Wednesday, December 07, 2005

General AI, The Turing Test, and Other Minds

The Turing Test, first put forth in Alan Turing's seminal paper “computing machinery and intelligence”, is recognized by most AI theorists as the ultimate test for human level general intelligence. Many, however, feel that this test is outdated and that it is not a fair test for machine intelligence because machines are so different from humans.

In this post I ask what the significance of this test is and I defend the rational behind it.

To really cut to the heart of the matter of why this test is so significant we need to take a look at a long standing problem in the history of philosophy. That problem is the problem of other minds.

It is recognized that there is no objective way to prove that anyone other than yourself has any subjective experience. This must seem like an odd thing to say. Most of us are inclined to say that of course other people have conscious thoughts – why we all do, right? Well, yes it is assumed that we all do but because no one can ever experience the world except through one's own mind we must infer this through the behavior of others. There is no absolute test to determine if the people you meet everyday have a mind but we assume they do based on their intelligent behavior.

Imagine a world where we doubted if others had a mind. What a terrible thought. So most of us who have considered this problem simply decide that everyone must have a mind, even if this decision is just made for pragmatic reasons (since no “scientific” evidence can be adduced to support it).

This raises a very interesting question. What would we think if something non-human could behave in a manner practically indistinguishable from human behavior? Well it would follow that unless we are prepared to start questioning whether or not our fellow humans have minds we must be prepared to accept that these non-human entities have minds if their behavior is truly indistinguishable from our own.

Now lets go a little deeper into this and examine just what a machine would need to be able to do in order to pass a Turing Test.

The main thing, and by far and away the most difficult, is mastering human natural language. Some people view the hilariously unintelligent behavior displayed by contemporary 'chat bots' and think that it must be possible to master language without being able to think. This kind of misunderstanding is implicit in John Searle's “Chines Room Argument”. This, however, is a gross misunderstanding. In order to actually have command of a language one must be able to think.

There are an infinite number of possible sentences that can be formulated. Think about that. That's the power of natural language. If you can think of an idea then there is some combination of words that can represent that idea. Every day you probably put together a string of words that was never uttered before. The point of what I'm trying to say is that true mastery of natural language cannot be accomplished through gimmicks and tricks. It requires genuine human human level intelligence. Try asking a 'chat bot' what the significance of Plato's “Allegory of the Cave” is and you will see what I mean. No machine today can even come close to thinking and anyone who claims they can hasn't the slightest idea what thought really is.

So what is 'thought' then, you ask? Thought is the ability to understand interrelation. This is part of what is so difficult about natural language. Somehow we manage to ascribe 'meaning' to a word by interrelating that word to a cluster of other word. Some words refer to specific objects in the environment but other word refer to more abstract relationships between objects. All of the words derive meaning and connotation from other words. The greater understanding one has of how all these words interrelate then the greater command over language and thought one acquires. So you see the ability to communicate and think are inextricably connected. I'm sure most of you are familiar with the stories of neglected children who never learned a language and the result was a failure to be able to think abstractly. Also it is of interest to point out that most anthropologists are in agreement that the story of the evolution of modern man from Homo Habalis is a story of the evolution of language. So we recognize that modern man (the only known example of general intelligence) first arose when the ability to communicate was fully developed and man is the only one known to posses this skill.

So in order for a machine to display intelligence indistinguishable from human intelligence it must actually understand the interrelations of human propositions about the world. If a machine could actually do this the ramifications would be astonishing. This is because the machine could then take all known propositions about our world that are written in natural language and almost instantaneously formulate any possible deductions between any possible combination of propositions. That would be amazing. That would be intelligence.

So if a machine is ever able to master natural language then it should easily be able to pass a Turing test. But I don't think that is necessarily what will convince us that the machine is intelligent. What will convince us is its ability to respond to any question with amazing answers by being able to synthesize and analyze all know propositions in order to come up with novel answers to difficult problems.

That's intelligence, and that is what we can expect from a machine capable of passing a Turing test.

For more reading on the Turing Test go here.


Passing a Turing test does not imply self awareness. Having a three way conversation including a non-self aware machine might be awkward.

By Blogger al fin, at 8/12/05 07:16  

The machine might not be self-aware but if its behavior betrayed this fact in any way you would be able to pick up on that and hence it would not pass a Turing test. I have a feeling, however, that self-awareness is a necessery prerequisite for natural language mastery. I actually just read something very interesting about this last night and I'm going to go ahead and right today's post about it. Thanks for the comments Al.

By Blogger Micah J. Glasser, at 8/12/05 12:05  

Here’s a list of tools you will need to start: Jewelers’ pandora jewellery wire cutters - If you can only afford one pair, get memory wire shears. pandora charms These are designed to make clean cuts on tough memory wire, so can also be used for pandora charms uk softer wires. Chain-nose pliers sometimes called cheap pandora charms needle-nose pliers – Very versatile for picking up and grasping small items, pandora charms sale bending eye pins, closing jumps rings, even closing crimp beads. discount pandora charms Round-nose pliers – Used for creating loops on beaded head and eye pins. Can also be used for winding your own jump rings and as the second pliers you’cheap pandora ll need for closing jump rings. Optional pliers – Wire-looping pliers which have several graduated circumferences to allow you to form perfectly uniform jump rings and loops in place of the pandora discount uk round-nose pliers mentioned above. Crimping pliers which have little notches to allow you to both flatten a crimp bead and then bend it to form a rounded finished look instead of the flat crimp you pandora uk get using the chain-nose pliers. As for materials, I recommend some assortment packs of beads in coordinating colors, some decorative metal spacers, seed beads in both silver and gold These can serve as spacers and beautifully set off pandora sale your other beads., tube-shaped crimp beads Buy the best you can find – these are what hold it all together!, head and eye pins. Other than that, let your choice of project be your guide. You might want some silver or pewter charms.

By Blogger 每当遇见你, at 7/10/10 23:22  

Post a Comment

<< Home