The Fourth Boom, Now

How the Japanese Society for Artificial Intelligence's “History Scroll of Artificial Intelligence” (2023) Divides the Booms and Winters

Over one year, the Japanese Society for Artificial Intelligence drew the history of AI as a picture scroll on the covers of its journal. Its division of time is: picture 1 = from the start of the field to the first boom (late 1950s–1960s, fading from the late 1960s as limits came into view); picture 2 = from the winter to the second boom; picture 3 = the second winter (around 1992 to around 2010); picture 4 = the first half of the third boom (around 2010–2016, triggered by the 2012 image recognition contest); picture 5 = the second half of the third boom (2017 on; “it may already be called the fourth AI boom” (translated); the original uses a formal word for “by now”) through the generative AI boom; picture 6 = the future as drawn by researchers.

What we still don’t know:Even in the scroll, the “Fourth” (translated) is only put as “may be fine to call it” (translated), and the society has not decided on it.

Japan's Ministry of Internal Affairs and Communications white paper (2024) says that with the spread of generative AI from around 2022, “it is also said that we have now entered the fourth AI boom” (translated)

A national white paper (the Ministry of Internal Affairs and Communications' White Paper on Information and Communications in Japan, 2024 edition) explained that the history of AI “has repeated booms and winters several times” (translated). It then wrote that “with the rapid spread of generative AI from around 2022” (translated) (generative AI is AI that creates text and images), “it is also said that we have now entered the fourth AI boom” (translated). It does not state this flatly, staying with the phrase “it is also said” (translated). In the same section it also writes of the third boom as “from the 2000s” (translated) and “the third AI boom that continues to this day” (translated). In the figure (Figure I-3-1-1), the band for the fourth boom starts in 2022, and the figure says “moving into a worldwide AI boom without passing through a winter era” (translated).

Teikoku Databank (2026): the fourth AI boom is from late 2022, and the next is “AI that executes”

A report by Teikoku Databank (TDB Review No. 46) positions the current boom as the “fourth AI boom” (translated) and divides the periods. The first boom is the late 1950s–1960s, the second is centered on the 1980s and runs to the early 1990s, the third is from around 2010, and the fourth is from late 2022 (generative AI centered on large language models). It then names “AI that executes” (translated), centered on AI agents and physical AI, as “the next AI boom” (translated).

What we still don’t know:Whether “the next AI boom” (translated) (the one that would be the fifth) will come is not yet known. The report itself writes “for convenience, we call it ‘the next AI boom’” (translated).

Japan's Ministry of Education (MEXT) white paper (2024): the third boom is dated from around 2010, and “it is also said that we are now approaching a ‘fourth boom’” (translated)

MEXT's white paper explained that AI technology has expanded its research and development efforts “while repeating booms several times” (translated). It places the first boom in the late 1950s–1960s, the second in the 1980s–1990s, and the third from around 2010. It then said that “in response to advances in generative AI technology and so on” (translated), “it is also said that we are now approaching a ‘fourth boom’” (translated). Like the Ministry of Internal Affairs and Communications' 2024 edition's “it is also said” (translated), it uses the hedged wording “it is also said” (translated). The figure it printed (Figure 1-1-1) comes from a report of the Japan Science and Technology Agency (JST); it draws the first to third booms as three rising curves, and beyond them it labels “Fourth-Generation AI” (translated) and “Trusted AI” (translated), without writing “the fourth boom” (translated).

What we still don’t know:JST's “Fourth-Generation AI” (translated) is a different way of counting from the “fourth boom” (translated) (a generation number of the technology). Even though both are “4” (translated), they point to different things.

Japan's Ministry of Internal Affairs and Communications stops writing “fourth” in its 2025 and 2026 white papers (in the range we looked at)

The 2024 white paper wrote “it is also said that we have now entered the fourth AI boom” (translated). But in the 2025 and 2026 editions, the word “boom” appears only once in the whole white paper. They stay with saying that in 2006, “since deep learning was proposed, the third AI boom has arrived” (translated), and that in 2022 generative AI “began to catch on in earnest” (translated), and do not call it “fourth.” Even in the same government's white paper, the way of naming changes from year to year.

Japanese Society for Artificial Intelligence roundtable (2023): asked whether this is the fourth boom, “Nobody has authorized it, right?” (translated), “It may be about a 3.1st boom” (translated), and AI may have become “a field that doesn’t cool down” (translated)

In a roundtable by the people who made the history picture scroll, Fujio Toriumi asked, “So is it now treated as the fourth AI boom after all?” (translated). Hitoshi Matsubara answered, “Some professors say so” (translated), and Tadahiro Taniguchi said, “Whoever says it first wins, in a way (laughs)” (translated). When Toriumi asked again whether anyone has actually decided it, Matsubara answered, “Nobody has authorized it, have they?” (translated) (authorize means to approve formally), and “But I also feel it has stayed up since the third boom” (translated). Mai Sugimoto of Kansai University's Faculty of Sociology said, “To begin with, globally, how to treat the second AI boom and the second winter after it is delicate” (translated) and “In Japan, I think the influence of the Fifth Generation Computer project was big” (translated), and added that “unless funding is cut off, it does not leave a picture (in the public mind) of having gone down” (translated); since funding never ran out from the third AI boom onward and it turned into the generative AI boom, she said “I don't think it counts as the boom having ended once” (translated). Toriumi said, “It may be about a 3.1st boom” (translated), and Sugimoto continued: “it has become a field that grows even without relying on a boom, a field that doesn’t cool down” (translated). The exchange ends with Toriumi's “Another winter might come (laughs)” (translated) and Sugimoto's reply, “We may have just planted a flag (laughs)” (translated) (“planting a flag” is an expression meaning an omen that something is about to happen).

What we still don’t know:If you mark the boundary between booms by whether funding was cut off, there is no boundary between the third boom and now. If you mark it by the size of public attention, there is a boundary in 2022–2023. The answer changes depending on which standard you use.

The phrase “fourth AI boom” never appears in the Diet records (within the range checked in September 2026)

Searching the Diet Record Search System, a national system in which every statement made in the Diet can be searched, “third AI boom” appears in 6 statements between 2017 and 2025. “Fourth AI boom” (written with kanji or with a digit numeral) appeared 0 times. In April 2025, a minister (Minoru Kiuchi; in the record his title is Minister of State for Special Missions, Cabinet Office) answered with “since the so-called third AI boom” (translated). In another statement the same month, the phrase “generative AI boom” (translated) was used. The Diet does not call the current boom by a number.

What we still don’t know:The record search picks up only exact matches of the words. Wordings with another word in between, such as “fourth-stage AI boom”, may not have been picked up (the combination “AI boom fourth” also gave 0 results).

In 2017, the “fourth AI boom” (translated) Meant a Future Boom Triggered by the Arrival of General-Purpose AI (AI That Can Do Many Kinds of Work)

In the preface to the special feature “AI Society Theory” (translated) in the journal of the Japanese Society for Artificial Intelligence, Koichi Takahashi and Tomohiro Inoue introduced Inoue's paper as one discussing “the fourth AI boom, which will probably be triggered by the emergence of general-purpose AI” (translated), a boom beyond the one at the time (general-purpose AI means AI that can do many kinds of work the way a person does). The same text also says that, as a boom in the literal sense, the boom of the time (the third) would peak in 2016 or 2017 and then settle into steady industrial work.

What we still don’t know:When the peak of the third boom came depends on what is measured (number of papers, funding, number of articles, and so on). We have not yet checked with numbers whether the 2017 forecast (a peak in 2016–2017) was right.

[Japan] In the Diet and in book titles, the second peak is in 1990 (about 3 years later than in English) and the third in 2019; rising again from 2023

From the Diet’s records of speeches and from Japanese book titles, we counted how many times “artificial intelligence” and “AI” (recounted with parts of other words removed, so that it means AI) appear. In book titles, “artificial intelligence” is highest in 1990 (37 books). In the Diet, “artificial intelligence” peaks in 1985 (7 mentions) and 1990 (9), and was almost 0 from 1993 to 2014. In the Diet, “AI” meaning AI went 11 in 2015 → 593 in 2019 → down to 336 in 2021 → 761 in 2023 → 2,424 in 2025 (the year the AI law was made). “Generative AI” was 0 until 2022 and reached 172 in 2023.

What we still don’t know:The reason the counts fell in 2020–2021 cannot be pinned down to one cause (it overlaps with the period of the infectious disease epidemic).

Yutaka Matsuo: “I think we can say from this year on that it’s the ‘fourth AI boom’” (translated) (2023) — We Found the Original Article

In an interview, Yutaka Matsuo of the University of Tokyo said that “with the arrival of ChatGPT it has been settled that AI will clearly change the world” (translated), and so “I think it is fine to say ‘fourth AI boom’ from this year” (translated). In the same interview he also says that he already knew AI had great power around 2020, when GPT-3 came out, and that he expects a bigger change than when the internet appeared.

What we still don’t know:We do not know whether Matsuo said “fourth” (translated) earlier than this. His slides (31 slides) for the Liberal Democratic Party meeting on 2023-02-17 and the materials for the 1st and 2nd meetings of the Cabinet Office AI Strategy Council (May 2023) did not contain the word “boom” (translated) itself (a separate record).

“Matsuo Said It Was the Dawn of the Fourth AI Boom” Grew Stronger in the Telling

In the original interview (March 13, 2023, by the NHK “Science ZERO” reporting team), Yutaka Matsuo's words were “I think it is fine to say ‘fourth AI boom’ from this year” (translated). The article's subheading was “The Dawn of the ‘Fourth AI Boom’” (translated), and the title of the NHK blog was “The Arrival of the Fourth AI Boom” (translated) (according to a note in a paper by Tagui Ichikawa of Hitotsubashi University). After that, the title of a Toyo Keizai article (April 2023) became “The Fierce Contest of the ‘Fourth AI Boom’ Opened by ChatGPT” (translated) (“fierce” here means intense). Ichikawa's paper (May 2023) wrote that the present situation is “as if” (translated) “A new fourth AI boom has dawned (Matsuo)” (translated). A text by Waichi Sekiguchi of the Japan Research Institute (June–July 2023 issue) said “Professor Yutaka Matsuo points out that ‘the AI boom has already entered the fourth’” (translated), and an Aizawa Investment University article (April 2024) wrote that Yutaka Matsuo “said in 2023 that it was ‘the dawn of the fourth AI boom’” (translated). His own phrasing, “fine to say” (translated), changed into “has entered” (translated) and “dawn” (translated) as it was quoted.

Yutaka Matsuo (July 2023): “It may be fair to say the fourth AI boom has arrived” (translated), but it is “a phenomenon of a different dimension from the word ‘boom’” (translated)

In an interview with Toyo Keizai, Yutaka Matsuo said, “There have been three AI booms in the past” (translated), and that “the spread of generative AI may be fine to call the arrival of the fourth AI boom” (translated), but that “What is happening now is a phenomenon of a different order from the word ‘boom’” (translated). He also said generative AI would spread through society all at once over the next few years. As with “it is fine to say” (translated) in the March interview, here too he uses the hedged phrasing “may be fine to say” (translated). At the same time, he does not see it as a boom (a fad that eventually ends).

What we still don’t know:Whether the view that it is “on a different level from the word ‘boom’” (translated) proves right (that is, whether no winter comes) is still unknown. Whether people also said “this time is different” during past booms needs to be checked by comparing with this site's records of each boom.

Yutaka Matsuo's February 2023 Slides for a Political Party Do Not Contain the Word “boom”

A full-text search of Yutaka Matsuo's slides “AI Evolution and Japan’s Strategy” (2023-02-17, 31 slides used at the 2nd meeting of the Liberal Democratic Party's “Project Team on the Evolution and Implementation of AI”) found neither the word “boom” nor the word “fourth.” His slides (33 slides) for the Cabinet Office's AI Strategy Council in May 2024 (9th meeting), and the materials (11 documents) for the 1st and 2nd meetings of the same council in May 2023, also had no “boom.” In the materials of national government meetings, the practice of counting booms by number is hardly used (among those we looked at).

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”).

U.S. AI policy turns sharply in 2023–2026 (changes in presidential executive orders)

The United States has no nationwide AI law and has set its policy through executive orders (orders the president gives to government agencies). Executive Order 14110 of October 2023 was titled “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” and required reports to the government for AI systems and others trained with more than 10^26 operations of computation. In January 2025 the next president revoked that order (14148) and ordered a review of obstructive regulations in order to “sustain and enhance America’s global AI dominance” (14179). In July he released “America’s AI Action Plan”. Executive Order 14365 of December called for a single federal framework, to avoid state-by-state AI laws becoming “a patchwork of 50 different regulatory regimes”, and had a body set up under the Attorney General to challenge state laws in court. Executive Order 14409 of June 2026 set up a mechanism that defines powerful AI that could be misused for cyberattacks as a “covered frontier model”, and a voluntary framework in which developers let the government try the model for up to 30 days before release. It also states clearly that no national licensing system will be created.

What we still don’t know:Even “America's AI Action Plan” of 2025, which pushes deregulation, points out that “the inner workings of frontier AI systems are poorly understood” and that they often cannot explain why they gave an answer, and says it will focus on research into this (PDF page 12).

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”).

“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).

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”.

GPT-4 arrives: it scores high on exams, while its maker declares it will not report the model's size or how it was built

OpenAI announced the text AI “GPT-4”. Its technical report said that on a test imitating the U.S. bar exam, GPT-4 reached “a score around the top 10% of test takers”. But the same report also declared that, given both the “competitive landscape” and the “safety implications” of large models like GPT-4, it would give “competitive landscape … safety implications … no further details about the architecture (including model size)”, the hardware, the amount of training computation, how the data was built, the training method, and so on. Because the GPT-2 and GPT-3 papers had stated the number of parameters, this was a big change from the earlier trend of sharing how models were made.

What we still don’t know:Because the developer stopped publishing size, data and computation, outside researchers can check less about large AI systems from GPT-4 onward (the report itself says these details are withheld).

In Spring 2023, Companies in the US and China Release Text AIs One After Another Within About a Month

Within three to four months of ChatGPT's release, other companies also put out text AIs one after another. Meta (the Facebook company) released LLaMA, handed out in full to researchers; Anthropic released Claude for companies; Baidu released Wenxin Yiyan, which is strong in Chinese (invitation only); and Google released Bard as “an early experiment” (translated) (first in the United States and the United Kingdom). The ways of releasing differed by company: handing out the model itself, lending it to companies, or starting with a few people.

Japanese researchers come together to start building “LLM-jp”, a Japanese text AI that is fully open, from its inner workings to its training data

The “LLM Study Group (LLM-jp)” launched by Japan's National Institute of Informatics (NII) in May 2023 drew more than 500 participants from universities and companies. In October, it released “LLM-jp-13B”, a text AI with 13 billion parameters that is strong in Japanese, fully open, including the texts used for training (the corpus). NII frankly wrote that its performance was “comparable to models previously released in Japan” (translated). In April 2024 it set up a dedicated center, and in December it released a model of the same size as GPT-3 (about 172 billion parameters), calling it “the largest model in the world that is fully open, including the training data” (translated). In April 2026 it released the next generation (LLM-jp-4), trained on about 12 trillion words.

China's DeepSeek releases its reasoning AI “DeepSeek-R1” in a form anyone can use; a week later, the share price of NVIDIA, the AI GPU company, falls 17% in one day

The Chinese company DeepSeek released “DeepSeek-R1”, a text AI that scored on a par with the U.S. company OpenAI's o1 on math and programming problems. It was distributed including its trained data (weights), under the MIT license, which allows commercial use and modification. A week later, on January 27, amid worries about tougher competition from DeepSeek and that investment in AI parts (GPUs) might level off, the share price of the U.S. company NVIDIA, which makes those parts, fell 17%. Its market value dropped by nearly $600 billion in one day, which was reported as the biggest one-day loss by a U.S. company at the time.

The Picture-Making AI “Stable Diffusion” Is Released in a Form Anyone Can Run on Their Own Computer

Stability AI, a company in the UK, released the trained model of “Stable Diffusion,” an AI that generates images from text, in a form anyone can download and use. According to the announcement, it can generate an image in seconds on a home GPU (a chip for graphics processing), and its license allowed commercial use (with usage restrictions). In the announcement for researchers 12 days earlier, it says development was led by Patrick Esser of Runway and Robin Rombach of CompVis, a research group at a German university. For training, pairs of images and captions collected from the internet (part of LAION-5B) were used.

Open letter says to “immediately pause for at least 6 months” the training of “AI systems more powerful than GPT-4”; six months later, no pause was reported

The U.S. organization Future of Life Institute issued an open letter calling on all AI labs to “immediately pause for at least 6 months” the training of “AI systems more powerful than GPT-4.” During the pause, it said, safety rules should be created, and if the labs could not pause, governments should step in. The letter ends with a seasonal metaphor: “enjoy a long AI summer, not rush unprepared into a fall”. Six months later, the organization's statement did not say that training had paused; it called for government regulation, and its president spoke of “The ongoing arms race”.

Geoffrey Hinton, one of the researchers behind today's AI, leaves Google and begins speaking about the dangers of AI

Geoffrey Hinton (then 75), who had worked on neural networks since the 1980s, revealed that he had left Google and began speaking about the dangers of AI. He told the BBC that, as far as he could tell, “Right now, they’re not more intelligent than us … I think they soon may be”. He also voiced concern about a “worst case scenario” in which bad actors misuse AI. At the same time, he said he did not want to criticize Google, which he said had acted in a “very responsible” way.

A one-sentence statement, “risk of extinction from AI should be a global priority … pandemics and nuclear war”, is signed by researchers and AI company leaders

The U.S. organization Center for AI Safety released a statement of just one sentence. Its content: the “risk of extinction from AI should be a global priority … pandemics and nuclear war” (that is, alongside other society-wide risks such as pandemics and nuclear war). Besides Geoffrey Hinton and Yoshua Bengio, representatives of Google DeepMind, OpenAI and Anthropic also signed. On the other hand, Yann LeCun (Meta), who also won the Turing Award, said such warnings were exaggerated. Other experts also said that the risk of human extinction is not realistic and distracts from problems that already exist, such as bias.

What we still don’t know:Even among the leading figures in the same field (the three 2018 Turing Award winners), views differ on how large the danger is. Which side is right is still unknown.

In a government survey, the share of people in Japan who have used generative AI goes 9.1% (fiscal 2023) → 26.7% → 58.8% (fiscal 2025). Lower than the U.S., Germany and China, but the biggest one-year rise

Every year, Japan's Ministry of Internal Affairs and Communications asks people in Japan, the U.S., Germany and China (and the UK too in fiscal 2023) whether they have ever used generative AI (AI that creates text and pictures), and publishes the results in the White Paper on Information and Communications. Japan's usage rate rose to 9.1% in fiscal 2023, 26.7% in fiscal 2024 and 58.8% in fiscal 2025. Over the same three years, the U.S. went 46.3% → 68.8% → 75.6%, Germany 34.6% → 59.2% → 75.6%, and China 56.3% → 81.2% → 93.6%. The 2026 white paper pointed out that Japan had the largest one-year rise. However, because people aged 15–19 were added to the survey from fiscal 2025, the conditions are not completely the same as in earlier years.

At the G7 Hiroshima Summit chaired by Japan, the “Hiroshima AI Process” is launched to discuss rules for generative AI

In May 2023, at the G7 summit of leaders of seven major countries held in Hiroshima, the leaders launched the “Hiroshima AI Process” and told their ministers to discuss generative AI by the end of the year. Example topics included protecting intellectual property (such as copyright), transparency, and countering foreign disinformation. According to the Japanese Ministry of Internal Affairs and Communications (MIC) page, in December a “Comprehensive Policy Framework”, which compiled guidelines and a code of conduct for organizations developing advanced AI, was agreed and endorsed by the G7 leaders. The circle of supporting countries and companies (the Friends Group framework) has kept growing through 2026.

Japan's Ministry of Education (MEXT) issues a generative AI guide for elementary, middle and high schools: “start with limited use”

In 2023, as ChatGPT spread, MEXT (Japan's Ministry of Education, Culture, Sports, Science and Technology) published the “Provisional Guidelines on the Use of Generative AI” (translated) for elementary, middle and high schools. It said that for now “it is appropriate to start with limited use” (translated), with trials at some schools that are ready. It says students should come to understand AI's strengths and weaknesses, and also that “AI has no self or personality” (translated), and that it is important not to leave everything to AI but to think for themselves. It warned that handing in text that AI wrote, as it is, for work such as book reports can count as cheating depending on the rules. In Ver. 2.0 in December 2024, a year and a half later, “provisional” was dropped from the title and “use” was changed to “effective use”.

A Japanese Agency for Cultural Affairs council compiles “General Understanding on AI and Copyright”: training done to deliberately make AI reproduce a particular artist's expression may fall outside the exception that needs no permission

A council of Japan's Agency for Cultural Affairs (a subcommittee that discusses copyright law) compiled its thinking on generative AI and copyright. Under Article 30-4 of the Copyright Act, as amended in 2018, AI training can in principle be done without permission. However, it concluded that this exception is considered not to apply where the purpose is mixed with enjoying the works, such as additional training aimed at “intentionally outputting all or part of the creative expression of works included in the training data” (translated) (training meant to deliberately bring out the expression of the works learned, in whole or in part). It also noted that a mere resemblance in “style” is not a copyright problem, but that care is needed when an AI is trained on only a few works by a specific creator and made to output that expression. The document says at its start that it is not legally binding.

Japan passes the “AI Act (AI Promotion Act)”, which pushes forward AI research, development and use, and creates an AI Strategy Headquarters led by the Prime Minister

Japan's Diet passed the “Act on the Promotion of Research and Development, and Utilization of Artificial Intelligence-Related Technologies” (nicknamed the AI Act or AI Promotion Act). The government's outline gives as reasons that “Japan’s AI development and use are lagging” (translated) and that “many citizens are anxious about AI” (translated), and sets the goal of “becoming the country where it is easiest to develop and use AI” (translated). It provides for setting up an “AI Strategy Headquarters” chaired by the Prime Minister with all ministers as members, deciding the national “AI Basic Plan”, having the government gather information, investigate cases of rights infringement, and give guidance and advice to businesses, and requiring businesses to cooperate with national measures. Unlike the EU's AI Act, which bans specific uses in detail, it was made as a law for “promotion” (pushing forward), as its title says.

Japan's government adopts its first “Artificial Intelligence Basic Plan”, and just 7 months later decides a second-term plan

Under the AI Act, in December 2025 the government adopted the “Artificial Intelligence Basic Plan: ‘Japan’s Revival’ through ‘Trustworthy AI’”. The plan itself writes that AI is not actively used in Japan's daily life and work, and that in development and investment “the lag is becoming more evident year by year” (translated), and it aims to make Japan “the country in the world where it is easiest to develop and use AI” (translated). As new technologies of 2025 it names “AI agents”, which carry out work on their own, and “physical AI”, which moves robots and the like. In July 2026 it drew up a second-term plan titled “Japan AX: Stronger and Richer” (AX stands for AI transformation, change brought about by AI), which also gives estimates of national and corporate investment through fiscal 2040.

What happened next:The check against reality comes in fiscal 2040. As of September 2026, it is not yet known.

The Japanese government launches the “AI Strategy Council”; the first meeting is held at the Prime Minister's Office, chaired by Yutaka Matsuo of the University of Tokyo

As generative AI spread, the Japanese government set up the “AI Strategy Council”, where experts discuss policy. The first meeting was held on May 11, 2023, at the Prime Minister's Office, with Prime Minister Fumio Kishida and cabinet ministers attending. The chair was Yutaka Matsuo of the University of Tokyo, and 7 other members took part. The minutes include the opinion that “a simple either-or debate [on whether to regulate AI by law (hard law)] is becoming somewhat unproductive” (translated). They also record opinions such as “we consider responsible AI development, ethics and social acceptance to be areas for cooperation” (translated) (a cooperation area is a field where parties work together instead of competing). The council continued until its 14th meeting in June 2025.

The Liberal Democratic Party team drafts an “AI White Paper”, saying “A new era of artificial intelligence has begun” and calling for a new national AI strategy

The Liberal Democratic Party's “Project Team on the Evolution and Implementation of AI”, set up in February 2023 (chair: Masaaki Taira), heard from experts seven times and compiled its proposals. It opens with “A new era of artificial intelligence has begun” (translated), citing the impact of ChatGPT as the reason. It quotes experts, such as “It is highly likely to affect almost all white-collar work” (translated) (Yutaka Matsuo) and “It is comparable to the invention of the internal combustion engine, semiconductors, and the internet” (translated) (Hiroaki Kitano). It then calls for deciding on a control tower for AI policy and quickly drawing up a new basic national strategy, and for pushing the discussion forward in forums such as the G7, which Japan was chairing.

Japan gets an institute to study AI safety (the AI Safety Institute), and also publishes the “AI Guidelines for Business” for companies

Prime Minister Fumio Kishida announced its creation at the AI Strategy Council in December 2023, and on February 14, 2024, the “AI Safety Institute (AISI)”, an organization to study methods and standards for evaluating the safety of AI, was launched with its secretariat at the Information-technology Promotion Agency (IPA). According to its official page, this followed a worldwide trend in which institutes with the same name were created in the UK and the U.S. In April, the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry published version 1.0 of the “AI Guidelines for Business,” which sums up what businesses that develop, provide or use AI should ideally observe. The guidelines point out that conversational generative AI is advancing the “democratization of AI” (translated), while new risks such as infringement of intellectual property and disinformation are arising.

At the first “AI Safety Summit” at Bletchley Park in the UK, 28 countries including China and Japan, and the EU, declare that frontier AI's “those capabilities are not fully understood and are therefore hard to predict”

At Bletchley Park in the UK, known as the place where codes were broken during World War II, the first international meeting to discuss the safety of frontier AI was held. 28 countries, including the United States, China and Japan, and the EU agreed to “The Bletchley Declaration”. The declaration pointed out that part of the danger of frontier AI comes from the fact that “those capabilities are not fully understood and are therefore hard to predict,” and it was especially worried about cyberattacks, biotechnology and disinformation. It also decided to meet again in 2024.

What we still don’t know:A joint declaration among governments stated clearly that, for frontier AI, “those capabilities are not fully understood and are therefore hard to predict” (November 2023).

The EU enacts the AI Act (which the EU describes as “the first-ever comprehensive legal framework on AI worldwide”); in 2026 it postpones the start of key parts

The European Union (EU) enacted a law, the “Artificial Intelligence Act” (AI Act), that sorts AI by level of risk and regulates it. Full application was set for August 2, 2026, in stages: prohibited uses and similar rules from February 2, 2025, and rules for large general-purpose AI and similar from August 2, 2025. But in July 2026 the EU put forward an amendment (the Digital Omnibus on AI) “as regards the simplification of the implementation of harmonised rules on artificial intelligence”. It postponed the start of regulation for high-risk AI used in hiring, education and similar areas to December 2, 2027, and for AI covered by product-safety regulation to August 2, 2028.

China puts interim rules for generative AI services into force, and in 2025 also makes labeling of AI-generated content mandatory

In July 2023, seven Chinese ministries, including the government body in charge of the internet (the Cyberspace Administration of China), promulgated the “Interim Measures for the Administration of Generative Artificial Intelligence Services” (interim rules on managing services that create text and pictures with AI), and they took effect on August 15. The rules cover services that generate text, images, audio and video for the general public inside China. Research and development that is not offered to the public was excluded. The basic policy is “equal emphasis on development and security” (translated); the rules require generated content to uphold the core socialist values, to prevent discrimination by ethnicity, gender, age and so on, and to improve the accuracy and reliability of generated content. They also set out that training must use data from lawful sources, and that services with influence on public opinion must undergo security assessments. In March 2025, four ministries announced a rule requiring text, images and other AI-generated content to carry a visible mark and a hidden mark inside the file; it took effect on September 1 (a mandatory national standard took effect at the same time).

U.S. AI policy turns sharply in 2023–2026 (changes in presidential executive orders)

The United States has no nationwide AI law and has set its policy through executive orders (orders the president gives to government agencies). Executive Order 14110 of October 2023 was titled “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” and required reports to the government for AI systems and others trained with more than 10^26 operations of computation. In January 2025 the next president revoked that order (14148) and ordered a review of obstructive regulations in order to “sustain and enhance America’s global AI dominance” (14179). In July he released “America’s AI Action Plan”. Executive Order 14365 of December called for a single federal framework, to avoid state-by-state AI laws becoming “a patchwork of 50 different regulatory regimes”, and had a body set up under the Attorney General to challenge state laws in court. Executive Order 14409 of June 2026 set up a mechanism that defines powerful AI that could be misused for cyberattacks as a “covered frontier model”, and a voluntary framework in which developers let the government try the model for up to 30 days before release. It also states clearly that no national licensing system will be created.

What we still don’t know:Even “America's AI Action Plan” of 2025, which pushes deregulation, points out that “the inner workings of frontier AI systems are poorly understood” and that they often cannot explain why they gave an answer, and says it will focus on research into this (PDF page 12).

Japan passes the “AI Act (AI Promotion Act)”, which pushes forward AI research, development and use, and creates an AI Strategy Headquarters led by the Prime Minister

Japan's Diet passed the “Act on the Promotion of Research and Development, and Utilization of Artificial Intelligence-Related Technologies” (nicknamed the AI Act or AI Promotion Act). The government's outline gives as reasons that “Japan’s AI development and use are lagging” (translated) and that “many citizens are anxious about AI” (translated), and sets the goal of “becoming the country where it is easiest to develop and use AI” (translated). It provides for setting up an “AI Strategy Headquarters” chaired by the Prime Minister with all ministers as members, deciding the national “AI Basic Plan”, having the government gather information, investigate cases of rights infringement, and give guidance and advice to businesses, and requiring businesses to cooperate with national measures. Unlike the EU's AI Act, which bans specific uses in detail, it was made as a law for “promotion” (pushing forward), as its title says.

National AI summits continue: Seoul (2024), Paris (2025), New Delhi (2026) — the U.S. and the UK are not on the Paris statement's list of signers

After the “AI Safety Summit” in the UK in November 2023 (record → At the first AI Safety Summit at Bletch…), meetings of national representatives continued every year. At the “AI Seoul Summit” in May 2024, leaders of 10 countries, including Japan, the U.S. and the UK, and the EU issued a declaration. It said “AI safety, innovation, and inclusivity are inter-related goals” (inclusivity means leaving no one behind). At the “AI Action Summit” in Paris in February 2025 (more than 100 countries took part), a statement on “Inclusive and Sustainable Artificial Intelligence for People and the Planet” was issued. It was signed by 63 countries, including China, Japan, India and South Korea, plus the African Union Commission and the EU (the U.S. and the UK are not on the list). The statement also states that this was the first time a summit discussed the issue of AI and energy. At the “AI Impact Summit” in New Delhi in February 2026, 92 countries and international organizations endorsed the “New Delhi Declaration on AI Impact”.

Three Japanese newspaper companies sue an AI search company in Tokyo District Court over “unauthorized use of articles”

In August 2025, the Yomiuri Shimbun Group sued the U.S. company behind the AI search service “Perplexity”, which answers questions with AI, in the Tokyo District Court. The Nikkei and the Asahi Shimbun then sued jointly. According to the Nikkei's announcement, both companies had placed “robots.txt” (a marker that tells machines not to use their articles) on their sites. The two companies claim that Perplexity ignored it, copied and stored the articles without permission, and had been using them in its answers since around June 2024 at the latest. They also claimed that it damaged their credibility by showing false content that differed from the articles while displaying the newspapers' names. Each of the two demanded 2.2 billion yen. This is a case of major Japanese newspapers suing a generative AI company after the 2018 amendment of the Copyright Act (under which AI training in principle needs no permission).

What we still don’t know:This site has not confirmed the outcome of the trial (as of September 2026). How Japanese courts will judge the difference between AI learning from articles and AI searching and using articles in its answers is still unknown.

A Japanese Agency for Cultural Affairs council compiles “General Understanding on AI and Copyright”: training done to deliberately make AI reproduce a particular artist's expression may fall outside the exception that needs no permission

A council of Japan's Agency for Cultural Affairs (a subcommittee that discusses copyright law) compiled its thinking on generative AI and copyright. Under Article 30-4 of the Copyright Act, as amended in 2018, AI training can in principle be done without permission. However, it concluded that this exception is considered not to apply where the purpose is mixed with enjoying the works, such as additional training aimed at “intentionally outputting all or part of the creative expression of works included in the training data” (translated) (training meant to deliberately bring out the expression of the works learned, in whole or in part). It also noted that a mere resemblance in “style” is not a copyright problem, but that care is needed when an AI is trained on only a few works by a specific creator and made to output that expression. The document says at its start that it is not legally binding.

China's DeepSeek releases its reasoning AI “DeepSeek-R1” in a form anyone can use; a week later, the share price of NVIDIA, the AI GPU company, falls 17% in one day

The Chinese company DeepSeek released “DeepSeek-R1”, a text AI that scored on a par with the U.S. company OpenAI's o1 on math and programming problems. It was distributed including its trained data (weights), under the MIT license, which allows commercial use and modification. A week later, on January 27, amid worries about tougher competition from DeepSeek and that investment in AI parts (GPUs) might level off, the share price of the U.S. company NVIDIA, which makes those parts, fell 17%. Its market value dropped by nearly $600 billion in one day, which was reported as the biggest one-day loss by a U.S. company at the time.

A giant U.S. AI computing center, “The Stargate Project”: announced as a plan that “intends to invest” $500 billion over 4 years

SoftBank Group, OpenAI and others announced “The Stargate Project,” a new company to build AI data centers in the U.S. for OpenAI. The announcement said it “intends to invest $500 billion over the next four years … $100 billion immediately”. The initial investors are SoftBank, OpenAI, Oracle and MGX, with SoftBank in charge of finances and OpenAI in charge of operations, and Masayoshi Son as chairman. In an announcement in September of the same year, it said planned power was close to 7 gigawatts and investment over the next three years was more than $400 billion. It then stated a “clear path to securing the full $500 billion, 10-gigawatt commitment … ahead of schedule” (the commitment announced in January, to be secured by the end of 2025).

A survey of 2,778 AI researchers in autumn 2023 shows forecasts for “human-level machine intelligence (HLMI)” moving 13 years earlier in one year (2060 → 2047)

In autumn 2023, the same research group (AI Impacts and others) surveyed 2,778 researchers who had published papers at leading AI conferences. The year in which “unaided machines outperforming humans in every possible task” becomes a 50-50 chance is 2047, which is 13 years earlier than in the same survey one year before (2022, when it was 2060). Over the six years from the 2016 survey to the 2022 survey, the figure had moved only one year, from 2061 to 2060. So, in a survey taken after ChatGPT, GPT-4 and others appeared, the experts' outlook shifted sharply in one year. On the other hand, when asked for the year when “all human occupations becoming fully automatable”, the answer was 2116 (2164 in the 2022 survey), so within the same survey the way of asking made a difference of nearly 70 years (the two questions were answered by different people, split by lottery: 1,714 people answered the tasks question and 774 the occupations question). The paper says, “We do not know what accounts for this gap in forecasts”.

What happened next:The years to check against reality are 2027 (the 10% year), 2047 and 2116. As of September 2026, it is not yet known.

What we still don’t know:The experts' own outlook moved 13 years in one year, and shifts by about 70 years depending on how the question is asked. When human-level AI will exist is still not known even to experts (on the difference caused by the wording, the paper itself writes “We do not know what accounts for this gap in forecasts”).

A 2016 survey of 352 AI researchers: “translating languages (by 2024), writing high-school essays (by 2026)”: those years have now come

Researchers at the University of Oxford and elsewhere asked researchers who had presented papers at two large machine learning conferences in 2015 in which year AI would reach or exceed human-level ability, and 352 answered (21% of those asked). Summing up the answers, the chance that “AI outperforming humans in all tasks” (human-level machine intelligence; called HLMI in the paper) exists reaches even odds in 45 years, and that “when all occupations are fully automatable” reaches even odds in 122 years. The answers differed greatly by region: 30 years for respondents in Asia and 74 years for those in North America. For individual tasks, translation (around 2024), a high-school history essay (around 2026), and driving a truck (around 2027) were among those given.

What happened next:Translation and essays have entered the predicted period. However, in the next survey by the same group in 2023 (→ A survey of 2,778 AI researchers in aut…), these two were still left among the items asking when they would be achieved. We have not found a judge's source that rules them achieved. Human-level AI (2061) is waiting for an answer.

Anthropic's CEO Amodei says, of powerful AI, “I think it could come as early as 2026”; the company's recommendation says “late 2026 or early 2027”

Dario Amodei, CEO of the AI company Anthropic, published the essay “Machines of Loving Grace”. He defined “powerful AI (I dislike the term AGI)” as AI that is, among other things, “smarter than a Nobel Prize winner across most relevant fields”. He then wrote, “I think it could come as early as 2026 … could take much longer”. He noted at the start that many of his guesses will be wrong in the details, but that he wants to give a useful estimate. In March of the next year, the company's recommendation submitted to the U.S. government stated, “we expect powerful AI systems will emerge in late 2026 or early 2027”.

What happened next:We are in the middle of the period when this prediction can be checked (it is now September 2026). This site has not found any source in which a third party judged that powerful AI has been completed.

Kurzweil's prediction that AI will reach human-level intelligence by 2029: in 2002 he bet $20,000 against Kapor, who took the opposite side, and he revisited it in a 2024 book

Inventor Ray Kurzweil, known for his predictions of the future, predicted in 1999 that AI would reach human-level intelligence by 2029 (according to the publisher's description). In 2002 he bet $20,000 with Mitch Kapor on the prediction (Kapor's side) that “By 2029 no computer … will have passed the Turing Test”. The rules were set in detail: three judges would each exchange text with four parties, a computer and three humans, for two hours each (24 hours in all), and the computer would pass if it fooled at least two of the three judges. In his 2024 book “The Singularity Is Nearer”, Kurzweil revisits this 2029 prediction.

What happened next:The result will be known in 2029. In 2025 there was a study in which GPT-4.5 was judged to be human 73% of the time in a test of 5-minute conversations. But that is a different test from the conditions of the bet (a test totaling 24 hours), and no ruling on the bet itself has been made (we found no record of a ruling on the bet).

Google DeepMind paper: “highly uncertain about the timelines … plausible that they will be developed by 2030”

Researchers at Google DeepMind wrote a paper on how to prepare for the dangers of powerful AI (AGI). As one of its assumptions, it says, “highly uncertain … but crucially, we find it plausible … developed by 2030”. Because the timeline may be very short, it aims for safety measures that work “anytime”, that is, in a form that can be put in place immediately when needed. The AI this paper assumes is one whose ability “matches or exceeds that of the 99th percentile of skilled adults … non-physical tasks”. Among the authors is Shane Legg, who in 2011 predicted that human-level AGI would arrive in 2028 on average, with 2025 the most likely year.

What happened next:The check against reality comes in 2030. As of September 2026, it is not yet known.

What we still don’t know:The researchers of the company building it write, “We are highly uncertain about the timelines”.

Prediction: AGI (human-level general-purpose AI) is “In the next five to ten years, I think.” (2025, Google DeepMind's Hassabis, in a TV interview)

In April 2025, on the U.S. CBS program “60 Minutes”, Demis Hassabis, head of Google DeepMind (2024 Nobel Prize in Chemistry), answered an interviewer. When asked “On track for AGI in--” (AGI is AI that can do any intellectual task like a human), Hassabis said, “In the next five to ten years, I think.” To the question “And in 2030 you will have what?” he answered with “everything around you ... nuanced and deep ways ... kind of embedded in your everyday life” (a system that understands everything around you in fine detail and deeply, and is more or less built into daily life). In the same interview, he pointed out that today's AI cannot yet come up with new questions or hypotheses on its own, has no curiosity, and probably still lacks a little imagination and intuition.

What happened next:The years the prediction refers to have not yet arrived (as of September 2026).

What we still don’t know:People and companies differ on what must be achieved to call something “AGI”, so the yardstick for judging whether it has been achieved is itself not settled (this site's estimate; the transcript introduces it as “with the versatility of a human mind”).

SoftBank Group's Masayoshi Son predicts that within the next 10 years a turning point will come when humanity is overtaken by artificial intelligence, and says he aims to realize ASI, “something 10,000 times beyond humanity’s wisdom” (translated)

At SoftBank Group's shareholders' meeting, Chairman and CEO Masayoshi Son said the group's mission is “the evolution of humanity” (translated), and set out “the realization of ASI (artificial superintelligence)” (translated) as the means. According to the company's article, AGI (general-purpose AI) is said to be something “said to be 10 times beyond humanity’s wisdom” (translated), and Son described ASI as “something 10,000 times beyond humanity's wisdom” (translated). Son said, “We seriously aim to realize ASI” (translated), and the article sums up his forecast as “within the next 10 years, a turning point will come when humanity is overtaken by artificial intelligence” (translated), changing all common sense. He also said that if ASI is achieved it may save humanity from cancer, stroke, accidents, disasters, pandemics and more. At the next year's (2025) shareholders' meeting he said, “In the age of ASI, SoftBank Group will become the world’s number one platform provider” (translated).

What happened next:The check against reality comes around 2034. As of September 2026, it is not yet known.

OpenAI's CEO Altman: “It is possible that we will have superintelligence in a few thousand days (!)” and “in 2025, we may see the first AI agents ‘join the workforce’”

OpenAI's CEO Sam Altman wrote in his essay “The Intelligence Age” on his own website, “superintelligence in a few thousand days (!); it may take longer … we’ll get there” (superintelligence means AI far smarter than humans). As the reason it got this far, he explained that “deep learning worked, got predictably better with scale, and we dedicated increasing resources”. In the January essay of the next year, “Reflections”, he wrote that “confident we know how to build AGI as we have traditionally understood it”, and that “first AI agents ‘join the workforce’ and materially change the output of companies” in 2025.

What happened next:(1) is waiting for a result. (2) The year 2025 has passed. This site has not found any source that judged whether it greatly changed companies' output.

In 2011, Shane Legg, who later became a Google DeepMind researcher, predicts “human level AGI … a mean of 2028 and a mode of 2025”

The AI researcher Shane Legg published on New Year's Eve 2011 the prediction he posted every year on his blog. He expressed the year in which human-level artificial general intelligence (AGI) will be built as a probability distribution, with a mean of 2028 and a most likely year of 2025 (assuming no major disaster such as nuclear war). He also wrote, “I expect to see an impressive proto-AGI within the next 8 years”. He explains a proto-AGI as a system that has acquired the basic abilities of vision, hearing, bodily control and language by learning rather than having them built in beforehand. In the comments, he himself wrote, “One complication with this prediction is that it might be impossible to falsify” (he goes on to say, for example, that if a large military built such a system it would have a strong reason to keep it secret), and “this might be my last annual prediction … kind of a lose-lose proposition” (a situation in which he loses whether the prediction comes true or not).

What happened next:The most likely year, 2025, has passed. This site has found no source in which a third party judged that human-level AGI had been achieved. The mean, 2028, is waiting to be checked. For the “proto-AGI” (translated) by 2019, he himself gave a standard for judging (something that a knowledgeable person, after a 1–2 hour briefing and some hands-on use, would think is within 10 years of human level), but this site has not found material judging whether anything met it.

US government report “Preparing for the Future of Artificial Intelligence”: for machines to show broad intelligence is “very unlikely ... comparable to or exceeding that of humans in the next 20 years”

In October 2016, the National Science and Technology Council (NSTC) of the US Executive Office of the President released the report “Preparing for the Future of Artificial Intelligence.” In 2016 it held five public workshops, and 161 comments came in from a public request for information. The report sets out 23 recommendations, and a national AI research and development strategic plan was released at the same time. The cover letter says that machines will increasingly exceed humans at particular tasks, but that for them to show broad intelligence is “very unlikely ... comparable to or exceeding that of humans in the next 20 years”. The main text also says that private-sector experts broadly agree that General AI (AI that can do any intellectual task like a human) “General AI will not be achieved for at least decades”, and that the committee agrees. It also says that concern about far-future superintelligence should have little impact on current policy.

What happened next:The target year (2036) has not yet arrived (as of September 2026). Since then, in 2023–2025, forecasts of around 2026–2030 have appeared from company representatives and company papers (records → OpenAI's CEO Altman: It is possible tha…, → Anthropic's CEO Amodei says, of powerfu… and → Google DeepMind paper: highly uncertain…). The results of expert surveys vary greatly depending on how the question is asked (record → A survey of 2,778 AI researchers in aut…).

What we still don’t know:The report itself notes that society would benefit from “accurate and timely methods for monitoring and forecasting AI developments”, and a note admits that this may be difficult. It then recommends continuing to survey experts' views regularly (pdf page 33).

Japan's government adopts its first “Artificial Intelligence Basic Plan”, and just 7 months later decides a second-term plan

Under the AI Act, in December 2025 the government adopted the “Artificial Intelligence Basic Plan: ‘Japan’s Revival’ through ‘Trustworthy AI’”. The plan itself writes that AI is not actively used in Japan's daily life and work, and that in development and investment “the lag is becoming more evident year by year” (translated), and it aims to make Japan “the country in the world where it is easiest to develop and use AI” (translated). As new technologies of 2025 it names “AI agents”, which carry out work on their own, and “physical AI”, which moves robots and the like. In July 2026 it drew up a second-term plan titled “Japan AX: Stronger and Richer” (AX stands for AI transformation, change brought about by AI), which also gives estimates of national and corporate investment through fiscal 2040.

What happened next:The check against reality comes in fiscal 2040. As of September 2026, it is not yet known.

2005, Kurzweil: “I set the date for the Singularity … as 2045”. By his own account, machine intelligence matching humans comes “Within a quarter century” (by 2030), and 2045 is the year of “one billion times more powerful than all human intelligence today”

Ray Kurzweil (introduced in the 2014 Journal of the Japanese Society for Artificial Intelligence as an “American futurist” (translated)), in a Q&A in which he answered questions about his 2005 book “The Singularity Is Near” (Japanese edition: “The Birth of the Post-Human”), wrote that within a quarter century the intelligence of machines would match the breadth and subtlety of human intelligence, and then go on to surpass it. He went on to call the point at which machine intelligence redesigns itself, grows faster and smarter at an ever greater pace, and technology advances too fast for unmodified human intelligence to keep up, “the Singularity”, and wrote “I set the date for the Singularity … as 2045”. The machine intelligence created in that year, he says, would be “one billion times more powerful than all human intelligence today”. As for the amount of computation needed, he also wrote that the work of a whole human brain needs about 10 quadrillion (10^16) calculations per second, that supercomputers will reach this around the end of the 2000s (two Japanese projects are aiming for it), and that by 2020 it will be available for about $1,000.

What happened next:The check on 2045 is still in the future (as of September 2026). Among the milestones along the way, on the forecast that supercomputers would reach 10^16 calculations per second (10 quadrillion), Japan's RIKEN “K computer” scored 10.51 petaflops in the TOP500 list of November 2011. It became the first supercomputer to reach “10 Petaflop/s, or 10 quadrillion calculations per second” (a little later than the forecast of “around the end of this decade”). For the forecast that by 2020 about $1,000 would buy 10^16 calculations per second, this site has not opened any material to check it.

What we still don’t know:The amount of computation needed for the work of a whole human brain differs by a factor of 100 depending on who estimates it (Kurzweil 2005: 10^16 calculations per second; Moravec 1998: 100 million MIPS = 10^14 instructions per second). How much computation the brain actually does is still not known.

Kurzweil's prediction that AI will reach human-level intelligence by 2029: in 2002 he bet $20,000 against Kapor, who took the opposite side, and he revisited it in a 2024 book

Inventor Ray Kurzweil, known for his predictions of the future, predicted in 1999 that AI would reach human-level intelligence by 2029 (according to the publisher's description). In 2002 he bet $20,000 with Mitch Kapor on the prediction (Kapor's side) that “By 2029 no computer … will have passed the Turing Test”. The rules were set in detail: three judges would each exchange text with four parties, a computer and three humans, for two hours each (24 hours in all), and the computer would pass if it fooled at least two of the three judges. In his 2024 book “The Singularity Is Nearer”, Kurzweil revisits this 2029 prediction.

What happened next:The result will be known in 2029. In 2025 there was a study in which GPT-4.5 was judged to be human 73% of the time in a test of 5-minute conversations. But that is a different test from the conditions of the bet (a test totaling 24 hours), and no ruling on the bet itself has been made (we found no record of a ruling on the bet).

Turing’s forecast: “in about fifty years' time ... 70 per cent chance ... after five minutes”

In his 1950 paper, Turing set down his own forecast. In about 50 years, he wrote, computers could be made to play the imitation game well, and an average interrogator would have no more than a 70% chance of identifying correctly after 5 minutes of questioning. In the same paragraph he also wrote that by the end of the century the use of words and people’s ideas would have changed, so that one could “speak of machines thinking” without being contradicted.

What happened next:According to a survey paper from 2000 (Saygin et al.), no program had yet won the grand prize in the contest with prize money that had been held since 1991 (the Loebner Prize), and the records show that the machines were usually spotted quickly. In other words, in the year of the forecast, it had not come true in the form predicted. Some said the results of the 2000 contest “shows Turing’s prediction to have been in error”, but in the same year, 2000, Copeland wrote that “‘About fifty years’ does not mean ‘exactly fifty years’”. Then in March 2025, researchers at a US university (UC San Diego) published in a paper (arXiv) the results of an experiment in which an interrogator chatted for 5 minutes with a human and a machine at the same time. Of two large language AIs told to play the part of a human, one (GPT-4.5) was judged to be the human 73% of the time and the other (LLaMa-3.1-405B) 56% of the time (since there were two choices, the interrogator correctly spotted the first one only in the remaining 27%). But this is the announcement of one research group, and we have not confirmed later evaluations on this site. The part about being able to “speak of machines thinking without expecting to be contradicted” has not been confirmed.

What we still don’t know:Whether a machine being judged human in the 2025 experiment means that it can “think” is not known (Turing himself, regarding the original question as not worth discussing, replaced it with the game).

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%).

Ford announces a plan to run large numbers of fully self-driving cars with no steering wheel or pedals in ride-hailing services in 2021 (2016)

In August 2016, the automaker Ford announced (in Palo Alto) a plan to run large numbers of “Level 4” vehicles (fully self-driving in limited areas) in ride-hailing and ride-sharing services in 2021. The plan was for the cars to have no steering wheel, no accelerator and no brake. Its president said “The next decade will be defined by automation of the automobile”. Around the same time, Japan's national plan and carmakers also announced self-driving plans targeting around 2020 (→ A national plan: self-driving on expres…).

What happened next:A service of this kind did not begin in 2021. On October 26, 2022, Ford announced that it had judged that putting Level 4 to profitable use at scale in the industry would come later than first expected. It booked a loss of 2.7 billion dollars on its investment in the self-driving company Argo AI, and shifted its development funds to Level 2+ and Level 3 (driver-assistance technologies in which a person watches over the driving).

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.

Researchers who peer inside AI sum up for themselves that “there are many open problems in the field that require solutions” (29 people from universities, institutes and companies)

29 researchers who study how each part inside a large AI computes (understanding its mechanisms, called mechanistic interpretability) compiled the “open problems” of the field. The authors' affiliations cut across companies and universities, including Apollo Research, Anthropic, Google DeepMind, MIT, Harvard University, Tel Aviv University and the University of Melbourne. While acknowledging recent progress, the paper wrote that “many open problems … require solutions before many scientific and practical benefits can be realized”. The points it lists include the following. What “features”, the main clue used for investigation, actually is remains unclear. At present AI's abilities are mainly checked with “black box” tests that look at the answers from outside. And it is not known whether methods that worked on small experimental AI can be extended to the largest frontier AI.

What we still don’t know:What we still don't know (points made by experts who study the inner structure themselves): The methods for reading what is computed inside large AI are still unclear even at the basic idea (what a “feature” is). Today's AI abilities are mainly checked by testing from the outside.

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.

The UK's central bank warns: “On a number of measures, equity market valuations appear stretched, particularly for technology companies” focused on AI

The committee that watches financial stability at the UK's central bank (the Bank of England) wrote in its October 2025 record, “On a number of measures, equity market valuations appear stretched, particularly for technology companies” that are focused on AI. Because stock indexes are concentrated in a small number of companies, it said markets could be hit hard “should expectations around the impact of AI become less optimistic”. As examples of what could trigger a fall, it lists AI performance or adoption disappointing, and shortages of power, data and materials. It also gave as an example a major breakthrough in ideas that makes large AI facilities unnecessary. In its July 2026 report it also wrote that “Equity prices have increased especially for AI-related stocks”, that “on some metrics, valuations have also become more stretched”, and that “The risk of a sharp correction in equity markets remains high”.

What we still don’t know:Whether a winter will come after the current boom is not known. As of July 2026, the central bank was writing that “The risk of a sharp correction in equity markets remains high.” Meanwhile, share prices kept rising. The answer has not yet come.

2025: in the same research firm's view for Japan, generative AI is at the “Peak of Inflated Expectations” (on the boundary with the trough of disillusionment in the figure)

The main text of Gartner Japan's “Hype Cycle for Future-Oriented Infrastructure Technologies in Japan, 2025” wrote that AI agents and generative AI are “Peak of Inflated Expectations” (translated). Meanwhile, the generative AI dot in the figure sits on the line between the peak and the trough of disillusionment, and the headline of one news article reported “Generative AI is in the trough of disillusionment” (translated).

What we still don’t know:Where the fourth boom is on the curve right now differs depending on who looks, and is not known.