Should I engage autopilot while approaching a school bus that just pulled over? Will the Tesla stop when the bus’s stop sign pops out? At what angle does that stop sign not get detected? What if the sun is behind the bus? Again, these are scenarios that a 16 year old can handle, and yet, there’s nobody on this forum who would be willing take a $1k bet that today, a Tesla would properly handle that situation in all angle, timing, lighting conditions, but are instead willing to say, those pesky kids, crossing from their school bus, are acceptable collateral damage… don’t want people to understand these corner cases, because Tesla can detect a few red pixels better than distracted drunk teen.

diffeomorphism 4y ago

I see that backfiring horribly, e.g. see reddit's /r/idiotsincars. If that is your standard it seems much more reasonable to expect that "normal driving" will require substantially higher insurance premiums or even special permits within like 20 years