Retired.

Retired titles, newest first

1. 2026 · AtCoder World Tour Finals, Heuristic — last human: Przemysław Dębiak (Psyho). OpenAI model; AtCoder World Tour Finals 2026 Heuristic, Tokyo: one 10-hour optimization problem, 12 human finalists; OpenAI first, far ahead of all 12 human finalists; the 'Humanity Prevails' award went unclaimed. Asterisk: OpenAI entered as a Human vs. AI exhibition. The banner names Dębiak as the 2025 champion and the last human to beat it, not as a 2026 finalist. Przemysław Dębiak (Psyho), Jul 2025: “Humanity has prevailed (for now!)” source
2. 2026 · Half marathon — last human: Jacob Kiplimo. Honor humanoid robot; Beijing E-Town humanoid robot half marathon, run alongside a human race; 50:26, against the human world record of 57:20. Asterisk: Not head-to-head: Kiplimo set 57:20 in Lisbon in March 2026 and didn't race. Remote-controlled robots ran with a 1.2× time handicap, and the robot time is not a ratified record. The 2025 winner took 2:40:42. source
3. 2025 · ICPC World Finals — last human: St. Petersburg State University. OpenAI reasoning models; ICPC World Finals, Baku, same 12 problems and five hours; OpenAI 12 of 12; best human team 11 of 12; Gemini 10 of 12. Asterisk: The AI entries were unofficial and unranked: same problems and time limit, but no full submission logs were published. St. Petersburg State are still the official champions. source
4. 2022 · Drone racing — last human: Alex Vanover, Thomas Bitmatta and Marvin Schaepper. Swift; Head-to-head races on a purpose-built track, Dübendorf airfield hangar, Jun 5–13; published in Nature, Aug 30, 2023; Won 15 of 25 races; fastest lap by half a second. Asterisk: A lab event on the team's own track, reported a year later. Swift lost 40% of its races. source
5. 2021 · Gran Turismo — last human: Takuma Miyazono. GT Sophy; Race Together 2021, second event: four GT Sophy cars against four top drivers over three races; Team score 104–52; Sophy took first and second in every race. Asterisk: Team scoring. Miyazono was the 2020 FIA Gran Turismo champion and the best-known of the four humans. source
6. 2021 · Crosswords — last human: Tyler Hinman. Dr.Fill; American Crossword Puzzle Tournament, held online, Apr 23–25; Top score: 12,825 to the best human's 12,810. Asterisk: Dr.Fill was an unofficial entrant, so Hinman took the trophy. Dr.Fill made three errors and still beat top human Erik Agard (no errors) by 15 points, then solved the playoff puzzle in 49 seconds, over two minutes faster than Hinman. source
7. 2020 · Protein structure prediction — last human: Human-led CASP groups. AlphaFold 2; CASP14 blind assessment; Median 92.4 GDT across all targets; about 90 counts as competitive with experiment. Asterisk: A field, not a person. CASP groups are people using software, and DeepMind's AlphaFold 1 had already led CASP13 in 2018. source
8. 2020 · Air-combat dogfight — last human: “Banger”, USAF F-16 pilot. Heron Systems AI agent; DARPA AlphaDogfight Trials final, simulated F-16 against F-16, streamed live; 5–0. Asterisk: Simulation only. The pilot said some of the rules and constraints of human training weren't there for the AI. His name was withheld. source
9. 2019 · Six-player no-limit poker — last human: Darren Elias and Chris Ferguson. Pluribus; Lab experiments published in Science: five pros plus one AI over 10,000 hands, and one pro against five copies over 5,000 hands each; Pluribus won in both formats; against five pros, about $5 a hand if each chip were $1. Asterisk: A private research experiment scored statistically, not a public contest with a title at stake. Date is publication. source
10. 2019 · Dota 2 — last human: OG. OpenAI Five; OpenAI Five Finals, San Francisco, best of three against the reigning International (TI8) champions; 2–0. Asterisk: Restricted rules: a limited hero pool and some game mechanics disabled. Five humans on one team, so the banner names a team. source
11. 2017 · Shogi — last human: Amahiko Satō. Ponanza; Denō-sen, two games against the reigning Meijin, April–May; 2–0. source
12. 2017 · Heads-up no-limit poker — last human: Dong Kim, Jason Les, Daniel McAulay and Jimmy Chou. Libratus; Brains vs. Artificial Intelligence: Upping the Ante, Rivers Casino, Pittsburgh, 120,000 hands over 20 days; Libratus up $1,766,250 in chips. Asterisk: Chips, not cash. The four were heads-up specialists rather than a reigning champion, since poker has no single title holder. source
13. 2016 · Go — last human: Lee Sedol. AlphaGo; Google DeepMind Challenge Match, five games, Seoul, Mar 9–15; 4–1. Asterisk: Lee was the decade's great champion but not the top-rated player. World No. 1 Ke Jie lost 3–0 to a newer AlphaGo at Wuzhen in May 2017. Lee Sedol, Nov 2019: “Even if I become the number one, there is an entity that cannot be defeated.” source
14. 2015 · Arimaa — last human: Jean Daligault, Lev Ruchka and Mathew Brown. Sharp; Arimaa Challenge: three best-of-three matches against top human defenders, for a prize of about $12,000; 7–2. Asterisk: Arimaa was invented in 2002 specifically to be hard for computers, with a standing prize for the first bot to beat top humans. source
15. 2011 · Jeopardy! — last human: Ken Jennings and Brad Rutter. Watson; Two-game IBM Challenge, taped in January, aired Feb 14–16; $77,147 to $24,000 (Jennings) and $21,600 (Rutter). Asterisk: An exhibition, not a regular game. Critics argued Watson's buzzer speed, more than its knowledge, decided it; Jennings called that edge the nature of the game. Ken Jennings, Feb 2011: “I, for one, welcome our new computer overlords” source
16. 2006 · Scrabble — last human: David Boys. Quackle; Best-of-five Human–Computer Showdown, Toronto; Boys won the first two games; 3–2 (final game 482–465). Asterisk: Boys was a former world champion (1995), not the reigning one. Maven had already beaten top player Adam Logan in a 1998 exhibition. David Boys, Nov 2006: “It's still better to be a human than to be a computer.” source
17. 1997 · Othello — last human: Takeshi Murakami. Logistello; Six-game match, NEC Research Institute, Princeton, Aug 4–7; 6–0. source
18. 1997 · Chess — last human: Garry Kasparov. Deep Blue; Six-game rematch, New York; Deep Blue won game 6 in 19 moves; 3½–2½. Garry Kasparov, Mar 1996: “I could feel—I could smell—a new kind of intelligence across the table.” source
19. 1994 · Checkers — last human: Marion Tinsley. Chinook; Man vs. Machine World Checkers Championship, Computer Museum, Boston; Six draws; Tinsley withdrew and the title passed to Chinook. Asterisk: Chinook never beat Tinsley for the title. He withdrew ill after six drawn games, was diagnosed with pancreatic cancer days later, and died on April 3, 1995. Chinook drew its follow-up match with Don Lafferty. source
20. 1979 · Backgammon — last human: Luigi Villa. BKG 9.8; Exhibition match to 7 points, Monte Carlo, the day after Villa won the world title; 7–1. Asterisk: Dice. Berliner himself said the program got the better rolls and Villa played better over the match. One short match, not a title defense. source
21. 1830 · Horse vs. locomotive — A horse, a legend. Race against a horse-drawn car on the Baltimore and Ohio Railroad The horse won after a belt slipped off the engine's blower; the B&O went to steam anyway. Legend: the B&O Railroad Museum tells it 'according to legend', and the date and details vary. source
22. c. 1870s · Steel-driving — John Henry, a legend. Race against a steam drill while building a railroad tunnel for the Chesapeake & Ohio, West Virginia (by legend, Big Bend Tunnel, 1870–72) He won, and in the song he died with his hammer in his hand. A folk legend. Historians disagree on where it happened, or whether it did: Scott Reynolds Nelson argues for Lewis Tunnel in Virginia. source

Still playing and reserved

• StarCraft II: AlphaStar reached Grandmaster, above 99.8% of ranked players, in 2019, but no sanctioned match against the reigning world champion followed. source
• Table tennis: Sony's Ace now beats ranked pros under ITTF rules, including world top-30 Miyuu Kihara, but has not beaten a world No. 1. source
• International Mathematical Olympiad: A tie at the top. AI systems scored a perfect 42/42 in 2026, but so did seven students, and you can't score higher than perfect. source
• Diplomacy: Meta's Cicero finished in the top 10% of an online blitz league in 2022, scoring more than double the human average, but it has not won a world championship. source
• Akutagawa Prize: Japan's top literary prize for new writers went to a human in 2024, though she said about 5% of the novel was lifted from ChatGPT. source
• Football (soccer): By mid-21st century, a team of fully autonomous humanoid robot soccer players shall win the soccer game, comply with the official rule of the FIFA, against the winner of the most recent World Cup. (Condition set by the RoboCup Federation (verbatim, from 'The road to RoboCup 2050')) source
• Marathon: A bipedal robot, fully autonomous, with no battery swaps, finishes 42.195 km on a course certified for record purposes faster than the men's marathon world record standing that day. (Condition written by retired.cloud) source
• Contract bridge: A partnership of programs, bidding and playing every board with no human help, finishes ahead of the reigning Bermuda Bowl champions over a duplicate match of at least 128 boards played under World Bridge Federation rules. (Condition written by retired.cloud) source
• Booker Prize: A novel whose text was written by an AI system, disclosed as such before the judging, wins the Booker Prize. (Condition written by retired.cloud) source
• Millennium Prize Problems: The Clay Mathematics Institute awards a Millennium Prize for a solution whose proof was found by an AI system, with no human supplying the key idea. (Condition written by retired.cloud) source

Forecasts

• Alan Turing, Oct 1950, for 2000: In about fifty years (around 2000), computers would play the imitation game well enough that an average interrogator has no better than a 70% chance of picking the machine after five minutes of questioning. Arguably met in March 2025, when a UC San Diego three-party test with five-minute chats judged GPT-4.5 to be the human 73% of the time. “I believe that in about fifty years' time it will be possible to programme computers, with a storage capacity of about 10^9, to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning.” — happened Mar 2025. source
• Herbert Simon and Allen Newell, 1958, for 1968: A digital computer would be world chess champion within ten years (talk to the Operations Research Society in 1957; published in Operations Research, 1958). Deep Blue beat Kasparov in 1997, about forty years later. “That within ten years a digital computer will be the world's chess champion, unless the rules bar it from competition.” — happened May 1997. source
• Herbert Simon, 1960, for 1980: Machines would be able to do any work a man can do within twenty years (first in The New Science of Management Decision, 1960; reprinted in The Shape of Automation for Men and Management, 1965). By 1980 nothing of the kind existed; the second AI winter followed. “machines will be capable, within twenty years, of doing any work that a man can do” — did not happen. source
• Marvin Minsky, 1967, for 1997: The problem of creating artificial intelligence would be substantially solved within a generation (Computation: Finite and Infinite Machines, 1967). Read as thirty years, that is 1997. The oft-quoted 1970 LIFE line about a machine with average human intelligence in three to eight years is disputed by Minsky himself, so it is not used here. “Within a generation, I am convinced, few compartments of intellect will remain outside the machine's realm—the problems of creating 'artificial intelligence' will be substantially solved.” — did not happen. source
• John Maynard Keynes, Oct 1930, for 2030: A hundred years on, wealth would be such that three-hour shifts or a fifteen-hour week could satisfy most people's need to work (Economic Possibilities for our Grandchildren, The Nation and Athenaeum, 11 and 18 October 1930). Standard full-time weeks in rich countries are still 35 to 40 hours. “Three-hour shifts or a fifteen-hour week may put off the problem for a great while.” — pending. source
• Ray Kurzweil, 1999, for 2029: A computer would pass the Turing test by 2029 (The Age of Spiritual Machines, 1999). In 2002 he bet Mitch Kapor $20,000 on it (Long Bets #1, resolves end of 2029, with an eight-hour test far stricter than a five-minute chat). A short five-minute three-party test was arguably passed in 2025; the bet is still open. — pending. source
• Rémi Coulom (author of Crazy Stone), May 2014, for 2024: Asked by Wired, after Crazy Stone beat a professional with a four-stone handicap, when a machine would win at Go without a handicap, the leading computer-Go programmer guessed about ten years. AlphaGo beat Fan Hui in October 2015 and Lee Sedol in March 2016. “I think maybe ten years. But I do not like to make predictions.” — happened Mar 2016. source
• 352 machine-learning researchers (Grace et al. 2016 survey of NIPS/ICML authors), 2016, for 2026: Median expert estimate that AI would write a high-school essay that earns high grades by 2026. ChatGPT (November 2022) and GPT-4 (March 2023) made passable essay-writing routine years ahead of the median. — happened Nov 2022. source
• Paul Christiano (bet with Eliezer Yudkowsky), Feb 2022, for 2025: Christiano put 8% on an AI built before the 2025 IMO reaching gold-medal level on it; Yudkowsky took at least 16%. In July 2025 both OpenAI and Google DeepMind reported gold-medal scores (35/42) on IMO 2025 with natural-language proofs under contest time limits. — happened Jul 2025. source
• Geoffrey Hinton, 2016, for 2021: At the Machine Learning and the Market for Intelligence conference in Toronto, Hinton said radiologist training should stop because deep learning would beat radiologists within five years. By 2021 radiologist demand and pay had risen; Hinton later said he was wrong on timing, not direction. “People should stop training radiologists now.” — did not happen. source
• Dario Amodei, Oct 2024, for 2026: "Powerful AI" (a model smarter than a Nobel laureate across most fields; a "country of geniuses in a datacenter") could arrive as early as 2026, with the caveat that it could take much longer (Machines of Loving Grace, October 2024). “I think it could come as early as 2026, though there are also ways it could take much longer.” — pending. source
• Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean (AI 2027), Apr 2025, for Mar 2027: The AI 2027 scenario has a superhuman coder (Agent-3, faster and cheaper than the best human engineers) by March 2027 and a superhuman AI researcher (Agent-4) by September 2027. The authors' December 2025 model update moved their superhuman-coder median out to about 2032. “We predict that the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution.” — pending. source
• 2,778 AI researchers (2023 Expert Survey on Progress in AI, AI Impacts), Oct 2023, for 2047: Aggregate forecast gives a 50% chance of high-level machine intelligence (machines better and cheaper than humans at every task) by 2047, thirteen years earlier than the same survey's 2060 answer in 2022. — pending. source
• 2,778 AI researchers (2023 Expert Survey on Progress in AI, AI Impacts), Oct 2023, for 2116: Aggregate forecast gives a 50% chance that all human occupations are fully automatable by 2116 (down from 2164 in the 2022 survey), with a 10% chance by 2037. The gap between this and the 2047 HLMI answer is a known puzzle of the survey. — pending. source

Sun readings (METR 50% task-completion time horizon)

• Feb 2019 · GPT-2: about 0 minutes (about 0 minutes to about 0 minutes) source
• May 2020 · GPT-3 (davinci-002): about 0 minutes (about 0 minutes to about 0 minutes) source
• Mar 2022 · GPT-3.5 (gpt-3.5-turbo-instruct): about 1 minutes (about 0 minutes to about 1 minutes) source
• Mar 2023 · GPT-4 (March 2023): about 4 minutes (about 2 minutes to about 8 minutes) source
• May 2024 · GPT-4o: about 7 minutes (about 4 minutes to about 12 minutes) source
• Jun 2024 · Claude 3.5 Sonnet: about 11 minutes (about 5 minutes to about 23 minutes) source
• Sep 2024 · o1-preview: about 20 minutes (about 12 minutes to about 33 minutes) source
• Oct 2024 · Claude 3.5 Sonnet (Oct 2024): about 21 minutes (about 10 minutes to about 40 minutes) source
• Dec 2024 · o1: about 39 minutes (about 22 minutes to about 1.1 hours) source
• Feb 2025 · Claude 3.7 Sonnet: about 1 hours (about 33 minutes to about 1.8 hours) source
• Apr 2025 · o3: about 2 hours (about 1.2 hours to about 3.3 hours) source
• May 2025 · Claude Opus 4: about 1.7 hours (about 57 minutes to about 2.8 hours) source
• Aug 2025 · Claude Opus 4.1: about 1.7 hours (about 59 minutes to about 2.7 hours) source
• Aug 2025 · GPT-5: about 3.4 hours (about 1.9 hours to about 6.8 hours) source
• Nov 2025 · Gemini 3 Pro: about 3.7 hours (about 2.3 hours to about 6.6 hours) source
• Nov 2025 · GPT-5.1-Codex-Max: about 3.7 hours (about 2.2 hours to about 6.6 hours) source
• Nov 2025 · Claude Opus 4.5: about 4.9 hours (about 2.6 hours to about 11 hours) source
• Dec 2025 · GPT-5.2: about 5.9 hours (about 3.2 hours to about 13 hours) source
• Feb 2026 · GPT-5.3-Codex: about 5.8 hours (about 3.2 hours to about 16 hours) source
• Feb 2026 · Claude Opus 4.6: about 12 hours (about 5.3 hours to about 73 hours) source
• Feb 2026 · Gemini 3.1 Pro: about 6.4 hours (about 4 hours to about 12 hours) source
• Mar 2026 · Claude Mythos Preview (early snapshot): about 17 hours (about 8.5 hours to about 55 hours) — a floor: at least 16 h source
• Jun 2026 · GPT-5.6 Sol: about 11 hours (about 5 hours to about 40 hours) — not a robust measurement source