• Quick note - the problem with Youtube videos not embedding on the forum appears to have been fixed, thanks to ZiprHead. If you do still see problems let me know.

Worried about Artificial Intelligence?

Indeed. AIs only need to drive better than people. Which we are IMHO pretty near to. Especially if you consider the current boom of AI. Obviously, cars are not LLMs .. but all the hardware and infrastructure designed and made for LLMs will also be useful for all AI applications.
 
If we're ever going to transition to flying cars, AI will absolutely be a necessary part of that transition. Most people can't handle a vehicle in two dimensions, let alone three.
 
arthwollipot said:
If we're ever going to transition to flying cars, AI will absolutely be a necessary part of that transition. Most people can't handle a vehicle in two dimensions, let alone three.
Click to expand...

You envision an army of traffic-enforcing Terminator bots, flying around destroying bad air-drivers? Yes, that sounds acceptable. They should utter things like "my way or the die way" in Arnold's accent when they shoot.
 
theprestige said:
Fly by wire is already a thing. You don't need AI to translate a pilot's instructions into safe maneuvers for the aircraft.
Click to expand...
Yeah I'm not talking about fly by wire. Fly by wire is (as far as I know but happy to be corrected if I'm wrong) where the link between controls and control surfaces is electronic rather than mechanical. There's no intelligence there, artificial or otherwise.

For a normal person with no more training than a 17-year old getting their driver's license to control an aerial vehicle would a lot more than mere fly-by-wire.

Regardless, this is science fiction for now. My point in bringing it up is that AI would necessarily be integrated into daily life in any kind of future we might be able to imagine. We don't need AGI (though it is my opinion that we'll create AGI just to see if we can), but we will need AI.
 
arthwollipot said:
Yeah I'm not talking about fly by wire. Fly by wire is (as far as I know but happy to be corrected if I'm wrong) where the link between controls and control surfaces is electronic rather than mechanical. There's no intelligence there, artificial or otherwise.

For a normal person with no more training than a 17-year old getting their driver's license to control an aerial vehicle would a lot more than mere fly-by-wire.

Regardless, this is science fiction for now. My point in bringing it up is that AI would necessarily be integrated into daily life in any kind of future we might be able to imagine. We don't need AGI (though it is my opinion that we'll create AGI just to see if we can), but we will need AI.
Click to expand...

Depending on your definition of intelligence.

Fly by wire systems are often configured to prevent control surfaces or the plane from exceeding predetermined limits. Manual controls allow pilots to exceed recommendations (typically pilots would try to do that in an emergency).

Modern fighter jets are so unstable, the flight control systems over-ride the pilots all the time, pilots are not able to fly the plane at all without that assistance.

(In a fighter, the ability to make sudden changes in direction is very important, and the built-in instability assists with that.)
 
Checkmite said:
This is an academic problem that doesn't apply to real life, though. In real life, a driver who sees an imminent collision with no completely clear way to dodge it instinctively slams on the brakes to try to stop or at least slow down, and if there isn't enough braking room to avoid the collision then they collide. People are hit by drivers all the time, and outside of a terrorist attack or something it's never because the driver actively made a choice to hit those people instead of hitting something else. By and large, if an unavoidable collision situation develops suddenly, the driver isn't going to have time to make such a choice, and neither will an AI. The "superpower" of an AI car is that it will be able to avoid avoidable accidents by just paying attention all the time and slowing down or stopping in plenty of time.

So no, I don't think an AI car has to be loaded with "values" about what things it's less-okay to run over. That's useless information. An AI should be programmed to not run over anything intentionally and that's it.
Click to expand...

Ah yes, the Kobayashi Maru dodge. I know Captain Kirk, and you're no Captain Kirk. Sooner or later an AI is going to have to decide whether to swerve or power through.

How do you train an AI to anticipate how a war crimes tribunal will judge a particular question of collateral damage?
 
Checkmite said:
This is an academic problem that doesn't apply to real life, though. In real life, a driver who sees an imminent collision with no completely clear way to dodge it instinctively slams on the brakes to try to stop or at least slow down, and if there isn't enough braking room to avoid the collision then they collide. People are hit by drivers all the time, and outside of a terrorist attack or something it's never because the driver actively made a choice to hit those people instead of hitting something else. By and large, if an unavoidable collision situation develops suddenly, the driver isn't going to have time to make such a choice, and neither will an AI. The "superpower" of an AI car is that it will be able to avoid avoidable accidents by just paying attention all the time and slowing down or stopping in plenty of time.

So no, I don't think an AI car has to be loaded with "values" about what things it's less-okay to run over. That's useless information. An AI should be programmed to not run over anything intentionally and that's it.
Click to expand...

I actually mostly agree with your point, but regarding the idea that humans never actually face the tradeoff you're talking about: my girlfriend's father was a truck driver (mostly driving from northern china to Inner Mongolia) for many years. He also happens to love dogs. One day a dog ran in front of his truck and he was faced with the choice of either swerve/slam on the brakes, or run over the dog. He chose the former, the truck ended up slamming into a wall, he broke his knee and damaged the truck, had to go through several months of bed rest, lost his job and became unhirable.

A bit of irony: In the poverty that ensued over the next few years, while her mother tried to support their family with what employment she could get, they had very little food. My girlfriend often tells stories about going into the fields to collect crickets that she'd bring home for her mother to cook, or stealing corn from farmer's fields.
One day she came home from school to find their family dog had been cooked for dinner. Her mother is much more practical than her father.
She refused to eat it, though.
 
Roboramma said:
I actually mostly agree with your point, but regarding the idea that humans never actually face the tradeoff you're talking about: my girlfriend's father was a truck driver (mostly driving from northern china to Inner Mongolia) for many years. He also happens to love dogs. One day a dog ran in front of his truck and he was faced with the choice of either swerve/slam on the brakes, or run over the dog. He chose the former, the truck ended up slamming into a wall, he broke his knee and damaged the truck, had to go through several months of bed rest, lost his job and became unhirable.
Click to expand...

I wasn't there so I can't say for sure obviously, but I strongly suspect that the choice he made wasn't "hit that wall instead of running over this dog", but rather more likely something along the lines of "avoid this dog by swerving into that apparently clear area" and as a result of swerving he lost control of the truck and unintentionally hit a wall. It wouldn't be the same situation as consciously deciding to run into a wall, IMO.
 
Checkmite said:
I wasn't there so I can't say for sure obviously, but I strongly suspect that the choice he made wasn't "hit that wall instead of running over this dog", but rather more likely something along the lines of "avoid this dog by swerving into that apparently clear area" and as a result of swerving he lost control of the truck and unintentionally hit a wall. It wouldn't be the same situation as consciously deciding to run into a wall, IMO.
Click to expand...

Oh absolutely, I don't think he intentionally ran into a wall to avoid the dog, but there's still an element of "take a risky action to avoid the dog", as swerving or hitting the brakes is risky. As you say, I doubt he had time to fully ascertain the level of the risk.
 
It seems to me that if AI were capable of making ethical choices there are a couple of ways they'd be more usefully ethical than humans: they'd be able to think much faster than humans, being computers, so what to a human would be a split-second mostly-instinct response would be to an AI a carefully thought-out reaction; and also afterwards at the inquest/inquiry/press conference an AI could just print out all the thinking it did at the time in question, from the inputs it received to the exact process of reasoning it used to reach the conclusions it did and take action on them. Unless it were canny enough to edit its own thought history to protect itself from liability, AI is going to be the perfect witness for or against itself.
 
Maybe - but, I don't know, I kind of think that presumes AI is going to be thinking through its trolley problems like a human would only faster, and I'm not sure that's true at all.

I think the confusion is partly due to the way we frame the issue - we talk about AIs having "alignment", and teaching them things like values and ethical principles and then, I guess, just letting them mull those over and come to conclusions? But I don't see that as likely or even prudent; I see the more realistic course of action being just a line of code telling the AI that if there's a human in the way it should stop, end of story.

In other words, I don't see us trying to "teach AI ethics" and allowing it to make ethical decisions, I see humans making all of the ethical decisions long ahead of time and then just giving the AI a set of standing instructions.
 
Checkmite said:
Maybe - but, I don't know, I kind of think that presumes AI is going to be thinking through its trolley problems like a human would only faster, and I'm not sure that's true at all.

I think the confusion is partly due to the way we frame the issue - we talk about AIs having "alignment", and teaching them things like values and ethical principles and then, I guess, just letting them mull those over and come to conclusions? But I don't see that as likely or even prudent; I see the more realistic course of action being just a line of code telling the AI that if there's a human in the way it should stop, end of story.

In other words, I don't see us trying to "teach AI ethics" and allowing it to make ethical decisions, I see humans making all of the ethical decisions long ahead of time and then just giving the AI a set of standing instructions.
Click to expand...

The issue with alignment is imho mostly language problem.
We can make normal programs behave as we want, because we can program them exactly. Programming languages have basically zero room for interpretation. There is still difference between what the programmer wanted, and what he actually programmed .. but for example another programmer can check it and fix it.
We can't do that with AIs. We program them with training data and processes used. And at least for now, it means there are huge differences between what we want to achieve, and what's actually going on. They might seem to do what we wanted .. but in some rare condition they might not. And it's typically not that they do something we didn't train them to do .. after complicated analyzing it's usually AI doing exactly what we trained them to do, except we didn't realize.
It goes way beyond LLMs and spoken language .. but LLMs demonstrate this nicely as we think we can easily change their behavior with spoken language commands .. and we can easily see how they don't often take unexpected take on it. It's like every sentence has some chance of being misunderstood .. and with every next sentence the chance goes higher, until it reaches practical certainty. There is universal feature of training process .. the longer we train, the bigger the chance an AI will find some logical shortcut which will completely destroy the desired function. It is usually measured by testing intended function and the training process is stopped before it happens .. but some form and unintended interpretation may still remain, and there is no way to check for it.
Alignment is not that AIs want to do something else than what we trained them. They do exactly that. Alignment is an issue of us not being able to train them properly. It's our inability to express the problem exactly, with all edge cases. And it's basically by definition. If we could, we wouldn't need AI to solve the problem.
 
There's a danger when talking about the Alignment Problem that the argument will ultimately end up at Roko's Basilisk, and I'm not sure that's all that productive.
 
Roko's basilisk is imho very specific and convoluted case, which requires many things to work before we even get there. AI will cause damage by unexpected behavior.
 
arthwollipot said:
There's a danger when talking about the Alignment Problem that the argument will ultimately end up at Roko's Basilisk, and I'm not sure that's all that productive.
Click to expand...

I don't think there's any danger of that at all, actually. If you listen to 80,000 hours or go on LessWrong you'll see a lot of discussion of issues related to the alignment problem, but basically never see Roko's Basilisk come up.
 
Checkmite said:
Maybe - but, I don't know, I kind of think that presumes AI is going to be thinking through its trolley problems like a human would only faster, and I'm not sure that's true at all.

I think the confusion is partly due to the way we frame the issue - we talk about AIs having "alignment", and teaching them things like values and ethical principles and then, I guess, just letting them mull those over and come to conclusions? But I don't see that as likely or even prudent; I see the more realistic course of action being just a line of code telling the AI that if there's a human in the way it should stop, end of story.

In other words, I don't see us trying to "teach AI ethics" and allowing it to make ethical decisions, I see humans making all of the ethical decisions long ahead of time and then just giving the AI a set of standing instructions.
Click to expand...

Just giving the AI a standard set of instructions is a particular method of alignment. In the vehicle case "Just hit the brakes" is a particular answer to a large set of problems. Its not the answer that humans have, who sometimes try to swerve out of the way, other times hit the brakes, other times just keep going through. Your solution here has the virtue of being simple. Its at least conceivable that it's worse than the human average, but it's also plausible to me that it's not. Is it worse than the optimal solution that we could implement with an AI system with much faster reaction times and enough compute to look at the difference between one situation and another? It seems reasonable that at least with some technologies it is.

But it's at least probably not much worse and potentially about the same as the status quo.

Does this generalize to other types of decisions AI systems will have to make? Can we just build AI systems to have a simple set of standard instructions whose outcomes are predictable? That seems plausible if you're designing self-driving cars or domestic robots for cleaning your home, or factory workers. It seems much less plausible if your system is making movies or writing novels (what's the analogue here of a simple set of standard instructions in choosing themes or or deciding on how various characters are portrayed?), or running a paperclip factory.

Perhaps we shouldn't design AI systems to do those tasks, and instead should simply design tools that humans can use to improve their efforts at doing those tasks. "Tool AI" is one suggested solution to the alignment problem, but there is some concern that it just ends up at the same point when the tools become complex enough and the human defers most decisions to the AI's expertise.
 
Roboramma said:
Does this generalize to other types of decisions AI systems will have to make? Can we just build AI systems to have a simple set of standard instructions whose outcomes are predictable? That seems plausible if you're designing self-driving cars or domestic robots for cleaning your home, or factory workers. It seems much less plausible if your system is making movies or writing novels (what's the analogue here of a simple set of standard instructions in choosing themes or or deciding on how various characters are portrayed?), or running a paperclip factory.
Click to expand...

I think that's more an illusion of a problem than an actual problem, coming from the fact that we use the single term "AI" to describe these two separate, completely unrelated types of systems that work in very different ways and are intended to do entirely different things, which is really weird and confusing. I don't think there is any common DNA whatsoever between for instance an automated car control program and a LLM. But we call them both "AI" and then we get bogged down trying to come up with principles or guidelines or solutions that somehow have to cover both at the same time.

For instance, the "trolley problem" belongs in a discussion about self-driving cars but is completely irrelevant to a discussion about an LLM like ChatGPT.
 
Last edited:
You must log in or register to reply here.

Back
Top Bottom