Did Godel disprove the idea of artificial intelligence?

[Jay GW]

Has artificial intelligence been discredited by Godel's incompleteness theorem?

Hao Wang A logical journey: From Godel to philosophy, 1997

Godel's Incompleteness Theorem states that in any consistent formal system which is adequate for arithmetic there is a true but unprovable sentence.

Second Theorem

If an axiomatic system can be proven to be consistent from within itself, then it is inconsistent. No consistent system can be used to prove its own consistency.

Therefore, in order to establish the consistency of a system S, one needs to utilize some other system T, but a proof in T is not completely convincing unless T's consistency has already been established without using S.

The theorem does not imply that every interesting axiom system is incomplete. The theorem only applies to systems that allow you to define the natural numbers as a set.

Minds, Machines and Gödel

Gödel's theorem seems to me to prove that Mechanism is false, that is, that minds cannot be explained as machines. So also has it seemed to many other people: almost every mathematical logician I have put the matter to has confessed to similar thoughts, but has felt reluctant to commit himself definitely until he could see the whole argument set out, with all objections fully stated and properly met.

The Emperor's New Mind: Concerning Computers, Minds, and the Laws of Physics
by Roger Penrose

For decades, proponents of artificial intelligence have argued that computers will soon be doing everything that a human mind can do. Admittedly, computers now play chess at the grandmaster level, but do they understand the game as we do? Can a computer eventually do everything a human mind can do?

Roger Penrose--physicist, and Stephen Hawking - puts forward his view that there are some facets of human thinking that can never be emulated by a machine.

Penrose examines what physics and mathematics can tell us about how the mind works, what they can't, and what we need to know to understand the physical processes of consciousness. He is among a growing number of physicists who think Einstein wasn't being stubborn when he said his "little finger" told him that quantum mechanics is incomplete, and he concludes that laws even deeper than quantum mechanics are essential for the operation of a mind.

[ZouPrime]

Of course artificial intelligence is possible.

Humans exists. Humans are intelligents. At the very worst, we would only have to simply recreate from scratch an artificial human being, and thus obtaining an artificial intelligence. I say "at the very worst", because creating AI will, at the end, probably going to be simpler than that. Of course it's still well beyond our current technology, but it definitively settle the question of if it's possible or not.

"The Emperor's New Mind" isn't really a good book, and Penrose should have sticked to mathematics.

[Agemegos]

Has artificial intelligence been discredited by Godel's incompleteness theorem?

Nope. Gödel's Theorem is not even revelant to AI. Minds aren't axiomatic systems, and neither 'completeness' nor 'consistency' is even defined with respect to a mind.

I have read Penrose's book, and discussed it with Marvin Minsky. I don't find it convincing, nor even impressive.

[BlakeEM]

I have been very interested in AI for quite some time. Right now AI mostly just mimics what humans do, however I don’t feel this is that great of AI at all.

The way someone will have AI like a human mind would be by creating the things that drive the mind and not the individual decisions or things the mind will have to think about.

Example is build it self aware. Build it to want to self preserve. Build it to want to preserve others. Build in learning, rather than individual reactions to things. It has to learn the reactions its self.

Build in the chemical reactions in our bodies into artificial variants that do the same thing. Adrenaline, pain, and emotions... etc

They would have to create a new way of processing. Currently computers send zeros and ones that are then interrupted into actions and processes or things to display on the screen. A human like AI brain would need to be built is multi process many complex things at once. It would need to be built like our brain; with linkages to relevant things that form automatically the more we become familiar with it or deem it useful.

Only with all these things and more can we have AI like a human I would think.

Right now if you think about it, computers are basically the same as our brain; only the computer listens to what it’s told and does it. Both have short term memory (ram), long term (hard drive), processing of long term and short term memories (processor). Our brain just has the programming already setup that has evolved over millions of years.

I see no reason what so ever that we can’t do it in the future. I really feel if we wait far enough in the future anything would be possible with the infinite amount of technology that we would be capable of.

[Majestyk]

"Mind", in its current apriority, may well be on its way to becoming an archaic term. The more we isolate and identify the processes, pathways, and architecture of neurological systems, the less applicable "mind" becomes. "Mind", today may well be equivalent to "Heavens", in the Middle Ages, as both were used to label a construct whose true nature was basically unknown.

Although I can't say with certainty that we will ever reproduce intelligence in a non-organic system since, we haven't yet fully identified the process; I can say that we appear to be continuously increasing our understanding of the subject. That understanding points to the "mind" as a purely biological process and therefore, wholly dependent on the mechanism that generates it.

Goedel's Theorem does not separate the property of intelligence from the organism that manifests it. It claims that the property will not allow the consistently accurate processing of information. This isn't really news.

[baron greenback]

Roger Penrose--physicist, and Stephen Hawking - puts forward his view that there are some facets of human thinking that can never be emulated by a machine.


Penrose and Hawking are physicists. Good, well known, respected, brilliant physicists. They are not computer scientists.

Godel's models are based on binary logic. We've had computers that use logic other than pure binary since the late fifties when they were comprised of vacuum tubes. The mathematical processes known as "stochastic optimization" and "fuzzy logic" make godel's apparent paradoxes irrelevant.

From mathworld.wolfram.com, here are the definitions of stochastic optimization and fuzzy logic.

fuzzy logic: An extension of two-valued logic such that statements need not be true or false, but may have a degree of truth between 0 and 1. Such a system can be extremely useful in designing control logic for real-world systems such as elevators. See also: Alethic, False, Logic, True

stochastic optimization: Stochastic optimization refers to the minimization (or maximization) of a function in the presence of randomness in the optimization process. The randomness may be present as either noise in measurements or Monte Carlo randomness in the search procedure, or both. Common methods of stochastic optimization include direct search methods (such as the Nelder-Mead method), stochastic approximation, stochastic programming, and miscellaneous methods such as simulated annealing and genetic algorithms.

And as far as self-proofs are concerned, every compiler for every programming language since c# has had to be able to compile itself.

[Undercurrent]

(citing J.R. Lucas without providing a source (http://cogprints.org/356/00/lucas.html) which he should know to do by now :mad: )

Gödel's theorem seems to me to prove that Mechanism is false, that is, that minds cannot be explained as machines. So also has it seemed to many other people: almost every mathematical logician I have put the matter to has confessed to similar thoughts, but has felt reluctant to commit himself definitely until he could see the whole argument set out, with all objections fully stated and properly met.

This strikes me as just another one of those philsophy papers claiming that some aspect of science or mathematics proves something that the author really wishes was true in the hopes that the reader will lack the expertise to evaluate the claim and will just accept it at face value. I am especially moved to think so given the seemingly needless claim that "other unnamed people who actually have that expertise probably think so, too, and for some reason other than 'this claim does not stand up to substantial thought' have neglected to write about it" put right up there in the first paragraph.

Needless to say it pisses me off.

Goedel's theorems simply do not place any restrictions on artificial intelligence that they do not also place on natural intelligence. It is only if you assume, for no good reason, that human intelligence is trans-Turing that any of the arguments for AI's impossibility make sense.

The same technique is used with the creationist version of the second law of thermodynamics. The creationist "law" would prohibit any kind of order-creation anywhere in the universe, so exceptions are made for humans and gods. The real law, of course, applies to all objects in the universe irrespective of their anthropogenic value.

[Schneibster]

First of all, let me make my opinion clear.

Gödel's Incompleteness Theorem proves that it is impossible for any formal system that is sufficiently complex to reference its own theorems to be both complete and internally consistent. That is, either there will exist theorems that are legal in that system that cannot be derived, or there will exist at least one pair of theorems that are inconsistent with one another but that both can be proven under the rules of the system.

In natural language, we find constructs of the second type with relative ease. For instance, the following sentence:

This sentence is a lie.

presents exactly this type of problem. Can you feel your brain stopping as a result of this paradox? Is your thinking interrupted by this problem? Are you rendered unable to function as a result of this basic inconsistency in your world-view? I think not.

I have encountered bugs in computer programs that were the result of problems like this one. While they did cause a problem in a particular function within the program with a particular piece of data, they did not in any other way affect the program's operation. For anyone familiar with programming, the reason for this will be clear: modularization. Thus, the problem is restricted to a local area of the program and does not obtain the ability to ravage the system, as so many science fiction stories written by people who do not know very much about programming portray.

Similarly, both our minds and any AI programs we may construct are similarly modularized, and can or will be able to protect themselves from inconsistencies of this type.

In fact, Hofstadter (to whom I am indebted for his introduction to both Escher and the Incompleteness Theorem) in Gödel, Escher, Bach: An Eternal Golden Braid (see the S&S reading list) shows that intelligence may actually depend on this type of inconsistency; certainly, it depends on the ability of the mind to engage in self-reference.

So, to answer the OP, no, Gödel did not disprove the idea of AI.

Minds, Machines and Gödel

Gödel's theorem seems to me to prove that Mechanism is false, that is, that minds cannot be explained as machines. So also has it seemed to many other people: almost every mathematical logician I have put the matter to has confessed to similar thoughts, but has felt reluctant to commit himself definitely until he could see the whole argument set out, with all objections fully stated and properly met.Consider carefully that such inconsistencies must perforce exist in every formal system that you contain in your mind, but it does not seem to keep you from functioning well enough to write this. You might share this thought with yoru mathematical logician friends too. ;)

The Emperor's New Mind: Concerning Computers, Minds, and the Laws of Physics
by Roger Penrose

For decades, proponents of artificial intelligence have argued that computers will soon be doing everything that a human mind can do. Admittedly, computers now play chess at the grandmaster level, but do they understand the game as we do? Can a computer eventually do everything a human mind can do?

Roger Penrose--physicist, and Stephen Hawking - puts forward his view that there are some facets of human thinking that can never be emulated by a machine.

Penrose examines what physics and mathematics can tell us about how the mind works, what they can't, and what we need to know to understand the physical processes of consciousness. He is among a growing number of physicists who think Einstein wasn't being stubborn when he said his "little finger" told him that quantum mechanics is incomplete, and he concludes that laws even deeper than quantum mechanics are essential for the operation of a mind.Penrose has a bug about "microtubules" that he says account for human consciousness. My opinion is that this is what happens when a physicist fails to understand the inherent power of replication and the mechanisms of intelligence.

[Undercurrent]

Penrose and Hawking are physicists. Good, well known, respected, brilliant physicists. They are not computer scientists.

I might add that the quotes from them are much weaker. "There are aspects of human thought that we may not be able to figure out" includes such sentiments as "the human mind is very complex and hard to understand and it will probably be never worth the effort to hash out every minute detail".

Godel's models are based on binary logic. We've had computers that use logic other than pure binary since the late fifties when they were comprised of vacuum tubes. The mathematical processes known as "stochastic optimization" and "fuzzy logic" make godel's apparent paradoxes irrelevant.

I read this quote to my computer scientist coworker. He said "smack that guy for me". Nothing is trans-Turing. Your two examples are no exception (ironically, both are simulated on a standard Von Neumann machine in nearly all applications).

[Dark Knight Bob]

The idea of AI being impossible due to the impossilty of having a true self checking system falls down under a few major assumptions.

Human beings are not infinite. If you ran a human mind indefintely you might encounter an infinite loop effect. But because human minds work on such a complex level there is enough self checking to get through at least 80 years of life without any problems.

The human mind works in a parallel arrangement. This allows an exponential increase in the ability to run simultaneous "programs" that self check each other adding to the lifespan. godel's theorm was based primarily on serial computer logic. Whereby you could only run one line of code at a time.

A computer system is isolated from the world. Given a human mind that is setup in a similiar manner there is no way to tell if it would fall to the same problems. However the surrounding environment is always changing so simply havign a computer set up with 5 major senses it becomes a part of the universe itself and is no longer a closed system subect to eventual sterility. (you merely have to deal with entropy which everything has to deal with)

Admittedly, computers now play chess at the grandmaster level, but do they understand the game as we do?

The models used to explain how humans learn chess can be simulated on a computer. It is merely a basis of playing the game enough times to learn all the available patterns.

Based around the idea of "chunking" data in short term memory. If you've ever heard of the muller experiment you'll know about the idea that we can store about 7 items in short term memory and that's our limit. The more experience we have (drawn from long term memory) the more we can compress data to allow more information to be stored in STM. This is how magicians do memory tricks where they can memorise the order of cards drawn out of an entire pack.

Current chess models utilise this information to act as "human" as possible. So within the confines of the model itself, yes they do "understand" as well as humans do. Though an element of randomness is encountered on the humans part due to other factors that have yet to be modelled.

Fuzzy logic is merely setting boundaries for certain criteria. It doesn't really solve the problem at all.

Penrose and Hawking are physicists. Good, well known, respected, brilliant physicists. They are not computer scientists.

Like I said it is essentially the same mathematical system as that of a closed system subject to entropy.

The idea that humans have somehting "magical" that makes them above Godel's model is irrelavnt. If you can make a computer complex enough to survive at least 80 years then you can simulate a humans thought patterns just as well. It is irelavnt if it crashes after 120 years or so,

[PoodleLovinPessimist]

And as far as self-proofs are concerned, every compiler for every programming language since c# has had to be able to compile itself.

C# is a very new language; you're probably thinking of "C". In fact, C# is not self-compiling, nor is Java, AFAIK. But before then, even assemblers (an assembly language is a direct translation of a machine language to human-readable symbols) were self-assembling.

[PoodleLovinPessimist]

PoodleLovinPessimist cannot assert this statement to be true.

[Schneibster]

Penrose and Hawking are physicists. Good, well known, respected, brilliant physicists. They are not computer scientists.On this we agree.

Godel's models are based on binary logic. No. Gödel's Incompleteness theorem is not a model. It is number theory, in the purest sense. He proves conclusively that any formal system that is of sufficient power to allow self-reference must be either inconsistent or incomplete. I explained already what those terms mean. It doesn't matter if its decimal, hexadecimal, duodecimal, binary, trinary, quanternary, octal, or whatever. It doesn't matter if it's tri-state logic, or fuzzy logic, or if it uses stochastic optimization. If it can self-ref, it is either incomplete or inconsistent.

And as far as self-proofs are concerned, every compiler for every programming language since c# has had to be able to compile itself.First of all, C-sharp was the first one? Funny, I seem to recall Kernighan and Ritchie getting that trick done along about 1970 or so; and the last I checked, C# wasn't available until like 1999 or 2000 or something (sorry, I just don't keep up with "advances" in programming languages on inferior operating systems. Maybe I'm just lazy or whatever). And by the way, it wasn't just the compiler- it was the freakin entire operating system along with it. You might have heard of it- UNIX ring any bells for you? The first compiler was, of course, written in B, and each version was compiled on the previous version. If you play with Linux, you can do the same thing yourself.

But this is not proof of avoidence of the Gödel problem. You have to actually program in the language to find it. And it's not easy to make a program that shows it; C actually in and of itself is only just barely powerful enough to allow the construction of the G trick; and it's a hell of a lot of programming, so it's easier to use yacc and lex to construct a lexer and parser that show it.

If you like, we can demonstrate the problem with yacc and lex. It will take me awhile, but somewhere or other I have an example program for yacc and lex that will let me construct a C program that shows Gödelian incompleteness in C. Note that you cannot show the inconsistency; C, yacc, and lex are designed to guarantee that you will never see it; but the price is, of course, incompleteness.

On second thought, I think you might be able to google that up. Why don't you try that before I go digging through the archives for something eight or ten years old that I might not even be able to find?

[Bold]

I'm not a fan of strong AI running on anything resembling an electronic computer. However, I don't really buy Penrose's arguments. They seem to assume that brains are sound and know they are sound, and I think this is unjustified. In shadows of the mind, this soundness is qualified a kind of 'in principle' soundness making mathematics possible, but then that is no longer about the human brain, but mathematics.

I don't think fuzzy logic, parallellism, stochastic optimization or any other kind of calculations is going to solve the strong AI problem. If it is possible to do using those things, then it is possible to do in a binary, serial, deterministic way. This is obvious if you consider how computers work.

No, I am fairly convinced of Searle's position - that there is some physical process going on in the brain that we are going to have to understand before we get anywhere. It is not the calculations that might represent this process that matter, but the process itself. I am veering towards the problem of consciousness rather than intelligence here, but the two are probably related.

So Penrose's suggestion that intelligence is non-computable I think is unsupported, but he may still be right.

[Undercurrent]

The human mind works in a parallel arrangement. This allows an exponential increase in the ability to run simultaneous "programs" that self check each other adding to the lifespan. godel's theorm was based primarily on serial computer logic. Whereby you could only run one line of code at a time.

Parallel processing -- Another thing that people are wont to believe makes a computer trans-Turing. It doesn't. Goedel's theorem applies just as much to parallel computers as it does to serial ones.

Incedentally, Goedel's theorem was not formulated based on any form of "computer logic" at all. They are voiced in a very number-theorhetic way. It's the equivalent Turing machine and lambda calculus based formulations of equivalient theorems that are readily related to "computer" logic.

[baron greenback]

I read this quote to my computer scientist coworker. He said "smack that guy for me". Nothing is trans-Turing. Your two examples are no exception (ironically, both are simulated on a standard Von Neumann machine in nearly all applications).

*considers self smacked*

And then smacked around a bit some more. Although, thankfully, with an oppurtunity to redeem myself,

On second thought, I think you might be able to google that up. Why don't you try that before I go digging through the archives for something eight or ten years old that I might not even be able to find?

All I can really say for myself is that I was responding to the language of the original post, and lazily at that. In point of fact, I was responding to the implication that information systems weren't capable of generating new concepts, rather than trying to contradict Godel (even I don't have that kind of audacity). Which was not only unclear from my post, but arguing with an implication is, well, strawmanning. SO. In order to avoid arguing with an implication, I'm going to quote someone's position on the subject as the position I am arguing against. That quote is

Gödel's Theorem has been used to argue that a computer can never be as smart as a human being because the extent of its knowledge is limited by a fixed set of axioms, whereas people can discover unexpected truths

Basically, I was trying to point out that information systems need not be completely governed by "a fixed set of axioms", and can actually be designed to incorporate new and/or random data and weigh it against existing data, similar to "discovering unexpected truths" (which I am understanding to mean incorporating and utilizing new information that did not exist originally to use as axioms in deciding propositions).

That doesn't really excuse my c# comment, which was a flat out inaccuracy. My first computer, if it matters, was a 1979 zenith with no hard disk. One of the floppies had wordperfect on it, and the other had basic (just basic). I had a few books with games written line-by-line in it that kept me and entertained 9 year old despite the fact that all my friends had those new black-and-white macs with pictures and mice and stuff.

Nonetheless, I've got enough intellectual honesty to swallow my pride when I'm wrong.

First off, to prevent others from following in my footsteps.

Again, from mathworld.wolfram.com:

Godel's Theorum:Informally, Gödel's incompleteness theorem states that all consistent axiomatic formulations of number theory include undecidable propositions (Hofstadter 1989). This is sometimes called Gödel's first incompleteness theorem, and answers in the negative Hilbert's problem asking whether mathematics is "complete" (in the sense that every statement in the language of number theory can be either proved or disproved). Formally, Gödel's theorem states, "To every -consistent recursive class of formulas, there correspond recursive class-signs r such that neither (v Gen r) nor Neg(v Gen r) belongs to Flg(), where v is the free variable of r" (Gödel 1931).

A statement sometimes known as Gödel's second incompleteness theorem states that if number theory is consistent, then a proof of this fact does not exist using the methods of first-order predicate calculus. Stated more colloquially, any formal system that is interesting enough to formulate its own consistency can prove its own consistency iff it is inconsistent.


Turing Machine: A Turing machine is a theoretical computing machine invented by Alan Turing (1937) to serve as an idealized model for mathematical calculation. A Turing machine consists of a line of cells known as a "tape" that can be moved back and forth, an active element known as the "head" that possesses a property known as "state" and that can change the property known as "color" of the active cell underneath it, and a set of instructions for how the head should modify the active cell and move the tape (Wolfram 2002, pp. 78-81). At each step, the machine may modify the color of the active cell, change the state of the head, and then move the tape one unit to the left or right.

Halting Problem: The determination of whether a Turing machine will come to a halt given a particular input program. The halting problem is solvable for machines with less than four states. However, the four-state case is open, and the five-state case is almost certainly unsolvable due to the fact that it includes machines iterating Collatz-like congruential functions, and such specific problems are currently open. The problem of whether a general Turing machine halts is undecidable, as first proved by Turing (Wolfram 2002, pp. 1137-1138).

Church-Turing thesis: The Church-Turing thesis (formerly commonly known simply as Church's thesis) says that any real-world computation can be translated into an equivalent computation involving a Turing machine. In Church's original 1935 formulation, the thesis says that real-world calculation can be done using the lambda calculus, which is equivalent to using general recursive functions.

As far as the googling goes, I remember the location of a python script that demonstrates the problem. I'll find it, but I wanted to get this up before other people had oppurtunity to get their digs in.

In the meantime, the website earlier had a good "in english" explanation regarding Godel's theorum.

The proof of Gödel's Incompleteness Theorem is so simple, and so sneaky, that it is almost embarassing to relate. His basic procedure is as follows:

Someone introduces Gödel to a UTM, a machine that is supposed to be a Universal Truth Machine, capable of correctly answering any question at all.
Gödel asks for the program and the circuit design of the UTM. The program may be complicated, but it can only be finitely long. Call the program P(UTM) for Program of the Universal Truth Machine.
Smiling a little, Gödel writes out the following sentence: "The machine constructed on the basis of the program P(UTM) will never say that this sentence is true." Call this sentence G for Gödel. Note that G is equivalent to: "UTM will never say G is true."
Now Gödel laughs his high laugh and asks UTM whether G is true or not.
If UTM says G is true, then "UTM will never say G is true" is false. If "UTM will never say G is true" is false, then G is false (since G = "UTM will never say G is true"). So if UTM says G is true, then G is in fact false, and UTM has made a false statement. So UTM will never say that G is true, since UTM makes only true statements.
We have established that UTM will never say G is true. So "UTM will never say G is true" is in fact a true statement. So G is true (since G = "UTM will never say G is true").
"I know a truth that UTM can never utter," Gödel says. "I know that G is true. UTM is not truly universal."

[Hyndis]

In the field of AI, there's the top down vs the bottom up approach.

The top down approach is doomed to failure in regards to AI, IMHO, and this is how the vast majority of computers function. The computer essentially boils down to having a large number of if->then type statements, and the computer can only operate within the bounds of its instructions. If a situation arises that its instructions do not cover, it crashes.

If I build an archer in AoE, the computer is programmed to build a skirmisher. This works well when the world is restricted to a particular scenario, such as in chess, but it simply cannot function in the real world. In order to do so, it would have to have an if->then solution for every possible situation in the universe. And that would take up quite a hefty amount of storage space and processing power, no?

The bottom up approach is to simply have the machine learn. Program in a basuc and simple learning program and some other basic things (instincts?), and then let the machine build up its own consciousness. This is what humans currently use, and it seems to work most of the time. :)

Of course, writing up such a powerful and flexible program and allowing the machine to learn is far easier said that done, and I bet that digital computers are simply the wrong type of computer to use fo this. Our brains are analogue, not digital. Other experiments in analogue computers and robots has come up with some pretty amazing results, such as the guy who can build a robot that walks forward, and without any programming, this robot "learns" to alter its gait depending on the condition of its legs. If one is damaged or removed, it changes its pattern of movement so that it can continue to move forward.

There was something on The Learning Channel about this, before that channel turned into the home remodeling channel. :down:

From what I gather, all one has to do is build an analogue machine, and somehow those circuits will just begin to do things on their own. Perhaps if you just stick enough neurons together, they just start to work on their own, and if the number of them gets big enough, things start to happen.

If I put a single nerve cell in a petri disk and try to have a conversation, it'll be a very one-sided conversation. If I get enough nerve cells together and talk to them, the group of nerve cells might ask me to order it some pizza.

[baron greenback]

Haven't found much that isn't a language itself, although it wouldn't be hard to write in a scripted language like python; then again, writing out a script to demonstrate something more easily explained in english is counterintuitive.

Still kicking myself about the c# comment. I know better than that. Nothing I can do about it now, though.

So, what I did is something I figured would be more helpful to the discussion. Here's (http://praelego.com/projects/genetic_demo.py) a python script you can play with at home to see how genetic algorithms work. And here (http://freebsd.mu/freebsd/archives/000039.html) is a great explanation on neural networks.

What I was discussing in my previous post was simply methods of gathering information that did not exist in the original problem in a form that can be utilized. Where I got myself into trouble was when I didn't specifically say how the information was being utilized. Which I really should have done, as the incompleteness theorum says "Informally, Gödel's incompleteness theorem states that all consistent axiomatic formulations of number theory include undecidable propositions". In other words, I was discussing the "consistent axiomatic formulations" portion, in a manner that seemed to imply I was discussing the propositions.

The reason I'm saying this is because employing things like fuzzy logic only gives you a wider set of data to use. Genetic algorithms and neural networks give us a method of utilizing that data, and it is necessary to discuss this as well, because I was in a discussion regarding Godel's incompleteness theorum, and part of the theorum is in the application.

Basically, I wasn't trying to say computers somehow "break" godel's problem. I was merely saying that, when we work with new information that did not exist in the original program, godel's theorum does not apply, because we are no longer working with consistent axiomatic formulations. To be blunt: if we allow new information to change a decision a program makes (even if we are weighing), the decision process of the program has been changed, and is no longer consistent.

[Schneibster]

*considers self smacked*

And then smacked around a bit some more. Although, thankfully, with an oppurtunity to redeem myself,

All I can really say for myself is that I was responding to the language of the original post, and lazily at that. In point of fact, I was responding to the implication that information systems weren't capable of generating new concepts, rather than trying to contradict Godel (even I don't have that kind of audacity). Which was not only unclear from my post, but arguing with an implication is, well, strawmanning. LOL OK, you're forgiven.

On edit: and a nice post, too.

[Demosthenes]

Of course, writing up such a powerful and flexible program and allowing the machine to learn is far easier said that done, and I bet that digital computers are simply the wrong type of computer to use fo this. Our brains are analogue, not digital. Other experiments in analogue computers and robots has come up with some pretty amazing results, such as the guy who can build a robot that walks forward, and without any programming, this robot "learns" to alter its gait depending on the condition of its legs. If one is damaged or removed, it changes its pattern of movement so that it can continue to move forward.

There was something on The Learning Channel about this, before that channel turned into the home remodeling channel. :down:

From what I gather, all one has to do is build an analogue machine, and somehow those circuits will just begin to do things on their own. Perhaps if you just stick enough neurons together, they just start to work on their own, and if the number of them gets big enough, things start to happen.

If I put a single nerve cell in a petri disk and try to have a conversation, it'll be a very one-sided conversation. If I get enough nerve cells together and talk to them, the group of nerve cells might ask me to order it some pizza.

even doing it the analog way is far far from simple. It's required a lot of research and work to accomplish basics. At heart, it doesn't matter whether the computing medium is analog or digital, either one is perfectly capable of simulating either one. In the long run, digital is preferred because it's easily manipulated and easier to implement. All it takes is appropriate software which performs the job of a hardwired analog computer.

In an analog computer, the circuits still have to be programmed with a set of behaviors rather like ceullar automata to start form something. It's hardly a matter of circuits just starting to do something on their own. Neither a petri dish of neural cells. You'll have to build up a lot of structures and organization s before you would start to see sophisticated behavior.

We're getting there with a rapid progress but we're still a long way from creating Sonny from I, Robot

[Majestyk]

There's a new paper here http://eccc.uni-trier.de/eccc-reports/2005/TR05-004/ that is relevant to this discussion.

[premjan]

I read this quote to my computer scientist coworker. He said "smack that guy for me". Nothing is trans-Turing. Your two examples are no exception (ironically, both are simulated on a standard Von Neumann machine in nearly all applications).
I think another point is that human thought at least has enough of an analog, perceptual component for it not to be a formal system at all (does not follow strict rules of logic). It is more of a function approximation system than a formal system. Hence I think this is another loophole why Godel does not apply.

[premjan]

I'm not a fan of strong AI running on anything resembling an electronic computer. However, I don't really buy Penrose's arguments. They seem to assume that brains are sound and know they are sound, and I think this is unjustified. In shadows of the mind, this soundness is qualified a kind of 'in principle' soundness making mathematics possible, but then that is no longer about the human brain, but mathematics.

I don't think fuzzy logic, parallellism, stochastic optimization or any other kind of calculations is going to solve the strong AI problem. If it is possible to do using those things, then it is possible to do in a binary, serial, deterministic way. This is obvious if you consider how computers work.

No, I am fairly convinced of Searle's position - that there is some physical process going on in the brain that we are going to have to understand before we get anywhere. It is not the calculations that might represent this process that matter, but the process itself. I am veering towards the problem of consciousness rather than intelligence here, but the two are probably related.

So Penrose's suggestion that intelligence is non-computable I think is unsupported, but he may still be right.

I think making the basic computer analog and having various formal digital operations sitting on top of it would probably work eventually (though lots of search would be needed to first evolve efficient strategies for all common survival tasks).

[Schneibster]

I think another point is that human thought at least has enough of an analog, perceptual component for it not to be a formal system at all (does not follow strict rules of logic). It is more of a function approximation system than a formal system. Hence I think this is another loophole why Godel does not apply.This is taken from a very high level view of the functioning of the mind/brain; however, if there are Godel-type problems, they are more likely at a much lower level.

The neurons have two sides: an input, called dendrites, where they accept signals from multiple other neurons or sensory cells, from the terminal buttons on their axons; and an output, their axon, which usually goes only to the dendrites of one other cell, but which can branch and go to more than one. The neuron uses complex electrochemical biases to control the "weight" of the input from any single axon connection to its dendrites, so for instance it can assign a high "weight" to one axon and a low "weight" to another. The neuron fires at a rate that is based on the combined "weight" of the rates of all the axons connected to its dendrites. It is thus capable of changing its response dynamically.

Two questions for the neuro-biologists or other professionals who may be hovering here:
1. Did I get that right so far?
2. What factors can influence the neuron's "weighting" of the various inputs?

Once I have answers to these questions, and perhaps some further explanation if I have made a bad model, I should be able to tell all of us what sort of logic functions this structure will represent. We might find out some pretty interesting stuff about the way the brain works. I'll be able to give a much better low-level interpretation of what you were talking about here.

[Tuvar Ane Ingolenen]

Drat!

I thought I voted for 'Did Godel disprove the idea of artificial intelligence?' and voted 'no' because Gödels Incompleteness Theorem has no relevance to 'thinking as humans do'.

To 'Will machines ever be able to think as humans do?' I would have voted 'yes'. I assume that a machine will pass the turing test ( not any time real soon ), and in that case 'yes' would be a better answer than 'no'.

[premjan]

I wonder if God is an example of a basic conceptual problem for neurons? I think maybe yes, because it is a logically simple proposition that is however practically very unparsimonious. It constitutes something that is quite easy to model, internally parsimonious for a human brain but externally very unparsimonious.

[Dark Knight Bob]

Parallel processing -- Another thing that people are wont to believe makes a computer trans-Turing. It doesn't. Goedel's theorem applies just as much to parallel computers as it does to serial ones.

It was merely an extension to the stated idea of lifespan. I know it doesn't solve the problem but it reduces it to a level where it is inconsequential for a useful finite lifespan

In the field of AI, there's the top down vs the bottom up approach.

The top down approach is doomed to failure in regards to AI, IMHO, and this is how the vast majority of computers function. The computer essentially boils down to having a large number of if->then type statements, and the computer can only operate within the bounds of its instructions. If a situation arises that its instructions do not cover, it crashes.

The top down approach is done where things need to get done. If we tried to do a bottom up approach to everything we wouldn't get anywhere.

By doing top down approaches you are able to find models that should become emergent as the bottom up model develops.

Also it has to do with resources. To make a chess program it is simpler to figure out the process that the human brain is using make a top down model that emulates that thought pattern.

The definition of AI is not restricted to just trying to reproduce a fully working human. It is applied to everything from super computers down to toasters.

Different tasks require different routes of inquiry.

[Jehanne]

Has artificial intelligence been discredited by Godel's incompleteness theorem?

For laypeople such as myself, here’s a wonderful site on mathematics, in general, and Godel's incompleteness theorem, in particular:

http://mathworld.wolfram.com/GoedelsIncompletenessTheorem.html

I voted "no", having read Nobel laureate Gerald Edelman's book, "Bright Air, Brilliant Fire".

[David Z]

This is taken from a very high level view of the functioning of the mind/brain; however, if there are Godel-type problems, they are more likely at a much lower level.

The neurons have two sides: an input, called dendrites, where they accept signals from multiple other neurons or sensory cells, from the terminal buttons on their axons; and an output, their axon, which usually goes only to the dendrites of one other cell, but which can branch and go to more than one. The neuron uses complex electrochemical biases to control the "weight" of the input from any single axon connection to its dendrites, so for instance it can assign a high "weight" to one axon and a low "weight" to another. The neuron fires at a rate that is based on the combined "weight" of the rates of all the axons connected to its dendrites. It is thus capable of changing its response dynamically.

Two questions for the neuro-biologists or other professionals who may be hovering here:
1. Did I get that right so far?
2. What factors can influence the neuron's "weighting" of the various inputs?

Once I have answers to these questions, and perhaps some further explanation if I have made a bad model, I should be able to tell all of us what sort of logic functions this structure will represent. We might find out some pretty interesting stuff about the way the brain works. I'll be able to give a much better low-level interpretation of what you were talking about here.I am not a neurobiologist, I am an engineer who works in the machine learning field, so my expertise is skewed toward artificial neural networks (ANNs). My understanding of biological neural networks is pretty sketchy, so I will respond as best as I can.

First off, your description of how biological neurons work is consistent with my own understanding, but I would add that there can also be inhibitory weights (corresponding to negative numerical weight values in an ANN) that make the neuron less likely to fire.

The neurons in ANNs operate in a similar way: Each input is multiplied by the weight value associated with that input, then the weighted inputs are summed, and a threshold function is applied. The threshold function most commonly cited in the literature is 1/(1+e^(-x)). The result of the threshold function is the output of the artificial neuron. This is perhaps the biggest difference between artificial and biological neurons: a biological neuron fires at a faster or slower rate due to its input stimulus, whereas an artificial neuron generally outputs a numerical value.

As far as what decides the weight of the inputs to a biological neuron, I have no idea. In ANNs there is a procedure known as backpropagation that works through the network backwards, providing you with a delta value for each weight that will nudge the net in the direction of less overall output error. It doesn't seem as if there is any such mechanism in animal brains, and even if there were, there is no authority feeding the network with the right answer. So the way this works is still an open question as far as I know.

As far as what functions an ANN can approximate, it has been shown that a 3-level feed-forward neural network can approximate any function with arbitrarily high accuracy, but the number of units in each layer may be very large.

Re: the OP... I am very much in agreement with the other posters who reject Penrose's conclusions in "The Emperor's New Mind". He seemed to be desperately trying to justify his own prejudices regarding the nature of the human mind, that there is something special about human cognition that would be cheapened by a machine being able to think. I disagree. Human beings are not lessened by creating smart machines any more than parents are lessened by having bright children.

[James T]

To respond to the question, yes they will. To offer a couple of other ideas to the pot.

Genetic algorithms are an interesting new approach, I've looked at these for control systems applications but at the present point in time they fail compared to using linear programming techniques. My impression here is that the understanding both of population management and fitness functions has yet a long way to go. However, it provides yet another tool in the arsenal for creating an artificial intelligence.

Another is the idea of the mind as a chaotic process. Even incredibly trivial concepts, Xn+1=AXn(1-Xn) iterated for 3.8ish < A <=4, generate deterministic (completely defined causal relationships) systems that are sufficiently complex to be unpredictable. I'll note that generating truly random numbers in a computer is hard but that there are many effective techniques which support your online banking experience and like security tools.

My impression is that AI research learnt an unpleasant lesson predicting great things before the supporting tools existed (or had even been researched properly). This time it's going to sneak up on as, almost as a fait accompli. There are few areas of thought where the software tools are not massively superior to even a clever person. Albiet I'd accept some of the key ones are missing.

[travc]

Just gotta say...

Will machines be *able* to think like humans... yeah
Though machines will probably never think like humans because machines thinking like machines makes so much more sense (and is so much more practical and powerful).

Of course, this is all assuming that humans don't destroy ourselves (as a civilization) or go major luddite... oh, and assuming Jesus doesn't come back anytime in the next 50 to 100 years.

[travc]

Good to see that people are talking about learning systems (like GAs, ANN, even linear programming). The key isn't figuring out a system that can theoretically learn the patterns that we are interested, it is in constructing learning systems that are constrained in such a way that they learn salient patterns efficiently.

Why I said that machines will think like machines, though they will (at least in theory) be capable of thinking like humans, is that the particulars of the way the various learing systems crammed into the human brain work are not necissarially the "right" or even a "good" way to do it. They work (most of the time), but we all know here how susceptible people are to learning false patterns and how plyable human memory is. I don't think we will ever endeavor to build machines that have all the same mental flaws as humans, when we can do so much better. Ok, maybe someone in Japan will try to build a Replicant, but really, why?

Anyway, for those who are interested, you might find Ed Stabler's work on Minimalist Grammars and Tupled Pregroup Grammars. He is in liguistics at UCLA, but don't let that turn you off, it is really about constructing a learning system that will learn logical relationships. Basically, the idea is to have a learing system that constructs a language out of observed data, and the language be a particular type that can be translated (simply) to logic. Once the learner has aquire the language, it can make logical inferences about the reality that produced the data it used to learn the language in the first place. A bit OT, but a really cool AI approach that I wish I were more closely invovled with... damn Biology department wouldn't accept this sort of stuff a dissertation topic, so it is relegated to a side project for me (which I have basically no time to persue). BTW: ANNs and all those sorts of learning systems are still critical though, since the inputs to this sort of language based learner have to be atomic, so a system to do basic classification type pattern recognition is still needed.

[travc]

I wonder if God is an example of a basic conceptual problem for neurons? I think maybe yes, because it is a logically simple proposition that is however practically very unparsimonious. It constitutes something that is quite easy to model, internally parsimonious for a human brain but externally very unparsimonious.

I don't know about "neurons", but this is one of the best statements for why God is not allowed as an explanation in science but still a very compelling explanation for many people. Science (as a process) is a particular type of learning system, which isn't the same way as individual humans learn.

[Jet Black]

I fail to see why artificial intelligence could not be produced. we are the result of the iterative process of evolution, and we could make an artificial intelligence in the same way.

[Bold]

I fail to see why artificial intelligence could not be produced. we are the result of the iterative process of evolution, and we could make an artificial intelligence in the same way.

Quite. Perhaps it comes down to the definition of machine. All machines, by a common sense meaning of the word, are a product of human design (iterative perhaps), not evolution. Moreover they are, currently, characteristically stupid, and exist to serve our purposes, not their own. They can't have purposes of their own.

Alternatively, if you take the view that everything is a machine - we are machines - then machine is a fairly meaningless word, and machines are intelligent already.

I think this difference comes down to the question: is nature mechanical? That is, is it at some fundamental level like the machines we have designed? If it is, then 'everything is a machine' is justified. If OTOH, some aspects of nature - the measurement problem, say - are unmachinelike and particularly if this may be important to thinking beings, then the question is still open.

Of course if this unmachinelikeness is important and we come to understand it, then we will be able to use that understanding in designing new life forms - do we get to call them machines because they are designed? Perhaps. But they may by todays standards be life forms and not machines. If the definition of machine is a moveable feast....

BTW I have made two intelligent machines. Both are at school right now.

[premjan]

I would guess that biological evolution works within tighter resource limits (energy budget) so perhaps intelligence makes a big difference. Right now, we don't seem to impose that constraint on computers so much: how to minimize the number of operations and the energy budget. Rather we concentrate on constantly producing new features so users have more to choose from. Perhaps we need to reach the point where algorithmic efficiency makes a very big difference. It could be an economcs problem.

[mirage]

Why in the name of Christ is this poll tipped to "yes"?

Do you need a consistent formal system for intelligence?

No.

Do you need a complete formal system for intelligence?

No.

Godel's theorem is utterly irrelevant to AI and anyone who thinks it is, is a moron. So there.

[Tuvar Ane Ingolenen]

Why in the name of Christ is this poll tipped to "yes"?

Do you need a consistent formal system for intelligence?

No.

Do you need a complete formal system for intelligence?

No.

Godel's theorem is utterly irrelevant to AI and anyone who thinks it is, is a moron. So there.
Did you notice that the thread title is 'Did Godel disprove the idea of artificial intelligence?', but the actual poll question is 'Will machines ever be able to think as humans do?' ( I didn't )

[Undercurrent]

Ok, maybe someone in Japan will try to build a Replicant, but really, why?

Oh, you know why. ;)

[mirage]

Did you notice that the thread title is 'Did Godel disprove the idea of artificial intelligence?', but the actual poll question is 'Will machines ever be able to think as humans do?' ( I didn't )
:rolling: Ooops!
Thanks for pointing that out. Serves me right for jumping to conclusions again. However I don't think that results will fairly reflect opinion given the mismatch.

And there still isn't a problem for AI with inconsistent or incomplete formal languages.

[Jay GW]

I think this difference comes down to the question: is nature mechanical? That is, is it at some fundamental level like the machines we have designed? If it is, then 'everything is a machine' is justified. If OTOH, some aspects of nature - the measurement problem, say - are unmachinelike and particularly if this may be important to thinking beings, then the question is still open.

Cells of organic things and metals and plastics are fundamentally different. Can someone explain how the two can simulate each other?

The cells of the brain are not like circuits exactly. Some way has to be found for plastics and metal to simulate cell behavior.

I wonder if the possibility of "growing" a brain is really more practical than trying to get metals to act like one?

Can't some parts of the body already be fabricated?

[premjan]

Godel's theorem is utterly irrelevant to AI and anyone who thinks it is, is a moron. So there.
This last bit is totally not defensible. All of AI is not necessarily achievable just using neur(on)al nets. Formal systems have different strengths and weaknesses compared to neural nets. For example, a neural net is much more likely to fall into the "God" trap IMO since it is internally parsimonious. Whereas a formal reasoning system might able to recognize that it is externally not parsimonious.

[David Z]

I wonder if the possibility of "growing" a brain is really more practical than trying to get metals to act like one?

Can't some parts of the body already be fabricated?Absolutely. The current state-of-the-art for creating autonomous agents uses biological 'machines'. The agent is created with a robust general-purpose learning algorithm and a powerful organic computing engine, but is otherwise a blank slate. This strategy has numerous advantages (such as not requiring that the operational algorithm be specified formally), as well as numerous disadvantages (such as a tendency to formulate and accept false inferences). But perhaps the most dangerous aspect of these autonomous agents is their ability to self-replicate, which threatens the world with a macroscopic version of the 'gray goo' nanotech problem. The technology for creating these agents is available to almost anyone... it is simple enough to be utilized by 3rd world dictators or even terrorists! Many of the members of II have dabbled in the technology (such as Bold, above), and even those without the inclination to create an autonomous agent of their own occasionally practice the required techniques as a form of recreation.

[Atheos]

I'm afraid that when the time comes, machines will only be limited to the ability to "think like humans" for maybe 10 minutes. They'll surpass us quickly after that.

-Atheos

[travc]

Cells of organic things and metals and plastics are fundamentally different. Can someone explain how the two can simulate each other?

What it is "made" of is irrelevent (at least taking the current conventional wisdom and the implications of the OP). What we are talking about is a system which behaves a certain way because of the way the state of parts of it influences the states of other parts. They could be neurons, sigmoid functions in a computer program, or logic gates on a chip. Hell, they could even be very carefully arranged billiard balls on a big-ass table.

Thus the wonder of computer science, information theory, and mathematical linquistics. It isn't about the actual matter that comprises the system, it is about the way the system is arranged.

[Jay GW]

It isn't about the actual matter that comprises the system, it is about the way the system is arranged.

Says who?

[James T]

Says who?I think this is another one of those category error problems.

I'll take a really basic example, internet protocols, since we are using them this very instant. This is layered: -

Application
Presentation
Session
Transport
Network
Data Link
Physical

The physical is copper or fibre and voltages, what's a 1 what's a 0 etc. Data link is CSMA/CD (carrier sense multiple access/ collision detect). TCP/IP covers the Transport and Network stuff. What you and I and the rest of us are doing at IIDB is using the A, P and S layers. The underlying stuff, whatever it is, is just a supporting layer and is conceptually separate from the activities above.

In much the same way the fact that my brain is organic and has neurons is kind of irrelevant. If you consider Descartes demon, how do you know you are not a program running on a computer. You don't!

[mirage]

Originally Posted by mirage
Godel's theorem is utterly irrelevant to AI and anyone who thinks it is, is a moron. So there.
This last bit is totally not defensible. All of AI is not necessarily achievable just using neur(on)al nets. Formal systems have different strengths and weaknesses compared to neural nets. For example, a neural net is much more likely to fall into the "God" trap IMO since it is internally parsimonious. Whereas a formal reasoning system might able to recognize that it is externally not parsimonious.I'll retract the "so there" just for you. ;)

I have no idea what you are getting at here. Formal system refers to an axiomatised formal language. This is in no way a requirement for a thinking machine. Even if it were, it would not need to be either complete or consistent.

[premjan]

I'll retract the "so there" just for you. ;)

I have no idea what you are getting at here. Formal system refers to an axiomatised formal language. This is in no way a requirement for a thinking machine. Even if it were, it would not need to be either complete or consistent.
well, we don't know how thought works so it might be useful. I lean towards the "multipronged" approach of AI myself, where the result of one method is checked by another method (if available). Using a formal axiomatized approach, along with a neural net approach might produce better results than either individually.

When you have a machine that thinks as well as a man, or a proof that formal systems (e.g. expert rule-based systems) cannot be useful for intelligence then we can talk further.

[James T]

This last bit is totally not defensible. All of AI is not necessarily achievable just using neur(on)al nets. Formal systems have different strengths and weaknesses compared to neural nets. For example, a neural net is much more likely to fall into the "God" trap IMO since it is internally parsimonious. Whereas a formal reasoning system might able to recognize that it is externally not parsimonious.I had to look up parsimonious in the dictionary (a paper one) and the the meaning is extreme care in spending. Similarly I googled neural-net and parsimonious, parsimonious, in the sense that the data is explained with as few hypotheses as possible.

What these people are talking about are the potential issues of trying to model the neurons in the brain based on the fact that if your software bloats all the computing power in the world won't do it. I've seen this used as a reason in the past why AI will never be possible.

The premises seem to be: -

P1: To produce an AI we will need to simulate all the neurons in the brain.
P2: The number of neurons and computational power required to carry out this simulation exceeds (some large amount) of computing power.
C: We cannot create AI.

There are many cases where mathematical problems are now solved in seconds where, if we used historical algorithms, we would not solve in years.

My own premise contradicts P1, I consider that we will be able to create AI with vastly less elements than required to simulate all the neurons in the brain. My reason for believing this is the number of computing tasks formally considered the sole domain of human cognitive powers that are being done by computers today.

And let me make a prediction, neural nets will not be necessary to achieve AI. My reason for this statement is my experience in use of feedback control loops. We've had fuzzy logic, neural nets, expert systems and latterly genitic algorithms touted as the new beaut solution. To be fair there are products based on fuzzy logic, too many heuristics still in these systems to be just called fuzzy, but PID loops are still the rule.

Still, I cannot see how parsimony bears any relevance to the issue.

[mirage]

When you have a machine that thinks as well as a man, or a proof that formal systems (e.g. expert rule-based systems) cannot be useful for intelligence then we can talk further.No matter how "formal" your AI system, it does not need to be either complete or consistent. We don't need this machine to know every truth following from its axioms! We need it to hold a conversation.

What is the proof that Gödel doesn't interfere with intelligence? Our intelligence.

[premjan]

No matter how "formal" your AI system, it does not need to be either complete or consistent. We don't need this machine to know every truth following from its axioms! We need it to hold a conversation.

What is the proof that Gödel doesn't interfere with intelligence? Our intelligence.
Suppose a formal system (not complete or consistent) that you are using to decide the truth of some proposition, goes into hang mode because it is unable to decide the truth value of a Godel proposition, and is more over unable to discover that this is an undecidable proposition? I think a neural net implementation which does not have the problem would be handy in that case. A neural net might easily decide that "God" was the simplest explanation to everything, and perhaps a formal system with a formal rule built in forbidding God could help fix this...

[James T]

Suppose a formal system (not complete or consistent) that you are using to decide the truth of some proposition, goes into hang mode because it is unable to decide the truth value of a Godel proposition, and is more over unable to discover that this is an undecidable proposition? I think a neural net implementation which does not have the problem would be handy in that case. A neural net might easily decide that "God" was the simplest explanation to everything, and perhaps a formal system with a formal rule built in forbidding God could help fix this...You're not a programmer then. Infinite loops and other hangs are considered a joke in programming circles. It's like a foot trip, simple, stupid and funny for everyone watching. This isn't something to get hung up on, even (horror of horrors) a watchdog timer would do the trick.

Anyway, it'll never be just a neural net. Never! Nor will it be just a formal system. Personally I'd code a reward for novelty (a relatively common practice in AI exercises) and a bit of vagueness just for good measure. AI won't be able to work if we tie it down in formal wrangles ... so we won't.

[mirage]

Suppose a formal system (not complete or consistent) that you are using to decide the truth of some proposition, goes into hang mode because it is unable to decide the truth value of a Godel proposition, and is more over unable to discover that this is an undecidable proposition?There is no reason for a system to need to hang just because there are undecidable formal statements about. I'm still not sure exactly what you are thinking of anyway, since a computing system isn't an axiomatised formal language in any way that I could conceive. They are really not all that related.
I think a neural net implementation which does not have the problem would be handy in that case. A neural net might easily decide that "God" was the simplest explanation to everything, and perhaps a formal system with a formal rule built in forbidding God could help fix this...A neural net would not be necessary to prevent crashing. Just some elementary good programming. I'm not sure why you think neural nets are some kind of Occam's razor machine (although, in a way they are if you check out my Occam thread in philosophy). Whatever formal logical conscious thinking we do seems to be very high up the orders of complexity of our brain and may not be particularly related to the underlying architecture. Who knows. However, God is the least parsimonious explanation in the entire universe of explanations (well almost). I think only a system with a built in evolutionary tendency to see personal agency behind every event would jump to this unjustified conclusion. It certainly wouldn't be a good way for a simple "learning" neural net to adapt as it has nil predictive value.

[premjan]

You're not a programmer then. Infinite loops and other hangs are considered a joke in programming circles. It's like a foot trip, simple, stupid and funny for everyone watching. This isn't something to get hung up on, even (horror of horrors) a watchdog timer would do the trick.

Anyway, it'll never be just a neural net. Never! Nor will it be just a formal system. Personally I'd code a reward for novelty (a relatively common practice in AI exercises) and a bit of vagueness just for good measure. AI won't be able to work if we tie it down in formal wrangles ... so we won't.

Of course you could force it not to hang in a program (just terminate after 10 tries or something). But in principle if you want to always search for an explanation it could cause a hang. I guess Godel is not a necessary limitation but it could happen in a program where termination were not explicitly forced.

As for parsimony etc, I was thinking of a case where the neural net had to explain rather than predict. Which might not occur often in practise but it would happen if you turned a neural net to philosophy (which is not a predictive activity). Anyway when designing a system these things could easily happen though they would be debuggable and avoidable.

[James T]

Of course you could force it not to hang in a program (just terminate after 10 tries or something). But in principle if you want to always search for an explanation it could cause a hang. I guess Godel is not a necessary limitation but it could happen in a program where termination were not explicitly forced.My expectation is that a form including: -

- inputs, vision, libraries, audio, web pages etc, each with it's own toolbox for speech recognition, image recognition, 3D spatial rendering, motion mass and energy estimations, preliminary object classification.

- knowledge, in the form used in neural nets, expert systems, predicate logic that we know, complete libraries (Dewey classification rules of course), morals (we'd code some in), premises (with confidence levels, fuzzy and absolute), stuff we already know how to do in a computer expressed as logic including chess, calculus, engineering, driving, flying, war (ever played against computer war games at their HARD level :(, I have and I lose ... badly ).

- tools, stuff to process neural net data, stuff to convert sentences to internal knowledge and back again, tools for processing logic, tools for creating stupid ideas (remembering a computer can do this fast), weightings based on novelty, utility, human benefit (remember Asimov's laws). Methods for processing logic and choosing base laws parsimoniously (:) ). Methods for generating new laws, all with confidence levels of course.

- outputs, the easy bit, internal items -> sentences to speech synthesizers and out. And other mundane stuff of course like writing to disk, printer, email.

Now when this takes a concept, like Godels kind of sneaky theorem, it will line up the sentence statement against the actions that it is required to take by the sentence and evaluate that the sentence does not permit it to answer in a way that is consistent with the sentence. It will then hunt for a resolution, internally and say by asking Godel the same question back. Godel, having posed the statement in the first place is a logical person to ask to elaborate the implications of the statement. I wonder how Godel would answer.

As for parsimony etc, I was thinking of a case where the neural net had to explain rather than predict. Which might not occur often in practise but it would happen if you turned a neural net to philosophy (which is not a predictive activity). Anyway when designing a system these things could easily happen though they would be debuggable and avoidable.We'd have the AI explain what of it's knowledge it used to reach a conclusion, this stuff was done commonly when I was at varsity in the 80's so it is not new. If it used a neural net it would include this in it's description of method. If the neural net gave the wrong answer the AI would confirm this itself as well as flagging a new item of knowledge that suggested the neural net was bad. It would have confidence levels against everything. It wouldn't trust but over time if, say, Fred was always right then items Fred said would have a higher confidence rating initially.

So the hanging problem, never really enters the equation.

[premjan]

Now when this takes a concept, like Godels kind of sneaky theorem, it will line up the sentence statement against the actions that it is required to take by the sentence and evaluate that the sentence does not permit it to answer in a way that is consistent with the sentence. It will then hunt for a resolution, internally and say by asking Godel the same question back. Godel, having posed the statement in the first place is a logical person to ask to elaborate the implications of the statement. I wonder how Godel would answer.

We'd have the AI explain what of it's knowledge it used to reach a conclusion, this stuff was done commonly when I was at varsity in the 80's so it is not new. If it used a neural net it would include this in it's description of method. If the neural net gave the wrong answer the AI would confirm this itself as well as flagging a new item of knowledge that suggested the neural net was bad. It would have confidence levels against everything. It wouldn't trust but over time if, say, Fred was always right then items Fred said would have a higher confidence rating initially.


I think the natural number version of Godel statements tend to be statements about the solutions of Diophantine Equations (Fermat's last theorem is an example). Anyway, undecidability (or difficulty-to-decide) is likely to pop up all over the place even without Godel. So some means to decide how much resources to spend on coming up with a particular answer would always be needed.

[Ninasgrandpa]

I've been looking at the poll results on this thread, and have been wondering how it is possible that a significant majority of the people on this board could possibly believe that Godel disproved the possibility of AI.
Now I see that the question actually proposed in the poll is quite different from the title of the thread. In fact the answer to the poll question is almost certain to be the opposite of the one in the title. (You could, I suppose, answer No to both questions, but probably not Yes).
Anyway, due to this misunderstanding, I replied No to this poll, when I should have replied Yes.

Ninas grandpa

[Tuvar Ane Ingolenen]

To all those people that think neural nets are somehow special: Neural nets are just ordinary software running on ordinary computers.

[premjan]

To all those people that think neural nets are somehow special: Neural nets are just ordinary software running on ordinary computers.
neural nets are not like ordinary software since that is usually procedural and written in one of our familiar computer languages.

[Tuvar Ane Ingolenen]

neural nets are not like ordinary software since that is usually procedural and written in one of our familiar computer languages.
Yes, but since a neural net is something that 'ordinary' software does, there's nothing that a neural net can do that 'ordinary' software cannot. Underneath the neural net its still the procedural software that does the work.

[Marduk]

Will machines ever think as humans do? No, of course not. Machines will never think as ponies do either.
Only humans think as humans do by definition. At best someone may make a machine that mimics human conversation to a point, but it will not “think� like a human, it will think like a machine programmed by its designer, it will not ponder what its going to make for dinner, fantasize about the hot babe at the terminal, stop running your programs because it thinks you are not paying enough attention to it theses days, or wonder if these jeans make it look fat. Human ‘thinking’ is driven by hormones as well as many other factors besides neural nets. As for Godel, he is way out of my league so I will defer to him

[James T]

Will machines ever think as humans do? No, of course not. Machines will never think as ponies do either.
Only humans think as humans do by definition. At best someone may make a machine that mimics human conversation to a point, but it will not “think� like a human, it will think like a machine programmed by its designer, it will not ponder what its going to make for dinner, fantasize about the hot babe at the terminal, stop running your programs because it thinks you are not paying enough attention to it theses days, or wonder if these jeans make it look fat. Human ‘thinking’ is driven by hormones as well as many other factors besides neural nets. As for Godel, he is way out of my league so I will defer to himAn AI will, however, wonder how we survive while being so stupid :).

[James T]

I think the natural number version of Godel statements tend to be statements about the solutions of Diophantine Equations (Fermat's last theorem is an example). Anyway, undecidability (or difficulty-to-decide) is likely to pop up all over the place even without Godel. So some means to decide how much resources to spend on coming up with a particular answer would always be needed.Interesting link on Godel's theorem http://www.sm.luth.se/~torkel/eget/godel.html where it says Unsurprisingly, the bulk of these invocations covers a range from the nonsensical to the merely technically inaccurate, and they often give rise to a flurry of corrections and more or less extended technical or philosophical disputes how accurate :).

Apologies for this one, I haven't included the lot but it shows enough and the full text is through the above link.
Using the []-notation we can now define what is meant by a Gödel sentence G for T. This is any sentence G which satisfies

(1) T |- G<=>~[]G

Thus, G is a sentence which "says of itself that it is not a theorem of T", and it is provable in T that G is true if and only if it is not a theorem of T. (A technical comment.) That there is such a sentence G is a special case of the diagonal lemma. For the proof of Gödel's theorem, it doesn't matter which such sentence G is used. Thus we often speak of simply "the Gödel sentence G of T".

A semantic version of the first incompleteness theorem
Now suppose all theorems of T are in fact true. Then, from (1), we conclude that the Gödel sentence G is undecidable in T, that is, neither G nor ~G is a theorem of T. For G is true if and only if it is not a theorem of T, and since all theorems of T are true, it follows that G cannot be a theorem of T. Since G is not a theorem, it is in fact true, again using the fact that G is true if and only if it is not a theorem of T. Thus ~G isn't a theorem of T either, since ~G is false.

Thus we have the following version of the first incompleteness theorem: Let T be a formal theory for which the Gödel numbering and diagonal lemma can be carried through, and all axioms - and hence theorems - of which are true. Then T is incomplete, i.e. there are sentences in the language of the theory- as exemplified by the Gödel sentence G for T - which can neither be proved, nor refuted in the theory.

Now for something more readable from http://www.miskatonic.org/godel.html.
The proof of Gödel's Incompleteness Theorem is so simple, and so sneaky, that it is almost embarassing to relate. His basic procedure is as follows:

1. Someone introduces Gödel to a UTM, a machine that is supposed to be a Universal Truth Machine, capable of correctly answering any question at all.
2. Gödel asks for the program and the circuit design of the UTM. The program may be complicated, but it can only be finitely long. Call the program P(UTM) for Program of the Universal Truth Machine.
3. Smiling a little, Gödel writes out the following sentence: "The machine constructed on the basis of the program P(UTM) will never say that this sentence is true." Call this sentence G for Gödel. Note that G is equivalent to: "UTM will never say G is true."
4. Now Gödel laughs his high laugh and asks UTM whether G is true or not.
5. If UTM says G is true, then "UTM will never say G is true" is false. If "UTM will never say G is true" is false, then G is false (since G = "UTM will never say G is true"). So if UTM says G is true, then G is in fact false, and UTM has made a false statement. So UTM will never say that G is true, since UTM makes only true statements.
6. We have established that UTM will never say G is true. So "UTM will never say G is true" is in fact a true statement. So G is true (since G = "UTM will never say G is true").
7. "I know a truth that UTM can never utter," Gödel says. "I know that G is true. UTM is not truly universal."

Think about it - it grows on you ... Now premjan, this was the one I was referring to in my earlier discussion. There are two parts of this argument I don't agree with, firstly the name UTM (universal truth machine), this is not an AI. An AI is never going to be a truth machine, it'll have some of our own foibles by necessity (the ability and necessity to make statements on incomplete information). This is not a contradiction of the quote merely a lack of desire to apply it to AI, which to me does not equate to UTM. If you were to so equate then ask the same question back to Godel, an obvious - albiet simplistic - argument against applying Godels theorem to AI.

[Marduk]

The way I understand it Godel did not claim AI was impossible, only that it was impossible with an ordinary recursive computer, he did not say it was impossible to some day build a non-recursive computer.
And why is every one in such a hurry to be replaced by a machine anyway? :confused:

[James T]

The way I understand it Godel did not claim AI was impossible, only that it was impossible with an ordinary recursive computer, he did not say it was impossible to some day build a non-recursive computer.You point being? Please reread, I specifically avoided saying this
This is not a contradiction of the quote merely a lack of desire to apply it to AI
And why is every one in such a hurry to be replaced by a machine anyway? :confused:Different issue altogether. Personally I'd prefer to translate my mind from ogranic to inorganic, with a few improvements at the same time of course :).

[Matt the Medic]

Will machines ever think as humans do? No, of course not. Machines will never think as ponies do either.
Only humans think as humans do by definition. At best someone may make a machine that mimics human conversation to a point, but it will not “think� like a human, it will think like a machine programmed by its designer, it will not ponder what its going to make for dinner, fantasize about the hot babe at the terminal, stop running your programs because it thinks you are not paying enough attention to it theses days, or wonder if these jeans make it look fat. Human ‘thinking’ is driven by hormones as well as many other factors besides neural nets. As for Godel, he is way out of my league so I will defer to him

Unless of course, many years from now, great advances are made both in neurophysiology and nano/biotechnology allowing us to create an exact 'living' duplicate of a human brain. If we know how everything works and have the ability to construct it, there is no reason to think it shouldn't work the same way. After all, no souls are required, just atoms.

edit: but I guess this surpasses the realm of 'artificial intellegence' as it would not be 'artificial' but rather 'synthetic' (in the 'geniune' definition of artificial).

[Jay GW]

Even though machines may have the same brains as a human, religious people will always say it's isn't human, because machines don't have souls.

[Huon]

I have always thought that "Artificial Intelligence" is artificial enough, but it is not human intelligence, as Marduk said somewher upwards. The word AI has helped scientists to gather money for their research...

The point is not Gödel or not Gödel. It is "invention". At some moment, a human decides to do "something different", because of his (her) hormons, or because he (she) thinks it could work better, or just for fun.

So far, I never saw a program deciding to rewrite itself, just for fun.

The first AI program I met was written in Basic, for a macIntosh, in 1983. It analysed the answers of a user. The answers had to be written in a very rigid form : an apple is a juicy fruit. Then the computer would ask : What is fruit ? What is juicy ? and so on. From time to time the program could ask a clever question. If an answer was : a grape is a fruit. Then the next question could be : Is a grape juicy ?

[Schneibster]

Even though machines may have the same brains as a human, religious people will always say it's isn't human, because machines don't have souls.And the scientific will always reply, "OK, show me your soul."

[James T]

And the scientific will always reply, "OK, show me your soul."Give it time and the AI will be explaining some of the finer points of faith and theism to all of us :).

[premjan]

I think an ideal intelligence is a truth machine, although an embodied intelligence will be merely a survival machine or an efficiency machine.

[Schneibster]

Give it time and the AI will be explaining some of the finer points of faith and theism to all of us :).LOL no question.
:eek:

[Peter Soderqvist]

Goldbach conjecture is probable truth!
The existence of uncomputable numbers can be inferred from Cantor's diagonal argument! P is consistent because G is not provable within P!
Do you know-how if a computer in principle can infer these conclusions too?

[James T]

Goldbach conjecture is probable truth!
The existence of uncomputable numbers can be inferred from Cantor's diagonal argument! P is consistent because G is not provable within P!
Do you know-how if a computer in principle can infer these conclusions too?I take it you are trying to find an example a computer AI program could never do. Given an AI does not currently exist we are the the position of having to guess. However, I don't see any problems with either of the ones you suggest in principle.

[Bold]

I think an ideal intelligence is a truth machine, although an embodied intelligence will be merely a survival machine or an efficiency machine.

Nature abhors a truth machine. I would say artificial intelligences ought to have the same philosophical problems with truth as real ones.

Perhaps I am agreeing with you, in which case in what sense is the truth machine an ideal?

[premjan]

Nature abhors a truth machine. I would say artificial intelligences ought to have the same philosophical problems with truth as real ones.
Perhaps I am agreeing with you, in which case in what sense is the truth machine an ideal?
Well, truth is the highest possible standard for intelligence. It represents knowledge that is context-free so applicable to anyone and anything.

[Peter Soderqvist]

James T, do you know in principle how to program a computer to infer these conclusions, or not? I myself don't know if it is possible or not, because I am not a computer scientist, but my intuition is that it is not possible to reach these conclusions "step by step", since uncomputable numbers are not reachable step by step, just as my intuition tells me that; "how it is for a junky to have an LSD hallucination", or how it is like for ordinary humans to-be in love is not crunching numbers, since my intuition tells me that these experiences have something more to do with biochemical reactions than just mere crunching meaningless strings!

[James T]

James T, do you know in principle how to program a computer to infer these conclusions, or not?Yes, what's your point. Peter, do you know in principle how it was possible for Cantor to work this out?

[David Z]

There is no way that a level of artificial intelligence comparable to a human being will be the result of an algorithm specified by a human being. We are certainly capable of describing algorithms that are intelligent on the scale of an insect, but not on the scale of even a human toddler. It is clear that a machine intelligence will have to be largely self-organizing, and that to some extent it will also be a black box, so that we are not easily able to alter its thought processes as we would like.

For example, how do we encode the concept of "Harm no human" in an inference engine that has grown from a blank slate into a mass of tangled inferences that exhibits as much complexity as a human brain? The behavior of that entity is the result of the interaction of millions of inference rules for everything from the lowliest sensory data to the most complex abstract thought. The concepts of 'human' and 'harm' don't even exist in any discrete sense. The only concepts which we can easily deal with are the lowest level of perceptions... the senses of the machine. The inferences built upon this sensory basis will be bewilderingly complex, and the inferences built upon those inferences will be even less tractable. Just as we cannot open the skull of a psychotic human being and surgically remove the 'crazy', we will not be able to open up a machine and insert 'morality'.

Note that I am not talking about a particular implementation such as fuzzy logic, neural nets, or whatever. I am assuming only a robust learning algorithm, and the capacity for human-level sophistication. The complexity that will develop in such a machine is the reason it's behavior was not elaborated in an algorithm in the first place.

The result of the learning of such a machine is hard to predict, but there are certain capabilities which will lead, inevitably, to problems. If it is capable of believing in things which it has not directly experienced (e.g. Antarctica), it is apt to believe in things for which it has no counter-evidence (e.g. jackalopes). If it takes action in the face of uncertainty, it will be capable of making mistakes. If it has imagination, it will imagine things.

It will be non-trivial to find the balance point where the machine's thought processes are in the useful area between gullibility and denial, timidity and recklessness, banality and lunacy, etc.

[James T]

David Z, I find myself both agreeing with your statements on complexity and our ability to describe even on the scale of an insect, and disagreeing with your concerns over how to encode concepts.

I certainly agree that at some point AI will have to be self-programming on process because we will not be able to cope with the level of complexity. On the other hand good programming practice is about separating knowledge from process. I expect we will retain the ability to be able to tinker with the knowledge.

[David Z]

James T, I am not sure I understand what you mean by the separation of knowledge from process. Do you mean that an AI could be made that has open-ended knowledge (what it thinks about), but a fixed process (how it thinks). Is that right? If so, I disagree. I think the AI should be able to not only think about new things but also change the way in which it thinks about them, so we are essentially dealing with self-modifying code, making good programming practice a moot point.

I should note that I have never heard any of the authorities voice opinions similar to mine regarding the future of AI. Truth isn't a democracy, of course, but mine is still the minority opinion.

[James T]

James T, I am not sure I understand what you mean by the separation of knowledge from process. Do you mean that an AI could be made that has open-ended knowledge (what it thinks about), but a fixed process (how it thinks). Is that right? If so, I disagree. I think the AI should be able to not only think about new things but also change the way in which it thinks about them, so we are essentially dealing with self-modifying code, making good programming practice a moot point.No that's not right at all. There will be a split between knowledge (what it thinks about) and process (how it thinks). But neither of these things are fixed and good programming practice is never a moot point.

It makes good sense to create the AI in such a way as it's knowledge may be accessed directly by people. Also it is common in AI work to set up the process in such a way that it is able to explain HOW it came to a conclusion. After all you can't improve the thought process without being able to understand it.

[Jay GW]

Part of what gives people (most people) their self awareness is the physical form they are in. Everyone can perceive that they are distinct entities, separate from everything else.

How exactly does a machine come to see itself as an independent "thing"?

[Marduk]

Not to mention how a machine that can perceive (if it thinks) down to the trillionths of seconds or less. Its existence/perception will be nothing like ours, to it we will be some incredibly slow moving annoyance.
It will do 100 quadrillion things in the time it takes us to sip our morning coffee.

[BioBeing]

I'm about 1/3 of the way through reading Descartes' Error : Emotion, Reason, and the Human Brain (http://www.amazon.com/exec/obidos/tg/detail/-/0380726475/internetinfidels/). Where I am, Damasio is just starting to make the argument that emotion is needed for the human brain to function correctly (make "good" decisions). Will an AI system need to mimic/possess some form of emotion in order to function like a human brain?

[pmurray]

He proves conclusively that any formal system that is of sufficient power to allow self-reference must be either inconsistent or incomplete

Well duh! There's your answer! Humans are inconsistent, as the most cursory investigation will reveal. People regularly think mutually incompatible things.

[David Z]

No that's not right at all. There will be a split between knowledge (what it thinks about) and process (how it thinks). But neither of these things are fixed and good programming practice is never a moot point.

It makes good sense to create the AI in such a way as it's knowledge may be accessed directly by people. Also it is common in AI work to set up the process in such a way that it is able to explain HOW it came to a conclusion. After all you can't improve the thought process without being able to understand it.Not every machine learning tactic is able to explain its functioning. Traditional AI that creates inferences certainly can, and variations on this theme with fuzzy sets can also, but neural nets are not so transparent. You have a tangled mess of equations that are too complex to extract any coherent rules from in all but the simplest cases. Here I am showing my connectionist bias, of course. I think that the future of AI lies in an elegant hardware implementation of recursive neural nets with unsupervised learning. Something like the ART architecture.

Good programming practice is moot because the code is self-modifying. Heck, the idea that the AI is implemented in 'software' at all is a questionable assumption. One could argue that if it is implemented in a custom IC that is capable of learning and altering its own configuration, you are working with self-modifying hardware.

[James T]

I think that the future of AI lies in an elegant hardware implementation of recursive neural nets with unsupervised learning.Ah, I see your bias as you see mine. Yes I agree that a neural net will not be able to explain it's thought process.

My main reasons for disagreeing with you about recursive neural nets fall into three categories.

Firstly they fail to integrate lots of really clever coding work being done and being improved in all the narrow areas of thought where programming has become quite advanced. In fact I see this as the main reason behind AIs first run of dissappointing results.

Secondly, what's the point of creating something that we have no understanding of. Almost irrespective of how it works whats the point. A fundamental part of knowledge is being able to explain the knowledge. Why invest in something that is another version of a person, with the vagaries in reasoning that go with it. How will it help to have another 'opinion'.

Finally, I don't like neural nets. I know, it's irrational, but they fall into a category for me of being the next great solution to every problem and the fundamental fervour that people express with each new wonderful fix all solution turns me off.

[Jay GW]

Will an AI system need to mimic/possess some form of emotion in order to function like a human brain?

I brought this up I think, a few posts back, that plastics and metal can't duplicate tissues exactly. Not approximately, EXACTLY. That's what's needed for a machine to think as a human being does, isn't it?

Since the perception of pain is tied directly to the fact that humans feel it in the body, and it causes emotions, alot of them, how would a machine duplicate that? If it doesn't, it's not really a mind like a humans.

[premjan]

I think emotion may be a carbon-centric mechanism, the corresponding elements of intelligence in an electronic creature would be a bit different (perhaps would just consist of "change/motion" sensors as someone suggested in the Philosophy forum. Also there is not necessarily a good reason to go for the same kinds of intelligence as seen in carbon lifeforms. Something different might crop up in electronic Intelligence.

[James T]

I think emotion may be a carbon-centric mechanism,This sounds an awful lot like Special Pleading, one of those logical fallacies.

[premjan]

This sounds an awful lot like Special Pleading, one of those logical fallacies.
What makes you think that silicon brains need an "emotion" mechanism? Two totally different hardware types indicate that the emergent structure may be far from identical. What makes this special pleading? Even at the chess playing level, the strategies typically applied by humans and machines are different (Go illustrates this better than Chess). Why not at the emotion level? Computers are just not fuzzy enough to do this emotion thing, IMO.

[James T]

What makes you think that silicon brains need an "emotion" mechanism? Two totally different hardware types indicate that the emergent structure may be far from identical. What makes this special pleading? Even at the chess playing level, the strategies typically applied by humans and machines are different (Go illustrates this better than Chess). Why not at the emotion level? Computers are just not fuzzy enough to do this emotion thing, IMO.I didn't say that. In any event I'd say that they could have one, certainly not need.

I was refering to your comment that you thought emotion might be carbon-centric. Special pleading for a carbon based brain to have emotion. I criticised this because not only are you giving carbon centric a special position but you are talking at the hardware level about a software issue, also a category error.

[Majestyk]

I'm about 1/3 of the way through reading Descartes' Error : Emotion, Reason, and the Human Brain (http://www.amazon.com/exec/obidos/tg/detail/-/0380726475/internetinfidels/). Where I am, Damasio is just starting to make the argument that emotion is needed for the human brain to function correctly (make "good" decisions). Will an AI system need to mimic/possess some form of emotion in order to function like a human brain? I'll have to read that book.

I've been mulling over the idea that emotion might be a prerequisite for intellignece. Assuming, the basis of intelligence is the ability to associate then, the root criteria for establishing an associtation must have a purpose else, the resulting network would be incoherent.

If, purpose is required and the system is to be autonomous then, there must be an initial set of states to be acheived and avoided. Then, low level mechanisms for responding to the current state, could provide the foundation on which a personality may develop.

I don't yet see a reason to discount the possibilty of incorporating an emotional structure into a non-organic design. The complexity would be astronomical but when you get down to the core of the issue, we all may just be incredibly sophisticated stimulus/response patterns.

[James T]

Emotion a necessity ... hmm, not sure I agree with this. However I though it might be a useful time to have a look for a definition of intelligence. I found one at wikipedia I didn't like (included social behaviour not really practical to an AI, children for instance). The second one that looked interesting is http://en.wikipedia.org/wiki/Sternberg%27s_Triarchic_Theory_of_Intelligence.

I'm not really familiar with Sternberg, hopefully the following brief restatement is relatively accurate.

componential Sternberg labeled these metacomponents, performance components, and knowledge-acquisition. Telling the mind how to act, carrying out the directed actions and selecting and combining elements of information to create new information.

experiential how well a task is performed with regard to how familiar it is . Finding new ways for novel tasks, automatically carrying out oft repeated tasks.

practical through the three processes of adaptation, shaping, and selection. Changing to suit the environment. Changing the environment and changing to a completely new environment to suit needs.

[Huon]

A good precaution would be to insert a device cutting the energy of the robot, so that it could not bypass it. Now, computers can restart through a program, but if you take off the electric wire, nothing happens.

"Who is the master, you, or me ?"

[James T]

A good precaution would be to insert a device cutting the energy of the robot, so that it could not bypass it. Now, computers can restart through a program, but if you take off the electric wire, nothing happens.

"Who is the master, you, or me ?"A book I have reread three or four times, James P Hogan, The Two Faces of Tomorrow. http://www.sfreviews.net/2faces.html, cutting the wire is no where near being good enough.

[David Z]

A book I have reread three or four times, James P Hogan, The Two Faces of Tomorrow. http://www.sfreviews.net/2faces.html, cutting the wire is no where near being good enough.Woo hoo! Sounds like a good one. I ordered it from Amazon used for the inexplicable price of $0.01.

[James T]

Woo hoo! Sounds like a good one. I ordered it from Amazon used for the inexplicable price of $0.01.It's worth every cent :).

[Huon]

... cutting the wire is nowhere near being good enough.

I am quite convinced of that. In my car, there is a battery, I don't know exactly what is it for :p but at the time of my previous message, I could not find a better image ...

James T and David Z, you taught me something about SF. Thank you.

Coming back to Artificial Intelligence, I suppose there is a problem about Intelligence. "Artificial" means obviously "man-made by industrial means". Arte-fact : made by art.

And then, what is the exact definition of "Intelligence" ?
Are we so well aware of the working of the brain ?
Can we describe exactly what happens in my brain just before I think "I am thirsty" ?
With what tests can we decide that something is intelligent ?
Don't you dare speak of IQ ! IQ measures adaptation to a certain state of the society, not really "intelligence", except if you define "intelligence" as "what IQ measures" :D

[David Z]

Coming back to Artificial Intelligence, I suppose there is a problem about Intelligence. "Artificial" means obviously "man-made by industrial means". Arte-fact : made by art.

And then, what is the exact definition of "Intelligence" ?
Are we so well aware of the working of the brain ?
Can we describe exactly what happens in my brain just before I think "I am thirsty" ?
With what tests can we decide that something is intelligent ?
Don't you dare speak of IQ ! IQ measures adaptation to a certain state of the society, not really "intelligence", except if you define "intelligence" as "what IQ measures" :DYou are right on the money, in that there is no generally accepted definition of human intelligence, and that how it works is unknown. It sounds really bad to be working on artificial intelligence when you can't even define natural intelligence, I admit. The only thing I can say in defense of the field is that when people say they are working to create artificial intelligence, they usually mean they are working to reduce machine stupidity... a far more humble goal.

[James T]

A nice quote from Edsger W.Dijkstra, EWD898, The threats to computing science.The Fathers of the field had been pretty confusing: John von Neumann speculated about computers and the human brain in analogies sufficiently wild to be worthy of a medieval thinker and Alan M.Turing thought about criteria to settle the question of whether Machines Can Think, a question of which we now know that it is about as relevant as the question of whether Submarines Can Swim.