1. There Was No Name Yet
In the introduction to a book published in 1948, Norbert Wiener, a mathematician at MIT, explains what his group had been facing: a set of problems in communication, control, and statistical mechanics common to machines and living tissue. The field had no common terminology, not even a name of its own.
A person coining a name for a field does not usually begin by declaring it nameless. Wiener did. By 1948 the discussions had a wartime history behind them, and several formal meetings of the Macy series had already been held. What the people called their work changed every so often.
So he made a name. He said it was a new Greek formation from the word for a steersman, to designate the theory of control and communication in the animal and the machine. He connected it to Maxwell's paper of 1868 on governors, saying that governor came by way of Latin from the same Greek root. He also noted that a ship's steering gear was an early mature instance of a feedback mechanism. The first thing he did in making a new word was give it a road back.
On the same page he dated the word to the summer of 1947. The form was not without precedent. In 1834 André-Marie Ampère had used its French form in a work on the classification of the sciences, for a proposed science of government.
One can see the shape of the enterprise in those years in the titles of the meetings.
From 1946 the group met in New York once or twice a year. The first meeting, March 8–9, 1946, was formally titled “Feedback Mechanisms and Circular Causal Systems in Biological and Social Systems.” The second, October 17–18 that year, used “Teleological Mechanisms and Circular Causal Systems.” The third was in March 1947. The fourth, October 23–24, 1947, changed the name again to “Circular Causal and Feedback Mechanisms in Biological and Social Systems.” The fifth was in March 1948.
When the title changed for the fourth meeting, the new word was three months old. It was not in the official title. There is a lag between making a name and taking it up, and in the meantime the work goes on.
The sixth meeting, in March 1949, kept the long name. Not until the seventh, in late March 1950, did cybernetics appear in the general title. The book had been out a year and a half. Meetings eight through ten kept the same title. The last was in April 1953.
A field had held four meetings before it had a name.
The book did well. By the end of 1949 it had sold around twenty thousand copies and had been reprinted three times in its first six months. A monograph on feedback and control had become a bestseller.
2. Nerve Nets
In December 1943 a journal of mathematical biophysics published a nineteen-page paper, “A Logical Calculus of the Ideas Immanent in Nervous Activity,” by Warren McCulloch and Walter Pitts. McCulloch had just turned forty-five. Pitts was twenty, had no degree, and knew logic. He had studied logic and mathematics informally in Chicago for some time. He entered McCulloch's circle around 1942 and for a while lived in his house.
Neurons are cells in the brain. In the model, a neuron receives a stimulus from other neurons. If it is enough, the cell fires; otherwise it does not. What is recorded is whether it fires, not how strongly. The junctions between neurons are called synapses. Some urge a neuron to fire, and others hold it back. The point the stimulation has to reach is the threshold.
The first formal assumption at the start of the paper is that the activity of a neuron is an all-or-none process. Four more follow. A fixed number of excitatory synapses have to be active within a period of what they call latent summation. Synaptic delay is the only delay that matters to the model. An inhibitory synapse, when active, prevents the neuron from firing altogether. The structure of the net does not change with time.
They did not simply say a neuron was a logic gate. They began by arguing that because the event was all or none, the activity of a single neuron could be represented by a proposition. They then required relations among the physiological events to correspond to logical relations among propositions. A neuron's response corresponds to a simple proposition, and the arrangement of synapses, thresholds, and inhibition determines whether the propositions are joined by conjunction, disjunction, or negation.
At the end they joined this to the paper from seven years earlier. Equip a net with a tape, a scanner connected to its incoming side, and a suitable output, they said, and the only numbers it can compute are those a Turing machine can compute. The converse holds too. In the same passage they called this a psychological justification of Turing's definition of computability, setting it beside Church's and Kleene's.
The paper on undecidability had been in print only seven years when it was brought over to the side of nerves as a point of reference. The direction was reversed. That paper stripped a machine out of the figure of a person. This one took that machine as a measure of nerve activity.
They added a qualification. A net with feedback loops but no tape or scanner can compute only some of what a Turing machine can, not all of it. They also expressly listed two difficulties for the initial formalization: facilitation and extinction on the one hand, learning on the other. The former temporarily changes how the same part of a net responds to a stimulus. In the latter, simultaneous activity in the past changes the net permanently, so that a stimulus once too weak to elicit a response later suffices.
Their method was to replace the changing nets with equivalent virtual nets whose connections and thresholds were fixed. Right after proposing it they gave a warning. Formal equivalence of this kind is not an explanation of what actually happens.
The sentence is unambiguous. They knew they were making a substitution, not giving an explanation.
Where the model would not go through, they made a stand-in that would and marked it as such beside the result. People have gone on doing the first part. The second is often dropped.
Their account of the physiological processes proper was not the same as the fixed-net model. They attributed facilitation and extinction to continuous changes in threshold, mentioning electrical and chemical variables, afterpotentials, and ion concentrations. Learning was a lasting change, one that persisted through sleep, anesthesia, convulsions, and coma.
When they came to learning later in the paper, they altered the assumption again. Actual learning seemed to require a permanent change in structure. The simplest mechanism they gave was the formation of new synapses or an equivalent local decrease in threshold. They then used a theorem to show that variable synapses could formally be replaced by circuits with loops.
The fifth basic assumption says the net does not change with time. Later in the same paper they say actual learning requires structural change. The way through is not to declare that the physiological structure does not really change. It is to prove that some changing nets can be simulated in formally equivalent ways by a larger fixed virtual net.
When discussing relative inhibition, they replaced a continuously changing threshold with a series of discrete thresholds and noted that only the approximation to the nearest integer mattered for excitation of the natural type. They supplied the qualification approximation themselves.
After a general realizability theorem, they said the proof was very rough for lack of space. Applying the condition in practice would mean examining a huge class of functions, so they offered a sufficient condition that was easier to use and which they thought would probably cover most practical purposes. Rough, probably, practical purposes: three restrictions in one place.
Nowhere in the paper did they say they had a complete model of the brain, call it an artificial brain, or say they had invented something. Their terms were nerve net, formal equivalence, calculus, virtual net.
3. In the Same Paragraph
On June 30, 1945, the first draft of an account of the structure of an electronic computer was completed. In Section 4.2, right after describing the neurons of higher animals, the author writes that he is following Pitts and McCulloch. He then gives the paper's title, journal, volume, year, and pages.
In the next few sentences he says he will disregard some of the more complicated aspects of a neuron's function. He lists them: thresholds, summation in time, relative inhibition, and aftereffects lasting beyond synaptic delay. Then he compares the neuronal function simplified in this way with telegraph relays and vacuum tubes.
The citation and the stripping away happen in the same paragraph.
He does not say the paper is wrong, or those complications do not exist, or he knows how to deal with them. He says he will leave them out of what follows.
The four things he names in order to leave them out are the very things the earlier authors had repeatedly marked as difficulties. They said the things were there and formal equivalence did not explain them. Their reader says the things are there and he will go around them.
Neither side conceals what it is doing. When the construction advances, what is pushed out of the way is often not forgotten. It is named and set aside.
What is set aside does not disappear by itself. It waits where it was for someone to come back and get it. Fewer people come back, because the road around it grows wider and wider.
The neuron compared with relays and vacuum tubes in the 1945 manuscript is the simplified one. Later accounts often say computers were built in the likeness of the brain. Four things had already been taken out of the brain in the manuscript.
Here the object is called an automatic computing system, or a high-speed automatic digital computing system. There is no brain in the name, no intelligence, just automatic.
4. Usually Thought to Require Thinking
On March 9, 1949, in New York, Claude Shannon of Bell Laboratories read a paper to a national convention of radio engineers. It came out in the Philosophical Magazine in March 1950 as “Programming a Computer for Playing Chess.” A year and a half earlier he had published a paper on communication in another journal, in two installments. It dealt with measures of information, entropy, channel capacity, and redundancy.
The paper opens with four reasons for choosing chess. They are not all the same kind of reason.
The first is the structure of the problem. The permitted moves and the final goal are clear.
The second is difficulty. It is complicated enough, but not so complicated that it cannot be dealt with.
The short sentence in the third is that playing chess well is “generally considered to require ‘thinking.’”
The fourth is that the discrete structure of chess fits the discrete operations of a digital computer.
The first, second, and fourth describe the thing itself. Only the third describes how people see it. In the original, it is generally considered to be so. That is a statement about reputation, not structure. Shannon puts the two kinds of reason alongside each other without making them one.
This is the first time in the series someone separates the reasons for choosing a problem into two kinds and explicitly says that one has to do with how other people see it.
The second half of the third reason is conditional. If the machine can play well, either people have to allow that a kind of mechanized thinking has taken place or they have to narrow what they call thinking.
Before he begins, the sentence lays out the mechanism this series has been watching. Once a machine does something, either people let it in or they move the door farther back. He does not say which will happen. He sets the two routes out. He puts this in his reasons for choosing the subject, not in his conclusion. He is explaining his choice, not asserting a result.
Then comes the technical part. Starting with about thirty legal moves in a typical position, Shannon gets about a thousand combinations for one move by White and one by Black. Over a game of about forty such turns that makes about ten to the power of 120 possible lines in the whole tree. This shows why searching them all is not workable. Farther on he returns to the thousandfold increase each time both sides have moved to show how even a little selection can greatly reduce blind trial and error.
Then he gives two strategies.
Type A systematically searches all legal moves. At a fixed depth it scores the positions with an evaluation function and works the choices back up the tree. His verdict is direct. The machine would be slow and weak. He works it out on a very optimistic assumption: one microsecond per position. Even looking three moves ahead for each side would take about a billion evaluations and more than sixteen minutes to make a move. No machine of the time was that fast. Even an imaginary one that was would give an unusable result.
Type B does not develop every branch equally. It picks forcing moves and promising lines, extending some lines until the position is fairly stable before evaluating it.
He adds a warning here: “It is not being suggested that we should design the strategy in our own image.” The machine should use its strengths in speed and accuracy and make up for its weaknesses in recognition and analysis.
That sentence fits one of his reasons for choosing the subject. He chose it because people thought it required thinking. When it came to doing it, the point was not to do it in our likeness.
The evaluation function gives a position a numerical estimate of how good it is, made from things one can calculate: relative material, pawn structure, mobility, king safety. Where the search cannot or is not going to push farther, use that approximate score.
The minimax process goes like this in words. When it is your move, choose the branch that makes the final evaluation as high as possible. When it is the opponent's move, assume they choose the branch that makes your evaluation as low as possible. Work back in that way, level by level, to find the move to make now.
Nowhere does the paper say unconditionally that the machine is thinking. The nearest it comes is still the conditional dilemma at the start. Its stated aim is to construct a routine that will let a general-purpose computer of the time play chess. What it gives is algorithms, strategies, evaluations, and estimates.
5. The Paper in the Drawer
In 1948 there was a typescript at Britain's National Physical Laboratory. The title page read “Intelligent Machinery: by A. M. Turing.” It ran to some twenty pages. It was written for people inside the laboratory, where Turing was then on the payroll.
Its opening asks whether a machine might show intelligent behavior. It does not put itself under any of the names for a field that only took hold later.
One kind of thing it describes is what Turing calls an unorganized machine: a large number of similar units connected in a relatively unsystematic, random way. He qualifies the phrase at once. It does not claim to be exact. What one person sees as organized another may see as unorganized. The idea of the last section is already showing through. Different people can classify the same thing differently.
An A-type unorganized machine consists of a large number of similar units. Each has two inputs and one output; the connections can be assigned at random. A finite A-type machine eventually cycles through its states. To make the B type, Turing replaces the connections in the A type with a different kind, for use in the later discussion of training and education.
What he imagines under education is not designing a finished intelligent machine outright. It is making a starting machine that can be changed, then using rewards, penalties, and interventions to change its organization. He likens the process to education and discusses unorganized machines alongside the cortex of a newborn. This went in the opposite direction to another approach of the time. Rather than work out what the machine should do and write the rules in advance, make something that can change and work out how to teach it. He distinguishes a machine out of the factory from an educated one. The first is more like an untrained initial state; the second has had its organization changed through experience and intervention from outside.
He gives a list of areas to try. In the original the items are games, the learning of languages, translation of languages, cryptography, and mathematics. The games include chess, noughts and crosses, bridge, and poker.
These are the things the field would return to for decades.
The heading of the last section of the twenty pages is “Intelligence as an Emotional Concept.”
His point is that whether we take something's behavior to be intelligent depends not only on the thing itself but, to a considerable degree, on the observer's state of mind and training. Once people can explain the behavior, predict it, or see a very simple plan behind it, they become less inclined to call it intelligent. One person can call the same thing intelligent and another can say it is not.
In 1948 one of the people involved wrote down the very mechanism this series has been following.
In the earlier essays the line had to be inferred from the material. Somebody draws a boundary for a machine. The machine does the thing, and the boundary moves a little farther back. This time there is no need to infer it. Someone has written it down.
In these sentences the boundary is not on the machine's side. It is on the side of the person looking, and it moves with what that person knows. Once people see how something works, they no longer count it as intelligent. Someone at the time had already put this in the last section of an internal report.
He goes straight on to an example. Put a person following written rules in another room, so that an opponent receives only the moves. Sometimes the opponent may not be able to tell whether it is the person or the rules on paper on the other side. In parentheses he calls this an idealized version and says he has made experiments along similar lines.
Putting a person who follows rules in a room to stand in for a machine will recur, in another form, two years later.
The report was not published that year. The director of the laboratory is said to have disliked it and called it a schoolboy essay. It first appeared in print in 1968.
Twenty years. The pages saying that intelligence was an emotional concept spent twenty years in a drawer.
In those twenty years the field acquired a name, its first laboratories, a first round of hopeful predictions, and a first retreat. The piece that located the boundary on the observer's side did not take part. By the time it appeared in print, the thing it described had happened several times already.
6. A Sonnet
On June 9, 1949, at the Royal College of Surgeons of England, Geoffrey Jefferson gave the Lister Oration. His subject was “The Mind of Mechanical Man.” Jefferson was a neurosurgeon speaking about his own field. The full text appeared in a medical journal on June 25.
The sentence most often cut short is on the last page. The crucial qualification is that a machine would have to write a sonnet or compose a concerto because of thoughts and emotions it felt, rather than by the chance fall of symbols.
There is another condition in the second half. The machine must not only make the work. It must know that it made it.
He goes on to unfold that condition in a list of subjective experiences: pleasure at success, grief when a valve burns out, a response to flattery, sex, distress at going wrong or not getting what one wants, and anger.
He is not setting a test of form. A sonnet in correct meter will not pass it. There are three levels: the felt thoughts and emotions it comes from, knowing it has written the sonnet, and the whole list of pleasure, grief, distress, and anger.
He has already discussed what machines can do with words, logic, and mathematics. His main line runs through opinions, creative thinking, awareness of what one is creating, and feeling. He is not denying that machines can work sums or handle signs. That part often drops away in quotation. What is left is the sonnet, and the short version sounds as if a machine could not write a poem.
On the following day, June 10, a newspaper ran an account titled “Mechanical Man Has No Mind.” On June 11 the same paper followed up with “From Calculus to Sonnets,” under the large heading “Mechanical Brain,” and another line below: a three-hundred-year-old question had an answer.
Turing's reply appears in the June 11 report. He does not think one can draw the line even at a sonnet, he says, though the comparison may be a little unfair.
What he doubts is the use of the sonnet as a line that cannot be crossed. He adds a joke: another machine may be better placed to appreciate a machine's work. The following year, in a philosophy journal, he quotes the lecture's sonnet-and-concerto argument when he takes up the objection from consciousness.
Another two and a half years pass. In January 1952 four men recorded a radio program, “Can Automatic Calculating Machines Be Said to Think?” The Third Programme broadcast it on January 14. The participants were Newman, Turing, Jefferson, and Braithwaite.
The title draws a precise verbal boundary. The things in question are still called calculating machines. What they argue over is the predicate: can those machines be said to think? The program does not start by naming them thinking machines.
At one point Braithwaite asks whether a machine can really learn by analogy. Turing says, “I certainly left a great deal to the imagination.” It is still an admission of a gap in the design. The education plan set down three years earlier is still on paper.
7. Things That Were Built
Around March 1948, at a hospital in Gloucestershire, Ross Ashby made a device.
It consisted of four coupled electromechanical units, made from parts of Air Force bombing-control equipment available after the war. Each unit had movable magnetic and electrical controls and a critical variable that could be affected by the output of the other units. When a parameter went beyond the preset region of stability, stepping and selector mechanisms changed the parameters of the connections until the whole system settled back into a stable state.
On January 24, 1949, an American weekly reported on it under the title “The Thinking Machine.” It was not yet a year old. A low black device, it had four square electrical units in a group. What one mainly saw in motion was magnets and indicators.
After an external or internal disturbance, it could rearrange its parameters and find a new stable configuration. It was not a computer that could handle language, play chess, or do arithmetic. There were no instructions inside it that could be swapped out. Its powers were in the wiring. The device was demonstrated and discussed at the ninth meeting in March 1952. The same year Ashby published a book of about 260 pages, Design for a Brain: The Origin of Adaptive Behaviour.
Around the same time, at a neurological institute in Bristol, Grey Walter made two little things that could move about, named Elmer and Elsie. He worked on them between Easter 1948 and Christmas 1949. He gave the class of machines a mock-Latin species name that meant something like machines that look around. The nickname tortoises came from their low shells and the way they crept.
The design was a wheeled chassis, a few vacuum tubes, a photocell, steering and drive motors, a contact in the shell to register collisions, and a battery. The transparent or translucent cover gave the creature the shape of a tortoise and served as an input for touch when it hit something. What Walter could get his hands on was the vacuum tubes and motors available after the war.
Those few controls combined to produce a surprising amount of behavior: approach a light, search, change direction after a collision, sway and veer away on getting close to a light source. What Walter wanted to show was how complex a behavior a few components like parts of a nervous system could produce. He was not trying to get a machine to follow a symbolic program and reason in general. It did not do sums, play chess, read letters, or remember anything. It just moved around the room, in a way that made it seem alive.
In May 1950 a popular science monthly published his “An Imitation of Life.” He had given public demonstrations in 1949 and 1950. Three later tortoises were shown at the national exhibition in 1951, where the catalogue still gave the mock-Latin name. One from that group survives in a museum today. It is about the size of a cat.
In November 1951 Dietrich Prinz's program ran on a commercial machine in Manchester. It solved mate-in-two problems, not a full game of chess, and used exhaustive search. The rules were simplified to accommodate the machine. It did not handle castling, en passant, or promotion. It had no visual board. The position went in on punched tape and the solution came out in print. In 1952 Prinz published an account under the title “Robot Chess.”
Christopher Strachey originally wrote his checkers program for another machine. In 1951 an early version did not have enough storage to finish a game. He moved the work to the commercial machine in Manchester and by the summer of 1952 it could play a full game. The operator moved pieces on a real board according to the output.
Go back another four years. On June 21, 1948, Manchester's small experimental machine successfully ran a set of instructions stored inside it for the first time. Keeping the task as data inside the machine, so it could be changed as needed, was what distinguished this class of electronic computer from devices whose instructions had to be arranged outside. In February 1951 the commercial machine was delivered to the University of Manchester.
8. A List of Names
Line up the names used between 1943 and 1952 and each brings a boundary of its own. The names are not synonyms. Each governs a different piece of ground. The one name that came later was the result of choosing one road through this collection. The paths not chosen had not been disproved. Fewer people went down them. As the people thinned out, the territory those paths covered receded from view.
Nerve net. What the 1943 paper itself repeatedly calls it. The authors give it strict limits: a net of formal neurons and synaptic relations that can be described in propositional logic and recursive expressions. They do not call it an artificial neural network. That name came later.
Automatic computing system. The term in the 1945 manuscript. It means a calculating device that carries out detailed instructions automatically.
Feedback mechanisms and circular causal systems. The formal title of the March 1946 meeting. Six months later the name with teleological took its place. In October 1947 the title went back to starting with circular causality and feedback. All three concern biological and social systems. All put the shape of causation in the title.
Cybernetics. Coined in the summer of 1947, published in 1948. The author's boundary covers whatever the image of steering can reach, made things and living things alike. He does not limit it to electronic digital computers or to nerve nets.
Electronic brain. A word from the platform and the newspaper, not from the people making the machines. On November 1, 1946, a newspaper ran “An Electronic Brain,” reporting a speech delivered the previous day in a London hotel. The speaker was Lord Mountbatten, a naval officer and that year's president of the British Institution of Radio Engineers. The stage, he said, was set for the development that sounded most like something from Wells: the electronic brain. Wells was the novelist who wrote The Time Machine. In the same speech Mountbatten said the machines actually in use already had memory of a sort. That February in America, newspaper accounts of the first electronic computer were already full of brains, giant brains, and such comparisons.
Six days later, on November 7, the same paper ran a letter under a title calling the electronic brain a misleading term and no substitute for thought. The writer was Douglas Hartree, the man who copied Lovelace's sentence in his 1949 book in the first essay. One sentence in the letter says that the machines can only do precisely what their operators tell them to do.
A name is said aloud from a platform. The next day it is a newspaper headline. Six days later someone says in the same paper that the name is wrong. The 1949 lecture objected to equating the machine and the brain in substance. This letter objects to the name itself.
There is an “only” in the letter. Three years later, it is not in the century-old sentence Hartree copies into his book.
Thinking machine. The title the American weekly gave the report on Ashby's device in January 1949. The same year Edmund Berkeley published Giant Brains, or Machines That Think.
Mechanical man. A phrase in the title of the 1949 lecture. Mechanical brain was on the newspaper page two days afterward.
Intelligent machinery. The title of the internal manuscript of 1948. It did not appear in print that year, so as a public name it had to wait until 1968.
Automata. A word in the title of a presentation at a 1948 conference, not formally printed until 1951.
Computing machinery and intelligence. The title of the 1950 paper. Not artificial intelligence, not machine intelligence.
Automatic calculating machines. The term in the 1952 broadcast. The first part is the same as before. The dispute is all in the predicate.
Before a field had a name, it had at least ten.
The names are distributed unevenly. The builders say automatic computing system or automatic calculating machine. The theorists say nerve net, automata, cybernetics. The papers say electronic brain, mechanical brain, thinking machine. Each group speaks its own language and draws its boundary in a different place. The builders draw it closest in; the newspapers farthest out. The theorists are in between, and do not agree with one another. To draw a boundary anyone will recognize, someone first has to bring these names together into one. That choosing is already a boundary. Who counts as inside and who is left outside is now no longer a matter of description but of belonging.
Neither of the two passages here that put the boundary on the side of the person looking circulated widely at the time.
One is in the last section of those twenty pages. Once people understand a thing, can predict it, and know the method underneath, it no longer looks intelligent. Give the same thing to a different observer and it changes category. These pages were in the drawer for twenty years.
The other is in the first section of the 1950 chess paper, the third reason for choosing the subject: playing chess well is generally thought to require thinking. It is an explanation of the choice in a technical paper. What people quote afterward is not usually that sentence but the number, ten to the power of 120. A number is easier to remember and use than a sentence, and one need not take sides. For decades the number has been brought out to show how difficult the problem is. The sentence says why the problem was picked. It does a different job.
The remainder was neither taken away in those years nor handed over. Two people pointed to it in turn. Once they had spoken, their sentences lay down again in the documents each came from. One was in a drawer for twenty years. One was in the first section of a paper about chess. The remainder cannot be eliminated, but it can go unmentioned for a long time. Not talking about it does not mean it is not there. The evidence it is there is that every few years someone says it again, in different words. In 1843 it was not being able to originate. In 1887 it was not being able to initiate. In 1948 the person looking made the decision. In 1950 it was generally considered to require thinking. These are not the same claim. All four point in the same direction.
In the words Turing, Shannon, Wiener, McCulloch and Pitts, and Jefferson wrote themselves, whether a machine could think always came with a question, a condition, the limits of a definition, or a standard of behavior.
Not one of their sentences stood bare.
The name had not arrived. The argument had been going on for six years. The people arguing had wooden tortoises, a black box made of four electromechanical units, a commercial computer that could solve a mate in two, an internal paper about educating a machine, a paper on programming one to play chess, and a book that had sold twenty thousand copies. The name they wanted would take another four years. When it arrived, some of the ten or so names above went into it, others were pushed aside, and some appeared only in the papers from then on. Nobody carried the two sentences that put the boundary with the observer into the new name.
The machines are still being built. Each generation says again where the boundary lies.