|
|
Recently there was a competition for the Loebner Prize, a prize in the field of artificial intelligence for the best performance in a challenge that is basically the classic Turing test: human judges communicate via computer with a chatbot and a human, and they have to determine which is which. This year, there was a neat result – in one of the rounds, a chatbot named Suzette actually managed to fool one of the judges.*
* Apparently what happened was a combination of two things. First, the human being tested was “all over the place -- confusing, vague”. Second, the judge tried to be clever by asking the same question (about who to vote for in some provincial election, which Suzette couldn’t answer because she didn’t know the candidates or the race) over and over and over, which Suzette was designed to notice, and to respond to by first getting irritated, then bored.
So, this is clearly turning out to be one of those philosophical problems that becomes a practical (scientific) problem, but as yet, philosophy has no clear answer.
Here’s the hypothetical: let’s assume they actually do manage to make a chatbot that is capable of passing as a human in conversation so well that it scores 50% on Turing tests (in other words, the judges are just guessing 50-50 which is the human and which is the chatbot). Would you call that chatbot “intelligent”?
If not, what would you require to consider a program “intelligent”? How would you test for those qualities? Why would these qualities only exist in an intelligent agent, and not just a very sophisticated but unintelligent program? Or, is intelligence simply a matter of sophistication – the more complicated you make a program, eventually it will become intelligent? And bear in mind, whatever test you design would have to be able to be passed by humans, too (unless you think humans aren't intelligent).
Note: don’t bother trying to define “intelligence” – that’s a rabbit hole with no end. Instead, just try to come up with a reasonable estimation of what we should consider intelligent, and why, and how to test for that.
Basically: What would you consider a valid test for intelligence, and why? What would the test be testing for, and why would only intelligences have that quality?
Judging intelligence is indeed a very hard thing to do, partly because, as you say, intelligence can mean a lot of different things. I think intelligence is a result of multiple parameters rather than a single one, which makes it even harder to judge.
I would say that for anything to be considered intelligent, it must not just be able to talk about some subject in the way that a human would, it would have to actually understand the subject as well. It should be able to talk rather than just mimic talking. This means that the program should be able to draw conclusions upon existing data, not just repeat pre-programmed sentences and responses. (In an educational context, this means that the student should be able to grasp a concept and see what it implies, rather than learn a set of premade answers.)
The problem when judging is that this is mostly a matter of how things are made rather than how they behave. A theoretical scenario:
Let's say we have a chatbot with premade answers to all questions and statements that could possibly come up in any conversation, ever, an infinite amount of data (in practice impossible, just theoretically). It would be impossible to seperate this chatbot from one that is actually capable of thinking.
One thing I am certain about is that intelligence is not limited to biological creatures (in theory, whether current artificial intelligence could be considered intelligent is another question).
I don't have an answer just yet; cognition isn't an area that I have studied to any depth... but I would put forth that the answer is probably not a single trait, but a suite of traits, much like life.
Good luck formulating responses everyone. This is a really tough question.
It is indeed a biggie - one of the biggest.
Rather than attempt an answer at this early stage, I'll offer a possible line of attack on the problem.
It can be argued (and has been argued) that the mind is a symbol manipulation device.
Since we know (or at least I think we do) that intelligence is a property of the mind, then it is reasonable to suggest that intelligence is a process of symbol manipulation.
| catscratches wrote: | | I would say that for anything to be considered intelligent, it must not just be able to talk about some subject in the way that a human would, it would have to actually understand the subject as well. It should be able to talk rather than just mimic talking. This means that the program should be able to draw conclusions upon existing data, not just repeat pre-programmed sentences and responses. (In an educational context, this means that the student should be able to grasp a concept and see what it implies, rather than learn a set of premade answers.) |
Let's say that i could write an "expert system" of such sophistication that it did have - in the model it has in its code - as good a mental model about a single topic as an average human would. For example, let's say it had internally a complete model of the functioning of an internal combustion engines, including how all relevant environmental factors (temperature, humidity) would affect the process. If you were to "talk" with this program about internal combustion engines, it would know as much or more as any human, and it would respond correctly to any question. For example, if you tried to ask a non-sequitur like "will the engine work as well in the rain?" it would respond intelligently and correctly (by telling you how the efficiency would change... and if you tried a non-sequitur like "will the engine work as well on a day with 'r' in the name?" it will correctly respond by saying that doesn't matter.
In other words, it will correctly respond to all intelligent questions, and correctly identify questions that are irrelevant.
Would this be intelligence?
| catscratches wrote: | The problem when judging is that this is mostly a matter of how things are made rather than how they behave. A theoretical scenario:
Let's say we have a chatbot with premade answers to all questions and statements that could possibly come up in any conversation, ever, an infinite amount of data (in practice impossible, just theoretically). It would be impossible to seperate this chatbot from one that is actually capable of thinking. |
Wouldn't that be intelligence? How would you know the difference, or, more precisely, does the difference matter?
In other words, say you had such a chatbot (and really, you wouldn't need it to have an infinite amount of knowledge, just a comparable amount to human knowledge... and some modern practical designs "feed" on the Internet to be trained - some using things like Tweets, believe it or not). It can do anything an intelligent human can do... why isn't it intelligent?
| Bikerman wrote: | | It is indeed a biggie - one of the biggest. |
i got tired of dealing with small potatoes... bring on the real questions... the BIG questions.
| Bikerman wrote: | Rather than attempt an answer at this early stage, I'll offer a possible line of attack on the problem.
It can be argued (and has been argued) that the mind is a symbol manipulation device.
Since we know (or at least I think we do) that intelligence is a property of the mind, then it is reasonable to suggest that intelligence is a process of symbol manipulation. |
So let's say hypothetically that we have a program like Maple that can do (mathematical) symbolic manipulation with fantastic accuracy and dexterity. Would that be a sign of intelligence? If so, certainly that's not enough, so what's missing?
I would say in order for a program to be truly intelligent(without defining it) it would be required to adapt and evolve. I agree that a large bulk of data is not enough. It is maybe all about how we use that data interactively to the stimuli given.
If asked a question, “Where does the gas go?” any chat bot could answer. (in the gas tank) Asking the same question a second time a program could select another thread to go to (in the gas filler tube) but it must do it with some learning capacity, I would assume to qualify as intelligent. (what is your problem? What side of the vehicle are you on?) could be the result of a programmed flow chart type logic however I think the real test of intelligent responses, if it was indeed possible for the program to do would be to go through many loops of possible scenario’s on its own to determine where the dialogue is going (chess). In other words follow the past logic and project it into the future which the human brain can do. (Predict the future and learn from it – possible?)
Science fiction writers I think had it nailed down with Data on the TV series Star Trek a machine who showed intelligence far above the actual starship computer. He actually took data and played with it to evolve on his own.
It makes me think of old window utilities (9x) whereby they were able to define a problem but not fix it. As computer systems are designed with decision making capabilities built in they become more intelligent needing less and less human intervention to function correctly. I suppose that is a form of artificial intelligence.
That's not a bad attempt. Using this criterion (adapt and evolve) is tricky though. Plants adapt and evolve, albeit on a longer timescale, and I certainly wouldn't want to call plants intelligent. We could specify adapt and evolve responses in communication - but then we run into another problem. Would a human without sensory input still be intelligent? The nearest we can come to that is probably deaf-blind children and I wouldn't want to say they were not intelligent.....
interesting starting point though.....
If it is indistinguishable from a human (who is an intelligent being), then it is indistinguishable from an intelligent being.
Simple logic,
All* humans are intelligent.
This machine is identical to a human. (As far as we can tell.)
Therefore, the machine is intelligent (as far as we can tell).
*The examples of non-intelligent humans can be ignored for this purpose.
Same thing as,
All these marbles are green.
This ball is indistinguishable from the marbles.
Therefore, the ball is green.
My point is, no matter how well you define intelligence, the machine will pass your test as long as it can pass Turing tests consistently. (Unless you define intelligence so strictly that many humans fail to qualify as well.)
-- And I suspect that there is no way to consistently pass Turing tests without actually being intelligent. It will be faking its humanity, but in order to do that convincingly, it needs to be intelligent.
(Also, I saw that news headline. The headline was extremely over-hyped, in that any realistic conversation with that chat-bot will quickly show it is not a human... All that happened is that this particular bot was designed to beat the simplistic trick the judge was using.)
| Indi wrote: | | catscratches wrote: | | I would say that for anything to be considered intelligent, it must not just be able to talk about some subject in the way that a human would, it would have to actually understand the subject as well. It should be able to talk rather than just mimic talking. This means that the program should be able to draw conclusions upon existing data, not just repeat pre-programmed sentences and responses. (In an educational context, this means that the student should be able to grasp a concept and see what it implies, rather than learn a set of premade answers.) |
Let's say that i could write an "expert system" of such sophistication that it did have - in the model it has in its code - as good a mental model about a single topic as an average human would. For example, let's say it had internally a complete model of the functioning of an internal combustion engines, including how all relevant environmental factors (temperature, humidity) would affect the process. If you were to "talk" with this program about internal combustion engines, it would know as much or more as any human, and it would respond correctly to any question. For example, if you tried to ask a non-sequitur like "will the engine work as well in the rain?" it would respond intelligently and correctly (by telling you how the efficiency would change... and if you tried a non-sequitur like "will the engine work as well on a day with 'r' in the name?" it will correctly respond by saying that doesn't matter.
In other words, it will correctly respond to all intelligent questions, and correctly identify questions that are irrelevant.
Would this be intelligence? |
I think it depends on how the program arrives at the answers. Consider it like a book where you have an index of questions. "Would the engine work in rain?", for example. Then you look up the relevant page listed in the index and read the answer. I would not consider this book intelligent (it would, however, be very handy). The difference between this and the computer program is only that the computer program automatically finds the question in the index, goes to the page and prints the answer. It is merely a set of premade answers and questions. If I ask it a question that is not listed yet, it wouldn't be able to respond.
Now, to make this program intelligent, I think it would be able to use sets of data and draw conclusions based on them. For example if I have the following two sets of data:
1. Red things reflect light with the wavelength of 630–740 nm.
2. A firetruck is red. (What? Not all firetrucks are red? Nonsense!)
Out of that information, it would be able to conclude that:
Firetrucks reflect light with the wavelength of 630–740 nm.
But yes, if the program was able to see what questions were relevant, etc., without having to rely on a set, definite index of questions, I think that would count as another way of drawing conclusions, just based on input rather than predefined data.
Now, I don't think it is important where in the process this happens. It can very well be an index of questions and answers as long as the index is generated (at least to some part) by drawing conclusions.
| Indi wrote: | | catscratches wrote: | The problem when judging is that this is mostly a matter of how things are made rather than how they behave. A theoretical scenario:
Let's say we have a chatbot with premade answers to all questions and statements that could possibly come up in any conversation, ever, an infinite amount of data (in practice impossible, just theoretically). It would be impossible to seperate this chatbot from one that is actually capable of thinking. |
Wouldn't that be intelligence? How would you know the difference, or, more precisely, does the difference matter? |
I think the difference is that one is intelligent whether the other just appears to be intelligent. It would be impossible to seperate them from a viewer's perspective. You can't really ever test how things are, just how they appear to be. Then you simply have to trust that they are what they appear to be. Whether it matters... I don't know. Probably not.
| Bluedoll wrote: | | I would say in order for a program to be truly intelligent(without defining it) it would be required to adapt and evolve. |
Evolution and adaptation during an organism's lifetime (rather than over the span of the species) is just classical conditioning. If that's all you need to be intelligent, then worms are intelligent.
| ocalhoun wrote: | If it is indistinguishable from a human (who is an intelligent being), then it is indistinguishable from an intelligent being.
Simple logic,
All* humans are intelligent.
This machine is identical to a human. (As far as we can tell.)
Therefore, the machine is intelligent (as far as we can tell). |
There's a fallacy in here. Being indistinguishable from something else does not mean being the same as something else, nor does it mean you have the same qualities as that other thing.
Suppose i had a violin made by Antonio Stradivari. Suppose i used a molecular replication technique to make an exact duplicate of that violin, atom for atom. It is now completely impossible to tell the two apart - they are literally indistinguishable. Is the second violin also made by Antonio Stradivari? No, of course not. Similarity is not the same as being identical.
The only way traits can carry over from one thing to a copy of that thing is if the copy process directly or indirectly takes that trait into account. In other words, molecular duplication of the violin will copy the traits of mass, volume, texture, colour and so on, because all of those things are directly or indirectly the properties of the things being duplicated - the molecules. It will not copy traits like the history of the violin.
Similarly, if you were to create something that is completely indistinguishable from a human, intelligence would only be carried over into the copy if the copy process somehow copied it. How would you know that it did? How would you know that intelligence was copied, if we can't even define what intelligence is?
So basically, if you've got something that you're calling a human analogue - something that you're saying is indistinguishable from a human - that's not enough. You have to actually be sure that the same human intelligence is in the analogue, and if you can't be sure that you put it there (because you can't define intelligence), the only way that you can know the analogue has human intelligence is if you test for it to make sure it's there.
Which, long story short, means you've come full-circle. You need to figure out a way to test for intelligence. So, you're right back where you started. All you've done is changed the nature of the problem... by complicating it. Now you're not just testing for intelligence. Now you're trying to compare intelligences - the intelligence of a human with the intelligence of you test subject. In other words, you don't only need a test for intelligence, you need a test for intelligence with a way to quantitatively score the result.
| ocalhoun wrote: | | (Also, I saw that news headline. The headline was extremely over-hyped, in that any realistic conversation with that chat-bot will quickly show it is not a human... All that happened is that this particular bot was designed to beat the simplistic trick the judge was using.) |
Actually, that's not all that happened. The judge was a doofus, but so was the other test subject. The other test subject was random, and incoherent.
| catscratches wrote: | I think it depends on how the program arrives at the answers. Consider it like a book where you have an index of questions. "Would the engine work in rain?", for example. Then you look up the relevant page listed in the index and read the answer. I would not consider this book intelligent (it would, however, be very handy). The difference between this and the computer program is only that the computer program automatically finds the question in the index, goes to the page and prints the answer. It is merely a set of premade answers and questions. If I ask it a question that is not listed yet, it wouldn't be able to respond.
Now, to make this program intelligent, I think it would be able to use sets of data and draw conclusions based on them. For example if I have the following two sets of data:
1. Red things reflect light with the wavelength of 630–740 nm.
2. A firetruck is red. (What? Not all firetrucks are red? Nonsense!)
Out of that information, it would be able to conclude that:
Firetrucks reflect light with the wavelength of 630–740 nm. |
You think that would be intelligence? If that's so, then all you would need to create an intelligence is a big enough database, and a very simple query engine.
For example. i could create a database entry for "red" that included facts like its wavelengths, its artistic interpretations and a couple of associated keywords (apple, blood, etc.)... and a database entry for "firetruck" that included facts like that it's a truck, it pumps water, etc., and that it's usually red. So now you ask it "what wavelengths of light does a firetruck normally reflect?"... it does a simple query on firetruck, finds out that it's normally red in colour, and do a second query on red to find out the wavelengths. Problem solved.
Is that intelligence?
| catscratches wrote: | | I think the difference is that one is intelligent whether the other just appears to be intelligent. It would be impossible to seperate them from a viewer's perspective. You can't really ever test how things are, just how they appear to be. Then you simply have to trust that they are what they appear to be. Whether it matters... I don't know. Probably not. |
That's evasive though, because although you can never make a TRULY direct measurement of anything, you can make measurements that are more direct than others. For example, i can measure the length of a wire by taking its material density, then measuring its volume using a displacement test, then calculating its cross sectional area by passing a current through it and using the resistivity, and finally use all that to calculate the length. Or, i can get an object calibrated to a defined length (like a ruler, calibrated in centimetres), and just sit them side-by-side and compare. Clearly the latter is a more direct measurement of length than the former, because it requires far fewer translations between the measurements and the conclusions.
If we can't measure intelligence directly, we have to find incidental features of intelligence that correlate directly to it. The stronger the correlation, the better the measurement. We can measure the length of a wire qualitatively by weighing it: the more it weighs, the longer it is... usually. Or, we can measure the length of a wire qualitatively by shooting something at a constant velocity along it, and noting the times that it crosses the beginning and end: the longer it takes, the longer the wire is... always. The latter is far better correlated to length.
So the question is: is being able to connect facts strongly correlated to intelligence... or weakly?
| Indi wrote: | | catscratches wrote: | I think it depends on how the program arrives at the answers. Consider it like a book where you have an index of questions. "Would the engine work in rain?", for example. Then you look up the relevant page listed in the index and read the answer. I would not consider this book intelligent (it would, however, be very handy). The difference between this and the computer program is only that the computer program automatically finds the question in the index, goes to the page and prints the answer. It is merely a set of premade answers and questions. If I ask it a question that is not listed yet, it wouldn't be able to respond.
Now, to make this program intelligent, I think it would be able to use sets of data and draw conclusions based on them. For example if I have the following two sets of data:
1. Red things reflect light with the wavelength of 630–740 nm.
2. A firetruck is red. (What? Not all firetrucks are red? Nonsense!)
Out of that information, it would be able to conclude that:
Firetrucks reflect light with the wavelength of 630–740 nm. |
You think that would be intelligence? If that's so, then all you would need to create an intelligence is a big enough database, and a very simple query engine.
For example. i could create a database entry for "red" that included facts like its wavelengths, its artistic interpretations and a couple of associated keywords (apple, blood, etc.)... and a database entry for "firetruck" that included facts like that it's a truck, it pumps water, etc., and that it's usually red. So now you ask it "what wavelengths of light does a firetruck normally reflect?"... it does a simple query on firetruck, finds out that it's normally red in colour, and do a second query on red to find out the wavelengths. Problem solved.
Is that intelligence? |
It would meet that criteria, yes. Though I think it would have to be slightly more advanced than that and understand conditions in those searches. Eg:
1. An apple becomes red when it ripens.
2. Red things reflect light with the wavelength of 630–740 nm.
It should not draw the conclusion that apples are red. It should draw the conclusion that apples are red when ripened.
Though that's quite an easy fix, I imagine.
As I said, that's only one critera, however.
| Quote: | | I think intelligence is a result of multiple parameters rather than a single one, which makes it even harder to judge. |
Another criteria would be learning. That's quite an easy one and has been done, as you said, by tying to extract information from Twitter feeds, user input and all kinds of different sources.
| Indi wrote: | That's evasive though, because although you can never make a TRULY direct measurement of anything, you can make measurements that are more direct than others. For example, i can measure the length of a wire by taking its material density, then measuring its volume using a displacement test, then calculating its cross sectional area by passing a current through it and using the resistivity, and finally use all that to calculate the length. Or, i can get an object calibrated to a defined length (like a ruler, calibrated in centimetres), and just sit them side-by-side and compare. Clearly the latter is a more direct measurement of length than the former, because it requires far fewer translations between the measurements and the conclusions.
If we can't measure intelligence directly, we have to find incidental features of intelligence that correlate directly to it. The stronger the correlation, the better the measurement. We can measure the length of a wire qualitatively by weighing it: the more it weighs, the longer it is... usually. Or, we can measure the length of a wire qualitatively by shooting something at a constant velocity along it, and noting the times that it crosses the beginning and end: the longer it takes, the longer the wire is... always. The latter is far better correlated to length.
So the question is: is being able to connect facts strongly correlated to intelligence... or weakly? |
I think it's only one criteria out of several. But it is necessary for intelligence. Since intelligence is an umbrella term for different qualitites, I don't think we can give a direct measure like that. We have to test different criteria rather than a single thing. But yes, if you could bundle several of these criterias together in one test and make it more direct, that would be sweet. I have no idéa how that would be done, though.
So, unless we can find some way to combine them in one test, I can't answer your original question. Though I think ansking for a single test to determine an umbrella term is quite unfair. 
In the case of a computer program, we really do have access to more direct measurement, though. Since we know the exact code (unless it's made to evolve, but that's not a criteria for intelligence, imo) we can know just how it works rather than how it appears to work. But well, that's cheating. ;p
One thing worth noting is that I don't believe an intelligent being has to be indistinguishable from a human being and that that should be part of the test. In my opinion, that's quite an egocentric viewpoint. There are other animals than humans who are intelligent but you wouldn't confuse eg. a dolphin for a human.
I think that with the latest advances in artificial intelligence, and especially with the human lookalike robot in Japan who works as an observing nurse, its only a matter of time before actual robots are built, and I hope not to replace us, but the tedious tasks that we have,  except politicians jobs.
I think basic intelligence is necessary,a nd will prove beneficial to probably police, who will use AI as lie detectors, and other kinds of tasks.
| Indi wrote: |
| ocalhoun wrote: | If it is indistinguishable from a human (who is an intelligent being), then it is indistinguishable from an intelligent being.
Simple logic,
All* humans are intelligent.
This machine is identical to a human. (As far as we can tell.)
Therefore, the machine is intelligent (as far as we can tell). |
There's a fallacy in here. Being indistinguishable from something else does not mean being the same as something else, nor does it mean you have the same qualities as that other thing.
|
True, but still, if the machine is indistinguishable from a human, then the intelligence of the machine will be indistinguishable from a human intelligence, no matter how you define that human intelligence.
...And I still suspect that actual intelligence of some sort would be required in order to convincingly simulate intelligence of any kind.
Much in the way a weak computer cannot convincingly simulate a powerful one, I don't think a weak intelligence could convincingly simulate a powerful intelligence.
OK let's try and have a go here.
The ability to distinguish differences, similarities and patterns in a useful way.
Intelligence comes in 2 types:
abilities which are learned from others probably accounts (opinion) for most of what we call intelligence. The ability to distinguish differences, similarities and patterns in occurrences in a way that allows us to derive useful information AND make a choice/action based on that useful information. Given enough dedication, we could develop a machine or set of machines that do this for us - the downside being that the machine(s) may actually reduce our personal intelligence as our reliance on the machine(s) increases. Still, learned intelligence has an enormous potential even though it is limited.
abilities to distinguish differences, similarities and patterns in a useful way which are learned from direct experience ungoverned by learning it from others. This occurs when an individual develops an ability to perceive useful indicators about a type of occurrence which has not been pointed out by someone else. This ability grows when someone is able to hold the mind and senses open to perceiving things which have not been deemed relevant by others AND continue to scan the data received to find out whether or not there is anything reliably informative to be derived from the data. This type of intelligence may not always be able to be taught to another until their faculties are acute enough to detect the data and as such may seem like nonsense to the perceivers peers, thus creating somewhat of a catch 22. Examples of this sometimes occur when an individual is deprived by whatever means of the normal sensory data input.
I don't see how the two are distinct.
The ability to discern patterns and extract meaning cannot be learned from others. People may teach you where such patterns exist, but you must discern them yourself, or simply take it on trust.
| catscratches wrote: | | So, unless we can find some way to combine them in one test, I can't answer your original question. Though I think ansking for a single test to determine an umbrella term is quite unfair. |
No one said it needs to be a test that measures a SINGLE metric. Just like a test for density that works by measuring both volume and mass, a test for intelligence could work by measuring a bunch of things. We just need to figure out what those things are, why they relate to intelligence, how to measure them and how to combine the results to get an intelligence estimate.
As a practical and topical example, suppose someone asked you about someone "Can they vote in a federal election?" (instead of "Are they intelligent?"). To test for that, you measure their age, you check for criteria like citizenship, legal standing, and mental capability, and if they are all up to snuff, the answer is yes. One test (for voting eligibility); multiple measurements of multiple metrics.
| ocalhoun wrote: | | True, but still, if the machine is indistinguishable from a human, then the intelligence of the machine will be indistinguishable from a human intelligence, no matter how you define that human intelligence. |
The fallacy is still there, and i don't think you understand why, so let me try and explain it another way.
Unless two things are identical - that is A ≡ B, or A and B share the same identity - then you cannot say two things are similar in a certain metric unless you test to see if they measure the same in that metric. In other words, if you have Bob and Steve... unless Bob is Steve you cannot say Bob and Steve have similar appearances unless you compare their appearances - for example, measuring each of their heights and comparing them, measuring their skin tones and comparing them, comparing the locations and sizes of their features, and so on. And in order to compare their appearances... you need a way to measure their appearances. Which is why you come full circle: you can't say two people have similar appearance until you find a way to measure appearance.
So there's your dilemma: unless your test subject is a human, you can't say it has similar intelligence to a human... until you measure the intelligence of both and compare them. Therefore, you need a way to measure intelligence.
So how do you measure intelligence? You compare it to human intelligence. How do you compare intelligences? You measure both intelligences and compare them. How do you measure intelligence? You compare it to human intelligence. ...
See? You're stuck in a circular problem.
| ocalhoun wrote: | ...And I still suspect that actual intelligence of some sort would be required in order to convincingly simulate intelligence of any kind.
Much in the way a weak computer cannot convincingly simulate a powerful one, I don't think a weak intelligence could convincingly simulate a powerful intelligence. |
Yes, real intelligence would be necessary to create a simulated intelligence, but the real intelligence in our case would be ours - human intelligence. Our intelligence is what creates the simulation, not the computer. The computer just runs the simulation.
Incidentally, a weak computer can simulate a more powerful one... it just can't do it at the same speed. But the fact that you need powerful enough hardware to run the simulation fast enough does not mean that the powerful hardware is where the intelligence comes from.
Consider a sophisticated physics simulation - and they exist. We don't have computers powerful enough to do these simulations in real time, but we can run them. They produce results indistinguishable from reality for whatever the problem domain is... but that doesn't mean physical reality exists within the machine, because you don't need to actually have something in order to convincingly and indistinguishably simulate it. All you need is that the person that designed the simulator understands the phenomenon well enough to design something that outputs results that are indistinguishable.
In other words, it is possible that one day we will understand what intelligence is well enough that we can produce a simulation of it that is indistinguishable from the real thing (and have hardware capable of running that simulation) - but that would just be a simulation of intelligence which doesn't necessarily mean that it is intelligence or that it has intelligence, any more than a simulation of soft-body physics actually is a physical universe, or has a physical universe inside (or even just a part of one).
| Bikerman wrote: | I don't see how the two are distinct.
The ability to discern patterns and extract meaning cannot be learned from others. People may teach you where such patterns exist, but you must discern them yourself, or simply take it on trust. |
That's the point I'm making Bikerman. If I show someone a pattern I can invariably get them to see it - show them enough times and they begin to notice it for themselves. Taking it on trust is both handy and potentially dangerous. Handy for the teacher/demonstrator as it allows the student/learner to be trained - potentially dangerous for the student/learner as (s)he could be getting led up the garden path. This is where discernment is of particular advantage.
Unfortunately, in my experience, most people take things on trust in a blind/poor-sighted way - it's trained in much the same way as the Pavlov's dog experiments. This is handy for running planned social structure using signs and symbols to indicate other scenarios (like a skeleton's head on a pole to act as a kind of scarecrow for oncoming enemies or a road-sign telling us that elderly people may be crossing the road ahead), however unless the student develops a pattern-testing mechanism, it's not really intelligence in the truest sense. Look at how people slow-down when they see a cardboard police car or how they stiffen up when confronted by a cardboard officer. What I'm saying here, is it's okay for someone to show a pattern, though the intelligent approach would be to test its relevance.
The 'other' (or better phrased 'higher') kind of intelligence is developing the ability to automatically scan for patterns and test the relevance of any which are found. I've come across many people, (even well educated people) who either clam up when the text books or their previous experience doesn't have an answer or they fly into wild speculations. Of course, there's nothing wrong with speculating as it 'keeps the eyes open' so to speak, though it's important to keep anything found in the speculations 'box' until proven otherwise.
I still don't see any difference. ALL people learn to spot patterns. We all have some inate pattern spotting ability, but that is not necessarily a good thing - it is frequently a bad thing. We are 'programmed' for example, to spot faces, even where they do not exist. Hence people often see faces in clouds, flames, and a variety of other situations. That is not, to me, a sign of intelligence and certainly not a 'higher' form.
Patterns which ARE useful are invariably learned - such as patterns in numbers etc. It is true that some people find it easier than others.
You are simply talking about degrees of the same ability, not two distinct skills/abilities.
| Bikerman wrote: | I still don't see any difference. ALL people learn to spot patterns. We all have some inate pattern spotting ability, but that is not necessarily a good thing - it is frequently a bad thing. We are 'programmed' for example, to spot faces, even where they do not exist. Hence people often see faces in clouds, flames, and a variety of other situations. That is not, to me, a sign of intelligence and certainly not a 'higher' form.
Patterns which ARE useful are invariably learned - such as patterns in numbers etc. It is true that some people find it easier than others.
You are simply talking about degrees of the same ability, not two distinct skills/abilities. |
| jeffryjon wrote: | The ability to distinguish differences, similarities and patterns in a useful way.
Intelligence comes in 2 types:
abilities which are learned from others probably accounts (opinion) for most of what we call intelligence. The ability to distinguish differences, similarities and patterns in occurrences in a way that allows us to derive useful information AND make a choice/action based on that useful information. Given enough dedication, we could develop a machine or set of machines that do this for us - the downside being that the machine(s) may actually reduce our personal intelligence as our reliance on the machine(s) increases. Still, learned intelligence has an enormous potential even though it is limited.
abilities to distinguish differences, similarities and patterns in a useful way which are learned from direct experience ungoverned by learning it from others. This occurs when an individual develops an ability to perceive useful indicators about a type of occurrence which has not been pointed out by someone else. This ability grows when someone is able to hold the mind and senses open to perceiving things which have not been deemed relevant by others AND continue to scan the data received to find out whether or not there is anything reliably informative to be derived from the data. This type of intelligence may not always be able to be taught to another until their faculties are acute enough to detect the data and as such may seem like nonsense to the perceivers peers, thus creating somewhat of a catch 22. Examples of this sometimes occur when an individual is deprived by whatever means of the normal sensory data input. |
As you can see from the above, your points have already been covered.
| Bikerman wrote: | | I still don't see any difference. ALL people learn to spot patterns. We all have some inate pattern spotting ability, but that is not necessarily a good thing - it is frequently a bad thing. We are 'programmed' for example, to spot faces, even where they do not exist. Hence people often see faces in clouds, flames, and a variety of other situations. That is not, to me, a sign of intelligence and certainly not a 'higher' form. |
Is this not the creative part of intelligence? Humans' ability to visualize must be a necessary part of intelligence in order to facilitate solutions such as designing the space shuttle for example. That ability may also be responsible for some humans seeing things that are not real, but that is an offsetting negative of its positive function in intelligence. Perhaps a robot also needs to have those attributes in order to facilitate creative thinking? But then if robots do, that may be the point where things could get out of control, as maybe we may not get the mix right for thinking creatively in an intelligent way only?
I would think that intelligence would include the ability to remember. For example if I enter my house there could be a robot called Eva who immediately remembers what time of the day it is. I may have told her in the morning that I would like her to remind me of X, Y, Z. Memory of an interpretive kind would be even a higher level of intelligence. I.e. those things I did not tell her to remember, but that she thinks I should think about. Like I am running out of supplies and need to renew my motor vehicle license. In other words she can shift through her own database memories and come to interpretations that are necessary to be communicated.
Visualising is extrapolating or creating something that doesn't (yet) exist. Pattern spotting is something different. You don't need to spot existing patterns in order to visualise something.
| jeffryjon wrote: | | As you can see from the above, your points have already been covered. |
No they haven't. Your dichotomy is between taught pattern recognition and experiential pattern recognition. I say that is a completely false dichotomy. Whether you learn it from others or you learn it some other way does not seem to me to be relevant, and certainly not a demarcation between two different types of intelligence.
| Indi wrote: |
So how do you measure intelligence? You compare it to human intelligence. How do you compare intelligences? You measure both intelligences and compare them. How do you measure intelligence? You compare it to human intelligence. ...
|
My point though, is if it is indistinguishable from a human, then it will pass the human intelligence test, no matter how you measure it.
-Because if there was any discernible difference between the two, they would not be indistinguishable.
(I'm taking indistinguishable to mean that the human and the machine are basically the same, no matter what measurement we use.)
| Quote: |
Yes, real intelligence would be necessary to create a simulated intelligence, but the real intelligence in our case would be ours - human intelligence. Our intelligence is what creates the simulation, not the computer. The computer just runs the simulation.
Incidentally, a weak computer can simulate a more powerful one... it just can't do it at the same speed. But the fact that you need powerful enough hardware to run the simulation fast enough does not mean that the powerful hardware is where the intelligence comes from.
|
Okay, perhaps I overemphasized the weak vs. powerful part of the analogy...
(Though I also assumed that our intelligent machine was not given the benefit of extra time to compute responses... Because if it were, then you could tell the difference between it and a human by just asking questions too rapidly for the machine to keep up.)
A computer that is simulating a different computer must -- at some level -- execute the code that the simulated computer runs (or at least a more efficient version of it)...
Similarly, I think that if you asked an AI a question that required intelligence to understand and answer, somewhere within the software, something resembling intelligence would need to solve the problem, before forwarding the answer to the part that formats the answer in a human-like fashion.
This might be irrelevant to the thread, but programs cannot become artificially intelligent. Artificial intelligence does not exist, not in the sense of intelligence comparable to a dog or a cat, certainly not human.
I've had this debate before with my computer science lecturer, it's a very tedious one.
What does exist that's dubbed as Artificial Intelligence is programmed intelligence, which isn't any different to any other program in the sense that it was designed by someone to perform a certain task, in doing so it's not any more intelligent than your calculator. It might have been designed cleverly, perhaps cleverly enough to be mistaken for genuine intelligence, but it's still just a program written by someone, any meaning we draw from it is purely biased.
This is where it is important to mention a book called Simulacrum And Simulation.
The whole idea of artificial intelligence is really just a delusion/science-fiction/fallacy. The reason is simple, funny enough it lies in the human mind. We need to make sense of things, for we are intelligent.
The word mountain, what does it mean? You might know what I mean by the word, but the underlying concept of the mountain, it's physical existence is different. The word is just a model for a 'real' physical phenomenon, a form of simplification, a limit. Without the ability to build abstract data models we wouldn't be able to reduce the problem enough (consciously speaking) to handle the data. This is the concept of simulation (brutally simplified). Any Artificial Intelligence program you write is just that - a simulation of an idea, which can perform a certain task. We just get carried away with being able to witness a programmed logic have an effect, as if it were intelligent. But all we're doing is translating our own intelligence into logic, logic into machine code. Any illusions we transpire are just that, in effect a good simulation. But it is false underneath.
As for the definition of intelligence, the accepted view is that it is the ability to spot patterns. I disagree, it is the ability to create patterns, make sense of events to create further patterns and only then the ability to spot patterns.
It is a messy topic, but quite simple if you stop and pause for a second. I find talk of artificial intelligence really annoying though. The kind of denial you meet is akin to that of theists, perhaps AI is the deity of geeks and nerds
Well, I think you are fundamentally wrong.
Firstly let's debunk this notion that computers can only do what they are programmed to do. It is dead wrong. Neural Networks, as one example, are not 'programmed' to do anything except learn. Nobody determines in advance what pathways a signal will follow and nobody determines in advance what the balance/probability weightings of the final 'artificial neurone paths' will be after a period of learning.
If your lecturer did not point this out then he isn't much of a lecturer.
Secondly this notion of 'simulacrum'. That is exactly what your thoughts are. When you think about a mountain there isn't a real mountain in your brain - there is an image (simulacrum) which may or may not be of a specific mountain.
Thirdly - we know the brain is an interconnected series of neurones (switches). Why, then, is there some principled difference between a biological set of switches and an artificially constructed set in silicon? There isn't, to my mind. The difference at the moment is that our capacity to build switches is still very limited so the neural networks we build are still pretty small and dim. That will change, and IS changing.
| Bikerman wrote: | Well, I think you are fundamentally wrong.
Firstly let's debunk this notion that computers can only do what they are programmed to do. It is dead wrong. Neural Networks, as one example, are not 'programmed' to do anything except learn. Nobody determines in advance what pathways a signal will follow and nobody determines in advance what the balance/probability weightings of the final 'artificial neurone paths' will be after a period of learning.
|
There's a galactic difference between Neural Networks and CPUs running instructions sets, so my point that software cannot be artificially intelligence stands as yet, unfazed. I thing you're fundamentally wrong, as well.
| Bikerman wrote: |
If your lecturer did not point this out then he isn't much of a lecturer.
|
Well to be fair to him, he sort of did, but I was in 1st year back then so he was probably afraid of crushing my little argument, along with potentially my desire to question things etc. Credit is due to him for recognizing that some students can succumb to such things, although I dare say I wouldn't be one of them 
| Bikerman wrote: |
Secondly this notion of 'simulacrum'. That is exactly what your thoughts are. When you think about a mountain there isn't a real mountain in your brain - there is an image (simulacrum) which may or may not be of a specific mountain.
|
That is not exactly what I was getting at, but yes (vaguely). And?
| Bikerman wrote: |
Thirdly - we know the brain is an interconnected series of neurones (switches). Why, then, is there some principled difference between a biological set of switches and an artificially constructed set in silicon? There isn't, to my mind. The difference at the moment is that our capacity to build switches is still very limited so the neural networks we build are still pretty small and dim. That will change, and IS changing. |
The difference is that we're building static circuits that can only perform certain calculations - that's all it is. You can't even remotely compare it to Neural networks! Neural networks have the grand unification between processing, storage and memory/energy retention, not to mention that it's a living thing which can change/evolve. Where as anything a software can achieve is a simulation, or essentially fast, fast calculations. Modern Object Orientated approach is based on modeling, or simulating problems and solutions for science's sake!
Sure there's changes occurring in the computing world, quantum computing, biological computing, I don't know what else. I don't know much about quantum computing, so I won't theorize too much on it. However, I reason the only way we'll achiever artificial intelligence is if we essentially create life, so for example chain together organic tissue and/or bacteria to create a new organism (perhaps modeled after the human brain), which might give rise to intelligence, but what we'll find then is that we won't be able to program it, beyond perhaps simulating it's instincts and afterwards enslaving it, but not program it. (and it won't necessarily be better at calculations than our brain) But guess what we have then, a replica or I dunno what's another word for it?
Without being physically real and without the grand unification of memory, processing memory/energy retention and essentially life/evolution/change Artificial Intelligence is not possible. But then it won't be all that artificial, rather very real, but conditioned (or created) artificially.
| The-Nisk wrote: | | Bikerman wrote: | Well, I think you are fundamentally wrong.
Firstly let's debunk this notion that computers can only do what they are programmed to do. It is dead wrong. Neural Networks, as one example, are not 'programmed' to do anything except learn. Nobody determines in advance what pathways a signal will follow and nobody determines in advance what the balance/probability weightings of the final 'artificial neurone paths' will be after a period of learning.
|
There's a galactic difference between Neural Networks and CPUs running instructions sets, so my point that software cannot be artificially intelligence stands as yet, unfazed. I thing you're fundamentally wrong, as well. |
You can implement a NN very easily using traditinal Von-Neumann architecture. You don't NEED to - you can build NNs in hardware as well as software. Your point does not stand. There are many AI systems using neural net software and/or hardware and they are NOT pre-programmed to produce a particular outcome as you said.
| Quote: | | That is not exactly what I was getting at, but yes (vaguely). And? |
So there is no principled difference between symbolic representation in a computer and in a brain.
| Quote: | | The difference is that we're building static circuits that can only perform certain calculations - that's all it is. You can't even remotely compare it to Neural networks! Neural networks have the grand unification between processing, storage and memory/energy retention, not to mention that it's a living thing which can change/evolve. Where as anything a software can achieve is a simulation, or essentially fast, fast calculations. Modern Object Orientated approach is based on modeling, or simulating problems and solutions for science's sake! |
Of course you can compare it to Neural Networks. Neural nets are not 'living' things - you can build them in a variety of ways. It is easy to construct a NN in software using a traditional V-N machine with a traditional MPU setup. The software then simply creates a switched matrix using standard tree nodes, and assigns probability algorithms to each nexus. Alternatively you can build a NN in silicon using artificial neurones and logically controlled bus pathways between the neurones.
In both cases the system 'learns' and is not predetermined. In other words your basic assumptions are completely wrong.
| ocalhoun wrote: | | My point though, is if it is indistinguishable from a human, then it will pass the human intelligence test, no matter how you measure it. |
But you still need the test. ^_^; Which is exactly the problem we're looking to solve here.
| ocalhoun wrote: | | (I'm taking indistinguishable to mean that the human and the machine are basically the same, no matter what measurement we use.) |
Aha, see this an excellent example of what i have been trying to explain. "... no matter what measurement we use..."? Really? So if i weighed a human and a clever machine and they weighed the same, then the machine is intelligent?
No? Oh. Well... then... what measurement should we be taking in order to compare their intelligences?
And BAM, there's your problem.
So let's say you decide to throw every measurement at it - rather than trying to narrow down what you should be measuring. Would that solve your problem? Nope.
See, for any finite number of measurements, you can never be sure that the measurement(s) you need isn't in that group of measurements you didn't take.
So let's get crazy and just think theoretical. What if you could take ALL measurements of both the human and the device - that is, what if you were able to have perfect knowledge to infinity of both. Would that be enough? Fraid not.
See, first, that presupposes that it's possible to measure intelligence... which, unless you know how to measure intelligence, you can never be sure is true. Having taken all ALL measurements won't help you if there are no measurements that can give you the information you want.
Furthermore, even if you assume that it is possible to measure intelligence, you're still stuck, because you have a discrimination problem. Comparing two items is not simply a matter of knowing which data to match... it is also knowing which data to throw away. If i just measured EVERYTHING about both a human and a machine, and they compare perfectly in every measurement except weight... does that mean that since they're not similar the machine isn't intelligent? Which data do i need, and which data do i ignore?
Basically, it boils down to this. If your suggestion is sound, then it teaches us only two things: 1) that humans are intelligent (because they're the comparison standard), and 2) that it is possible to measure and compare intelligence. Thing is, both of those things have been assumed since the start. So in other words... you've gotten us nowhere. ^_^; Indeed, the idea follows directly logically from the initial problem: humans are intelligent, and we assume there is a way to measure intelligence, so if a machine scores the same as a human on an intelligence test it must be intelligent. Well, great, but how do we measure intelligence?
| ocalhoun wrote: | | (Though I also assumed that our intelligent machine was not given the benefit of extra time to compute responses... Because if it were, then you could tell the difference between it and a human by just asking questions too rapidly for the machine to keep up.) |
Would that mean then that the machine isn't intelligent? Does speed matter?
| ocalhoun wrote: | A computer that is simulating a different computer must -- at some level -- execute the code that the simulated computer runs (or at least a more efficient version of it)...
Similarly, I think that if you asked an AI a question that required intelligence to understand and answer, somewhere within the software, something resembling intelligence would need to solve the problem, before forwarding the answer to the part that formats the answer in a human-like fashion. |
If there were such a question - a question that REQUIRES intelligence to answer, and cannot be generated even by a good simulation of intelligence - what would it look like? Is such a question even theoretically possible?
| The-Nisk wrote: | This might be irrelevant to the thread, but programs cannot become artificially intelligent. Artificial intelligence does not exist, not in the sense of intelligence comparable to a dog or a cat, certainly not human.
I've had this debate before with my computer science lecturer, it's a very tedious one.
What does exist that's dubbed as Artificial Intelligence is programmed intelligence, which isn't any different to any other program in the sense that it was designed by someone to perform a certain task, in doing so it's not any more intelligent than your calculator.
,,, |
Yes, we all know that true artificial intelligence - or artificial general intelligence - does not exist yet, and what we call AI today are just limited approximations that are not really intelligent (or that only appear intelligent in limited situations).
But to say that artificial general intelligence is ridiculous - on par with belief in deities - is... frankly... ridiculous. We know how the neurons in our brain work, and we can create physical analogues quite easily - or we can create logical analogues in software. We have done both of these things. All we have to do now is create enough of these artificial neurons (let's say, the same amount in a human brain), and wire them up correctly (let's say, the same way that they are wired in a human brain), and we'll have a complete, artificial human brain. And if our intelligence comes from your brain... well, job done. We'll have an artificial intelligence. Exactly what point in that description do we jump from reasonable to religiously absurd?
The only reason we don't have a strong AI today is because our current computers are tens of thousands of times (or more, depending on what measure you use, possibly millions of times) too simplistic to model a human brain. That situation CANNOT last. It is foolish to think that it will given the pace of modern technology. You can debate when it will happen - estimates by people like Kurzweil might be a bit unrealistic - but you cannot reasonably argue that it will never happen. It can, it will, and, frankly, it must... the only question is when.
Of course, all of those people who insist that it won't be for a long time because modelling the human brain is so very hard are missing a very obvious point. We may not have to model the human brain to create intelligence. Most of our biological systems are needlessly complex and highly redundant... why assume our brain is different? Plus, we certainly don't need to model the parts of the brain that deal with the nervous system and such. In other words, we may be able to create a strong, complete, general artificial intelligence decades before we are capable of creating a model of the human brain. Again, at exactly what point in all of that do we jump off the reason train?
Yes, general artificial intelligence or AI-complete intelligence don't yet exist. But they could, and it is not completely outside the realm of reason that they could exist quite soon. All that would need to be true for that to happen is two things: first, that it is true that the human brain is far more complex than it needs to be to support intelligence, and second, that someone figures out a way to simplify the problem to a scale that contemporary hardware can handle.
That's why this question is important. The Turing test is a lousy test for intelligence - most children would fail it, and even very intelligent adults would fail it if their language skills in the testing language are weak. Surely we can do better.
| Indi wrote: | | ocalhoun wrote: | | My point though, is if it is indistinguishable from a human, then it will pass the human intelligence test, no matter how you measure it. |
But you still need the test. ^_^; Which is exactly the problem we're looking to solve here.
|
And, basically, my answer is that the Turing test, administered by a variety of judges, both educated and ignorant, is the best theoretically possible test for machine intelligence.
(At least for machines designed to mimic humans, that is.)
| Quote: |
| ocalhoun wrote: | | (I'm taking indistinguishable to mean that the human and the machine are basically the same, no matter what measurement we use.) |
Aha, see this an excellent example of what i have been trying to explain. "... no matter what measurement we use..."? Really? So if i weighed a human and a clever machine and they weighed the same, then the machine is intelligent?
|
I thought we were measuring the intelligence of a chatbot, with the chat window being our only way to interact with it.
If we're allowed the full arsenal of scientific tests, then it must for all practical purposes be a human, in order to be indistinguishable from one.
| Quote: |
you've gotten us nowhere. ^_^; |
Quite likely. Just trying to make the point that it likely isn't possible to get anywhere.
| Quote: | so if a machine scores the same as a human on an intelligence test it must be intelligent. Well, great, but how do we measure intelligence?
|
It doesn't matter, because no matter what type of test it is, the test result will be indistinguishable from a human's test result.
(Since we've made the presumption that the machine is indistinguishable from a human.)
If you somehow think of a test that shows a difference between the machine and the human, we've invalidated the beginning assumption that they are indistinguishable.
(And if no test can show any difference, what use is there in devising different tests?)
| Quote: |
| ocalhoun wrote: | | (Though I also assumed that our intelligent machine was not given the benefit of extra time to compute responses... Because if it were, then you could tell the difference between it and a human by just asking questions too rapidly for the machine to keep up.) |
Would that mean then that the machine isn't intelligent? Does speed matter?
|
Speed only matters in that the lack of it could be used to distinguish between the machine and the human.
There's no theoretical reason you couldn't have an intelligence that functions in geologic time, taking thousands of years to complete a single thought.
Such an intelligence would be easy to distinguish from a human though.
| Quote: |
Is such a question even theoretically possible?
|
I'm not sure... Any specific question can just be specifically programmed for without using any intelligence; it would need to be unpredictable at some level...
| Quote: |
That's why this question is important. The Turing test is a lousy test for intelligence - most children would fail it, |
Before they reach a certain level of development, I'd regard that as a valid result, that they're not intelligent.
| Quote: | | and even very intelligent adults would fail it if their language skills in the testing language are weak. Surely we can do better. |
Well, offer the test subject a choice in languages, if needed.
A truly magnificent AI simulation could also simulate human language difficulty though.
| ocalhoun wrote: |
| Quote: |
That's why this question is important. The Turing test is a lousy test for intelligence - most children would fail it, |
Before they reach a certain level of development, I'd regard that as a valid result, that they're not intelligent. |
Why are children before a certain age not intelligent?
Intelligence is the ability to learn about, learn from, understand, and interact with one’s environment. It consists of a number of specific abilities, which include these specific abilities.
It is an underlying ability which enables an individual to adapt to and function effectively within a given environment. i can judge intelligence to be the capacity of an individual which helps him/her to be able to understand any condition and make possible adjustment.
I induce that comes from http://giftedkids.about.com/od/glossary/g/intelligence.htm
The problem with quoting as you have is that you left off a very important part of that second sentence by putting a full stop where you did, instead of listing the specific abilities...that means the sentence becomes meaningless and confusing.
| catscratches wrote: | | ocalhoun wrote: |
| Quote: |
That's why this question is important. The Turing test is a lousy test for intelligence - most children would fail it, |
Before they reach a certain level of development, I'd regard that as a valid result, that they're not intelligent. |
Why are children before a certain age not intelligent? |
Because they're not fully developed yet, of course.
It develops gradually, of course, so you can't put any clear differentiation point though.
(I think we can all agree that a newborn is not intelligent, while a teenager is... For some part of the continuum between, the child is not intelligent.)
Also,
@The original topic of the discussion:
Before you can test for intelligence, you must first define intelligence.
Once clearly defined, it would be far easier to test for.
| ocalhoun wrote: | | Indi wrote: | | ocalhoun wrote: | | My point though, is if it is indistinguishable from a human, then it will pass the human intelligence test, no matter how you measure it. |
But you still need the test. ^_^; Which is exactly the problem we're looking to solve here.
|
And, basically, my answer is that the Turing test, administered by a variety of judges, both educated and ignorant, is the best theoretically possible test for machine intelligence.
(At least for machines designed to mimic humans, that is.)
| Quote: |
| ocalhoun wrote: | | (I'm taking indistinguishable to mean that the human and the machine are basically the same, no matter what measurement we use.) |
Aha, see this an excellent example of what i have been trying to explain. "... no matter what measurement we use..."? Really? So if i weighed a human and a clever machine and they weighed the same, then the machine is intelligent?
|
I thought we were measuring the intelligence of a chatbot, with the chat window being our only way to interact with it.
If we're allowed the full arsenal of scientific tests, then it must for all practical purposes be a human, in order to be indistinguishable from one. |
Best theoretically possible? That is a huge claim.
But even setting that aside, who says we are only interested in looking for human intelligences? And even if we are, the Turing test isn't all that effective even in that case.
Again, you're off track. We're not looking for humans, we're looking for intelligence. Your test, therefore, is useless. (And even if we were looking for humans, your test is useless, because a DNA test could find a human far easier than a Turing test.)
| ocalhoun wrote: | | Quote: |
you've gotten us nowhere. ^_^; |
Quite likely. Just trying to make the point that it likely isn't possible to get anywhere. |
Yeah, see, you could have made that point far more effectively, and wasted less of the thread's time, if you had just said, "We're never going to get anywhere here." Would have done just about as much good for the discussion, too.
| ocalhoun wrote: | | Quote: | so if a machine scores the same as a human on an intelligence test it must be intelligent. Well, great, but how do we measure intelligence?
|
It doesn't matter, because no matter what type of test it is, the test result will be indistinguishable from a human's test result.
(Since we've made the presumption that the machine is indistinguishable from a human.)
If you somehow think of a test that shows a difference between the machine and the human, we've invalidated the beginning assumption that they are indistinguishable.
(And if no test can show any difference, what use is there in devising different tests?) |
It does matter, because even if we were looking for "human-like intelligences" (which, i can't see why we would restrict ourselves like that, but let's roll with that), without knowing how to measure intelligence we can't be sure our test result is indistinguishable from a human result... because we can't distinguish unless we can measure... because all "distinguishing" means is being unable, when measuring two things, to assert that the two measurements differ. Two things are "indistinguishable" in a certain metric if the difference between them in that metric is below the threshold of our measuring device.
For example, if we have two massive objects and a scale that can measure to ±0.00001 g, and we weigh one and it is 10.12345 while the other is 10.12345... then the two objects are indistinguishable by weight as far as we're concerned. If one is actually 10.123451 while the other is 10.123454, then they are different, but we cannot distinguish that difference. If we had a scale that could measure that extra decimal place, then they would no longer be indistinguishable.
That's what indistinguishable means. It doesn't mean: "Well,... we'll sorta... do some sciencey stuff to them... run some... 'tests'... i guess... i dunno, sciencey tests... and if we can't figure out a way to tell them apart using our sciencey tests... well, then we'll say their mass is indistinguishable. Hopefully, one of our sciencey tests measured mass, somehow." That's about as scientific as ghost hunting, and about as useful.
FURTHERMORE your whole thesis falls apart because the human standard is not a single value... even humans can be distinguished from each other in a Turing test! So what do we do? Say that our test intelligence has to be able to beat ANY human in a Turing test? Even children or the mentally handicapped? If that's the standard, my Casio calculator can be programmed to pass a Turing test, and is therefore intelligent. Okay, so do we say the test intelligence has to beat the AVERAGE human standard? If that's the case, then you are defining 50% of humans as unintelligent. What if we say that the test intelligence has to beat ALL humans? If that's the case then not only will we will fail intelligences that are far smarter than the average human... all humans will no longer be counted as intelligent.
Bottom line, dude, your test blows.
| ocalhoun wrote: | Speed only matters in that the lack of it could be used to distinguish between the machine and the human.
There's no theoretical reason you couldn't have an intelligence that functions in geologic time, taking thousands of years to complete a single thought.
Such an intelligence would be easy to distinguish from a human though. |
Again, you're not in tune with the problem. Your suggestions are completely ridiculous - you're not putting any real thought into them at all, and this is a philosophy forum: thought is the point. You talk about being unable to distinguish from a human as being the standard, when that could be accomplished simply visually. Then, you seem to have silently assumed that physique is not a factor in an intelligence test, but you haven't justified that assumption any more than you've justified anything else - and now you're saying that mental speed is a factor... or isn't it... you really don't know. On the one hand you say "sure, an intelligence could function in geologic time", on the other you say that since that would make it distinguishable from a human it would fail your putative intelligence test... which, unless your intelligence test is a complete farce, would mean that it wouldn't be intelligence. You're completely confused about where you are and what you're trying to say. Do you really believe that comparing a prospective intelligence to a human is a good way to measure intelligence? Do you not believe that intelligences could be smarter or dumber than humans in such a way as to make them easily distinguishable... yet still be intelligent? Do you not believe that even intelligences on an equal level could think so differently that they would be unable to pass for human... yet still be intelligent? If you answered no to any of these questions... then what is the point of anything you've been saying?
| ocalhoun wrote: | | Quote: |
Is such a question even theoretically possible?
|
I'm not sure... Any specific question can just be specifically programmed for without using any intelligence; it would need to be unpredictable at some level... |
It wouldn't take much effort to build up a database of all the possible questions a huge number of people might ask, and the expected answers. The set of questions a given single person might ask will very likely be a subset of that set. In other words, it would not be a big deal to create a big-ass database with a simple query function that could pass your test. (Whether it would be fast enough to maintain the illusion of human speed is another issue, but that's just a hardware problem, not a design issue.)
| ocalhoun wrote: | | Quote: |
That's why this question is important. The Turing test is a lousy test for intelligence - most children would fail it, |
Before they reach a certain level of development, I'd regard that as a valid result, that they're not intelligent. |
That's a dangerous Pandora's box to open. i really wouldn't go down that road until you've given it a lot more thought.
| ocalhoun wrote: | | Quote: | | and even very intelligent adults would fail it if their language skills in the testing language are weak. Surely we can do better. |
Well, offer the test subject a choice in languages, if needed.
A truly magnificent AI simulation could also simulate human language difficulty though. |
Wonderful, but that doesn't solve the issue of people with poor language skills. What if they have no fluent language? Are they no longer intelligent - because they'd fail your test, after all. What if they have a handicap that prevents fluency (by the way, there's a hint of what awaits you in that Pandora's box i mentioned above)? What if you sustained a head injury against your linguistic centre and end up with dysphasia- you are perfectly capable of thinking, but can no longer translate those thoughts into words?
| ocalhoun wrote: | Before you can test for intelligence, you must first define intelligence.
Once clearly defined, it would be far easier to test for. |
No, not really. Certainly not completely. They were measuring the variation of gravity centuries before anyone even suggested the phrase "curvature of space time", and Galileo was measuring gravity around a century and a half before Newton (before Galileo even had a word for gravity!). You don't need to define something before you can discern it, you just need to be able to discern it.
Sometimes it is enough to know what something is not, without knowing what it is. You know a being's mass has no relation to its intelligence. That's a start. So let's keep chipping away and see what's left.
Sometimes it's even possible to "define" something by a measurement, with no understanding of what it really is. For example, the way temperature was handled before the dawn of statistical mechanics, where "temperature" was just defined as the difference between an object when it was hot, and when it was cold.
There is nothing yet that implies the problem is insoluble. It might be, but there's no reason to believe that it is yet. We even have multiple ways we can go about attacking it. You seem to have given up on it before even taking a serious stab at it.
| Indi wrote: |
Best theoretically possible? That is a huge claim.
|
For human-mimicking intelligences, yes.
| Quote: |
But even setting that aside, who says we are only interested in looking for human intelligences? And even if we are, the Turing test isn't all that effective even in that case.
Again, you're off track. We're not looking for humans, we're looking for intelligence. Your test, therefore, is useless. (And even if we were looking for humans, your test is useless, because a DNA test could find a human far easier than a Turing test.)
|
Ah, so I suppose the original post has just gotten me completely off-track... I thought the discussion was about human-mimicking intelligences from the get-go.
If we want to determine any sort of intelligence, that is indeed a much more difficult task.
| Quote: |
"We're never going to get anywhere here." Would have done just about as much good for the discussion, too.
|
I try to avoid making random statements without any supporting argument though.
| Quote: |
Bottom line, dude, your test blows.
|
For testing intelligence that does not attempt to mimic human intelligence, yes it does.
I didn't think that was what we were talking about though.
| Quote: |
It wouldn't take much effort to build up a database of all the possible questions a huge number of people might ask, and the expected answers.
|
I think that would take considerable effort, actually.
A- There are a LOT of possible combinations that could make up likely questions.
B- Context counts, the response to one question should influence the response to the next, adding another layer of complexity. (The question "How long do you think that will take?" for example, would need a wide variety of different answers depending on the previous conversation.)
| Quote: |
That's a dangerous Pandora's box to open. i really wouldn't go down that road until you've given it a lot more thought.
|
When have I ever shied away from such things?
| Quote: |
What if you sustained a head injury against your linguistic centre and end up with dysphasia- you are perfectly capable of thinking, but can no longer translate those thoughts into words?
|
Then I would fail a Turing test, perhaps validly so, since I would no longer be able to mimic human intelligence.
| Quote: |
| ocalhoun wrote: | Before you can test for intelligence, you must first define intelligence.
Once clearly defined, it would be far easier to test for. |
No, not really. Certainly not completely. They were measuring the variation of gravity centuries before anyone even suggested the phrase "curvature of space time", and Galileo was measuring gravity around a century and a half before Newton (before Galileo even had a word for gravity!). You don't need to define something before you can discern it, you just need to be able to discern it.
|
You do need to define something before you can measure it though.
Even Galileo's experiments in measuring gravity would not turn out well if he didn't at least crudely define what he was measuring.
(Even if that happens to be 'the speed at which this ball falls to the ground when I drop it'.)
If you define gravity as nebulously as intelligence is usually defined, then it would also be extremely difficult to test for.
| Quote: |
Sometimes it's even possible to "define" something by a measurement, with no understanding of what it really is. For example, the way temperature was handled before the dawn of statistical mechanics, where "temperature" was just defined as the difference between an object when it was hot, and when it was cold.
|
So, pick an arbitrary intelligence test, and say that defines what intelligence is.
Since the test defines what intelligence is, the test will be perfect, never having any false results.
The problem with this approach is that the test's results may not agree with our fuzzily-determined ideas of what is intelligent. In order to correct that problem, we must codify and strictly define those ideas of what is intelligent. Once that is done, we can easily design a test that agrees closely with those ideas.
(Or, just throw away the ill-defined ideas of intelligence, and accept the test result as the only measure.)
| Quote: |
You seem to have given up on it before even taking a serious stab at it. |
For testing human-mimicking systems for intelligence, nothing beats a Turing test.
For testing other types of intelligence, there is a multi-step process that must be followed to achieve a satisfactory result:
1- Define intelligence
2- Design a test to test for elements defined in step 1
3- Apply test to a variety of systems
4- Determine if the test results match the notions of what is intelligent or not, as defined in step 1.
You simply can't do step 2 without doing step 1 first.
| Quote: | You do need to define something before you can measure it though.
Even Galileo's experiments in measuring gravity would not turn out well if he didn't at least crudely define what he was measuring.
(Even if that happens to be 'the speed at which this ball falls to the ground when I drop it'.)
If you define gravity as nebulously as intelligence is usually defined, then it would also be extremely difficult to test for. |
No you don't.
Here's a scenario :
I make a stick of an alloy of tin, bismuth and steel. I hook it up to a scope and I observe that I'm picking up what appears to be EM radiation. Naturally I assume it is em and my stick is acting as an antenna...what else? but to be sure I stick it in a Faraday cage and, blow me, I still get a reading.
I am now measuring something which I have no clue about. I can measure it very accurately but I don't know what I'm measuring and whence it originates....
| Bikerman wrote: | | I can measure it very accurately but I don't know what I'm measuring and whence it originates.... |
You don't know where it originates, but you do know what you're measuring: current and/or voltage as displayed on your scope.
Now, we can make up a word for the radiation ('tibistation' anyone?), but for anybody to know what we're talking about, we need to define this word.
We could define it as 'the output of the scope when connected to our special stick,' and that works fine to begin with.
Since we now have a well defined quality to test for, we can test other objects for tibistation, using the same standard.
If 'tibistation' is defined as 'appearing to output EM radiation both inside and outside of a Faraday cage', then it is simple to devise a test for this quality: Just connect your scope to various objects both inside and outside of Faraday cages... Then it will be easy to determine what objects display tibistation and which do not.
That is different. You are talking about communicating, not measuring. The fact remains that you can most certainly measure something with no knowledge of what it is that you are measuring...
I am indeed measuring potential difference and current flow but that is simply a proxy, and I know it is caused by something, but I don't know what.
| Bikerman wrote: | That is different. You are talking about communicating, not measuring. The fact remains that you can most certainly measure something with no knowledge of what it is that you are measuring...
I am indeed measuring potential difference and current flow but that is simply a proxy, and I know it is caused by something, but I don't know what. |
Of course, but suppose:
A- You detected current, but no measurable voltage.
B- You told another scientist so he could duplicate your results, but you only said 'I detected a signal'.
If that other scientist recreates the experiment, but measures voltage, he'll detect no signal, and falsely believe that the experiment failed.
Why is this? Because the word 'signal' was poorly defined, making it much more difficult to test for said signal.
Before you can effectively test for the signal, you must first define what the signal is.
In much the same way, suppose you conducted an experiment testing some subject for intelligence. You find intelligence unexpectedly, and tell another scientist so he can duplicate the experiment. However, if you don't define intelligence as you were testing for it, the other scientist may use a very different test for intelligence and falsely determine that the same subject is not intelligent.
| ocalhoun wrote: | | Bikerman wrote: | That is different. You are talking about communicating, not measuring. The fact remains that you can most certainly measure something with no knowledge of what it is that you are measuring...
I am indeed measuring potential difference and current flow but that is simply a proxy, and I know it is caused by something, but I don't know what. |
Of course, but suppose:
A- You detected current, but no measurable voltage.
B- You told another scientist so he could duplicate your results, but you only said 'I detected a signal'.
If that other scientist recreates the experiment, but measures voltage, he'll detect no signal, and falsely believe that the experiment failed. |
You are struggling here. Why would I neglect to tell him to measure current?
I don't need to know anything about the thing being measured in order to give a highly detailed description of the apparatus.
| Quote: | | In much the same way, suppose you conducted an experiment testing some subject for intelligence. You find intelligence unexpectedly, and tell another scientist so he can duplicate the experiment. However, if you don't define intelligence as you were testing for it, the other scientist may use a very different test for intelligence and falsely determine that the same subject is not intelligent. |
The same applies - I only need to describe the experiment, not the quantity being measured.
| Bikerman wrote: |
I don't need to know anything about the thing being measured in order to give a highly detailed description of the apparatus.
|
In the analogy, the vagueness of your description of the experiment is an analogue to the vagueness inherent in the word 'intelligence'.
...Though I do suppose the analogy may have been weakly built.
To save time, I've just pulled a few applicable definitions of intelligence from the internet:
| Quote: |
1-the ability to comprehend; to understand and profit from experience
2-Intelligence is an umbrella term describing a property of the mind comprehending related abilities, such as the capacities for abstract thought, reasoning, planning and problem solving, the use of language, and to learn.
3-Capacity of mind, especially to understand principles, truths, facts or meanings, acquire knowledge, and apply it to practice; the ability to learn and comprehend; An entity that has such capacities;
4-an underlying ability which enables an individual to adapt to and function effectively within a given environment. |
Contained within these definitions are various attributes that when combined constitute intelligence.
Since intelligence is (apparently) so multi-faceted, any test for it must also be multi-faceted.
A battery of individual tests would need to be conducted, such as a test for the ability to learn from past experience, and a test for the ability to plan ahead.
In the end though, your tests can only be as accurate as your definition.
If the definition being used disagrees with your notion of what intelligence is, then the results of the tests will also disagree with your notion of what is intelligent or not.
| ocalhoun wrote: | | Quote: |
1-the ability to comprehend; to understand and profit from experience
2-Intelligence is an umbrella term describing a property of the mind comprehending related abilities, such as the capacities for abstract thought, reasoning, planning and problem solving, the use of language, and to learn.
3-Capacity of mind, especially to understand principles, truths, facts or meanings, acquire knowledge, and apply it to practice; the ability to learn and comprehend; An entity that has such capacities;
4-an underlying ability which enables an individual to adapt to and function effectively within a given environment. |
Contained within these definitions are various attributes that when combined constitute intelligence.
Since intelligence is (apparently) so multi-faceted, any test for it must also be multi-faceted.
A battery of individual tests would need to be conducted, such as a test for the ability to learn from past experience, and a test for the ability to plan ahead.
In the end though, your tests can only be as accurate as your definition.
If the definition being used disagrees with your notion of what intelligence is, then the results of the tests will also disagree with your notion of what is intelligent or not. |
I like some of the above, though as you say, defining intelligence in a nutshell is not easy. Even the above can be interpreted in so many different ways.
1. A squirrel learning to store nuts can understand the need to store food for the winter and profit from taking a larger harvest to supplement through the difficult months - that is a form of intelligence for squirrels, but is it showing intelligence of an individual squirrel? Sometime, somehow, there might have been an Albert Nutstein who figured it out, so at best it would show that each surviving squirrel can learn from other squirrels.
2. Dogs are certainly capable of dreaming which is a kind of abstract thought process. When people read a good story book, we can hallucinate the scenes quite vividly (though not necessarily what the author was seeing). Many birds and mammals show signs of problem solving, language and can certainly learn through training. To me at least the question for any judgement relating to 'higher' intelligence would be whether an entity could consciously invoke abstract thoughts and learn something with practical use from those abstract thoughts. Examples of this would be an engineer 'building' a machine in the mind with such efficiency that it's very near or a perfect representation of a model that will work in practicality - alternatively a musical composer who can 'create' an orchestra or band and produce something new and highly entertaining.
4. The ability to adapt to new environments is something with great interest when judging intelligence. The ability to 'think on your feet' so to speak can be used to survive and/or prosper where others will fail.
Personally I believe that intelligence can be grown and it's theoretically possible to create a machine that can meet all aspects of intelligence, the danger of course is that creating a machine with the ability to develop real intelligence is that it would also develop a 'survival and prospering' mentality and that could conflict with our own and result in some kind of Frankenstein's monster scenario. Really we should question whether creating something with the potential to be equally and possibly more intelligent than our own race is in fact an intelligent thing to do - this especially when the machine may develop the ability to clone itself with a complete install of the knowledge etc that we as biological entities take years to gain.
| ocalhoun wrote: | In the end though, your tests can only be as accurate as your definition.
If the definition being used disagrees with your notion of what intelligence is, then the results of the tests will also disagree with your notion of what is intelligent or not. |
Clearly.This problem is highlighted with reference to animals. We have no satisfactory measure of intelligence in animals. There are crude measures - brain mass/body mass for example *
The general question, however, is not quantitative - we are not asked to measure the quantity of intelligence, just to test whether it exists or not.
*and various modified versions which are more sophisticated but still rely on mass.
| ocalhoun wrote: | | Indi wrote: |
Best theoretically possible? That is a huge claim.
|
For human-mimicking intelligences, yes. |
Even then, no. Because, as i've pointed out, the Turing test doesn't test intelligence. Surely if you had dysphasia you still have your intelligence, but there's no way you could pass a Turing test. What about someone raised by wolves who didn't have any human language - are they unintelligent, or just unable to communicate? i would say the latter, but if the Turing test is a test of intelligence then that person would fail.
Clearly something's wrong with calling the Turing test a test for intelligence. And mere fact that i can judge intelligence intuitively better than a Turing test shows that the Turing test cannot be the best theoretically possible. Clearly we can do better. The only question is how.
| ocalhoun wrote: | | Ah, so I suppose the original post has just gotten me completely off-track... I thought the discussion was about human-mimicking intelligences from the get-go. |
i don't know where you got that idea from. i explicitly suggested that even if they made a chatbot that could pass for a human, it might still not be intelligent: let’s assume they actually do manage to make a chatbot that is capable of passing as a human in conversation so well that it scores 50% on Turing tests (in other words, the judges are just guessing 50-50 which is the human and which is the chatbot). Would you call that chatbot “intelligent”? i even took the time to point out that this would have to be a test that humans could pass, too, so it seems a little bizarre to assume that finding humans or human-like intelligences was the goal.
| ocalhoun wrote: | | If we want to determine any sort of intelligence, that is indeed a much more difficult task. |
Not necessarily. If we can test for intelligence, it would take more effort to identify an intelligence as specifically human-like. The problem is that you are still labouring under the assumption that the Turing test is a test for intelligence. It's not. It may require intelligence to pass, but that doesn't make it a test for intelligence, just like the "touch your nose with your fingers" sobriety test requires hands to work, but doesn't test for the existence of hands (because you can have hands and pass or fail the sobriety test, and you can be sober or drunk with or without hands).
| ocalhoun wrote: | | Quote: |
"We're never going to get anywhere here." Would have done just about as much good for the discussion, too.
|
I try to avoid making random statements without any supporting argument though. |
Then make a statement and back it up with an argument. Don't argue aimlessly for several posts while pretending to be supporting a completely different claim, then suddenly state that your thesis all along has been something different.
| ocalhoun wrote: | | Quote: |
It wouldn't take much effort to build up a database of all the possible questions a huge number of people might ask, and the expected answers.
|
I think that would take considerable effort, actually.
A- There are a LOT of possible combinations that could make up likely questions.
B- Context counts, the response to one question should influence the response to the next, adding another layer of complexity. (The question "How long do you think that will take?" for example, would need a wide variety of different answers depending on the previous conversation.) |
That's all that chatbots are (actually, chatbots even more complicated than that!), and - again to the example in the first post - they can almost pass Turing tests. So... back to the original question... if they could pass the test, would they be intelligent?
| ocalhoun wrote: | | Quote: |
What if you sustained a head injury against your linguistic centre and end up with dysphasia- you are perfectly capable of thinking, but can no longer translate those thoughts into words?
|
Then I would fail a Turing test, perhaps validly so, since I would no longer be able to mimic human intelligence. |
No, that's false logic. You still have all the "human" intelligence you had before. No intelligence was lost. You just lost the ability to communicate like a human. Therefore, the Turing test is not testing for human-like intelligence, it is testing for human-like communication.
And it's easy to prove that, too. Suppose you were on one side of the blind in a Turing test, and you had this conversation:- You: What is the meaning of life?
- Subject: My life is meaningful and not think only one.
- You: What does that mean?
- Subject: It means believing that I question is meaningless.
- You: Does that mean that you think it is meaningless to question anything.
- Subject: No, I do believe is that all that is asked.
- You: So belief is what matters?
- Subject: Conviction, must be based on reason and evidence, it is meaningless.
- You: Conviction is meaningless?
- Subject: No, but support the evidence given by reason and conviction must be proportional.
Well? Are you talking to an intelligence? By a standard Turing test, this would pretty clearly be a fail. (And if you don't believe this is a fail, believe me when i say that i can make the responses even more convoluted by using more complex Japanese grammar... or using a less mature translater than Google's 日本語->English mode.)
However... all that happened was that i wrote my answers in Japanese and passed them through Google translate. This is what the conversation looked like from my side:- You: What is the meaning of life?
- Subject: I don't believe that life has only one meaning.
- You: What does that mean?
- Subject: It means that I believe that the question is meaningless.
- You: Does that mean that you think it is meaningless to question anything.
- Subject: No, I believe that everything should be questioned.
- You: So belief is what matters?
- Subject: Belief must be based on reason or evidence, or it is meaningless.
- You: Conviction is meaningless?
- Subject: No, but conviction should be proportional to support given by reason and evidence.
See? It's not intelligence that you are testing in a Turing test, it is the communication skills (in this case, the ability of the translater to translate my responses).
That's the bottom line. The Turing test is not a test for intelligence. Never was, never will be. It can't even reliably detect human intelligence, let alone "human-like" intelligence. Our intuition is better. Therefore, there is no way it can be the best theoretically possible test.
| ocalhoun wrote: | You do need to define something before you can measure it though.
Even Galileo's experiments in measuring gravity would not turn out well if he didn't at least crudely define what he was measuring.
(Even if that happens to be 'the speed at which this ball falls to the ground when I drop it'.)
If you define gravity as nebulously as intelligence is usually defined, then it would also be extremely difficult to test for. |
No, you don't. Galileo didn't crudely define anything (nobody defined gravity in any way until Newton came around). He just observed that different masses fell at the same rates. He never said "something causes that to happen" (which would be identifying gravity, the very first stage of defining it), he just observed the phenomenon and measured it. Galileo did not define anything as "the speed at which this ball falls to the ground when I drop it"... he just measured. Galileo did even less toward defining gravity than we've already done for defining intelligence... yet he was measuring it, and even making qualitative observations about it.
Or, take the electricity example. People were identifying the presence of electricity since the time of ancient Greeks (and even before). They were even using it to do stuff (albeit stupid stuff like shocking people to cure them from diseases). They had no idea of voltage or current, although they were aware that some materials conducted and others didn't. They didn't even have a name for the phenomenon. But, they did know that some materials could be "charged" by rubbing them, and others couldn't.
There's your analogy right there. We don't need to know what static electricity is, nor do we even need to presume it exists, in order to identify which materials can shock after rubbing. Same goes for intelligence - we don't need to know what it is, or even presume it exists, in order to recognize that we can do some things that others animals and machines can't, that we know of no other cause for. However, we have already taken it upon ourselves to hypothesize a "thing" that gives us those abilities, and give that "thing" a name: intelligence.
We don't need to do that, though. We can just observe that humans have capabilities that animals and machines, don't. We can categorize those abilities and identify which abilities are caused by known traits (for example, our ability to speak is caused by our language and communication faculties), and whatever remains (for example, our language and communication faculties allow us to share and understand ideas, but where do the ideas come from and how do we understand them), we can test for. All of that without defining intelligence.
THEN we can define intelligence as the sum of those traits, or a subset of them, or something else, and with that definition, we could improve and refine our testing procedure. But we don't need to do that to start creating a test now... and if we do that, and create a reasonable test, we can call whoever passes that test intelligent as a provisional definition.
| ocalhoun wrote: | | So, pick an arbitrary intelligence test, and say that defines what intelligence is. |
Sure, so all we need is an intelligence test. Again, back to square one.
| ocalhoun wrote: | Since the test defines what intelligence is, the test will be perfect, never having any false results.
The problem with this approach is that the test's results may not agree with our fuzzily-determined ideas of what is intelligent. In order to correct that problem, we must codify and strictly define those ideas of what is intelligent. Once that is done, we can easily design a test that agrees closely with those ideas.
(Or, just throw away the ill-defined ideas of intelligence, and accept the test result as the only measure.) |
That's a rather strange problem you are creating... and there's no reason to create it. Why would you have to be so anal-retentive about absolutely defining intelligence by the test? There's no need for that.
If we have a provisional test there's no reason we can't recognize that it is provisional, and that sometimes it will screw up. In fact, that's the logical thing to do, because then we can analyze how and when it fails, and refine it.
| ocalhoun wrote: | | For testing human-mimicking systems for intelligence, nothing beats a Turing test. |
Not even close to true. ^_^; Already it has been suggested that we could look at the guts of the system to see how it mimics intelligence, and use that to determine whether it is mimicking intelligence, or actually is intelligence. Even that vaguely defined test is much more accurate than the Turing test, because it would correctly identify cheating, and communication impairment.
The average human 10 year-old is probably a better test for intelligence than the Turing test. They can identify the difference between machine "intelligence" and real intelligence well enough to recognize things that are intelligent versus things that have been cleverly programmed.
| ocalhoun wrote: | For testing other types of intelligence, there is a multi-step process that must be followed to achieve a satisfactory result:
1- Define intelligence
2- Design a test to test for elements defined in step 1
3- Apply test to a variety of systems
4- Determine if the test results match the notions of what is intelligent or not, as defined in step 1.
You simply can't do step 2 without doing step 1 first. |
Again, not true. In fact, you procedure could look like this:
- Start with a random test.
- Apply the test to a variety of subjects.
- Compare the results to our fuzzy, undefined notion of intelligence.
- If they are identical, congratulations, you have a provisional test for intelligence. You're done.
- If they are not, make adjustments to the test to correct for errors.
- Go back to 2.
See? No need for definition. In fact, in the procedure you described, step 1 is superfluous because of what you do in step 4.
This is how most early scientific work is done, too. Before properties like time, temperature, inertia, etc. were defined, we had a vague notion of what they were, and designed instruments to measure them, and then defined the properties by the measurement. (For time, for example, we went from "we need to know what part of the day it is", and defined all of our time measurements from that - seconds and hours were define as fractions of a day. Later we decided to define time as a constant property, so the time units should be absolute and not vary with the time of year... and still later our idea of its constancy was called into question. To this day we still don't have a proper definition for time, and still define time by the measurement... but the important thing is that they measured time for centuries without even considering that time was an actual "thing", let alone defining what that thing was.)
| Indi wrote: |
We don't need to do that, though. We can just observe that humans have capabilities that animals and machines, don't. We can categorize those abilities and identify which abilities are caused by known traits (for example, our ability to speak is caused by our language and communication faculties), and whatever remains (for example, our language and communication faculties allow us to share and understand ideas, but where do the ideas come from and how do we understand them), we can test for. All of that without defining intelligence.
|
And categorizing the traits that it is composed of is fundamentally different than 'defining' how?
| Quote: |
THEN we can define intelligence as the sum of those traits, or a subset of them, or something else, and with that definition, we could improve and refine our testing procedure. But we don't need to do that to start creating a test now... and if we do that, and create a reasonable test, we can call whoever passes that test intelligent as a provisional definition.
|
A bit of a semantic quibble, but if you're only testing for various traits, without defining those traits as intelligence, you're not testing for intelligence yet.
Isn't it only an intelligence test once you define what you're testing for as 'intelligence'?
| Quote: |
If we have a provisional test there's no reason we can't recognize that it is provisional, and that sometimes it will screw up. In fact, that's the logical thing to do, because then we can analyze how and when it fails, and refine it.
|
But how will you determine if it failed or not, much less how?
You need something to compare the results to.
Like it or not, the thing you're comparing them to is your definition of intelligence, even if it isn't precisely defined yet.
The vagueness of the definition will then make your comparison of test results erratic and subject to the differences of opinion held by different testers.
| Quote: |
This is how most early scientific work is done, too. Before properties like time, temperature, inertia, etc. were defined, we had a vague notion of what they were, and designed instruments to measure them, and then defined the properties by the measurement. [...] but the important thing is that they measured time for centuries without even considering that time was an actual "thing", let alone defining what that thing was.) |
So, just apply the same concept; define the property (intelligence) by the measurement, then refine the measurement by how it agrees with a vague definition.
-Just saying that the process will be smoother if you could compare the measurement with a precise definition.
| Bikerman wrote: |
The general question, however, is not quantitative - we are not asked to measure the quantity of intelligence, just to test whether it exists or not.
|
Of course.
Yet, it doesn't seem to be a black-and-white issue, there seems to be a grey area between 'intelligent' and 'not intelligent'.
Without a precisely defined point on that continuum where we can say 'below this is unintelligent, and above this is intelligent', we'll have to deal with quantitative measures... And deal with vagueness like 'somewhat intelligent,' 'slightly intelligent,' 'more intelligent than ____, but less than ___,' et cetera.
The question originally posed was "How would you judge intelligence?" It asks for opinion and as such all comments are valid.
Is a horse or a dog intelligent? How can we establish that without sufficient communication between the observer and the horse or dog? For all we know, the animal has superior intelligence to the observer. We can only judge those things which are communicated. Does a dog or horse adapt to training? - in various levels, yes so that fact may allow the observer to deduce that one horse or dog is of superior intelligence to the other, but how can we be sure? Maybe one horse or dog doesn't wish to conform to the training. Maybe one horse or dog sees the problems in conforming as being superior in importance than the benefits of conforming - after all, they may not want to be human-trained animals for whatever reasons of their own.
This is the difficulty in trying to assess intelligence. It relies on several factors. An ability for the subject to communicate that intelligence to the testers, the abilities to adapt sufficiently to make those communications and the willingness to show what intelligence there is. I certainly wouldn't show certain things to certain people as I believe they would harm themselves or others - does that mean I don't have intelligence in these areas? - no - does it mean they may presume that I have no intelligence in these areas? Possibly, and that may be preferred above proving my case. Why should others not do the same?
Let's say a dog knows everything there is to know about making a nuclear bomb, a car, a bicycle. it may decide that teaching or sharing this knowledge with us will work as a disadvantage since humans are perceived by this particular dog as dangerously stupid. This works in much the same way, that people may decide not to build an intelligent machine because it may disadvantage our human-ness in some way.
| jeffryjon wrote: |
Is a horse or a dog intelligent? How can we establish that without sufficient communication between the observer and the horse or dog? For all we know, the animal has superior intelligence to the observer. |
No... They often have more intelligence than we give them credit for, but they are not more intelligent than us.
Otherwise, answer me this:
My horse knows where her food is stored (in an open bin), and is EXTREMELY fond of eating.
Once, she got out of her fenced pasture.
Did she a) go raid the food bin, b) go visit other horses in nearby pastures, or c) run off into the road, and almost get run over?
Given that she chose 'c'... I'm guessing she's not super-intelligent.
| ocalhoun wrote: | | jeffryjon wrote: |
Is a horse or a dog intelligent? How can we establish that without sufficient communication between the observer and the horse or dog? For all we know, the animal has superior intelligence to the observer. |
No... They often have more intelligence than we give them credit for, but they are not more intelligent than us.
Otherwise, answer me this:
My horse knows where her food is stored (in an open bin), and is EXTREMELY fond of eating.
Once, she got out of her fenced pasture.
Did she a) go raid the food bin, b) go visit other horses in nearby pastures, or c) run off into the road, and almost get run over?
Given that she chose 'c'... I'm guessing she's not super-intelligent. |
Oh but Ocalhoun, couldn't it just be that your horse loves and respects you soooooooo much that it couldn't even consider making your life harder? Anyways, since ya want to reincarnate as a hoss, ya must think they're intelligent enough where it counts
Seriously though, the point I was making is that trying to assess intelligence properly is hindered by any lack of communication. If we look at Steven Hawkins as an example - would we have become aware of his intelligence had we not developed mechanisms to allow him and us to communicate fully. He, and I'm sure many others, could easily be missed if there's insufficient communication.
One way of interpreting the original question is "What methodologies would you use to try judging intelligence", which to me is the most important aspect related to the subject. That being the case for me at least, the first step would be ensure that we have a thorough and accurate means of communicating with the subject(s). In this endeavour, we'd also benefit from developing friendships with the subject(s), as any mistrust or fear on their parts could easily cause us to be misled if the subject(s) decide to play dumb.
| ocalhoun wrote: | | Indi wrote: |
We don't need to do that, though. We can just observe that humans have capabilities that animals and machines, don't. We can categorize those abilities and identify which abilities are caused by known traits (for example, our ability to speak is caused by our language and communication faculties), and whatever remains (for example, our language and communication faculties allow us to share and understand ideas, but where do the ideas come from and how do we understand them), we can test for. All of that without defining intelligence.
|
And categorizing the traits that it is composed of is fundamentally different than 'defining' how? |
The difference between identity and equality is fundamental in any logical field. For example, mathematics has two different operators to distinguish them: = and ≡.
Take π for example.
π ≡ c ÷ d
Where c is the circumference of a circle, and d is the diameter (with some extra technical caveats).
But:
π = Q q ÷ (4 F ε₀ d²)
Where Q and q are point charges, F is the force between them, d is the distance between them and ε₀ is the electric constant.
Now here's a thought experiment. What if we found that the ratio between the circumference and diameter of a circle was five times what we thought? How would these formulae change?
They would change to this:
π ≡ c ÷ d
π = Q q ÷ (20 F ε₀ d²)
Now what if we found that the charge produce was half what it should be for a given force? What would happen to the formulae then?
They would change to this:
π ≡ c ÷ d
π = Q q ÷ (2 F ε₀ d²)
You see? Definition is definition, and represents an identity relationship. When something is defined as something else, then the relationship between the two is an axiomatic fact, and can never be changed by anything but changing the definition itself.
Equality is a post hoc observation. When something is noted as equal to something else, that is interesting, but not binding.
Take Barack Obama for example. Obama is equal to the first black president of the US... but is not defined as the first black president of the US. "First black president of the US" is defined by one set of parameters, and Obama is defined as something else, but they just happen to be equal. If Obama had lost the election, neither of the two definitions would change, but the equality would be broken.
π has many characteristics that we can test for: it's about 3, it is less than 22/7, the 4th digit is 1, it is irrational, it is transcendental... none of those things define π, but they are things we can observe following from the definition.
Now here's the wicked part. Without knowing the definition of π, like if civilization collapsed and future thinkers just found a few scribbled equations with π in them, we can know many things about it just from observations of how it is used. For example, suppose we had no definition of π, and tested Coulomb force... this would give us a good approximation of π. If we know that π is between 3 and 22/7, we have more information to test for. Or we can use the normal (Gaussian) distribution to measure/calculate π, etc.
For a less esoteric and more real world example, consider Obama again. Do you know the "definition" of Obama? Do you know the precise formulation that identifies him, and him alone? Do you even know what that definition would entail (would it include his DNA? his life history? etc.)? i don't. But i do know how to figure out whether Obama is here. i can ask "is the first black president of the US here?", and even if someone doesn't know who Obama is, they can pick out the first black president, and i can use the equality to tell me that that's Obama. Thus, i can find Obama, without knowing Obama.
The same is true for intelligence. i don't need a definition of intelligence. i just need to know what traits it would have to have, and i can find many of them without a definition.
| ocalhoun wrote: | | Quote: |
THEN we can define intelligence as the sum of those traits, or a subset of them, or something else, and with that definition, we could improve and refine our testing procedure. But we don't need to do that to start creating a test now... and if we do that, and create a reasonable test, we can call whoever passes that test intelligent as a provisional definition.
|
A bit of a semantic quibble, but if you're only testing for various traits, without defining those traits as intelligence, you're not testing for intelligence yet.
Isn't it only an intelligence test once you define what you're testing for as 'intelligence'? |
That's not a semantic quibble, it's a good point. Without a precise definition we can never be sure that we're testing for intelligence... but even without a definition we can be pretty sure. If we know the traits that intelligence must come with, and the traits that don't matter, we can know - even without knowing the definition for intelligence - that a given test doesn't test for intelligence. For example, i think we all agree that intelligence is not a function of height. Therefore, if a test uses height, it's not really a test for intelligence.
The Turing test, for example, tests for a multitude of things that obviously have no relation to intelligence, such as the knowledge of idioms and social conventions, and other linguistic stuff.
| ocalhoun wrote: | | Quote: |
If we have a provisional test there's no reason we can't recognize that it is provisional, and that sometimes it will screw up. In fact, that's the logical thing to do, because then we can analyze how and when it fails, and refine it.
|
But how will you determine if it failed or not, much less how?
You need something to compare the results to.
Like it or not, the thing you're comparing them to is your definition of intelligence, even if it isn't precisely defined yet.
The vagueness of the definition will then make your comparison of test results erratic and subject to the differences of opinion held by different testers. |
Because you will be looking for traits that you know an intelligent being must have, while not being swayed by traits that you know they don't need. If you don't find those key traits, you know - even without defining intelligence - that the subject can't be intelligent. For example, the presence of a thinking apparatus - i don't need a definition of intelligence to know that a thinking apparatus is required for intelligence. A plain rock is not intelligent.
Without a precise definition of intelligence, we can never be sure something that passes the test is really intelligent, but we can be damn sure that anything that fails it isn't.
For example, suppose we had traits A-Z. We know from intuition and reason that, what whatever the definition of intelligence is, it must have traits A, B and C... and that traits X, Y and Z are superfluous. If we had the definition, we would know that intelligence is actually defined as having traits A through F... but we don't know that.
So we have a test that tests for trait A. Does this test for intelligence? Yes, but not completely, because not everything that has A is intelligent (it might have A-D, but not E and F). However, everything that fails (doesn't have A) isn't intelligent. Not bad, but we can do better.
So we have a test that tests for traits A and Z. Does this test for intelligence? No. Even without the definition we can know that, because we know Z can't possibly factor into intelligence.
So we have a test that tests for traits A and K. Does this test for intelligence? Maybe. We don't know for sure whether K matters. It may, or it may not. We can't trust this test.
So we have a test that tests for traits A, B and C. Does this test for intelligence? Yes, but not completely. It will never fail anything intelligent, but it may pass things that are not intelligent. Still, this is the best we can do without our current knowledge.
Basically, no, we are not comparing them to "our definition of intelligence". We don't have a definition. We don't even have a vague definition. We do have an idea of what the definition will have to include, and what it will have to ignore, but our knowledge there is not exhaustive. Still, it's a start.
And no, it doesn't follow that the test will be erratic or subjective. The Turing test is, sure, but the Turing test sucks. It's literally terrible. Even without knowing the precise definition of intelligence, we can know some of the traits that definition must include. All we need to do is test for those traits and no others. Subjectivity does not come into play, unless the measurements themselves are subjective, and there is no reason they have to be. (And, in fact, they shouldn't be, which is another reason the Turing test sucks so hard.)
| ocalhoun wrote: | | Quote: |
This is how most early scientific work is done, too. Before properties like time, temperature, inertia, etc. were defined, we had a vague notion of what they were, and designed instruments to measure them, and then defined the properties by the measurement. [...] but the important thing is that they measured time for centuries without even considering that time was an actual "thing", let alone defining what that thing was.) |
So, just apply the same concept; define the property (intelligence) by the measurement, then refine the measurement by how it agrees with a vague definition. |
Why? Why define intelligence by the measurement? There's no need to. At least, not yet.
It is possible to measure something to just to find out about it, without having to use the measured data to pigeonhole the subject. That wasn't the case with temperature/time/etc.... they had to make their provisional definitions, because they had to use their measurements for practical purposes. We don't. We're not interested in giving citizenship to machine intelligences (for example)... yet.
| ocalhoun wrote: | | -Just saying that the process will be smoother if you could compare the measurement with a precise definition. |
Well, of course. But in the absence of a definition, we can get by with the traits we know the definition must have. It's not perfect, but it's better than guessing.
You're so wrapped up in this definition thing that you're not seeing the forest for the trees. Do you know of some traits that intelligence... however it is defined, because no-one knows yet... must have? Then list those things, explain why intelligence - whatever it may be - must have those traits, and then devise a test to test for those traits. Simple.
Does your test also test for traits that intelligence doesn't need to have? Then in that case, your test is no good. (That's why the Turing test fails.)
See? Simple.
By the way, i'm a huge fan of Turing, and i think the dude got a raw deal, and i don't mean to call his reputation into account by saying his AI test is idiotic. i haven't read the original paper, but to my understanding, all Turing was trying to do was explain that it was possible to generate intelligence mechanically, using the test as a thought experiment to demonstrate his argument without having to define intelligence in any way (by the argument that if you think a human is intelligent, and you can't tell the difference, etc.)... not that his test should be a general-purpose method of testing for intelligence. Actually running the Turing test looking for real-world results is about as stupid as running the Schrödinger cat experiment.
| Indi wrote: |
Well, of course. But in the absence of a definition, we can get by with the traits we know the definition must have. It's not perfect, but it's better than guessing.
|
Which are...?
That's what I'm getting at here.
| ocalhoun wrote: | | Indi wrote: |
Well, of course. But in the absence of a definition, we can get by with the traits we know the definition must have. It's not perfect, but it's better than guessing.
|
Which are...?
That's what I'm getting at here. |
Is it? Because what you're getting at seems to change every post.
Alright, just off the top of my head:
- Intelligence must be non-physical. Proof: due to the nature of what we are talking about when we talk about intelligence. So any test that relies on physical characteristics cannot be a test for intelligence.
- Intelligence cannot rely on linguistic prowess. Proof: we don't consider people unintelligent when they have communication impediments, like dyslexia or dysphasia. So any test that relies on linguistic prowess cannot be a test for intelligence (the Turing test fails here).
- Intelligence cannot rely on social knowledge or awareness. Proof: society varies wildly across the globe, and across history, but we don't believe that any of that variation affects intelligence, therefore intelligence must be independent of it.
- Intelligence must include the ability to attach meaning to symbols. Proof: computers can manipulate symbols far better than humans, but we do not consider them intelligent because they cannot attach meaning to those symbols.
And so on....
If we figure out more characteristics, and if we can design a test to look for those characteristics, then we will have a workable test for intelligence, without needing a formal definition.
| Indi wrote: |
Is it? Because what you're getting at seems to change every post.
|
Of course it does. This is a two-way discussion, yes?
(Well, more than two.)
| Quote: |
If we figure out more characteristics, and if we can design a test to look for those characteristics, then we will have a workable test for intelligence, without needing a formal definition. |
But wouldn't 'figuring out more characteristics' and 'defining' be practically the same?
If you prefer, I'll amend my statement to,
"You have to figure out what characteristics define intelligence before you can test for intelligence."
(Supposing you want to use a very strict definition of 'define'.)
| ocalhoun wrote: | | But wouldn't 'figuring out more characteristics' and 'defining' be practically the same? |
No. In the limiting case - the case where you find ALL characteristics - you have effectively defined intelligence. But you don't need to find ALL characteristics of intelligence to test for intelligence.
Consider an example of elemental magnesium. Elemental magnesium is defined as a material made up solely of atoms with atomic number 12. From that definition, we can infer dozens of characteristics, such as its reactivity and so on.
But now suppose we don't have that definition - we don't know what "magnesium" is, precisely. However, we do know many characteristics of this material we have observed: it is silvery white, it burns with a brilliant white light, it's very light and strong, etc. (None of this is strange, because, after all, magnesium was identified, isolated and named long before they knew what atoms were, and before they even had a notion of the periodic table... precisely by observing its characteristics.) Obviously this list of characteristics isn't exhaustive - there is no mention there of spectral emissions, for example - so this list of characteristics cannot completely describe magnesium... but it can be used to identify it a very large percentage of the time.
In other words, if i put a lump of metal in front of you, and you didn't know what defined magnesium (you didn't know it's atomic makeup), but you did know that it was silvery white, that it burns with a brilliant white light and that it's very light, you could simply weigh it, burn it and look at it, and conclude with reasonable certainty that it is magnesium.
With absolute certainty? No, but that's okay. We're not looking for perfection, just good enough. Without a definition for intelligence we cannot be perfectly sure we've found it, but we can be reasonably sure just by looking for the traits we know it has to have.
It's just like a classic mystery story problem. You find enough clues (characteristics) to uniquely identify the perpetrator... but that doesn't mean you've described the perpetrator. You know whoever did the deed had access to the room, was unaccounted for during certain hours, and was left-handed... that may be enough to narrow down the perpetrator to a single person, but that's hardly the same as a definition of the perpetrator - you don't even know their gender from that info (for example)!
| ocalhoun wrote: | If you prefer, I'll amend my statement to,
"You have to figure out what characteristics define intelligence before you can test for intelligence."
(Supposing you want to use a very strict definition of 'define'.) |
No, you don't. You just have to figure out some of the characteristics intelligence must have (and you don't need the definition to do that).
Ideally, you want characteristics that nothing else you know of has, so if you find those characteristics, you can be reasonably sure you've found intelligence. (As with the magnesium example: not many things burn with a brilliant white light, so if you have something that does, you probably have magnesium. Of course, the more characteristics you test for - such as if you also test it's density etc. - the more certain your conclusion.)
You don't need a definition. It would be nice, sure, but we can do without it.
|
