All History of AI

In the 1985 Diet, AI is explained as being said to be “the second boom now” (translated)

Takemochi Ishii of the University of Tokyo was called to the Diet (a committee of the House of Representatives) and explained, “This AI is currently said to be in its second boom” (translated). He said the first was an effort centered on the United States from the mid to late 1960s, and the second was Japan's Fifth Generation Computer project. The idea of counting booms was already around in the middle of the second boom.

What we still don’t know:We do not know who started the phrase “the second boom” (translated) before 1985 (this is the oldest instance found within the Diet records).

1984: “AI Winter” was spoken of as the name of the worst thing that might come

At a panel discussion at AAAI-84, the moderator, Drew McDermott, said that researchers were worried that expectations for AI were too high and that it would someday end in trouble. As the worst-case scenario he gave this picture: in five years the strategic computing initiative fails miserably, Japan's Fifth Generation also does not work out, the Japanese government quickly pulls out of computers, the start-up companies all go bankrupt, and no money goes to anything related to AI. He described this as “This condition, called the ‘AI Winter’ by some”. He urged them to warn themselves so that it would not come to that, and to tell the world. In other words, “The AI Winter” was a phrase that appeared in the middle of the second boom as a warning about the future. Calling the decline in research funding in the 1970s the “first AI winter” is a label added afterward.

What we still don’t know:Who first said “AI Winter” is not known (even in the discussion it was “by some”). What book the small value for 1983 in the Ngram comes from has not been confirmed.

Calling booms by number, “first, second … boom”, is mostly a Japanese way of counting (English numbers the “AI winter”)

Word frequencies in English-language books (Google Books Ngram) show that “AI winter” appears from 1983 and rose sharply in the late 2010s. Expressions that put numbers on winters, such as “first AI winter” and “second AI winter”, also exist. On the other hand, “second AI boom”, “third AI boom” and “fourth AI boom” did not appear in countable amounts. In Japanese, “first artificial intelligence boom” (translated) can be seen in a 1982 magazine, “the second artificial intelligence boom” (translated) in a 1985 book, and “the second boom” (translated) in the Diet in 1985. From 2014, the first, second and third AI booms were used in an academic society journal, and they also appear in a chapter title of a 2015 book by Yutaka Matsuo and in government white papers from 2016 on.

What we still don’t know:We have not looked at how this is counted in Chinese (such as phrases for a “third wave” of AI) or Korean. Ngram covers only books through 2022, so we do not know the English expressions from 2023 onward.

Turing, “On Computable Numbers”: defining what it means to be “computable” with a machine in the head

Alan Turing published the paper “On Computable Numbers, with an Application to the Entscheidungsproblem”. He simplified the steps of a person calculating with paper and pencil down to a “machine” that has a tape divided into squares (in place of paper), a head that reads and writes one square at a time, and only a finite number of states (Sections 1 and 9). He defined a number whose decimal digits this machine can write out as “computable”. He further showed that one machine (a universal machine) can compute any computable sequence if it receives the description of another machine on its tape (Section 6). And he showed that there is no method that always decides, by a fixed procedure, whether a logical formula can be proved (the answer to the decision problem) (Section 11). The next year this model was named the “Turing machine”, and it is regarded as a foundational model of the theory of computation and of computer science.

The McCulloch and Pitts paper: writing the brain’s nerve cells as a circuit of switches that fire or do not fire, in formulas

Warren McCulloch and Walter Pitts published the paper “A Logical Calculus of the Ideas Immanent in Nervous Activity”. Because the activity of a nerve cell (neuron) is “all-or-none” (it fires or it does not), they reasoned that the workings of a neural network can be described by propositional logic (logic with only two values, true and false). As premises for the calculation they set that firing is all-or-none, that a neuron fires when a certain number of synapses are stimulated at the same time, that if even one inhibitory synapse is active the neuron does not fire, and that the structure of the network does not change with time (page 118). They further wrote that if a tape and a reading device are added, this network can compute the same numbers as a Turing machine (page 129). It is regarded as a starting point of later research on artificial neurons.

What we still don’t know:The paper itself cautions that the “formal equivalence”, that a learning network can be replaced by a network whose structure does not change, is not an explanation of facts about the real brain (page 117). That real neurons do not work exactly as these premises assume (all-or-none, unchanging structure) can be used as an example for understanding the difference between a model and the real thing.

Turing publishes a paper discussing machine intelligence through the “imitation game”

The British mathematician Alan Turing published a paper that replaces the question “Can machines think?” with the “imitation game”. The original is a game in which an interrogator tries to guess, using only written exchanges, which of a man and a woman in another room is which. Turing turned it into the question of how often the interrogator would be wrong if a machine played the man’s part. The paper also discusses answers to 9 objections and the idea of a “child machine” that is raised by education.

What we still don’t know:We have not confirmed on this site who first used the name “Turing test”, or when.

Kempelen’s “Chess-Playing Turk”: a person was hiding inside

Kempelen, a nobleman from Pressburg in Hungary, built a machine at the court in Vienna for Empress Maria Theresa: a figure in Turkish dress that played chess on top of a cabinet. It toured Europe and America and amazed many people, but in fact a skilled chess player was hidden in the cabinet and moved it. As early as 1784 a book was published saying that a person was hidden inside and playing, and in 1821 a book appeared that explained with diagrams how the person hid. The machine was destroyed in a fire in Philadelphia in 1854.

The proposal for a summer workshop that put forward the name “Artificial Intelligence”

Four people, McCarthy, Minsky, Rochester and Shannon, drew up a proposal to hold a workshop on “artificial intelligence” 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 “so precisely described that a machine can be made to simulate it”. 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 Navy's announcement and the newspaper headline “NEW NAVY DEVICE LEARNS BY DOING”

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.

What we still don’t know:The article does not make clear whose statement within the Navy the Navy's expectations came from.

Prediction: “within ten years a digital computer will be the world’s chess champion” (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 “unless the rules bar it from competition”).

The A* (A-star) pathfinding method — searching by the distance so far plus an estimate to the goal

Peter Hart, Nils Nilsson and Bertram Raphael published the paper “A Formal Basis for the Heuristic Determination of Minimum Cost Paths”. For the problem of finding the lowest-cost route in a graph of connected points, they built a theory that brings in hints from the problem itself (a heuristic: a shortcut based on experience or estimates) and showed that one search method is optimal (from the paper's abstract). According to Nilsson's recollection, for the path search of the robot “Shakey,” they took the idea of using the “airline distance” to the goal as the estimate (Nilsson) and added to it the idea of “the distance traveled so far” (Raphael). Hart then showed that as long as the estimate never exceeds the actual remaining distance, the shortest route is always found. This search method was named A*, and today it is used in places such as car navigation route guidance and games (Nilsson 2010).

ELIZA, a program that converses in text. One aim of its author's paper was to explain away a magic that seems impressive

Joseph Weizenbaum of MIT presented in a paper ELIZA, a program that can converse with a computer in text. It finds keywords in a sentence typed by a person and rearranges them according to rules to produce a reply. The role of the conversation partner can be changed by swapping a “script”, and as of 1966 there were scripts in Welsh and German as well as English. At the start of the paper, the author writes, in effect, that if the inside of the program is explained in plain words the magic crumbles and it is seen to be just a collection of procedures, and that this paper is meant for that purpose.

Automatic reader-sorter for handwritten postal codes (Toshiba and the Ministry of Posts, 1968)

On July 1, 1968, the day the postal code system began, the automatic postal code reader-sorter (TR-3 and TR-4), which Toshiba had developed under the guidance of the Ministry of Posts, was shown at the Tokyo Central Post Office. It is said to be the world's first machine to recognize digits that people had written freely by hand and sort mail with them. The developers analyzed about 300,000 handwritten digits collected from across the country to find which quirks in letter shapes appear most easily, and used an optical filter to erase the red frame so that only the strokes of the digits were extracted. It read the first 3 digits at a speed of about 6 items per second, and characters it could not recognize were marked as “reject” and passed to people.

What we still don’t know:The statement that it was “the first in the world” (translated) comes from the accounts of Japanese academic societies, museums and companies. We have not opened any source that compares it with machines from the same period in other countries.

Simon's prediction “machines will be capable, within twenty years, of doing any work” (1965 book／originally a 1960 paper)

In his 1965 book, Herbert Simon wrote that within 20 years machines would be able to do any work a human can do. In the 1960 paper it was based on, he wrote that in a period much shorter than 25 years machines would have the “technical capability” to replace any human function in organizations. Immediately afterward, however, he went on to say that what is technically possible is separate from how it is actually used. He then assumed that, in the long run, nearly everyone who wants to work can work (full employment), and argued that the division of labor between people and machines is decided by “comparative advantage” (each side takes on what it is relatively better at).

What happened next:No source was found saying that by 1985 machines could do any human work. Nilsson (2010) says that, even as of 2010, Simon's similar statement from 1957 (that the range of problems machines can handle would become as wide as the human mind) was still far from coming true. Simon himself, on the other hand, thought of technical capability and employment separately and expected that, over the long run, unemployment would not rise (in his 1966 reply he wrote that the unemployment rate was below 4% and falling).

From which year to which year was the “First AI Winter”? The answer differs by source. Some historians say “There Was No ‘First AI Winter’”

Japan's white paper from the Ministry of Internal Affairs and Communications (2016) calls the 1970s “the winter years” (translated), and Wikipedia gives 1974 to 1980. On the other hand, the researcher Nilsson (2010), who wrote a history of AI, calls about 1975 to 1985 the “boomtimes” of AI and says the “AI winter” came in the late 1980s. And the historian Haigh (2023) argued “There Was No ‘First AI Winter’,” pointing out among other things that the membership of SIGART, the AI special interest group of the computing society ACM, nearly tripled from 1973 to 1978. In fact, an early example of the term “AI Winter” is its use in 1984 as a word expressing worry about a winter that might be coming.

What we still don’t know:We have not found figures that line up total AI research funding in the 1970s by country and by year. Whether it was an “AI winter” changes with what you use as the yardstick (the budgets of large labs／the membership of the society／how the words were used)

The book “Perceptrons” by Minsky and Papert shows mathematically the limits of simple neural networks

A book by Marvin Minsky and Seymour Papert of MIT that examined, with mathematics, the power of the “perceptron,” the learning machine Rosenblatt built in the 1950s (a way of doing pattern recognition with a neural network). They proved that if the number of points each part can see is limited, the machine cannot recognize things more complicated than “one or the other but not both” (XOR), such as “parity” (whether the number of points is even or odd) and “connectedness” (whether a figure is a single connected piece). About multilayer networks, they wrote their “intuitive judgement that the extension is sterile,” while also saying that confirming (or dismissing) this is an important research problem (p.231, as cited by Pollack).

The UK's Lighthill Report: “In no part of the field ... the major impact that was then promised”

A report by James Lighthill, a leading figure in fluid dynamics, who surveyed AI research alone at the request of a British research funding body (the Science Research Council, SRC). He sorted AI into A (automation), B (the “bridge” work such as building robots) and C (studies of the brain and nerves using computers), and while he rated A and C, he was harsh on B. As a major reason for the lack of progress, he pointed to the failure to notice the phenomenon in which the number of combinations to be searched grows explosively (the combinatorial explosion). After this report, trust in AI research in Britain and its funding fell sharply.

Kaoru Nakano (University of Tokyo)'s “Associatron”: a brain-inspired memory that recalls the whole when shown a part

Kaoru Nakano of the University of Tokyo presented the “Associatron,” a mechanism that stores information as patterns of 0s and 1s spread across a whole network and recalls the whole from a part of it. The larger the clue, the more accurately it recalls, and the smaller the clue, the more vaguely. It is a simplified model of the network of nerves in the brain, and he showed it could be built with elements connected only to their neighbors working at the same time. He ran experiments on a computer and also showed examples of applying it to concept formation and to games.

MYCIN: an AI of “IF–THEN” rules that advises doctors on choosing drugs for infections. Its evaluation matched specialists, yet it was not used routinely in hospitals

At Stanford University, Edward Shortliffe, then a medical student, and his colleagues developed MYCIN, a program that advises doctors on choosing drugs for bacterial infections (blood infections, and later meningitis). They built doctors' knowledge into it as rules of the form “if (a test result is ...) then (the germ is probably ...)” (about 200 at first, about 500 by 1978), and made it able to explain the reasons for its judgments. In a 1979 evaluation, for 10 meningitis patients, 8 infectious disease specialists from outside Stanford scored the prescriptions of MYCIN and of doctors without knowing whose they were. 65% of MYCIN's prescriptions were “rated as acceptable by the evaluators”, while five Stanford specialists scored 42.5 to 62.5% (average 55.5%). Even so, the developers' book records that “MYCIN was never used routinely in patient-care settings”.

What we still don’t know:We have not checked who first brought up the matter of law and liability, or when.

1984: at an AI conference exhibit hall, an ad says “We’ve built a better brain” — a record of the mood that this time it was the real thing

Science writer George Johnson described the exhibit hall at the August 1984 conference of the American Association for Artificial Intelligence (AAAI). Most researchers had come to discuss small steps forward, but some exhibiting companies acted as if the age-old dream of making intelligent companions had already come true. The brochure for TIMM, an AI tool for entering experts' knowledge, boasted “We’ve built a better brain” and said that expert systems never get sick and never resign or retire early. The reporter wrote that the sellers' words sometimes sounded like “prophets proclaiming a new age”.

What we still don’t know:The article does not name the company that sold TIMM.

Sendai's subway opens, using “predictive fuzzy control” (translated) (AI with rules that handle vagueness) for automatic train operation

The Sendai city subway (the ninth city subway in Japan) opened on July 15, 1987. The automatic train operation system developed by Hitachi adopted “predictive fuzzy control” (translated). It used AI (knowledge engineering) rules written in “IF–THEN form” (translated) (the form 'if so-and-so, then do so-and-so') to replace the skills of expert drivers. The article says that better ride comfort, better stopping accuracy and energy saving were achieved, and introduces this predictive fuzzy control as “the first in the world” (translated) (the words of the article by the company that made it).

What we still don’t know:“First in the world” (translated) is the words of an article by the company that made it, and has not been confirmed in other sources.

The 1981 forecast: it is said that in the early 1990s computers with eyes, ears, and a mouth, that understand speech and can do automatic translation and interpreting, will appear

The report of the survey before the plan (1979–1981; chaired by Tohru Moto-oka of the University of Tokyo) stated that the goal was to complete a prototype in the year 1990. A booklet explaining it for the general public wrote that fifth-generation computers are “said to appear in the early 1990s” (translated) and that research must start early to “realize a prototype in 1990” (translated) (to build a trial machine). As the pictures of what they aimed for, it listed eight functions, including “computer with eyes, ears and a mouth” (translated) that can handle sound, figures, images, and characters as they are; “computer that understands what is said” (translated) that can be spoken to in everyday language (and so can also do automatic translation and interpreting); computers that have common sense and solve unfamiliar problems by reasoning; and computers that can be used without writing programs.

What happened next:The project ran in fiscal years 1982–1992. According to the 1993 final evaluation report, the goal for inference speed (100 million per second) was met and exceeded, at over 150 million per second. On the other hand, the report says that systems that interact with people by voice or images turned out to need a huge amount of background knowledge, so partway through the project narrowed its focus to a base for language research and did not get as far as building concrete systems. In 1992 the director, Kazuhiro Fuchi, said that building application systems such as machine translation had been dropped from the goals when the project began in 1982.

What we still don’t know:The “Statement of R&D Topics” (translated) (1981; 7 groups, 26 tasks), which are said to set numerical goals such as the number of words translated and accuracy, could not be found and have not been checked.

Fifth Generation results: the Parallel Inference Machine PIM and the language KL1 (the largest machine had 512 computers; the five models together exceeded 1,000)

The Fifth Generation project developed the “Parallel Inference Machine PIM”, which links many computers to carry out inference (drawing conclusions from rules) at the same time, and its dedicated language “KL1”, and presented them at the international conference FGCS'92 in June 1992. KL1 was designed by Takashi Chikayama, based on the mechanism of the language GHC devised by researcher Kazunori Ueda (from a talk by director Kazuhiro Fuchi). Five models of PIM with different specifications were built; the largest, PIM/p, used 512 computers and could perform 156 million inferences per second. Counting all five models together, there were more than 1,000 computers. The final evaluation states that the target of 100 million inferences per second was exceeded.

What we still don’t know:We have not confirmed how PIM was used after the project (how many machines ran where, and for how many years)

The story that “the Fifth Generation Computer failed”: what the people involved, the evaluation committee and outside researchers each wrote

In June 1992, right after the international conference that wrapped up the project (FGCS'92), an editorial in the newspaper The Baltimore Sun (written for American companies) wrote of the project's “failure ... finally acknowledged June 1 at a Tokyo seminar”. But in the keynote speech at that conference, the director, Kazuhiro Fuchi, introduced the situation by saying “reviews have appeared in Japan and abroad saying that this project is a failure” (translated). He then explained that big images such as “trying to settle the hard problems of artificial intelligence (AI) in 10 years” (translated) and “trying to build a human-level machine translation system” (translated) had spread from the 1981 conference before the project. He said that the project begun in 1982 had a more modest content (parallel inference), and that he believed it was a success. The promotion committee's final evaluation (1993) also said it had “fully met the initial expectations” (translated). A report of the Ministry of International Trade and Industry's research institute (1995) wrote that calling it a failure because it did not become a product is an opinion that does not fully understand the project's aims (basic research and international contribution). At the same time, the same report also writes that the structure of the computers made is not expected to be used in future technology, because of the progress of small computers. A history book by an outside researcher (Nilsson 2010) writes that many people think most of the results now have only historical significance.

What we still don’t know:Where the reputation of “failure” first spread from in the United States (for example the June 1992 New York Times article “‘Fifth Generation’ Became Japan’s Lost Generation”) has not been confirmed, because we could not open that article.

A paper on training multilayer neural networks with backpropagation appears in Nature

Rumelhart, Hinton and Williams announced a way of training neural networks that have “hidden layers” between input and output. The difference between the output and the correct answer (the error) is passed back from the output side to the input side step by step, and the strengths of the connections are updated bit by bit (passing the errors backward, as in the paper's title “Learning Representations by Back-Propagating Errors”, is backpropagation). They showed by experiments that the hidden layers come to represent features that nobody taught them.

What we still don’t know:People could read what is represented inside the hidden layers in this paper's small example (a family tree). But in large networks, only parts of it are understood even now (this connects to the unit on the fourth boom).

The worst-case scenario of “The AI Winter” drawn in a 1984 discussion ── checking the answers afterward

At the conference of the American Association for Artificial Intelligence (AAAI), a panel discussion was held titled “The ‘Dark Ages’ of AI – Can We Avoid Them or Survive Them?” The moderator, Drew McDermott, said that behind the bustle in industry, there was deep unease among people who had researched for many years, from a worry that “perhaps expectations about AI are too high ... will eventually result in disaster”. As the worst-case scenario he drew this picture: in five years America's “strategic computing initiative” fails, Japan's Fifth Generation also hits a dead end, start-up companies go under, funding for AI stops, and everybody changes the names of their research to something else. He introduced the fact that some people call this state “The AI Winter.” But he also said he thought it unlikely. In the discussion, another researcher (Roger Schank), when asked whether he thought a catastrophe would come, said that he told them it would. On the other hand, the view “Will there be an AI Winter? I think there may be an AI dusk” was also put forward (the same person went on to say that the bubble might not burst but shrink a little).

What happened next:Afterward, development of the autonomous land vehicle (ALV), one of the plan's showpieces, was cut off in April 1988. The head of the responsible division, who took up the post in 1987, also cut off parts of the AI program. The plan disappeared from DARPA's budget in 1993. Membership in the American Association for Artificial Intelligence peaked at 16,421 in 1987 and settled at 4,000–5,000 by 1996. Many of the AI companies developing expert systems disappeared, were acquired, or changed their business. Japan's Fifth Generation Computer Project ended its main stage in fiscal 1992. Nilsson writes that these events piled up in the mid-to-late 1980s and came to be called “The AI Winter.”

What we still don’t know:Whether one of the scenarios, “Everybody hurriedly changes the names of their research projects to something else”, really happened in the 1990s has not been confirmed in original sources (a candidate dealt with in u08)

LSTM: a neural network that can remember the earlier parts of a long sentence or sound is announced

Hochreiter and Schmidhuber announced LSTM, a neural network for time-series data (sequences of sounds or letters) with a built-in “memory cell” (a box that remembers) and “gate units” that open and close what goes into and out of the box. With earlier methods, the signal used for correction in backpropagation either faded away (approached zero) or blew up as it went back in time, so they could not remember things from long ago. LSTM could hold on to and use clues from the past more than 1000 steps away.

Vapnik and colleagues create the support vector machine, which separates data by drawing the widest road between them

When drawing a line to separate two kinds of points, you draw the line down the middle of the widest road, as far as possible from both kinds of points. Then even points seen for the first time are separated well. Vapnik and colleagues made it so that this idea can be computed from only the few points at the edges of the road (the support vectors: the points that hold up the road). On the handwritten digits of MNIST it reached an error rate of 1.1% (Cortes and Vapnik), close to the neural network LeNet. By around 2000, this kind of method had become the mainstream in practical use.

Fei-Fei Li and colleagues at Princeton University present ImageNet, millions of named photos; the people who checked them were workers hired over the internet

Li and colleagues presented ImageNet, built on WordNet, a dictionary that organizes concepts in a hierarchy, and collecting an average of 500–1,000 photos for each of about 80,000 groups of nouns (the paper's example is “Burmese cat”). At the time of publication it had reached 12 areas, 5,247 categories and 3.2 million images. The images were collected with search engines, and workers around the world, paid through Amazon Mechanical Turk, checked one by one whether the target object was really in each picture, deciding by the votes of several people.

What we still don’t know:The paper does not state the number of workers, the countries they lived in, or the amount they were paid (unconfirmed).

In the image-recognition competition ILSVRC 2012, the University of Toronto's deep convolutional neural network (SuperVision) wins by a wide margin. The runner-up is a University of Tokyo team

In a competition to identify 1,000 kinds of objects from photos, the team “SuperVision” of Krizhevsky, Sutskever and Hinton held its error rate, under a rule allowing 5 guesses, to 15.3% (16.4% if only the data supplied by the competition is used). The runner-up, a University of Tokyo team (ISI), scored 26.2%, a gap of about 11 points. They trained an 8-layer neural network with 60 million parameters on 2 gaming GPUs (graphics chips, made by NVIDIA). In a later paper, the competition's organizers wrote that “This was a turning point for large-scale object recognition” for research on recognizing a wide variety of objects.

Hinton and colleagues present a way to train a deep neural network “one layer at a time”. It is seen as an important step toward “deep learning”

Neural networks with many stacked layers could not be trained well by backpropagation alone. Hinton and colleagues showed a method in which, first, one layer at a time and without being given correct answers, the network learns the “features” of the data (pre-training), and then the whole network is adjusted by backpropagation. It reached an error rate of 1.25% on the handwritten digits of MNIST (the Neural Computation paper), and it was also better than earlier methods at representing images and documents with a small number of values (the Science paper).

The $1 million contest to predict movie tastes, the “Netflix Prize” (2006–2009): the winner was decided by a 20-minute margin

In October 2006, Netflix, which then rented out movies on DVD, opened a contest. It offered a $1 million prize to anyone who could predict movie ratings 10% more accurately than its own recommendation system (Cinematch). It distributed more than 100 million ratings (1 to 5 stars) of about 18,000 movies by about 480,000 people, with names replaced by numbers. About three years later, on September 21, 2009, the team BellKor's Pragmatic Chaos won. The second-place team had the same score (error 0.8567, a 10.06% improvement), but the side that submitted about 20 minutes earlier became the winner (under the rule that a tie goes to whoever submitted first).

AlphaGo beats Lee Sedol, one of the world's top players, 4 games to 1 (Seoul)

In March 2016 in Seoul, South Korea, AlphaGo played a five-game match with no handicap against Lee Sedol, a 9-dan player who had won 18 world titles. AlphaGo won 4 games to 1. In Game 4, Lee Sedol won with a surprising move at move 78. DeepMind announced that it would donate the 1 million dollar prize to education and Go organizations and to UNICEF.

The two papers behind “Monte-Carlo Tree Search”, which made Go programs strong

In Go, there are many possible moves in a single position, and evaluating a position is hard. So search methods like those used for chess did not make programs strong. In 2006, a paper by Coulom of France presented a new search method (Monte-Carlo Tree Search, MCTS). It plays out a game at random many times from a position to the end (Monte Carlo), compares moves by their win rates, and examines the more promising branches more deeply. With his 9×9 Go program Crazy Stone, Coulom won a computer Go tournament (KGS). In the same year, Kocsis and Szepesvári of Hungary proposed UCT, a method for deciding which branch to try and how much. Ten years later, the AlphaGo paper (2016) cited these two as the MCTS literature, and combined MCTS with two neural networks (one predicting the next move and one evaluating positions).

The Information Processing Society of Japan declares an end to its project, saying computer shogi has “caught up with top professional players”

From 2010 the Information Processing Society of Japan (IPSJ) had been running a project aiming to achieve “computer shogi that beats top professional players” (note 1). In October 2015 it declared the project finished, saying “analysis results show that it has caught up with top professional shogi players as of 2015”. “Caught up” did not mean winning every time; it was explained to mean that, over many games, the computer is likely to come out ahead. At that point, however, a direct match against a top professional had not yet taken place.

What happened next:About a year and a half later, on April 1 and May 20, 2017, the shogi software PONANZA beat Amahiko Sato, the Meijin (the top title holder; he played as Eiō), in both games. The organizers ended the Denōsen matches with this contest.

Oxford University study: “47 percent of total US employment is in the high risk category” (risk of being replaced by computers; 2013)

In September 2013, Carl Benedikt Frey and Osborne of the University of Oxford published a paper estimating the probability that computers would replace workers in 702 occupations in the United States. It says that 47% of U.S. employment falls “in the high risk category”. The figure was later taken as “a prediction that jobs will disappear” and became a target of checking (for example, the title of a 2022 ITIF article, “Oops: The Predicted 47 Percent of Job Loss From AI Didn’t Happen”, from a separate record this site keeps, checking predictions about work against what happened). The details and the later checking are recorded in a separate study, apart from this site, on checking predictions about work in the age of AI.

The Paper on “the Transformer”, the Basis of Today's Text AI, Is Published

Eight people, researchers at Google and a University of Toronto student, presented a new AI architecture (the Transformer) built only from “attention mechanisms,” which pay attention to which words in a sentence relate to which other words. The paper tested it on tasks such as translating English into German and French, and it achieved better results than before with shorter training time. The later GPT and BERT are built on this architecture (confirmed in their respective papers).

The Conversational AI “ChatGPT” Is Released as a Free “research preview”

OpenAI released the conversational AI “ChatGPT”. The announcement says it can answer follow-up questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests. Inside, it is a sibling of GPT-3.5 whose training finished in early 2022, tuned by having humans score its answers. It was free during the “research preview,” and the stated purpose was to collect users' feedback.

What we still don’t know:As one reason it is hard to fix mistakes, the announcement gives that “during RL training, there’s currently no source of truth” (RL means reinforcement learning). Why and when plausible-sounding mistakes (later called “hallucinations”) happen was something the developers, even at this point, still wrote “Fixing this issue is challenging”.

“ChatGPT Reached 100 Million Users in 2 Months” Is Not a Company Announcement but an Estimate by a Bank's Analysts from Web Visit Counts

In February 2023, news spread around the world that “ChatGPT was used by 100 million people two months after release and became the fastest-spreading app in history.” Traced back, the source was a memo by analysts at the investment bank UBS. Using figures from the web analytics company Similarweb, it concluded that ChatGPT “is estimated to have reached 100 million monthly active users in January” (translated) (the quoted words are the wording of the Reuters article that reported the memo). Reuters wrote “estimate,” and Similarweb had not answered requests to confirm the figures at that time. In the body of The Guardian's article, it is presented as Similarweb's figure of “about 590m visits in January from 100 million unique visitors” (translated).

A Paper on a Law Where Performance Rises at a Fixed Rate as Models Grow — but the Authors Themselves Note “Our trends must eventually level off”

Researchers at OpenAI and Johns Hopkins University showed that a language model's results (measured by how hard it is to guess the next word) improve in a clean, fixed form (a power law) as the number of parameters, the amount of training data and the amount of computation increase. This form held over more than seven orders of magnitude (10 million times). This idea became one of the foundations of the later race to build ever larger AI.

What we still don’t know:What is known: in the range studied up to 2020, performance rose by a fixed rule as scale increased. What is not known: how far this continues and why it happens (the paper itself writes that the trends “must eventually level off” and that “we do not have a solid theoretical understanding”).

Will AI Win Gold at the Math Olympiad by 2025? In 2022 Two Researchers Forecast Under 8% and Over 16%; the Result Came in July 2025

The AI researchers Paul Christiano and Eliezer Yudkowsky differed in their views on the progress of AI. So, as a prediction that can be settled one way or the other, they publicly gave their probabilities that, at the International Mathematical Olympiad (a high-school competition), an AI “made before the IMO is able to get a gold … one of 2022-2025” (translated) (the rules, such as time, follow the IMO Grand Challenge). Christiano said under 8%, and Yudkowsky said 16% or more. For solving the hardest problem, Christiano predicted under 4%. In July 2025, Google DeepMind announced that an enhanced version of Gemini had fully solved 5 of the 6 problems and earned 35 points out of 42 (a gold-medal score), formally graded and confirmed by the competition's graders. It wrote its answers in ordinary language within the 4.5-hour time limit.

What happened next:In July 2025, an enhanced version of Google DeepMind's Gemini earned 35 points out of 42 (a gold-medal score). The competition's graders formally graded it, and the president of the International Mathematical Olympiad confirmed it as “a gold medal score” (translated) (however, according to the blog, the confirmation was of the answers, not a check of how the AI works). It was done within the 4.5-hour time limit, and the hardest problem (Problem 6) was not solved (5 of 6 problems). Because both predictions are probabilities, neither can be called right or wrong outright, but the actual result was closer to the side that had put the probability higher (over 16%).

International report by 96 experts (including an advisory panel chosen by 30 countries and others): “The inner workings of these models are largely inscrutable, including to the model developers” and “experts disagree on what to expect even in the coming months and years”

Following the Bletchley meeting (2023), 96 experts chaired by Yoshua Bengio (including an advisory panel chosen by 30 countries and the UN, EU and OECD) compiled an international report on AI safety (Hiroaki Kitano took part from Japan). The report explains that general-purpose AI models are “not programmed in the traditional sense” but are built through learning, and wrote that “The inner workings of these models are largely inscrutable, including to the model developers” and that “interpretability techniques for explaining why … any given output remain severely limited”. On future progress, it said the capabilities of general-purpose AI “could advance slowly, rapidly, or extremely rapidly”, and that experts “disagree about the pace … different views on the promise of further ‘scaling’”. The second report of February 2026 also points out that “researchers cannot reliably predict when specific capabilities will emerge” and that “experts disagree about whether exponential increases in inputs will continue” (inputs: computation, algorithm improvements and data).

What we still don’t know:What we still don't know (points made by experts chosen by the countries): (1) The inner workings of modern large AI (the inner mechanisms are largely unreadable even to the developers, and explanation techniques are severely limited). (2) The pace of future progress (anything from slow to extremely rapid is possible; experts disagree about how far scaling will keep working). The 2026 report shows the same uncertainty continuing.

A study finding that AI predictions mostly fall “15 to 25 years in the future” (the paper's first page now carries the note “The findings in this paper are based on a dataset error”)

A paper that collected and analyzed 95 predictions of when AI would arrive. The most common prediction was “15 to 25 years in the future”, and it concluded that predictions by experts and non-experts cannot be told apart. The first page of the publicly available PDF now carries the note “The findings in this paper are based on a dataset error”. The error came from a misreading of the column classifying whether someone is an expert (pointed out by AI Impacts in 2016).

What we still don’t know:Whether predictions cluster around 20 years ahead changes depending on how the data are built. It is not clear.

Tracing a national teaching material's prediction, “By 2025, performance will reach the point where a single human brain can be fully simulated” (translated), back to its source — it began as a line drawn on a supercomputer-speed figure in Kurzweil's 2005 book

Teacher-training material (2020) from the Ministry of Education, Culture, Sports, Science and Technology for the high-school course “Information II” (Japan's high-school informatics course that follows Information I), based on a document from the Ministry of Internal Affairs and Communications' 2014 White Paper on Information and Communications, says this: if computing speed keeps rising as it is, by 2025 it will reach the performance needed to fully simulate (imitate and run inside a computer) one human brain. Tracing the source, the White Paper carried this statement not in its main text but in Figure 3-1-1-4, “Improvement in Computing Speed”. The figure's source is a document presented at a Ministry of Internal Affairs and Communications meeting on February 13, 2014 by Mr. Kitani (NTT Data Technology Development Headquarters, according to the document's cover). That document says it was made from a figure in Ray Kurzweil's 2005 book The Singularity Is Near. The original figure plots supercomputer speeds (FLOPS: how many decimal-number calculations per second) as dots, with a straight line (Trendline) that doubles every 1.2 years, and at the point where that line reaches 10¹⁹ FLOPS it carries the note “Required for Human Brain Neural Simulation for Uploading (2025)”. In other words, the original was not brain simulation itself but a forecast of computing speed.

What happened next:Checking computing speed: by the same measure as the original figure (the HPL value used in the TOP500 supercomputer performance ranking; the original figure also cites TOP500 as its source), as of November 2025 the leader, El Capitan (USA), is at 1.809 exaflops (about 1.8×10¹⁸ operations per second), only about one fifth (1/5.5) of 10¹⁹ FLOPS. LineShine (China), which took first place in June 2026, is at 2.198 exaflops and has not reached it either. On the other hand, in HPL-MxP (formerly HPL-AI), a different ranking that measures a method mixing in calculations with fewer digits (mixed-precision), three machines passed 10¹⁹ FLOPS in November 2025: El Capitan (16.680), Aurora (11.643) and Frontier (11.390). 10¹⁶ FLOPS (predicted in the figure for 2013) had been reached by Japan's “K computer” in November 2011 at 10.51 petaflops. Brain simulation results: in 2013, using all the compute nodes of the “K computer”, a simulation of 1.73 billion nerve cells and 10.4 trillion synapses (about 1% of the human brain's neural circuitry) took 40 minutes to reproduce 1 second of brain activity (RIKEN and others). In 2018, RIKEN wrote that even the highest-performance supercomputer of the time could not simulate a whole human brain, and announced an approach toward achieving it on next-generation supercomputers from 2020 on. In November 2025 the Allen Institute, the University of Electro-Communications and others announced that they had used “Fugaku” to simulate in detail the whole cerebral cortex of a mouse (about 10 million nerve cells, 26 billion synapses and 86 brain regions). The Allen Institute's Anton Arkhipov said that the long-term goal is to model the whole brain, and eventually a model of the human brain. None of the announcements from the parties concerned that I opened on this site said that a whole human brain had been simulated by 2025.

What we still don’t know:How much computation is really needed to simulate a whole human brain differs greatly depending on who estimates it. A note in another chapter of the same 2014 White Paper gives a forecast, from IBM's Modha in 2008, that simulating a human brain would become possible in 2018 at 38.5 petaflops (about 4×10¹⁶). That is a gap of about 260 times from Kurzweil's figure of 10¹⁹. The speed needed changes with how finely nerve cells are imitated (for example, the 2025 simulation on “Fugaku” was announced as imitating even the flow of electricity inside the branches of nerve cells). This site has not been able to confirm whether there is no example anywhere of a whole human brain being simulated by 2025 (I opened only three announcements from the parties concerned).