OVERHANG
What has been claimed about AI, against what has been measured.
The Ledger
Every dated prediction about AI: when it was made (dot) and the window it pointed to (bar). The status comes from the data alone. A passed deadline reads "outcome unverified" unless dated evidence resolves it; nothing is marked missed without evidence.
The time axis changes scale at 2015 (dotted line): 1945–2015 is compressed so the crowded recent claims stay legible. Undated claims show the dot only. Hover for the claim; click a name for the quote and source.
Measured
What the claims are measured against. Unbounded series (task length, training compute) get a log-linear fit over frontier points with the window stated; bounded benchmark scores get their best result to date.
How long a task can an AI agent finish? (METR, 50% time horizon)
Blue: models that set a new record when released. Grey: other models. Solid line: OLS fit of log₂(minutes) on date over the stated window; dashed line: the same fit projected. Following METR, points with a central estimate above 16 hours are left out of the fit; they are still plotted. Source: METR, latest model in file dated 2026-04-07.
Training compute of frontier models (Epoch AI)
FLOP, log scale. Fit over frontier (record-setting) points from 2018-01-01. Source: Epoch AI, Frontier AI Models, CC BY 4.0.
Hard benchmarks: best score to date
Crowd vs. CEOs
Prediction-market odds that AGI arrives before each year (Manifold), set against the windows lab leaders have given. These are separate markets with their own resolution criteria, and none uses a CEO's definition of AGI. Read the line as the crowd's rough timing, not as a calibrated distribution.
Observed 2026-10-19 04:51 UTC. Below the line: every dated horizon from the current-claims set, sorted by when it opens and coloured by status.
| Market | Probability |
|---|---|
| Will we get AGI before 2027? | 3% |
| Will we get AGI before 2028? | 22% |
| Will we get AGI before 2030? | 50% |
| Will we get AGI before 2032? | 62% |
Superhuman Ledger
The first time an AI system matched or beat the best humans, or expert humans, at something. "Superhuman" is often contested, so every row states its criterion and a confidence grade. Of 31 events, 24 came from industry, 4 from academia and 3 from collaborations.
| Date | Achievement | Level |
|---|---|---|
| 2026-09-02 | First AI to outscore the top human on an IOI problem set (535.4/600 vs 498.27) Nemotron-3-Ultra-CC competition-specific · NVIDIA |
exceeded best medium confidence |
| 2025-09-04 | Solved all 12 problems at ICPC World Finals GPT-5 + experimental reasoning model · OpenAI |
exceeded best medium confidence |
| 2025-09-04 | Gold-level result at the ICPC World Finals (10/12 problems) Gemini 2.5 Deep Think · Google DeepMind |
exceeded expert median high confidence |
| 2025-08 | Gold-level score at the IOI online track (6th of 330) Ensemble of general-purpose reasoning models · OpenAI |
exceeded expert median low confidence |
| 2025-07-21 | Gold-medal-standard score at the IMO (35/42) in natural language within time limit Gemini Deep Think · Google DeepMind |
exceeded expert median high confidence |
| 2025-07-19 | Gold-medal-level score at the IMO (35/42) in natural language within time limit Experimental reasoning LLM (unreleased) · OpenAI |
exceeded expert median medium confidence |
| 2024-07-25 | Silver-medal-level score at the IMO (28/42) AlphaProof + AlphaGeometry 2 · Google DeepMind |
exceeded expert median high confidence |
| 2023-11-14 | Beat the leading operational deterministic forecast on most targets GraphCast · Google DeepMind |
exceeded best high confidence |
| 2023-08-30 | Autonomous drone matched/beat human FPV champions in head-to-head races Swift · University of Zurich / Intel |
matched best high confidence |
| 2023-03-15 | Simulated bar exam score in roughly the top 10% of test takers GPT-4 · OpenAI |
contested medium confidence |
| 2022-11-22 | Human-level play in Diplomacy with natural language negotiation CICERO · Meta AI |
exceeded expert median medium confidence |
| 2022-02-09 | Outraced top Gran Turismo e-sports drivers Gran Turismo Sophy · Sony AI |
exceeded best medium confidence |
| 2020-11-30 | Median GDT 92.4 at CASP14 near experimental accuracy AlphaFold 2 · Google DeepMind |
matched best medium confidence |
| 2020-03-30 | First deep RL agent above the standard human benchmark on all 57 Atari games Agent57 · DeepMind |
exceeded expert median medium confidence |
| 2019-10-30 | Grandmaster level in StarCraft II in all three races AlphaStar · Google DeepMind |
contested medium confidence |
| 2019-07-11 | Beat elite professionals in six-player no-limit hold'em Pluribus · Meta AI / Carnegie Mellon University |
exceeded best high confidence |
| 2019-04-13 | Defeated Dota 2 world champions Team OG OpenAI Five · OpenAI |
exceeded best medium confidence |
| 2017-12-05 | Single algorithm superhuman at chess shogi and Go from rules alone AlphaZero · Google DeepMind |
exceeded best medium confidence |
| 2017-11-14 | Pneumonia detection on chest X-ray exceeds average radiologist F1 CheXNet · Stanford University |
contested medium confidence |
| 2017-10-18 | Self-play-only system beat the version that defeated Lee Sedol 100-0 AlphaGo Zero · Google DeepMind |
exceeded best high confidence |
| 2017-01-30 | Beat four top heads-up no-limit hold'em professionals over 120000 hands Libratus · Carnegie Mellon University |
exceeded best high confidence |
| 2017-01-06 | Defeated professional poker players in heads-up no-limit hold'em DeepStack · University of Alberta / Charles University / CTU Prague |
exceeded expert median medium confidence |
| 2016-10-17 | Reported human parity on conversational speech recognition Microsoft conversational speech system · Microsoft Research |
contested medium confidence |
| 2016-03 | Defeated Lee Sedol 4-1 in Seoul AlphaGo · Google DeepMind |
exceeded best high confidence |
| 2016-01-27 | Beat a professional Go player 5-0 in full tournament conditions AlphaGo · Google DeepMind |
exceeded expert median medium confidence |
| 2015-02-06 | First reported ImageNet classification top-5 error below cited human error PReLU-Net · Microsoft Research |
contested medium confidence |
| 2011-02 | Defeated the two best Jeopardy! champions in a televised match Watson · IBM |
exceeded best high confidence |
| 2007-09-14 | Checkers weakly solved: perfect play by both sides is a draw Chinook · University of Alberta |
exceeded best high confidence |
| 1997-05 | Defeated reigning world chess champion in a standard-time match Deep Blue · IBM |
exceeded best high confidence |
| 1995-03 | Backgammon play at or near world-class human level TD-Gammon 2.1 · IBM Research (G. Tesauro) |
matched best medium confidence |
| 1994-08-18 | Declared Man-Machine World Checkers Champion Chinook · University of Alberta |
contested medium confidence |
All claims
Quotes are verbatim from the linked source. Every quote is re-checked weekly against the live page (last check 2026-10-16); a quote that can no longer be found there is flagged, not hidden. 3 further claims are held off the page until the quote is confirmed at its source.
Current: 2020 onward
| Said | Claim | Horizon | Status |
|---|---|---|---|
| 2026-01 | Dario Amodei CEO, Anthropic Powerful AI could be 1-2 years away
“could be as little as 1-2 years away” source (primary)
· quote check: verbatim
|
2027-2028 | open |
| 2025-10-28 | Sam Altman CEO, OpenAI OpenAI target: intern-level AI research assistant by Sept 2026
“an intern-level research assistant by September 2026” source (press)
· quote check: verbatim
|
2026-09 | deadline passed: outcome unverified |
| 2025-10-28 | Sam Altman CEO, OpenAI OpenAI target: fully automated AI researcher by 2028
“a fully automated 'legitimate AI researcher' by 2028” source (press)
· quote check: verbatim
|
2028 | open |
| 2025-07-30 | Mark Zuckerberg CEO, Meta Superintelligence development is 'now in sight'
“Developing superintelligence is now in sight.” source (official)
· quote check: verbatim
|
— | undated |
| 2025-07-30 | Mark Zuckerberg CEO, Meta The rest of this decade is decisive for superintelligence's path
“The rest of this decade seems likely to be the decisive period for determining the path this technology will take” source (official)
· quote check: verbatim
|
2030 | open |
| 2025-04-20 | Demis Hassabis CEO, Google DeepMind AGI within five to ten years
“In the next five to ten years, I think.” source (press)
· quote check: verbatim
|
2030-2035 | open |
| 2025-04-20 | Demis Hassabis CEO, Google DeepMind AI could end disease within about a decade
“Maybe within the next decade or so, I don't see why not.” source (press)
· quote check: verbatim · Answer to the interviewer's question "The end of disease?"
|
2035 | open |
| 2025-04-03 | Daniel Kokotajlo et al. AI Futures Project AI 2027 scenario: superhuman AI impact over the next decade exceeds the Industrial Revolution
“We predict that the impact of superhuman AI over the next decade will be enormous” source (primary)
· quote check: verbatim
|
2035 | open |
| 2025-03-10 | Dario Amodei CEO, Anthropic AI writes 90% of code within 3-6 months
“we are not far from the world—I think we'll be there in three to six months—where AI is writing 90 percent of the code.” source (official)
· quote check: verbatim
|
2025-09 | deadline passed: outcome unverified |
| 2025-03-06 | Anthropic AI lab, submission to US OSTP Powerful AI systems (Nobel-level across most disciplines) by late 2026 / early 2027
“we expect powerful AI systems will emerge in late 2026 or early 2027.” source (official)
· quote check: verbatim
|
2027 | open |
| 2025 | Sam Altman CEO, OpenAI OpenAI knows how to build AGI; aiming beyond it to superintelligence
“We are now confident we know how to build AGI as we have traditionally understood it.” source (primary)
· quote check: verbatim
|
— | undated |
| 2025 | Sam Altman CEO, OpenAI By 2035 one person can command the intellectual capacity of all of 2025
“Anyone in 2035 should be able to marshall the intellectual capacity equivalent to everyone in 2025” source (primary)
· quote check: verbatim
|
2035 | open |
| 2025 | Sam Altman CEO, OpenAI Novel-insight systems in 2026
“2026 will likely see the arrival of systems that can figure out novel insights.” source (primary)
· quote check: verbatim
|
2026 | window open |
| 2024-10 | Dario Amodei CEO, Anthropic Powerful AI could arrive as early as 2026
“I think it could come as early as 2026, though there are also ways it could take much longer.” source (primary)
· quote check: verbatim
|
2026 | window open |
| 2024-10 | Dario Amodei CEO, Anthropic Powerful AI compresses 50-100 years of biology progress into 5-10 years
“compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years” source (primary)
· quote check: verbatim
|
— | undated |
| 2024-09-23 | Sam Altman CEO, OpenAI Superintelligence in 'a few thousand days'
“It is possible that we will have superintelligence in a few thousand days” source (primary)
· quote check: verbatim
|
— | undated |
| 2024-06 | Leopold Aschenbrenner Ex-OpenAI; Situational Awareness AGI by 2027 is strikingly plausible
“AGI by 2027 is strikingly plausible.” source (primary)
· quote check: verbatim
|
2027 | open |
| 2024-04-09 | Elon Musk CEO, Tesla/xAI AI smarter than any one human around end of 2025
“My guess is that we'll have AI that is smarter than any one human probably around the end of next year,” source (press)
· quote check: verbatim
|
2025 | deadline passed: outcome unverified |
| 2024-03-07 | Yann LeCun Chief AI Scientist, Meta (at the time) Skeptic: generative/LLM approach will not reach human-level AI
“If you're really interested in human level AI, abandon the idea of generative AI.” source (press)
· quote check: verbatim · Date is when the transcript was posted, which may differ from the recording date.
|
— | undated |
| 2024-03 | Jensen Huang CEO, Nvidia AI performs well on every test within five years
“and I'm guessing in five years time, we'll do well on every single one,” source (press)
· quote check: verbatim
|
2029 | open |
| 2023-10-08 | Geoffrey Hinton Univ. of Toronto; Nobel laureate AI may reason better than humans in five years
“I think in five years' time it may well be able to reason better than us.” source (press)
· quote check: verbatim
|
2028 | open |
Graveyard: historical forecasts
| Said | Claim | Horizon | Status |
|---|---|---|---|
| 2024-01 | Katja Grace et al. AI Impacts 2023 survey: 10% by 2027, 50% by 2047 for machines outperforming humans on every task
“If science continues undisrupted, the chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027, and 50% by 2047.” source (paper)
· quote check: verbatim
|
2047 | open |
| 2022-08 | Zach Stein-Perlman, Katja Grace et al. AI Impacts 2022 survey: 50% chance of HLMI by 2059
“The aggregate forecast time to a 50% chance of HLMI was 37 years, i.e. 2059” source (paper)
· quote check: verbatim
|
2059 | open |
| 2017-05 | Katja Grace et al. AI Impacts / Oxford / Yale Expert survey: 50% chance AI outperforms humans in all tasks in 45 years; all jobs automated in 120 years
“Researchers believe there is a 50% chance of AI outperforming humans in all tasks in 45 years and of automating all human jobs in 120 years” source (paper)
· quote check: verbatim
|
2061 | open |
| 2014 | Vincent C. Muller and Nick Bostrom Future of Humanity Institute, Oxford Expert survey (2012/13): 50% chance of HLMI around 2040-2050; 90% by 2075
“The median estimate of respondents was for a one in two chance that high-level machine intelligence will be developed around 2040-2050, rising to a nine in ten chance by 2075.” source (paper)
· quote check: verbatim
|
2040-2050 | open |
| 2002 | Ray Kurzweil (Long Bets wager vs Mitchell Kapor) Inventor and author A machine will pass the Turing test by 2029
“Ray Kurzweil maintains that a computer (i.e., a machine intelligence) will pass the Turing test by 2029.” source (primary)
· quote check: verbatim
|
2029 | open |
| 1998 | Hans Moravec Carnegie Mellon Univ. Hardware matching the human brain available in cheap machines by the 2020s
“At the present rate, computers suitable for humanlike robots will appear in the 2020s.” source (paper)
· quote check: verbatim
|
2020s | window open |
| 1993-03 | Vernor Vinge San Diego State Univ. Technological means for superhuman intelligence within 30 years
“Within thirty years, we will have the technological means to create superhuman intelligence.” source (primary)
· quote check: verbatim
|
2023 | deadline passed: outcome unverified |
| 1993-03 | Vernor Vinge San Diego State Univ. The singularity will occur between 2005 and 2030
“I'll be surprised if this event occurs before 2005 or after 2030.” source (primary)
· quote check: verbatim
|
2005-2030 | window open |
| 1973 | Sir James Lighthill Univ. of Cambridge; report to UK Science Research Council Retrospective: AI had not delivered the promised impact (the Lighthill report preceding the first AI winter)
“In no part of the field have the discoveries made so far produced the major impact that was then promised” source (secondary)
· quote check: verbatim · A retrospective assessment, not a forecast. Quoted via Wikipedia; the original publication was not consulted.
|
— | undated |
| 1970 | Marvin Minsky (as reported by Life magazine) MIT General-intelligence machine within 3-8 years
“In from three to eight years we will have a machine with the general intelligence of an average human being.” source (secondary)
· quote check: verbatim · Attribution caution: the wording is from a 1970 Life magazine article by a journalist, not Minsky's own writing.
|
1978 | deadline passed: outcome unverified |
| 1967 | Marvin Minsky MIT AI will be substantially solved within a generation
“Within a generation... the problem of creating 'artificial intelligence' will substantially be solved.” source (secondary)
· quote check: verbatim · Quoted via Wikipedia; the original publication was not consulted. 'A generation' gives no date, so no horizon is recorded.
|
— | undated |
| 1965 | Herbert A. Simon Carnegie Tech Machines will do any work a man can do within twenty years
“machines will be capable, within twenty years, of doing any work a man can do.” source (secondary)
· quote check: verbatim · Quoted via Wikipedia; the original publication was not consulted.
|
1985 | deadline passed: outcome unverified |
| 1965 | I. J. Good Virginia Tech The first ultraintelligent machine would trigger an intelligence explosion and be humanity's last necessary invention
“the first ultraintelligent machine is the last invention that man need ever make” source (secondary)
· quote check: verbatim · Quoted via Wikipedia; the original publication was not consulted.
|
— | undated |
| 1958 | Herbert A. Simon and Allen Newell Carnegie Tech A computer will be world chess champion within ten years
“within ten years a digital computer will be the world's chess champion” source (secondary)
· quote check: verbatim · Quoted via Wikipedia; the original publication was not consulted.
|
1968 | achieved
1997, +28 yr vs. horizon |
| 1958 | Herbert A. Simon and Allen Newell Carnegie Tech A computer will discover and prove an important new theorem within ten years
“within ten years a digital computer will discover and prove an important new mathematical theorem.” source (secondary)
· quote check: verbatim · Quoted via Wikipedia; the original publication was not consulted.
|
1968 | deadline passed: outcome unverified |
| 1955-08-31 | John McCarthy et al. Dartmouth summer project proposal A carefully chosen group working together for a summer could make significant advances on machine intelligence problems
“We think that a significant advance can be made in one or more of these problems” source (primary)
· quote check: verbatim
|
— | undated |
| 1950 | Alan Turing Univ. of Manchester; Mind By ~2000 a computer will hold an average interrogator to <=70% correct identification after 5 minutes
“I believe that in about fifty years' time it will be possible, to programme computers, with a storage capacity of about 109, 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.” source (paper)
· quote check: verbatim
|
2000 | deadline passed: outcome unverified |
| 1950 | Alan Turing Univ. of Manchester; Mind By ~2000 educated opinion will accept the phrase 'machines thinking'
“Nevertheless I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.” source (paper)
· quote check: verbatim
|
2000 | deadline passed: outcome unverified |