Artificial intelligence, great and terrible. Part two
Start Next
Like the previous one, this note is a review of articles and comments to them here on Habré, grouped into several topics.
Perhaps the funniest, but, oddly enough, is still a widely discussed topic. Actually, there is simply nothing to talk about here: is it not clear to someone that if the AI starts to get out of the diaper at least a little (and even earlier - even now!), The military and every kind of “state security” will immediately be imposed on it? And they will teach the AI exactly to kill , not bothering with any nonsense on the topic of "pseudo-ethics." Is it not clear that these “laws” in principle do not work and cannot work, which Asimov himself knew very well? Is it not clear that the main threat from the AI is just linked to human control? And what, we will believe in fairy tales, asking: "Will we give the military robots a license to kill?" Does anyone ask us? Or: "The largest IT companies will not allow AI to escape from the control of a person." Is someone asking them? Or: "The well-known AI developers around the world have agreed not to create smart weapons." Stupid will create? Or: "The UN expert called on the world community to slow down the creation of combat robots with artificial intelligence." But did the large countries capable of creating such robots ever consider the opinion of the UN? So let's stop discussing obvious nonsense - this is not even a kindergarten, this is a junior nursery.
')
For some reason, this nonsense is considered the criterion of AI. And not even just an AI, but a strong AI! At the same time, nobody cares that the person from the 19th century will not pass the Turing test in our time. As well as the Chinese with Indian appraisers. In general, for the pre-determined composition of the jury, the Turing test will not pass the majority of the population of our planet. The test is based on the assumption that the intellect is inherent only to humans - even animals are denied intelligence! - he doesn’t check the presence of intelligence at all, but the similarity of Homo sapiens being tested for representatives in the eyes of “examiners” of the same kind. Moreover, in the almost equally stupid "IQ test" computers have already bypassed people.
A cursory analysis of the conversations during the test says that in order for the computer to recognize the presence of intelligence, he should not only lie and pretend to be an idiot - it is also a good idea to use bad language and make grammatical mistakes. And then his chances of becoming an “intellectual” will increase dramatically. And another point: no one argues that the specific intelligence of different people are different? There are geniuses, there are idiots and there are other normal and not very people. So intelligence - is it a quantitative characteristic or qualitative ? And what if the AI does pass the Turing test, will it have intelligence? Nothing like that! This only means that in animals of the species Homo sapiens with an exorbitantly inflated CHS, elected as a judge, there is no such intelligence for nothing! And if at some point the average intellectual level of humanity falls quite low - to about the level of an orangutan (apparently, it has been for years, not decades), then almost all such machines will have artificial intelligence.
The state of California has forbidden chat bots to pretend to be people: one of the most difficult problems of artificial intelligence is solved elegantly and simply. The prohibition is explained by the fact that when a deception is revealed, a very negative reaction of the client always follows. And vice versa: if a chat bot immediately warns who he is (more precisely, WHAT he is), the person still very easily animates him and ascribes human properties to him. It has long been known: a real programmer with a computer is not even on “you”, but on “you, scum”!
Well, and a wonderful comment in one of the branches: “The Turing test is meaningless. He tests not the ability to think, but the ability to pretend to be human. This is equivalent to offering a person to behave like a dog - so as to deceive other dogs, and if it does not work out, then consider a person more stupid than a dog. ” Subscribe!
There are a lot of discussions like “Doing Artificial Intelligence” or “Everything you need to know about AI in a few minutes” or “It's been (!) About four months since I seriously (!) Got into Artificial Intelligence.” There is nothing to comment here either - it remains only to sigh: what a shit they call AI now!
The opposite tendency is to consider the development of AI to be incredibly difficult or even impossible in principle. That von Neumann architecture does not fit - give them quantum computers, then "as we already know (!), A computer is not able to create something distinctively new, or, in other words, not capable of engaging in creative activity," or "cannot exist RR algorithm, since the algorithm is created to solve a specific problem and the scope of its application is limited to this task . ” In the extreme case, it will require an unmeasured amount of dough - of course, for yourself, loved ones: “The system we have invented will require several billion dollars to develop and build, and in order to make it compact, it will include several advanced innovations” . Here I will focus only on one “theory”, which amused me very much: “There are a number of problems (class NP) that it is simply inexpedient to solve algorithmically. And even: It is APPROXIMATE that is not described algorithmically. There are ways to approach the traveling salesman. But this is NOT an algorithmic solution to the problem. ”
Actually, the answer was given in the same place: “A person cannot accurately solve even solvable NP-complete problems, but from the AI he wants to solve algorithmically unsolvable”. I have a program for solving the traveling salesman problem, there is even a long-standing publication on this topic (“Software for solving of TSP”), and I argue that not only “NP tasks should be solved algorithmically,” but that approximation is PERFECTly described algorithmically. I do not know how anyone, but MY methods of the APPROACHED solution of a traveling salesman are quite algorithmic.
And no one knows! And the versions are less and less. It is hard to think of something that we have IT, but nobody else had it: the games have disappeared, the Turing test is ridiculous, the IQ test is failed ... What is left?
Training: AI must be self-taught. Having realized self-training, he will have to learn the rest himself. The algorithm entered into the system remains unchanged, and self-learning occurs due to the constant complication of the processed information, due to the constant transition from a low level of abstraction to a higher one.
Bullshit! Algorithms should be a lot! They are MANY, and therefore, to perceive the mind as ONE algorithm, especially unchanged, means to stupidly task to the level of complete impossibility of creating at least some kind of AI. As you know, none of the Mowgli children for some reason did not “self-educate” anything. And even Lomonosov went from Kholmogor to Moscow by no means to educate himself, namely, to look for teachers. What can we say about us, who are almost kicked into schools? First, someone will teach yourself, and then demand it from the AI!
Task setting: algorithms can solve problems, and a full-fledged intellect can set goals for itself and find the most effective ways to achieve them.
About 20 years ago I teased supporters of this thesis as follows: “The ability to set tasks is a task that must be solved.” In addition, any idiot "with one gyrus, and that trace from the cap" in five minutes so many tasks ponasavit - rake evaporated!
Algorithm compilation: the ability to find an algorithm for solving a problem, the method of solving which is unknown to you. And even: the ability to solve arbitrary tasks formulated in natural language, the conditions of which are poorly formalized, and the initial data are not collected in a ready-made set .
So how? Can people brag about it? Many have found "problem solving, the solution of which is not known"? I now consider myself an algorithmist, and I assert that finding an algorithm is a trivial matter that does not require a single drop of intelligence! No more difficult than setting a task! But to find a good algorithm or the right to set the task is much more difficult. But this is not the most difficult. But telling the others that the algorithm is REALLY good is already an almost impossible task! If someone does not believe - look at the next branch, where I just proposed an algorithm for solving the problem, the solutions to which are not just known, but are calculated in dozens!
Sense of humor: There are various discussions on the topic “Microsoft developed AI with a sense of humor” or “ can the computer understand sarcasm? ”. But since no one knows not only what “intellect” is, but also what “sense of humor” is, all this reasoning is in fact all the same chatter about anything.
Emotions: the algorithm can not feel anything. AI has neither suffering nor thoughts. Only juggling bits in memory.
Let's turn it around: Man cannot feel anything, has neither suffering nor thoughts. Only juggling pulses between neurons. Emotions are fairly easy to program, and the presence of emotions is fundamentally impossible to distinguish from their imitation. This is all a consequence of the very error when information is considered material, for example: “Information is material by definition and in essence. Information is the transfer of energy between particles. Information and energy are, in fact, equivalent concepts. ” I do not even want to comment on such nonsense, just ask: “Do animals have emotions? And the intellect? Oh, no? Then why even get to the bottom of emotions? Cut them off with Occam's razor!
Consciousness: an obvious fact for its owner (I know that I have a consciousness), and completely unprovable for everyone else (people around me will not see the difference between me and a complex system that imitates my reactions to external signals) ”.
Here is exactly the same situation: the formal signs of consciousness are unknown to society. So there is also nothing to talk about here.
We know almost nothing about the IE, but we are already building the AI, so let's take a little look at the “natural neural network”. To begin with, we discard the thesis that “the human brain is analog.” Personally, I think that the world is generally discrete and all this nonsense about the “analog brain” is precisely nonsense. We cannot even measure the analog signal at all! After all, it must be compared with some kind of reference scale, and any scale is discrete by definition. So we can get any information ONLY digitally!
The second important property of the brain is the ability to forget: the old data is erased in order to give way to a new one. Personally, I do not believe that “with the help of hypnosis from memory, a mountain of everything forgotten is drawn out”, because there is no guarantee that these memories really correspond to the way things really were. I also do not believe in brain tales about “petaflops” (vision gives 10 ^ 6 bits / s, touch - 10 ^ 5 bits / s, hearing - 10 ^ 4 bits / s), which was brilliantly refuted here, in Habré: “The volume of incoming information is greatly exaggerated. Imagine an average person reading an average book of 300 pages. Reading without interruption for food, toilet, sleep and sex. Suppose that it will take a day. We get 24 * 60 * 60 * 10 ^ 6 bits = 864000 ^ 6 bits of incoming information. At the same time, in a text-based digital format, without loss of quality, this volume is unlikely to exceed 1 megabyte. ” Subscribe!
And one more good phrase: “The first AI, which can be called intelligence, will not be the one who will solve the ethical problem, but the one who refuses to solve it, because he has no mood today or the wires are itching.”
There are still three large topics that are not covered by AI: neural networks, variations on SkyNet and my favorite databases. But this is the material for the next note.
Like the previous one, this note is a review of articles and comments to them here on Habré, grouped into several topics.
The laws of robotics Asimov
Perhaps the funniest, but, oddly enough, is still a widely discussed topic. Actually, there is simply nothing to talk about here: is it not clear to someone that if the AI starts to get out of the diaper at least a little (and even earlier - even now!), The military and every kind of “state security” will immediately be imposed on it? And they will teach the AI exactly to kill , not bothering with any nonsense on the topic of "pseudo-ethics." Is it not clear that these “laws” in principle do not work and cannot work, which Asimov himself knew very well? Is it not clear that the main threat from the AI is just linked to human control? And what, we will believe in fairy tales, asking: "Will we give the military robots a license to kill?" Does anyone ask us? Or: "The largest IT companies will not allow AI to escape from the control of a person." Is someone asking them? Or: "The well-known AI developers around the world have agreed not to create smart weapons." Stupid will create? Or: "The UN expert called on the world community to slow down the creation of combat robots with artificial intelligence." But did the large countries capable of creating such robots ever consider the opinion of the UN? So let's stop discussing obvious nonsense - this is not even a kindergarten, this is a junior nursery.
')
Turing Test
For some reason, this nonsense is considered the criterion of AI. And not even just an AI, but a strong AI! At the same time, nobody cares that the person from the 19th century will not pass the Turing test in our time. As well as the Chinese with Indian appraisers. In general, for the pre-determined composition of the jury, the Turing test will not pass the majority of the population of our planet. The test is based on the assumption that the intellect is inherent only to humans - even animals are denied intelligence! - he doesn’t check the presence of intelligence at all, but the similarity of Homo sapiens being tested for representatives in the eyes of “examiners” of the same kind. Moreover, in the almost equally stupid "IQ test" computers have already bypassed people.
A cursory analysis of the conversations during the test says that in order for the computer to recognize the presence of intelligence, he should not only lie and pretend to be an idiot - it is also a good idea to use bad language and make grammatical mistakes. And then his chances of becoming an “intellectual” will increase dramatically. And another point: no one argues that the specific intelligence of different people are different? There are geniuses, there are idiots and there are other normal and not very people. So intelligence - is it a quantitative characteristic or qualitative ? And what if the AI does pass the Turing test, will it have intelligence? Nothing like that! This only means that in animals of the species Homo sapiens with an exorbitantly inflated CHS, elected as a judge, there is no such intelligence for nothing! And if at some point the average intellectual level of humanity falls quite low - to about the level of an orangutan (apparently, it has been for years, not decades), then almost all such machines will have artificial intelligence.
The state of California has forbidden chat bots to pretend to be people: one of the most difficult problems of artificial intelligence is solved elegantly and simply. The prohibition is explained by the fact that when a deception is revealed, a very negative reaction of the client always follows. And vice versa: if a chat bot immediately warns who he is (more precisely, WHAT he is), the person still very easily animates him and ascribes human properties to him. It has long been known: a real programmer with a computer is not even on “you”, but on “you, scum”!
Well, and a wonderful comment in one of the branches: “The Turing test is meaningless. He tests not the ability to think, but the ability to pretend to be human. This is equivalent to offering a person to behave like a dog - so as to deceive other dogs, and if it does not work out, then consider a person more stupid than a dog. ” Subscribe!
AI is just
There are a lot of discussions like “Doing Artificial Intelligence” or “Everything you need to know about AI in a few minutes” or “It's been (!) About four months since I seriously (!) Got into Artificial Intelligence.” There is nothing to comment here either - it remains only to sigh: what a shit they call AI now!
AI is hard
The opposite tendency is to consider the development of AI to be incredibly difficult or even impossible in principle. That von Neumann architecture does not fit - give them quantum computers, then "as we already know (!), A computer is not able to create something distinctively new, or, in other words, not capable of engaging in creative activity," or "cannot exist RR algorithm, since the algorithm is created to solve a specific problem and the scope of its application is limited to this task . ” In the extreme case, it will require an unmeasured amount of dough - of course, for yourself, loved ones: “The system we have invented will require several billion dollars to develop and build, and in order to make it compact, it will include several advanced innovations” . Here I will focus only on one “theory”, which amused me very much: “There are a number of problems (class NP) that it is simply inexpedient to solve algorithmically. And even: It is APPROXIMATE that is not described algorithmically. There are ways to approach the traveling salesman. But this is NOT an algorithmic solution to the problem. ”
Actually, the answer was given in the same place: “A person cannot accurately solve even solvable NP-complete problems, but from the AI he wants to solve algorithmically unsolvable”. I have a program for solving the traveling salesman problem, there is even a long-standing publication on this topic (“Software for solving of TSP”), and I argue that not only “NP tasks should be solved algorithmically,” but that approximation is PERFECTly described algorithmically. I do not know how anyone, but MY methods of the APPROACHED solution of a traveling salesman are quite algorithmic.
What is AI?
And no one knows! And the versions are less and less. It is hard to think of something that we have IT, but nobody else had it: the games have disappeared, the Turing test is ridiculous, the IQ test is failed ... What is left?
Training: AI must be self-taught. Having realized self-training, he will have to learn the rest himself. The algorithm entered into the system remains unchanged, and self-learning occurs due to the constant complication of the processed information, due to the constant transition from a low level of abstraction to a higher one.
Bullshit! Algorithms should be a lot! They are MANY, and therefore, to perceive the mind as ONE algorithm, especially unchanged, means to stupidly task to the level of complete impossibility of creating at least some kind of AI. As you know, none of the Mowgli children for some reason did not “self-educate” anything. And even Lomonosov went from Kholmogor to Moscow by no means to educate himself, namely, to look for teachers. What can we say about us, who are almost kicked into schools? First, someone will teach yourself, and then demand it from the AI!
Task setting: algorithms can solve problems, and a full-fledged intellect can set goals for itself and find the most effective ways to achieve them.
About 20 years ago I teased supporters of this thesis as follows: “The ability to set tasks is a task that must be solved.” In addition, any idiot "with one gyrus, and that trace from the cap" in five minutes so many tasks ponasavit - rake evaporated!
Algorithm compilation: the ability to find an algorithm for solving a problem, the method of solving which is unknown to you. And even: the ability to solve arbitrary tasks formulated in natural language, the conditions of which are poorly formalized, and the initial data are not collected in a ready-made set .
So how? Can people brag about it? Many have found "problem solving, the solution of which is not known"? I now consider myself an algorithmist, and I assert that finding an algorithm is a trivial matter that does not require a single drop of intelligence! No more difficult than setting a task! But to find a good algorithm or the right to set the task is much more difficult. But this is not the most difficult. But telling the others that the algorithm is REALLY good is already an almost impossible task! If someone does not believe - look at the next branch, where I just proposed an algorithm for solving the problem, the solutions to which are not just known, but are calculated in dozens!
Sense of humor: There are various discussions on the topic “Microsoft developed AI with a sense of humor” or “ can the computer understand sarcasm? ”. But since no one knows not only what “intellect” is, but also what “sense of humor” is, all this reasoning is in fact all the same chatter about anything.
Emotions: the algorithm can not feel anything. AI has neither suffering nor thoughts. Only juggling bits in memory.
Let's turn it around: Man cannot feel anything, has neither suffering nor thoughts. Only juggling pulses between neurons. Emotions are fairly easy to program, and the presence of emotions is fundamentally impossible to distinguish from their imitation. This is all a consequence of the very error when information is considered material, for example: “Information is material by definition and in essence. Information is the transfer of energy between particles. Information and energy are, in fact, equivalent concepts. ” I do not even want to comment on such nonsense, just ask: “Do animals have emotions? And the intellect? Oh, no? Then why even get to the bottom of emotions? Cut them off with Occam's razor!
Consciousness: an obvious fact for its owner (I know that I have a consciousness), and completely unprovable for everyone else (people around me will not see the difference between me and a complex system that imitates my reactions to external signals) ”.
Here is exactly the same situation: the formal signs of consciousness are unknown to society. So there is also nothing to talk about here.
Brain
We know almost nothing about the IE, but we are already building the AI, so let's take a little look at the “natural neural network”. To begin with, we discard the thesis that “the human brain is analog.” Personally, I think that the world is generally discrete and all this nonsense about the “analog brain” is precisely nonsense. We cannot even measure the analog signal at all! After all, it must be compared with some kind of reference scale, and any scale is discrete by definition. So we can get any information ONLY digitally!
The second important property of the brain is the ability to forget: the old data is erased in order to give way to a new one. Personally, I do not believe that “with the help of hypnosis from memory, a mountain of everything forgotten is drawn out”, because there is no guarantee that these memories really correspond to the way things really were. I also do not believe in brain tales about “petaflops” (vision gives 10 ^ 6 bits / s, touch - 10 ^ 5 bits / s, hearing - 10 ^ 4 bits / s), which was brilliantly refuted here, in Habré: “The volume of incoming information is greatly exaggerated. Imagine an average person reading an average book of 300 pages. Reading without interruption for food, toilet, sleep and sex. Suppose that it will take a day. We get 24 * 60 * 60 * 10 ^ 6 bits = 864000 ^ 6 bits of incoming information. At the same time, in a text-based digital format, without loss of quality, this volume is unlikely to exceed 1 megabyte. ” Subscribe!
And one more good phrase: “The first AI, which can be called intelligence, will not be the one who will solve the ethical problem, but the one who refuses to solve it, because he has no mood today or the wires are itching.”
There are still three large topics that are not covered by AI: neural networks, variations on SkyNet and my favorite databases. But this is the material for the next note.
Source: https://habr.com/ru/post/449838/
All Articles