I understand a software monoculture is bad, but if Toyotas crash to the left, and Fords crash to the right, the kind of accident you have will depend on the type of car.
If the crash is unavoidable, then that would arguably be a good thing, assuming Toyota designed their left side to withstand more impact, and Ford designed their right side similarly....
A human eye can discern far more contrast differences than a camera does, so whether she was visible to a human eye while in the shadow is questionable.
That's arguable. You'll often read people claiming that the human eye has a dynamic range of 20 stops. As I understand it, that's technically correct (the best kind of correct), in that the human eye has a range of 20 stops if you include everything from fully dark-adapted vision with the iris wide open all the way to full day vision with the iris as closed as it can be. The problem is that it takes half an hour to go from that second state back to the first, i.e. the human eye has nowhere near 20 stops at any single point in time.
If you measured a camera in the same way, my Canon 5D Mark IV would have almost 14 stops of dynamic range before you add the iris in the lens, plus six stops from the lens (assuming an f/2.8 lens), making it roughly equivalent to the human eye even before you start cranking up the ISO sensitivity.
Adding further complexity is the fact that the human eye has its maximum dynamic range when there is no bright light hitting it. In the daytime, it has only about a 10-stop range. The question of how much a set of car headlights reduces your eye's night vision abilities is left as an exercise for the reader.
Ahhh. Right, buy car X, 20% less likely to run over pedestrians or kill occupants based on our patented algorithm that no one else can use. Well, you said right there "inarguably", so I guess I can't argue with that logic.
I didn't say information shouldn't be shared. And I'm not a fan of patents in general. But it is important for the implementations to be independent, and for most of the training data to be independent.
To use the operating systems example, we have Linux and several *BSD variants. There have been lots of bugs in one that don't affect the other and vice versa, because they were designed independently. We still have too much of a monoculture even in server OSes, mind you, but it is better than it could be. Imagine if Linux had never happened and AT&T UNIX and BSD hadn't forked from one another after basically writing independent implementations of most of the interesting bits. And ask yourself how much better off we would be if there were four or five popular crypto implementations (with compatible API) instead of everybody using OpenSSL. Heartbleed would have been a much smaller crisis.
Like Elon Musk said on Colbert's show, if you're in a sinking boat, and you have a design for a better bucket, you share the design. You don't set up multiple groups with different bucket designs and another committee to identify the minimum standards for a bucket. I may have misquoted him a little, but you get the idea.
Well, yes, but if you have a thousand boats that are on land, but that stand a decent risk of sinking someday, you let multiple people design different buckets, so that if one of them turns out to have a problem with the bottom falling out, only some of your boats sink. Human drivers aren't perfect, but they're a long way from being a sinking boat. We'd certainly like to achieve self-driving ubiquity, but getting there isn't an emergency.
Think of crawling out from under a pile of progressively larger cars.
Sure, you can rip your way straight up and out. But that leaves a vertical tunnel through the cars. You also push a certain number of cars up and out with you.
Instead you mostly follow the layers of cars, slowly making your way upward and outward. As you pass, gravity causes the cars that are displaced to collapse back in on the path of travel.
Personally, I think regulation is required. It's great if Google/Alphabet/Waymo is having success with their cars, or Lyft, or Tesla's experience with autopilot, but if we're going to have these cars on the road they should all be running the same software, it needs to be a collaborative effort. They can compete on human amenities inside the car, the software at a minimum (maybe sensors as well) should be a cooperative process where they share information and develop together. At the end, it gets certified by the government, so that all autonomous cars are controlled by the same software.
Assuming I'm understanding your proposal correctly, go ask a security researcher how the Windows monoculture has worked out, and you'll get a better understanding of why that's actually an exceptionally bad idea.
Imagine this scenario: In a hypothetical future, vehicles have ditched their steering wheels entirely, because cars drive themselves. All the car companies use the same models, with the same software underneath. Someone discovers an exploitable bug in the software or in the machine learning models — a way to trick cars into suddenly speeding up to 100 MPH and bricking the computer in that state, for example. Now you have a monoculture of tech, and there are a**holes flashing lights at cars or putting stickers on stop signs or doing whatever it is that causes the cars to misbehave, and instead of having five or ten percent of the cars misbehave, every single car does, and the human population potentially collapses overnight.
Worse, even if the problem got caught quickly, you wouldn't be able to say, "Everyone with a vehicle made by [insert car company here] needs to stop driving his or her car until it gets an update," because the problem would affect every car on the road, plus every delivery truck, etc. The economic impact would be catastrophic even if folks were only stranded for a day or two. If it took a week to fix the bug, people would be starving in the streets.
No, having multiple competing technologies is inarguably a good thing, though sharing ideas with each other at a high level wouldn't necessarily be bad. And there should be a standard battery of tests that every car has to pass before it can be allowed to drive without a trained safety driver, and the various companies should constantly be adding to those tests to ensure a certain minimum level of safety across the industry. But the emphasis should be on minimum standards, not creating a single, standard set of software. Otherwise bad things will happen.
Yes and no. You're seeing a recording that almost certainly has a lot less dynamic range than the raw pixel data from the sensor. I would assume that the self-driving tech uses raw pixel data, not a JPEG/MPEG-compressed approximation thereof.
If you take a photo in RAW mode on a DSLR, you can crank the gain up by two or three stops and see all sorts of stuff in the shadows that would otherwise not be visible within the color gamut of your monitor or a JPEG rendering. And even with the smaller cameras that they use in cars, you'd probably still have at least one stop worth of additional useful data down in the mud.
So there's a decent chance that there is visual information that isn't visible in the recording, but that the computer vision system could "see". Thus, we can't really judge whether a person could have seen the pedestrian any more than we can judge whether the car should have been able to see her, because you lose too much information in the recording. The best we can really do is guess until Uber actually takes the raw recorded data and analyzes it.
The problem with your logic, and the place where all such arguments inevitably fall down, is that employees don't get to choose their pay level. If older employees could say, "I don't need this stress. Demote me to junior grade and reduce my pay accordingly," and then slack off, their experience would make them worth keeping. They can't, and their higher pay grade is basically tied to their age, which means canning the people because of their pay rate is age discrimination, just under a different name.
The USA is around 5% of the population of the planet. We can't be 1% unless you do some very strange math involving all the dead people on Earth who ever lived.
This is why you dump the brine somewhere that has an actual current, i.e. not right at the coast, and not in an inlet. That way, it gets dissipated over a large swath of ocean.
Of course, the "A" in "AMD" stands for age-related, which is to say that having half your life ahead of you is pretty rare. Most people who get it do so in their 60s or 70s.
Now, perhaps a human might have realized the limited sight lines and the possible pedestrian conflict and slowed down before arriving at this location?
Or used the next lane over. Either way, inadequate visibility is a road design problem.
Twitter actually ran an ad about how chicken gets to your table, on these tweets about someone getting killed while crossing the road. I reported it as "I don't like this ad", because there's no "This is highly inappropriate in this context" option. Lovely.
Well, my theory was that the cable failure let it sag until one end was no longer on top of its post. The sudden part in that theory was the whole "slipping past the edge of the post and dropping straight down" part.
I would argue that there should be enough cable redundancy so that one or two failures wouldn't bring down the whole bridge, which IMO makes it at least partially a design issue, too, but....
- The engineering team apparently did analyze the cracks and concluded they were not a safety concern. We don't know that this is wrong, the bridge might have collapsed because of cable tightening and the cracks that showed up before that were irrelevant.
I think it's more likely that the cracking had the same root cause. I got the impression from various news articles that the cables lost tension for some reason. Perhaps the cables inside the bridge itself snapped or some welds broke, allowing the whole bridge to sag until they hit a point where the cables were unable to slip any further. Then, when the crew tried to tighten it, the cables pulled completely loose, allowing it to sag well beyond what it should have, and then the end of the bridge slid off the top of the post and fell straight down, and all the support posts punched through the top as the top part pancaked on the bottom part upon impact.
But that's just a gut feeling from looking at the result and reading news stories (which may or may not be even slightly accurate), so take that with a grain of salt.
If he were in the South, I'd assume he was cooking meth. Not sure about Wisconsin. Either way, that seems to be the usual cause of random buildings exploding these days.
Slashdot Mirror
Or we'd have someone walking out ahead of the self-driving cars carrying a flag.
If the crash is unavoidable, then that would arguably be a good thing, assuming Toyota designed their left side to withstand more impact, and Ford designed their right side similarly....
... says the anonymous coward, while pointing to an article that indicates that investigators don't yet know whether autosteer was even turned on.
Good to know that we'll be able to continue using Linux, so long as the round function is changed to compute the square root instead.
That's arguable. You'll often read people claiming that the human eye has a dynamic range of 20 stops. As I understand it, that's technically correct (the best kind of correct), in that the human eye has a range of 20 stops if you include everything from fully dark-adapted vision with the iris wide open all the way to full day vision with the iris as closed as it can be. The problem is that it takes half an hour to go from that second state back to the first, i.e. the human eye has nowhere near 20 stops at any single point in time.
If you measured a camera in the same way, my Canon 5D Mark IV would have almost 14 stops of dynamic range before you add the iris in the lens, plus six stops from the lens (assuming an f/2.8 lens), making it roughly equivalent to the human eye even before you start cranking up the ISO sensitivity.
Adding further complexity is the fact that the human eye has its maximum dynamic range when there is no bright light hitting it. In the daytime, it has only about a 10-stop range. The question of how much a set of car headlights reduces your eye's night vision abilities is left as an exercise for the reader.
I didn't say information shouldn't be shared. And I'm not a fan of patents in general. But it is important for the implementations to be independent, and for most of the training data to be independent.
To use the operating systems example, we have Linux and several *BSD variants. There have been lots of bugs in one that don't affect the other and vice versa, because they were designed independently. We still have too much of a monoculture even in server OSes, mind you, but it is better than it could be. Imagine if Linux had never happened and AT&T UNIX and BSD hadn't forked from one another after basically writing independent implementations of most of the interesting bits. And ask yourself how much better off we would be if there were four or five popular crypto implementations (with compatible API) instead of everybody using OpenSSL. Heartbleed would have been a much smaller crisis.
Well, yes, but if you have a thousand boats that are on land, but that stand a decent risk of sinking someday, you let multiple people design different buckets, so that if one of them turns out to have a problem with the bottom falling out, only some of your boats sink. Human drivers aren't perfect, but they're a long way from being a sinking boat. We'd certainly like to achieve self-driving ubiquity, but getting there isn't an emergency.
Think of crawling out from under a pile of progressively larger cars.
Sure, you can rip your way straight up and out. But that leaves a vertical tunnel through the cars. You also push a certain number of cars up and out with you.
Instead you mostly follow the layers of cars, slowly making your way upward and outward. As you pass, gravity causes the cars that are displaced to collapse back in on the path of travel.
Happy?
Assuming I'm understanding your proposal correctly, go ask a security researcher how the Windows monoculture has worked out, and you'll get a better understanding of why that's actually an exceptionally bad idea.
Imagine this scenario: In a hypothetical future, vehicles have ditched their steering wheels entirely, because cars drive themselves. All the car companies use the same models, with the same software underneath. Someone discovers an exploitable bug in the software or in the machine learning models — a way to trick cars into suddenly speeding up to 100 MPH and bricking the computer in that state, for example. Now you have a monoculture of tech, and there are a**holes flashing lights at cars or putting stickers on stop signs or doing whatever it is that causes the cars to misbehave, and instead of having five or ten percent of the cars misbehave, every single car does, and the human population potentially collapses overnight.
Worse, even if the problem got caught quickly, you wouldn't be able to say, "Everyone with a vehicle made by [insert car company here] needs to stop driving his or her car until it gets an update," because the problem would affect every car on the road, plus every delivery truck, etc. The economic impact would be catastrophic even if folks were only stranded for a day or two. If it took a week to fix the bug, people would be starving in the streets.
No, having multiple competing technologies is inarguably a good thing, though sharing ideas with each other at a high level wouldn't necessarily be bad. And there should be a standard battery of tests that every car has to pass before it can be allowed to drive without a trained safety driver, and the various companies should constantly be adding to those tests to ensure a certain minimum level of safety across the industry. But the emphasis should be on minimum standards, not creating a single, standard set of software. Otherwise bad things will happen.
Yes and no. You're seeing a recording that almost certainly has a lot less dynamic range than the raw pixel data from the sensor. I would assume that the self-driving tech uses raw pixel data, not a JPEG/MPEG-compressed approximation thereof.
If you take a photo in RAW mode on a DSLR, you can crank the gain up by two or three stops and see all sorts of stuff in the shadows that would otherwise not be visible within the color gamut of your monitor or a JPEG rendering. And even with the smaller cameras that they use in cars, you'd probably still have at least one stop worth of additional useful data down in the mud.
So there's a decent chance that there is visual information that isn't visible in the recording, but that the computer vision system could "see". Thus, we can't really judge whether a person could have seen the pedestrian any more than we can judge whether the car should have been able to see her, because you lose too much information in the recording. The best we can really do is guess until Uber actually takes the raw recorded data and analyzes it.
ssssssssssssPOP
My head asplode.
The problem with your logic, and the place where all such arguments inevitably fall down, is that employees don't get to choose their pay level. If older employees could say, "I don't need this stress. Demote me to junior grade and reduce my pay accordingly," and then slack off, their experience would make them worth keeping. They can't, and their higher pay grade is basically tied to their age, which means canning the people because of their pay rate is age discrimination, just under a different name.
The USA is around 5% of the population of the planet. We can't be 1% unless you do some very strange math involving all the dead people on Earth who ever lived.
This is why you dump the brine somewhere that has an actual current, i.e. not right at the coast, and not in an inlet. That way, it gets dissipated over a large swath of ocean.
I wish I could say that I saw it coming.
Of course, the "A" in "AMD" stands for age-related, which is to say that having half your life ahead of you is pretty rare. Most people who get it do so in their 60s or 70s.
Get this guy a hacksaw.
But in all seriousness, that's seriously awesome. Was this wet AMD or dry AMD? Because the description doesn't sound like either one.
Or used the next lane over. Either way, inadequate visibility is a road design problem.
Twitter actually ran an ad about how chicken gets to your table, on these tweets about someone getting killed while crossing the road. I reported it as "I don't like this ad", because there's no "This is highly inappropriate in this context" option. Lovely.
It's reverse-alphabetical. Or something. Just wait. Azure Systems will leave them all in the dust in a few years.
Well, my theory was that the cable failure let it sag until one end was no longer on top of its post. The sudden part in that theory was the whole "slipping past the edge of the post and dropping straight down" part.
I would argue that there should be enough cable redundancy so that one or two failures wouldn't bring down the whole bridge, which IMO makes it at least partially a design issue, too, but....
I think it's more likely that the cracking had the same root cause. I got the impression from various news articles that the cables lost tension for some reason. Perhaps the cables inside the bridge itself snapped or some welds broke, allowing the whole bridge to sag until they hit a point where the cables were unable to slip any further. Then, when the crew tried to tighten it, the cables pulled completely loose, allowing it to sag well beyond what it should have, and then the end of the bridge slid off the top of the post and fell straight down, and all the support posts punched through the top as the top part pancaked on the bottom part upon impact.
But that's just a gut feeling from looking at the result and reading news stories (which may or may not be even slightly accurate), so take that with a grain of salt.
Seriously. Now if the weed dealers were dealing outside the school... oh, wait.
If he were in the South, I'd assume he was cooking meth. Not sure about Wisconsin. Either way, that seems to be the usual cause of random buildings exploding these days.
So how do they unlock them? :-D