In 1955, a four-person proposal put the name on the page. In the five years that followed came a program that searches for proofs of theorems, a machine that learns from examples, a programming language for handling symbols, and a Japanese translation machine, along with a prediction that something would happen “within ten years.”
011955 · Name and proposal
A Proposal for
Where, and alongside what outlook, was the name born?
On August 31, 1955, four people (John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon) wrote a proposal for a workshop to be held at Dartmouth College the following summer and applied to the Rockefeller Foundation for funding. The words in its title were . The proposal starts from the conjecture that any feature of learning or intelligence can in principle be .
The seven boxes on the right are the research topics listed in the proposal. Press one to see what is inside. Language, neural networks, the size of computation, machines that improve themselves. Many of the spreads in this chapter pick up where one of them leaves off.
They planned for two months and ten participants. If chosen scientists spent a summer together on it, they wrote, . They asked for a budget of $13,500, but the foundation granted $7,500.
The items in the budget table add up to $13,000, which is $500 short of the $13,500 total written in the table (checked in the reprint on McCarthy's website; the reason for the gap is not given).
What you think you knowThe popular story: was a name everyone agreed on The original source says:It was the name McCarthy chose, and there were objections even then. Newell and Simon used a different name for years, .
What you think you knowThe popular story:The Dartmouth participants meant to finish AI in one summer The original source says:The proposal says . It does not say they would finish it.
Why ?
In a later book review, McCarthy wrote that one reason he chose the term was to avoid confusion with . He found its focus on (a mechanism that watches the result and corrects its action) beside the point. Other names in use included automata theory, complex information processing, and thinking machines.
For the Japanese term for artificial intelligence, the oldest example found in a paper is from Nagoya University in December 1962, and its first appearance in the minutes of the National Diet was in February 1983.
Sources & confidence (4)A Proposal for
31 August 1955Verified in the originalSources disagree
The proposal for a summer workshop that put forward the name
Four people, McCarthy, Minsky, Rochester and Shannon, drew up a proposal to hold a workshop on at Dartmouth College the following summer, and asked the Rockefeller Foundation for funding. The proposal started from the conjecture that every aspect of learning, and any other feature of intelligence, can in principle be . It lists research topics such as the use of language by machines, the forming of concepts, and self-improvement. The amount requested was $13,500 in total.
The popular story:People tend to say “they meant to complete AI in a single summer” The original source says:The proposal only says that if a carefully selected group of scientists worked on it together for a summer, on the list. It does not say they would complete it (proposal p. 2)
2 sources and how they were checked
Where sources disagree:Date the proposal was submitted: the cover says August 31, 1955 (proposal p. 1)／ “sent on September 2” (Grace Solomonoff 2016 p. 6, who says she saw the foundation’s records)／ “submitted in August 1955” (Nilsson 2010 p. 77)
The Dartmouth proposal expected that if the 10 selected people worked together for a summer (2 months), in one or more of the problems, such as getting machines to use language.
The popular story:People tend to say “the Dartmouth people said they would complete AI in a summer” The original source says:The original proposal says
What happened next:The workshop was held in the summer of 1956. But the grant from the foundation was about half of what was requested ($7,500), and many of the participants came for short visits in turn, with only 3 people there for the whole period. McCarthy later said that because of lack of funds and everyone’s own circumstances, the plan for everyone to gather for the whole summer was not realized. At the 50th anniversary conference in 2006, he admitted that the joint work had not turned out as hoped. He said the main reason the workshop did not go as expected was that . On the other hand, he also pointed out that important results came out around the same time, such as Newell and Simon’s theorem-proving program.
In a later book review, McCarthy wrote that one reason he coined the term was to avoid being lumped together with . Cybernetics was one of the names for thinking machines at the time (Grace Solomonoff 2016). McCarthy said he found it off the mark that cybernetics leaned toward (a mechanism that looks at the result and corrects the movement). He also wrote that he wanted to avoid both looking up to Wiener as a teacher and arguing with him.
The popular story:People tend to think was a name that everyone agreed on The original source says:It was a name McCarthy chose, and there was opposition from the start. Newell and Simon used a different name, , for years (Nilsson 2010 p. 79, citing McCorduck 1979 p. 97; McCorduck’s book is unchecked)
3 sources and how they were checked
Where sources disagree:Year of the review: 1989 in McCarthy’s own list. Some accounts say 1988 (unconfirmed)
The Japanese term for : in papers from 1962, in the National Diet from 1983
A search of J-STAGE (Japan's database of academic papers) turned up, in a December 1962 paper by Moriya Oda of Nagoya University, the phrase (translated) (a sentence introducing a survey by Minsky and others). Meanwhile, at a 1966 symposium at Tohoku University, the theme was written in English as . In the National Diet proceedings, the Japanese term for had zero hits through 1982. Its first appearance was in February 1983, when an expert witness brought it up in the phrase (translated).
What we still don’t know:J-STAGE's full-text search cannot search old scanned papers that have no text data, so there may be uses from before 1962. Newspapers (paid article databases such as the Asahi and Yomiuri) and books of the time have not been checked.
What the figure shows:The 1955 proposal began with the idea that features of learning and intelligence could be described precisely enough for a machine to imitate, and said “a significant advance can be made in one or more of these problems” across 7 tasks, without promising to finish them. The funding was only about half the amount requested.
021956 · Dartmouth
What Was the Summer Workshop Like?
What kind of gathering was “the big conference where AI was born,” really?
In the summer of 1956, researchers interested in thinking machines gathered on the top floor of the mathematics building at Dartmouth College. According to a note by participant Ray Solomonoff, it ran from around June 18 to August 17, about 8 weeks. Rather than a conference with a fixed schedule, it was a gathering where people came and went, visiting and talking.
In the figure on the right, press “Details: plan and records” to see the plan McCarthy sent to the foundation on May 26 (purple frames), with the people who were “there that day” according to notes by Solomonoff and Trenchard More (navy dots) laid over it. Only 3 people were there for the whole period, and 2 of the 11 people planned for never came. Between 3 and 8 people or so took part in the daily discussions.
There were also attempts to get everyone working on a single problem. On August 7, McCarthy proposed chess, but it did not become a joint project. McCarthy later said that, because funds were short and everyone had other business, the plan for everyone to be together all summer never came true.
No exact roster survives (McCarthy lost his own list). The dots mark only days recorded in the notes, so a day without a dot does not mean the person was absent.
What you think you knowThe popular story:The Dartmouth workshop was a large conference where AI was born The original source says:In fact it was a summer gathering where people came and went, with about 3 to 8 people in the daily discussions. The foundation's funding was about half the amount requested, and no report from the workshop was ever published.
Six weeks or eight?
It is often written as “6 weeks,” but the notes of Solomonoff, Minsky and More agree on about 8 weeks.
In a note at the end of the summer, Solomonoff wrote that the workshop was not very stimulating, but listed among its gains the idea that the problem of search might be important.
Sources & confidence (1)What Was the Summer Workshop Like?
18 June 1956Verified in the originalSources disagree
The Dartmouth summer workshop (regarded as the starting point of artificial intelligence research)
In the summer of 1956, researchers gathered on the top floor of the Dartmouth College mathematics department and discussed thinking machines. It was less a conference with a fixed schedule than a gathering where people came and went, and the daily discussions involved about 3 to 8 people. Only three people stayed for the whole period: John McCarthy, Marvin Minsky and Ray Solomonoff. Allen Newell and Herbert A. Simon brought the results of their program that proves logic theorems (Logic Theorist).
The popular story:People tend to think of “a big conference where AI was born” The original source says:in fact it was a summer gathering where people came and went, with about 3 to 8 people on any given day. Only 3 people were there the whole time. The foundation paid only about half of the $13,500 that was requested ($7,500), and no report of the workshop was ever published.
What we still don’t know:An exact list of participants has not survived (McCarthy lost his own list, according to Grace Solomonoff 2016 p.9). Who stayed for how many days can be partly worked out only from notes by Ray Solomonoff and Trenchard More.
5 sources and how they were checked
Where sources disagree:Length: about 8 weeks, from around 6/18 to 8/17 (Ray Solomonoff's notes; Grace Solomonoff 2016 p.9)／6 weeks (Nilsson 2010 p.78 and others). Attendee list: 32 people on Trenchard More's list (including people who did not come and people who showed interest)／20 people on Ray's list (Grace Solomonoff 2016 p.9-10). Trenchard More's stay: “two weeks, separately” (Moor 2006 p.87)／ “three weeks” (Grace Solomonoff 2016 p.8, p.10).
What the figure shows:Dartmouth was a summer gathering of about 8 weeks with visitors coming and going, rather than a large conference. Only 3 people stayed throughout, and daily discussions involved about 3–8 people.
People taking part in the daily discussions in week 6 (July 23–27), from More's notes
031956 · Search
What If We Just Build Every Proof?
If we build every possible proof in order, won't we eventually reach the theorem we want?
Allen Newell, Cliff Shaw and Herbert Simon took on the logic theorems in the book . From five axioms, three rules build new theorems: substitution (putting an expression in for a letter), replacement (rewriting into an equivalent form), and detachment (from A and “A implies B,” get B).
The simplest method is to build every possible proof, shortest first, and wait for the target expression to appear. The three called this the British Museum algorithm and used it as a point of comparison. By the report's numbers, there are 42 expressions up to the 4th step and 246 up to the 8th. To collect all of Chapter 2 might take something like 100 million, and they estimated it would take hundreds of thousands of years.
Checking candidate answers one after another is called search. Brute force has too many candidates, so you need rules of thumb for narrowing the range you check (heuristics). The figure on the right builds expressions using this site's way of counting.
expressions at step n ≈ (expressions at step n−1) × (number of ways to change one spot)
ways to change
turn p into ¬p, swap → with ¬…∨, and so on
2.02 and 2.07
In the report, with this method 2.02 came out 79th and took about 158 seconds. The Logic Theorist took about 10 seconds, using substitution.
In the figure, 2.07 never appears unless you allow substitution of one letter for another (turning q into p).
Sources & confidence (1)What If We Just Build Every Proof?
12 July 1956Verified in the originalSources disagree
Logic Theorist, a program that for proofs of logic theorems
Allen Newell, J. C. Shaw and Herbert A. Simon developed a program that searches for proofs of logic theorems, called . They estimated that generating every possible proof by brute force would take hundreds of thousands of years, so they built in heuristics (rules of thumb for searching) that pick promising paths the way a human does. Running on RAND's computer JOHNNIAC, it succeeded in proving 38 (73%) of the first 52 theorems in Chapter 2 of Whitehead and Russell's co-authored work Principia Mathematica.
3 sources and how they were checked
Where sources disagree:When it ran: a hand simulation was done on 1955-12-15 (Simon's recollection). It ran on JOHNNIAC a little before the Dartmouth workshop (summer 1956) (Nilsson p.113). The report of the results is a manuscript dated 1957-01-11. The paper was published at the Western Joint Computer Conference in February 1957 (the ACM bibliographic page gives 1957-02-26, pp.218-230; confirmed there, though the full text could not be opened on ACM).
What the figure shows:Generating every possible expression from shortest to longest makes the number grow at each stage, so the desired theorem takes a long time to appear. Shortcuts based on useful judgment, called heuristics, are therefore needed.
Expressions per step (the vertical axis grows by factors of 10)
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_british_museum.py
041956 · Logic Theorist
Searching Backward From the Answer
How can a program look for a proof the way a person does, with educated guesses?
Instead of building everything, the Logic Theorist (LT) searched backward from the expression to be proved. It tried four methods in turn. Substitution puts expressions in for letters of an axiom or an already proved theorem, to see whether it can match the problem exactly. Detachment looks for a theorem “A implies B” whose B matches the problem, and makes A a new, smaller problem. Forward chaining and backward chaining also create smaller problems.
In the 1957 report, it found proofs for 38 of the first 52 theorems of Chapter 2. Seventeen needed substitution alone, and 19 were solved in 2 steps. 2.01 took about 10 seconds and 2.45 about 12 minutes. On 2.31 the search ran out of moves after about 23 minutes. The figure on the right is a small LT built for this site. Press "Run to the end" and the tree grows downward; the green path is the proof.
The report offered an analogy. Moving forward from the axioms is like looking for one needle in a haystack; working backward from the target is more like looking for the path by which the needle leaves the stack.
The LT in the figure differs from the 1956 program in the details. The report says 2.13 cannot be proved with the four methods, but because this site's example puts expressions into shape differently and uses different conditions for the provided theorems, it finds a proof of 2.13 too when you check “Use provided theorems too”.
Are more known theorems always better?
In the report, 2.48 could be proved in 2 steps when all the earlier theorems were given, but the effort was 51,450. With only the axioms and 2.16 it took 3 steps, yet the effort was about one third. The report says that too much information, like too little, can be a burden on the solver.
Sources & confidence (1)Searching Backward From the Answer
12 July 1956Verified in the originalSources disagree
Logic Theorist, a program that for proofs of logic theorems
Allen Newell, J. C. Shaw and Herbert A. Simon developed a program that searches for proofs of logic theorems, called . They estimated that generating every possible proof by brute force would take hundreds of thousands of years, so they built in heuristics (rules of thumb for searching) that pick promising paths the way a human does. Running on RAND's computer JOHNNIAC, it succeeded in proving 38 (73%) of the first 52 theorems in Chapter 2 of Whitehead and Russell's co-authored work Principia Mathematica.
3 sources and how they were checked
Where sources disagree:When it ran: a hand simulation was done on 1955-12-15 (Simon's recollection). It ran on JOHNNIAC a little before the Dartmouth workshop (summer 1956) (Nilsson p.113). The report of the results is a manuscript dated 1957-01-11. The paper was published at the Western Joint Computer Conference in February 1957 (the ACM bibliographic page gives 1957-02-26, pp.218-230; confirmed there, though the full text could not be opened on ACM).
What the figure shows:Instead of generating everything, the Logic Theorist searched backward from the expression it wanted to prove and found a path to the same form as a known theorem.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_logic_theorist.py
051957 · A Machine That Learns
Fix the Weights Only When It Is Wrong
Can a machine learn how to tell things apart just by being shown examples?
In 1957, Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory, began work on the with support from the Office of Naval Research. It is an artificial neuron that imitates the connections between nerve cells: it multiplies each input (0 or 1) by a weight (how heavily the input counts), adds them up, and outputs 1 if the total is at least a threshold.
The way it learns is simple. Only when it is wrong, it nudges the weights a little in the direction of the right answer. In the figure on the right, place ● and ○ and press "Keep learning". The purple line turns and shifts in parallel, and the number of mistakes (purple rings) changes. Henry Block, a mathematician at Cornell, proved that if a line that separates the two groups exists, this correction method is sure to find one.
The history of neural networks starts here. Problems that one line cannot separate are in Chapter 5, and stacking layers is in Chapter 8.
y＝1 if w₁x₁ ＋w₂x₂ ≧ t, otherwise 0. When wrong: w ← w＋η(correct − y)x, t ← t − η(correct − y)
t
threshold
η
size of the correction (0.5 in the figure)
purple line
the boundary where w₁x₁ ＋w₂x₂ equals t
What he himself thought
In a 1961 report, Rosenblatt wrote that the perceptron is above all a model of how the brain works, and that inventing AI was not its main purpose.
Sources & confidence (1)Fix the Weights Only When It Is Wrong
July 1957–1960Verified in the originalSources disagree
A machine that learns by correcting itself when it is wrong: the perceptron (Frank Rosenblatt)
Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory, studied the , modeled on the neural circuits of the brain. It multiplies the pattern of dots that is input (an image) by weights and adds them up. When its answer is wrong, it adjusts the weights a little, and so it learns how to tell things apart just by being shown examples. In 1960, the Mark I, a machine that takes an image as input and recognizes shapes (its weights were motor-driven ), was unveiled. Rosenblatt himself wrote that this was above all a model for investigating how the brain works.
3 sources and how they were checked
Where sources disagree:Beginning: “as the PARA project in 1957” (Nilsson p.92)／Navy support began in July 1957 (Rosenblatt himself, 1961 p.viii). The first report (said to be dated January 1957) has not been opened, so the month is left undecided.
What the figure shows:The perceptron adjusts its weights slightly toward the correct answer only when it makes a mistake. If a straight line can separate the groups, this update rule is guaranteed to find one.
Mistakes per pass
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_perceptron_line.py
061957 · Predictions
Four Predictions: “Within Ten Years”
How did researchers in 1957 see the next ten years?
On November 14, 1957, at a dinner meeting of a conference in Pittsburgh, Simon gave a talk written with Newell (published in a journal the next year). He summed up that today's computers can solve problems that have no fixed solution method in ways resembling people, and can learn from experience. Then he predicted that four things would happen within ten years.
A computer would become world chess champion (, as he put it). A computer would discover and prove an important new mathematical theorem. A computer would write music that critics accept. Most theories in psychology would take the form of computer programs. He also said that .
There were results to build on: a program that proves logic theorems, a computer that plays beginner's chess, and the University of Illinois's ILLIAC, which composes music. In the figure on the right, advance the years to see the year of the answer check (1968) and what happened after.
A computer beat a reigning world chess champion in 1997, about 40 years after the prediction. A computer has never won the world champion title itself.
Sources & confidence (4)Four Predictions: “Within Ten Years”
14 November 1957Verified in the original
Prediction: (1957 lecture)
Herbert A. Simon and Allen Newell predicted that within ten years a computer would become the world chess champion (unless the rules barred it from competition).
What happened next:As of April 1967, one of the strongest chess programs (MIT's MAC HACK VI) played at the level of a mid-ranked amateur (1450 on the American rating scale). In 1968, International Master David Levy and some researchers made a bet on whether a computer could beat Levy within ten years. In 1978 Levy beat the computer and won the bet. A computer beat a reigning world champion (Kasparov) in a six-game match about 40 years after the prediction, in May 1997 (IBM's Deep Blue, 3.5 to 2.5). Note that a computer has never won the title of human world champion itself (the prediction also says ).
14 November 1957Verified in the originalSources disagree
Prediction: (a computer, in the 1957 lecture)
Herbert A. Simon and Allen Newell predicted that within ten years a computer would discover and prove an important new mathematical theorem.
The popular story:People tend to say “in 1976 a computer proved the four color theorem” The original source says:in the published text, the procedure for sorting the cases was made by humans by hand, and the computer's role was to check a large number of configurations one by one (Appel & Haken 1976 p.712)
What happened next:Within the ten years, no record of an event that could be called this was found. In 1976, Kenneth Appel and Wolfgang Haken proved the (the open problem, dating from around 1852, of whether any map can be colored with 4 colors). But the line of the proof was built by humans, and the computer's role was to check nearly 2,000 configurations. In October 1996, the theorem-proving program EQP automatically found, in about 8 days, a proof of the , which had gone unsolved since the 1930s. In both cases the problem itself had been posed by humans long before. Nilsson (2010) also writes that the prediction about proving theorems could be said to have come true before long (he does not say which event he means).
What we still don’t know:Whether a computer can be said to have achieved depends on how much of the work is counted as the computer's achievement. There is no clear criterion.
2 sources and how they were checked
Where sources disagree:Differing judgments: this site's search found no record of it happening within ten years／Nilsson 2010 p.163 writes that it could be said to have come true before long
14 November 1957Verified in the originalSources disagree
Prediction: (a computer, in the 1957 lecture)
Herbert A. Simon and Allen Newell predicted that within ten years a computer would compose music that critics would accept as having considerable value as art. In the lecture they also mention that the University of Illinois computer ILLIAC was already composing.
What happened next:No source for checking the answer was found. There is no standard for measuring whether critics accepted the music, and no record exists of anyone judging the prediction right or wrong. The paper itself states that, at the time of the lecture (1957), the University of Illinois's ILLIAC was already composing. Nilsson (2010) writes that the prediction about composing could be said to have come true soon after it was published (he does not give his criterion for judging).
What we still don’t know:There is still no settled way to judge whether music made by a machine has been accepted as art.
2 sources and how they were checked
Where sources disagree:Differing views on checking the answer: this site's search found no record of music being “accepted by critics” within ten years／Nilsson 2010 p.163 writes that it could be said to have come true before long
Herbert A. Simon and Allen Newell predicted that within ten years most theories in psychology would be written in the form of computer programs (or statements about the properties of programs).
What happened next:No source for checking the answer was found. In their 1958 reply, the two men wrote that treating programs as serious candidates for psychological theories . No source that counts whether were in program form in 1968 has been opened.
What we still don’t know:No source has been found that counts how many psychological theories took the form of programs.
What the figure shows:A 1957 lecture predicted 4 developments within 10 years. None had happened by the 1968 check, and chess alone reached a clear outcome, about 40 years later.
071958 · Critique and Reply
They Wrote the Judgment Day Themselves
What does it take to make a prediction checkable?
After the talk appeared in the journal, Richard Bellman criticized it in a letter. According to the two authors' reply, the criticism was that the predictions were . In their 1958 reply the two wrote, .
They explained that they attached a deadline and made the predictions concrete so that they would not become unfalsifiable predictions like those of Nostradamus. But to check a prediction, besides what will happen and by when, you also need a criterion for what counts as being right. In the figure on the right, try removing the deadline and the criterion. With chess, the result is clear. With the theorem, music and psychology predictions, even after the deadline people would disagree in their evaluations.
About the psychology prediction, the two also wrote in their reply, of the claim that the programs are strong candidates for psychological theories, that .
The chess criterion (beating a reigning world champion in a match) was set provisionally by this site.
What you think you knowThe popular story:Early AI researchers made big claims with no basis The original source says:Even then there was criticism, and the authors explained that they had attached deadlines so the claims could be checked, even naming the day of judgment themselves (around January 1, 1968).
Sources & confidence (2)They Wrote the Judgment Day Themselves
June 1958Verified in the original
A criticism of the prediction, and a reply that it
Richard Bellman criticized Herbert A. Simon and Allen Newell's prediction in a letter to a journal (according to the two men's reply, he called it ). In their reply the two wrote: . They explained that they set a deadline and made the prediction specific so that it would be something that could be checked, not a vague prediction like Nostradamus's.
The popular story:People tend to think “the early AI researchers made big claims with no basis” The original source says:there was already criticism at the time. The two men explained that they made the prediction with a deadline so it could be checked, and they themselves named the judgment date (1968-01-01).
Herbert A. Simon and Allen Newell predicted that within ten years most theories in psychology would be written in the form of computer programs (or statements about the properties of programs).
What happened next:No source for checking the answer was found. In their 1958 reply, the two men wrote that treating programs as serious candidates for psychological theories . No source that counts whether were in program form in 1968 has been opened.
What we still don’t know:No source has been found that counts how many psychological theories took the form of programs.
What the figure shows:Checking a prediction requires not just what will happen and by when, but also a rule for deciding what counts as success. Without that rule, even the deadline cannot settle whether it was right or wrong.
081958 · The Navy Demonstration
Mark on the Card: Right or Left?
On the day the newspaper wrote big expectations, what did the machine actually do?
In July 1958, the U.S. Navy showed reporters a demonstration in Washington. A simulation of the perceptron ran on the Weather Bureau's large IBM 704 computer, and after about 50 trials it could tell whether a mark on a card was on the right or the left. The special-purpose machine had not even been built yet.
The next day's New York Times headline was . The article reported that the Navy expected this “embryo” eventually to be .
The figure on the right is a small re-creation of that day's demonstration (the card size and marks were chosen by this site). Press "Run 50 more" and, on the map of weights, a pattern grows in which the right half votes for right and the left half votes for left. What the machine learned is only which squares vote which way. It does not know the words right or left.
What you think you knowThe popular story:Rosenblatt said “perceptrons walk, talk and are conscious” The original source says:What the article said it expected was . He himself said that , but in his 1961 report he criticized the press treatment as being like .
Sources & confidence (2)Mark on the Card: Right or Left?
7 July 1958Verified in the original
The Navy's announcement and the newspaper headline
In July 1958, the U.S. Navy gave a demonstration for reporters in Washington. The Weather Bureau's large computer, an IBM 704, became able to tell whether a mark on a card was on the right or the left after about 50 trials. The wire-service UPI's article reported that the Navy expected this “embryo” (a stage before birth) to be able, in time, to walk, talk, see, write, reproduce itself and be conscious of its existence. In the article, Frank Rosenblatt said it would be the first machine to think like the human brain, that it would be wrong at first but get wiser with experience, and that it might become an exploration craft sent to other planets. The dedicated machine (the perceptron) did not yet exist, and was planned to be built over about a year for $100,000.
The popular story:People tend to say “Rosenblatt said ‘the perceptron will walk, talk and be conscious’” The original source says:according to the article, the one that held that hope was (in the UPI wire story). In the article Rosenblatt himself also made big forecasts: that , and that . But in his 1961 report he criticized the way newspapers covered it in 1958, saying it was like , and wrote that the perceptron is above all a model of the brain. Also, the demonstration was not on a dedicated machine but a simulation on a general-purpose computer, the IBM 704 (the dedicated machine had not yet been built).
What we still don’t know:The article does not make clear whose statement within the Navy the Navy's expectations came from.
July 1957–1960Verified in the originalSources disagree
A machine that learns by correcting itself when it is wrong: the perceptron (Frank Rosenblatt)
Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory, studied the , modeled on the neural circuits of the brain. It multiplies the pattern of dots that is input (an image) by weights and adds them up. When its answer is wrong, it adjusts the weights a little, and so it learns how to tell things apart just by being shown examples. In 1960, the Mark I, a machine that takes an image as input and recognizes shapes (its weights were motor-driven ), was unveiled. Rosenblatt himself wrote that this was above all a model for investigating how the brain works.
3 sources and how they were checked
Where sources disagree:Beginning: “as the PARA project in 1957” (Nilsson p.92)／Navy support began in July 1957 (Rosenblatt himself, 1961 p.viii). The first report (said to be dated January 1957) has not been opened, so the month is left undecided.
What the figure shows:The newspaper raised great expectations, but the machine that day only learned in about 50 trials to tell whether a card's mark was on the right or left. All it learned was which side each square should vote for.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_perceptron_cards.py
091958 · LISP
Nested Parentheses for Expressions
How do you compute with symbolic expressions themselves, not numbers?
In the fall of 1958, McCarthy and Minsky started the at MIT. They wrote no proposal; asked what they needed, they reportedly answered one room, two programmers, one secretary, and one keypunch machine. That same fall they began building LISP (list processing), a programming language for handling strings of symbols.
In LISP, expressions are written with nested parentheses, like (TIMES X (PLUS X A) Y). The first word says what to do and the rest are what it is done to. The basic operations are only a few: car takes the first item, cdr takes the rest, cons attaches a new first item, and so on. A 1960 paper showed how to write differentiation of expressions using only these, conditional branching, and functions that call themselves.
In the figure on the right, select “X(X+A)Y · Paper example”, uncheck “Simplify 1 and 0 (site addition)”, and press "Differentiate with respect to X (DIFF)". The same answer as in the paper appears as a tree. ONE for 1 and ZERO for 0 stay in the result. Languages in the LISP family were used at the center of AI research for a long time afterward.
d(u·v·w)/dX＝u′·v·w＋u·v′·w＋u·v·w′
′
the derivative with respect to X. DIFF builds this form while keeping the nested parentheses
What the language was for
LISP was made for the plan of the , which draws conclusions from advice written in words (the next spread).
Sources & confidence (2)Nested Parentheses for Expressions
September 1958Verified in the original
Development of LISP, a programming language for symbol processing
John McCarthy and the MIT AI research group developed LISP (short for list processing), a programming language that makes it easy to handle sequences of symbols (words and formulas) rather than numbers. It was originally built for the project, which draws conclusions from advice written in words. Afterward, LISP and its descendants were used for many years as the center of AI research.
September 1958Verified in the originalSources disagree
The begins at MIT (McCarthy and Minsky)
John McCarthy and Marvin Minsky started the with the support of MIT's Research Laboratory of Electronics. There was no written plan; when the director asked what they needed, they reportedly answered: one room, two programmers, one secretary and one machine for punching holes in cards. A report from January 1959 states the aims as getting machines to solve problems that require intelligence, and, for that, cutting down the effort of by clever methods.
3 sources and how they were checked
Where sources disagree:When Minsky came to MIT: McCarthy's recollection (1979) says Minsky was an assistant professor in the mathematics department when it started in autumn 1958／Nilsson 2010 p.157 says he “joined MIT a year later (1959)”, and p.81 says that as of November 1958 he was “a staff member at Lincoln Laboratory, on the way to becoming an MIT assistant professor”. In the report of 1959-01-15, the roster lists Prof. M. L. Minsky.
What the figure shows:LISP handles expressions as nested parentheses, so a few operations such as taking the first part or the rest can calculate with expressions themselves, including differentiation, rather than just with numbers.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_mini_lisp.py
101958 · Common Sense
Deriving the First Step to the Airport
If facts are written as sentences, can a machine derive actions from them?
In November 1958, the symposium was held at the National Physical Laboratory in England. There McCarthy presented . The idea (the Advice Taker) is to write facts and advice as logical formulas and reason from them, and when a conclusion of the form “do …” comes out, to carry it out.
The paper's example is a problem of getting from a desk at home to the airport. From 17 premises, including at(I, desk) (I am at my desk), at(car, home) (the car is at home), walkable(home) (you can walk within the house) and drivable(county) (you can drive within the county), it derives 8 conclusions in turn, and ends with do(go(desk, car, walking)), that is, first walk to the car.
In the figure on the right, press "Derive next". If you remove one premise, no plan can be found.
The paper says it cannot fully answer how to gather the 17 premises or how to derive from them. The order of derivation in the figure follows the order of the conclusions in the paper.
Sources & confidence (2)Deriving the First Step to the Airport
24 November 1958Verified in the original
An international conference on the in Britain
An international conference on research into the mechanization of thought was held at the National Physical Laboratory in Teddington, England. John McCarthy presented (a proposal for a program that works out its own conclusions and actions from facts and advice written in words). Marvin Minsky presented a summary of AI methods, and Oliver Selfridge presented (a proposed learning mechanism in which small “demons” each look for a part of a shape and shout, and letters and the like are recognized by how loud the shouting is).
Development of LISP, a programming language for symbol processing
John McCarthy and the MIT AI research group developed LISP (short for list processing), a programming language that makes it easy to handle sequences of symbols (words and formulas) rather than numbers. It was originally built for the project, which draws conclusions from advice written in words. Afterward, LISP and its descendants were used for many years as the center of AI research.
What the figure shows:Writing facts and advice as formulas allows step-by-step reasoning to produce an action such as “First walk to the car.” Removing 1 premise prevents it from finding a plan.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_advice_taker.py
111958 · Pattern Recognition
Recognizing Letters by Demons' Shouts
Can we split a letter-recognition mechanism into many small parts?
At the same symposium, Oliver Selfridge presented a mechanism called (meaning the place where all the demons gather). The data demons at the bottom each look for part of a letter (a horizontal bar, a vertical bar, a curve and so on) and shout loudly in proportion to how confidently they find it. The cognitive demons above collect the shouts that match their own letter and shout in turn, and the decision demon at the top answers with the letter that has the loudest voice.
He also proposed a way to learn. Change the demons' megaphone volumes slightly from outside, pick the adjustment that raises performance most, and repeat (hill climbing). In the figure on the right, press "Keep tuning" and the accuracy rises from volume 0.
The structure that picks out features and combines them with weights is the same mechanism as in the perceptron and in later neural networks.
voice of cognitive demon k＝Σi volumek,i × voice of data demon i
volumek,i
the weight with which demon k listens to demon i (the part that learns)
answer
the k with the loudest voice
The figure and the sources
The letters (A, E, L, O, T) and the demons in the figure are an example made by this site. The description of the mechanism comes from Nilsson (2010), who wrote a book on the history of AI (Selfridge's paper itself could not be checked).
What came next
According to Nilsson, a 1960 paper by Selfridge and Ulric Neisser recognized 10 kinds of handwritten characters using 28 features and was trained on 330 examples. The two wrote that its accuracy was only about 10% lower than humans reading them.
Sources & confidence (1)Recognizing Letters by Demons' Shouts
24 November 1958Verified in the original
An international conference on the in Britain
An international conference on research into the mechanization of thought was held at the National Physical Laboratory in Teddington, England. John McCarthy presented (a proposal for a program that works out its own conclusions and actions from facts and advice written in words). Marvin Minsky presented a summary of AI methods, and Oliver Selfridge presented (a proposed learning mechanism in which small “demons” each look for a part of a shape and shout, and letters and the like are recognized by how loud the shouting is).
What the figure shows:Small demons look for parts of letters, and their shouts are combined after multiplying by their loudness, or weight, to identify the letter. Learning can happen simply by changing the loudness a little and keeping changes that improve results, a method called hill climbing.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_pandemonium.py
121959 · GPS
Shrinking the Gap to the Goal
Can we build a way of solving that works on any problem?
Newell, Shaw and Simon built the on the basis of the Logic Theorist and reported it as a preprint for a Paris conference in June 1959 (the report is dated December 30, 1958). It was called general because they built the content of the problem (logic, algebra, trigonometric formulas and so on) separately from the way of solving.
The heart of the method is means-ends analysis. Compare the current state with the goal, find the “differences,” and choose a move that reduces them. If the conditions for using a move are not met, making those conditions true becomes a new, smaller goal. In this way goals are created inside goals, and the same procedure is applied again and again at many levels.
The first problem in the figure on the right is one this site made up: making buttered toast. If you swap the table saying which move reduces which difference, the same way of solving works on a different problem.
One aim of the report was to bring the program's behavior closer to records of university students thinking aloud while solving logic problems. The examples shown in the report were symbolic logic, elementary algebra and trigonometric formulas.
Sources & confidence (1)Shrinking the Gap to the Goal
30 December 1958–1959Verified in the original
GPS (General Problem Solver), a program aimed at a way of solving that works on any problem
Allen Newell, J. C. Shaw and Herbert A. Simon presented , a development of Logic Theorist. It separates the content of a problem (logic, formulas, trigonometric identities and so on) from the method of solving, so that the same method can be used on many kinds of problems. The methods are means-ends analysis, which finds the between the current state and the goal and picks moves that reduce them, and first making a rough plan. Another aim was to come close to the spoken thoughts and written steps of university students solving problems.
What the figure shows:The system finds a difference between the current state and the goal and chooses an action that reduces it; if the action's conditions are missing, they become new smaller goals. Replacing the table lets the same solving method handle another problem.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_gps_means_ends.py
131959 · Japan
Ware ga ongaku o konomu (I like music)
What kind of “thinking machines” were being built in Japan at that time?
Around the summer of 1958, at the Electrotechnical Laboratory, a plan began for “Yamato”, a special-purpose computer that translates English into Japanese. They had no computer with a large memory for holding a dictionary, so they built a dedicated machine using the circuits of the transistor computer ETL Mark IV. The budget was 6 million yen, and the memory was a magnetic drum of 820,000 bits. In February 1959, when they fed in a paper tape punched with I like music, it typed out (romanized from the katakana output).
To avoid the difficulty of Japanese input (Chinese characters and word boundaries), they chose translation from English into Japanese. Even so, they had to move the verb to the end of the sentence, move prepositions, and add particles such as ga and o. For pronouns, I became ware, you became nare, and he became kare. In the figure on the right, step through the translation procedure one stage at a time.
Hiroshi Wada, who directed the plan, recalled that the 13 seconds until it began to type the first translation felt very long. The results were presented in June 1959 at the first International Conference on Information Processing in Paris.
Except for the translation of the pronouns, the dictionary and rules in the figure are an example made by this site. The sources do not say how Yamato translated I have some eggs in my hand., which was shown in Paris.
Japan at the time
The Electrotechnical Laboratory had been building trial transistor computers since about 1954. At the University of Tokyo, the PC-1, a computer using the parametron, a component invented in Japan, was completed in 1958. Wada also wrote that for the Paris exhibition, a transistor computer and a parametron computer were flown in from Japan and made to run.
Sources & confidence (2)Ware ga ongaku o konomu (I like music)
February 1959Verified in the originalSources disagreeJapan
The Electrotechnical Laboratory's dedicated English-to-Japanese machine translator “Yamato” (I like music → Ware ga ongaku o konomu, printed in katakana)
The Electrotechnical Laboratory developed a dedicated machine for translating English into Japanese, called “Yamato”. To avoid the difficulty of entering Japanese (kanji and word spacing), the team chose English to Japanese, and the translation was output in katakana. The first sentence it translated, , came out as (translated), and it took 13 seconds before printing began. It was presented at an international conference in Paris in June 1959.
2 sources and how they were checked
Where sources disagree:Start of the research: around Showa 32 (1957) (Wada's recollection)／the plan dates from around the summer of 1958 (Information Processing Society of Japan). Example sentence shown at the Paris conference: “I have some eggs in my hand.” (Wada's recollection). The Information Processing Society of Japan page does not give an example sentence.
Eiichi Goto, a graduate student at the University of Tokyo, invents the “Parametron”, a computing component unique to Japan
Eiichi Goto, a graduate student in Hidetoshi Takahashi’s laboratory at the University of Tokyo, invented a new component for expressing 0 and 1, the “Parametron”. It makes use of the fact that the oscillation of a circuit of a coil and a capacitor settles into one of two states (the phase is either 0 or 180 degrees). At the time, research and development of parametrons was also going on at Nippon Electronic Measuring Instruments, the Nippon Telegraph and Telephone Public Corporation, Kokusai Denshin Denwa, Hitachi, Fujitsu, NEC and others, and at the University of Tokyo the parametron computer PC-1 was completed in 1958.
What the figure shows:“Yamato” was a computer built for English-to-Japanese translation: it looked up words, moved the verb to the end, and added particles such as ga and o to print I like music as “Ware ga ongaku o konomu” (Japanese for “I like music,” translated).
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_yamato_translate.py
141959 · Machine Learning
Fixing Its Own Scoring Rules
Can a program become stronger than the person who wrote it?
Arthur Samuel of IBM kept up research on computers learning from experience, using checkers (a game on an 8×8 board where you jump over pieces to capture them). A learning version was demonstrated on television on February 24, 1956. The summary of his 1959 paper says that after 8 to 10 hours of play, the program could play better than the person who wrote it.
The score of a position is a sum of features such as the difference in pieces, how many moves are available, and whether the center is held, each multiplied by a weight. The program looks several moves ahead, choosing the larger score for itself and the smaller for the opponent, and carries scores upward (minimax). One of its ways of learning is to fix the weights so that the score before looking ahead gets closer to the score carried up after looking ahead. In the figure on the right, press "Learn".
Not writing out the whole way of solving, but letting the machine learn from examples and experience, is machine learning.
w ← w＋η(carried-up score − score before looking ahead) × feature
w
weight
η
size of the correction
score
the sum of feature × weight (the numbers in the figure are an example made by this site)
The real program
It chose and used the best 16 from 38 features. It usually looked ahead about 3 moves. This learning method of fixing the weights is regarded as a forerunner of temporal-difference learning (TD learning) (Nilsson 2010).
The term
It is often said that Samuel coined the term machine learning in 1959. We can confirm that the term is used in the title and summary of the 1959 paper, but we have found no source that checks whether this was its first use.
At the Dartmouth summer workshop as well, Samuel talked about checkers on August 6. Solomonoff's note records that Samuel was interested in the problem of learning more than in checkers itself.
Sources & confidence (1)Fixing Its Own Scoring Rules
1 July 1959Verified in the original
Samuel's
Arthur Samuel of IBM wrote up research in which a computer learns from experience using checkers (a game on an 8×8 board where pieces jump over and capture each other; called draughts in Britain). According to the abstract of the paper, all you need to give it is the rules, a rough policy, and a list of features that seem relevant. After 8 to 10 hours of play, it is said to play better than the person who wrote the program. Among the methods used was automatically adjusting, during play, the weights used to evaluate positions in look-ahead search.
What we still don’t know:It is often said that Samuel coined the term machine learning in 1959. We confirmed that it is used in the title and abstract of the 1959 paper, but we found no source that confirms whether this was its first use.
What the figure shows:Samuel's checkers program adjusted its own weights so its score before looking ahead moved closer to the score brought back from future positions. Learning from experience instead of having every solving step written down is machine learning.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_samuel_weights.py
151960 · Mark I
Random Wiring, Motor-Turned Knobs
What if we build a learning machine as a real machine, not a program?
On June 23, 1960, Mark I, the special-purpose perceptron machine, was shown to the public for the first time. Its weights were adjusted by motor-driven (potentiometers), and the wiring was connected through a large patch panel. According to the 1961 report, it had 512 A-cells and 8 R-cells that give the answer. Wightman and Martin assembled it, and Hay carried out the experiments.
The structure has three stages. From the points of the retina (S) to the A-cells, the connections just go to randomly chosen points, and this part does not learn. Only the weights from the A-cells to the R-cells are learned. Following Experiment 1 in the same report, the figure on the right uses a retina shaped like a ring, with top and bottom and left and right connected, and has the machine tell horizontal bars from vertical bars.
If the retina is wired directly, without passing through the A-cells, the mistakes never vanish however many passes it learns. Whatever the weights, the average score over all the horizontal bars and the average score over all the vertical bars come out the same. With 80 random A-cells in between, cells that pick up combinations of points get mixed in, and the mistakes reach 0 in a few passes.
The report also records an experiment in which, even when the teacher deliberately gave wrong answers 30% of the time, the machine learned to tell E from X 100% of the time (a different method from error-correcting learning). Mark I is now at the Smithsonian Institution.
July 1957–1960Verified in the originalSources disagree
A machine that learns by correcting itself when it is wrong: the perceptron (Frank Rosenblatt)
Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory, studied the , modeled on the neural circuits of the brain. It multiplies the pattern of dots that is input (an image) by weights and adds them up. When its answer is wrong, it adjusts the weights a little, and so it learns how to tell things apart just by being shown examples. In 1960, the Mark I, a machine that takes an image as input and recognizes shapes (its weights were motor-driven ), was unveiled. Rosenblatt himself wrote that this was above all a model for investigating how the brain works.
3 sources and how they were checked
Where sources disagree:Beginning: “as the PARA project in 1957” (Nilsson p.92)／Navy support began in July 1957 (Rosenblatt himself, 1961 p.viii). The first report (said to be dated January 1957) has not been opened, so the month is left undecided.
What the figure shows:Connecting the retina directly cannot separate horizontal and vertical bars, no matter how many training rounds run. Adding randomly wired A cells between them reduces errors to 0 within a few rounds.
Mistakes per pass (16 bars per pass)
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_mark1_alpha.py
161960 · People and Computers
Sharing One Computer Among Everyone
Before we make machines do all the thinking, how should people and computers team up?
In March 1960, J. C. R. Licklider published the paper . Like a fig tree and the insect that carries its pollen helping each other, he imagined a partnership in which people set the goals, form hypotheses and decide the criteria for evaluation, while the computer takes on the routine work. He recorded his own work in the spring and summer of 1957 and found that about 85% of his thinking time went to preparing to think (looking things up, calculating, drawing graphs and so on).
One of the technologies needed for this is time-sharing (a mechanism that divides one computer's time finely among many people). He argued that the large computers of the day were too fast and too expensive to think together with just one person, so their time should be shared among many people. In the figure on the right, compare it with waiting in line. Even if the total time the computer spends processing is the same, the wait for the first reply becomes much shorter.
What you think you knowThe popular story:Licklider predicted that machines would think by 1980 The original source says:The paper's reports an estimate from an Air Force conference. About the length of the era of people and computers working together, he himself left a wide range: .
Sources & confidence (1)Sharing One Computer Among Everyone
March 1960Verified in the original
Licklider's — before leaving thinking to machines alone, an era is coming in which people and computers team up
J. C. R. Licklider published the paper . Just as fig trees and the insects that carry their pollen live by helping each other, he imagined a relationship in which people and computers cooperate closely. Humans would set the goals, form hypotheses, and decide the criteria for judging and evaluating, while the computer would take on the routine work. He argued that making this happen requires progress in time-sharing systems (where many people use one computer at the same time), memory devices, programming languages, and input and output devices. The paper points out that programs for things like theorem proving and chess already match humans in limited areas, and that a program that proves plane-geometry theorems works at the same speed as a Brooklyn high-school student and makes similar mistakes. He then admitted that machines might take the leading role in thinking in the distant future, but argued that for a long period before then, people and computers would cooperate to produce intellectual progress.
The popular story:People sometimes quote “Licklider predicted that machines would think by 1980” The original source says:the 1980 in the paper reports an Air Force study group's estimate (the year machines alone could do much of the thinking or problem solving of military significance). It is, in the paper, an estimate that . Licklider's own claim is that the long period before that is an era of people and computers working together, and he leaves that period of 15 years wide open: .
What the figure shows:Time-sharing divides 1 computer's time into small pieces and distributes them. The total processing time stays the same, but the wait for a first response becomes much shorter.
171976 · Checking the Answer
Four Colors: Human Logic, Machine Labor
When can we say a computer has proved a theorem?
The prediction said that a computer . For those ten years, we find no record of an event that would let us say so. In 1976, Kenneth Appel and Wolfgang Haken proved the (the problem, dating from about 1852, of whether any map can be colored with 4 colors so that neighbors differ). But people built the line of proof and the procedure for classifying the shapes, and the computer's role was to check nearly 2,000 shapes one by one.
The figure on the right has the computer color a map, an example made by this site. It colors the countries one at a time, and when no color is left it goes back one step (backtracking). A, B, C and D are all next to each other, so no matter how you try, they cannot be colored with 3 colors. A search computation that colors one map and the proof of a theorem that holds for all maps are different things.
In October 1996, the proof-search program EQP found a proof, in about 8 days and entirely automatically, of the , which had gone unsolved since the 1930s. In both cases, people had already posed the problem itself.
What you think you knowThe popular story:In 1976 a computer proved the four color theorem The original source says:In the text of the announcement, people built the classification procedure by hand and the computer's role was to check many shapes one by one.
Is there a single answer?
Nilsson (2010) writes that the theorem and music predictions could be said to have come true soon after they were made (without saying which events he means). The answer changes with how much of the work you credit to the computer, and there is no fixed criterion.
Sources & confidence (1)Four Colors: Human Logic, Machine Labor
14 November 1957Verified in the originalSources disagree
Prediction: (a computer, in the 1957 lecture)
Herbert A. Simon and Allen Newell predicted that within ten years a computer would discover and prove an important new mathematical theorem.
The popular story:People tend to say “in 1976 a computer proved the four color theorem” The original source says:in the published text, the procedure for sorting the cases was made by humans by hand, and the computer's role was to check a large number of configurations one by one (Appel & Haken 1976 p.712)
What happened next:Within the ten years, no record of an event that could be called this was found. In 1976, Kenneth Appel and Wolfgang Haken proved the (the open problem, dating from around 1852, of whether any map can be colored with 4 colors). But the line of the proof was built by humans, and the computer's role was to check nearly 2,000 configurations. In October 1996, the theorem-proving program EQP automatically found, in about 8 days, a proof of the , which had gone unsolved since the 1930s. In both cases the problem itself had been posed by humans long before. Nilsson (2010) also writes that the prediction about proving theorems could be said to have come true before long (he does not say which event he means).
What we still don’t know:Whether a computer can be said to have achieved depends on how much of the work is counted as the computer's achievement. There is no clear criterion.
2 sources and how they were checked
Where sources disagree:Differing judgments: this site's search found no record of it happening within ten years／Nilsson 2010 p.163 writes that it could be said to have come true before long
What the figure shows:A computer can color 1 map by trying colors and going back when stuck, but calculating for 1 map is different from proving something about every map. Even for the four-color theorem, people designed the reasoning.
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u03_four_colors.py
Closing
What Is Still Unknown in This Chapter
No exact list of the Dartmouth workshop's participants survives. From a few surviving notes, we can only piece together, bit by bit, who came and when. It is often said that Samuel coined the term machine learning in 1959, but no source has been found that checks whether that was its first use.
Of the four predictions, only chess got a clear answer, about 40 years later. For the theorem, music and psychology predictions there was no criterion for what would count as right, and evaluations still differ from person to person. To make a prediction checkable, you need not only a deadline but also a criterion for judging.
In these five years, two roads started running side by side: searching with rules of thumb using symbolic rules (Logic Theorist, GPS, LISP), and learning by correcting weights from examples (the perceptron, Pandemonium, Samuel). They connect directly to the algorithms of Information I (Japan's required high-school informatics course) and to the ideas behind the machine-learning basics of Information II (the national Information II teaching materials cover the perceptron in their Learning Unit 17) (this site's own view). In the next chapter we look at how far AI that searches and estimates got in the 1960s, and what predictions were made.