OOTS #1172 - The Discussion Thread

[mjasghar]

Please delete if this breaks rules
Recently in the uk the post office was privatised
Now they introduced a software system that managed accounts and tolls for all the post offices - many of which are franchises run by ordinary people who are more about providing a service in rural areas that aren’t economical
The system went wrong and accused people of embezzling
The post office has special legal methods for historical reasons so the people were effectively told the software is never wrong so you are guilty
This ended up with people in jail (some still are) or being forced to put their own money to balance it and only now is being addressed due to a class action taking it to the highest court
So pardon me if I think self driving cars is rubbish
In fact the whole Boeing 737 Max is the same issue - the software took over because the sensor got it wrong and the company hadn’t explained adequately how to override and saved money by only having one sensor

[littlebum2002]

I actually don't think you even NEED that much machine learning to have working self-driven cars - at least in the scenario that ALL cars are computer-controlled.




I'd go out on a limb here and say that I think computer driven cars would already be safer in my scenario.

City A: only human controlled cars.
City B: only computer controlled cars (setup done with technology available today, with highest standards in mind, and with people who want to do a good job).

Which city do you think would be safer?

City B, by a WIDE margin. Computer controlled cars are already safer than human cars, and they don't yet have the ability to communicate with each other. Once they can do that, they will be so safe we'll wonder how we managed to do without them. Imagine a car having to brake for a piece of debris on the highway and all the cars behind it immediately braking as well, instead of each car braking only when they react to the brake lights in front of them, until one person is on their phone and causes a pileup.

I never will understand how this scenario is even possible (and it is quite possible, and has been done with individual cars on the road right now). Why the heck do cars allow remote instructions while operating? Or indeed, any kind of outside instruction that doesn't need to be confirmed by someone in the vehicle prior to applying? The whole scenario is not unlike xkcd's antivirus in a voting machine adage.



Grey Wolf

I mean, vehicles will definitely need to communicate with other vehicles. See my scenario above. I'd imagine there must be some sort of a way to block off that communication from the internal computer of the car, therefore giving a hacker who compromises the information the car is receiving from taking over control of the car,but I don't know enough about computer systems to say for sure.

The autopilot might be a better driver than the typical human

If we could take current autopilot technology and run it for 20 years against 1000 randomly chosen drivers and record every single accident, every narrowly avoided accident, every time they veer out of their lane, every time they slam on their brakes, every time they aggressively change lanes or accelerate, and every time they hit a pothole or speed bump at a high speed or swipe a curb and potentially damage their car; I would bet every dollar I have now and every dollar I will earn over those 20 years that the autopilots will outperform all 1000 of those drivers. They never get distracted, they never get tired, they never get angry, they never get drunk, they have almost instantaneous reaction time. There is absolutely no chance that, over an extended period of time, a human driver would EVER outperform an autopilot.

And that is using today's technology. Imagine how amazing they will be 20 years from now, when the difference between that technology and today's technology is comparable to the difference between an iPhone X and that Nokia that came out 20 years ago that could only play Snake and the memory game.

[Jannoire]

Been re-reading the strip and it got me thinking: what happens with someone who is turned to stone? I mean, is he dead and his soul goes for the afterlife, or his soul stays in the petrified body until it's destroyed or he "dies of old age" (since in DnD people have a precise expiration date)?

If we suppose Durkon's petrification means he's dead, was his death honorable? Yes, he died sort of in battle, and fighting for the sake of all dwarves, but he also died breaking dwarven law, so I can see this going one way or another...

And what if the crux of their backup plan wasn't just about the hammer returning, but also about him dying and going whererver he needs to go? (and I have no idea where that should be).

Since the petrification only lasts until the meeting is over, I'd wager the soul doesn't go anywhere.
Plus, way back in the beginning, Nale told Celia that she was petrified for thousands of years and she believed it. She may be an outsider and petrification might not be that common, but I guess that's how it would work.
Be turned to stone, lose all your senses and wake up a splitsecond later, whether it lasted centuries or a minute.

Edit: better example would be Haley, who got turned to stone during the fight in Bleedingham. She was petrified mid-sentence and completed the sentence upon restoration. She didn't know where she was and needed to be filled in by elan.

[Grey_Wolf_c]

So pardon me if I think self driving cars is rubbish

I fail to see the connection. At all. "Laws screw over people" is a universal issue, not one particularly connected to the issues of AI driving. It doesn't take much googling to find similar catch 22 scenarios for regular car driving.

In fact the whole Boeing 737 Max is the same issue - the software took over because the sensor got it wrong and the company hadn’t explained adequately how to override and saved money by only having one sensor

No. The Max 737 problem started way before the software made it difficult to override. The core of the problem was that air transportation companies wanted to skimp on retraining their pilots. Which meant keeping the 737 cabin, rather than update it to a modern one with more complex controls and feedbacks. So when they had to lift the wings to fit larger engines, and thus put in a new tilt controller, they had to do so with 60s era UI, which meant a pilot light.

And then they made the pilot light optional, again to allow the buyer to save money by not buying it.

Blaming the problem on the software rather than in companies being skinflints is missing the whole point.

I mean, vehicles will definitely need to communicate with other vehicles.

Your example of AI training against 1000 driviers doesn't seem to rely on the statement that "vehicles will definitely need to communicate with other vehicles".

If they need to transmit information, they can do so with the same methods we use to day: blinking lights. Which could be made more complex, if need be (for example, have a number indicating the lane they wish to occupy, rather than simply "right or left"), but that's not the same as transmitting instructions on how the other car must act. Just information on their own intentions.

Grey Wolf

[Dion]

So pardon me if I think self driving cars is rubbish

I think you underestimate just how quickly computers are learning to adjust their behavior to maximize key performance indicators. You give them some measurement to optimize, let them train themselves to optimize it, and set them free. Computers are unbelievably awesome at tasks like that!

Once we ask computers to build personal transportation devices that maximize the only two performance indicators that matter, I think you’ll be amazed at the results!

[mjasghar]

Oh I see
This is the all those people that end up dead as the AIs learn are acceptable or even desirable deaths argument

[Grey_Wolf_c]

Oh I see
This is the all those people that end up dead as the AIs learn are acceptable or even desirable deaths argument

Wow, straight for the strawman.

Good bye.

Grey Wolf

[mjasghar]

I think you underestimate just how quickly computers are learning to adjust their behavior to maximize key performance indicators. You give them some measurement to optimize, let them train themselves to optimize it, and set them free. Computers are unbelievably awesome at tasks like that!

Once we ask computers to build personal transportation devices that maximize the only two performance indicators that matter, I think you’ll be amazed at the results!

That’s what someone said
Let the AIs learn by themselves
No consideration of collateral damage as a result and even saying only 2 performance indicators matter
And going back to the 737 max - the issue is that sensor said the plane was stalling so the software took over regardless of the pilots
So if a self driving car sensors go wrong due to something as simple as bird poo suddenly dropping on them what happens?
Do you say oh the driver can take over? That would require a constantly alert driver who can override at any time for any reason - which invalidates the whole idea and causes issues with liability
It’s amazing that people can read a comic which puts forward a situation where blindly following a system leads to disaster are arguing for blindingly putting their lives in the hands of a system

[Schroeswald]

If they need to transmit information, they can do so with the same methods we use to day: blinking lights. Which could be made more complex, if need be (for example, have a number indicating the lane they wish to occupy, rather than simply "right or left"), but that's not the same as transmitting instructions on how the other car must act. Just information on their own intentions.

Grey Wolf

I mean we have things that can make communication quicker and lacking blinking lights, we already use them while driving even though we shouldn’t.

[Grey_Wolf_c]

That’s what someone said
Let the AIs learn by themselves
No consideration of collateral damage as a result and even saying only 2 performance indicators matter

Still a strawman.

And going back to the 737 max - the issue is that sensor said the plane was stalling so the software took over regardless of the pilots

No, that is not "the issue". Already explained.

So if a self driving car sensors go wrong due to something as simple as bird poo suddenly dropping on them what happens?

Same thing that happens if that happens to a human driver's windshield.

I mean we have things that can make communication quicker and lacking blinking lights, we already use them while driving even though we shouldn’t.

Fair. But my greater point is that any system which employs said information should not be capable of simply passing on instructions to the car. I.e. The car takes into consideration "that car over there is radioing its intention to change lanes, do I want to slow down", not simply "some radio signal from outside told me to turn off the engine right now, even though I'm in the middle of a highway, so I will"

(radio here used in the generic sense, not necessarily involving actual radio bandwidth signals)

Grey Wolf

[brian 333]

Oh I see
This is the all those people that end up dead as the AIs learn are acceptable or even desirable deaths argument

Assuming the many currently dying is more acceptable than seeking a remedy.

No technology has ever been pain free. Look up water heaters and clothes washing machines. Back when they came out they were killers too.

[mjasghar]

The evidence shows that the new Maneuvering Characteristics Augmentation System (MCAS) repeatedly pushed the nose down on both flights, driven by erroneous angle of attack (AoA) data from a defective sensor, while pilots struggled against the automation. The original flight crew manual and training materials omitted MCAS. After the Lion Air accident, Boeing issued a service bulletin, mandated by an FAA Airworthiness Directive, to update the crew manual with instructions and procedures to shut off the system.
Pilots struggled against the automation
I’m not sure why you are ignoring this and just shrugging it off as something else

[littlebum2002]

Your example of AI training against 1000 driviers doesn't seem to rely on the statement that "vehicles will definitely need to communicate with other vehicles".

If they need to transmit information, they can do so with the same methods we use to day: blinking lights. Which could be made more complex, if need be (for example, have a number indicating the lane they wish to occupy, rather than simply "right or left"), but that's not the same as transmitting instructions on how the other car must act. Just information on their own intentions.

Grey Wolf

In order to get the most safety and efficiency we can, they will need to communicate with other vehicles. My example was just saying that the technology we have now is already safer than human drivers. I was not saying that we can't make them even safer still.

It's not just about cars riding behind other cars (although that is a scenario in which communication between cars would drastically improve the driving experience). There are other ways in which communication between vehicles will make them safer and more efficient. For instance, this type of communication could lead to having intersections with no red lights or stop signs. Instead the cars communicate with a server at each intersection which ensures that the cars navigate the intersection as efficiently as possible without any accidents.

[mjasghar]

Assuming the many currently dying is more acceptable than seeking a remedy.

No technology has ever been pain free. Look up water heaters and clothes washing machines. Back when they came out they were killers too.

{scrubbed}

[Grey_Wolf_c]

The evidence shows that the new Maneuvering Characteristics Augmentation System (MCAS) repeatedly pushed the nose down on both flights, driven by erroneous angle of attack (AoA) data from a defective sensor, while pilots struggled against the automation. The original flight crew manual and training materials omitted MCAS. After the Lion Air accident, Boeing issued a service bulletin, mandated by an FAA Airworthiness Directive, to update the crew manual with instructions and procedures to shut off the system.
Pilots struggled against the automation
I’m not sure why you are ignoring this and just shrugging it off as something else

The issue is that 737 cabins are simply not capable of delivering the information about its software effectively, especially if the UI systems are optional. The problem is that Boeing needs to accept this and retire the design, but its customers don't want to have to retrain their pilots, so Boeing bends over backwards to try and retain the old system. So the software developers were given an impossible task of delivering information through an hilariously limited method. And then the bean counters allowed even that system to not be there. The issue is money, not software.

{scrub the post, scrub the quote}

And the strawmen jut keep coming.

In order to get the most safety and efficiency we can, they will need to communicate with other vehicles. My example was just saying that the technology we have now is already safer than human drivers. I was not saying that we can't make them even safer still.

Not disagreeing. But again, also isn't what I was going on about: there is a line between "offering information to the cars around for them to incorporate into their decisions" (which runs the full spectrum from blinkers to whatever it takes to not need traffic lights) and "giving instructions to the cars from outside that they must obey". Weirdly, we already do have the latter, and I just can't understand why.

Grey Wolf

[mjasghar]

When someone says it’s okay for people to die to get progress it’s hardly a straw man
Someone else must die for the greater good

[Grey_Wolf_c]

When someone says it’s okay for people to die to get progress it’s hardly a straw man

When someone hasn't said anything of the sort and you insist that they have, it is.

Grey Wolf

[D.One]

And since we've entered into AI discussion, is it possible that opening a hole in the ceiling is a strategy to call for a Skynet of allies?

[mjasghar]

When someone hasn't said anything of the sort and you insist that they have, it is.

Grey Wolf

{scrubbed}
Quote Originally Posted by brian 333 View Post
Assuming the many currently dying is more acceptable than seeking a remedy.

[Grey_Wolf_c]

And since we've entered into AI discussion, is it possible that opening a hole in the ceiling is a strategy to call for a Skynet of allies?

Durkon's daddy was a savant of mining. He knew exactly where to hit to cause a cave in.

Durkon's mom tells Durkon that he's just like his daddy when he throws the hammer to break the ceiling.

It will be really strange if the point is to do anything other than bring down the dome.

{scrub the post, scrub the quote}

The best I can do at this point is to politely direct you to the forum rules.

Grey Wolf

[danielxcutter]

V is in the chamber. From the looks of it, the domed room is right up against the edge of the mountain, as the hole in the ceiling opens straight to daylight. Unless V can Dimension Door to outside the mountain, it doesn't look like they'll be able to do anything with that hole.

For one, V probably hasn't bought those D-Door boots since they're not wearing them, and for another, vampire domination is a (Su) ability - Dispel Magic won't work.

Finally, some NPC's that aren't stupid enough to fall for it. I do so hate when the plot only works if the characters are idiots especially if it's for the same tired joke over and over again. And that ceiling is about to collapse on all the dominated elders who haven't voted yet.

Oh gods beyond, you have no idea how much I agree. Heck, the only reason the non-dominated elders were willing to look into the vampire problem after the vote was because they'd never imagined that their vote could carry such repercussions, and when things actually started to go weird, they figured out something was fishier than a room full of sahuagin almost right away.

[D.One]

Durkon's daddy was a savant of mining. He knew exactly where to hit to cause a cave in.

Durkon's mom tells Durkon that he's just like his daddy when he throws the hammer to break the ceiling.

It will be really strange if the point is to do anything other than bring down the dome.

Grey Wolf

Oh, I agree with that. I think there's more to it than just "bring down the ceiling", but I agree with that. (the goal of the skynet joke was to divert the discussion from a theme that seemed to be getting way down the offense lane...)

[Worldsong]

Been re-reading the strip and it got me thinking: what happens with someone who is turned to stone? I mean, is he dead and his soul goes for the afterlife, or his soul stays in the petrified body until it's destroyed or he "dies of old age" (since in DnD people have a precise expiration date)?

If we suppose Durkon's petrification means he's dead, was his death honorable? Yes, he died sort of in battle, and fighting for the sake of all dwarves, but he also died breaking dwarven law, so I can see this going one way or another...

And what if the crux of their backup plan wasn't just about the hammer returning, but also about him dying and going whererver he needs to go? (and I have no idea where that should be).

Going by 3.5e rules petrified is considered unconscious. From how the rules read you can stay petrified forever and still be perfectly fine as long as you're eventually turned back to flesh, with the only harm being if, upon being turned to flesh, parts of the statue/body are missing.

Also Durkon's specific petrification is stated to only last for as long as the meeting does, so that'd be a rather short trip for his soul if it went anywhere in the meantime.

[Frozenstep]

Also Durkon's specific petrification is stated to only last for as long as the meeting does, so that'd be a rather short trip for his soul if it went anywhere in the meantime.

Enough time to grab a pint while he waits?

[mjasghar]

Going back to cars - the biggest increase in survivability has been safety features
First seatbelts and now airbags
Airbags in particular have meant people walking away from even high speed accidents
Making accidents more survivable or letting computers decide acceptable levels of death - I know what I would choose

[Worldsong]

Going back to cars - the biggest increase in survivability has been safety features
First seatbelts and now airbags
Airbags in particular have meant people walking away from even high speed accidents
Making accidents more survivable or letting computers decide acceptable levels of death - I know what I would choose

I choose both at the same time.

Now someone show me where it says you can't have both.

EDIT: For the record I don't believe the main feature of car AIs would be that they decide who lives and who dies. And frankly even if they did, if you're in a situation where it seems unavoidable that someone gets hurt or dies a human also has to decide which option is best, otherwise the trolley problem would never have been conceived in the first place regardless of how accurate it is.

Is someone suggesting that with car AIs we're gleefully developing programs that would purposefully drive us to our deaths OUTSIDE of scenarios where a human would also be choosing who gets to live and who possibly dies?

[mjasghar]

So who programs the software that decides what an acceptable fatality rate is?
And who stops some rich person getting the software hacked so their life is always the priority?
Maybe this is a cultural thing but I’m amazed so many people are willing to let a computer rule their life

[CriticalFailure]

It's crazy to me that the Boeing thing was able to happen - my understanding is that the software was made more aggressive and the back up sensors were removed without there being additional review and oversight? I guess it goes to show how important it is to have these kinds of design decisions appropriately reviewed.

[mjasghar]

I choose both at the same time.

Now someone show me where it says you can't have both.

EDIT: For the record I don't believe the main feature of car AIs would be that they decide who lives and who dies. And frankly even if they did, if you're in a situation where it seems unavoidable that someone gets hurt or dies a human also has to decide which option is best, otherwise the trolley problem would never have been conceived in the first place regardless of how accurate it is.

Is someone suggesting that with car AIs we're gleefully developing programs that would purposefully drive us to our deaths OUTSIDE of scenarios where a human would also be choosing who gets to live and who possibly dies?

I mean, if you haven’t actually paid attention that’s exactly what is happening
https://www.technologyreview.com/s/6...olley-problem/

[Peelee]

Maybe this is a cultural thing but I’m amazed so many people are willing to let a computer rule their life

You think you already don't? In cars, no less?

During a crash, the vehicle's crash sensors provide crucial information to the airbag electronic controller unit (ECU), including collision type, angle, and severity of impact. Using this information, the airbag ECU's crash algorithm determines if the crash event meets the criteria for deployment and triggers various firing circuits to deploy one or more airbag modules within the vehicle.

Spoiler: And if you want to see how important those decisions are, check it:

Show