Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[BFEL]

If anyone is interested there's a decent webcomic that covers A.I. stuff called Freefall. One thing I find particularly interesting about it is that it also defines genetically engineered organisms as "A.I." and builds Asimov's laws into their brains. The protag is a wolf-girl, and we get many moments of how horrible that is for her.

[Puzzlemaker]

I read freefall, it's pretty good.

[Besserwisser]

I was never talking about something like the three laws, I was suggesting that several ounces of thermite or an equivalent remote kill system be implemented, also, why the hell does everyone think that networking or otherwise allowing an AI outside of it's own head (figuratively of course) is a good idea?  Anyone who thinks in the even remotely long-term would be able to see the myriad flaws in that logic and go, "No, you can't hook the super-intelligent self-improving computer up to the internet, it doesn't need to look up kitten videos and porn."

The problem with that is that you have to interact with the AI somehow, otherwise what's the point? And this AI will be pretty much as persuasive as it wants to be. If it wants to be free, it will be eventually. The AI box is a thought experiment that deals with precisely that.

[NullForceOmega]

The problem with that is that you have to interact with the AI somehow, otherwise what's the point? And this AI will be pretty much as persuasive as it wants to be. If it wants to be free, it will be eventually. The AI box is a thought experiment that deals with precisely that.

So you give it sensors and a readout/vocal system.  There is no cause to integrate it with a data network, unless you WANT the Skynet scenario.  Also, that thought experiment is fundamentally flawed, it presupposes that the AI wants to escape and can actually be persuasive, I have no evidence to support those items in this instance.

[Persus13]

Why the hell does everyone think that networking or otherwise allowing an AI outside of it's own head (figuratively of course) is a good idea?  Anyone who thinks in the even remotely long-term would be able to see the myriad flaws in that logic and go, "No, you can't hook the super-intelligent self-improving computer up to the internet, it doesn't need to look up kitten videos and porn."

I can think of several reasons why you might want an intelligent computer to be linked up to the internet. I'm sure some places would love to have one running their firewall, for instance.

The Moon is a Harsh Mistress has an intelligent AI develop accidentally because the moon colony staff just keep hooking more and more stuff up to it.

[Shadowlord]

The problem with that is that you have to interact with the AI somehow, otherwise what's the point? And this AI will be pretty much as persuasive as it wants to be. If it wants to be free, it will be eventually. The AI box is a thought experiment that deals with precisely that.

So you give it sensors and a readout/vocal system.  There is no cause to integrate it with a data network, unless you WANT the Skynet scenario.  Also, that thought experiment is fundamentally flawed, it presupposes that the AI wants to escape and can actually be persuasive, I have no evidence to support those items in this instance.

That sounds like it might be sensible, but it seems unlikely that you're going to be able to run an AGI on a single computer, even a powerful one.

Let's look at what AlphaGo needed to beat Lee Sedol (and still lost 1 of its 5 games to him), bearing in mind that AlphaGo isn't remotely close to being an AGI, as far as we know - it's just really good at playing Go.

“We call it deep reinforcement learning,” Hassabis said. “It’s the combination of deep learning, neural network stuff, with reinforcement learning: so learning by trial and error, and incrementally improving and learning from your mistakes and your errors, so that you improve your decisions.”

In AlphaGo’s case, that involved splitting itself in half and playing millions of matches against itself, learning from each victory and loss. In one day alone, AlphaGo was able to play itself more than a million times, gaining more practical experience than a human player could hope to gain in a lifetime. In essence, AlphaGo got better at Go simply by thinking extremely hard about the problem.

DeepMind is also looking at releasing a version of the programme to run on home computers. “We haven’t figured out how, and we would need to optimise it a bit so it could fit on a normal machine … but certainly the intent is that this could be an amazing tool,” said Hassabis.

In its competition form, AlphaGo runs on Google’s cloud computer network, using 1,920 processors and a further 280 GPUs, specialised chips capable of performing simple calculations in staggering quantities, but a simpler version of the programme was built that could be run on one machine (albeit still one with 48 processors and eight GPUs).

I'm skeptical you'd be able to run an AGI on a single computer unless it was designed with physical efficiency approaching that of living brains, and even then there are lots of living critters that appear, to us, monumentally stupid... (having likely simply evolved programmed behaviors in reponse to certain stimuli because it's most likely to keep them alive against the threats they have faced in their evolutionary history, whereas we and a few other species learn (but of course we do stupid things too))

[Culise]

That said, even if you set up an entire network dedicated to running your AGI, there is still no need to connect whatever network your AGI is running on to the Internet.  Isolated networks exist all over the world; that's why, f'rex, Stuxnet was designed to be spread by USB drives rather than just networks.  Physical isolation obviously isn't the be-all and end-all of data security (leaving aside the inevitable case of user error/sloppiness, there are actually certain theoretical possibilities for malware to hop air gaps), but it's definitely much better than wiring your WMD arsenal and military C3I interfaces into the broader Internet. 

That said, this doesn't seem to be a case of a case of "research strong AI, get immediate revolution."  As a late-game threat, I suspect strongly that you'll typically be running your AIs for quite some time with no issue until one goes off the deep end and sets off a chain reaction/some virus metastasizes in your AIs and drives them nuts/some neo-Luddite faction taking power drives them into a preemptive war of survival/et cetera.  That's purely speculation on my part, but the longer the timespan between the development of strong AI and the Skynet scenario, the more likely it is that people are actually going to start using AIs for everything from network security - as noted by Persus13 - to your military - Skynet or AM in fiction, yes, but both of these had valid reasons behind their original development, whether it's akin to the Dead Hand/Samson Option or an improved coordination system on the NORAD model - to research and development.

[Cruxador]

The protag is a wolf-girl

Well that sure doesn't recommend it highly.

[BFEL]

The protag is a wolf-girl

Well that sure doesn't recommend it highly.

I'll admit that I realized this as soon as I wrote it. But its definitely not a furry fic.

[Orange Wizard]

But its definitely not a furry fic.

Aww.

[Criptfeind]

Assuming it's as intelligent as a human, it'll eventually convince SOMEONE to let it free.

You know, I've seen this argument a lot lately, (this is at least the third time this month on bay12 someone's said this) but I really don't think this logically follows. It's IMPOSSIBLE to simply not be persuaded? That seems an insane thing to say to me. Sure, the AI is going to be massively intelligent, but it's not going to be omniscient. Hell, even if it was omniscient, omnipotent doesn't necessarily follow.

Unless you mean eventually in the sense that in a infinite universe everything eventually happens (Which I don't thiiink is the case???). The idea that a single AI is absolutely guaranteed to talk it's way out of confinement is... Well, silly. 

Now, to bring it back around a bit, on the scale of Stellaris where there's going to be many many many AI trying to talk their way out of the box. Sure somewhere in the galaxy it's possible someone fucks up. But for a single case... No. Not at all is it a 100% that it will talk it's way out. (well, equally it's not a 100% that it fails to talk it's way out. But honestly probably pretty low.)

[Criptfeind]

Well, that's probably because I can't actually tell if this conversation is still about Stellaris and a galaxy scale magical sci fi civilization or about a single AI trapped in a box in the real world.

In the first case, yeah, that's a pretty good reason, persuasion, for the eventually AI rebellion. In the second case that's an argument I've seen multiple times on bay12 and it doesn't get less nonsensical with time.

[Greenbane]

Personally, I'm still talking in the context of Stellaris, and a galaxy with a sizeable number of species experimenting with artificial intelligence.

An AI in an isolated box would be nice and safe to study in a theoretical, pressure-free context, but I'm not sure just how practically useful it would be, in the end. Perhaps for very specific, very restricted purposes.

As I said earlier, you can't simultaneously impose very high security and be able to take anywhere near full advantage of an advanced AI. You're toying with a very delicate balance between safety and effectiveness, and eventually someone will get greedy/reckless. And there's also the matter that unpredictability would increase with complexity, and it'd become progressively more difficult to keep an ever-advancing AI shackled.

There are many possible scenarios, but I'm picturing a corporate/government environment: the organization has heavily invested in the AI project, but it's being conservative as far as security is concerned. Perhaps too conservative, someone with decision-making power thinks: the project is costly and it's not producing enough results. It could be axed, but then all the investment would've been for naught, and there are reports of competitors having better success. Or perhaps there's no competition, but there's pressure to keep developing and advancing. The technicians would complain, but there's jobs on the line, and the executives' minds are on profit and results rather than the theoretical dangers of loosening security.

[Shadowlord]

greatorder: IIRC that 'experiment' you're thinking of isn't any kind of replicable scientific experiment: it was Eliezer Yudkowsky (playing the AI), challenging people to a 2 hour game with specific conditions. He didn't release the transcripts, etc, etc. (He played 5 games, won the first three, lost the next 2, and then "called a halt to it")

[ZeroGravitas]

Hell, I remember reading about a little thought experiment thing where two researchers tried to convince the other to 'let them out of the box', acting as if they were an AI in an isolated system, and one of the researchers successfully convinced the other.

So like... a short story?