OK this is bordering in philosophy, so I might post something similair in the philosophy section, but this will deal more with the scientific side.

How would one program a computer to:
(1) be self-aware
(2) come to a "best possible" solution when no perfect solution is findable
(3) has a desire to learn
(4) can feel emotion
(5) sends its thoughts through logical loops which consider previous memories and thoughts
(6) can formulate opinions based on facts

But that's not what's going on. We have only the output to evaluate. To change the analogy, imagine that we are measuring some form of output of the game (say, the sound). Now, if we have two boxes, one with a real game being played inside, and one with a simulation being played, and we listen to both. Why would one be qualitatively different from the other?

Edit: To anticipate on a discussion that may follow: If you were to find out that a friend of yours was in fact (and had always been) a robot, would you consider yourself deceived for all the years you had known him/her?

This post has been edited by zaragosa: Sep 3 2004, 01:06 PM

That is exactly what is going on. The fact that the process is black boxed just means that it is even more difficult to know whether the system in question exists inside of the set of circuits and makes it even less likely.

Let's look at your example. From outside the box they may sound the same, but one is simply reproducing the sounds based on a programmed and highly algorithmic pattern. The other is producing them as a consequence of playing the game. In the first case the game itself doesn't have to be played anywhere, in the second it is being played in real time. That's a huge difference.

Why are they different? One is a model, one is an actual system. One is using a set of algorithms to anticipate and produce sounds in an attempt to mimic the game, the other actually involves a game being played.

Now, given a set of changing inputs and a set of outputs (rather than just analyzing a set of outputs), we can analyze it the same way we analyze a black boxed circuit and, as a matter of faith, I do not believe a computer can beat a set of human judges. Enter the Turing Test. Show me an AI which is close to passing an unrestricted Turing Test, since you seem to think we are "practically there" on creating a strong AI.

Last year the first place of the AI's in the Loebner Prize was Jabberwock. Not even close.

You:...Where do you live?
Jabb:...In the Great Realm of Imagination, care to meet me there?
You:...Sure, I'll meet you at half a figment around the bend from the Disney ride.
Jabb:...You think you can meet me? Not that I think you can't, guest

You:...You're flawed.
Jabb:...Sure thing. Most of our suspicions of others are aroused by our knowledge of ourselves. But I just felt I had to make that point strongly.

You're missing the point. I'm saying that intelligence is not in the process, but in the output itself (which is of course identical if the model is good). Consider: If you were to find out that a friend of yours was in fact (and had always been) a robot, would you consider yourself deceived for all the years you had known him/her?

I must disagree. Intelligence cannot simply be the output. Otherwise, we could say that computers today are "intelligent". Same with calculators. We throw in some numbers, and they return answers. We could do the same with a mathematician working with pen and paper. Both will return an answer (Presumebaly correct), but its a far cry to say that calculators are intelligent.

This post has been edited by Nalvaros: Sep 5 2004, 07:40 AM

Zaragosa, you have claimed that we are "practically there."

Please show me where we have come close to having a computer "win" an unrestricted Turing Test such as the one used in the Loebner Prize.

Sure, computers today aren't intelligent by any practical definition of the word, but if intelligience is in more than the output, how would we tell today if a computer program, which seemed to display all the attributes of real intelligence (self awareness, capable of learning, <insert other criteria here>) in it's output, is really intelligent or not? Assuming we understood perfectly all that went on inside of it's `brain', we wouldn't be able to compare it to our own (or any model of how intelligence should work on the lower level).

practically, adv.
3: All but; nearly; almost

The output rendered by computers today isn't yet the same as that from beings we would normally classify as 'intelligent'.

I quote myself:



Emphasis added.

I never implied that the evolution towards AI was a gradual process. The hardware being relevant only to the degree that it can make fast calculations, the last hurdle is software. So, in my view, what's left is a conceptual, not a technological feat. That's why I say I think we're practically there.

i disagree, intelligence is output. even with humans, all output is a result of input at some point. doesn't mean stimulus->immediate response. hell, you could be walking down the street and remember when you were 5 year old something happened and this causes you to thing about that experience and some others and produce an output, and this thinking may be intentional or even hard-wired (the primitive brain is involved). so in that matter inteligence is a "black box" type scenario

"Thinking" or "conscious"? Or are they the same?

Cause then you get into the whole realm of what consciousness is, and to my knowlege, no one really has a good definition of that...

Neither 'conscious', nor 'thinking' or 'intelligence' have ever been satisfactorily defined in my opinion. By diagnostically used definitions of intelligence (Wechsler tests etc.), even the earliest computers with some elementary programming and a functioning interface have us beat hands down.

Familiar with Searle's Chinese Room?

For those who don't know.

Baumgartner summarizes it:

In other words, we need a conceptual leap that we haven't made yet and which may not even be makable.

That is not, in my book, "practically" there by any stretch of the imagination.

Yes. I could reformulate it like this: As soon as we could program intelligence into a computer, we could 'program' it into a contraption with water running through tubes and valves. Surely that cannot be intelligence!
If we are to take that as an argument, it is an emotional one at best.


Conceptual leaps happen every day. Someone might have stumbled across AI a few months ago and might now be working on getting it published (or patented...). Since we are already where we need to be (and beyond) hardware-wise, and what is left is essentially sitting down and programming (and educating the thing), and with no reason to assume anything so supernatural that we could not reproduce it, I stand by 'practically there'. But you may call it what you like.

"You insist that there is something a machine cannot do. If you will tell me precisely what it is that a machine cannot do, then I can always make a machine which will do just that!"
- Von Neumann

Nope, you are confusing the basic argument with a later implementation which was to counteract a specific objection. Using that extension, however, Searle argues that the system could be internalized completely within the operator and the operator would still have no understanding of Chinese in order to pass the test.

As Searle points out for that situation: the operator does not need to understand Chinese in order to give the illusion of understanding it based on a set of dictionary responses (this is the fundamental mechanism that modern AIs use in an effort to pass the unrestricted Turing Test, for the record). That is the foundation of the Chinese Room argument, and it holds water (if you will pardon the pun) quite nicely--the operator neither knows nor understands any Chinese, will not learn Chinese simply by passing these symbols back and forth, but to the outside observer will appear fluent.



Requiring one to get from point A to point B without any evidence that we are approaching such is, at best, a matter of blind faith.



It is just a small leap to gravitons as well...

Yet we haven't found them yet. We might find them tomorrow, but we might never find them. They might not even exist. Saying that we have "practically found them" would be scientifically irresponsible, and saying that we will discover them "any day now" is a matter of pure and unadulterated faith.



Quantum Entanglement != supernatural
Hypercomputability != supernatural (Penrose's argument).
The theory that cognition is noncomputable != supernatural.

Simply, the statement that the applicability of the Church-Turing Thesis has been overstated does not require appealing to anything vaguely supernatural.

This post has been edited by Llywelyn: Sep 10 2004, 11:00 AM

Why does that matter? I don't think anyone is trying to say that any individual part of the human brain possesses intelligence (or understanding).


Which is why I've added qualifiers to every post in the line of "in my opinion," and "I (personally) think." I wouldn't dream of denying that this is a belief. One which I believe to be a reasonable one (for reasons I have laid out), but a belief nonetheless.


If it can be sufficiently described, it can be simulated. If it cannot, what is the point of discussing it (beyond passing time)?

This post has been edited by zaragosa: Sep 10 2004, 02:58 PM

Searle already countered this objection when he claimed that the system could be internalized to the operator.



The leader of ALICE bot project, a finalist in the upcoming 2004 Loebner Prize and winner in IIRC 2000 and 2001, has said that "believe that in about fifty years' time it will be possible, to programme computers, with a storage capacity of about 109, to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning." This is not exactly a glowing endorsement of the "right around the corner" theory. IIRC he has also said that ALICE, while significantly more advanced, is not that fundamentally different than the original ELIZA.




Can consciousness be "sufficiently described"?

It also can't necessarily be computationally simulated unless it meets with certain criteria (such as being a discrete state machine). Hence where I talked about the applicability of the Church-Turing thesis.

what is knowledge?

what is meaning?




what is the difference between Data and Information.
Data is useless 1's and 0's - un ordered - perhaps in a stream - with patterns within but unresolved. - raw stuff.

Information is organised data. A network of facts.

Organisation adds meaning because organisation transforms raw 1's and 0's into something more: more than the sum of its parts.

At which point the operator, possessing the entire system, would be indistinguishable from a human, which was the goal.


Until otherwise demonstrated, I don't see why not. I'm working on it ;)

Aren't there some kind of artificial intelligence with computer games anyway?

In a limited way, just like in chess, yes.

Maybe we can build on that.

IMO, the way chess (et al.) AI works is much too simplistic to be used as a starting point for real AI. IIRC, chess AI is just a combination of heuristics and look-ahead stuff.

Well, humans use heuristics, too, we're just better at them. Computers get much of their playing strength from number crunching and encyclopaedic knowledge. For now, we haven't gotten computers to understand more abstract chess concepts like power bases and field stress...

But then again, I used to teach chess, and I didn't get some kids to understand those either...