1. Four Predictions
November 14, 1957, the Penn-Sheraton Hotel in Pittsburgh, the national meeting of an operations research society. There was a speech at dinner. Herbert Simon gave it. The paper he read was the work of two people, Simon and Allen Newell.
Both names appeared in the preceding essay. They wrote the paper brought to Hanover in the summer of 1956, about a program that found proofs of logical propositions. At the time, the computer implementation was still being coded. This time their subject was not one program. It was something larger: where a field should go next. The room was full of people who did operations research.
The trade of operations research then was to make work more economical: schedule staff, schedule production, schedule transport. It relied on algorithms, mostly procedures one could follow step by step. What Simon had to tell the people in the room was that there were problems with no ready-made method of solution, and a machine could still take a run at them.
The talk was published the following year in an operations research journal under the title “Heuristic Problem Solving: The Next Advance in Operations Research.” The word heuristic will come up a great deal in this essay. It means a method that does not guarantee an answer. Instead of following every possibility to the end, try the leads that look promising. If one does not work, go back and try another. Much of human problem solving goes this way.
Simon starts with three statements about the present. First, digital computers can already take on some problem-solving tasks for which no algorithm is available. Second, their processes closely parallel the ones people use to solve problems. Third, within limits, machines can improve their performance through experience. He specifies that this does not merely mean storing successful patterns; it includes processes by which the machine rewrites its own program.
All three are about things that have already happened. They are the ground for the four that follow. The third says a machine can improve through experience. Ten years later the author of the checkers paper would write in his own follow-up that getting the machine to come up with a new way of seeing things for itself was still unresolved. The four predictions had to cover the distance from some tasks for which no ready algorithm was available to a world champion.
Then come the four. On the basis of these developments and given the speed of progress in the field, he says, he is willing to make the following predictions, all to come true within ten years. They have one deadline: within ten years.
First: a digital computer will become world chess champion. This comes with an exception: unless the rules bar it from competing.
Second: a digital computer will discover and prove an important new mathematical theorem on its own. Two things are required, discovery and proof, not simply checking a proof a person supplies. That can be done by algorithm.
Third: a digital computer will write music that critics accept as having considerable aesthetic value. The criterion sits on the critics' side, not the machine's. Writing notes is not enough. The critics have to say yes.
Fourth: most theories in psychology will take the form of computer programs or of qualitative statements about the characteristics of computer programs.
Three predictions give the machine a particular achievement; one gives a discipline a new shape. In their reply to a critic the following year, the two authors say the best time to judge would be around January 1, 1968.
Give a prediction a deadline and the books can be checked. All four have a deadline, and it is the same one. That is rare in this series. What people said in the earlier essays usually came without a date: cannot originate, cannot initiate, undecidable, possible in principle. A statement with a deadline is a different kind of thing. It hands itself over to the calendar.
2. Thinking Machines
The same speech says other things, and it does not take the same degree of care with each. After the four predictions Simon applies three verbs directly to the machines that already exist: think, learn, create. He writes about machines that think and means the ones already running, not ones still to come. In the same paragraph he extends the range of what they will be able to do in the visible future. He does not turn visible into a number of years.
Farther on, when he asks whether machines will replace managers, he makes no ten-year prediction. He puts it under a condition: can the cost per unit of problem-solving capacity compete with that of managers? In the same paragraph he qualifies it twice. The prediction is difficult to make, and if the change comes it will probably come gradually.
So in one paper a person gives four concrete forecasts ten years to come true, puts a larger possibility under a condition, and puts the boundary of ability under a word without a scale. The care is not distributed evenly. One place is loose, another tight. People quoting the paper later usually take the loosest places.
There is another comparison. Training a person to solve problems scientifically ordinarily takes about twenty years, he says. Copying a procedure learned by one machine, and shown to work, onto another computer would be a small matter hardly worth mentioning.
Then he uses a phrase about the time when machines will have minds. He is talking about copying such minds. A person takes twenty years; a machine can come by the same thing in a copy. First the machine has to have what will be copied. The sentence passes over that premise as something that will happen in the future.
Put this beside the boundaries drawn earlier in the series and it points the other way. In 1843 the machine could not begin. In 1887 it could not find its own questions. In 1936 a limit was proved. This sentence does not ask whether the machine can have a mind. It takes that as something that will happen and turns to a technical question afterward: making a copy.
The boundary is not in the sentence. The sentence stands on the far side and looks back. The earlier essays asked where it lay. This one no longer asks. It asks what to do after crossing it.
In the same talk he lists things people are already doing: programs that discover proofs rather than merely check them, checkers, chess, engineering design, musical composition, and pattern recognition. He expressly puts the theorem program alongside what is often called the creative, intuitive work of a scientist looking for a proof.
When he gets to music, he adds two qualifications. One is I believe, and one is I am told by a competent judge. His example is a piece of music made by a computer at a university. He does not judge the piece himself. He passes along somebody else's judgment. In the same after-dinner speech a man uses two layers of reported uncertainty for the merits of a computer's music, and only one exception, about eligibility, for the emergence of a world champion within ten years.
3. A General Problem Solver
The program for logical proofs could do one thing. It could prove the numbered propositions in Principia Mathematica. After that, the same people wanted to do something larger: a problem-solving program that did not come with a particular problem attached.
Its name was General Problem Solver. At an international conference on information processing in Paris in 1959, Newell, Shaw, and Simon presented “Report on a General Problem-Solving Program.” In April 1961 the RAND Corporation printed “GPS, A Program That Simulates Human Thought.” The title does not say intelligence. It says simulates human thought.
On the first page of that paper, the subject is tightly circumscribed: how certain people attempt to solve certain simple formal problems. The first sentence identifies the subject as the psychology of human thinking. The authors supply all three restrictions: certain people, certain problems, simple and formal.
A program with general in its name narrows the range that far on its first page. Between the general in the name and the qualifications in the text lies the whole paper. What people later quote is usually the name.
The task it actually demonstrates at length is transforming expressions in symbolic logic. Logical expressions are on the table. A few given logical rules can be brought into play. The thing to watch is the difference: here one connective has to be replaced, there an extra term has to be added. These differences are things a machine can compare.
That is what lets the whole procedure move. The difference between what is in hand and the goal first has to be something one can point to. Otherwise the remainder of the procedure has nothing to work on. How far the method can go depends on whether that difference can be written in a form a machine recognizes. What can be written down goes in; what cannot stays outside.
The method is called means–ends analysis. Compare what you have with what you aim to have, and pick out a difference. From a table matching differences to means, pick a means likely to reduce that difference. If it cannot be used yet, make the condition needed to use it into a new, smaller goal. Once the subgoal is achieved, apply the means. Compare the result with the original goal again and see what still differs. The subgoal itself is pursued in the same way.
This does look like what people do. You want to go somewhere and find you have no car. First get a car. To get a car you need money, so first get money. People call it figuring something out. The program makes it a sequence it can execute. The sequence can recur inside itself: a subgoal can have a subgoal, level after level.
The method does not try every means in turn. The 1961 paper ranks differences by difficulty. If a subgoal is harder than the goal above it, do not try it. If a goal appears after an easier one, it too may be passed over.
So the program actively lets go of some routes. Among them are routes that would have worked. That is a heuristic's business: give up some ways to get there so that the work as a whole can finish.
The two authors compared what a person said aloud, sentence by sentence, while solving a problem with the route the program took. They reported the mismatches as well as the matches. Have the person think aloud while working and make a record. Then print out the steps the program took. Put the two on the table side by side and count where the steps agree and where they do not.
Psychological experiment and program output are now on the same table. The people who did this wrote the program as a theory about human beings. Most of those who took up the work later treated it simply as a program.
At the end of that passage, the authors say something quite blunt. They acknowledge that the program has many degrees of freedom that can be adjusted to fit the data. There is no good way then to tell how many of those freedoms are relevant. Some of the agreement with human thinking may have been tuned in, and the authors themselves cannot count how much.
When enough things in a model can be adjusted, it gets easier to make it match the data in front of you. The admission is at the end of their own paper. It is not a fault someone else pointed out.
So in their account of 1961 it is two things at once: a computer program and a psychological theory of how people think. The aim is not to make an intelligence like a person's. It is to explain how a person thinks.
The two came to be treated as one. If a program is written as a theory about people, you check it against people to see whether it is right. If it is written as a tool, you check it against the work to see whether it is any good. Same code, two questions, two ways to keep score. The authors chose the first in 1961. The field went the second way. The choice determines what counts as success and who is owed an explanation when it fails.
They wrote it in a programming language they had put together themselves, made to work with symbols and lists. It was in the same line as the language used for the logical proof program in the previous essay.
4. Machine Learning
Another thing often mentioned from these ten years is a program that played checkers, written by Arthur Samuel. Checkers is a smaller problem than chess. The pieces start out alike. They move diagonally and capture by jumping, and at the far edge of the board they become kings. The rules can be told in a sentence or two. Playing well is not easy. It fits Shannon's reasons for choosing chess six years earlier: the permitted moves and the final goal are clear, and the difficulty is enough to matter. There is another advantage. It is easy to keep score. You can see who won and who lost, and look at the game record to see where a player went wrong. The field would keep picking problems like this. When the score is easy to keep, it is easy to tell people about. The thing people talked about most from these ten years was something you could score.
In July 1959, a company's research and development journal published his “Some Studies in Machine Learning Using the Game of Checkers.” The words in the title are machine learning. Not machine intelligence and not thinking machines. At the 1957 dinner the words had been heuristic problem solving; on the 1955 proposal they had been artificial intelligence. Here the word is learning.
Samuel writes carefully about what he can show. A computer can be programmed to learn to play checkers better. The shape of the claim is this: a person writes in the rules of the game, the way moves are generated, how far to look ahead, and the framework for scoring positions. Inside this structure people have set up, the program learns to play better than the person who wrote it.
Beating the person who wrote it sounds important. Its weight depends on how well that person played. Samuel was not a strong checkers player. In his paper it is a precise statement. Taken outside it often becomes a machine overtaking a human being.
As for what counts as learning, he uses a criterion of behavior: getting a machine to engage in behavior that would be described as learning if it took place in a person or an animal. He does not first settle whether a machine actually has a mental state called learning.
The move has come up several times in the series. An inaccessible thing is exchanged for one within reach, and the argument goes on. In 1936 a note took the place of a state of mind. In 1950 a game took the place of a question. Here a test of behavior takes the place of the word learning itself. Each time something is exchanged, the argument goes a step farther. What was taken out stays where it was.
Samuel's sentence does not say the machine is learning. It says the machine does something, and the same behavior elsewhere would be called learning. Change the subject and the sentence can stand.
He uses two methods. The first is called rote learning. The program stores positions it actually encounters in its search, along with the results of its analysis. If it comes to a stored position the next time it looks ahead, it can use the old result as information from farther down the tree. It can look farther in the same amount of time. It is not memorizing published games. It is memorizing calculations it has already done. The more positions it plays through, the more accounts it has stored, and the farther it can see in the same time.
The second is generalization. A person specifies parameters for judging how good a position is. Each is multiplied by an adjustable coefficient. The sum is the score. As the program plays, it adjusts the coefficients to make the scoring more accurate. Which features to look at is a person's decision. How much weight each should carry is the machine's adjustment. The person sets the question, the machine sets the score.
Both ways have shortcomings, and Samuel writes them down himself. Improvement by rote is slow but steady, and more useful in the opening and the endgame. Generalization learns faster but comes to a plateau. Where it stops depends on whether the features people supplied are up to the job. It is better in the middle game, worse in the opening and at the end.
The list of features is written by a person. The machine can weight these things very accurately. It cannot come up with a new feature to add. The restriction stays with it and has not been overcome by the end of the ten years.
5. That One Game
In 1962 the program won a game against Robert Nealey. That game became one of the most talked-about events of the whole decade. A machine beat a person, and the person had a title. The shape is easy to remember, tell, and use. Over the next sixty years, every so often a new story would be told in the same shape. The project would change, the game would change, the opponent would change. The shape stayed: a machine, a person with a title, a game, a result. When it took this shape in 1962, the title was supplied by a company's bulletin. Thirty-five years later, somebody went through the tournament records.
In August 1962 a company research bulletin described Nealey as a former Connecticut checkers champion and one of the country's leading players. A company research bulletin is written for people on the outside.
Samuel's 1959 paper was collected in a book in 1963, and the record of the game was printed with it. A game record is evidence of a particularly solid kind. There are the moves in black and white. Sixty years later one could give them to another program to work out again. A record that people later can recalculate and an assessment they can repeat are different kinds of thing. The former grows more exact as it sits. The latter grows smoother. What gets passed down is usually the latter.
They played six more games afterward. The six took more than five months. Nealey looked at a board at home and thought. Having chosen a move, he wrote it on a postcard and mailed it to the company's research center in Yorktown Heights. The machine usually took ten or twenty seconds to choose a move and could take about five minutes in a difficult position. A minute or so was the average. Nealey averaged three minutes a move.
Six games, more than five months. The actual thinking time on both sides together would probably not fill an afternoon. The rest was postcards on the road. A machine and a man played for five months across the postal system. This is what a person playing a machine looked like in the early 1960s. No screen, no two people facing one another, no spectators. There were postcards, one after another, traveling back and forth.
Nealey wrote about the 1962 game himself afterward. The machine had played a very fine endgame, he said. He had not met an opponent like it in an endgame since 1954. These are the opponent's own words, and they were quoted far more often than the results of his next six games.
6. The Games Afterward
Samuel gives the results of those six games in his own paper of 1967: five draws and one loss. The loss was the program's, after its victory in 1962. Five draws and a loss is how it looks from the program's side. On Nealey's side it is a win and five draws. The man who had lost one game in 1962 did not lose any of the next six.
In the same passage Samuel gives Nealey a title different from the one in the company bulletin of 1962. He calls him the United States blind checkers champion. Within five years one person is described by two different titles in two documents, both from people at the same institution. The first title comes in the year the program wins. The second comes in the paper that reports five draws and a loss.
There were other games. In 1965 the program played four games by correspondence with world champion Hellman. Hellman won all four. By 1967, a Pacific Coast champion named Hanson had beaten the current program twice. These are in Samuel's own paper. The losses are the writer's own report.
A historian of the game later checked the tournament records. Nealey won the Connecticut title in 1966. He had never competed in a top-level event such as the national championship. In the familiar later version he was ranked fourth in the country.
The same researcher went through the game of 1962 with a much stronger program. At a critical move Nealey took a position he could have drawn and turned it into a losing one. Right afterward Samuel's program made a bad mistake too and gave him his chance of a draw back. He did not take it. So this is what the game looked like: both sides made mistakes. The machine's mistake was not used against it.
In November 1967 Samuel published the follow-up in the same journal. He lists five marked shortcomings of the earlier program. The man listing them is the one who wrote it. It has been running a decade or more.
One, it cannot generate useful new features for evaluation on its own. Two, it is wrong to assume that the effects of features add up linearly. Three, it learns slowly. Four, its means of pruning and stopping the search tree are inadequate. Five, it has no means of shifting its overall strategy with the tactical situation. The first and fifth are ends of the same problem: it cannot devise a new view for itself, and it cannot change the way it plays to fit a position.
In the new version Samuel changes how combinations of features map to the value of a position and adds more complete pruning. On the very same page he says the basic limits of machine learning have not been overcome, and the program still cannot outplay checkers masters.
The new version also has a way to learn from existing records of expert play. About 250,000 positions from master games were punched onto cards, then put on magnetic tape. The program goes through the games, picking a move for itself, comparing it with the move the book recommends, then adjusting how it evaluates positions. A training run takes about ten hours.
Samuel gives this kind of learning a qualification too. In effect it assumes that the book's move is the one right move. That can fail. A position can have several good moves, and the book itself can be wrong.
There is another line in the conclusion. The final goal he had set eight years before, getting the program to produce the features it needed, still looked as distant as it had then. A decade on, the same person, the same task, the two ends of a sentence. Put his two papers next to each other and you have a set of accounts he kept for himself.
7. The Name Did Not Retreat
One line has run through the first six essays in this series. Before the machine does a thing it counts as intelligence. Once the machine has done it, the thing is renamed. In 1948 someone put it down in the last section of an internal report. Once a thing is understood, its behavior predictable and the method below it plain, the thing no longer seems intelligent.
Among the firsthand texts used in this essay, there is no clear statement that people stopped counting a particular capacity as intelligence as soon as a machine acquired it.
There is a string of examples in the other direction. In the speech of 1957 the theorem program's work is put alongside the intuitive work of scientists. Checkers and chess are on the list of achieved results in heuristic problem solving, with nothing saying those two do not count as intelligent.
In 1959 Samuel called his checkers paper machine learning. The machine was already playing. Doing so had not expelled it from the word learning.
In 1960 Newell, Shaw, and Simon titled another paper “A Variety of Intelligent Learning in a General Problem Solver.” Intelligent and learning appear alongside one another in the title. The general problem solver was already running.
In January 1961 Marvin Minsky published “Steps Toward Artificial Intelligence” in the proceedings of an electrical engineering society. By then the term could go straight into the title of a research paper. It was no longer an expression used inside the proposal from 1955.
In 1962 Samuel also used it in the title of a paper. It was the year his program won that game. After the win he did not change what he called his field.
Over those ten years the name was also becoming institutional. MIT had first a research project under that name, then a laboratory. Stanford had a laboratory of the same name too. The words moved from a proposal to the titles of papers, then to a sign on a door. It did not take a space vacated by an achievement to move them at any stage.
At the start of a book in 1967 Minsky lists activities that had by then appeared: playing chess, learning, non-numerical mathematical problems, expressions in ordinary language. He says these used to be confined to human intelligence. He does not follow that with a line saying a machine has done them and they therefore no longer belong to intelligence. They used to be confined there, he says. Machines can now do them too.
On the same page there is a line that has been quoted many times. In a generation, he is convinced, few compartments of intellect will remain outside the domain of machines. The second half says that the problem of creating artificial intelligence will be substantially solved. There are three qualifications. The time is a generation, not three years, ten, or twenty. The strength of the belief is personal: I am convinced. The extent of the result is substantially solved, not fully solved.
As for the widely repeated line that artificial intelligence is what has not been done yet, an early source located here is a book from 1979 about logic, music, and art. Its writer attributes it to somebody else and gives it that person's name as a theorem. The explanation is that once a mental function has been programmed, people soon stop treating it as an essential part of real thinking. The person it is named for said later that the version in circulation was not his exact wording. So the line that goes around as a law of the field leads back to a book in 1979 that credits somebody who does not altogether accept the credit. Another saying has it that once a thing works, nobody calls it artificial intelligence. That line has not turned up in the material used here from 1957 to 1970.
A good many things were done in those ten years, and this line did not move back. Playing checkers was done and still called learning. Proving theorems was done and still called intelligence. The book of 1967 still lists them as things that had previously belonged to people.
Something else did give way. Claims about ability gave way. Eight years later the man who set the final goal in the checkers paper said it was still where it had been. The four predictions had reached the point at which their authors said it would be time to look back. The name had not retreated. The abilities had not grown as fast as forecast.
This differs from the direction of the earlier essays. There, what was accomplished pushed the boundary back. In this set of texts, things that were done did not leave intelligence one by one along that line.
8. Three to Eight Years
On November 20, 1970, a popular illustrated magazine ran a piece headlined “Meet Shakey, the First Electronic Person.” The subtitle called it a machine with a mind of its own. The reporter was Brad Darrach.
Shakey was a machine at a research institute that could make its way around a room. It moved on wheels and saw with a camera. The reporter went to see it. The headline supplied the word person. The article also says what it could do: move between a few rooms and push a box that was in its way. The headline was for the people at the newsstand.
The piece quotes something the reporter says Minsky recently told him. The first sentence in quotation marks is that in three to eight years we will have a machine with the general intelligence of an average human being.
What follows is a set of everyday abilities to say what ordinary human intelligence means: read Shakespeare, grease a car, play office politics, tell a joke, get in a fight. Put the five things together and you have a list of things a living person might do in a day. It is not a technical specification. You can check the books against a specification. An everyday list is harder to check.
The quotation goes on to say that the machine could thereafter educate itself very quickly, first to the level of genius, then into a range where its abilities would be hard to reckon. Three to eight years to the ordinary person. There is no time given for the steps after that.
The same part includes the prospect of losing control. If machines take control, people may not get it back. With luck the machines might keep people around. There is no if or perhaps or under the right conditions in the first sentence.
Set it beside the four predictions from thirteen years earlier. In 1957 they came with a deadline and one exception, and before them the speaker said he was willing to make them. In the 1970 sentence there is an interval, three to eight years. There is nothing else.
The tone of the line is supplied by the reporter, who says Minsky spoke with quiet certainty. A later account says Minsky strongly denied saying or endorsing this passage. A statement printed as a quotation and a statement actually spoken are two different things. Later the line was often taken out as an example of the extravagant things the field had said.
Two people who wrote down four predictions supplied four standards at a dinner. When the date they proposed for checking them came around, no reckoning from them going through the four in turn has turned up here. Shortly before that point the last page of the checkers paper says the final aim looks as far away as it did eight years before.
The one game in 1962 is what people remember. The six games by correspondence, the four losses to the world champion, and the two losses to the Pacific Coast champion are all written down by the same person in the same paper. They are not what people remember. The shape of a thing in memory is not its shape on the page.
These records did not count out the remainders one by one. The people involved still counted some things they had got done as intelligence. What they had failed to do, they acknowledged in their separate papers.
The deadline changed, and so did what the talk was about. The ten years in 1957 concerned four things. The generation in 1967 concerned how many compartments of intellect would still be left. The three to eight years printed in 1970 concerned whether a machine could be like an ordinary person. One came at a dinner at an annual meeting, one at the start of a book, and one as a direct quotation in a popular illustrated magazine. You cannot use them to measure rates of progress. What can be compared is where the qualifications stand and how many of them survive.
A statement with a deadline can leave the page where it was printed without taking its conditions with it one by one. The sentences that are quoted over and over and taken as the field's promises are often the ones easiest to lift out.
The machines are still being built. Each generation says again where the boundary lies.