[Discussion] Implications of an AI society

[hal]

Originally Posted by Anthony

Copying an intelligence leads to a copy that will not differ significantly from the behavior of the original, at least until a considerable time period has passed. Thus, multiple AI copies working together are far less effective than multiple independent AIs, because they don't each think in their own way.

Instead of talking about a theoretical situation, we should all agree to a set of parameters for this discussion.

For instance, if the year is 2080, in theory, that would be around TL 10 in general. So let's pick a computer system, what its primary purpose is, fill in the data such as cost, weight, location (mobile or immobile) etc.

Then we should discuss the actual AI or Mind Emulation residing on the system and why it is there, etc.

Thoughts?

[Anthony]

Realistically, the solution is to cap both human and AI IQ, such that, at the same complexity, the max IQ of an AI is the same as the max IQ of a ghost. However, as long as we're stuck with humans allowing uncapped IQ, it doesn't make sense for AIs to be capped.

[Daigoro]

Human IQ is soft-capped at 20, isn't it?

But also, think about the different paths to getting a high IQ. For a human, it involves, say, 30 years of life experience and education. If you want an AI to have such a high IQ, you'd have to mimic that or program it somehow, which means it's not going to be cheap or easy to do.

[Kax]

Originally Posted by hal

I hadn't realized you were using the Socratic method when you asked the question.

:)

Apologies. Will use the Socratic font in future.


:P

[Kax]

Originally Posted by Daigoro

Human IQ is soft-capped at 20, isn't it?

But also, think about the different paths to getting a high IQ. For a human, it involves, say, 30 years of life experience and education. If you want an AI to have such a high IQ, you'd have to mimic that or program it somehow, which means it's not going to be cheap or easy to do.

The brain is involved too; I personally know people with 60 years of life experience I'd be loathe to trust doing their own taxes using a manual.

And adversarial/combinatorial neural net programming (assuming a good sim of the real world with at least a few pre-existing other AIs) can give a thousand years of experience in a few weeks.

[jeff_wilson]

Originally Posted by TGLS

All you'd have to do is spin up a few thousand AIs, letting them run for a few months, and then testing them for IQ (or some other indicator).

Citation needed.

Originally Posted by Kax

And adversarial/combinatorial neural net programming (assuming a good sim of the real world with at least a few pre-existing other AIs) can give a thousand years of experience in a few weeks.

Citation needed.

[TGLS]

Originally Posted by hal

Instead of talking about a theoretical situation, we should all agree to a set of parameters for this discussion.

For instance, if the year is 2080, in theory, that would be around TL 10 in general. So let's pick a computer system, what its primary purpose is, fill in the data such as cost, weight, location (mobile or immobile) etc.

Then we should discuss the actual AI or Mind Emulation residing on the system and why it is there, etc.

Thoughts?

Alright. Zotech, one of the early developers of AI, is working on a new version of the SecuritCommanderPRO AI (let's say it commands security drones) that they sell. Zotech recently purchased a brand new (TL 10) Megacomputer to handle all of their needs. It is Complexity 11. The SecuritCommanderPRO AI is a Non-Volitional AI with an IQ of 8, implying a Complexity of 6.

A software engineer hits upon the idea that the AI can self improve beyond its normal constraints, and so runs 1000 instances of the AI in a simulated environment so they can self improve (1000 instances of AI would require Complexity 9, and let's say 1000 simulations for each of them is also Complexity 9. Maybe they interact for practice.). He will allow this to run for three months, where upon he will test whether the AIs have actually shown any meaningful improvement over the other. Come three months later, at least a few of them are showing some kind of improvement (perhaps some have what GURPS would call Talent or higher IQ).

[Flyndaran]

I'd assume most A.I's learn in a similar way to us. As in they need interactions with other intelligences and the environment. They can't drastically self improve by just sitting in a "bare room" thinking.

[hal]

For those interested, I opened up a new thread regarding the rules as written in GURPS ULTRA-TECH as relates to AIs.

The new twist has to do with the Legality class of an AI whose IQ falls between any given range. The reason I posted that to a new thread instead of this one, is to open up a sort of "post playtest" kind of environment where those things in the rules cause you (the reader of GURPS ULTRA-TECH) to say "why wasn't this hashed out in playtest in the first place?" without necessarily saying "This should have been play tested further" Gotta be polite ya know?

In any event, reproduced here...

Originally Posted by hal

if the computer's IQ can't be ascertained based on its complexity limited IQ taboo trait, and the AI can refuse to submit to the test, then what?

And, if you really want to get into the nitty gritty? *evil grin*

Biological characters do not have to worry about being illegal on the basis of their IQ. Why do the player character AIs have to worry about that? To put a proper spin on this - the whole thing does not compute.

Looking at the numbers as well as the LC issues involved per GURPS, we get the following:

LC4 if IQ 6-8 (meaning that there is no issues involved at all)
LC3 if IQ 9-14 (Meaning that the AI has to be licensed)
LC2 if IQ 15-19 (Meaning that this is restricted, on the same order as military weapons)
LC1 if IQ 20+ (meaning Military Grade, restricted like heavy weapons)

Think about that for a second. Those rules imply that there are laws pertaining to AIs to the extent that IQs between 8 to 14 are largely acceptable (save for the fact that in relation to a human being, only the DUMB ones are allowed to exist unregistered).

Why do I get the feeling that the rules as written is trying to do two mutually incompatible things at the same time? On one hand, as an NPC "thing" that can be owned, they are regulated (by law) and treated as seriously as handguns might be, military grade rifles might be, or even military weapons might be. But no player character ever has to take that kind of an issue unless they are second class citizens who can be owned.

As I vet the book more thoroughly, with an eye towards actively BUILDING an NPC via the rules in GURPS ULTRA-TECH, I'm running into things that make me wonder "What where they thinking!"

So, back to you - HOW does one test an AI that is volitional that also deliberately avoids answering any questions that might give its IQ away and potentially make its own existence even harder?

What does the government usually do when it classifies something as LC 2 and finds anyone in violation of said possession laws?

[hal]

Originally Posted by Flyndaran

I'd assume most A.I's learn in a similar way to us. As in they need interactions with other intelligences and the environment. They can't drastically self improve by just sitting in a "bare room" thinking.

Unless an AI has an altered time rate advantage, they will be living in the same real time as human beings do. I think the reason Jeff was asking for a "Cite" a few responses upstream, is that the question of "how long does it take for an AI to evolve" needs to be carefully examined within the rules as written, or failing that, offer some reasonable alternatives based on the idea that any "robot" AI can be treated as a character. If so, then they have to have the same basic limitations as do biological entities.

Which brings up another issue. Throughout the years of reading books and watching movies, the joke has been that many of the alien intelligences are simply humans in rubber suits. How does one describe an alien intelligence that isn't a human in a rubber suit so to speak?