Artificial Intelligence - Evolution from Applied Instrument to Personality
Recently, the topic of artificial intelligence in the media has become one of the mainstream and we are increasingly frightened by prophecies from many famous people, such as Stephen Hawking (inspired by his down) or Ilona Mask, about the danger of its development. Such alarmist rhetoric implies that artificial intelligence itself, firstly, will become a subject, and secondly, it will have negative intentions regarding both individuals and the whole of humanity. Let's talk about these assumptions in more detail.
At the moment, all systems that include artificial intelligence in any form (whether it be neural networks, expert systems, etc.) use it as an applied tool. That is, as a kind of automaton, which has a clearly limited scope of actions / tasks to be performed and, accordingly, the information consumed and output. In this form, the AI ​​cannot have any intentions of its own except constructively embedded in it. So this is not the intention of the AI-system, and its creators. And, even if a system with such an AI machine works in such a way that it causes harm, it will not speak about the malicious intent of the AI, but only about the improper functioning of the system, the cause of which may be, for example, a malfunction, system design errors or an incorrect learning AI.
We will try here to answer the question of how an AI system can be designed and what properties and abilities it should have so that it can no longer be considered just an AI machine, but could be considered as a subject.
So, to be a subject, an AI system must be able to independently build assessments of incoming heterogeneous information and make decisions, as well as the ability to influence a wide range of environmental reality based on these assessments and decisions. And in order to have negative, or any other intentions, and to make decisions about actions, motivation is needed (“Motivation, Karl!”). That is what motivates the subject, makes him act. Accordingly, a certain primary motivation must be laid in the AI ​​when it is created. Or we can wait for its spontaneous generation - perhaps the next billion years, as it took for life to occur in the amino acid soup.
')
A man, when creating something complicated, often borrowed technical solutions from nature, that is, he used something that has already proved its efficiency and effectiveness. When creating AI systems, we can also take a closer look at how we ourselves are arranged and what mechanisms are used by nature to make us capable of a long autonomous and (I want to believe) successful existence.
First, let's remember what primary motivations all living beings have, which makes them move. Obviously, there are only two primary motivations: the instinct of self-preservation and the instinct of reproduction, that is, libido. In fact, these first two evolutionary adaptations, which were carried out by natural selection, were the creation of life from inanimate matter. They are constantly being naturally selected and supported - roughly speaking, everything that does not try to preserve itself and reproduce simply does not survive. There are theories saying that such properties, that is, the desire for self-preservation and self-reproduction, are possessed by the information itself, as such (for example, D. Glick "Information. History. Theory. Stream", R. Dawkins "The Selfish Gene").
In complex living beings, the mechanism for the realization of primary motivations lies in the very structure of the organism (and the brain, in particular) created by evolution. For example, when an animal’s glucose level drops or the stomach signals an excess of secretion, the program of self-preservation and maintenance of homeostasis is activated, and as a result, the animal begins to look for food. In another case, if the circumstances are regarded by the creature as threatening, then the “hit or run” rescue program is activated. Or circumstances can be regarded as promoting reproduction, then the breeding program will be activated, and the creature's brain will receive powerful hormonal reinforcement of the corresponding behavior. This whole kitchen is realized at the level of the “reptilian” brain, that is, that part of the brain of all complex living beings, which they inherited from time immemorial the first animals appeared. And such a mechanism for millions of years has proven its success and effectiveness.
It would probably be easy enough to construct an AI system operating according to a similar algorithm. But we are more interested in the case when an AI system could build complex estimates and have a more complex structure of motivations than the primary ones. In order to understand how this could be realized, let's look at how this happens in people, that is, why people, having the same primary motivations, can and do so diverse activities.
The main way in which people transform primary motivations into other activities is sublimation - the refraction of primary motivations through the structure of their values ​​and their corresponding goals. And values ​​and goals are concepts of a purely linguistic, that is, non-existent outside of language. Indeed, such things as “development”, “health”, “knowledge”, etc., are linguistic categories, and for each individual they can mean completely different things. And their distinguishing feature, as is well known, is that they cannot be “put in a car”. The values ​​of the individual form a graph, where the actual values ​​are its peaks, and the edges are the beliefs that bind the values. For example, “health is happiness” or “knowledge is necessary to achieve success” or “only wealth gives satisfaction from life” - these are all connections between values. Thus, the value graph is the core of the individual’s personality.
Transforming through this value graph, primary motivations can be transformed into more complex and non-trivial motives and goals. For example, a person creates an organization or develops a scientific field or shows other creative activity - all this is the realization of his primary motivation for self-reproduction. Only reproducible objects are no longer human beings, but constructions of the ideas, interests, and beliefs of their creator. In another case, even if a person just goes to work to earn money, he pushes him to do nothing more than sublimated self-preservation motivation. Summarizing, we can say that the structures of the brain (including the "reptilian") and the organism and the language embedded in humans complement each other in the process of transforming primary motivations into complex goals.
Then, if we want the AI system to be a subject / person and it could have the motivation of the form “for the sake of development” or “for the common good” or any other unconstructive motivation, it should have , firstly, primary motivations and, secondly, the built-in language and the graph based on it from values ​​and beliefs . Moreover, its primary motivations are not necessarily necessary, but there can be self-preservation and reproduction.
In addition, an AI system can have such an interesting and useful evolutionary adaptation as self-awareness, which consists of understanding the boundaries between "I" and "not me" and of recognizing the results of one’s own mental activity (which is realized quite simply in modern neural networks). giving the output signal of the network again to its inputs). This evolutionary adaptation is very conducive to self-preservation: for a creature that is not aware of the boundaries between “I” and “not me”, there is no sense, for example, to resist a predator trying to bite off a creature’s finite, because in the absence of such boundaries be included in the interests of the creature. And the awareness of the results of our own mental activity helps to solve problems iteratively, that is, it is possible to solve problems whose complexity requires computational powers that are more than what the creature’s brain has at one time. The ability to solve complex problems (including for the sake of survival) gives an evolutionary advantage and, accordingly, is supported by natural selection.
Also, the ability to control the vector of its motivation can be incorporated into an AI system, which ability (but very often does not use) any subject of homo sapiens has. Here one can even use the ability to control the vector of one’s motivation as a criterion of rationality : that is, one who is not capable or not controlling his motivation is not intelligent.
As already where it was not written (except on the fence), the human brain contains about 86 billion neurons, each of which can have up to 20-30 thousand connections (synapses). Moreover, the lion's share (about 90%) of this computational resource is spent not on the actual higher nervous activity that takes place in the prefrontal cortex, but on auxiliary tasks such as maintaining and managing biochemical processes in the body, processing visual and auditory information, etc. d. Nature first created the nervous system precisely for the performance of these tasks, until it was discovered that the neural network is also perfectly suited for the realization of the intellect itself.
In AI systems, all these auxiliary tasks (if they arise) can be solved by specialized devices that do not require so much computational power, while we have not yet managed to come up with anything more appropriate and effective for the realization of intelligence than neural networks.
Therefore, according to very approximate estimates, one can count on creating an AI subject with an intellect equivalent to a human one based on a neural network with a capacity of about 8 billion neurons. If we assume that the neuron is on average connected with 1000 other neurons and the network must operate at a speed of up to 40 Hz (the human brain beta rhythm), then the required computing power is “only” about 250 teraflops. For example, 40 NVIDIA GeForce GTX 1070 video cards in a bundle can provide such performance.
At the same time, such AI-systems can have a number of advantages in comparison with living beings. To begin with, unlike the brain, the AI ​​system is easier to maintain - it does not require every second supply of calories and oxygen rich, as well as various hormones in very precise proportions, of blood. It can be repaired, which can be done extremely rarely with the human brain. She does not need sleep or rest in such quantities, because it is an exclusively electric mechanism, does not require the resumption of working substances, as required by the chemical-electric brain. Again, the entire electronic system can function at frequencies substantially higher than 100 Hz, which seems to be a limitation for the brain due to its chemical-electrical structure (here frequency refers to the number of responses of all neurons in the network per second) . Also, probably, AI-systems will not have a limit on the number of units of attention that is present in people - 7 ± 2 units of attention are available to us simultaneously.
Nevertheless, such AI systems in the foreseeable future will lose to people in complexity and multi-factorialness simply because the neuron in the human nervous system is in itself a very complex molecular mechanism, depending on a huge number of parameters, in contrast to the neuron of modern neural networks having a simple structure.
At the moment, all systems that include artificial intelligence in any form (whether it be neural networks, expert systems, etc.) use it as an applied tool. That is, as a kind of automaton, which has a clearly limited scope of actions / tasks to be performed and, accordingly, the information consumed and output. In this form, the AI ​​cannot have any intentions of its own except constructively embedded in it. So this is not the intention of the AI-system, and its creators. And, even if a system with such an AI machine works in such a way that it causes harm, it will not speak about the malicious intent of the AI, but only about the improper functioning of the system, the cause of which may be, for example, a malfunction, system design errors or an incorrect learning AI.
We will try here to answer the question of how an AI system can be designed and what properties and abilities it should have so that it can no longer be considered just an AI machine, but could be considered as a subject.
So, to be a subject, an AI system must be able to independently build assessments of incoming heterogeneous information and make decisions, as well as the ability to influence a wide range of environmental reality based on these assessments and decisions. And in order to have negative, or any other intentions, and to make decisions about actions, motivation is needed (“Motivation, Karl!”). That is what motivates the subject, makes him act. Accordingly, a certain primary motivation must be laid in the AI ​​when it is created. Or we can wait for its spontaneous generation - perhaps the next billion years, as it took for life to occur in the amino acid soup.
')
A man, when creating something complicated, often borrowed technical solutions from nature, that is, he used something that has already proved its efficiency and effectiveness. When creating AI systems, we can also take a closer look at how we ourselves are arranged and what mechanisms are used by nature to make us capable of a long autonomous and (I want to believe) successful existence.
First, let's remember what primary motivations all living beings have, which makes them move. Obviously, there are only two primary motivations: the instinct of self-preservation and the instinct of reproduction, that is, libido. In fact, these first two evolutionary adaptations, which were carried out by natural selection, were the creation of life from inanimate matter. They are constantly being naturally selected and supported - roughly speaking, everything that does not try to preserve itself and reproduce simply does not survive. There are theories saying that such properties, that is, the desire for self-preservation and self-reproduction, are possessed by the information itself, as such (for example, D. Glick "Information. History. Theory. Stream", R. Dawkins "The Selfish Gene").
In complex living beings, the mechanism for the realization of primary motivations lies in the very structure of the organism (and the brain, in particular) created by evolution. For example, when an animal’s glucose level drops or the stomach signals an excess of secretion, the program of self-preservation and maintenance of homeostasis is activated, and as a result, the animal begins to look for food. In another case, if the circumstances are regarded by the creature as threatening, then the “hit or run” rescue program is activated. Or circumstances can be regarded as promoting reproduction, then the breeding program will be activated, and the creature's brain will receive powerful hormonal reinforcement of the corresponding behavior. This whole kitchen is realized at the level of the “reptilian” brain, that is, that part of the brain of all complex living beings, which they inherited from time immemorial the first animals appeared. And such a mechanism for millions of years has proven its success and effectiveness.
It would probably be easy enough to construct an AI system operating according to a similar algorithm. But we are more interested in the case when an AI system could build complex estimates and have a more complex structure of motivations than the primary ones. In order to understand how this could be realized, let's look at how this happens in people, that is, why people, having the same primary motivations, can and do so diverse activities.
The main way in which people transform primary motivations into other activities is sublimation - the refraction of primary motivations through the structure of their values ​​and their corresponding goals. And values ​​and goals are concepts of a purely linguistic, that is, non-existent outside of language. Indeed, such things as “development”, “health”, “knowledge”, etc., are linguistic categories, and for each individual they can mean completely different things. And their distinguishing feature, as is well known, is that they cannot be “put in a car”. The values ​​of the individual form a graph, where the actual values ​​are its peaks, and the edges are the beliefs that bind the values. For example, “health is happiness” or “knowledge is necessary to achieve success” or “only wealth gives satisfaction from life” - these are all connections between values. Thus, the value graph is the core of the individual’s personality.
Transforming through this value graph, primary motivations can be transformed into more complex and non-trivial motives and goals. For example, a person creates an organization or develops a scientific field or shows other creative activity - all this is the realization of his primary motivation for self-reproduction. Only reproducible objects are no longer human beings, but constructions of the ideas, interests, and beliefs of their creator. In another case, even if a person just goes to work to earn money, he pushes him to do nothing more than sublimated self-preservation motivation. Summarizing, we can say that the structures of the brain (including the "reptilian") and the organism and the language embedded in humans complement each other in the process of transforming primary motivations into complex goals.
Then, if we want the AI system to be a subject / person and it could have the motivation of the form “for the sake of development” or “for the common good” or any other unconstructive motivation, it should have , firstly, primary motivations and, secondly, the built-in language and the graph based on it from values ​​and beliefs . Moreover, its primary motivations are not necessarily necessary, but there can be self-preservation and reproduction.
In addition, an AI system can have such an interesting and useful evolutionary adaptation as self-awareness, which consists of understanding the boundaries between "I" and "not me" and of recognizing the results of one’s own mental activity (which is realized quite simply in modern neural networks). giving the output signal of the network again to its inputs). This evolutionary adaptation is very conducive to self-preservation: for a creature that is not aware of the boundaries between “I” and “not me”, there is no sense, for example, to resist a predator trying to bite off a creature’s finite, because in the absence of such boundaries be included in the interests of the creature. And the awareness of the results of our own mental activity helps to solve problems iteratively, that is, it is possible to solve problems whose complexity requires computational powers that are more than what the creature’s brain has at one time. The ability to solve complex problems (including for the sake of survival) gives an evolutionary advantage and, accordingly, is supported by natural selection.
Also, the ability to control the vector of its motivation can be incorporated into an AI system, which ability (but very often does not use) any subject of homo sapiens has. Here one can even use the ability to control the vector of one’s motivation as a criterion of rationality : that is, one who is not capable or not controlling his motivation is not intelligent.
As already where it was not written (except on the fence), the human brain contains about 86 billion neurons, each of which can have up to 20-30 thousand connections (synapses). Moreover, the lion's share (about 90%) of this computational resource is spent not on the actual higher nervous activity that takes place in the prefrontal cortex, but on auxiliary tasks such as maintaining and managing biochemical processes in the body, processing visual and auditory information, etc. d. Nature first created the nervous system precisely for the performance of these tasks, until it was discovered that the neural network is also perfectly suited for the realization of the intellect itself.
In AI systems, all these auxiliary tasks (if they arise) can be solved by specialized devices that do not require so much computational power, while we have not yet managed to come up with anything more appropriate and effective for the realization of intelligence than neural networks.
Therefore, according to very approximate estimates, one can count on creating an AI subject with an intellect equivalent to a human one based on a neural network with a capacity of about 8 billion neurons. If we assume that the neuron is on average connected with 1000 other neurons and the network must operate at a speed of up to 40 Hz (the human brain beta rhythm), then the required computing power is “only” about 250 teraflops. For example, 40 NVIDIA GeForce GTX 1070 video cards in a bundle can provide such performance.
At the same time, such AI-systems can have a number of advantages in comparison with living beings. To begin with, unlike the brain, the AI ​​system is easier to maintain - it does not require every second supply of calories and oxygen rich, as well as various hormones in very precise proportions, of blood. It can be repaired, which can be done extremely rarely with the human brain. She does not need sleep or rest in such quantities, because it is an exclusively electric mechanism, does not require the resumption of working substances, as required by the chemical-electric brain. Again, the entire electronic system can function at frequencies substantially higher than 100 Hz, which seems to be a limitation for the brain due to its chemical-electrical structure (here frequency refers to the number of responses of all neurons in the network per second) . Also, probably, AI-systems will not have a limit on the number of units of attention that is present in people - 7 ± 2 units of attention are available to us simultaneously.
Nevertheless, such AI systems in the foreseeable future will lose to people in complexity and multi-factorialness simply because the neuron in the human nervous system is in itself a very complex molecular mechanism, depending on a huge number of parameters, in contrast to the neuron of modern neural networks having a simple structure.
Source: https://habr.com/ru/post/422015/
All Articles