h2g2 The Hitchhiker's Guide to the Galaxy: Earth Edition

if I suppose, that this theory is behind the researches of artifical intelligence, using neural networks which are able to "learn" and their capacity (in theory) is unlimited...

Complexity has not historically been behind AI research but you are right, IMO, to make the connection. "Goo old fashioned AI" has been the result of a mainly top down approach. Artificial Life, on the other hand, some say (and I tend to agree) the main source of future AI advances, has been linked intimately with complexity almost since it's conception.

If by artificial life you're refering to all those robots they're building to mimic things that are alive, I'm so with you on that one. I've been thinking of ways to reverse-engineer the brain for so long, and those things are deeply inspirational.

i think the question is, wether the essence of human intelligence manifests in creativity, or not. Because if it does, I doubt that any artificial intelligence systems could ever "create" something "new", just think of any kind of art. Any attempt in this direction will neglect originality. The way the human brain can disperse things (events=logical steps with result), and put the parts together in a way that is unique. Can AI ever reach that? Or: will we notice if it does?

> I think the question is, wether the essence of human intelligence manifests in creativity, or not.

That depends how *you* define "intelligence" and there a lots of definitions as I was reminded in the "Playing God" threads.

> Because if it does, I
> doubt that any artificial intelligence systems could ever "create" something "new", just think of any kind of art. Any
> attempt in this direction will neglect originality.

There are three interpretations. 1. You accept that AI will reach this level, 2. You don't accept that AI will reach this level, 3. You work to find out.

As an interesting aside, ALife experiments have been yielding results that have 'surprised' their experimenters for at least a decade, finding novel (and often inexplicable) solutions to problems. Is that not "creating" something "new"? We're not talking art here, but what is the "logical" problem with AI being artistic?

It has been said to me that "we'll never model human thought because so many of our decisions are based on 'emotion'". Non-sequituer. Why can't we model emotions? It's only brain doping. Why can't we model the release of dopamine and seratonin et al. and have them modify an AI's behaviour? The answer is, of course, that there is no logical reason why we can't.

> The way the human brain can disperse things (events=logical steps
> with result), and put the parts together in a way that is unique. Can AI ever reach that?

Why not?

> Or: will we notice if it does?

Why shouldn't we? If the results are original we surely would, wouldn't we?

Re: emotions

All thought is brain doping at a certain level. Life is just DNA doping : )

Re: AI creativity

Si's right: AIs are already creative, in the sense that they acheive novel solutions to problems. What are the programs that engineers use to "evolve" solutions to their problems? I can't remember the name, but those guys are creative as Leonardo : )

I think the real test of AI will be artificial emotions, which will be really tricky and controversial to test...

Well,guys...thank you for the thorough answer - I am not an expert in this field.
Still, I cannot imagine that you can "model" emotions ( manifestating in pieces of art), because it's not just doping. They are the result of previous experiences (associations with past events), state of mind in the moment (chemistry, metabolism), etc. Too many factors with too difficult weights.

> Still, I cannot imagine that you can "model" emotions ( manifestating
> in pieces of art), because it's not just doping. They are the result of
> previous experiences (associations with past events), state of mind
> in the moment (chemistry, metabolism), etc. Too many factors with
> too difficult weights.

Experience (assoc w/ past events) - Memory: has been done.
State of mind - the current state of the AI (firing neurons or "true" clauses whatever the implementation is) also been done.

The problem is overall complexity, we've got all the ingredients.

It's funny with Si - he always makes my arguments A) before me and B) better than I would anyway - almost makes me wonder why I bother : )

That said, it's generally my belief that anything's possible. For instance, pre-historic iron filings would never have believed it possible that fleshy sacks of blood wrapped around a bone scaffold would be able to paint the Sistine Chapel's ceiling, but it happened, and given the complexity of the universe, it was essentially a digital choice: either it would or it wouldn't happen. In this reality, it did.

In the future reality I plan on being a part of, AIs will successfully model "human" experiences in every way, except that they'll be AI experiences.

Or to say that better, there will be AIs built by humans or our minions that will model human experiences to an indistinguishable degree. Wow, every time I say that, it gets more obscure.

Anyway, it'll happen for the same reason as the Sistine Chapel - someone's already imagined it, so it's become a digital choice for all them little quantum choice-makers: either it'll happen or it won't. Since people are trying to make it happen, those little quantum choice-makers will eventually let it happen.

My (somewhat crazy) opinion, anyhow...

>>Si's right: AIs are already creative, in the sense that they acheive novel solutions to problems. What are the
>>programs that engineers use to "evolve" solutions to their problems? I can't remember the name, but those guys
>>are creative as Leonardo : )
I'm not sure if they count as "AI", but I think neural nets would fit that bill. The idea is that you connect a number of input factors that you think might be important and let your program connect the inputs through a network of hidden layers that eventually spit out an output. You then set your system up with a big mess of known inputs and outputs and let it tweek the strength of each link in its network until it can reproduce the training set-- on a good day it will then be able to come up with the right outputs for new inputs. Often it is very difficult to figure out why a given neural network is able to solve the problems you give it, which raises some interesting questions about whether one would like to go ahead and use a "black box" technology which appears to work better than a lot of other available methods, but for reasons that we don't understand.

Right, what I'm thinking of is kind of a particular form of neural net... hold on...

my source tells me I'm thinking of genetic algorithms. The main difference in function between a neural net and a genetic algorithm us that a neural net looks over historical data and makes conclusions, genetic algorithms create new data and then test it for usefulness.