In this case, it's academic. The operator wasn't watching the road.
If you touch the brakes everytime there is a possible obstacle, you would be on them _all_day_ (in city traffic or side streets). Your back to 'no additional information'.
As always, you have to be careful not to train to artifacts. But artifacts exist in both virtual worlds and real world sensors. An advantage of virtual testing is you can isolate subsystems (or not). e.g. You don't need to render the scene and feed that to the camera/image analysis pipeline, you can just feed to object locations to the navigation.
Shitty drivers can't drive where they _learned_. Adults that have been around can navigate between massholes and origonions. (the two extremes of American bad drivers, still pikers globally, Mumbai laughs at Boston driver's shenanigans.)
How do you propose to validate a pipelined set of neural nets' training? This isn't just a message loop and a giant case statement.
I'd start by setting up a transparent, adversarial system in a virtual world. Let the public earn money by providing data sets that trigger bad behavior in the 'AI'. Good fun.
If the purpose is to test, you can't disconnect it before it makes a mistake.
But video shows the 'safety operator' was not watching the road. Which is understandable. Try watching a CNC machine with you hand hovering over the e-stop button, see how long you last. I guarantee you, it won't be an 8 hour shift of alert watching. Then again, he was in the car. You'd likely watch a lot better if your skull was in reach of the tool.
Testing autodrive cars on the road is hubris. Build an artificial test environment first, let it run for virtual giga car-years. _Constantly_ collect new data for it with manually driven cars and traffic cams. Don't bulshit yourself, you won't be 'done and ready' in a year. Test it against Mumbai traffic...check that, maybe not, that's how you get Stephen King movies realized.
I'd say just tune it to have a high false disconnect rate. In other words, force the driver to take control very very regularly, make turning it back on more than on button.
But then, what's the point?
I bet liability lawyers, plus promises made and cash accepted, end at that place (shitty 'autodrive' that disconnects constantly, to keep you awake).
One example of someone outside the EU having their accounts frozen (without being declared a terrorist org)?
The EU doesn't have the teeth.
I would suggest not keeping money in EU banks, duh. Unless your family has an old numbered swiss account that's still covered by confidentiality. But if your an old money tax evader, your likely not running a web site.
Slashdot Mirror
In this case, it's academic. The operator wasn't watching the road.
If you touch the brakes everytime there is a possible obstacle, you would be on them _all_day_ (in city traffic or side streets). Your back to 'no additional information'.
As always, you have to be careful not to train to artifacts. But artifacts exist in both virtual worlds and real world sensors. An advantage of virtual testing is you can isolate subsystems (or not). e.g. You don't need to render the scene and feed that to the camera/image analysis pipeline, you can just feed to object locations to the navigation.
Shitty drivers can't drive where they _learned_. Adults that have been around can navigate between massholes and origonions. (the two extremes of American bad drivers, still pikers globally, Mumbai laughs at Boston driver's shenanigans.)
Many cyclists are assholes, future Darwins, but you can't just run them over. She was in the road long enough to cross a lane and a half.
Roads are dirty places, the better your sensor array, the more signals it will have to see and decide, hopefully correctly, to ignore.
It's a software guy alright.
It's you! We've pinned the whole deal on whatever code we can prove you wrote and abandoned. Bet you didn't expect that code to end up in an 'AI'!
How do you propose to validate a pipelined set of neural nets' training? This isn't just a message loop and a giant case statement.
I'd start by setting up a transparent, adversarial system in a virtual world. Let the public earn money by providing data sets that trigger bad behavior in the 'AI'. Good fun.
If the purpose is to test, you can't disconnect it before it makes a mistake.
But video shows the 'safety operator' was not watching the road. Which is understandable. Try watching a CNC machine with you hand hovering over the e-stop button, see how long you last. I guarantee you, it won't be an 8 hour shift of alert watching. Then again, he was in the car. You'd likely watch a lot better if your skull was in reach of the tool.
Testing autodrive cars on the road is hubris. Build an artificial test environment first, let it run for virtual giga car-years. _Constantly_ collect new data for it with manually driven cars and traffic cams. Don't bulshit yourself, you won't be 'done and ready' in a year. Test it against Mumbai traffic...check that, maybe not, that's how you get Stephen King movies realized.
I'd say just tune it to have a high false disconnect rate. In other words, force the driver to take control very very regularly, make turning it back on more than on button.
But then, what's the point?
I bet liability lawyers, plus promises made and cash accepted, end at that place (shitty 'autodrive' that disconnects constantly, to keep you awake).
I understand that programmatically telling a blowing plastic bag from a child's toy is difficult.
But she (and her bike) were clearly large enough to damage the vehicle. Even if the code saw her as debris, the car should have avoided it.
I think the code had to have dismissed her as lens flair or something similar.
And yet you blame Facebook, after you do those very things.
You realize we're talking about bog standard SD cards?
The vendor locking ship sailed years ago. IIRC even Sony is using standard flash.
We all know you're just repeating derp you heard somewhere else.
Intel used to be a great engineering company, not for 20 years now. CPU is commodity, Intel is now a marketing company.
What? Chinese grad students did _not_ have polit officers in the USA 20 years ago, or 30 years ago, or 40. Perhaps before that, I don't know.
They did generally watch what they said in public and especially to any media.
Cardiovascular Russian roulette is the world's most popular X-sport.
An autopilot that 'works great', until it doesn't is the _fucking_problem_.
Humans are shitty at 'sitting and watching', strangely idiots are actually better at this kind of task.
Your fucking scumbag government doesn't have the balls.
One example of someone outside the EU having their accounts frozen (without being declared a terrorist org)?
The EU doesn't have the teeth.
I would suggest not keeping money in EU banks, duh. Unless your family has an old numbered swiss account that's still covered by confidentiality. But if your an old money tax evader, your likely not running a web site.
Long since happened. I'm describing the current status quo.
I will note that _they_ still haven't been able to block piratebay.
More like 'Go fuck yourself eurotrash'.
What would you say to an American cop that wanted to search your EU located servers based on American laws?
That's the same answer the EUcrats will get.
The little ones will ignore the EU, just as they ignore laws from Thailand and Saudi.
We'll see if the EU 'great firewalls' them in mass.
This law doesn't say that. Next the EU will block websites that simply ignore their overreach.
That will be the 'screw you'. But VPNs.
Perhaps the VPN operators in the EU will comply, but I wouldn't count on it.
You don't have to comply, any more then you have to comply with Saudi laws (unless you set foot in their jurisdiction).
Unless you have a server or office in the EU, they don't have shit.
The law should be ignored by all non-EU web sites.
Exactly, just don't have a presence in the EU and they can pound sand.
They are responsible for their own investment stupidity.