Making Artificial Intelligence

    Prologue

    About four months have passed since I became seriously interested in the problem of Artificial Intelligence. At first, the desire was not so obvious - I just wanted to write a small program with natural-language interaction as a training. But the more I thought about this goal, the more it became more complicated and, in the end, changed beyond recognition, it became more like Artificial Intelligence rather than what it was at the beginning. Rummaging through the net, I did not find anything worthwhile on this topic and decided to continue to develop my original idea.

    At the moment, the result of numerous applications of analysis and synthesis is a certain amount of information in the form of concepts, partial models of architecture, implementation ideas and forecasts. So I wanted to share with the public.

    I want to say right away that I believe in my idea and develop it not where I will lead, but so that it can be really realized, and in the very near future. So that the result can be called Artificial Intelligence (or Reason, I do not share these concepts), while stupidly not giggling and not looking somewhere to the side with a guilty look.
     
    The first article will be devoted to the basic concept, which I, without thinking for a long time, called “The Concept of Human AI”.
     
    I guess how such statements can sound for the average habrauha, so that I'm mentally ready for almost anything!
     

    Legend

    After the 56th, when the first concept of Artificial Intelligence was announced to the world, a lot of water flowed. Scientists, and after them simple inhabitants, were divided into several directions. And each of them has its own view of the question, its definition, its own criteria. Someone is based on the idea that AI is already invented, while on the contrary, it is impossible to invent AI. For reference, it would be possible to lay out all these definitions, but I see no reason to engage in copy-paste from the same Wikipedia. It will not give anything. Most people have their own views on this issue and will defend what is consistent with their views.
     
    Therefore, I will take advantage of this right and briefly state my opinion, what I mean when I write “Artificial Intelligence”. In general terms, this corresponds to the definition of “strong artificial intelligence” introduced by John Searle. It says that AI should not be a model of the human mind, but the mind itself, to the extent that the human mind is the mind. Yes, the wording is rather confusing, but I think you get the point.
     

    List of unused literature

    Every self-respecting speaker should have a similar section, so I also think - what makes me worse! Only the meaning of this section is slightly opposite for completely objective reasons.
     
    Here I would like to clarify my statement from the prologue about the fact that "I did not find anything worthwhile on the net." First of all, I do not consider any of the main lines of research promising for the creation of Artificial Intelligence (see Symbols) in the near future. For other purposes - to health, much is already being used.
     
    Some areas of cybernetics were considered in my first article, " Painting without an Artist ." In it, descriptions of technologies are interspersed with my comments, each of which makes it clear why we will not achieve the goal with this technology. In the second articleNeurocybernetics became the object of criticism. Already by tradition, it contains a description of my view on this area and thoughts on why I need to dig in another direction to create AI.
     
    Behind the scenes are still the ideas of science fiction writers and various stereotypes. Consider the most popular of them.
     
    So, in the first place among the misconceptions about AI, I put serial models of robots, popular in science fiction films of the 90s. This also includes most of the robots that I respect Azimov and, for example, S-3PO. I united them according to the following criterion - they are all created with an intellect already wired, programmed for certain actions in certain conditions, which is either completely incapable of rewriting its program (invariant) or is capable, but very slightly. In addition, each of these robots has cinematic memory.
     
    What is the crazyness of this image, which is the goal of cybernetics? If these creatures are described by the near, then I do not mind - this is quite possible. But this is not AI. For a weak-minded robot, you can write some algorithm, but it will be very far from the human mind. But to write a program that gives the robot such properties that it passes the Turing test and generally thinks no worse than a person is impossible. It’s the same as developing a training program for a person of preschool age, so that in a week (month, year) he masters the knowledge and experience of an adult.
     
    Well, in the second place I have fantastic mega-computers: Skynet, UAVs from Stealth, VIKI from I, Robot, Set from Universal Soldier 2. I think you can continue the list yourself. In principle, they do not particularly differ from the robots considered just above, but they are perceived not as something humanlike, but as a program that can be run on a computer. Many hope that such things will hang on servers with web search engines, on their own machine as part of any IDE or even as an OS. Moreover, some purely human properties are expected from them: thinking, the ability to solve non-trivial tasks, make difficult choices, and the like.
     
    What is wrong here? Indeed, everything looks beautiful and convincing. On the other hand, in such hypothetical systems they want to use only some of the human properties, while others, because of their futility, are not welcomed and ignored. For example, I perceive this situation as follows: “I like apples, so bring me a kind of apple tree that would consist of apples alone.” In other words, we want to take a part of the intellect, and that this part, in spite of everything, be a full-fledged intellect.
     
    I propose to dwell on this. But if such a need arises, some well-known theory or hypothesis that I have not mentioned comes up, it may well be here, among others, in the list of unused literature.
     

    Key Points

    In a way, they follow from the previous section. Our Human AI must have the maximum number of properties of its creator. After launch, it should have the same knowledge, that is, almost none, and the same skills that can generally be called the ability to self-learn. Further it will work autonomously.
     
    The learning process will take place in the same sequence as a person’s, starting with acquaintance with the outside world, learning speech, then development and complication of the picture of the world goes on. All these stages should occur in the same way as in humans. Therefore, an important criterion is communication with people. Without it, AI simply cannot develop. It follows that after the AI ​​is created and launched, it will take several years to train it. Strictly speaking, he will continue to study in the same way, if the necessary conditions exist.
     
    One would think that it would be much more efficient to first arrange some block of information in a special way, and then simply copy it to the AI. But the human mind works with information much more complicated than a search engine. AI must itself, in the same ways as a person, obtain this information and then arrange it, again just like a person.
     
    I can’t say for sure how fast training will take place. Indeed, the most important information, especially “at an early age”, the AI ​​will receive from communicating with people like any child. So in the first stage, this will limit the speed of learning. Further, when he reaches the mental development of at least a 15-20 year old person, he will be able to gather a lot of information from books, articles and other textual information carriers. With such a development, this information can already affect his views, give knowledge that he can take and use.
     
    Here the question is concluded - at what speed will training take place? Can a limitation in a person’s learning speed be a consequence of the peculiarities of his mind? How much can you speed up the process with the same functionality and efficiency? The partial models that I created are not yet so perfect as to give answers to these questions. But already now we can say that the option in which the AI ​​after launching will go online, wiser in a couple of minutes, and in a couple of minutes - launch nuclear missiles, disappears. As we train him, he will become so.
     
    As for training, the question may arise - can AI outperform humans? Forget for a moment the prophecies of science fiction writers and sensibly assess the situation. In this case, intelligence must be perceived as something concrete. But the specific intelligences of different people are different. There are geniuses, there are idiots, and there are other normal and not very people. So Artificial Intelligence can be on equal terms among not the most stupid representatives of humanity, and surpass them. After all, he goes the same way from birth, has the same incentives and cognitive abilities, he also has organized memories. So ask yourself the question, what prevents you from becoming the smartest in the world?
     
    If we create AI identical in structure with the human mind, then we will automatically endow it with all the properties and functions characteristic of a person. This includes feelings, and intuition, the ability to lie and forgive - everything that relates to the characteristics of the mind.
     
    Creating something that functions exactly like our intellect is much simpler than posing as a god and from scratch inventing a system that could develop autonomously (without interfering with architecture) from the moment it is launched. That's when we do Human AI, when we understand how it works, then we will know what and how to optimize, if at all something can be optimized.
     

    How will it look like

    Everything described in the previous section is nothing more than a philosophy, an idea that must permeate all stages of development and must be present in the final result. This section will be devoted to how this final result will look.
     
    Platform

    We have already decided that we will have the maximum functional similarity in the architectures of AI and human minds. This means that absolutely everything related to the human mind must be present in Artificial Intelligence. I emphasize that the similarity must be functional.
     
    If we need something “absolutely” human, then here some problems may arise. Man lives not only with his intellect, thoughts, dreams. His head wears a body that constantly makes itself felt and intervenes in our sacred thoughts and dreams. However, without this intervention, we would not be ourselves, since many processes in the body somehow affect our Self. Something affects only superficially, for example, when a bad tooth prevents us from concentrating, but there are also things whose influence is somewhat deeper like hormones. We will also not forget about innate fears, instincts, etc. If you try to copy all this diversity, you will have to create, in addition to the intellect itself, an exact copy of the entire human body, and as a result we will get a person. I personally don’t see the point of it (except self-assertion) and I don’t see the possibility of achieving this in an “unnatural” way. Yet we are not gods or even UFOs.
     
    Therefore, the abstraction level of the model (system) should be higher than the level of cells (neurons). In this case, the created model will be platform independent. The priority platform is naturally a computer. The level of abstraction should be such that the entire system can be implemented in an object-oriented language. And since any information can be encoded using text, in order to organize the memory, it is only necessary to develop a method of modulation (coding) and markup. The main advantages of this solution: saving on expensive extraterrestrial technology due to the use of conventional computers with conventional hard disks for storing memory and saving on programmers working with an extraterrestrial programming language, in favor of the good old C and the like.
     
    Interaction

    We decided on the platform. Now we begin to describe the properties that it should possess in the first place. Let me bring Minski to the account of Artificial Intelligence:
    Research in the field of artificial intelligence is aimed at developing programs that solve such problems that people now cope better with, because they require the involvement of such functions of the human brain as the ability to learn based on perception, the special organization of memory and the ability to draw conclusions based on judgments.
    Here it is - perceptual learning ability. In order for the intellect to receive information, something must supply it with it. In humans, this function is performed by the sense organs. But to reproduce each of them will be difficult. However, this complexity cannot be compared with the problem of storing the received information. After all, it is necessary not only to record and reproduce it, but to perform more complex actions on it, for which it must be interpreted and stored in a special way, as it is stored in humans.
     
    Which of the developments in this area currently exist? Probably the most important of them and, probably, therefore the most developed one is vision, namely pattern recognition. But the level of its development does not allow using it to the full, that is, the way we do. So the first thing we have to give up is eyesight. For most people, vision is the main source of information, so they need to find a worthy replacement.
     
    We’ll think about why now we can try to organize a normal “human” memory. Perhaps many will say “text” - and be right. Although a person does not have such a separate body, but we decided that the similarity would not be stupid one to one. Already there are knowledge storage systems in the form of just text. Of course, in the form as they are now, they cannot be used, since, again, a person uses their analogues in a different way, but this is already something.
     
    Speech can be attributed to the textual form of communication, and partly, hearing. After a small study of the literature on teaching blind children, it turned out that a prerequisite for the formation of speech is the presence of non-textual information coming from the senses. So far, we have not decided on any of them, but it turns out that you cannot learn speech without information coming from the sensory organs, since words cannot cling to one another and stick only to each other (... and here I remembered knowledge systems ...). Be sure to receive some information besides the words, so that they have something to associate with in memory.
     
    Perception

    As noted earlier, it is not yet possible to realize any of the human senses, unless only externally. So, we need to come up with a completely new body that will provide information about the outside world.
     
    I admit, at first I had an idea to use a virtual world created on a computer in the form of a three-dimensional interactive model. A kind of Matrix for AI. Moreover, there are many of them now, of which only one Half Life 2 is worth. It will be simpler than ours, and eyesight will be easier to organize. But even here a lot of problems will arise, since we don’t know how a person’s visual memory is organized, and it will take a very long time to find out by trial and error on such a complex model, even for such a relatively simple world.
     
    It is not necessary that the final version of AI be equipped with only one sense organ. At the first stage, the main task of this body is to help develop and configure learning, memory, and other mechanisms to match human ones. From here the main requirements to it follow.
     
    It should be simple, but giving ample opportunities in interaction with AI. It should not distract from work on tuning the memory by its complexity, forcing it to deal with cumbersome recognition algorithms and interpretation methods for subsequent storage instead. At the same time, the technologies on the basis of which it will be built should be sufficiently studied and developed so that it is not necessary to invent something to use it.
     
    And then one sunny day after one and a half kilometers and one and a half albums, I realized what it should be. Vector graphics.
     
    Yes, yes, yes, it is simple, free and it has a whole bunch of advantages. Although this is a purely computer thing, it can serve as a simplified analogue of human vision and partially even touch. It gives tremendous freedom of action, since it is possible to form a whole environment from different objects with different colors. Thinking a little more, you can create dynamic objects, allowing them to move, change, interact, and give rise to new objects.
     
    Applied to the real world, objects will serve as analogues of toys for a small AI, manipulating with which it will receive various information. These objects can imitate real objects and some of their properties, so that AI can be taught their names. In short, the scope for imagination is huge. With just such a system, much more can be achieved than just the development of an annual child. Indeed, in the vector you can organize numbers, maybe even music ...
     
    Well, now the portrait can be considered complete. Outwardly, our humane Artificial Intelligence will look like a window in which there is an area for drawing, a field for entering text, buttons ...
     
    It may look like a hybrid of Corel Draw and Tamaguchi, but in fact it is a wonderful platform for testing and finalizing various ideas, which, I must say, are already numerous.
     

    Epilogue

    It is dedicated to those who have read to the end and have already forgotten what was at the beginning.
     
    This is not a science fiction idea or forecast. This concept was conceived and developed in order to be implemented. I do not deny that AI can be implemented in other ways. Just in my opinion, which I as detailed as possible, other directions will require much more costs both in time and in money.
     
    In principle, the main part of what is written in this article was formed several months ago. So most of the time I was doing what I was trying to understand how Artificial Intelligence should work (see Legend), how its systems should be arranged, what methods and sciences will help to figure this out. Despite the fact that it is still far from ready-made models, there are still many considerations and ideas. If readers have such a desire, I will write an article on this topic.

    Table of Contents:

    The History of Artificial Intelligence, Part 1. Painting without an artist.

    Artificial Intelligence History, Part 2. Neural Network AI - Inevitable or Impossible?

    Making Artificial Intelligence

    Also popular now: