5 ms·

Maybe semantics, but the biggest problem isn't necessarily trusting the machine to do what's expected. The problem is the machine being able to predict what ped

Maybe semantics, but the biggest problem isn't necessarily trusting the machine to do what's expected. The problem is the machine being able to predict what pedestrians and other drivers are going to do. This isn't just a problem for driverless cars, it's become a real problem for people, too. The rules of the road are there for a reason, people! When you wave someone into traffic, or have your emergency flashers on while still moving in the rain, or wait too long at a busy 4-way stop, you're not being courteous or defensive or cautious, you're being unpredictable! (And most likely inefficient, as well.) Worse, (and this is going to be driverless cars' Achilles heel, IMHO) is people pulling into or across traffic (or otherwise failing to yield) based only on the expectation that the vehicle with the right-of-way will brake and do whatever is required to avoid a wreck. If they do it in front of a driverless car while trying to cross a busy highway, then people making this assumption are going to get killed.

jonplackett 7y ago
Or it might go the other way and driverless cars will be push-overs on the road and brake for everyone all the time. There’s a lot of decisions to be made about how they should react to that kind of stuff. Tesla’s last automation talk was interesting about that - they’re basically training the cars to act like a human driver would, looking for tiny signals that might warn of unexpected behaviour.
- beerandt 7y ago
  Good point. And it could have both problems. There's no reason I can think of to assume mutual exclusion. >act like a human driver Anything AI making this or a similar claim is automatically suspect to me. >looking for tiny signals that might warn of unexpected behaviour I suppose this is progress, except that it sort of implies that the AI wasn't already doing that. Which to me says the programmers are still having to anticipate too many possible scenarios.
- rayiner 7y ago
  Tesla still reliably runs into stationary objects and fails to brake for (dummy) children, so take all that with a grain of salt.
  - beerandt 7y ago
    Really? Stationary objects? I don't have a lot of faith in Tesla, but that's surprising.
jakeogh 7y ago
I agree. The humans are the bugs thing is an excuse that the pre-programmed proponents keep leaning on. That hypothetical has the unintentional implicit frame; as if it's the non pre-programmed humans that are the problem, when clearly a "self driving" car should be able to handle that mundane situation (although it never will). Adding rules to abstract away the humans is the natural next step that their bad assumptions lead to.
foota 7y ago
I actually think driverless cars will do better here. They will likely be able to recognize cars deviating from their current path in O(hundreds of milliseconds) and avoid a collision, whereas if a car from the opposite direction lane starts drifting into mine seconds before we hit there's little I can do.
- beerandt 7y ago
  I mentioned this elsewhere, but I'll elaborate- I don't think the detection or reaction time is the problem, it's the concept of useful recognition (and prediction). Is that first movement/deviation (that an AI car can certainly measure sooner than I can see) going to be a swerve within their lane or is the car going to come all the way over? It's not the watching for the trigger, but the intuition of what the trigger is, well, triggering. Humans have an uncanny ability in predicting the overall movement before any perceivable deviation. Of course I'm not saying it's magic; there's obviously some kind of signal to our subconscious that we're able to process and make useful. Perhaps AI can pick up on some of that. But it's not necessarily the "first movement" of the swerve that is most useful. And the "signal" that humans pick up on may not even be something externally measurable by the car. Like knowing of a particular interstate exit that drivers always seem to almost miss, but swerve across lanes to catch at the last second. There are also signals that are externally measurable, but have to be learned (seen on a previous trip, or warned by a friend). Like knowing about a bad pothole ahead, and knowing that the car in front of you is going to turn the corner, slam on the brakes and then swerve to avoid it. These are the kinds of things that I suspect AI will struggle to keep up with in the foreseeable future. And humans rely on this type of intuition way more than I think most people realize.
  - jodrellblank 7y ago
    Is this the 'disconnected lever' problem of AI, applied to vehicle driving? Of course I'm not saying it's magic; there's obviously some kind of signal to our subconscious that we're able to process and make useful. That signal is connected to some much larger dataset of years of human experience - things like what a distracted driver might do, what a tired driver might do, what an old unmaintained vehicle looks like, what hints at an out-of-state or out-of-country driver who might not be aware of local routes or problems, how people tend to behave differently at dawn/rush hour/day/night. An AI car should be able to tell which is a busy junction pretty easily, but watching a car swerve in a way that might indicate the driver is not paying full attention will make you more cautious around that car for a long time up ahead. Still, AI cars which don't tailgate, will get an ENORMOUS safety advantage over the people who drive 2ft behind me at highway speeds. That alone might be enough to overcome every other benefit. "AI car doesn't get involved in multi-car pileup", "AI car didn't get shunted at traffic lights", "AI car let me in at that junction" will keep happening and keep boosting people's confidence in them little by little.
visarga 7y ago
> The problem is the machine being able to predict what pedestrians and other drivers are going to do. That's something machine learning can do pretty well, actually. There has been a lot of research on this topic, and not applied just to self driving - natural language generation, stock prediction, playing games with multiple agents, even content recommendation do future prediction based on past information.
- beerandt 7y ago
  >natural language generation, stock prediction, playing games with multiple agents, even content recommendation I'm not familiar with the games, but from what I've seen with all the other things listed, that's not a technology that I'm anywhere near ready to trust my life with.

jonplackett 7y ago

Or it might go the other way and driverless cars will be push-overs on the road and brake for everyone all the time. There’s a lot of decisions to be made about how they should react to that kind of stuff. Tesla’s last automation talk was interesting about that - they’re basically training the cars to act like a human driver would, looking for tiny signals that might warn of unexpected behaviour.

beerandt 7y ago
Good point. And it could have both problems. There's no reason I can think of to assume mutual exclusion. >act like a human driver Anything AI making this or a similar claim is automatically suspect to me. >looking for tiny signals that might warn of unexpected behaviour I suppose this is progress, except that it sort of implies that the AI wasn't already doing that. Which to me says the programmers are still having to anticipate too many possible scenarios.
rayiner 7y ago
Tesla still reliably runs into stationary objects and fails to brake for (dummy) children, so take all that with a grain of salt.
- beerandt 7y ago
  Really? Stationary objects? I don't have a lot of faith in Tesla, but that's surprising.

jakeogh 7y ago

I agree. The humans are the bugs thing is an excuse that the pre-programmed proponents keep leaning on. That hypothetical has the unintentional implicit frame; as if it's the non pre-programmed humans that are the problem, when clearly a "self driving" car should be able to handle that mundane situation (although it never will). Adding rules to abstract away the humans is the natural next step that their bad assumptions lead to.

foota 7y ago

I actually think driverless cars will do better here. They will likely be able to recognize cars deviating from their current path in O(hundreds of milliseconds) and avoid a collision, whereas if a car from the opposite direction lane starts drifting into mine seconds before we hit there's little I can do.

beerandt 7y ago
I mentioned this elsewhere, but I'll elaborate- I don't think the detection or reaction time is the problem, it's the concept of useful recognition (and prediction). Is that first movement/deviation (that an AI car can certainly measure sooner than I can see) going to be a swerve within their lane or is the car going to come all the way over? It's not the watching for the trigger, but the intuition of what the trigger is, well, triggering. Humans have an uncanny ability in predicting the overall movement before any perceivable deviation. Of course I'm not saying it's magic; there's obviously some kind of signal to our subconscious that we're able to process and make useful. Perhaps AI can pick up on some of that. But it's not necessarily the "first movement" of the swerve that is most useful. And the "signal" that humans pick up on may not even be something externally measurable by the car. Like knowing of a particular interstate exit that drivers always seem to almost miss, but swerve across lanes to catch at the last second. There are also signals that are externally measurable, but have to be learned (seen on a previous trip, or warned by a friend). Like knowing about a bad pothole ahead, and knowing that the car in front of you is going to turn the corner, slam on the brakes and then swerve to avoid it. These are the kinds of things that I suspect AI will struggle to keep up with in the foreseeable future. And humans rely on this type of intuition way more than I think most people realize.
- jodrellblank 7y ago
  Is this the 'disconnected lever' problem of AI, applied to vehicle driving? Of course I'm not saying it's magic; there's obviously some kind of signal to our subconscious that we're able to process and make useful. That signal is connected to some much larger dataset of years of human experience - things like what a distracted driver might do, what a tired driver might do, what an old unmaintained vehicle looks like, what hints at an out-of-state or out-of-country driver who might not be aware of local routes or problems, how people tend to behave differently at dawn/rush hour/day/night. An AI car should be able to tell which is a busy junction pretty easily, but watching a car swerve in a way that might indicate the driver is not paying full attention will make you more cautious around that car for a long time up ahead. Still, AI cars which don't tailgate, will get an ENORMOUS safety advantage over the people who drive 2ft behind me at highway speeds. That alone might be enough to overcome every other benefit. "AI car doesn't get involved in multi-car pileup", "AI car didn't get shunted at traffic lights", "AI car let me in at that junction" will keep happening and keep boosting people's confidence in them little by little.

visarga 7y ago

> The problem is the machine being able to predict what pedestrians and other drivers are going to do. That's something machine learning can do pretty well, actually. There has been a lot of research on this topic, and not applied just to self driving - natural language generation, stock prediction, playing games with multiple agents, even content recommendation do future prediction based on past information.

beerandt 7y ago
>natural language generation, stock prediction, playing games with multiple agents, even content recommendation I'm not familiar with the games, but from what I've seen with all the other things listed, that's not a technology that I'm anywhere near ready to trust my life with.