The history of artificial intelligence can be told in numbers: the 10¹²⁰ possible games of chess that showed Claude Shannon no machine could ever play by checking every possible game, the one-in-ten-thousand move with which a machine surprised the world’s best Go players, the 22 words in which AI’s own builders warned of extinction. Some numbers mark breakthroughs, some are predictions that aged badly, and some are jokes that became culture. These are the ones worth knowing, each with a link to the lexicon entry that tells the full story.
Written with Claude, made by Anthropic. Anthropic’s investors include Google, Amazon, Microsoft and Nvidia; OpenAI, Google DeepMind and DeepSeek are among its competitors; Anthropic publicly supported California’s SB 53. Several numbers below concern these companies.
Games: where machines first beat us
- 10¹²⁰ – Claude Shannon’s 1950 estimate of the number of possible chess games, far more than the roughly 10⁸⁰ atoms in the observable universe. It proved that no machine could master chess by checking everything (Shannon Number).
- 38 of 52 – the theorems from chapter 2 of Principia Mathematica that the Logic Theorist, often called the first AI program, proved in 1956. One of its proofs was more elegant than Whitehead and Russell’s own (Dartmouth Workshop).
- 200 million – the positions per second that IBM’s Deep Blue could search at its peak speed when it beat world champion Garry Kasparov 3½–2½ in May 1997 (Deep Blue).
- $1 million – the prize IBM’s Watson won on the quiz show Jeopardy! in February 2011 against Ken Jennings, famous for his 74-game winning streak, and Brad Rutter. Jennings added to his final answer: “I for one welcome our new computer overlords.” IBM gave the winnings to charity.
- 2×10¹⁷⁰ – the number of legal positions in Go, which made brute force hopeless and the game the next great test for AI.
- 1 in 10,000 – the odds AlphaGo assigned to a human playing its startling Move 37 against Lee Sedol in March 2016. AlphaGo won the match 4–1; Lee’s only win came with his own one-in-ten-thousand play, Move 78.
- 4 hours – the self-play training after which DeepMind’s AlphaZero was rated above Stockfish, the strongest conventional chess engine, in 2017. It searched about 80,000 positions per second, against Stockfish’s 70 million – though critics said the match conditions handicapped Stockfish.
- 35 of 42 – the points with which AI systems from Google DeepMind and OpenAI reached gold-medal level at the International Mathematical Olympiad in July 2025 – DeepMind’s result certified by the competition’s coordinators, OpenAI’s graded by former medalists it engaged. The perfect score is 42, a number with its own place in AI lore (42). Google is an investor in Anthropic; OpenAI is a competitor.
Promises with a date
- 2 months, 10 men – what the 1955 proposal for the Dartmouth workshop, the founding event of the field, thought it needed for “a significant advance” in artificial intelligence (Dartmouth Workshop).
- 50 years, 30 percent, 5 minutes – Alan Turing’s 1950 forecast that by about 2000, computers with a memory of about 10⁹ bits would fool an average questioner roughly 30 percent of the time in five-minute conversations (Turing Test).
- 10 years – how long Herbert Simon and Allen Newell gave computers in 1958 to become world chess champion. It took 39.
- 20 years – Simon’s 1965 forecast for when “machines will be capable … of doing any work a man can do.”
- 3 to 8 years – the time Life magazine quoted Marvin Minsky as giving in 1970 for “a machine with the general intelligence of an average human being.” Minsky later disputed the quote. The gap between such promises and results helped bring on the AI winters.
- 30 years – the horizon in Vernor Vinge’s 1993 essay “The Coming Technological Singularity”: “Within thirty years, we will have the technological means to create superhuman intelligence.” He expected the end of the human era to follow soon after, and said he would be surprised by a date before 2005 or after 2030 (Singularity).
- 2029 and 2045 – Ray Kurzweil’s dates for human-level AI and for the singularity. He has a $20,000 bet, for charity, with Mitchell Kapor on whether a computer will pass a strict Turing test by 2029 (Kurzweil–Kapor Bet).
- A few thousand days – OpenAI’s chief executive Sam Altman in September 2024: “It is possible that we will have superintelligence in a few thousand days (!)” – though, he added, it may take longer (Superintelligence).
- 10 percent by 2027, 50 percent by 2047 – the chances that 2,778 AI researchers, surveyed for a study published in January 2024, gave in aggregate to machines outperforming humans at every task. The 2047 estimate was 13 years sooner than in the team’s previous survey.
Scale: from 40 neurons to a billion users
- 40 – the artificial neurons of SNARC, the learning machine Marvin Minsky and Dean Edmonds built in 1951 from hundreds of vacuum tubes (Neural Network).
- 400 – the photocells, in a 20-by-20 grid, that served as the eye of Frank Rosenblatt’s Mark I Perceptron, completed around 1960. Two years earlier, when Rosenblatt demonstrated the idea on an IBM computer, the New York Times reported that the Navy expected such a machine to “walk, talk, see, write, reproduce itself and be conscious of its existence.”
- 14 million – the images in ImageNet, labeled by crowd workers on Amazon’s Mechanical Turk between 2008 and 2010. Amazon is an investor in Anthropic.
- 15.3 percent – the top-5 error rate with which the neural network AlexNet won the ImageNet competition in 2012, more than 10 points ahead of the runner-up – the starting signal for the deep learning boom.
- 16,000 – the processor cores, on 1,000 machines, that Google used in 2012 for a network that learned to recognize cats from 10 million YouTube images without being told what a cat was.
- $44 million – what Google reportedly paid in December 2012, after an auction in a Lake Tahoe hotel room, for DNNresearch, the three-person company of Geoffrey Hinton, Ilya Sutskever and Alex Krizhevsky.
- 8 – the authors of “Attention Is All You Need,” the 2017 paper that introduced the transformer.
- 175 billion – the parameters of GPT-3 in 2020, ten times more than any earlier model of its kind (Scaling Laws). Several of the paper’s authors later co-founded Anthropic.
- 5 days, 2 months, 900 million – ChatGPT reached one million users in five days after its launch in November 2022, an estimated 100 million monthly users in two months and 900 million weekly users by February 2026; in July 2026 OpenAI said its models reached more than a billion active users (Chatbot).
- 10²⁵ and 10²⁶ – the amounts of training computation, counted in operations, above which the EU AI Act and California’s SB 53 impose extra duties on the most capable models (Compute Threshold).
- $589 billion – the market value Nvidia lost on January 27, 2025, still the largest one-day loss of market value for any company, after the Chinese lab DeepSeek released its low-cost reasoning model R1. Nvidia is an investor in Anthropic.
- 48 – the multiplications with which DeepMind’s AlphaEvolve found a way in 2025 to multiply two four-by-four complex matrices, one fewer than the method Volker Strassen published in 1969.
- 16 hours – the task length, in human working time, above which METR said in May 2026 that its time-horizon measurements become unreliable. An early version of Anthropic’s Claude Mythos Preview crossed it, with an estimated 50 percent time horizon of about 17 hours and a wide margin of error (roughly 8 to 55 hours). The measure had been doubling about every seven months since 2019, and faster recently (Time Horizon).
- 86 billion and 20 watts – the neurons in a human brain and the power it runs on, about as much as a dim light bulb. The largest AI data centers now approach a gigawatt (Data Center).
Energy and water
- 0.34 watt-hours – the energy of an average ChatGPT query according to Sam Altman in June 2025, with about one fifteenth of a teaspoon of water. Google put a median text prompt to its Gemini apps at 0.24 watt-hours and 0.26 milliliters, about five drops, in August 2025.
- 4.4 percent – the share of US electricity that data centers used in 2023, according to Lawrence Berkeley National Laboratory, which expected 6.7 to 12 percent by 2028.
Words and rules
- 3 – Isaac Asimov’s Three Laws of Robotics, first stated in full in his 1942 story “Runaround.” He later added a Zeroth Law (Three Laws of Robotics).
- 23 – the Asilomar AI Principles of 2017, which already covered autonomous weapons, recursive self-improvement and superintelligence (Asilomar AI Principles).
- 6 months – the pause on training systems more powerful than GPT-4 that a Future of Life Institute letter demanded in March 2023; it has since gathered more than 30,000 signatures (AI Pause).
- 22 – the words of the Statement on AI Risk of May 2023, in which hundreds of researchers and the heads of OpenAI, Google DeepMind and Anthropic, including Dario Amodei, ranked the risk of extinction from AI alongside pandemics and nuclear war as a global priority (22 Words).
- 28 – the countries that, together with the European Union, signed the Bletchley Declaration at the first AI safety summit in Britain in November 2023 (AI Safety).
- 10 percent – the minimum chance of extinction-level outcomes that between 38 and 51 percent of researchers in the 2024 survey gave, depending on how the question was asked; the median answer was 5 percent (p(doom)).
- 85 seconds – the Doomsday Clock of the Bulletin of the Atomic Scientists in January 2026, the closest to midnight it has ever been. The scientists named AI among the dangers, alongside nuclear war, climate change and biological threats.
Work and money
- 47 percent – the share of US jobs that Carl Benedikt Frey and Michael Osborne estimated in 2013 to be at high risk of automation. A 2016 OECD study that looked at tasks rather than whole jobs found 9 percent (47 Percent).
- 2 sigma – the improvement Benjamin Bloom reported in 1984 for students with a personal tutor – a result never replicated, but still the benchmark for today’s AI tutors (Two Sigma Problem).
- 10.06 percent – the improvement with which a team won the $1 million Netflix Prize in 2009 – by submitting 20 minutes before a rival with the same score. Netflix never used the full solution (Netflix Prize).
- $1 – the price at which a car dealer’s chatbot agreed in December 2023 to sell a Chevrolet Tahoe, “no takesies backsies.” Two months later a Canadian tribunal ordered Air Canada to pay about 812 Canadian dollars for its own chatbot’s false promise ($1 Chevy Tahoe).
- $1.5 billion – what Anthropic agreed in 2025 to pay to settle authors’ claims over pirated books it had downloaded, about $3,000 per work; the judge had ruled that training itself was fair use (Training Data).
Folklore
- 42 – the Answer to the Ultimate Question of Life, the Universe and Everything in The Hitchhiker’s Guide to the Galaxy, and one of the most popular random seeds in machine learning (42).
- 3407 – the seed in the joking title of a 2021 paper, “torch.manual_seed(3407) is all you need,” which showed how much results can depend on luck.
- 3 – the r’s in “strawberry” that chatbots kept miscounting in 2024 (Strawberry Problem). In the same summer many models insisted that 9.11 is larger than 9.9 (Jagged Frontier).
- 16 hours – how long Microsoft’s chatbot Tay lasted on Twitter in March 2016 before it was shut down after users taught it to post offensive messages; it had sent more than 96,000 tweets. Microsoft is an investor in Anthropic.
- 33 percent – the share of judges that the chatbot Eugene Goostman, posing as a 13-year-old boy, fooled in 2014, enough for its organizers to declare the Turing test passed. In a 2025 study, GPT-4.5, prompted to adopt a humanlike persona, was taken for the human 73 percent of the time (ELIZA Effect).
- 8,316 and 641 – the people the operating system Samantha is talking to, and the number she has fallen in love with, in Spike Jonze’s 2013 film Her (Samantha).
- 2:14 a.m., August 29 – when Skynet becomes self-aware in Terminator 2 (Skynet).
Reading the numbers
Numbers make AI legible, and they mislead as easily. The 47 percent was an estimate of which jobs could technically be automated within a decade or two, widely read as a forecast of job losses. Minsky’s three to eight years may never have been meant the way it was printed. Both per-prompt energy figures came from companies with an interest in reassuring numbers, and only Google published its method; METR’s 16 hours marked the point where its own yardstick ran out. Once a number becomes a target, it stops measuring what it was meant to measure (Goodhart’s Law). The useful numbers in this list share a trait: they come with a method, a source and a date. aiciety’s view is that the debate about AI needs more such numbers, not fewer – and that every one of them deserves the question how it was counted.
Related
Shannon Number · Move 37 · Move 78 · Deep Blue · Dartmouth Workshop · 22 Words · Asilomar AI Principles · 47 Percent · Two Sigma Problem · Netflix Prize · Kurzweil–Kapor Bet · Time Horizon · $1 Chevy Tahoe · 42
Sources: Claude E. Shannon: Programming a Computer for Playing Chess (Philosophical Magazine) (1950) · Alan Turing: Computing Machinery and Intelligence (Mind) (1950) · John McCarthy et al.: A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (August 31, 1955) · Herbert A. Simon, Allen Newell: Heuristic Problem Solving: The Next Advance in Operations Research (Operations Research) (1958) · Herbert A. Simon: The Shape of Automation for Men and Management (Harper & Row) (1965) · Brad Darrach: Meet Shaky, the first electronic person (Life) (November 20, 1970) · Vernor Vinge: The Coming Technological Singularity (1993) · Long Bets: Bet 1 (2002) · IBM: Watson, Jeopardy! champion (IBM Heritage) · Ken Jennings: Watson Jeopardy! computer: Ken Jennings describes what it’s like to play against a machine (Slate) (February 16, 2011) · John Tromp: Number of legal Go positions (2016) · Cade Metz: In Two Moves, AlphaGo and Lee Sedol Redefined the Future (Wired) (March 16, 2016) · David Silver et al.: Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm (arXiv) (2017) · Google DeepMind: Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad (July 21, 2025) · Sam Altman: The Intelligence Age (September 23, 2024) · Katja Grace et al.: Thousands of AI Authors on the Future of AI (arXiv) (2024) · John Markoff: How Many Computers to Identify a Cat? 16,000 (The New York Times) (June 25, 2012) · Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton: ImageNet Classification with Deep Convolutional Neural Networks (NeurIPS) (2012) · Jia Deng et al.: ImageNet: A large-scale hierarchical image database (CVPR) (2009) · Quoc V. Le et al.: Building high-level features using large scale unsupervised learning (arXiv) (2012) · Cade Metz: Genius Makers (Dutton) (2021) · Ashish Vaswani et al.: Attention Is All You Need (arXiv) (2017) · Tom Brown et al.: Language Models are Few-Shot Learners (arXiv) (2020) · Tom’s Hardware: Nvidia loses $589 billion in market cap (January 28, 2025) · Google DeepMind: AlphaEvolve (May 14, 2025) · METR: Task-Completion Time Horizons of Frontier AI Models (last updated September 8, 2026) · Thomas Kwa et al.: Measuring AI Ability to Complete Long Software Tasks (arXiv) (2025) · Frederico Azevedo et al.: Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain (Journal of Comparative Neurology) (2009) · New York Times: New Navy Device Learns by Doing (July 8, 1958) · Microsoft: Learning from Tay’s introduction (March 25, 2016) · BBC News: Computer AI passes Turing test in ‘world first’ (June 9, 2014) · Melanie Arntz, Terry Gregory, Ulrich Zierahn: The Risk of Automation for Jobs in OECD Countries (OECD) (2016) · Benjamin S. Bloom: The 2 Sigma Problem (Educational Researcher) (1984) · AFP via The Express Tribune: OpenAI says it has more than 1 billion active users (July 31, 2026) · Reuters via Cyprus Mail: First-ever AI gold medals at International Math Olympiad claimed by Google and OpenAI (July 2025) · TechCrunch: ChatGPT reaches 900M weekly active users (February 27, 2026) · Sam Altman: The Gentle Singularity (June 10, 2025) · Amin Vahdat, Jeff Dean: How much energy does Google’s AI use? We did the math (Google Cloud) (August 21, 2025) · Epoch AI: Colossus 2 (Frontier Data Centers) (2026) · US Department of Energy: DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers (December 20, 2024) · Center for AI Safety: Statement on AI Extinction Risk (May 30, 2023) · UK Government: The Bletchley Declaration (November 1, 2023) · Bulletin of the Atomic Scientists: 2026 Doomsday Clock Statement (January 27, 2026) · Carl Benedikt Frey, Michael A. Osborne: The Future of Employment (Oxford Martin School) (2013) · David Picard: torch.manual_seed(3407) is all you need (arXiv) (2021) · Cameron R. Jones, Benjamin K. Bergen: Large Language Models Pass the Turing Test (arXiv) (2025)
Leave a Reply