FUTURE TYPES OF MACHINES THAT THINK

The machines that already exist show that some processes of thinking can already be performed very quickly:

We can expect other processes of thinking to come up to high speed through the further development of thinking machines.

Automatic Address Book

Nowadays when we wish to send out announcements of an event, like going to South America for a year, we may copy the addresses of our friends onto the envelopes by hand. In the future, we can see our address book as a spool of magnetic tape. When we wish to send out announcements, we put a stack of blank envelopes into the machine that will read the magnetic tape, and we press a button. Out will come the envelopes addressed.

If we wish to select only those friends of ours whose last names we put down on a list, we can write the list on another magnetic tape, place it also in the machine, and set a few switches. Then the machine will read the names on the list, find their addresses in the address-book tape, and prepare only the envelopes we want. If a friend’s address changes, we can notify the machine. It will find his old address, erase it, and enter the new address.

Automatic Library

We can foresee the development of machinery that will make it possible to consult information in a library automatically. Suppose that you go into the library of the future and wish to look up ways for making biscuits. You will be able to dial into the catalogue machine “making biscuits.” There will be a flutter of movie film in the machine. Soon it will stop, and, in front of you on the screen, will be projected the part of the catalogue which shows the names of three or four books containing recipes for biscuits. If you are satisfied, you will press a button; a copy of what you saw will be made for you and come out of the machine.

After further development, all the pages of all books will be available by machine. Then, when you press the right button, you will be able to get from the machine a copy of the exact recipe for biscuits that you choose.

We are not yet at the end of foreseeable development. There will be a third stage. You will then have in your home an automatic cooking machine operated by program tapes. You will stock it with various supplies, and it will put together and cook whatever dishes you desire. Then, what you will need from the library will be a program or routine on magnetic tape to control your automatic cook. And the library, instead of producing a pictorial copy of the recipe for you to read and apply, will produce a routine on magnetic tape for controlling your cooking machine so that you will actually get excellent biscuits!

Of course, you may have other kinds of automatic producing machinery in your home or office. The furnishing of routines to control automatic machinery will become a business of importance.

Automatic Translator

Another machine that we can foresee would be used for translating from one language to any other. We can call it an automatic translator. Suppose that you want to say “How much?” in Swedish. You dial into the machine “How much?” and press the button “Swedish,” and the machine will promptly write out “Hur mycket?” for you. It also will pronounce it, if you wish, for there would be little difficulty in recording on magnetic tape the pronunciation of the word as spoken by a good speaker of the language. The machine could be set to repeat the pronunciation several times so that the student could really learn the sound. He could learn it better, probably, by hearing it and trying to say it than he could by using any set of written symbols.

Automatic Typist

We now come to a possible machine that uses a new principle. This principle is that of being able to recognize signs. This machine would perceive writing on a piece of paper and recognize that all the a’s that appear on the paper are cases of a, and that all the b’s that appear on it are instances of b, and so forth. The machine could then control an electric typewriter and copy the marks that it sees. The first stage of this machine would be one in which only printed characters of a high degree of likeness could be recognized. In later stages, handwriting, even rather illegible handwriting, might be recognizable by the machine. We can call it an automatic typist.

The elements of the automatic typist would be the following:

1. Phototubes (electronic tubes sensitive to the brightness of light), which could sense the difference between black and white (these already exist).

2. A memory of the shapes of 52 letters, 10 digits, and punctuation marks. Fine distinctions would be required of this memory in some cases—like the difference between the numeral 5 and the capital letter S.

3. A control that would cause the machine to tune itself, so that a good matching between the marks it observed and the shapes it remembered would be reached.

4. A triggering control so that, when the machine had reached good enough matching between its observations and its memory, the machine would proceed to identify the marks, read them, and transfer them.

5. An electric typewriter, which would respond to the transferred instructions. (This also already exists.)

This machine is perhaps not so farfetched as it might seem. During World War II, gun-aiming equipment using the new technique radar reached a high stage of development. Many shots that disabled and sank enemy ships were fired in total darkness by radar-controlled guns. On the glowing screen in the control room, there were two spots, one that marked the target and one that reported the point at which the gun was aimed. These two spots could be brought almost automatically into agreement. In the same way, a report from a phototube telling the shape of an observed mark and a report from the memory of the machine telling the shape of a similar mark could be compared by the machine for likeness and, if judged enough alike, could be approved as identical.

Even the phrase “enough alike” can be applied by a machine. During World War II, tremendous advances were made in machinery for deciphering enemy messages. Machines observed various features and patterns in enemy messages, swiftly counted the frequency of these features, and carried out statistical tests. Then the machines selected those few cases in which the patterns showed meaning instead of randomness.

A machine like the automatic typist, if made flexible enough, would be, of course, extremely useful. A great load of dull office work is now being thrown on clerks whose task is to translate from writing and typing into languages that machines can read, such as punch cards. At the present time, if punch-card machines are widely used in a big company, the company must employ large numbers of girls whose sole duty is to read papers and punch up cards. A still bigger chore is the work of typists in all kinds of businesses whose main duty is to read handwriting, etc., and then copy the words on a typewriter.

Each square in the grill
is watched by a phototube.

Fig. 1. Scheme for distinguishing A and H by 15 phototubes.

Research has already begun on various features of the automatic typist because of its obvious labor-saving value. For example, many patents have been issued on schemes for dividing the area occupied by a letter or a digit into an array of spots, with a battery of phototubes each watching a spot. The reports from the phototubes together will distinguish the letter or digit. For example, if we consider A and H placed in a grill of fifteen spots, 5 long by 3 wide ([see Fig. 1]), then the phototubes can distinguish between A and H by sensing black or white in the spot in the middle of the top row. When we consider how easily and swiftly a human being does this, we can once more marvel at the recognizing machine we all carry around with us in our heads.

Automatic Stenographer

Another development that we can foresee is one that we can call the automatic stenographer. This is a machine that will listen to sounds and write them down in properly spelled English words. The elements of this machine can be outlined:

1. Microphones, which can sense spoken sounds (these already exist).

2. A memory storing the 40 (more or less) phonetic units or sounds that make up English, such as the 23 consonant sounds,

pblng
fvmth
tdnr
szhy
kg w
chj
shzh (heard in “pleasure”)

and the 17 vowel sounds,

LongShortOther
A (“ate”)a (“cat”)ar (“are”)
E (“eat”)e (“end”)aw (“awe”)
I (“isle”)i (“in”)er (“err”)
O (“owe”)o (“on”)ow (“owl”)
U (“cute”)u (“up”)oi (“oil”)
OO (“roof”)oo (“book”)

3. A collection of the rules of spelling in English, containing many statements like

The sound b is always spelled b

The sound sh may be spelled sh (ship), s (sugar), ti (station), ci (physician), ce (ocean) or tu (picture) and other statements based on context, word lists, derivation, etc. These are the statements by means of which a good English speller knows how to spell even words that he hears for the first time.

4. A triggering control so that, when the machine reaches good enough matching between its observations of sounds, its memory of sounds, and its knowledge of spelling rules, the machine will identify groups of sounds as words, determine their spelling, and report the letters determined.

5. An electric typewriter, which would type the reported letters.

With this type of machine, you would dictate your letters into a machine (now existing) that would record your voice. Then the record would be placed on the automatic stenographer, and out would come your letters written and spaced as they should be.

Automatic Recognizer

We can foresee a recognizing machine with very general powers. Suppose that we call it an automatic recognizer ([see Fig. 2]). It will have the following elements:

1. Input. This element will consist of a set of observing instruments, capable of perceiving sights, sounds, etc. There will be ways of positioning or tuning these instruments.

2. Memory. This element will store knowledge. It may store the patterns of observations that we are interested in; or it may store general rules on how to find patterns of observations that we will be interested in. It will contain knowledge about acceptable groups of patterns, about actions to be performed in response to patterns, etc.

3. Program 1. The element “Program 1” performs a set of standard instructions. Under these instructions, the machine:

Compares group after group of observations with the information in the memory.

Compares these groups with patterns furnished, or seeks to organize the observations into patterns.

Counts cases and tests frequencies.

Finds out how much matching with patterns there is.

Tunes the observing instruments in ways to increase matching.

Fig. 2. Scheme of an automatic recognizer.

4. Program 2. The element “Program 2” performs another set of standard instructions. Under these instructions, the machine, if it is tuned well, matches sets of observations one after another with the patterns and so reads them.

5. Triggering Control. This element shifts the control of the machine from Program 1 to Program 2. It does this when the machine reaches “good matching.” We shall set the meaning of this into the machine in much the same way as we set “warm” into a thermostat.

6. Output. This element performs any action that we want, depending on recognized patterns read and any other knowledge or instructions stored in the memory.

The automatic recognizer will be capable of extraordinary tasks. With microphones and a large memory, this type of machine would be able to hear a foreign language spoken and translate it into spoken or written English. With phototubes and with an expanded filtering and decoding capacity as in deciphering machines, the automatic recognizer should be able to read a dead language, even those (such as Minoan or Etruscan) that have so far resisted efforts to read it. The machine would derive rules for the translation of the language and translate any sample.

An automatic recognizer could perhaps be equipped with many sensitive, tiny observing instruments that could be placed around or in the brain and nervous systems of animals. Then the machine might enable us to find out what activity in the nervous system corresponds with what activity in the animal.

TYPES OF PROBLEMS THAT MACHINES
WILL SOLVE IN THE FUTURE

We turn now to the second question regarding the future of machines that think: What types of problems can we foresee as solved by these machines?

Problems of Control

Probably the foremost problem which machines that think can solve is automatic control over all sorts of other machines. This involves controlling a machine that is running so that it will do the right thing at the right time in response to information. For example, suppose that you are mowing a lawn with a mowing machine. You watch the preceding strip so as to stay next to it. You watch the ends of the strips, where you turn around. If a stick is caught in the cutting blade, you stop and take it out. Now it is entirely possible to put devices on the mowing machine so that all these things will be taken care of automatically. In fact, in the case of plowing a large field, a tractor-plow can be equipped with a device that guides it next to the preceding furrow. Thus, once the first furrow around the edge has been made, riderless tractors will plow a whole field and stop in the middle.

For another example, take a gas furnace for heating steam to keep a house warm. Such a furnace has automatic controls, which respond to the following information whenever reported:

In fact, your own meaning of “warm” can be put into the control system: you set the dial on your thermostat at the temperature that “warm” is to be for you.

In the future many kinds of automatic control will be common. We shall have automatic pilots for flying and landing airplanes. We shall have automatic missiles for destructive purposes, such as bombing and killing, and for constructive purposes, such as delivering mail and fast freight. An article in the magazine Fortune for November 1946 described the automatic factory ([see Supplement 3]). This is a factory in which there would be automatic arms for holding stuff being manufactured, and automatic feed lines for supplying material just where it is needed. All this factory would be controlled by machines that handle information automatically and produce actions that respond to information.

This prospect fills us with concern as well as with amazement. How shall we control these automatic machines, these robots, these Frankensteins? What will there be left for us to do to earn our living? But more of this in the next chapter.

Problems of Science

Other problems for which we can foresee the use of machines that think are the understanding, and later the controlling, of nature. One of these problems is weather forecasting and weather control.

The Weather Brain

We can imagine the following type of machine—a weather brain. A thousand weather observatories all over the country observe the weather at 8 a.m. The observations are fed automatically through a countrywide network of communication lines into a central station. Here a giant machine, containing a great deal of scientific knowledge about the weather, takes in all the data reported to it. At 8:15 the weather brain starts to calculate; in half an hour it has finished, having produced an excellent forecast of the weather for the whole country. Then it proceeds to transmit its forecast all over the country. By 8:50 every weather station, newspaper, radio station, and airport in the country has the details. In October 1945, Dr. V. K. Zworykin of the Princeton Laboratories of the Radio Corporation of America proposed solving the problem of weather forecasting in this way by a giant brain.

The weather brain will have a second stage of application. From time to time and here and there, the weather is unstable: it can be triggered to behave in one way or another. For example, recently, pellets of frozen carbon dioxide—often called Dry Ice—have been dropped from planes and have caused rain. In fact, a few pounds of Dry Ice have apparently caused several hundred tons of rain or snow. In similar ways, we may, for example, turn away a hail storm so that hail will fall over a barren mountain instead of over a farming valley and thus protect crops. Or we may dispel conditions that would lead to a tornado, thus avoiding its damage. Both these examples involve local weather disturbances. However, even the greatest weather disturbances, like hurricanes and blizzards, may eventually be directed to some extent. Thus the weather may become to some degree subject to man’s control, and the weather brain will be able to tell men where and when to take action.

Psychological Testing

Another scientific problem to which new machinery for handling information applies is the problem of understanding human beings and their behavior. This increased understanding may lead to much wiser dealing with human behavior.

For example, consider tests of aptitudes. If you take one of these tests, you may be asked to mark which word out of five suggested ones is nearest in meaning to a given word. Or your test may be 40 simple arithmetical problems to be solved in 25 minutes. Or you may be given a sheet with 20 circles, and be asked to put 3 dots in the first, 7 dots in the second, 4 dots in the third, 11 dots in the fourth, and so on, irregularly; you may be given a total of 45 seconds to do this as well as you can. Now, if a vocational counselor gives you one of these tests, and if you get 84 out of 100 on it, he needs to know just what he has measured about you. Also, he needs to know whether he can reasonably forecast that, as a result of your grade of 84, you will be good at writing articles, or good at supervising the work of other people, or good at designing in a machine shop. He needs to know the records of people with scores of about 84 on this test and to have evidence supporting his forecasts.

If we wish to make the most use of the tests, we need to carry out a good deal of statistics, mathematics, and logic. For example, it will turn out that answers to some questions are much more significant than answers to others, and so we can greatly improve the quality of the tests by keeping only the more significant questions. Powerful machinery for handling calculations will be very useful in the field of aptitude testing.

But, you may ask, what if the person analyzing your answers has to use interpretations and judgments? If the judgments and interpretations can be expressed in words, and if the words can be translated into machine language, then the machine can carry out the analysis. Usually the difference between a rule and a judgment is simply this: a rule in a case in which it is hard to express all the factors being considered is called a judgment.

Psychological Trainer

It is conceivable that machines that think can eventually be applied in the actual treatment of mental illness and maladjustment. Consider what a physician does. In treating a psychiatric case, such as a neurosis, a physician uses words almost entirely. He asks questions. He listens to the patient’s answers. Each answer takes the physician nearer and nearer to a diagnosis. By and by the physician knows what most of the difficulty is. Then he must present his knowledge slowly to the patient, gradually guiding the patient to understanding. It is a psychological truth that telling a man in ten minutes what is wrong with him does not cure him. The physician seeks to free the patient from the tormenting circles of habit and worry in which he has been trapped. Often the diagnosis is short and the treatment is long; the reasons for the neurosis may soon be clear to the physician, but they may take months to become clear to the patient.

Now let us consider the following kind of machine as an aid to the physician. We might call this kind of machine a psychological trainer, for in many ways it is like the training machines used in World War II for training a pilot to fly an airplane. The psychological trainer would have the following properties:

1. The machine is able to show sound movies—produce pictures and utter words.

2. It is able to put before the patient: situations, problems, questions, experiences, etc.

3. It is able to take in responses from the patient.

4. It is able to receive a program of instructions from the physician.

5. Depending on the responses of the patient and on the program from the physician, the training machine can select more material to put before the patient.

6. The training machine produces a record of what it presented and of how the patient responded, so that the physician and the patient can study the record later.

What sort of films would the machine hold? The machine could be loaded with a number of films which would help in the particular type of neurosis from which the patient was suffering.

What sort of responses could the patient make? The patient might have buttons in front of him which he could press to indicate such answers as:

YesI don’t knowRepeat
NoIt dependsGo ahead

Also, the patient might hold a device—like a lie detector, perhaps—which would report his state of tenseness, etc., and so report what he really felt.

Where would the machine’s questions come from? From one or more physicians very clever in the treatment of mental illness.

Suppose that the patient is inconsistent in his answers? The machine, discovering the inconsistencies, could return to the subject and ask related questions in a different way. As soon as several questions related to the same point are answered consistently, the machine could exclude groups of questions that no longer apply and could proceed to other questions that would still apply.

Patients would vary in their ability to go as fast as the machine could. So from time to time the machine would ask questions to test the effect of what it had presented; and, depending on the answers, the machine would go faster or would bring in additional material to clarify some point.

This machine might have a few advantages over ordinary treatment. For example, with the machine, treatment does not depend on the physician’s making the right answer in a split second, as it may in a personal interview. Also, the patient might be franker with the machine than with the physician, for it might be arranged that the patient could review his record, and then decide whether to confess it to his physician.

Such a machine would enable physicians to treat many more patients than they now can. In fact, it is estimated that nearly 50 per cent of persons who consult physicians are suffering only from mental illness. Such a machine would therefore be a great help.

Problems of Business

Another large group of problems for which we can foresee the use of machines that think is found in business and economics.

For example, consider production scheduling in a business or a factory. The machine takes in a description of each order received by the business and a description of its relative urgency. The machine knows (that is, has in its memory) how much of each kind of raw material is needed to fill the order and what equipment and manpower are needed to produce it. The machine makes a schedule showing what particular men and what particular equipment are to be set to work to produce the order. The machine turns out the best possible production schedule, showing who should do what when, so that all the orders will be filled in the best sequence. What is the “best” sequence? We can decide what we think is the best sequence, and we can set the machine for making that kind of selection, in the same way as we decide what is “warm” and set the thermostat to produce it!

On a much larger scale, we can use mechanical brains to study economic relations in a society. Everything produced in a society is made by consuming some materials, labor, equipment, and skill. The output produced by one man or factory or industry becomes the input for other men, factories, industries. In this way all economic units are linked together by many different kinds and degrees of dependence. The situation is, of course, complicated: it changes as time goes on and as people want different things produced. Economists have already set up simple models of economic societies and have studied them. But with machines that think, it will be possible to set up and study far more complicated models—models that are very much like the society we live in. We can then answer questions of economics by calculation instead of by arguments and counting noses. We shall be able to solve definitely such problems as: “How will a rise in the price of steel affect the farming industry?” “How much money must be paid out as wages and salaries so that consumer purchasing power will buy back what industry produces?”

Machines and the Individual

What about the ordinary everyday effects of these machines upon you and me as an individual? We can see that the new machinery will apply on a small scale even to us. Small machines using a few electronic tubes—much like a radio set, for example—and containing spools of magnetic wire or magnetic tape will doubtless be available to us. We shall be able to use them to keep addresses and telephone numbers, to figure out the income tax we should pay, to help us keep accounts and make ends meet, to remember many things we need to know, and perhaps even to give us more information. For there are a great many things that all of us could do much better if we could only apply what the wisest of us knows.

We can even imagine what new machinery for handling information may some day become: a small pocket instrument that we carry around with us, talking to it whenever we need to, and either storing information in it or receiving information from it. Thus the brain with a motor will guide and advise the man just as the armor with a motor carries and protects him.

Chapter 12
SOCIAL CONTROL:
MACHINES THAT THINK
AND HOW SOCIETY MAY CONTROL THEM

It is often easier for men to create a device than to guide it well afterwards: it is often easier for a scientist to study his science than to study the results for good or evil that his discoveries may lead to. But it is not right nor proper for a scientist, a man who is loyal to truth as an ideal, to have no regard for what his discoveries may lead to.

This principle is now being widely recognized. Many scientists today—both as individuals and as groups, and especially the atomic scientists—are considering the results of their scientific discoveries; and they are sharing in the effort to render those results truly useful to humanity.

It would be easy to leave out of this book any discussion of how machines that think may be controlled, any consideration of how they may be made truly useful to humanity. But that would be hardly right or proper. In concluding a book such as this one, that touches on many aspects of machines that think, we need to consider what can and should be done to make such machines of true benefit to all of humanity.

So, we come to the most important of all our questions: What sort of control over machines that think do we need in human society?