Timeline of AI timelines

This is a timeline of AI timelines, the study of advances in artificial intelligence, in particular when artificial general intelligence will be created.

Sample questions

The following are some interesting questions that can be answered by reading this timeline:

• What are the earliest quantitative predictions about when machines would achieve human-level intelligence?
  • Sort the full timeline by "Year of prediction" and look at the oldest rows.
  • You will see predictions by Alan Turing (by 2000), Herbert A. Simon (by 1985), and I. J. Good (before the year 2000), all made before the field of AI had produced any working systems of note.

• How have expert survey estimates of AI timelines changed over time?
  • Sort the full timeline by "Year of prediction" and look for rows with "survey" in the Predictor column.
  • You will see a consistent pattern of shortening timelines: the Future of Humanity Institute Winter Intelligence Survey (2011) placed 50% probability of Human-level machine intelligence at 2050; the AI Impacts ESPAI (2016) placed it at 2061; the 2022 ESPAI at 2059; and the 2023 ESPAI at 2047 — a 13-year shift in 14 months following the release of ChatGPT.

• What methodologies have been used to forecast AGI timelines, and when were they developed?
  • Sort the full timeline by "Concept" to group methodology rows together.
  • You will see three main approaches: hardware extrapolation / Biological anchors (Hans Moravec 1988, Nick Bostrom 1998, Ray Kurzweil 2001–2005, Ajeya Cotra 2020); expert elicitation via surveys (Future of Humanity Institute 2011, Müller–Bostrom 2012–13, AI Impacts 2016–2023); and capability benchmarking (AlexNet 2012, AlphaGo 2016, MMLU 2020, METR 2025).

• What are the most pessimistic and most optimistic predictions in the timeline?
  • Sort the full timeline by "Predicted year" to see the full range.
  • The most optimistic include Elon Musk's 2023 prediction of superintelligence by 2026 and Dario Amodei's "as early as 2026." The most pessimistic include Robin Hanson's fractional progress methodology implying timelines beyond 2100, and the Walsh 2017 survey's robotics experts placing 90% probability of HLMI at 2118.

• What are the main methodological critiques of AI timeline forecasting?
  • Sort the full timeline by "Concept" and look for rows with "critique" or "methodology" in the Concept column.
  • You will find critiques spanning five decades: Stuart Armstrong and Sotala (2012) documenting systematic prediction failures; Eliezer Yudkowsky (2017) arguing there is no fire alarm signal for artificial general intelligence; Yudkowsky (2021) arguing biological anchors are fundamentally invalid; titotal (2025) finding mathematical errors in the AI 2027 timelines model; and Greenblatt (2025) translating Anthropic's qualitative prediction into falsifiable benchmarks.

• What do mass expert sign-ons reveal about implicit timeline assumptions?
  • Sort the full timeline by "Concept" and look for rows with "mass expert sign-on" in the Concept column.
  • You will find two major sign-ons from 2023: the Future of Life Institute's "Pause Giant AI Experiments" open letter (March, 30,000+ signatures), which encodes a short-timeline assumption by calling for an immediate pause; and the Center for AI Safety's one-sentence "Statement on AI Risk" (May, signed by the CEOs of OpenAI, Anthropic, and Google DeepMind simultaneously), which goes further by asserting that the risk of extinction from AI is already comparable in priority to nuclear war and pandemics. The contrast between the two statements — and the contrast between both and the formal expert surveys conducted at the same time — illustrates how implicit timeline claims in advocacy documents can diverge substantially from elicited probabilistic estimates.

• Which predictions from the 1990s and 2000s have aged best?
  • Sort the full timeline by "Year of prediction" and look at rows from 1993–2005.
  • Vernor Vinge's 1993 prediction of superhuman intelligence before 2030 and Ray Kurzweil's 2005 prediction of human-level AI by 2029 are both still live as of the mid-2020s, whereas Hans Moravec's 1988 prediction of strong AI in supercomputers by 2010 proved incorrect on its target date.

Big picture

Time period	Development summary	More details
1950s–1960s	Founding optimism and the first AGI predictions	The formal establishment of artificial intelligence as a field at the 1956 Dartmouth conference is accompanied by an embedded prediction that general machine intelligence is solvable in the near term. Pioneers including Alan Turing, Herbert A. Simon, and I. J. Good make the earliest quantitative forecasts: Turing predicts machines will pass his imitation game by 2000; Simon predicts machines will do any human work within twenty years; Good introduces the concept of an intelligence explosion and judges an ultraintelligent machine more likely than not before the century's end. These predictions reflect the broad optimism of the post-Dartmouth period and would prove dramatically incorrect, contributing to the first AI winter of the 1970s.
1970s–1990s	Hardware extrapolation and the biological anchors methodology	As AI research matures and early optimism gives way to a more sober assessment of the remaining obstacles, a new forecasting methodology emerges: estimating the computational capacity of the human brain and projecting when hardware would match it. Hans Moravec's 1988 and 1997 predictions exemplify this approach, placing human-level AI at 2010–2030 based on Moore's law extrapolation. Nick Bostrom's 1998 paper applies the hardware argument more rigorously, concluding that superintelligence is likely within the first third of the twenty-first century. Ray Kurzweil's 2001 Law of accelerating returns generalizes this into an exponential framework that would prove highly influential, and his 2005 book The Singularity Is Near brings the concept of the technological singularity to mainstream audiences with specific predictions of human-level AI by 2029 and the singularity by 2045. The methodology — later formalized as "biological anchors" — becomes the dominant paradigm for quantitative AGI forecasting for the following two decades. Vernor Vinge's 1993 essay had earlier coined the term "singularity" and predicted superhuman intelligence before 2030.
2000s–2016	Expert surveys and the institutionalization of timeline forecasting	The question of when artificial general intelligence will arrive shifts from individual predictions to systematic expert elicitation. The FHI Winter Intelligence Survey (2011), the Müller–Bostrom survey of ~550 experts (2012–2013), the Walsh survey (2017), and the landmark AI Impacts Expert Survey on Progress in AI (2016, 352 researchers, median HLMI ~2061) establish the survey as the primary empirical instrument for assessing the distribution of expert opinion. These surveys consistently reveal wide disagreement, with individual estimates spanning decades, and a persistent pattern in which domain expertise correlates with longer predicted timelines. The 2012 AlexNet result at ImageNet provides the first major empirical signal that benchmark progress can be discontinuous, foreshadowing the scaling era. Methodological critiques by Stuart Armstrong and Kaj Sotala (2012) document systematic failure modes in AI timeline predictions.
2017–2022	Scaling, compute trends, and the shortening of expert timelines	The publication of OpenAI's "AI and Compute" post (2018) documents a 3.4-month compute doubling time, establishing compute growth as a central empirical anchor for forecasting — later corrected to ~5–6 months by Epoch AI's 2022 analysis of a larger dataset. Eliezer Yudkowsky's 2017 "fire alarm" essay argues that no socially legible warning signal for AGI will exist, and his 2021 critique argues the biological anchors methodology is fundamentally invalid. Ajeya Cotra's biological anchors draft report (2020) and Gwern Branwen's scaling hypothesis essay (2020) provide the most influential articulations of the view that scaling alone may be sufficient for transformative AI. The MMLU benchmark (2020) and the Metaculus #3479 question provide quantitative records of how fast AI capability and forecaster expectations are moving. The 2022 AI Impacts ESPAI (738 researchers) finds the median HLMI estimate has fallen to ~2059 — eight years shorter than 2016.
2023–present	Convergence of short timelines and the emergence of falsifiable forecasting	The release of ChatGPT and GPT-4 triggers a dramatic revision of expert expectations: the 2023 AI Impacts ESPAI (2,778 researchers) finds the HLMI median has fallen to 2047, a thirteen-year shift in fourteen months. Geoffrey Hinton and Yoshua Bengio separately revise their timelines from "30–50 years" to "5–20 years" in 2023, resign from or reduce ties with major AI labs, and co-found the International Dialogues on AI Safety. Leading figures at frontier AI labs — Dario Amodei, Sam Altman, Elon Musk, Demis Hassabis, Jensen Huang — make public predictions placing transformative or superintelligent AI between 2026 and 2030. The AI 2027 scenario (2025) and Toby Ord's "Broad Timelines" (2026) represent opposing responses to this convergence: the former commits to a concrete quantitative narrative with a modal 2027 date, the latter argues that the correct epistemic response is a broad distribution with a median around 2038. Methodological critiques by titotal and Ryan Greenblatt push the field toward more falsifiable intermediate predictions anchored to empirical benchmarks such as METR's task horizon metric.

Inclusion criteria

This timeline covers predictions, forecasts, surveys, methodological frameworks, and empirical benchmarks relevant to the question of when artificial general intelligence or transformative AI will be developed. The following criteria guide what is included and excluded.

Included

• Quantitative timeline predictions by named individuals or groups, giving a specific year or range of years for the arrival of human-level AI, AGI, superintelligence, transformative AI, or a closely related concept. The predictor should be a researcher, technologist, institution, or forecasting community with some claim to relevance — either professional expertise, institutional standing, or demonstrated forecasting track record.

• Expert surveys that elicit probabilistic timeline estimates from populations of AI researchers or related specialists, provided they use a defined methodology and report aggregate results. Surveys with fewer than approximately 50 respondents are included only if they are historically significant (e.g. the AGI-11 conference survey) or methodologically novel.

• Methodological frameworks for estimating AGI timelines, including biological anchors, hardware extrapolation, semi-informative priors, fractional progress estimation, and capability benchmarking. A framework warrants a row if it introduces a distinct approach that is subsequently cited by other forecasters, regardless of whether it produces a specific predicted year.

• Benchmark milestones that provide empirical inputs to timeline reasoning, such as AlexNet (2012), MMLU (2020), and METR's task horizon metric (2025). These are included when the benchmark's trajectory is explicitly used in forecasting arguments, not merely as a record of AI capability.

• Methodological critiques of AI timeline forecasting that identify systematic failure modes, mathematical errors, or epistemological problems with existing approaches, provided they are sufficiently detailed and cited. Opinion pieces or polemics without analytical content are excluded even if widely read.

• Compute trend analyses that quantify the rate of growth in AI training compute and inform timeline extrapolation, such as the OpenAI "AI and Compute" post (2018) and the Epoch AI three-era analysis (2022).

• Prediction markets and forecasting platform questions whose community forecast trajectory over time constitutes a notable record of how informed forecasters updated in response to AI developments, such as Metaculus question #3479.

Excluded

• Market revenue forecasts and commercial AI adoption projections, which predict economic outcomes rather than capability thresholds.

• Consumer technology speculation and journalistic roundups that aggregate or restate existing predictions without adding original analysis.

• Predictions by individuals with no relevant expertise or institutional standing, or predictions made in passing without elaboration.

• Vague qualitative statements that do not imply a specific timeframe or probability, even from prominent figures — for example, statements that AGI is "coming eventually" or "decades away" without further quantification.

• Posts and essays whose primary contribution is sociological commentary on the AI forecasting community (e.g. arguments about funding incentives or groupthink) rather than original predictions or methodological analysis.

• Endorsements of others' predictions, which are noted within the relevant row rather than given separate entries.

Timeline

Year of prediction	Predicted year	Concept	Predictor	Details
1950	2000	Machine intelligence (imitation game)	Alan Turing	English mathematician and computer science pioneer Alan Turing publishes "Computing Machinery and Intelligence" in the journal Mind, posing the question "Can machines think?" and proposing to replace it with what he calls the "imitation game": a test in which a human judge, communicating only through text, must distinguish between a human and a computer. The paper predicts that within approximately 50 years — by around 2000 — it will be possible to program computers to play the imitation game well enough that an average interrogator would have no more than a 70% chance of correctly identifying the machine after five minutes of questioning. Turing also predicts that by the end of the century, educated opinion would have shifted enough that speaking of machines thinking would no longer invite contradiction. The paper is widely regarded as the founding document of artificial intelligence as a philosophical and scientific discipline, and the imitation game would come to be known as the Turing test, becoming the most cited benchmark concept in AI for decades. The specific 2000 prediction is generally considered not to have been met, though the question of what constitutes passing the Turing test remains contested; large language models in the 2020s reignited debate about whether the threshold had finally been crossed.
1956	~1956	Artificial general intelligence	John McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon	The Dartmouth Summer Research Project on Artificial Intelligence, organized by mathematician John McCarthy together with Marvin Minsky, Nathaniel Rochester, and Claude Shannon, marks the formal founding of artificial intelligence as a research field. The 1955 proposal for the conference, submitted to the Rockefeller Foundation for funding, states that "a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer" — an implicit prediction that the core problems of machine intelligence were tractable enough to make decisive progress on in weeks. McCarthy would later reflect that the participants "thought that we would be able to solve the problem of intelligence within one summer." The conference introduces the term "artificial intelligence," establishes a shared research agenda, and produces early programs including the Logic Theorist and early work on chess-playing machines. The summer's optimism would prove dramatically misplaced: the problems of general machine intelligence would resist solution for decades, and the gap between the conference's ambitions and actual progress would contribute to the first AI winter in the 1970s. The Dartmouth conference is nonetheless regarded as the founding moment of AI as a discipline, making its embedded timeline assumption — that general intelligence was an engineering problem solvable in the near term — the original and most consequential overestimate in the history of AI forecasting.
1965	1985	Strong AI	Herbert A. Simon	American economist, cognitive scientist, and AI pioneer Herbert A. Simon — who had co-developed the Logic Theorist and General Problem Solver programs with Allen Newell at Carnegie Mellon, and who would later receive both the Turing Award (1975) and the Nobel Prize in Economics (1978) — writes in The Shape of Automation for Men and Management that "machines will be capable, within twenty years, of doing any work a man can do." The prediction reflects the broad optimism of the post-Dartmouth period, during which many AI researchers believed general machine intelligence was a matter of engineering effort rather than fundamental scientific obstacles. The claim originates in Simon's 1960 book The New Science of Management Decision and is reprinted in the 1965 volume; it is frequently cited in isolation, though in context Simon also argued that humans would retain comparative advantage over machines in many tasks for the foreseeable future. The prediction would prove incorrect by its 1985 target date, and is now widely cited as an exemplar of the overconfidence that characterised early AI forecasting — a pattern that would contribute to the first AI winter in the 1970s.
1965	2000	Superintelligence	I. J. Good	British mathematician and cryptologist I. J. Good, who had worked alongside Alan Turing at Bletchley Park during World War II and later collaborated with him on early computer design, publishes "Speculations Concerning the First Ultraintelligent Machine" in Advances in Computers. Good defines an ultraintelligent machine as one that can far surpass all the intellectual activities of any human however clever, and argues that since machine design is itself an intellectual activity, such a machine could design even better machines, triggering an "intelligence explosion" that would leave human intelligence far behind. In the paper's concluding remarks, Good judges it more probable than not that an ultraintelligent machine would be built within the twentieth century. The paper originates the concept later known as the technological singularity and would prove enormously influential, with later thinkers including Vernor Vinge and Ray Kurzweil reprinting and building on Good's argument.
1993	2005–2030	Superintelligence / technological singularity	Vernor Vinge	American mathematician and science fiction author Vernor Vinge presents "The Coming Technological Singularity: How to Survive in the Post-Human Era" at the NASA VISION-21 Symposium in March 1993, later published in Whole Earth Review. The essay opens with the prediction that "within thirty years, we will have the technological means to create superhuman intelligence" — placing the event by approximately 2023 — and states that Vinge would be surprised if it occurred before 2005 or after 2030. The essay argues that once greater-than-human intelligence exists, it will improve itself recursively, producing technological progress that accelerates beyond human comprehension — a transition Vinge compares to the unknowable space-time at the centre of a black hole, coining the term "singularity" for it. The essay builds on Vinge's 1983 Omni op-ed and becomes one of the most widely cited documents in the history of AI forecasting, directly influencing Ray Kurzweil's subsequent work and the broader discourse around I. J. Good's intelligence explosion concept.
1988	2010, 2030	Strong AI	Hans Moravec	In Mind Children: The Future of Robot and Human Intelligence, Carnegie Mellon roboticist Hans Moravec predicts human-level intelligence for supercomputers by 2010 and for personal computers by 2030. The prediction rests on extrapolating Moore's law trends in computing hardware: Moravec estimates the computational capacity of the human brain and projects when affordable machines would match it, reasoning that hardware parity would be sufficient to produce strong AI. The book also envisions a longer-term future in which superintelligent machines evolve independently of human biology. The supercomputer prediction would prove incorrect by its target date, becoming a frequently cited example of the pitfalls of pure hardware-extrapolation reasoning for AGI timelines.
1997	2020s	Strong AI	Hans Moravec	Carnegie Mellon roboticist Hans Moravec, whose 1988 book Mind Children had made an earlier hardware-extrapolation prediction that supercomputers would reach human-level intelligence by 2010, publishes "When Will Computer Hardware Match the Human Brain?" in the Journal of Evolution and Technology. Applying a similar methodology to his earlier work, Moravec estimates the human brain's processing capacity at approximately 100 million MIPS — derived by scaling from studies of retinal computation — and uses Deep Blue's 1997 defeat of chess champion Garry Kasparov as a calibration point, noting it operated at roughly 1/30 of estimated total human performance. Extrapolating Moore's Law trends, he projects that hardware of comparable power to the human brain will be available in inexpensive machines by the 2020s. The prediction does not rely on detailed models of human cognition, and comparisons between hardware capability and general intelligence remain uncertain due to wide variation in estimates of the brain's computational capacity and the gap between hardware parity and actual cognitive performance. The hardware prediction would prove broadly accurate — computers matching Moravec's MIPS estimate became widely available in the 2020s — but the arrival of human-level AI did not automatically follow, illustrating the limits of the hardware-sufficiency assumption underlying the biological anchors approach.
1997		Artificial general intelligence (term)	Mark Gubrud	Physics graduate student Mark Gubrud, then at the University of Maryland, presents "Nanotechnology and International Security" at the Fifth Foresight Conference on Molecular Nanotechnology in Palo Alto. The paper warns that breakthroughs in nanotechnology and advanced AI would redefine warfare and potentially exceed the destructiveness of nuclear weapons, and calls on nations to abandon the warrior tradition in the face of these risks. To distinguish the kind of powerful, human-level intellect he is concerned about from the narrow expert systems that dominate AI research at the time, Gubrud introduces the phrase "advanced artificial general intelligence" — apparently the first documented use of the term. The coinage goes largely unnoticed for several years. Around 2002, researchers Shane Legg and Ben Goertzel, searching for a label for human-like machine reasoning that was more precise than "strong AI" or "real AI," independently converge on "artificial general intelligence" and begin popularizing it; it is only around 2005, when Gubrud himself surfaces in online AGI community discussions citing his 1997 paper, that his priority is established — to the surprise of those who believed the term had originated with Legg and Goertzel. The term would go on to become arguably the central organizing concept of the AI field in the 2020s.
1998	~2033	Superintelligence	Nick Bostrom	Swedish philosopher Nick Bostrom, then at the London School of Economics and later founder of the Future of Humanity Institute at Oxford, publishes "How Long Before Superintelligence?" in the International Journal of Futures Studies. The paper is one of the earliest systematic treatments of the hardware argument for AI timelines: Bostrom surveys estimates of the processing power of the human brain, projects when computer hardware will achieve comparable performance by extrapolating hardware performance trends, examines bottom-up software approaches modelled on biological brains, and considers how fast superintelligence might be developed once human-level AI is achieved. The paper concludes that superhuman artificial intelligence is likely within the first third of the twenty-first century — implying a rough target of around 2033 or earlier — while carefully noting that the estimate depends on uncertain inputs at each step and could be substantially wrong in either direction. The paper establishes Bostrom as a central figure in systematic AI timeline reasoning more than a decade before the publication of Superintelligence: Paths, Dangers, Strategies (2014), and is an important early synthesis of the hardware-anchored forecasting methodology that would later be developed more rigorously by Kurzweil, Moravec, and Cotra.
2001	2023, 2049, 2059	Artificial general intelligence	Ray Kurzweil	Inventor and futurist Ray Kurzweil publishes his essay "The Law of Accelerating Returns," articulating a framework for predicting technological progress that would become one of the most cited — and contested — arguments in AI forecasting. Building on his 1999 book The Age of Spiritual Machines, Kurzweil argues that the rate of progress in information technology grows exponentially rather than linearly, generalizing Moore's law beyond transistor density to computing power, genome sequencing, brain scanning resolution, and other fields. Applying this framework to brain emulation, he estimates the human brain's computational capacity at approximately 2×1016 calculations per second — derived by multiplying 100 billion neurons by 1,000 connections per neuron by 200 calculations per second per connection — and projects that a $1,000 computer will match this figure around 2023, that the same cost will buy the processing power of all human brains combined by 2049, and that this capacity will cost one cent by 2059. Kurzweil himself notes that compute parity with the brain is a necessary but not sufficient condition for human-level AI, and that the software of intelligence is of equal or greater importance. Critics argue that the biological-anchor methodology conflates hardware performance with general cognitive capability, and that the brain's computational capacity estimates carry wide uncertainty. The essay would prove highly influential in rationalist and effective altruist communities, and Kurzweil's 2029 prediction for human-level AI — elaborated in his 2005 book The Singularity Is Near — became a widely debated benchmark.
2005	2029, 2045	Human-level AI / technological singularity	Ray Kurzweil	Inventor and futurist Ray Kurzweil publishes The Singularity Is Near, the most widely read book in the history of AI timeline forecasting and the work that brings the concept of the technological singularity to mainstream audiences. Building on his 2001 "Law of Accelerating Returns" essay and his 1999 book The Age of Spiritual Machines, Kurzweil makes two central predictions: that computers will pass the Turing test — achieving human-level conversational AI — by 2029, and that the singularity itself will arrive by 2045, the point at which machine intelligence merges with human intelligence and accelerates beyond comprehension. He grounds both predictions in his law of accelerating returns, arguing that information technologies improve exponentially rather than linearly, and projects that a $1,000 computer will match the human brain's estimated computational capacity of 1016 calculations per second "by around 2020." The book distinguishes Kurzweil's vision from Vinge's in one important respect: where Vinge predicts an abrupt discontinuity, Kurzweil predicts a gradual ascent. The 2029 and 2045 dates become the most cited specific AI timeline predictions of the following two decades, and both remain live reference points in AI forecasting discussions in the 2020s — particularly following the rapid advances in large language models from 2022 onward.
2011 (January)	2050	High-level machine intelligence (HLMI)	Future of Humanity Institute (survey)	The Future of Humanity Institute (FHI), a multidisciplinary research centre at Oxford University founded in 2005 by philosopher Nick Bostrom to study existential and civilisational risks, conducts the Winter Intelligence Survey during its AGI impacts conference held at St Anne's College, Oxford in January 2011. The conference is one of the first dedicated academic events to examine the societal implications of artificial general intelligence, bringing together AGI researchers and philosophers to assess both the mechanics and the safety and ethical ramifications of advanced AI — reflecting a growing conviction at FHI that AGI timelines were short enough to warrant serious institutional attention. Participants estimate a median 50% probability of human-level machine intelligence by 2050, with responses indicating approximately a 10% chance in the 2015–2030 range, a 50% chance around 2040–2080, and a 90% chance between 2100 and 2250. The wide distribution reflects deep uncertainty among even specialist participants. The survey represents one of the earliest systematic attempts by an academic institution to elicit probabilistic AI timeline estimates, predating the larger AI Impacts surveys that would follow from 2016 onward, and helping establish expert elicitation as a methodological tool in AI forecasting.
2011 (August)	<2030	Artificial general intelligence (AGI)	AGI-11 Conference participants (survey)	A two-question survey conducted by documentary filmmaker and AI risk author James Barrat — whose 2013 book Our Final Invention would become a widely read popular treatment of AGI risk — and Ben Goertzel, a prominent AGI researcher, chairman of the Artificial General Intelligence Society, and CEO of the SingularityNET platform, polls 60 participants at the AGI-11 conference, the Fourth Conference on Artificial General Intelligence. The conference draws researchers specifically focused on building human-level AI systems, making the participant pool more optimistic about near-term AGI than the broader machine learning research community. Results indicate that nearly half of respondents believe AGI would be achieved before 2030, almost 90% anticipate AGI appearing before 2100, and around 85% expect it to benefit humanity. The survey is a small convenience sample rather than a systematic expert elicitation, and its results reflect the self-selected optimism of a dedicated AGI research community rather than the broader AI field — a distinction that becomes clearer when compared with the more cautious estimates produced by larger surveys of the general AI research population conducted by AI Impacts from 2016 onward.
2008		Takeoff speed / methodology debate	Eliezer Yudkowsky and Robin Hanson	Eliezer Yudkowsky, co-founder of the Machine Intelligence Research Institute (MIRI), and economist Robin Hanson, then at George Mason University, conduct an extended public debate on the blog Overcoming Bias from late 2007 through 2009, later compiled by MIRI into "The Hanson–Yudkowsky AI-Foom Debate" e-book. The debate does not produce a specific timeline prediction but establishes the conceptual vocabulary — fast versus slow takeoff, local versus global intelligence explosion, recursive self-improvement — that structures virtually all subsequent AI timeline reasoning. Yudkowsky argues that once an AI system reaches a sufficient capability threshold it will improve its own cognitive architecture recursively at accelerating speed, producing a discontinuous "foom" event in which a single AI project rapidly acquires a decisive advantage over the rest of the world. Hanson argues that this framing is structurally misleading: economic history suggests that productivity-enhancing technologies diffuse broadly, that improvement requires verification and error-correction which bottlenecks recursive improvement, and that brain emulation — running a copy of a human mind at faster than biological speed — is a more plausible pathway to transformative AI than recursive self-improvement of an initially sub-human system. The debate also contests whether analogy (Yudkowsky's preferred method: examining the history of optimization processes) or outside-view reference-class reasoning (Hanson's preferred method) is the more reliable guide for forecasting unprecedented events. Hanson's fractional progress methodology, introduced in his 2012 informal survey at the UAI conference, is a direct outgrowth of his Foom debate position. Yudkowsky's 2017 "fire alarm" essay and 2021 biological anchors critique both invoke concepts developed in the Foom debate. Nick Bostrom would later acknowledge replacing much of the Foom debate's terminology with his own framework in Superintelligence (2014).
2012 (August)	>2100	High-level machine intelligence (HLMI)	Robin Hanson (fractional progress methodology)	Economist and George Mason University professor Robin Hanson publishes "AI Progress Estimate" on his blog Overcoming Bias, reporting the results of an informal survey he conducts among senior AI researchers at the Uncertainty in Artificial Intelligence (UAI) conference. Hanson asks attendees who have been in their subfield for at least twenty years to estimate what fraction of the distance from where AI was twenty years ago to human-level ability has been covered, and whether progress has been accelerating or decelerating. Respondents consistently estimate that 5–10% of the distance has been covered in twenty years, without noticeable acceleration. On a naive linear extrapolation, covering the remaining 90–95% of the distance at this rate implies at least a century until human-level AI — a substantially longer timeline than most contemporaneous predictions. Hanson continues collecting responses informally through 2017. The methodology — asking practitioners to estimate fractional subfield progress and extrapolating linearly to HLMI — becomes known as the fractional progress or extrapolation argument for AI timelines; it is later incorporated as one survey component in the AI Impacts 2016 ESPAI. When AI Impacts applies the same methodology to the 2016 ESPAI dataset, the implied median timeline is 36 years (year 2056), far shorter than the 372 years (year 2392) implied by Hanson's convenience sample of senior researchers — a divergence AI Impacts attributes partly to Hanson's sampling of exclusively long-tenured researchers, who consistently give more pessimistic estimates than the broader researcher population.
2012 (September)		Benchmark progress / deep learning inflection point	Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton	At the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) — an annual competition founded by Stanford computer scientist Fei-Fei Li in 2010 to benchmark image classification algorithms — a team from the University of Toronto comprising graduate students Alex Krizhevsky and Ilya Sutskever and their PhD supervisor Geoffrey Hinton submits AlexNet, a deep convolutional neural network, achieving a top-5 error rate of 15.3% against the second-place entry's 26.2% — a margin more than ten times larger than year-on-year improvements had previously produced. The result, described by researcher Yann LeCun at the time as "an unequivocal turning point in the history of computer vision," validates deep learning at scale and triggers a rapid reorientation of AI research toward neural networks. The convergence of three factors makes AlexNet possible: the ImageNet dataset itself (millions of labelled images), GPU computing through Nvidia's CUDA platform, and improved training methods for deep networks. The result directly informs AI timeline reasoning by demonstrating that benchmark performance can improve discontinuously rather than incrementally when a new approach is applied at scale — a pattern that would underpin the scaling hypothesis and shorter-timeline predictions throughout the following decade. Hinton, Krizhevsky, and Sutskever would all later join Google, with Sutskever going on to co-found OpenAI and Hinton receiving the 2024 Nobel Prize in Physics for foundational contributions to neural networks.
2012		Methodology critique / meta-analysis of AI predictions	Stuart Armstrong and Kaj Sotala	Stuart Armstrong, a researcher at the Future of Humanity Institute at Oxford, and Kaj Sotala, a researcher at the Singularity Institute (later renamed the Machine Intelligence Research Institute), publish "How We're Predicting AI — or Failing To" in the proceedings of the Beyond AI conference. The paper assembles a database of 257 AI predictions spanning the 1950s to 2012, of which 95 qualify as timeline predictions, drawn from a dataset compiled by Singularity Institute researchers Jonathan Wang and Brian Potter. The paper argues on both theoretical and empirical grounds that AI timeline predictions are systematically unreliable: theoretically, because the task requires estimating the difficulty of problems that are not well understood, and because predictors anchor on their own research area rather than the field as a whole; empirically, because the 95 timeline predictions in the database contradict each other widely and — the paper's headline finding — are statistically indistinguishable from non-expert predictions. This expert/non-expert finding is subsequently identified by AI Impacts as resting on a spreadsheet construction error, a conclusion Armstrong acknowledges. Despite the retracted finding, the paper's theoretical framework — its taxonomy of prediction failure modes and its decomposition schemas for analysing AI forecasts — remains a standard reference in the AI forecasting methodology literature, and the underlying dataset becomes a resource for subsequent survey work.
2012–2013	2040–2050	High-level machine intelligence (HLMI)	Vincent C. Müller and Nick Bostrom (survey)	Philosopher Vincent C. Müller, affiliated with the Future of Humanity Institute at Oxford and Anatolia College in Thessaloniki, and FHI director Nick Bostrom, conduct a survey of approximately 550 AI experts across four groups: participants of the PT-AI conference on philosophy and theory of AI; participants of dedicated AGI conferences; members of the Greek Association for Artificial Intelligence; and the top 100 AI authors by citation count according to Microsoft Academic Search. The survey is motivated by a desire to clarify the actual distribution of expert opinion on HLMI timelines at a time when concern about advanced AI risks was growing in some quarters while being dismissed as science fiction in others. The median estimate across respondents places a 50% probability of HLMI around 2040–2050, rising to a 90% probability by 2075. Respondents also estimate roughly a one in three chance that this development turns out to be bad or extremely bad for humanity. The survey is published in 2016 as part of Müller's edited volume Fundamental Issues of Artificial Intelligence, shortly after Bostrom's Superintelligence: Paths, Dangers, Strategies (2014) had brought questions of advanced AI risk to a much wider audience, lending the survey's findings additional retrospective significance.
2016 (June–August)	2061	High-level machine intelligence (HLMI) / task-specific milestones	AI Impacts (survey)	Katja Grace and John Salvatier of AI Impacts, in collaboration with Allan Dafoe of Yale, Baobao Zhang, and Owain Evans, conduct a survey of 352 machine learning researchers drawn from attendees of the 2015 Neural Information Processing Systems (NIPS) and International Conference on Machine Learning (ICML) conferences — at the time the two most selective venues in the field. The survey, published as "When Will AI Exceed Human Performance? Evidence from AI Experts" (arXiv:1705.08807), asks respondents both about specific task-level milestones and about high-level machine intelligence overall. For individual tasks, respondents estimate median dates of AI outperforming humans at language translation (2024), writing high-school essays (2026), driving a truck (2027), working in retail (2031), writing a bestselling novel (2049), and performing surgery (2053). For HLMI overall, respondents estimate a 50% probability of AI outperforming humans in all tasks within approximately 45 years — around 2061 — and full automation of all human jobs within 120 years, with Asian respondents estimating substantially earlier dates than North American respondents. The survey establishes the methodological baseline from which AI Impacts' subsequent 2022 and 2023 surveys explicitly measure change, making it the anchor point for tracking how expert expectations have shifted over time.
2016 (March 9–15)		Benchmark progress / AI capability inflection point	Google DeepMind (AlphaGo vs. Lee Sedol)	The five-game match between AlphaGo, a deep learning system developed by Google DeepMind, and Lee Sedol, widely considered the strongest Go player of the preceding decade and winner of 18 world titles, takes place at the Four Seasons Hotel in Seoul from March 9 to 15, 2016. AlphaGo wins 4–1, watched live by an estimated 200 million people worldwide. Go had been treated by AI researchers as a near-term benchmark target for decades — its extraordinarily large search space meant that the brute-force search methods that had succeeded in chess were inapplicable, and the conventional view as late as 2014 was that human-level Go performance was ten years or more away. AlphaGo's first professional victory (a 5–0 win over European champion Fan Hui in October 2015) had already surprised the field; the victory over Lee Sedol — by a wider margin than expected, and accompanied by creative moves that expert commentators described as not of human character — represented a qualitative shock. DeepMind's own post-match blog described the result as achieving "a major milestone for artificial intelligence a decade earlier than many predicted." The match triggers a wave of researcher belief updates: post-match commentary from AI researchers documents revision of timelines for AI capability more broadly, and the event is later cited as the first major public event — before the release of GPT-3 and ChatGPT — in which the AI research community was confronted with benchmark performance substantially ahead of prior expectations. The AlphaGo result reinforces the pattern established by AlexNet in 2012: that benchmark performance in AI can improve discontinuously when a new approach is applied at scale, making smooth linear extrapolation an unreliable guide to AI progress.
2016 (March)	>2041	Superintelligence	Association for the Advancement of Artificial Intelligence fellows (survey)	Oren Etzioni, then CEO of the Allen Institute for Artificial Intelligence and Professor of Computer Science at the University of Washington, conducts an anonymous survey on behalf of the Association for the Advancement of Artificial Intelligence (AAAI), posing a question to 193 fellows — researchers recognized as having made significant, sustained contributions to the field — about the timing of superintelligence, using Nick Bostrom's definition from Superintelligence (2014): an intellect much smarter than the best human brains in practically every field. Of the 80 respondents (a 41% response rate), 67.5% expect superintelligence to be achieved but not within 25 years; none expect it within the next 10 years, 7.5% estimate a timeframe of 10–25 years, and 25% believe it will never occur. Etzioni publishes the results in the MIT Technology Review under the headline "No, the Experts Don't Think Superintelligent AI Is a Threat to Humanity," framing the survey as a corrective to what he considers media overstatement of AI risk. The article and its framing draw a published rebuttal from AI researchers Stuart Russell and Allan Dafoe, who argue that the survey design and interpretation understate the significance of even low-probability catastrophic outcomes.
2016 (June 9)	~2100+	High-level machine intelligence (HLMI)	Bill Gates	Microsoft co-founder Bill Gates, in a video interview, states that achieving human-level artificial intelligence will take at least five times longer than the timeline proposed by Ray Kurzweil, who at the time predicts human-level AI by 2029. Gates's multiplier implicitly places human-level AI around or beyond 2100. Gates distinguishes three milestones: substitution of physical and visual labor, AI capable of skilled professional tasks such as writing contracts or medical diagnosis, and finally superintelligence surpassing humanity as a whole. He places himself among those who consider the timing of superintelligence impossible to predict reliably, while still judging it serious enough to warrant attention even within a 50-year timeframe. The comments follow a pattern of public statements by technology figures including Elon Musk and Stephen Hawking raising concern about long-term AI risk, and reflect the growing influence of Nick Bostrom's Superintelligence (2014) in shaping how technologists framed the question.
2017 (January–February)	2026	High-level machine intelligence (HLMI)	Non-experts (Toby Walsh 2017 survey)	Toby Walsh, professor of artificial intelligence at the University of New South Wales and a prominent advocate for the regulation of autonomous weapons, conducts a survey of 849 participants from late January to early February 2017, motivated by growing public concern about automation and technological unemployment. Respondents are divided into three groups: 200 AI experts drawn from two major AI conferences, 101 robotics experts who are fellows or authors of the Institute of Electrical and Electronics Engineers Robotics & Automation Society, and a non-expert group. All respondents classify 70 occupations by automation risk and estimate timelines for the arrival of high-level machine intelligence, defined as computers performing human professions at least as well as a typical human. Non-expert respondents estimate a 10% probability of HLMI by 2026, considerably more optimistic than either group of domain experts, a divergence the paper notes may mean society has more time to prepare for automation than public concern implies.
2017 (January–February)	2109	High-level machine intelligence (HLMI)	AI experts (Toby Walsh 2017 survey)	AI expert respondents in the Walsh 2017 survey — 200 authors drawn from the 2015 Association for the Advancement of Artificial Intelligence conference and the 2011 International Joint Conference on Artificial Intelligence — estimate a 90% probability of achieving high-level machine intelligence by 2109, placing the near-certainty threshold more than nine decades away. This estimate is substantially later than the non-expert 10% estimate of 2026, illustrating a recurring pattern in AI timeline surveys in which domain expertise correlates with longer predicted timelines.
2017 (January–February)	2118	High-level machine intelligence (HLMI)	IEEE Robotics & Automation Society fellows (Toby Walsh 2017 survey)	Robotics expert respondents in the Walsh 2017 survey — 101 participants who are either Fellows of the Institute of Electrical and Electronics Engineers Robotics & Automation Society or contributors to the 2016 IEEE International Conference on Robotics and Automation (ICRA) — estimate a 90% probability of achieving high-level machine intelligence by 2118, placing the near-certainty threshold more than a century away and even further than the AI expert group's estimate of 2109.
2017 (October 13)		Methodology critique / epistemology of AI forecasting	Eliezer Yudkowsky	Eliezer Yudkowsky, co-founder of the Machine Intelligence Research Institute (MIRI), publishes "There's No Fire Alarm for Artificial General Intelligence" on the MIRI website, a long-form essay arguing that there is no socially legible event that will serve as a reliable warning that AGI is imminent. Drawing on the social psychology of pluralistic ignorance — experiments showing that people fail to react to smoke filling a room when others appear unconcerned — Yudkowsky argues that a fire alarm's function is not to provide evidence of fire but to create common knowledge that it is socially safe to react. He contends that AGI lacks any equivalent: the epistemic state of "I don't see how to do the thing" feels identical whether AGI is fifty years away or two years away, because progress at the leading edge is invisible to those not directly involved. Historical examples ground the argument: Wilbur Wright said powered flight was fifty years away in 1901, two years before Kitty Hawk; Enrico Fermi expressed 90% confidence that a nuclear chain reaction was impossible in 1939, three years before he oversaw the first one at the University of Chicago. The essay argues that this pattern is structural rather than correctable, because timing is a function of peak knowledge at the leading project rather than average knowledge in the field. It becomes a standard reference in AI safety discourse for the argument that the absence of an obvious signal is not evidence that AGI is far away, and that waiting for a fire alarm before acting on AI safety is a strategy with no activation condition.
2018 (May 16)		Compute trends / framework for AGI forecasting	OpenAI (Dario Amodei and Danny Hernandez)	OpenAI researchers Dario Amodei and Danny Hernandez publish "AI and Compute," a blog post analysing the amount of computation used in the largest AI training runs since 2012. The post finds that training compute has been increasing exponentially with a 3.4-month doubling time since the advent of deep learning — far faster than the roughly 2-year doubling time of Moore's Law — representing a 300,000× increase over six years. Landmark systems charted include AlexNet, AlphaGo, and AlphaGo Zero. The post does not itself make a specific AGI timeline prediction, but concludes that "as long as this trend continues, it's worth preparing for the implications of systems far outside today's capabilities," framing compute growth as an actionable signal for forecasting. The post becomes a foundational empirical reference in AI timeline reasoning: Ajeya Cotra's biological anchors report, Gwern Branwen's scaling hypothesis essay, and numerous EA Forum analyses cite it as a key input. Amodei would later co-found Anthropic in 2021. Subsequent analyses by Epoch AI and others find the 3.4-month figure may have overstated the trend by conflating a distinct large-scale model regime with the broader trend, placing the corrected doubling time closer to 5–6 months for the period through 2022.
2019	>2029	Artificial general intelligence (AGI)	Gary Marcus and Ernest Davis	Cognitive scientist Gary Marcus, Professor Emeritus at New York University and founder of the machine-learning company Geometric Intelligence (acquired by Uber in 2016), and computer scientist Ernest Davis, professor of computer science at NYU's Courant Institute and a leading authority on commonsense reasoning, publish Rebooting AI: Building Artificial Intelligence We Can Trust. The book argues that achieving human-level general intelligence is far more complicated than the field's boosters claim, contending that the achievements of AI thus far have occurred in closed systems with fixed rules, and that the real world's openness and complexity will require fundamental advances beyond pattern recognition — including robust commonsense reasoning, causal understanding, and physical intuition — that current approaches do not provide. The book explicitly challenges the timelines of Kurzweil and others, arguing that getting to genuinely general intelligence "will require an immense amount of foundational progress — not just more of the same." Marcus would go on to challenge Elon Musk to a $100,000 bet in 2022 that AGI would not arrive by 2029, and issue further structured bets with AI researchers and practitioners through 2024, making him one of the most publicly active advocates of longer AGI timelines in the field. His skeptical position, widely mocked during the height of LLM enthusiasm in 2022–2023, drew renewed attention from 2025 onward as leading researchers including Richard Sutton began publicly questioning whether large language models could serve as a path to AGI.
2020	2020s	Scaling hypothesis / pathway to AGI	Gwern Branwen	Gwern Branwen, a pseudonymous independent American researcher and writer who maintains the long-form essay site gwern.net and is a prominent figure in rationalist and AI forecasting communities, publishes "The Scaling Hypothesis" following the release of GPT-3 by OpenAI in May 2020. The essay articulates what Gwern calls the "strong scaling hypothesis": that intelligence is "just" simple neural units and learning algorithms applied to diverse experiences at sufficient scale, and that as increasing computational resources permit running such algorithms at the necessary scale, neural networks will become progressively more intelligent without requiring fundamentally new algorithms or architectures. Gwern distinguishes this from what he characterises as a "weak" version of the hypothesis — held by organisations such as DeepMind — which holds that the right algorithms must first be found, with compute merely aiding their discovery. Drawing on empirical scaling laws and connecting to historical hardware forecasts by Hans Moravec, he contends that sub-human general intelligence becomes feasible in the 2020s, while cautioning that substantial uncertainty remains about the exact trajectory. The essay also critically argues that the AI research community lacks a coherent model of how AI progress happens, rendering their public timeline forecasts unreliable. The scaling hypothesis as articulated by Gwern would prove highly influential: the rapid capability gains of GPT-4, Claude, Gemini, and other large language models through the 2020s broadly vindicated the framework's core predictions, making the essay a key reference point in retrospective accounts of how the field's trajectory was foreseen.
2020 (August)	2047	Artificial general intelligence (AGI)	LessWrong community (aggregate)	LessWrong, a rationality-focused community blog founded in 2009 that has served as a major hub for AI safety discussion and forecasting, hosts a "Forecasting Thread: AI Timelines" created by community members Amandango, Daniel Kokotajlo, and Ben Pace on August 22, 2020, inspired by a recent AGI timeline update by AI researcher Alex Irpan and motivated by the hypothesis that visualising and comparing individual distributions could improve collective forecasting. The thread uses Elicit, a probabilistic forecasting tool developed by the research organisation Ought, allowing participants to enter probability distributions over AGI arrival dates. Fourteen users submit timeline forecasts; an aggregation weighted by normalised community votes on each comment yields a median estimate of approximately June 2047 for the arrival of human-level AGI. The distribution reflects wide disagreement, with substantial probability mass both before 2040 and after 2100. The organisers note several limitations: the question ("Timeline until human-level AGI") is defined imprecisely, allowing participants to forecast under different definitions of AGI, and the sample is small and self-selected from a community already engaged with AI safety, making it unrepresentative of the broader AI research population.
2020 (September)	2031–2050	Transformative AI (TAI)	Ajeya Cotra	Ajeya Cotra, a senior researcher at Open Philanthropy who had previously studied computer science at UC Berkeley, publishes a draft report on AI timelines that becomes one of the most discussed forecasting frameworks in the AI safety community. The report's central methodology — later termed the "biological anchors" approach — constructs a probability distribution over how much computation would be required to train a transformative AI system using 2020-era algorithms. Cotra builds six anchor distributions, each rooted in an analogy to biology: two of the most prominent are the "lifetime anchor," which grounds training compute estimates in the amount of computation the human brain performs during childhood development, and the "evolution anchor," which grounds estimates in the total computation evolution performed in producing humans from the first neurons. Combining these anchors with projections of compute costs and willingness to spend, the report yields a 10% probability of transformative AI — defined as AI with economic and societal impact comparable to the Industrial Revolution — by the early-to-mid 2030s (approximately 2031–2036 depending on assumptions), and a median estimate of around 2050. Cotra explicitly emphasises substantial uncertainty in both directions: earlier arrival is possible via cheaper training pathways or algorithmic breakthroughs, while later arrival is possible due to non-compute bottlenecks, regulation, or deployment barriers. The report is widely debated in LessWrong and EA Forum communities and draws a prominent critical response from Eliezer Yudkowsky, who argues the biological anchors methodology systematically overestimates the compute required for transformative AI.
2020 (September)		Benchmark methodology / LLM capability measurement	Dan Hendrycks et al.	UC Berkeley researcher Dan Hendrycks and collaborators publish "Measuring Massive Multitask Language Understanding," introducing the MMLU benchmark — 15,908 multiple-choice questions spanning 57 subjects from elementary mathematics and US history to professional law and clinical medicine. The benchmark is motivated by the observation that existing benchmarks such as GLUE (benchmark) and SuperGLUE are becoming saturated: frontier models are approaching or exceeding human performance on them, yet it is unclear whether this reflects genuine broad understanding or narrow pattern matching. MMLU is designed to require broad knowledge across subjects and to remain challenging for the foreseeable future. At launch, GPT-3 (175 billion parameters) achieves 43.9% accuracy against an estimated human expert baseline of approximately 90%, establishing a clear and measurable gap. The benchmark becomes a de facto standard for comparing large language models: GPT-4 achieves 86.4% in March 2023, at which point MMLU itself begins to saturate, prompting successor benchmarks. The trajectory — from 43.9% in 2020 to 86.4% in 2023 — provides a concrete empirical signal used by timeline forecasters to assess the pace of AI capability gain, though critics note that MMLU performance is highly sensitive to prompting choices and that saturation does not necessarily imply human-level reasoning across the tested domains.
2020–2024	2031–2070	Weakly general AI	Metaculus community (ongoing)	Metaculus, an online forecasting platform that aggregates probabilistic predictions from a self-selected community of forecasters — not a prediction market and not limited to domain experts, though with a track record on near-term questions that compares favourably to other aggregation methods — hosts a long-running question (question #3479) asking when the first weakly general AI system will be devised, tested, and publicly announced. The question defines weakly general AI using four operational criteria: reliably passing an adversarial Turing test, scoring at or above the 75th percentile of human students on a standard SAT mathematics section, achieving 90% accuracy on the Winograd schema challenge, and demonstrating general robotic capabilities such as autonomously assembling a scale model from instructions. The community prediction's trajectory over time is itself a notable record of how informed forecasters updated in response to AI progress: the median prediction stood at approximately 2070 in early 2020, fell to around 2058 by early 2022, then dropped sharply to approximately 2040 by July 2022 — an 18-year shift in four months coinciding with a wave of AI capability demonstrations — and continued falling to approximately 2031 by mid-2024. Critics note the definition has structural problems: the robotics requirement may make it harder to satisfy than a definition based on remote work or scientific research, while the absence of requirements for long-horizon agency may make it easier to satisfy than a robust AGI definition. The question's trajectory has been cited as a broader indicator of the AI forecasting community's shifting expectations, even if the specific definition it uses remains contested.
2021 (August 17)	2036, 2060, 2100	Transformative AI (PASTA)	Holden Karnofsky	Holden Karnofsky, co-founder of charity evaluator GiveWell and the philanthropic foundation Open Philanthropy, publishes his AI timeline estimates as part of his "Most Important Century" blog series on Cold Takes. Karnofsky defines the key concept as PASTA — a "process for automating scientific and technological advancement" — meaning AI systems capable of automating all the human activities needed to drive scientific and technological progress, and argues that its development would represent the most consequential event in human history. Drawing on technical reports from Open Philanthropy's Longtermist Worldview Investigations team, particularly work by Ajeya Cotra and Tom Davidson, he forecasts a greater than 10% chance of PASTA-like transformative AI within 15 years (by 2036), an approximately 50% chance within 40 years (by 2060), and an estimated two-thirds probability within the 21st century (by 2100). The series represents a notable shift in how a major philanthropic institution publicly frames AI risk, moving from vague concern to explicit probabilistic forecasting, and contributes to the broader mainstreaming of AI safety as a philanthropic priority.
2021 (December 1)		Methodological critique of biological anchors	Eliezer Yudkowsky	Eliezer Yudkowsky publishes "Biology-Inspired AGI Timelines: The Trick That Never Works" on LessWrong and the AI Alignment Forum, a long dialogue-format essay tracing a recurring methodological failure in AGI timeline forecasting across four decades. The essay reconstructs a lineage of attempts to use biological analogies to estimate AGI arrival: Hans Moravec's 1988 estimate anchored to the compute of a human brain; an unnamed internet poster's approximately 2006 estimate anchored to the total compute used by evolution to produce the human brain; and Ajeya Cotra's 2020 biological anchors report anchored to the compute required to train a neural network of brain size using gradient descent. Yudkowsky argues that all three commit the same error: they identify one of three factors relevant to AGI timing — available hardware, which is graphable — while assuming the other two are estimable from biology when they are not. The second factor, the rate of progress in knowledge about how to build AI, resists quantification; the third, the hardware required to achieve AGI at a given level of such knowledge, is an entirely unknown background parameter that biology does not constrain. Each generation of forecasters believes it has corrected the previous generation's mistake, but Yudkowsky argues the methodology is fundamentally invalid rather than merely incorrectly parameterised. The essay also invokes "Charles Platt's Law" — the observation, generalised in the 1980s, that AGI forecasts always place strong AI approximately thirty years from the date of the forecast — and notes that the 2020 biological anchors report's median of 2050 conforms to it exactly. Holden Karnofsky publishes a reply and Ajeya Cotra comments directly in the thread. The essay is curated by LessWrong and becomes a central reference point in debates over the validity of compute-based AGI forecasting.
2021 (March)	2036–2100+	Artificial general intelligence (AGI)	Tom Davidson (Open Philanthropy)	Tom Davidson, a researcher at the philanthropic foundation Open Philanthropy, publishes "Report on Semi-Informative Priors," a detailed technical treatment of how to estimate AGI timelines from an outside-view perspective — that is, without relying on inside-view knowledge of how AI research works, but instead applying base-rate statistical reasoning about how long ambitious technological projects take. The report's starting point is Laplace's rule of succession, a classical statistical tool that treats each year of AI research as a trial and derives a probability of success based only on the number of years elapsed and the fact that AGI has not yet been achieved. Davidson applies this framework to AI research starting from the Dartmouth conference (1956), yielding an uninformative prior, then refines it into a "semi-informative prior" by incorporating contextual judgements about how ambitious the goal is, what reference class of technologies AGI belongs to, and how to handle the trial-definition problem — whether a trial is a day, a year, or a decade. The resulting estimates place the probability of AGI by 2036 at approximately 8% (range: 1%–18%), rising to 40% by 2100 under the semi-informative prior. Davidson explicitly notes he was not the first to apply Laplace's rule to AGI timelines, citing earlier informal applications in EA community discussions, but provides the most rigorous and comprehensive treatment to date. The report is widely discussed in effective altruist and AI safety communities and becomes a standard reference in debates about outside-view versus inside-view AGI forecasting methodology.
2022 (February 11)		Compute trends / empirical correction of AI and Compute	Epoch AI (Sevilla et al.)	Jaime Sevilla, Lennart Heim, Anson Ho, Tamay Besiroglu, Marius Hobbhahn, and Pablo Villalobos of Epoch AI, an independent research organisation focused on the study of AI progress, publish "Compute Trends Across Three Eras of Machine Learning" on arXiv, later presented at the International Joint Conference on Neural Networks. The paper assembles a dataset of over 120 notable ML models and analyses their training compute from 1952 to 2022, identifying three distinct eras: a Pre-Deep Learning era with a ~2-year compute doubling time; a Deep Learning era beginning with AlexNet in 2012 with a ~6-month doubling time; and a Large-Scale era beginning around 2015–2016 — marked by models such as AlphaGo and GPT-3 — with a ~10-month doubling time and compute requirements 10–100 times greater than the Deep Learning trend. The paper's principal empirical contribution is a correction of OpenAI's influential 2018 "AI and Compute" post, which had reported a 3.4-month doubling time: with approximately ten times more data and an extended period, Epoch AI finds a corrected doubling time of ~5–6 months overall, attributing the discrepancy to OpenAI's smaller sample and its conflation of the distinct large-scale model regime with the broader trend. The paper becomes the standard empirical reference for compute trends in AI timeline forecasting, replacing the 2018 OpenAI figure in most subsequent analyses.
2022 (May 30)	2029	Artificial general intelligence (AGI)	Elon Musk	South African-born American entrepreneur Elon Musk — CEO of electric vehicle manufacturer Tesla and aerospace company SpaceX, co-founder of OpenAI (which he departed in 2018), and at the time one of the most followed accounts on Twitter — states via a tweet in reply to Twitter co-founder Jack Dorsey that artificial general intelligence would be achieved by 2029. The prediction is made in the context of a public conversation about AI timelines during a period of rapidly growing public interest in AI capabilities following the release of GPT-3 and early demonstrations of large language models. The prediction is criticised by cognitive scientist and AI sceptic Gary Marcus, who publicly challenges Musk to a $100,000 bet that AGI would not be achieved by 2029, arguing that Musk's timeline reflects a fundamental misunderstanding of the remaining technical challenges. Musk's commercial interests in autonomous driving through Tesla give his AI timeline claims particular weight and scrutiny beyond those of a typical commentator.
2022 (June–August)	2059	High-level machine intelligence (HLMI)	AI Impacts (survey)	The 2022 Expert Survey on Progress in AI (ESPAI), conducted by AI Impacts, analyzes responses from 738 machine learning researchers regarding progress toward high-level machine intelligence. The median estimate places a 50% probability of HLMI at approximately 37 years in the future, around 2059, representing a decrease of about eight years compared with the 2016 survey. Respondents also estimate a median 5% probability of extremely negative outcomes, such as human extinction, and 69% indicate that AI safety research should be prioritized more than current efforts.
2023	2028–2043	Superintelligence / human-level AI	Geoffrey Hinton and Yoshua Bengio	Geoffrey Hinton, British-Canadian computer scientist and co-inventor of backpropagation who shared the 2018 Turing Award with Bengio and Yann LeCun and would receive the 2024 Nobel Prize in Physics for foundational contributions to neural networks, and Yoshua Bengio, Canadian AI researcher and founding scientific director of Mila (the Quebec AI Institute), separately revise their AGI timelines dramatically following the emergence of large language models. Hinton, who had previously estimated AGI was "30 to 50 years or even longer away," resigns from Google in May 2023 specifically to speak freely about AI risks, and posts on Twitter that he now predicts "5 to 20 years but without much confidence." During Nobel Week in Stockholm in December 2024, he states there is "a good chance — a 50% chance — we'll get AI smarter than us" within that window, implying a median around 2033–2034 and a range of approximately 2028–2043. Bengio similarly revises his estimate, stating in 2023 a 95% confidence interval of 5–20 years for superhuman intelligence. Both Hinton and Bengio sign the May 2023 Center for AI Safety statement declaring that "mitigating the risk of extinction from AI should be a global priority on par with other societal-scale risks such as pandemics and nuclear war," and jointly co-found the International Dialogues on AI Safety (IDAIS) in October 2023 with Chinese AI researchers Andrew Yao and Ya-Qin Zhang. Bengio's timeline revision is notable because he had previously been among the more cautious of the Turing Award trio, while Yann LeCun maintains that human-level AI requires fundamental breakthroughs not yet in sight and resists offering specific timelines.
2023 (March 28)	2023–2024 (implicit)	Mass expert sign-on / implicit short-timeline consensus	Future of Life Institute (open letter)	The Future of Life Institute (FLI) publishes "Pause Giant AI Experiments: An Open Letter" one week after the release of GPT-4, calling on "all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4." The letter states that "AI systems with human-competitive intelligence can pose profound risks to society and humanity" and that "this is not science fiction" — framing the pause as an emergency response to present and imminent danger rather than speculative future risk. The letter receives more than 30,000 signatures, including prominent academics and researchers (Yoshua Bengio, Stuart Russell), technology figures (Elon Musk, Steve Wozniak), and public intellectuals (Yuval Noah Harari). As a mass expert sign-on it functions as an implicit survey: the signatories are collectively asserting that AI capabilities are advancing fast enough to require an immediate emergency pause, encoding a shorter-timeline assumption than most formal expert surveys had produced at equivalent dates. The letter is notable for its institutional breadth — spanning academic AI, AI safety, and industry — and for the gap between its implicit timeline assumptions and the responses from the labs it targets: OpenAI, Google DeepMind, and Anthropic all decline to pause, and several AI researchers publish prominent critiques arguing the letter conflates distinct risks, overstates near-term danger, and would be counterproductive. The letter nonetheless contributes to shifting public AI risk discourse, and is cited as a catalyst for subsequent legislative and regulatory responses in the United States and European Union.
2023 (May 30)		Extinction risk consensus / implicit short-timeline statement	Center for AI Safety (mass sign-on)	The Center for AI Safety (CAIS), a San Francisco-based nonprofit focused on reducing societal-scale risks from artificial intelligence, publishes a one-sentence "Statement on AI Risk": "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." The statement is signed by a historically broad coalition including Geoffrey Hinton, Yoshua Bengio, Demis Hassabis (CEO of Google DeepMind), Sam Altman (CEO of OpenAI), Dario Amodei (CEO of Anthropic), Ilya Sutskever (then Chief Scientist at OpenAI), and Shane Legg (Chief AGI Scientist and co-founder of Google DeepMind) — the simultaneous sign-on of multiple frontier AI lab chief executives being noted at the time as unprecedented. The statement is deliberately minimal, intended as CAIS notes to "create common knowledge of the growing number of experts and public figures who also take some of advanced AI's most severe risks seriously" while accommodating signatories with differing specific beliefs about mechanisms and timelines. Its implicit timeline content is stronger than the March 2023 FLI Pause letter: where the FLI letter called for a pause on the grounds that AI is advancing too fast, the CAIS statement asserts that the risk of extinction is already comparable in priority to nuclear war and pandemics — a framing that presupposes transformative AI is close enough to warrant civilisational-scale concern. Critics including Yann LeCun, who declined to sign, argued that the statement was premature and that current deep learning systems are not on a path to the kind of general intelligence that could pose extinction risk.
2023 (October)	2047, 2116	High-level machine intelligence (HLMI) / Full automation of labor (FAOL)	AI Impacts (survey)	AI Impacts conducts its 2023 Expert Survey on Progress in AI (ESPAI) with 2,778 researchers drawn from six top AI publication venues — roughly four times the size of the 2022 survey — making it probably the largest expert survey of AI researchers ever conducted. Respondents are asked about two differently worded definitions of human-level AI: high-level machine intelligence (HLMI), defined as when unaided machines can accomplish every task better and more cheaply than human workers, and full automation of labor (FAOL), defined as when all occupations are fully automatable. The aggregate forecast places a 50% probability of HLMI by 2047, down thirteen years from 2060 in the 2022 survey, and a 50% probability of FAOL by 2116, down 48 years from 2164. The scale of the shift is striking: between the 2016 and 2022 surveys, the HLMI median had moved by only about one year, whereas in the fourteen months between the 2022 and 2023 surveys — a period that included the release of ChatGPT and GPT-4 — it moved by thirteen. Seventy percent of participants indicate that AI safety research should be prioritized more highly than it currently is.
2023 (November 29)	2026	Superintelligence	Elon Musk	In an interview, Elon Musk — who had founded AI company xAI in July 2023, months before this statement — predicts that artificial intelligence would surpass the intelligence of the smartest human within approximately three years, implying a timeline of around 2026. The prediction is notably more aggressive than his 2022 tweet placing AGI arrival at 2029, and is made shortly after the public release and rapid adoption of ChatGPT had substantially raised mainstream expectations about AI progress. Musk's commercial stake in short AI timelines through xAI, which was at this point in direct competition with OpenAI and Google DeepMind, is relevant context for evaluating the prediction.
2024 (February 17)	~2029	Superintelligence / human extinction	Eliezer Yudkowsky	Eliezer Yudkowsky, co-founder of the Machine Intelligence Research Institute (MIRI) and a foundational figure in the AI safety movement, tells the Guardian that if forced to assign probabilities, his sense is that humanity's remaining timeline before a potentially extinction-level AI event looks more like five years than 50 — with two years and ten years as the plausible range. The statement is made in the context of a broader Guardian feature on AI scepticism, which profiles both Yudkowsky's existential-risk position and a separate current of neo-Luddite critics focused on nearer-term harms such as job displacement, workplace surveillance, and corporate power concentration. Yudkowsky's view is not a prediction of when superintelligence arrives but of how little time he believes remains before a loss of human control becomes irreversible. The claim is consistent with positions he had stated publicly since at least 2022, including a widely discussed Time magazine op-ed calling for a global halt to AI development. Yudkowsky's timeline estimates are disputed by many AI researchers, who consider both the probability and the imminence of such outcomes to be significantly overstated.
2024 (March 19)	~2029	Artificial general intelligence (AGI)	Jensen Huang	Jensen Huang, CEO of Nvidia — the semiconductor company whose graphics processing units supply the compute infrastructure for virtually all frontier AI training runs — states at Nvidia's annual GTC developer conference that AGI is approximately five years away, implying a timeline of around 2029. Huang is careful to note that the prediction depends on how AGI is defined, and declines to make a more specific claim until the questioner can specify what AGI means in context, offering examples such as passing a legal bar exam, a logic test, or a pre-medical exam as possible operationalisations. In the same session, Huang argues that AI hallucinations — the tendency of models to produce plausible but factually incorrect outputs — are solvable through retrieval-augmented generation, a technique that grounds model responses in verified sources. The statement attracts attention partly because Nvidia's commercial interests are directly tied to the pace of AI development: faster AI progress means greater demand for Nvidia hardware, making Huang's bullish timeline estimate relevant context for evaluating its source.
2025 (March)		Benchmark methodology / autonomous task completion	METR	AI safety evaluation organisation METR (Model Evaluation and Threat Research) publishes its task horizon benchmark, measuring the longest duration software engineering and research tasks that frontier AI agents can complete with 50% reliability — a metric directly tied to the question of how close AI is to autonomous knowledge work. The key finding is that the task horizon doubles approximately every six months from 2023 onward: from tasks taking a few minutes in early 2023 to tasks taking several hours by late 2024, with a 2025 update revising the doubling time to approximately three to four months as progress accelerates. Unlike knowledge-testing benchmarks such as MMLU, the task horizon metric is explicitly designed to track progress toward the kind of sustained autonomous agency that would constitute economically transformative AI, making it a direct input to timeline forecasting frameworks. The benchmark becomes one of the most cited empirical references in short-timeline arguments: at a doubling time of six months, extrapolation implies AI systems capable of month-long autonomous tasks — a rough proxy for autonomous AI research — by the late 2020s, though METR notes the extrapolation carries wide uncertainty and that "messy" real-world tasks remain harder than the cleaner benchmark tasks suggest.
2024 (October)	2032	Artificial general intelligence (AGI)	Ray Kurzweil	In The Singularity Is Nearer, a follow-up to his 2005 book The Singularity Is Near, inventor and futurist Ray Kurzweil updates his central AGI timeline prediction, moving it from 2045 — which had been his signature forecast since 2005 — to 2032. The revision reflects Kurzweil's assessment that the pace of AI progress since 2020, driven by large language models and scaling, has exceeded what his earlier exponential extrapolations anticipated. Kurzweil maintains his core methodology — extrapolating exponential trends in computing power and algorithmic efficiency — but applies updated estimates of current capability levels and the remaining gap to human-level intelligence. The book provokes debate about whether the earlier 2045 date was ever a serious prediction or a rough illustrative figure, and whether the shift to 2032 reflects genuine new information or post-hoc rationalisation of progress that had already occurred.
2024 (October)	2026–2027	Powerful AI / AGI	Dario Amodei	Dario Amodei, co-founder and CEO of AI safety company Anthropic — who previously led the team at OpenAI that built GPT-2 and GPT-3 before departing to found Anthropic in 2021 — publishes "Machines of Loving Grace: How AI Could Transform the World for the Better," a 14,000-word essay articulating both a timeline and a vision for what he calls "powerful AI." Amodei explicitly avoids the term AGI, preferring "powerful AI," which he defines as an AI model smarter than a Nobel Prize winner across most disciplines, capable of any task a human could perform remotely, able to act autonomously over minutes, hours, days, or months, and deployable in large numbers simultaneously — summarised as "a country of geniuses in a datacenter." He states that such a system "could come as early as 2026, though there are also ways it could take much longer." In Anthropic's March 2025 submission to the White House Office of Science and Technology Policy, this is strengthened to "we expect powerful AI systems will emerge in late 2026 or early 2027." The essay focuses more on what happens in the five to ten years after powerful AI arrives — predicting compressed progress in biology, mental health, and economic development — than on the timeline itself, and explicitly acknowledges the risks alongside the upside. The prediction is notable as an institutional position from a frontier AI lab rather than an individual speculative claim, and draws critical attention for the internal inconsistency between the relatively casual "as early as 2026" phrasing in the essay and the stronger institutional language in the OSTP submission.
2024 (September 23)	~2030–2035	Superintelligence	Sam Altman	Sam Altman, CEO of OpenAI — the company behind ChatGPT and the GPT series of models — publishes "The Intelligence Age," a personal blog post laying out his vision of AI's trajectory. The post's most widely discussed claim is that "it is possible that we will have superintelligence in a few thousand days," a deliberately vague timeframe that translates to roughly 5–14 years from the post's publication, implying a range of approximately 2030–2038. Altman grounds the prediction in the success of deep learning, arguing that the algorithm can genuinely learn any distribution of data and that capabilities continue to improve predictably as more compute and data are applied — an explicit endorsement of the scaling hypothesis. He warns that if AI infrastructure is not built out sufficiently, advanced AI will become "a very limited resource that wars get fought over and that becomes mostly a tool for rich people." In a separate Bloomberg interview in early 2025, Altman stated more specifically that he believes AGI would "probably get developed during [Trump's] term," implying a date before January 2029, while acknowledging that "AGI has become a very sloppy term." Critics note that the post's timing — published during OpenAI's $6.5 billion fundraising round — gives it a promotional dimension that complicates its interpretation as a purely analytical forecast.
2024–2025	~2030	Artificial general intelligence (AGI)	Demis Hassabis	Demis Hassabis, co-founder and CEO of Google DeepMind — a chess prodigy, cognitive neuroscientist, and AI researcher who co-founded DeepMind in 2010 and received the 2024 Nobel Prize in Chemistry for AlphaFold's contribution to protein structure prediction — makes a series of public statements placing AGI on a 5–10 year horizon. In November 2024 at Axios' AI+ Summit in San Francisco, Hassabis states that "AGI, probably the most transformative moment in human history, is on the horizon" and predicts it could come by 2030. In April 2025 he tells CBS News AGI could arrive in five to ten years. At the India AI Impact Summit in early 2026, he reiterates the five-year estimate and compares AGI's coming impact to ten times the Industrial Revolution at ten times the speed. Hassabis consistently emphasises that his estimates apply to systems meeting a genuinely full human cognitive range — including creativity, continual learning, and robust understanding — and notes that many existing AI systems fall short of this bar in important ways, placing him in a somewhat more cautious position than contemporaries such as Altman and Amodei who apply shorter timelines. Critics on the EA Forum note an apparent tension between Hassabis's detailed enumeration of remaining technical obstacles and his relatively near-term timeline, questioning whether the gaps he identifies could plausibly be closed so quickly.
2021 (August 6)	2026–2027	Artificial general intelligence / transformative AI	Daniel Kokotajlo	Daniel Kokotajlo, then a researcher at OpenAI, publishes "What 2026 Looks Like" on LessWrong, a year-by-year narrative scenario depicting AI development from 2022 through 2026. The essay is written using an incremental methodology — each year's history is written before the next is begun — intended to avoid the retrospective bias of conventional fiction, where endings are designed before beginnings. The scenario predicts: multimodal transformers succeeding GPT-3 by 2022; hype cycles and chatbot proliferation by 2023; AI-enabled propaganda and censorship becoming routine by 2024; agentic "bureaucracies" of chained model calls (later known as agent frameworks) enabling complex multi-step tasks by 2025; and full AI assistant capabilities arriving by 2026. A December 2022 author note clarifies that the essay's intended but unwritten continuation placed AI takeover in 2027 and a technological singularity around 2028–2029, making the 2026 story the first half of a short-timeline scenario rather than a standalone forecast. Among the essay's specific predictions confirmed before ChatGPT: the rise of chain-of-thought reasoning, inference-time scaling, AI chip export controls, and training runs costing $100 million or more. The post is subsequently cited as the methodological predecessor to AI 2027 (2025), which Kokotajlo co-authored after leaving OpenAI, and Kokotajlo's 2021 compute prediction for 2024 (5×10²⁵ FLOPs for the most expensive training run) was confirmed as exactly Epoch AI's central estimate for that year. The essay's retrospective accuracy, combined with its pre-ChatGPT origin date, made it one of the most widely cited examples of successful near-term AI forecasting in the AI safety community.
2023 (June 6)	>2043 (0.4% by 2043)	Transformative AGI	Ari Allyn-Feuer and Ted Sanders	Ari Allyn-Feuer, a director of AI product at GSK, and Ted Sanders, a research engineer at OpenAI, publish "Transformative AGI by 2043 is <1% Likely" as a submission to the Open Philanthropy AI Worldviews Contest. The paper estimates a 0.4% probability of transformative AGI — defined as AI capable of performing nearly all economically valuable tasks at human cost or less — by 2043, arguing that achieving this requires ten jointly necessary events whose cascading conditional probabilities multiply to this low figure. The ten events span software breakthroughs (inventing algorithms for transformative AGI, 60%; inventing a way for AGIs to learn faster than humans, 40%), hardware scaling (inference costs below $25/hour per human equivalent, 16%; cheap quality robots, 60%; massive chip and power scaling, 46%), and avoidance of derailment from regulation, AI-caused delay, war, pandemics, and economic depression. The paper's most contested premise is the 16% inference cost estimate, which the authors ground in a biological anchors-style argument that human-equivalent AGI would require 8–10 orders of magnitude more inference compute than GPT-4. Critics including Daniel Kokotajlo and jimrandomh argue the cascading structure double-counts shared uncertainties, that current LLMs are closer to AGI than the hardware anchor implies, and that the paper underestimates how quickly AI progress compounds once initial breakthroughs occur. The paper represents one of the most systematic attempts to apply Fermi-style decomposition to AGI timelines and remains a reference point in debates about whether AGI by 2043 is a reasonable default expectation.
2025 (April 3)	2027–2028	Artificial general intelligence (AGI) / superintelligence	Daniel Kokotajlo et al. (AI Futures Project)	The AI Futures Project, a nonprofit research organisation, publishes "AI 2027," a detailed scenario forecast produced by Daniel Kokotajlo — a former OpenAI researcher whose earlier 2021 scenario essay had correctly predicted the rise of chain-of-thought reasoning, inference scaling, AI chip export controls, and $100 million training runs more than a year before ChatGPT — alongside blogger Scott Alexander (of Slate Star Codex), forecaster Eli Lifland (ranked first on the RAND Forecasting Initiative all-time leaderboard), AI policy researcher Thomas Larsen (founder of the Center for AI Policy), and Harvard computer science student Romeo Dean. The scenario is informed by approximately 25 tabletop exercises and feedback from over 100 experts in AI governance and technical AI work, and is endorsed by AI researcher Yoshua Bengio. Rather than a single prediction, "AI 2027" is a concrete, quantitative narrative scenario depicting one plausible trajectory of AI development, written iteratively from the present through two alternative endings ("race" and "slowdown"). The scenario's modal prediction at publication is that a superhuman coder — an AI system capable of performing any software engineering task a leading human engineer can do, faster and more cheaply — arrives in 2027, followed by generally superintelligent AI in 2028. The timelines forecast appendix, built on METR's time horizon work, models these milestones via extrapolated capability trends rather than expert elicitation or hardware anchoring. A footnote added in November 2025 clarifies that 2027 was the modal year at time of publication but that the authors' median forecasts are somewhat longer. The scenario methodology — narrative scenario planning grounded in quantitative capability extrapolation — is explicitly distinguished by the authors from both expert survey and biological anchors approaches, and represents a distinct third methodology in the AI timelines literature.
2025 (June 19)		Methodology critique	titotal (pseudonymous)	A computational physicist writing under the pseudonym "titotal" publishes "A deep critique of AI 2027's bad timeline models" on LessWrong, a detailed methodological analysis of the quantitative forecasting appendices underlying the AI 2027 scenario. The critique focuses on the two timeline models in the forecast: the "time horizon extension" model and the "benchmarks and gaps" model. Key findings include: the "superexponential" curve used in the model is guaranteed to reach infinity within a fixed number of years regardless of starting parameters, such that even a nanosecond initial time horizon produces a 2026–2027 median for superhuman coders; the conceptual justifications offered for superexponentiality are either weak or argue against the hypothesis; the "exponential" model is itself superexponential once intermediate R&D speedups are factored in; the intermediate speedup equation, when backcasted to 2022, implies a rate of AI progress inconsistent with the authors' own stated estimates; and the RE-Bench logistic curve fitting presented as the basis for the "benchmarks and gaps" model does not appear in the actual simulation code, with the saturation date instead set manually by the forecasters. The post also documents a graph widely circulated on social media as representing the AI 2027 prediction that was not generated by the timelines model and used different parameters than the published forecast. The AI 2027 team responds constructively, awarding a $500 bounty, acknowledging several communication failures and methodological issues, and committing to model updates. The episode represents one of the most detailed public exchanges on forecasting methodology in the AI timelines literature to date.
2025 (November 3)		Methodology critique	Ryan Greenblatt	Ryan Greenblatt, a researcher at Anthropic, publishes "What's up with Anthropic predicting AGI by early 2027?" on LessWrong, examining and critiquing Anthropic's official prediction that powerful AI will emerge in late 2026 or early 2027. The post operationalizes the vague term "powerful AI" into three testable capabilities — full automation of AI R&D, virtually full automation of remote scientific R&D across most relevant fields, and automation of the vast majority of remote white-collar work — and derives a quantitative intermediate timeline that would need to be on track for the prediction to hold. Greenblatt estimates the probability of powerful AI by early 2027 at approximately 6%, primarily because achieving it would require AI task-horizon doubling times far shorter than the roughly 7-month doubling rate observed in the ongoing METR time-horizon trend. A comparison table shows his median expectation for December 2026 (a 1.75× research engineering multiplier) diverges dramatically from the ~50× multiplier implied by the Anthropic timeline. The post notes that Anthropic's prediction as stated in recommendations to the White House OSTP was stated more confidently than in Dario Amodei's original "Machines of Loving Grace" essay, and calls on proponents of short timelines to make falsifiable intermediate predictions. The post is among the most systematic attempts to translate a qualitative AI capability forecast into a benchmark-anchored, falsifiable timeline with explicit intermediate milestones.
2026 (March 19)	2038 (median); 80% interval ~2029–2126	Transformative AI	Toby Ord	Toby Ord, a philosopher at the University of Oxford and author of The Precipice, publishes "Broad Timelines" on the EA Forum, arguing that the correct epistemic response to lasting expert disagreement on AI timelines is to hold a broad probability distribution over when transformative AI will arrive, rather than committing to either short or long timelines. Ord defines transformative AI as the threshold at which AI systems would be powerful enough to take over the world if misaligned, or have doubled the global rate of scientific and technological progress. He presents his own distribution with a median of 2038 and an 80% interval ranging from approximately three years to one hundred years from the time of writing, and documents similar broad distributions held by Daniel Kokotajlo (80% interval 2027–2050+), Ajeya Cotra, Ege Erdil, and the Metaculus community forecast. The post argues that compressing such distributions into single numbers systematically misleads audiences, and that even proponents of short timelines like Kokotajlo assign only 27% probability to transformative AI arriving by their modal year. Ord draws two practical implications: that uncertainty is not an excuse to assume a preferred timeline, and that long-horizon projects such as founding research fields, writing books, or building organizations retain substantial expected value even under significant probability of near-term transformative AI. The post provoked responses from Ryan Greenblatt, who argued most workers in AI safety should still focus on short timelines given their greater leverage, and from Eli Lifland, who discussed how the AI 2027 scenario's 2027 date was chosen as a modal rather than median scenario.

Numerical and visual data

Predicted year vs. year of prediction

The following scatter plot displays AI timeline predictions from this timeline, plotting each predictor's year of publication on the horizontal axis against their predicted year for human-level machine intelligence, AGI, or transformative AI on the vertical axis. Three categories of predictor are distinguished: individual predictions (blue circles), expert surveys (green squares), and forecasting community aggregates (brown triangles). The dashed red line represents Platt's Law — the empirical observation, named by Eliezer Yudkowsky after Charles Platt, that AGI forecasts consistently place strong AI approximately 30 years from the date of the forecast regardless of when the forecast is made. Points above the line represent predictions more pessimistic than Platt's Law; points below it represent predictions more optimistic. The compression of points in the lower-right corner — predictions made between 2022 and 2026 targeting dates as early as 2026–2032 — reflects the dramatic shortening of expert expectations following the release of large language models.

Expert survey HLMI medians over time

The following chart plots the aggregate median prediction for high-level machine intelligence (HLMI) — defined in each survey as the point at which unaided machines can accomplish every task better and more cheaply than human workers — from six major expert surveys conducted between 2011 and 2023. The blue line connects the 50% probability ("median") estimates from surveys of the broader AI research community; the brown diamond shows the Walsh 2017 AI expert group's 90% probability threshold (2109), which is not a median but a near-certainty estimate and sits on a different scale from the other points. The 2022→2023 drop of twelve years — from ~2059 to ~2047 — is the largest single-step shift in the series and coincides with the fourteen-month period following the release of ChatGPT and GPT-4.

Metaculus #3479 community median over time

The following chart shows the trajectory of the community median prediction on Metaculus question #3479 — "When will the first weakly general AI system be devised, tested, and publicly announced?" — from January 2020 to June 2024. The question defines weakly general AI using four operational criteria: reliably passing an adversarial Turing test, scoring at or above the 75th percentile on a standard SAT mathematics section, achieving 90% accuracy on the Winograd Schema Challenge, and demonstrating general robotic capabilities. The median fell from approximately 2070 in January 2020 to approximately 2031 by June 2024 — a decline of nearly four decades in four and a half years. The sharpest single inflection occurs between January 2022 and July 2022, when the median dropped approximately 18 years in six months, coinciding with a wave of AI capability demonstrations including the public release of instruction-following large language models. The subsequent decline from mid-2022 onward is more gradual, reflecting continued but less dramatic updating. The trajectory is widely cited as one of the most concrete quantitative records of how the broader forecasting community revised its expectations in response to AI progress, though the question's specific definition — particularly its robotics requirement — means the forecast may not be directly comparable to predictions about purely cognitive AI milestones.

Compute doubling times by era

The following chart displays the training compute doubling time — the number of months required for the amount of compute used in the largest AI training runs to double — across three distinct eras identified by Epoch AI's 2022 analysis of over 120 notable machine learning models. The pre-deep learning era (before 2012) saw a doubling time of approximately 24 months, broadly consistent with Moore's law. The deep learning era (2012–2015), initiated by the AlexNet result at ImageNet, accelerated to a ~6-month doubling time — four times faster than the prior era. The large-scale era (2015–2022), marked by models such as AlphaGo and GPT-3 requiring 10–100 times more compute than the deep learning trend, settled at a ~10-month doubling time. The dashed red reference line at 3.4 months shows the figure originally reported in OpenAI's influential 2018 "AI and Compute" post — a figure that sat to the left of even the fastest era's bar. Epoch AI's analysis attributes the discrepancy to OpenAI's smaller dataset and its conflation of the large-scale model regime with the broader trend. The corrected figures have since replaced the 2018 estimate as the standard empirical reference in AI timeline forecasting.

Meta information on the timeline

How the timeline was built

The initial version of the timeline was written by Sebastian.

Funding information for this timeline is available.

Feedback and comments

Feedback for the timeline can be provided at the following places:

• FIXME

What the timeline is still missing

• Progress in artificial intelligence ✔ (assessed: Wikipedia benchmark milestone article, no primary timeline predictions not already covered, no rows to mine)
• Sebastian: https://medium.com/accelerated-intelligence/top-ai-experts-predict-artificial-superintelligence-in-3-5-years-now-what-fbfe00204c7f ✔ (assessed: Medium roundup article by Michael Simmons, no original predictions, all underlying predictions already covered)
• Sebastian: https://techcrunch.com/2024/03/19/agi-and-hallucinations/?fbclid=IwAR2_TayMg3U_HrIk21tXokSeTiOAkvnF5pU8KjkS9b_AHoVcQVOpxTnCH6g ✔
• probably go back to people like Turing✔, IJ Good✔, or whoever who first came up with ideas around AGI; what did they say about when it would be created?
• dartmouth conference and initial optimistic ideas on when AGI would be created✔
• would be cool to include the various "standalone"/"self-contained" AI timeline methods and when they were first used https://wiki.issarice.com/wiki/Category:AI_timelines_arguments ✔
• https://openai.com/blog/ai-and-compute/ ✔
• https://www.gwern.net/Scaling-hypothesis ✔
• what are the various AI benchmarks and how do progress on those benchmarks look like? ✔
• the stuff in this blog post (especially the stuff in the table) https://www.cold-takes.com/where-ai-forecasting-stands-today/ ✔
• this thread might be notable https://www.greaterwrong.com/posts/hQysqfSEzciRazx8k/forecasting-thread-ai-timelines ✔
• look through posts here https://www.greaterwrong.com/tag/ai-timelines to see if any seem important ✔
• similarly, check https://forum.effectivealtruism.org/topics/ai-forecasting ✔ (assessed: tag index page, all notable posts already covered)
• maybe look at https://www.greaterwrong.com/s/B9Qc8ifidAtDpsuu8 ✔ (assessed: pedagogical explainer sequence, no original predictions, no row)
• controversy like https://www.lesswrong.com/posts/KnQs55tjxWopCzKsk/the-ai-timelines-scam ✔ (assessed: sociological critique of AI funding incentives by jessicata, no original timeline prediction, no row)
• https://www.lesswrong.com/posts/ax695frGJEzGxFBK4/biology-inspired-agi-timelines-the-trick-that-never-works ✔
• look at metaculus threads and how predictions have changed over time e.g. [1]✔, [2]✔ (assessed: essay contest, no primary prediction, no row), [3]✔ (assessed: covered by #3479 row, no separate row needed)
• something something gary marcus ✔
• https://intelligence.org/2017/10/13/fire-alarm/ ✔
• probably include some famous people's views like Bill Gates https://lukemuehlhauser.com/bill-gates-on-ai-timelines/ ✔
• https://en.wikipedia.org/wiki/Ray_Kurzweil#Predictions ✔
• might be able to copy some rows from Timeline of artificial intelligence, Timeline of AI safety, Timeline of machine learning ✔ (assessed: all three cover institutional/technical history rather than timeline predictions; no rows to mine)
• the AI Impacts experts survey paper from 2016 https://arxiv.org/pdf/1705.08807.pdf ✔
• look through posts at https://aiimpacts.org/category/ai-timelines/ ✔ (assessed: all notable items already covered; Gruetzemacher 2019 survey borderline, no row added)
• https://thenextweb.com/news/what-is-the-technological-singularity#.tnw_4ol0E0Vr (assessed: explainer article, no original prediction, no row) ✔
• https://voicebot.ai/2023/08/16/generative-ai-revenue-will-reach-1-3t-in-2032-chart/ (assessed: market revenue forecast, out of scope, no row) ✔
• https://www.knowyourmobile.com/ai-intelligence/ai-2033-a-glimpse-into-the-future-of-artificial-intelligence/ (assessed: consumer tech speculation, no original prediction, no row) ✔
• Daniel Kokotajlo / AI 2027 scenario (ai-2027.com) ✔
• Vernor Vinge 1993 "The Coming Technological Singularity" ✔
• Ray Kurzweil 2005 The Singularity Is Near ✔
• Stuart Armstrong and Kaj Sotala 2012 "How We're Predicting AI — or Failing To" ✔
• Eliezer Yudkowsky 2017 "There's No Fire Alarm for AGI" ✔
• Epoch AI 2022 "Compute Trends Across Three Eras of Machine Learning" ✔
• Geoffrey Hinton and Yoshua Bengio 2023 timeline revisions ✔
• Stuart Russell — assessed, no specific quantitative timeline prediction found, no row ✔

See also

AI timelines and forecasting

• Timeline of AI safety — comprehensive institutional history of the AI safety field, covering the organisations and research programmes whose predictions appear in this timeline
• Timeline of artificial intelligence — broad history of AI as a field, covering the technical milestones that inform timeline reasoning

AI labs and systems

• Timeline of OpenAI — history of the lab behind GPT-3, GPT-4, and ChatGPT, whose researchers (Amodei, Altman, Hernandez) appear in multiple rows
• Timeline of Anthropic — history of the lab behind Claude, whose institutional predictions (Amodei, Greenblatt) appear in this timeline
• Timeline of ChatGPT — detailed history of the model whose release triggered the most dramatic expert timeline revision in the survey record
• Timeline of Google Gemini — history of Google's frontier model series, whose CEO (Hassabis) appears in this timeline
• Timeline of xAI — history of Elon Musk's AI company, whose founder appears twice in this timeline
• Timeline of AlphaGo — history of the Go-playing system whose 2016 victory over Lee Sedol constituted a major timeline-updating event for many researchers
• Timeline of Machine Intelligence Research Institute — history of MIRI, the organisation behind Yudkowsky's forecasting methodology critiques

Technical foundations

• Timeline of machine learning — history of ML research, covering the AlexNet result and other benchmark milestones that ground timeline predictions
• Timeline of transformers — history of the transformer architecture underlying the large language models that triggered expert timeline revisions from 2022 onward

Policy and ethics

• Timeline of AI policy — history of AI governance and regulation, whose trajectory is shaped by the short-timeline predictions documented here
• Timeline of AI ethics violations — history of AI harms, providing context for the safety-motivated predictions in this timeline