Artificial intelligence boom (2020s)

Things happened, and the rest is history

1. 01

   The Dartmouth workshop names the field

   Event

   What happened

   In 1955 John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon propose a summer study at Dartmouth College on making machines simulate aspects of intelligence. The proposal uses the term artificial intelligence. The workshop takes place in the summer of 1956.

   ↓ The rest is history

   The meeting is widely seen as the founding moment of AI as a research field. Its early optimism about fast progress was not met, and funding fell sharply in the 1970s and again in the late 1980s, periods often called AI winters. The field survived on narrower, practical methods. The 1955 proposal is still preserved and widely cited.

   Sources

   1. PrimaryMcCarthy, J., M. L. Minsky, N. Rochester, and C. E. Shannon. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. Proposal, 31 August 1955.
   2. BookRussell, Stuart, and Peter Norvig. Artificial Intelligence: A Modern Approach. Hoboken: Pearson, 2020.Standard textbook with a history of the field.
2. 02

   Deep Blue defeats Kasparov

   Event

   What happened

   In May 1997 IBM's chess computer Deep Blue beats world champion Garry Kasparov in a six-game match in New York, winning 3.5 to 2.5. It is the first loss by a reigning world champion to a computer in a standard match.

   ↓ The rest is history

   Deep Blue relied on specialised hardware and brute-force search rather than learning, so it was not the kind of AI behind the later boom. It did show that machines could beat top humans at a task long seen as a mark of intelligence. Later breakthroughs in games came from learning systems instead. The match ended on 11 May 1997.

   Sources

   1. BookCampbell, Murray, A. Joseph Hoane Jr., and Feng-hsiung Hsu. "Deep Blue." Artificial Intelligence 134, no. 1-2 (2002): 57-83.The designers' own account.
   2. BookHsu, Feng-hsiung. Behind Deep Blue: Building the Computer That Defeated the World Chess Champion. Princeton: Princeton University Press, 2002.
3. 03

   ImageNet and the deep learning breakthrough

   Event

   What happened

   In 2009 researchers led by Fei-Fei Li publish ImageNet, a database of millions of labelled images. In September 2012 a neural network called AlexNet, built by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton and trained on graphics processors, wins the ImageNet recognition contest by a wide margin.

   ↓ The rest is history

   The result convinced many researchers that deep neural networks, large datasets and GPU computing worked well together. Companies began hiring researchers and buying hardware on a large scale. Hinton, Yoshua Bengio and Yann LeCun received the 2018 Turing Award for their work on deep learning. AlexNet's paper was presented at the NIPS conference in December 2012.

   Sources

   1. ArticleKrizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "ImageNet Classification with Deep Convolutional Neural Networks." Advances in Neural Information Processing Systems 25 (2012): 1097-1105.
   2. ArticleLeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. "Deep Learning." Nature 521 (2015): 436-444.
4. 04

   AlphaGo defeats Lee Sedol

   Event

   What happened

   In March 2016 DeepMind's program AlphaGo beats Lee Sedol, one of the world's best players of Go, by four games to one in Seoul. AlphaGo combines deep neural networks with tree search and was trained partly on human games and partly by playing itself.

   ↓ The rest is history

   Go has far more possible positions than chess, and experts had expected computers to need many more years. The match drew huge public attention, especially in East Asia, and showed the strength of learning-based methods. DeepMind later built systems that learned without human game data. The match took place from 9 to 15 March 2016.

   Sources

   1. ArticleSilver, David, et al. "Mastering the Game of Go with Deep Neural Networks and Tree Search." Nature 529 (2016): 484-489.
5. 05

   The transformer architecture

   Event

   What happened

   In 2017 a team of researchers at Google publishes a paper titled Attention Is All You Need. It describes the transformer, a neural network design that processes whole sequences of text at once using a mechanism called attention.

   ↓ The rest is history

   Transformers trained faster on parallel hardware than earlier designs and scaled well with more data and computing power. They became the basis for large language models such as BERT (2018) and the GPT series from OpenAI, and later for image and audio systems. The paper was posted in June 2017 and presented at NIPS that December.

   Sources

   1. ArticleVaswani, Ashish, et al. "Attention Is All You Need." Advances in Neural Information Processing Systems 30 (2017): 5998-6008.
6. 06

   Scaling up language models: GPT-3

   Event

   What happened

   In May 2020 OpenAI publishes a paper describing GPT-3, a language model with 175 billion parameters trained on very large amounts of text. It can write text, answer questions and follow examples given in a prompt without being retrained for each task.

   ↓ The rest is history

   The paper supported the view that larger models trained on more data and computing power keep improving, a trend described in earlier scaling studies. Access to GPT-3 was first limited to selected users through a paid interface. It raised concerns about misuse, bias and the energy cost of training. It led directly to later, larger systems.

   Sources

   1. ArticleBrown, Tom B., et al. "Language Models Are Few-Shot Learners." Advances in Neural Information Processing Systems 33 (2020): 1877-1901.
7. 07

   ChatGPT reaches the public

   Event

   What happened

   On 30 November 2022 OpenAI releases ChatGPT, a free chatbot built on its GPT-3.5 models and tuned with human feedback to follow instructions. Within two months it is reported to have reached an estimated 100 million users. Image generators such as DALL-E 2, Midjourney and Stable Diffusion had appeared earlier that year.

   ↓ The rest is history

   The release made generative AI familiar to ordinary people and set off a race among technology firms to launch rival products and invest in chips and data centres. Schools, newsrooms and offices began setting rules for its use. Lawsuits over training data and copyright followed. GPT-4 arrived in March 2023.

   Sources

   1. PrimaryOpenAI. Introducing ChatGPT. Company announcement, 30 November 2022.
   2. BookBommasani, Rishi, et al. On the Opportunities and Risks of Foundation Models. Stanford: Stanford Center for Research on Foundation Models, 2021.Report that named and surveyed large general-purpose models.
8. 08

   Governments begin to regulate AI

   Event

   What happened

   In November 2023 the United Kingdom hosts the AI Safety Summit at Bletchley Park, and 28 countries and the European Union sign the Bletchley Declaration. In March 2024 the European Parliament approves the EU Artificial Intelligence Act, which sets rules by level of risk. The Act enters into force in August 2024.

   ↓ The rest is history

   These were among the first broad efforts to govern AI. Approaches differ, with the EU using binding law, the United States relying more on executive orders and voluntary commitments, and China issuing rules on areas such as generative AI and recommendation systems. Debate continues over how to balance safety, innovation and competition.

   Sources

   1. PrimaryThe Bletchley Declaration by Countries Attending the AI Safety Summit. Policy declaration, 1 November 2023.
   2. PrimaryRegulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence. Official Journal of the European Union, 2024.
9. 09

   The Turing test as the measure of machine intelligence

   Popular belief

   The legend

   Many people believe that a machine that passes the Turing test, by fooling a human into thinking it is a person, is truly intelligent, and that this was the agreed goal of AI. Chatbots are often said to have passed it.

   ↓ The rest is history

   Alan Turing's 1950 paper proposed an imitation game as a way to sidestep the question of whether machines can think. He did not present it as a formal test or a requirement. Some chatbots have fooled judges in short trials, such as Eugene Goostman in 2014, but those claims were disputed and the test is not a standard benchmark. Researchers now use many specific tests, and what such results show is still debated.

   Sources

   1. ArticleTuring, A. M. "Computing Machinery and Intelligence." Mind 59, no. 236 (1950): 433-460.
10. 10

    The Mechanical Turk, a thinking machine

    Folk tale

    The legend

    In 1770 Wolfgang von Kempelen shows the Austrian court a chess-playing automaton, the Turk. It is said to play and beat skilled opponents, including, in later tales, Napoleon and Benjamin Franklin, by machine thinking alone.

    ↓ The rest is history

    The Turk was a trick. A human chess player hid inside the cabinet and moved the pieces through a mechanism. It toured Europe and America for decades and was exposed in print in the 1820s and 1830s, with Edgar Allan Poe among those who wrote about it. It was destroyed in a fire in Philadelphia in 1854. Amazon later named its crowdwork service Mechanical Turk after it, and the boom's reliance on hidden human labelling echoes the story.

    Sources

    1. BookStandage, Tom. The Turk: The Life and Times of the Famous Eighteenth-Century Chess-Playing Machine. New York: Walker, 2002.
11. 11

    The machines will wake up and take over

    Popular belief

    The legend

    A common story holds that AI is about to become conscious and turn on its makers, as in the film The Terminator (1984) or the play R.U.R. (1920), which gave the world the word robot.

    ↓ The rest is history

    Fears of artificial beings rebelling run through fiction from Frankenstein (1818) onward. Current systems are statistical models that produce outputs from patterns in data, and there is no scientific agreement that they are conscious. Researchers do disagree seriously about long-term risks. A 2023 statement signed by many AI scientists and executives called reducing the risk of extinction from AI a global priority, though others called that focus overstated and argued that present harms such as bias and misinformation matter more.

    Sources

    1. BookBostrom, Nick. Superintelligence: Paths, Dangers, Strategies. Oxford: Oxford University Press, 2014.
    2. PrimaryCenter for AI Safety. Statement on AI Risk. Open statement, 30 May 2023.
12. 12

    The first AI winter was the end of AI

    Popular belief

    The legend

    It is often said that AI research was abandoned after early failures, and that the current boom came out of nowhere.

    ↓ The rest is history

    Funding and interest fell in the mid-1970s after critical reports such as the Lighthill Report in Britain in 1973, and again around 1987 to 1993. Work did not stop. Neural network research continued with figures such as Hinton, Bengio and LeCun, and expert systems and statistical methods found commercial uses. The 1986 popularisation of backpropagation and LeCun's handwriting recognition work in the 1990s were steps toward today's systems.

    Sources

    1. BookNilsson, Nils J. The Quest for Artificial Intelligence: A History of Ideas and Achievements. Cambridge: Cambridge University Press, 2010.
    2. BookLighthill, James. Artificial Intelligence: A General Survey. London: Science Research Council, 1973.Report that cast doubt on AI research funding in Britain.