Richard Hamming: Chapter 7. Artificial Intelligence - II

“The goal of this course is to prepare you for your technical future.”

Hi, Habr. Remember the awesome article "You and your work" (+219, 2442 bookmarks, 389k reads)?

So Hamming (yes, yes, self-monitoring and self-correcting Hamming codes ) has a whole book based on his lectures. We translate it, because the man is talking.

This book is not just about IT, it is a book about the thinking style of incredibly cool people. “This is not just a charge of positive thinking; it describes the conditions that increase the chances of doing a great job. ”

We have already translated 23 (out of 30) chapters. And we are working on the publication "in paper."

Chapter 7. Artificial Artificial Intelligence - II

In this book, we mainly touch on the benefits of computers in the intellectual field, and not in the mechanical, for example, in production. In the field of mechanics, computers allow us to produce better, preferable and cheaper products. In some areas, this assistance is very significant, for flights to the moon a lot would be hard to do without computers. AI can be considered as an addition to robotization - it mainly refers to the intellectual side of a person, and not to the physical, although of course both parts work closely together in many projects.

Let's go back to the beginning of the reasoning and re-examine the components of the machine and man. Both machine and man are made up of atoms and molecules. Both the car and the man consist of the main parts; machines, among other things, have accumulation and switching devices (valves), and man consists of organs. Machines have large structures, arithmetic units, memory, control units, input-output devices, and a person consists of bones, muscles, blood vessels, the nervous system, etc.

Consider some aspects more closely. It is known that in large systems new effects may occur. For example, it is believed that there is no friction between molecules, but in large systems this effect is detected - this is an effect that occurs when small parts are organized into a large system.

It is worth noting that when an engineer develops a device identical to that existing in nature, he does it differently. For example, we have airplanes, in general, with a fixed wing, while the birds flap their wings. We also did a little bit more: airplanes definitely fly higher and faster than birds. Nature has not invented the wheel, and man uses it in many devices. Our nervous system is relatively slow, transmitting signals at a speed of several hundred meters per second, transmitting signals to computers at a speed of about 186,000 miles per second.

Third, before further discussion of AI, we note that the human brain consists of many, many components that are connected with each other by nerves. We want to define thinking as a process that can be performed by the human brain. Explanations of past failures in programming thinking of machines consisted in insufficient size of machines, their speed, m, etc. Some began to believe that if the car is big enough, it will automatically think! Remember, it seems that the problem lies in writing the correct program, and not in building a large machine, a new effect will appear in such a program - thinking from unthinking details! Actually, this is all that is thinking! This is not something separate, but an artifact of a large system.
Let's return to the AI ​​applications. There is a program for proving geometric theorems studied in the course of classical geometry. The program was asked to prove the well-known theorem “If two sides in a triangle are equal, then its two angles are equal,” see fig. 7.1.



Probably you are trying to split the upper corner in half, prove the similarity of two triangles and derive from this the equality of two angles. Some may split the third side of the triangle in half, draw a line to the opposite corner, and also come to a conclusion about similarity. The proof that the program produced did not use any additional constructions. When comparing triangles ABC and CBA, it was concluded that self-similarity and equality of angles.
Someone will find this proof elegant, correct and unexpected. It is possible that the people who wrote the program of proving theorems did not know it; it is widely unknown, although it is discussed in the footnotes of textbooks on geometry. Someone said that the program showed "originality", "novelty", which was not incorporated into it by the developers, "creativity" and other remarkable qualities.

A more in-depth analysis shows that programmers laid the algorithm according to which the program first tries to prove the theorem, and then builds additional lines. If you were taught to act the same way, many of you would find this proof elegant? Thus, this evidence has been programmed. As I noted earlier, what is the course on geometry as not loading into the training programs on evidence? Irrelevant. This applies to people, but for the machine it is enough to load the program once and no longer need to repeat endlessly and repeat, and still forget something!

Did Samuel’s chess program show originality when she beat state vhecker champion, making an unexpected move? If not, can you say that you have originality? What should be the test to separate you from the computer program?
Some may argue that the chess program has learned, and the program for proving theorems has shown “creativity,” “originality,” or something else. These are just a couple of examples of similar written programs. It is hard to prove that the program has declared properties; as soon as the program executes something, it immediately correlates with the step of the algorithm embedded in it, even if random numbers are inserted into the program. Thus, the paradox turns out: the fact of the existence of the program confirms that this is a certain mechanical process and denies the manifestations of other properties. From this point of view, a car will never demonstrate that it is more than a “car” in the classical sense; There is no way to demonstrate, for example, that a car can "think."

Proponents of hard AI believe that a person is a machine and, therefore, all possible actions of a person in the intellectual sphere can be copied by machine. As I noted earlier, when a machine demonstrates certain properties, most readers automatically assume that these properties are not manifestations of human features. Two questions immediately arise. First, is it true? Secondly, how confident are you yourself that you are not just a set of molecules in an energy field, and the whole world is just a molecule revolving around another molecule? If you believe in some otherworldly, mystical forces, how do they affect the movement of molecules, and if they do not, then in what way influence our world? Physicists have described all the forces existing in nature, or there are still unexplored forces? That's a very difficult question. (At present, in 1994, it is believed that the Universe is 90-99% composed of so-called dark matter, of which only the ability for gravitational attraction is known.

We now turn to the consideration of computer applications in the field of culture. At the dawn of the computer revolution, Max Mathew and John Pierce from Bell Telephone Laboratory studied the generation of sound using a computer. As it became clear later, the sampling frequency is determined by the maximum frequency of the reproduced sound. People can hear sound with a frequency of up to 18 kHz and only at a young age, adults also talk on the telephone and recognize sound at a frequency of less than 8 kHz. Quantizing a sound music track does not provide great opportunities. The sequence of playing "music" is as follows: the computer calculates the sound track values ​​at each time interval, presents this value as a voltage, and applies a smoothing filter. Pure tone is represented as a simple sine wave. The combination of frequencies determines the instrument and its characteristic “sound” (increase in sound intensity at the beginning and notes and extinction at the end). With the help of various software tools it is possible to get various notes and music recorded for later playback. With this, it is not necessary to create music in real time, the computer can play music with the speed it needs, not necessarily with constant, the actual speed is reached after the end of editing and playback on the audio device.

Why are we only talking about playing notes? Why not teach a computer to compose music? In the end, there are many rules for writing music. Researchers at sound generation taught computers to compose music with the help of such rules and accidental random number generator. Now there are computers that can play music and compose music; There are many similar tracks on the radio and TV. This approach is cheaper, controlled and allows you to create sounds that do not emit any of the existing musical instruments. Any sound from a music track can be created by a computer.

Computers support music playback and creation. With the exception of small details (sampling frequency, the number of quantization levels, which can be increased at the expense of cost), composers have access to listen to any sounds that can only exist at any frequencies, in any combinations, at any tempo and volume. In fact, the “best quality of recorded music” is digital. In the future, there will be no significant technical improvements in this area. Many people have digital players and they reproduce sound much better than old analog players.

The machines help the composer to quickly (almost instantly) hear the written music. Previously, the composer often had to wait for years until he became famous, and his music began to sound from the stage, and not only in his imagination. Now composers can quickly come up with new directions. While reading a magazine about computer music, I got the impression that modern composers widely use and customize computer programs, and that there are many ways to create music with the machine.

There are more opportunities for conductors. In the past, when conducting music, the conductor tried to get the best from the musicians, often the final recording was obtained from various pieces of different recordings, different microphones. Now the conductor can get exactly the recording that he wants, up to milliseconds, to the desired tone, and taking into account the individual concepts about the sound quality of a particular instrument. In the end, musicians do not always perform the same passage (part of a piece of music) every time is perfect.
We will continue to consider the influence of computers, how they take us back from the world of things to the world of ideas, how they complement and empower people.

One of the aspects of AI that I'm interested in is what a person and a computer can do together, and there should be no competition between them. Of course, robots replace many people in terms of performing routine work. Indeed, robots perform routine work much better, while unloading people’s time for more “human” tasks. Unfortunately, many are not ready for such a competition with cars - many can not do anything except this or that routine work. There is a widespread belief that with proper training, people will be able to compete with robots. However, I doubt that you can take, for example, miners and turn them into useful programmers. I have some estimates about the percentage of people who can practice programming in the classical sense; Of course, if we assume that interaction with an ATM or a telephone is programming (data input from a person affects the program being executed), then many can be called programmers. But if programming is considered to be a classic activity on thoughtful analysis, detailed specification, then there are big doubts about the percentage of people who can compete with a computer on a par.

The use of computers has led both to a decrease in the number of jobs and to the creation of new ones, it is hard to say which is more. But it is clear that, on average, more high-level jobs have appeared and many low-level jobs have disappeared. Again, some believe that in the future, most people can be prepared to perform high-level work, it seems to me that this opinion is unproven.

In addition to programs in the field of games, geometry, music, there were programs in the field of algebra: they are more “manageable” than “single” programs and depend on human interaction at a particular stage. It is curious that it was possible to create a “single” program in the field of geometry, and it is impossible in the field of algebra. One of the unsolved problems is the simplification of expressions. When studying algebra, you might not have noticed that for the action “simplify expression” there were no clear rules for “simplification”; and if they did, these rules were obviously complex. For example, the expression is considered



cannot be simplified, but the expression



can!

We constantly use the word “simplify”, but its meaning depends on what we are going to do next. For example, if you are going to further integrate in the calculations, that you break the expression into small parts, at the same time combining them into a convenient product or relation.

A similar “human-controlled” program for the synthesis of chemical compounds was also developed. It turned out to be very useful, as it counted: (1) possible routes of synthesis, (2) cost, (3) time required for the reaction, (4) effective yield of the synthesized compound. Thus, this program helps to open many different ways of synthesizing new compounds or rediscover old ones due to changes in the cost of the necessary raw materials.

The machines were replaced by an unreliable person looking through a microscope during most medical tests. In most cases, the use of machine methods faster, more reliable and cheaper. Perhaps further machines will be able to diagnose and thus replace the doctors. In fact, in this case, the machine will probably diagnose faster than the doctor! There is nothing new in this idea: kits for self-diagnosis of certain diseases were sold. This approach merely represents an improvement in the self-test kit and a method for prescribing a treatment plan.

Doctors are human, therefore, they are unreliable. Often, in the case of rare diseases, the doctor may first encounter such a disease, but descriptions of all diseases can be entered into the machine, and it will never forget them. Based on the symptoms, the machine can either make a diagnosis with some probability, or assign clarifying analyzes for further diagnosis. In the long term, the machine, taking into account the likelihood of diagnoses being set up (which can be adjusted during epidemics), can lead to a better “patient reception” than a doctor of average or even high qualification. It should not be forgotten that a physician can physically treat a limited number of people.

Among others, one of the main problems is legal. The law forgives doctors for making mistakes, if they acted, in legal terms, “with due diligence” - they are just people. But who will be to blame for the machine making a mistake? A machine? Programmer? Experts who formed the rules? Those who formulated these rules in more detail? Those who put them in algorithms? Or those who programmed them? In case of an incorrect diagnosis made by the machine, it is possible to conduct a detailed analysis of the entire program; such an analysis cannot be carried out in terms of making the wrong decision by the doctor. I believe that in the future there will be many auxiliary programs for making a diagnosis by a doctor, but for a long time there will always be a person between the patient and the machine. There is a certain increase in programs that allow you to make a diagnosis on your own, but legal questions arise for such programs.

For example, I doubt that the patient will be able to independently (through the program) write a prescription for the necessary medications without the participation of the doctor. Probably, you noticed that in all licenses for distributed software it is exempt from any, let me emphasize again, from any responsibility! In this area, the main problem will be legal, not engineering.

In modern hospitals one can see a large penetration of machines into the field of medicine: medicine is very aggressively using the capabilities of machines to improve performance, reduce costs, improve accuracy and speed. In hospitals, cars keep records of finances, schedules, and keep records; even private doctors began to work with the use of certain machines. To some extent, this is the fault of the federal authorities, which oblige to conduct bureaucratic correspondence in electronic form.

In many hospitals in the intensive care wards and other, if necessary, installed computer monitors. Machines do not get bored, react quickly, instantly transmit a signal to the nurse's post if necessary. It is doubtful that a nurse would be able to constantly perform the duties that she performs with the monitor.

In the field of mathematics, one of the earliest programs for the symbolic transformation of formulas was a differentiation program for calculating higher order derivatives. With this program it is possible to calculate the first 20 members of the power series of a complex function. As you should know, differentiation is a simple formal problem with a small number of rules.While studying it could not seem so, but it is necessary to distinguish the task of direct differentiation from further simplification and work with derivatives. Another program for working with symbolic formulas was a program for converting coordinates - necessary for controlling satellites, radars, etc.

James Slegle (https://en.wikipedia.org/wiki/James_Robert_Slagle) wrote an analytic integration program. The algorithm of the program is similar to those algorithms that are taught in mathematics courses. The program can compete with the average MIT graduate in terms of integrals that can be solved, and in terms of the correctness and non-redundancy of the algorithms used. Since then, computer integration programs have improved significantly, it was assumed that a program based on the well-known Rish algorithm would appear, which could integrate any function, if possible. However, after many years of waiting, I did not see such a program. There are analytical integration programs that consider finite integrals or prove that the expression cannot be integrated.

Computers in the form of robots have invaded the production lines of complex products such as tablets, etc. Now computers are assembled by robots that are controlled by other computers, chips in integrated circuits are designed primarily by human-controlled computers. No human mind is in a position to qualitatively locate millions of transistors in a chip, this is a hopeless task. Definitely in the design program laid some degree of artificial intelligence. In limited areas where there are no accidents, robots are effective, but where there are unexpected events, robots can face serious problems. A routine response to a non-routine event can lead to disaster.

Another obvious area of ​​application for computers is a robot that is resistant to a more aggressive environment than even people equipped with protective clothing, such as fire. If during the performance of such work the robot is destroyed, then this is not the same as the death of a person. The minesweepers with remote control appeared on the armament of the navy; The fleet constantly uses robots during the passage of deep seas, now there are many non-mined areas at sea.

Let's think again about playing chess with a computer. Robots were constantly becoming more effective and, until they beat the world chess champion, seemed like a waste of time. It used to be that it was necessary to analyze all possible combinations of moves, and not the way a person plays chess. Computers are now analyzing millions of combinations per second, but a person, according to psychological research, analyzes up to 50 maximum 100 combinations to make a decision on a course. Thus, at least, so it is considered, a person thinks completely differently when playing chess! We do not even know how!

In other games, the cars were more successful. For example, I tell everyone about the backgammon robot that was able to beat all the champions. But some games with simple rules, for example, the game of Go, turned out to be very difficult for a machine solution.
To summarize, the machines turned out to be capable of playing games and similar activities, but for some games they were completely incapable. The way the machine plays can be described as “doing a large number of calculations,” and not a game based on understanding. We started teaching computer games to understand human thought processes, but the initial goal was distorted, and research continued towards developing programs that could benefit.

Let me repeat, you can not afford to ignore the field of artificial intelligence, the use of knowledge in any other area can both improve computer applications and lead to a big fiasco!

Now is the time to explain the difference between logical and psychological novelty. In the course of their work does not produce a logical novelty, but they definitely produce a psychological novelty. Programmers constantly notice new effects from the programs they write! But can a person produce a logical novelty? A careful analysis of the stories of great discoveries shows that they were made on the basis of an analysis of past experiences. The circumstances led to success; it's just a psychological, not a logical novelty. Do not all your new achievements come from past experience? Is logical novelty possible in its purest form?

No need to think that logical novelty is something banal. In any science, after the initial assignment of postulates, definitions, and logic, all further conclusions are a psychological novelty; in conclusions after the initial assignment there is no logical novelty!

It is believed that if we use a chance generator when making decisions, we can break the vicious circle of two molecules, and where can we get a random source of events, if not from the material world with molecules?

It is claimed the information is contained in the source of random events. This statement is based on mental experience with monkeys and typewriters. Monkeys are in front of typewriters and at random times click on random buttons. It is believed that at some point in time one of them will be able to print all the books from the British library in the order in which they stand on the shelves! Sooner or later the monkey will correctly press the first letter, in fact, with an infinite number of attempts to happen infinitely often.

Among these attempts will be correct with the second letter, and so on. With the endless working hours of the monkeys, the right combination of symbols will arise.

Experience is proof that knowledge is contained in random event sources, and you can extract them by writing the correct program to “recognize” the information. For example, sooner or later a new physical theory will arise from the noise stream, and you will be able to detect it if you filter the stream of random numbers! This logic is incontrovertible — it's hard to believe in such a reality! The truth is that you cannot always recognize “information” even if you see it.

It has long been considered that "free will" is a myth. Under given circumstances and at a given time, you are who you are and act the way you act. This argument sounds convincingly against the background that you believe in your free will. Let's try to clarify this issue with the help of experiment. It seems that you can not put a convincing experiment to illustrate this problem. The truth is that we constantly choose between the two lines of behavior. The teacher has to believe that he says the right words to his students. Parents have to believe that they are raising their children properly. Still, the feeling of freedom sits deep in us, and we reluctantly refuse it, but we can easily deny from others!

There are many more examples to discuss the question “Can machines think?”

In conclusion, perhaps thinking should be determined not by what we think, but by how we think. When I observe how a child learns to multiply two or three-digit numbers, I get the impression that he is thinking. When I do it myself, it looks like a certain "conditioned reflex." When a computer does the same multiplication, it does not seem to be thinking. In the words of the old song, "this is not what you do, but how you do." Perhaps in the field of thinking, we confuse the notions of the result and the method of production, which leads to some difficulties in the field of AI research.

Proponents of a rigid theory of artificial intelligence consider this or that theory successful only after confirmation. Such an approach without a proper assessment of the facts has affected many researchers. Believing that “the result is a measure of thinking” allows many people to believe that they can think, but the machine does not.

The situation regarding computers and thinking has become ambiguous. At the same time, we can assume that computers may or may not think. We want to believe in the thinking of machines, because they help us. And do not want to believe in order to preserve their own value. Computers have bypassed us in many areas, speed, accuracy, reliability, cost, reaction speed, they are free from boredom, easily forget the old and learn new things, can work in an aggressive environment, free from personal problems, we would like to exceed them in some areas - they are our creations! For example, if some computer program can do a job much better than doctors, then where will they stay? And where will you stay?

Two main conclusions:

1. , , .
2. , , , .

In the two previous chapters, I wrote the question of the limits of hardware and software, but in these two chapters, questions of artificial intelligence are addressed superficially. We just do not know what we are talking about; concepts are not defined and it is not clear whether they will be determined in the near future. We are forced to use a language to describe the computer language; this recursiveness leads to complexity and a decrease in severity. The main question of the chapters on AI - the question of the limits of software - remains open to be very important in your career. The study of the field of AI requires accuracy and weighted conclusions, as many researchers have made obvious wrong conclusions.

To be continued...

Who wants to help with the translation, layout and publication of the book - write in a personal or mail magisterludi2016@yandex.ru

By the way, we also launched another translation of the coolest book - “The Dream Machine: The History of Computer Revolution” )