7 ms·

This self-driving system promise keeps getting delayed! It looks to me this is a case of a donkey led by a carrot on a stick (investors are the donkey here). Wh

This self-driving system promise keeps getting delayed! It looks to me this is a case of a donkey led by a carrot on a stick (investors are the donkey here). What I never understand is what happens when a new situation that the system did not previously train for happen? what if this situation is deadly? Will all the Uber cars decide to jump off a cliff because of a wrong signalization or bad weather? What will the Teslas decide? Nature is very capable of creating 1 thing and multiply it, but it does not. The reason is it is more secure to have variation. If a problem happens, it will not affect every member (same what is happening with COVID by the way) Creating a mono-brain will soon prove disastrous. You just need a previously not seen situation or a crazy guy that gets imaginative and that will induce the death of hundreds

holoduke 6y ago
I am convinced that it takes another 100 to 200 years before we really have smart self driving cars. Maybe it will never happen as I think it will be much more likely that we have self flying drones before long before. Airspace being regulated heavily and thus much more predictable for a machine.
- drcross 6y ago
  What? 30 years ago we didn't have the internet. 100 years ago there was barely any electricity. Are you insane?
  - sillysaurusx 6y ago
    No need for the personal swipe!
- officehero 6y ago
  There is a push to upgrade road infrastructure with sensors so might be sooner. You're totally right about regulation though - engineers are only solving a "technical" problem, but what about the non-technical ones? Where is the brigade of lobbyists needed to get something like this through legal?
- papito 6y ago
  In 100 years the self-driving cars will be on Mars.
- cercatrova 6y ago
  100 to 200 years is a very long time. Look at the exponential nature of technology, and even deep learning. Only in 2012 had the image network reached 90% accuracy levels, and now we have all sort of networks for all sorts of things, including style transfer for videos.
  - officehero 6y ago
    The big problem is that the populace do not understand statistics. As soon as a self driving car kills someone the witch hunt begins and the whole fleet needs to be retracted until investigation is over. Can your image network solve this problem?
    - cercatrova 6y ago
      It's like nuclear power. It just has to continue regardless of how many deaths since the number of deaths will be lower than other methods, in this case coal and fossil fuel deaths, and in the case of self driving, manual driving accidents.
      - erklik 6y ago
        and like nuclear power, it is the politics that will kill* it, not the technological progress. * - Yes, Nuclear is not dead, but is not being used anywhere close to the potential it has i.e. effectively dead.
  - ntsplnkv2 6y ago
    There's nothing to suggest technological growth will be exponential forever. It's just as likely we've gotten all the easy stuff the last century and now everything else is much harder.
    - cercatrova 6y ago
      Maybe so, but there isn't much evidence of a slowdown while there is a lot of evidence of acceleration, such as time to new tech, re: hunter gatherer to agriculture to industrial revolution to information revolution to potentially genetic/biotech/AI revolution in the future.
- ska 6y ago
  I think humans are really poorly prepared to make accurate "100" year predictions. Otoh, we're pretty bad at 10 year ones too - the original self driving hype being a prime example. Lots of otherwise intelligent people fell into magical thinking there.
LeoPanthera 6y ago
> What I never understand is what happens when a new situation that the system did not previously train for happen? This happens to human drivers all the time, and humans are really bad drivers even in situations that have happened before. I fully expect AI drivers to be safer than human drivers.
- tintor 6y ago
  Human drivers can adapt to a new unseen before situation.
  - lelandbatey 6y ago
    While potentially true, it's a trade off it seems proponents are willing to take. The same kind of trade off happens with other modes of transport with very rigid constraints on movement, things like trains. A train can keep going, stop, or maybe with enough time reverse if there's tons of time and a strong need. We don't stand around worrying saying "BUT WHAT IF SOMETHING'S ON THE TRACKS?!?", because the answer is "well then, the train will hit it. That's the tradeoff for being able to move freight so efficiently."
  - reverend_gonzo 6y ago
    And software propagate learned knowledge perpetually, whereas each human has to learn it over and over again. On the other hand, software can make mistakes at scale whereas each human makes mistakes differently. It’s not so cut and dry.
  - warent 6y ago
    I think what this boils down to is that the volume of vehicle-related injury/death will decrease, and the few remaining cases will be really strange and unlikely to have occurred with human drivers.
  - jfoster 6y ago
    Sometimes they try to drive through floods. They also seem to randomly run red lights. Also, what's with the way they keep veering out of their lanes?
grecy 6y ago
> What I never understand is what happens when a new situation that the system did not previously train for happen? I think most people ask this kind of question because they assume the entire system is some kind of big if-else-if-else statement. The question about swerving and hitting the kid or the grandma follows the same logic. What we have to realize is that a neural net can and will react to things it hasn't specifically been told about using heuristics and "similar" stuff, just like human drivers. (see video of Tesla FSD avoiding a deer recently [1]) This is exactly like a new, or student driver. They have not yet encountered every possible situation, but they're able to apply reason and logic to new things and the more they drive, the better they get. So although they've never seen a deer jump out, they know it's good to avoid things when possible BUT NOT to swerve into oncoming traffic if there is traffic. So even a new-ish driver who has never seen a deer on the road will do a decent job of slamming on the brakes and swerving if save to do so. Same goes if a tire bounces towards you on the freeway, etc. etc. Tesla are working towards billions and billions of miles of this kind of "learning driving", where the system keeps getting better and better. And even after those billions of miles there will be situations it has never seen before, but the required response will be extraordinarily similar to things it has seen before. Chances are it will do a better job than you or I, because it's reaction time is close enough to 0 not to matter, and emotions won't cloud it's judgment if something really extraordinary happens (plane lands on highway, flood washes away road, etc) [1] https://www.youtube.com/watch?v=pWJoe8hwu_I https://www.youtube.com/watch?v=pWJoe8hwu_I
- resu_nimda 6y ago
  This is exactly like a new, or student driver. While I agree with your overall point, this is not true. AIs often fail in very inhuman ways. Their "thought process" and learning capabilities resemble ours only superficially. That said, humans also make a ton of lethal mistakes on the road and I do think we're getting close to a point where it might even out on balance. But people are more forgiving of relatable human failure modes than bizarre machine ones.
- jowday 6y ago
  That video isn't an example of this. Tesla's statements about fleet learning are pretty much theoretical - they're talking about a system they want to build someday. Right now, their motion planning and prediction is all hardcoded robotics. See Karpathy's talk here: https://medium.com/self-driving-cars/annotated-karpathys-autopilot-talk-3e6270a21f8d https://medium.com/self-driving-cars/annotated-karpathys-aut...
Shivetya 6y ago
well the one issue that should be obvious to all is that self driving cars suffer worse from the problem people suffer from, signage and markings on roads is still either incomplete or in some cases in error. People screw up every day driving and the lives lost is very high but going by how my Tesla drives I am going to say impede traffic by slowing down. My car does slow down, safely, when it gets confused or contradictory feedback, and I am curious what the in beta FSD software will do. Their new technique looks vastly improved. Traffic aware cruise control which many cars have along with limited lane keep assist can go a long way to making the drive safer but many don't use TACC except on highways Look at it this way, its coming. However it is all one sided because again I emphasize, we need better and more consistent signage and road marking. US wise, the HOV and Express lanes are probably the most perfect use case for self driving cars. Get into the lane and it should be hands off until exit.
kevincox 6y ago
Uber's has. But Waymo is currently giving self driving rides. By some definitions self-driving is here. Obviously there is still a ton of work to go from easy-to-drive areas to global but it seems like that is a huge notable step.

holoduke 6y ago

I am convinced that it takes another 100 to 200 years before we really have smart self driving cars. Maybe it will never happen as I think it will be much more likely that we have self flying drones before long before. Airspace being regulated heavily and thus much more predictable for a machine.

drcross 6y ago
What? 30 years ago we didn't have the internet. 100 years ago there was barely any electricity. Are you insane?
- sillysaurusx 6y ago
  No need for the personal swipe!
officehero 6y ago
There is a push to upgrade road infrastructure with sensors so might be sooner. You're totally right about regulation though - engineers are only solving a "technical" problem, but what about the non-technical ones? Where is the brigade of lobbyists needed to get something like this through legal?
papito 6y ago
In 100 years the self-driving cars will be on Mars.
cercatrova 6y ago
100 to 200 years is a very long time. Look at the exponential nature of technology, and even deep learning. Only in 2012 had the image network reached 90% accuracy levels, and now we have all sort of networks for all sorts of things, including style transfer for videos.
- officehero 6y ago
  The big problem is that the populace do not understand statistics. As soon as a self driving car kills someone the witch hunt begins and the whole fleet needs to be retracted until investigation is over. Can your image network solve this problem?
  - cercatrova 6y ago
    It's like nuclear power. It just has to continue regardless of how many deaths since the number of deaths will be lower than other methods, in this case coal and fossil fuel deaths, and in the case of self driving, manual driving accidents.
    - erklik 6y ago
      and like nuclear power, it is the politics that will kill* it, not the technological progress. * - Yes, Nuclear is not dead, but is not being used anywhere close to the potential it has i.e. effectively dead.
- ntsplnkv2 6y ago
  There's nothing to suggest technological growth will be exponential forever. It's just as likely we've gotten all the easy stuff the last century and now everything else is much harder.
  - cercatrova 6y ago
    Maybe so, but there isn't much evidence of a slowdown while there is a lot of evidence of acceleration, such as time to new tech, re: hunter gatherer to agriculture to industrial revolution to information revolution to potentially genetic/biotech/AI revolution in the future.
ska 6y ago
I think humans are really poorly prepared to make accurate "100" year predictions. Otoh, we're pretty bad at 10 year ones too - the original self driving hype being a prime example. Lots of otherwise intelligent people fell into magical thinking there.

LeoPanthera 6y ago

> What I never understand is what happens when a new situation that the system did not previously train for happen? This happens to human drivers all the time, and humans are really bad drivers even in situations that have happened before. I fully expect AI drivers to be safer than human drivers.

tintor 6y ago
Human drivers can adapt to a new unseen before situation.
- lelandbatey 6y ago
  While potentially true, it's a trade off it seems proponents are willing to take. The same kind of trade off happens with other modes of transport with very rigid constraints on movement, things like trains. A train can keep going, stop, or maybe with enough time reverse if there's tons of time and a strong need. We don't stand around worrying saying "BUT WHAT IF SOMETHING'S ON THE TRACKS?!?", because the answer is "well then, the train will hit it. That's the tradeoff for being able to move freight so efficiently."
- reverend_gonzo 6y ago
  And software propagate learned knowledge perpetually, whereas each human has to learn it over and over again. On the other hand, software can make mistakes at scale whereas each human makes mistakes differently. It’s not so cut and dry.
- warent 6y ago
  I think what this boils down to is that the volume of vehicle-related injury/death will decrease, and the few remaining cases will be really strange and unlikely to have occurred with human drivers.
- jfoster 6y ago
  Sometimes they try to drive through floods. They also seem to randomly run red lights. Also, what's with the way they keep veering out of their lanes?

grecy 6y ago

> What I never understand is what happens when a new situation that the system did not previously train for happen? I think most people ask this kind of question because they assume the entire system is some kind of big if-else-if-else statement. The question about swerving and hitting the kid or the grandma follows the same logic. What we have to realize is that a neural net can and will react to things it hasn't specifically been told about using heuristics and "similar" stuff, just like human drivers. (see video of Tesla FSD avoiding a deer recently [1]) This is exactly like a new, or student driver. They have not yet encountered every possible situation, but they're able to apply reason and logic to new things and the more they drive, the better they get. So although they've never seen a deer jump out, they know it's good to avoid things when possible BUT NOT to swerve into oncoming traffic if there is traffic. So even a new-ish driver who has never seen a deer on the road will do a decent job of slamming on the brakes and swerving if save to do so. Same goes if a tire bounces towards you on the freeway, etc. etc. Tesla are working towards billions and billions of miles of this kind of "learning driving", where the system keeps getting better and better. And even after those billions of miles there will be situations it has never seen before, but the required response will be extraordinarily similar to things it has seen before. Chances are it will do a better job than you or I, because it's reaction time is close enough to 0 not to matter, and emotions won't cloud it's judgment if something really extraordinary happens (plane lands on highway, flood washes away road, etc) [1] https://www.youtube.com/watch?v=pWJoe8hwu_I https://www.youtube.com/watch?v=pWJoe8hwu_I

resu_nimda 6y ago
This is exactly like a new, or student driver. While I agree with your overall point, this is not true. AIs often fail in very inhuman ways. Their "thought process" and learning capabilities resemble ours only superficially. That said, humans also make a ton of lethal mistakes on the road and I do think we're getting close to a point where it might even out on balance. But people are more forgiving of relatable human failure modes than bizarre machine ones.
jowday 6y ago
That video isn't an example of this. Tesla's statements about fleet learning are pretty much theoretical - they're talking about a system they want to build someday. Right now, their motion planning and prediction is all hardcoded robotics. See Karpathy's talk here: https://medium.com/self-driving-cars/annotated-karpathys-autopilot-talk-3e6270a21f8d https://medium.com/self-driving-cars/annotated-karpathys-aut...

Shivetya 6y ago

well the one issue that should be obvious to all is that self driving cars suffer worse from the problem people suffer from, signage and markings on roads is still either incomplete or in some cases in error. People screw up every day driving and the lives lost is very high but going by how my Tesla drives I am going to say impede traffic by slowing down. My car does slow down, safely, when it gets confused or contradictory feedback, and I am curious what the in beta FSD software will do. Their new technique looks vastly improved. Traffic aware cruise control which many cars have along with limited lane keep assist can go a long way to making the drive safer but many don't use TACC except on highways Look at it this way, its coming. However it is all one sided because again I emphasize, we need better and more consistent signage and road marking. US wise, the HOV and Express lanes are probably the most perfect use case for self driving cars. Get into the lane and it should be hands off until exit.

kevincox 6y ago

Uber's has. But Waymo is currently giving self driving rides. By some definitions self-driving is here. Obviously there is still a ton of work to go from easy-to-drive areas to global but it seems like that is a huge notable step.