Turing Test Breakthrough / AI thread

[BLANKBLANK 4,414]

http://www.independent.co.uk/life-style/gadgets-and-tech/computer-becomes-first-to-pass-turing-test-in-artificial-intelligence-milestone-but-academics-warn-of-dangerous-future-9508370.html

 

An interesting development in the pursuit of artificial intelligence. I sincerely doubt it is a sentient artificial intelligence, but how close are we to creating a sentient artificial lifeform?

 

Aside from the article, I think this thread would be a good place to discuss AI, its feasibility and of course what sentient intelligence itself is. Let's get deep people!

[Duke87 1,070]

This writer scathingly begs to differ on this matter.

 

To be fair, he has a valid point: there is no standard definition of "passing the Turing Test", and the achievement of the program here is not particularly amazing. Indeed, its level of success in duping judges is largely dependent on its claim that it is a 13 year old for whom English is not his first language. It is a decent chatbot program, but hardly anywhere close to truly being able to imitate a human.

[BLANKBLANK 4,414]

This writer scathingly begs to differ on this matter.

 

To be fair, he has a valid point: there is no standard definition of "passing the Turing Test", and the achievement of the program here is not particularly amazing. Indeed, its level of success in duping judges is largely dependent on its claim that it is a 13 year old for whom English is not his first language. It is a decent chatbot program, but hardly anywhere close to truly being able to imitate a human.

 

Oops! It wasn't the first time the papers have misled me.

 

Got trolled badly. Oh well.

[hym 358]

 

This writer scathingly begs to differ on this matter.

 

To be fair, he has a valid point: there is no standard definition of "passing the Turing Test", and the achievement of the program here is not particularly amazing. Indeed, its level of success in duping judges is largely dependent on its claim that it is a 13 year old for whom English is not his first language. It is a decent chatbot program, but hardly anywhere close to truly being able to imitate a human.

 

Oops! It wasn't the first time the papers have misled me.

 

Got trolled badly. Oh well.

 

 

In general, the world of AI largely disregards the Turing test as a popular, but worthless, test of intelligence.  There have been quite a few computer science PhD papers on the problems with the test and why it should be relegated to history.  One of the most well-known rebuttals of the Turing test is John Searle's Chinese Room.

 

Those who have read and remember my earlier comments on the issue of AI, particularly those relating to the concept of the Singularity, know that I think the concept is absolute bunk.  It is, in the words of one prominent computer scientist, "Intelligent Design for the 140+ IQ crowd."  The Chinese Room metaphor is probably the most succinct argument for why I consider the idea of the Singularity to be preposterous.  No matter how intelligent we think an AI is, it does not understand what it is doing.  It is following a set of rules given to it by a human being, and it will never exceed the bounds that those rules place on it.  Even when we design self-optimizing AI, it only exhibits first level understanding (which is nothing more than conditional logic).  Not only has no one achieved second level understanding for AI, we don't even understand how one would achieve such a thing.

 

I think this is one of those areas where people attribute significance to things simply because they don't know any better.  They see something they don't understand, so they anthropomorphize it in an effort to understand it, and subsequently ascribe to it qualities that it does not possess.

 

EDIT:  Perhaps it might be helpful to throw in this consideration.  Video games are an area where there is quite a bit of work on development of "intelligent AI."  One of the most common adages of AI development for video games is that the goal of the AI systems are not to be intelligent, but rather, to give the illusion of intelligence.  A lot of AI development, such as the chatbot in this article, has nothing to do with developing anything remotely intelligent, but rather, tricking people into believing that it is intelligent.

[Easy Bakes 253]

they should go back to Analog systems

[AkemiAzula 280]

Define sentient. It strikes me that both our brains which are organic computers and their mechanic counterparts arose from the same big bang. Logically if sentience can arise in flesh why not metal. Unless were arguing for souls which would redirect somewhat to the religion thread. At any rate intelligence of inorganic machines will mosy likely achieve singularity within our century imho.

[LexusInfernus 908]

In general, the world of AI largely disregards the Turing test as a popular, but worthless, test of intelligence.  There have been quite a few computer science PhD papers on the problems with the test and why it should be relegated to history.  One of the most well-known rebuttals of the Turing test is John Searle's Chinese Room.

 

Those who have read and remember my earlier comments on the issue of AI, particularly those relating to the concept of the Singularity, know that I think the concept is absolute bunk.  It is, in the words of one prominent computer scientist, "Intelligent Design for the 140+ IQ crowd."  The Chinese Room metaphor is probably the most succinct argument for why I consider the idea of the Singularity to be preposterous.  No matter how intelligent we think an AI is, it does not understand what it is doing.  It is following a set of rules given to it by a human being, and it will never exceed the bounds that those rules place on it.  Even when we design self-optimizing AI, it only exhibits first level understanding (which is nothing more than conditional logic).  Not only has no one achieved second level understanding for AI, we don't even understand how one would achieve such a thing.

 

I think this is one of those areas where people attribute significance to things simply because they don't know any better.  They see something they don't understand, so they anthropomorphize it in an effort to understand it, and subsequently ascribe to it qualities that it does not possess.

 

EDIT:  Perhaps it might be helpful to throw in this consideration.  Video games are an area where there is quite a bit of work on development of "intelligent AI."  One of the most common adages of AI development for video games is that the goal of the AI systems are not to be intelligent, but rather, to give the illusion of intelligence.  A lot of AI development, such as the chatbot in this article, has nothing to do with developing anything remotely intelligent, but rather, tricking people into believing that it is intelligent.

 

You assume that AI is just trying to mimic intelligence. Indeed, such a thing would never set the singularity in motion. The thing is, programs mimicking intelligence are not real AI's. The goal of AI research is to come up with a program that not just appears intelligent, but is intelligent, because it can understand the answers it gives. A true AI would be conscious of itself in a similar way humans are conscious of themselves. To use your Chinese room example, it would be putting someone who actually understands Chinese into that room. 

 

And why would that be impossible? Tell me, why do you understand the English language? What makes you understand it? Its your brain. A small part of your brain. But what is your brain? Neurons and synapses, which operate under an electric currents and chemical reactions. So what is your brain if not a hyper advanced biological computer? And what does that make you, if not the biological equivalent of an AI? You are by all accounts little more than the representation of your brain in its current form.

 

So, at least one way for us to make a real AI is to make a super computer, which consists of god knows how many smaller computer programs, operating together which would form its own consciousness. 

[A Nonny Moose 8,261]

Well, I've been in the computer game since 1962 and have been involved in a lot of exotic things, including chess automata.  After all this time, I still believe that Artificial Intelligence is an oxymoron.

 

A sentient computer will be able to reject commands given by anyone, will have reasoning power of its own, and will have at least the desire if not the capability to reproduce.  I think as long as we are stuck in the digital world, this won't happen.  An analogue system, on the other hand, is a whole new ball-game.  No matter how many on-off switches you have in your "neural network" you will never emulate the analog system found in a pint of gray gunk inside your skull.

[LexusInfernus 908]

Human like intelligence does not mean human like preferences. But indeed, a true AI would have 'free will' of a similar level as a human being. 

 

And that grey matter in your skull is basically just a biological computer chip. Of course, for now, more advanced given its size. But that is simply a matter of time before we can replicate it. 

[hym 358]

@LexusInfernus:

The system you described is effectively the architecture for the Google search system and we know that the search system is not the least bit intelligent as you have described an intelligent system.

Would you care to provide a detailed, actionable process through which AI researchers could turn the system you described into an system that meets the standard of intelligence you provided?

[LexusInfernus 908]

@LexusInfernus:

The system you described is effectively the architecture for the Google search system and we know that the search system is not the least bit intelligent as you have described an intelligent system.

Would you care to provide a detailed, actionable process through which AI researchers could turn the system you described into an system that meets the standard of intelligence you provided?

How do you mean, Google is not intelligent? Sure, it hasn't reached human levels of intelligence. But I think one could make the case it has about the intelligence level of certain small animals or insects. It lacks awareness of itself and its actions, but so do many animals, who just follow the programming provided by their genes. 

 

Google lacks complexity, but if you were to add more and more processes, with the goal of creating an actual singular intelligence instead of a highly efficient search engine, you would eventually reach a point where Google would start to recognize what its doing. I think though, it would need a second intelligence to evolve with it though, before it becomes intelligent and conscious enough to start off the singularity. An AI would be a unique form of intelligence, and it would need something like it to share its experience and reach full recognition of itself. 

 

So for AI researchers, they would basically have to follow how the human brain works. It has areas that are devoted to specific tasks. So build small programs that focus on one task, and as they work in conjunction with each other, they will result in some sort of consciousness, provided that there are enough of those processes. 

[BLANKBLANK 4,414]

@LexusInfernus:

The system you described is effectively the architecture for the Google search system and we know that the search system is not the least bit intelligent as you have described an intelligent system.

Would you care to provide a detailed, actionable process through which AI researchers could turn the system you described into an system that meets the standard of intelligence you provided?

 

I do appreciate your point that to date no computer system has exceeded its limits or the specification of its human masters. But what if, with ever more advanced feedback operations (all operations comprise of feedback), networking between operations and with the dawn of quantum computers. What if, given all of that complexity, a spark of evolution happens? The system takes on a life of its own?

 

The human brain is just a biological computer after all, with networking between areas of the brain. But what gives the human brain that spark? Which allows it to generate sentience?

[A Nonny Moose 8,261]

The point is that the human brain is not digital.  It is analogue and very much interconnected. 

 

A quantum computer is simply one that can have whatever state the qubits want to have at the time of access.  Quantum mechanics is something like God playing at dice, and I don't buy it for a moment.  It makes a mockery out of any concepts of reality.  And, by the way, at the moment these goodies operate at the temperature of liquid Helium.  Not very practical, and you sure won't have one of those on your desk.  The whole business reminds me of Seymour Cray who said "Parity is for farmers".

 

I know there was some classified work being done on photonic computers, but since I no longer have clearance, I don't know what happened there and if it was still classified and I did know, I couldn't say.

 

As for true man-made self-aware gadgets, I'd rather see a Ghaldron-Hesthor interdimensional transporter.  We've a much better chance of discovering para-time than constructing A.I. but it does keep some creative minds at work.

[LexusInfernus 908]

The human brain is just a biological computer after all, with networking between areas of the brain. But what gives the human brain that spark? Which allows it to generate sentience?

 

Id say the complexity of the network results in sentience. Sentience would simply be generated by the brain activity taken together.

 

It would make sense given what we know about the brain. Deactivate parts of the brain, and you reduce peoples sentience or completely change their personality.  

 

The point is that the human brain is not digital.  It is analogue and very much interconnected. 

Is the brain analogue? It works with neurons, neurons can either fire a signal, or they don't. That sounds very 'on-off' or '0-1' or digital to me. The neurons signal also remains the same. A stronger input does not result in a stronger signal, only in an increased frequency of firing. 

[A Nonny Moose 8,261]

Good point, but are the neuron signals the only things going on?  There's still a lot of open water in there.  The full histology of these cells is not understood.

[AkemiAzula 280]

Sentience. I think therefore I am sentient.

I dont know how to define or recreate sentience.

[LexusInfernus 908]

Good point, but are the neuron signals the only things going on?  There's still a lot of open water in there.  The full histology of these cells is not understood.

Well then there are the neurotransmitters. This bit is analog, as its not just an on-off distinction. The reaction is based on the amount of neurotransmitters released, which can vary. Basically the brain is both digital and analog. 

 

The question is, can't programs be programmed in such a way that their behavior is essentially analog? 

[A Nonny Moose 8,261]

Well, how would you formulate a neurotransmitter in a digital computer?  It has less than half of the required equipment.  Neuron behaviour varies, and the same set of neurotransmitters are not always used.  Neurotransmitters are chemicals that carry messages at an incredibly fast and voluminous rate over very short distances.  Just think about the operation of vision.

[hym 358]

How do you mean, Google is not intelligent? Sure, it hasn't reached human levels of intelligence. But I think one could make the case it has about the intelligence level of certain small animals or insects. It lacks awareness of itself and its actions, but so do many animals, who just follow the programming provided by their genes. 

 

Intelligence purists will tell you that intelligence is defined as the ability to correctly reason through illogical means, thus AI, by definition, cannot be intelligent.  However, some people consider that a meaningless distinction, so let's move past that.

 

To go back to the Chinese room metaphor, Google's search system is just like the man in the room.  He (Google's search system) is passed questions (queries from users like us).  He takes the "Chinese" you have "slipped under the door" to a set of well-defined rules (the book in the metaphor), writes answers (your search results), and "slips it back under the door."  At no time does he understand the "Chinese" he is writing.  When Google engineers improve the search algorithms, they are not making him more intelligent.  They are merely giving him a better book to allow him to write higher quality "Chinese."

 

So, no, Google's search system is not intelligent.  It fakes intelligence very well, thanks to the work of a lot of engineers that have created an incredibly detailed and complex system for data analysis, but that system does not understand a thing of what it is analyzing.

 

Google lacks complexity, but if you were to add more and more processes, with the goal of creating an actual singular intelligence instead of a highly efficient search engine, you would eventually reach a point where Google would start to recognize what its doing. I think though, it would need a second intelligence to evolve with it though, before it becomes intelligent and conscious enough to start off the singularity. An AI would be a unique form of intelligence, and it would need something like it to share its experience and reach full recognition of itself. 

 

Out of curiosity, if such an approach worked, wouldn't we have done it by now?  Google has the money to buy the smartest people in the world, and the first organization to produce truly intelligent AI would become rich beyond comparison.  That alone would be incentive enough to pursue such an endeavor if it was considered even remotely possible.  Even if Google for some strange reason had no interest, any AI researcher who seriously believed such a thing was possible would be begging for a chance to implement such a system simply for the fame it would garner.  And yet, no one is attempting this.

 

Your proposal for how to implement an intelligent AI roughly boils down to the following steps:

 

1. Create a bunch of stuff
2. Get stuff talking with stuff
3. ???
4. Intelligent AI

 

The problem with proposals for developing intelligent AI is that they all, at some point, involve that "???" step somewhere.  Somewhere in the process of developing the AI, you have to imbue it with the ability to think for itself, and no one (not even Ray Kurzweil) understands how to do this.  Many serious AI researchers don't consider it possible.  (Others will say that it is, but we are nowhere near close to knowing how.)  For what it is worth, there is a running quip in the AI community that there are two types of AI researchers: those who believe in God and those who are going to believe it eventually.  It is no exaggeration that there are AI researchers who have ultimately come to believe that humanity was created by God simply because they can find no other explanation for why humans can think.

 

It is this general reason that I said a lot of AI research isn't focused on creating truly intelligent AI, but rather, mimicking intelligence.  A lot of researchers don't believe a truly intelligent AI is possible, and if you don't believe that is possible, what is the point of trying to make one?  Most of the AI research we do is not focused on developing intelligent systems, but rather, designing systems that fake intelligence well enough that there is no meaningful difference between the two.  Despite being theoretically inferior to an AI that actually thinks for itself, an AI that fakes intelligence actually has a lot more practical value.  For this reason, one is able to reasonably argue that the goal of AI research isn't to create a truly intelligent AI, but rather fake intelligence really well (much like what is done in video games).

 

So for AI researchers, they would basically have to follow how the human brain works. It has areas that are devoted to specific tasks. So build small programs that focus on one task, and as they work in conjunction with each other, they will result in some sort of consciousness, provided that there are enough of those processes. 

 

For what it is worth, neural networks allow us to develop AIs that function in much the same method as the human brain.  We've been researching that for nearly 70 years.  So let's look at what we have achieved in those years.  Neural network AIs are, without a doubt, one of the powerful AI models ever developed.  They are also incredibly fragile.  A single bad input can cause "catastrophic unlearning" and destroy the AI.  Neural networks can leverage Moore's Law and distributed computing to discover things that their human creators would die before discovering.  They can also sit there and spit out garbage ad infinitum, sometimes from almost identical data sets.  Like the man in the Chinese room, they don't understand what they are doing; they rely on the rule book being correct for them to provide accurate answers.

 

To take it out of the abstract and into a more real world example, take IBM's Watson.  Without a doubt one of the most powerful AI systems ever developed.  It too, has absolutely no understanding of what it is doing.  The "rule book" that it used to play Jeopardy! is so fragile to bad information that IBM engineers literally hand programmed data into it.  When Watson was repurposed to help fight lung cancer, those same IBM engineers had to hand feed decades of medical cases into the system to prevent it from engaging in catastrophic unlearning.  Watson is incredibly powerful, but it does not understand the questions it is asked to answer.

 

I know there was some classified work being done on photonic computers, but since I no longer have clearance, I don't know what happened there and if it was still classified and I did know, I couldn't say.

 

There is public work being done on the subject.  Someone a while back claimed to be working on a light powered CPU that natively recognized an 8 state "bit."  Intel is believed to be working on such a processor as a successor to silicone and Group III/V metal based processors.  (Some speculate that Intel's fiber optic Thunderbolt protocol is based on the same research.)

[A Nonny Moose 8,261]

Interesting story about the first successful self-aware system.  After it had been fully loaded and given a self-sufficient power source some one asked it "Is there a God?" and the answer was "Now there is.".

[LexusInfernus 908]

Intelligence purists will tell you that intelligence is defined as the ability to correctly reason through illogical means, thus AI, by definition, cannot be intelligent.  However, some people consider that a meaningless distinction, so let's move past that.

 

To go back to the Chinese room metaphor, Google's search system is just like the man in the room.  He (Google's search system) is passed questions (queries from users like us).  He takes the "Chinese" you have "slipped under the door" to a set of well-defined rules (the book in the metaphor), writes answers (your search results), and "slips it back under the door."  At no time does he understand the "Chinese" he is writing.  When Google engineers improve the search algorithms, they are not making him more intelligent.  They are merely giving him a better book to allow him to write higher quality "Chinese."

 

So, no, Google's search system is not intelligent.  It fakes intelligence very well, thanks to the work of a lot of engineers that have created an incredibly detailed and complex system for data analysis, but that system does not understand a thing of what it is analyzing.

And you think a goldfish understands the things it does? Do you think it understands the concepts of hunger, or eating food, or swimming, or breathing? It just does those things, without ever truly understanding what it does. Goldfish are living Chinese rooms. But would you say that goldfish lack all forms of intelligence? I wouldn't say that a goldfish lacks intelligence completely. 

 

If you measure intelligence by using the concept of the Chinese room (thus, the ability to understand your own actions) only a very small number of species are 'intelligent' to a certain degree (great apes and maybe dolphins). But what are all the other animals and bugs then? Very elaborate natural programs? Biological automatons? 

 

Still, that only proves that the singularity IS possible, since nature has done it herself. We, humanity, are natures own singularity. We have reached a point of intelligence that made us so smart that we became capable of improving ourselves without the help of outside evolution. We use tools, we build things, and we have changed the face of this earth completely. And if nature can do it, so can we. 

 

 

The problem with proposals for developing intelligent AI is that they all, at some point, involve that "???" step somewhere.  Somewhere in the process of developing the AI, you have to imbue it with the ability to think for itself, and no one (not even Ray Kurzweil) understands how to do this.  Many serious AI researchers don't consider it possible.  (Others will say that it is, but we are nowhere near close to knowing how.)  For what it is worth, there is a running quip in the AI community that there are two types of AI researchers: those who believe in God and those who are going to believe it eventually.  It is no exaggeration that there are AI researchers who have ultimately come to believe that humanity was created by God simply because they can find no other explanation for why humans can think.

 

It is this general reason that I said a lot of AI research isn't focused on creating truly intelligent AI, but rather, mimicking intelligence.  A lot of researchers don't believe a truly intelligent AI is possible, and if you don't believe that is possible, what is the point of trying to make one?  Most of the AI research we do is not focused on developing intelligent systems, but rather, designing systems that fake intelligence well enough that there is no meaningful difference between the two.  Despite being theoretically inferior to an AI that actually thinks for itself, an AI that fakes intelligence actually has a lot more practical value.  For this reason, one is able to reasonably argue that the goal of AI research isn't to create a truly intelligent AI, but rather fake intelligence really well (much like what is done in video games).

Yep, nonsense. If these 'scientists' honestly start believing in God because they can't figure out how humans came to think, they were hacks to begin with. Honestly, not being able to figure out the question does not mean that therefor God did it. 

 

As for believing in the possibility of something, well, a little more than a 100 years ago, a lot of scientists were also convinced that it was impossible for people to fly using other means than hot air balloons/zeppelins. And look how incredibly wrong they turned out to be. But had humanity followed your advice, we wouldn't have had planes. Really, history is littered with examples of people believing that something was impossible, only to proven utterly wrong on every count. And in so many of those cases, the only thing that was preventing them from doing it sooner was some engineering problem. I bet that the only thing that stands in our way from developing a true AI is an engineering problem. One that I believe will be solved within the next 35 years. And think about it, this is one of the quickest developing fields. 

 

 

For what it is worth, neural networks allow us to develop AIs that function in much the same method as the human brain.  We've been researching that for nearly 70 years.  So let's look at what we have achieved in those years.  Neural network AIs are, without a doubt, one of the powerful AI models ever developed.  They are also incredibly fragile.  A single bad input can cause "catastrophic unlearning" and destroy the AI.  Neural networks can leverage Moore's Law and distributed computing to discover things that their human creators would die before discovering.  They can also sit there and spit out garbage ad infinitum, sometimes from almost identical data sets.  Like the man in the Chinese room, they don't understand what they are doing; they rely on the rule book being correct for them to provide accurate answers.

 

To take it out of the abstract and into a more real world example, take IBM's Watson.  Without a doubt one of the most powerful AI systems ever developed.  It too, has absolutely no understanding of what it is doing.  The "rule book" that it used to play Jeopardy! is so fragile to bad information that IBM engineers literally hand programmed data into it.  When Watson was repurposed to help fight lung cancer, those same IBM engineers had to hand feed decades of medical cases into the system to prevent it from engaging in catastrophic unlearning.  Watson is incredibly powerful, but it does not understand the questions it is asked to answer.

For what its worth is that we barely grasp how neural networks in our brain work, so it should be no surprise that when we try to build them into computers, they don't work so well yet. First, we need a better understanding of our brain. Sure, they have been trying to do that for decades now, but for the most part of those decades, they did not have the imaging machines necessary to come up with an accurate picture. Even now, MRI scans are limited in truly mapping the brain. But, again, this is just an engineering problem. Within 5-15 years scientists will find ways to increase the level of detail and accuracy of brain scan, or come up with entirely new and better methods all together. 

 

Once we fully understand how the brain works, we fully understand how intelligence works and where it comes from. And then, we just need to find a way to turn that into programming. 

[A Nonny Moose 8,261]

The problem with physically imaging the brain, is it doesn't tell the full story.  Until we devise a method of following every chemical interaction, success will remain elusive.  These reactions may occur in nanosecond time or less.  I can't imagine anyone being able to do this in a living brain.

[LexusInfernus 908]

The problem with physically imaging the brain, is it doesn't tell the full story.  Until we devise a method of following every chemical interaction, success will remain elusive.  These reactions may occur in nanosecond time or less.  I can't imagine anyone being able to do this in a living brain.

One, that is just an engineer problem. Ray Kurzweil assumes that with the creation of nanobots, we can effectively have nanobots map the human brain with unmatched precision without damaging the brain itself. 

 

And second, those reactions are actually quite slow. Its based on chemicals. It takes some time. Those reactions are observable (and we have been doing that for some time now). 

[A Nonny Moose 8,261]

There is no such thing as "just an engineering problem".  If that was true, why are we still playing rocket propulsion?  It might be an engineering chemistry problem with all kinds of strange and wonderful physics involved as well.  As for nanobots, I'll believe it when it has some concrete results that can be duplicated in a lab.

[LexusInfernus 908]

There is no such thing as "just an engineering problem".  If that was true, why are we still playing rocket propulsion?  It might be an engineering chemistry problem with all kinds of strange and wonderful physics involved as well.  As for nanobots, I'll believe it when it has some concrete results that can be duplicated in a lab.

There is such a thing as just an engineering problem, but not every problem is just an engineering problem.