AI Timelines
AI timelines refer to forecasts about when artificial intelligence systems will achieve transformative capabilities. Timeline estimates vary widely across methodologies and forecasters, with median predictions ranging from the late 2020s to beyond 2100. Since 2022, median estimates across multiple forecasting frameworks have shortened substantially, though the causes and implications of this shift remain contested.
Definitions and Operationalizations
Transformative AI
Coefficient Giving defines transformative AI (TAI) as "AI powerful enough to bring us into a new, qualitatively different future" comparable to the agricultural or industrial revolutions.1 This definition focuses on economic and societal impact rather than specific technical capabilities.
High-Level Machine Intelligence
Expert surveys typically use "High-Level Machine Intelligence" (HLMI), defined as "when unaided machines can accomplish every task better and more cheaply than human workers."2 This operationalization allows researchers to elicit comparable forecasts across different survey waves.
AGI and Other Definitions
The term "Artificial General Intelligence" (AGI) lacks a single agreed-upon definition. Google DeepMind researchers proposed defining AGI "in terms of capabilities, rather than processes" and suggested a levels-based framework distinguishing between performance (Narrow to Superhuman) and generality (Narrow to General).3 Different forecasting efforts use different operationalizations, making direct comparison of "AGI timelines" difficult without examining specific definitions.
AI R&D Automation
A more recent operationalization focuses specifically on the fraction of AI research and development tasks that AI systems can perform autonomously. This milestone is distinct from HLMI or TAI in that it targets the self-referential capacity of AI systems to accelerate their own development—rather than general economic productivity or human-parity performance across all tasks.
METR researcher Thomas Kwa's 2026 model defines AI R&D automation as a logistic function in log compute, capturing the fraction of AI R&D labor that AI systems can replace at a given level of effective compute.4 The logistic form is described as conservative: automation asymptotically approaches but never reaches 100%, preventing superexponential growth unless compute inputs are themselves superexponential.5 By this operationalization, approximately 95% AI R&D automation corresponds to AI systems achieving a 14-year "time horizon" on METR's coding task suite—a benchmark measuring the length of tasks an AI can complete with 80% reliability.6
The AI Futures Project's more detailed model (AIFM) uses a related milestone it calls "Automated Coder" (AC) and "Superhuman AI Researcher" (SAR) as intermediate steps, and Davidson and Eth (2025) analyze "AI Systems for AI R&D Automation" (ASARA) as a potential trigger for a feedback loop in AI capability growth.7 These related operationalizations share the core intuition that AI-driven AI research represents a qualitatively distinct threshold with non-linear implications for subsequent progress.
Major Forecasting Methodologies
Biological Anchors
Ajeya Cotra's 2020 biological anchors framework estimates TAI timelines by comparing required computational resources (measured in FLOP) to biological systems.8 The method involves:
- Estimating the effective compute required to train transformative models using various "anchors" (analogies to biological computation)
- Forecasting when sufficient compute will be affordable given historical cost trends
- Accounting for uncertainty through probability distributions over multiple anchors
Scott Alexander's analysis of the framework reported probability distributions of 10% chance of TAI by 2031, 50% by 2052, and approximately 80% by 2100.9
In a 2022 update, Cotra shortened her timeline estimates based on recent progress, shifting her median from ~2050 to ~2040, with updated distributions of 15% by 2030, 35% by 2036, 50% by 2040, and 60% by 2050.10
Critics of biological anchors note that the framework requires estimating numerous uncertain parameters, some of which critics characterize as arbitrarily chosen.11 Eliezer Yudkowsky specifically argued that biological analogies may not capture the relevant algorithmic structure of AI capability gains.12 Proponents respond that the framework's value lies in making assumptions explicit and quantifiable rather than hiding them in qualitative judgments.
Compute-Centric Models
Tom Davidson developed a compute-centric forecasting framework that models the relationship between computational resources, algorithmic progress, and economic growth. His model for Coefficient Giving estimates a median time of approximately 3 years from 20% to 100% automation of cognitive tasks, with superintelligence potentially arriving within a year after full automation.13
Davidson's modeling approach links training compute inputs to a "cognitive task automation fraction"—the share of economically valuable cognitive work that AI can perform at human-competitive cost. The framework treats algorithmic progress and hardware scaling as complements: each unit of compute provides larger capability gains when paired with algorithmic improvements. Key parameters include the rate of compute scaling, the rate at which algorithms improve effective compute, the shape of the automation-versus-compute curve, and the degree to which AI labor can substitute for human labor in research tasks. Davidson's 2023 report documents sensitivity analyses across these parameters, finding that median timelines are most sensitive to assumptions about how quickly AI systems can automate the research tasks needed to improve AI systems themselves.14
Davidson's earlier work on semi-informative priors used reference class forecasting to estimate AGI probability. His 2021 analysis suggested a central estimate of approximately 4% probability of AGI by 2036, with a preferred range of 1–10%.15
In 2025, Davidson and Daniel Eth at Forethought extended this line of work to examine whether AI R&D automation could trigger a "software intelligence explosion" (SIE)—a runaway feedback loop in which AI systems improve their own capabilities faster than human oversight can track.7 Their analysis considers whether AI can become substantially more capable through software improvements alone (better architectures, training methods, data, and scaffolding), without requiring corresponding hardware investments. They identify "complementarity"—how substitutable AI labor is for human labor in research tasks—as a key open parameter. If AI and human research labor are highly substitutable, the feedback loop could be rapid; if they are highly complementary (each requiring the other), the loop would be slower.
Scaling Laws as Foundational Inputs
Compute-centric models depend on empirical relationships between training compute, model size, data, and performance—commonly called "scaling laws." Kaplan et al. (2020) documented power-law relationships between these quantities for language models, finding that performance improved smoothly and predictably as compute scaled.16 Hoffmann et al. (2022), in the "Chinchilla" paper, identified that prior large models were undertraining relative to the amount of available data, and that for a given compute budget, optimal performance requires scaling model size and training tokens in roughly equal proportion.17 The Chinchilla result implied that GPT-3-scale models trained on 10× more data would outperform GPT-3 by a significant margin—a finding that influenced subsequent training decisions across the field.
Epoch AI analysis indicates that the introduction of the Chinchilla scaling laws alone accounted for the equivalent of 8 to 16 months of algorithmic progress, while the transformer architecture accounted for the equivalent of approximately two years of algorithmic progress.18 These results illustrate why scaling laws are not merely descriptive but have predictive value for timeline models: they provide a basis for estimating how much capability improvement should be expected from a given increase in compute, absent novel architectural advances.
A key uncertainty for timeline models is whether current scaling laws will continue to hold at higher compute levels, or whether they will exhibit phase transitions, diminishing returns, or qualitative changes. Epoch AI's analysis of frontier models suggests the 4–5× per year compute growth trend remained intact through 2024, but the period 2025–2030 introduces new constraints (power, manufacturing, data) that make extrapolation less certain.19
The AI Futures Model (AIFM)
The AI Futures Project (Eli Lifland, Brendan Halstead, Alex Kastner, and Daniel Kokotajlo) published the AI Futures Model in December 2025, a 33-parameter quantitative framework modeling AI capability growth and the transition through key milestones including Automated Coder (AC), Superhuman AI Researcher (SAR), and artificial superintelligence (ASI).20
The AIFM's all-things-considered distribution yields a 10th percentile of 2027.5, a median of 2032.5, and a 90th percentile of 2085 for its primary milestone.20 The model estimates approximately a 9% chance of rapid capability growth by the end of 2026, and approximately a 6% chance that no AGI-level system appears by 2050. The model does not account for hardware R&D automation or broad economic automation, which the authors note means it may underestimate the probability of sub-10-year outcomes.20
The AIFM uses task-level capability thresholds to separate the question of when AI systems could perform specific research sub-tasks from the question of when this would translate into overall AI R&D automation. This decomposition requires estimating approximately 33 parameters, including the fraction of AI R&D labor that requires human judgment, the rate at which AI systems can automate narrow versus broad research tasks, and the feedback coefficient linking AI capability to AI research speed. The authors calibrated parameters using expert elicitation, benchmark data, and prior quantitative forecasting work.
A January 2026 update from the AI Futures team clarified that their all-things-considered median for the Automated Coder milestone is approximately January 2035—about 1.5 years later than raw model output—and that the team was never confident an AGI milestone would occur in 2027 specifically.21
The AI 2027 Report
The AI 2027 report, published April 3, 2025 by the AI Futures Project (Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean), presents a narrative scenario analysis of AI development through 2027 and beyond.22 The scenario is distinct from the AIFM quantitative model: whereas the AIFM produces probability distributions over milestone dates, the AI 2027 report presents a single illustrative scenario intended to explore what rapid-development pathways might look like concretely. The scenario was informed by trend extrapolations, wargames, expert feedback, experience at OpenAI, and prior forecasting work.
The report uses METR's time horizon data as a core input: the length of coding tasks AI systems can complete with 80% reliability doubled approximately every 7 months from 2019 to 2024, and approximately every 4 months from 2024 onward. Extrapolating this trend, the report's scenario posits that AI systems could succeed with 80% reliability on tasks taking skilled humans multiple years by approximately March 2027. The report introduces an "R&D progress multiplier" concept measuring how many months of AI progress are achieved in one month when AI systems assist research.
The scenario forecasts that 2027-era coding agents would be capable enough to substantially accelerate AI R&D itself, potentially triggering an intelligence explosion reaching superintelligence by early 2028. Kokotajlo has stated a personal probability of approximately 0.7 for catastrophic AI outcomes; Alexander has stated approximately 0.2. These figures represent the authors' personal safety probability estimates and are not outputs of the AI 2027 forecasting model itself.22
A January 2026 update clarified that the team's all-things-considered median for the Automated Coder milestone is approximately January 2035—roughly 1.5 years later than model output—and that the April 2025 scenario was never intended as a confident point prediction.21
A Simpler AI Timelines Model (Kwa 2026)
Thomas Kwa, an alignment researcher at METR, published a simplified timelines model in February 2026 as a METR research note, cross-posted to LessWrong and the Alignment Forum.4 The model was constructed in approximately 15 hours and uses 8 parameters, compared to the AIFM's 33 parameters.
Core structure and methodology. Rather than estimating when AI systems will be capable enough to automate specific AI R&D sub-tasks (as the AIFM does), Kwa's model maps directly from effective compute to an automation fraction using a logistic function. This eliminates the need to estimate task-level capability thresholds, reducing the number of modeling assumptions. Compute growth is parameterized at 2.6× per year through 2029, based on Epoch AI estimates, then decelerating from 2× to 1.25× per year between 2030 and 2058. Algorithmic efficiency grows separately via a parameter v (automation velocity), estimated at 0.876 logits per year. Research progress is modeled as a Cobb-Douglas function of labor and compute, with human labor treated as the bottleneck at early automation levels.
Key outputs. The model's median prediction is greater than 99% automation of AI R&D by late 2032. Most simulations produce a 1,000× to 10,000,000× increase in AI efficiency and 300×–3,000× research output by 2035. The model estimates that automation was approximately 20% at the start of 2025 and approximately 37.5% (1.6× uplift) by the start of 2026.5
Conservative assumptions. The model adopts several stated constraints: no superexponential time horizon growth, no full (100%) automation, and adherence to Amdahl's Law (no substitutability between labor and compute). These choices are described as preventing artificially aggressive outputs.
Relationship to other models. Unlike biological anchors (which anchor to neural/evolutionary compute analogies) or semi-informative priors (which use outside-view historical base rates), Kwa's model is an inside-view compute extrapolation anchored to METR's empirically measured time horizon metric. The author notes that "any reasonable timelines model will predict superhuman AI researchers before 2036, unless AI progress hits a wall or is deliberately slowed," with the main exceptions being scenarios in which compute and human labor growth slow in ~2030 with no architectural breakthroughs.6
Limitations acknowledged. Kwa explicitly states he does not put high weight on the exact predicted timelines given limited time spent on parameter values, and that additional models would be needed to characterize long-timeline cases. The model does not separate research taste from software engineering as distinct skills, so it addresses timelines but not takeoff dynamics specifically.5
Expert Surveys
The 2022 Expert Survey on Progress in AI contacted 4,271 researchers who published at NeurIPS or ICML 2021, receiving 738 responses (17% response rate).23 The aggregate forecast estimated 36.6 years until HLMI from the survey date.
A 2023 survey of 2,778 AI researchers found:24
- 50% probability of machines outperforming humans in every task by 2047
- 10% probability by 2027
- Median estimate 13 years shorter than the 2022 survey
- Between 38% and 51% of respondents assigned at least 10% probability to advanced AI leading to outcomes "as bad as human extinction"
The survey found variation across questions and methodologies, with estimates differing depending on how questions were framed.25
Community Forecasting
Metaculus aggregates predictions from thousands of forecasters. Metaculus forecasters' current estimate is August 2033 for the announcement of the first general AI system.26 A separate Metaculus question on transformative AI shows a current median estimate of December 2041, drawn from 168 forecasters.27 28
An aggregated AGI timeline dashboard drawing from multiple Metaculus questions, as well as Manifold Markets and Kalshi, tracks combined forecasts for AGI arrival across several operationalizations of the question.29 Manifold Markets uses different resolution criteria than Metaculus, which observers attribute to differing operationalizations of AGI rather than genuine disagreement about underlying probabilities.30
Superforecasters from the XPT (Expert and Public Forecasting Tournament) have produced longer timelines than Metaculus community estimates, with median estimates closer to 2047, illustrating that forecasting methodology and population selection produce meaningfully different outputs from the same underlying evidence base.28
Different Metaculus questions about AGI produce different median estimates partly because they use different resolution criteria—some require autonomous AI passing specific benchmarks, others require a human judge's assessment, and others specify economic output thresholds. Comparing estimates across these questions requires attention to which specific milestone each question resolves on.
Prediction market estimates face additional interpretive considerations: thin liquidity in AI-related markets can mean that individual large trades move prices substantially, and resolution criteria ambiguity may create incentives to bet on narrower or more certain operationalizations rather than the underlying capability question of interest. These structural factors mean prediction market estimates may not straightforwardly reflect participant beliefs about capability timelines.
AI Lab Leadership Statements (2025–2026)
Public statements from major AI lab leadership have become a widely cited input to timeline discourse, though they carry interpretive complications including potential motivated reasoning, definitional ambiguity around "AGI," and PR considerations.
OpenAI. Sam Altman published a blog post stating: "We are now confident we know how to build AGI as we have traditionally understood it," and indicated OpenAI was beginning to focus on superintelligence.31 In a Bloomberg interview, Altman stated he believes "AGI will probably get developed during [Trump's] term." He has also stated OpenAI expects models to function as AI "research interns" within 2026 and as fully independent researchers by approximately 2028, with a stated goal of "an automated AI researcher by March 2028."32 Altman has noted that "AGI has become a very sloppy term."
Anthropic. In March 2025 recommendations to the White House Office of Science and Technology Policy, Anthropic stated it expects "powerful AI systems will emerge in late 2026 or early 2027."33 Dario Amodei's essay "Machines of Loving Grace" offered a more hedged version: "I think it could come as early as 2026, though there are also ways it could take much longer." Amodei has described AI systems matching the collective intelligence of "a country of geniuses" within a few years as a target scenario.33 At Davos in January, Amodei stated he sees a form of AI "better than almost all humans at almost all tasks" emerging in the "next two or three years."34 Anthropic has been described as the only major AI company with publicly stated official AGI timelines.33
Google DeepMind. Speaking at a briefing in DeepMind's London offices, Demis Hassabis stated that he thinks AGI — which is as smart or smarter than humans — will start to emerge in the next five or ten years.35
These statements should be interpreted alongside the fact that AI lab leaders have commercial and reputational incentives that may affect public timeline statements in either direction, and that their operationalizations of "AGI" differ from those used in formal forecasting frameworks.
Current Forecast Summary
As of early 2026, major forecasting sources report:
| Source | Milestone | Median Estimate | Range/Distribution | Method |
|---|---|---|---|---|
| Cotra (2022) | TAI | 2040 | 15% by 2030, 50% by 2040, 60% by 2050 | Biological anchors |
| Expert Survey (2023) | HLMI | 2047 | 10% by 2027, 50% by 2047 | Researcher survey |
| Metaculus Aggregate (2026) | AGI | 2031 | 80% CI: 2027–2045 | Community forecasting |
| Davidson (2021) | AGI | >2036 | 1–10% by 2036 | Semi-informative priors |
| AI Futures Model (Dec 2025) | AGI/AC | 2032.5 | 10th pct: 2027.5; 90th pct: 2085 | Compute + task modeling |
| Kwa/METR (Feb 2026) | 99% AI R&D automation | ≈2032 | Most runs: 2028–2036 | 8-parameter logistic model |
| AI 2027 (Apr 2025, updated Jan 2026) | Automated Coder | ≈2035 (adjusted) | Scenario-based | Trend extrapolation + expert judgment |
These estimates are not directly comparable due to different milestone definitions, methodologies, and operationalizations of "transformative AI," "AGI," or "AI R&D automation."
Key Uncertainty Factors
Algorithmic Progress and Compute Availability
Epoch AI tracks training compute for frontier models as a key driver of capability growth. GPT-4 (released March 2023) was the first model trained at the 10²⁵ FLOP scale, at approximately 2×10²⁵ FLOP; Gemini Ultra was estimated at approximately 5×10²⁵ FLOP; and the first model estimated above 10²⁶ FLOP was Grok-3 from xAI, released February 2025.36 As of June 2025, over 30 publicly announced models from 12 developers have been identified at or above the 10²⁵ FLOP threshold.36
Epoch AI's analysis of training compute growth found a rate of approximately 4.1× per year (90% CI: 3.7× to 4.6×) between 2010 and May 2024.19 About two-thirds of performance improvements in language models over this period came from increases in model scale, with algorithmic progress accounting for the remainder.19
Algorithmic efficiency improvements compound with compute scaling. Epoch AI's study of 231 language models found that the compute needed to achieve a given level of performance has halved roughly every 8 months (95% CI: 5–14 months), outpacing Moore's Law hardware doubling time of approximately 2 years.18 A Shapley value decomposition found that 60–95% of language model performance gains came from increased compute and training data, with algorithmic innovations accounting for 5–40%.18 A separate MIT study using Epoch AI benchmark data estimated algorithmic efficiency gains at approximately 3× per year after controlling for hardware improvements, with three independent estimates (Epoch AI 8-month halving, Amodei ~4×/year, and the MIT analysis ~3×/year) reaching similar conclusions with halving times of 8, 6, and 7.5 months respectively.37
Epoch AI research notes that hardware and algorithms function as complements rather than substitutes: the most transformative algorithmic shifts have been compute-dependent, yielding 10–50× compute-equivalent gains or more but only when sufficient hardware is available.38 This complementarity has implications for timeline models: scenarios that assume algorithmic progress could substitute for slower compute scaling may underestimate the coupling between the two.
Training compute for state-of-the-art models increased by a factor of 4–5× per year from 2010 to 2024.39 Epoch AI's 2024 analysis of scaling feasibility suggests training runs of 2×10²⁹ FLOP will likely be feasible by 2030, representing a roughly 10,000× scale-up relative to models at time of writing.40 This projection was derived from Monte Carlo simulations incorporating four constraints: power availability, chip manufacturing capacity, data scarcity, and the "latency wall." The constraint assessed most likely to bind first is power, followed by chip manufacturing capacity.40
GPU production through 2030 is projected to expand 30–100% per year, with median projections suggesting manufacturing capacity sufficient for approximately 100 million H100-equivalent GPUs—enough to support a 9×10²⁹ FLOP training run.40 Power constraints at a single campus level (1–5 GW by 2030) would support 10²⁸ to 3×10²⁹ FLOP training runs; using a distributed US network (2–45 GW), the feasible range extends to 2×10²⁸ to 2×10³⁰ FLOP.40
A "latency wall" at approximately 2–3×10³¹ FLOP may be reached within approximately 3 years of the 2024 paper, where data movement between chips dominates arithmetic computation time, making efficient use of hardware impossible without alternative network topologies, reduced communication latencies, or aggressive batch size scaling.41
An important interaction exists between the latency wall and algorithmic efficiency trends: if algorithmic efficiency continues improving at 3× per year, the effective capability achievable within a given hardware budget increases correspondingly, partially offsetting hardware constraints. However, Epoch AI's analysis emphasizes that compute-dependent algorithmic advances require sufficient hardware to realize their gains—meaning that hardware bottlenecks may suppress capability growth even if algorithmic research continues.38
Data Constraints
Epoch AI estimates the total effective stock of human-generated public text data at approximately 300 trillion tokens (90% CI: 100 trillion to 1,000 trillion tokens), including only sufficiently high-quality data and accounting for the possibility of multi-epoch training.42 The amount of text data fed into AI language models has been growing approximately 2.5× per year.42 If current trends continue, language models will fully utilize this stock somewhere between 2026 and 2032—or earlier if models are intensively overtrained.42 This range was updated from an earlier 2022 Epoch AI estimate that had predicted exhaustion by 2024, based on methodological refinements and updated understanding of data quality thresholds.43
The data constraint creates a potential transition from a compute-constrained to a data-constrained training regime. The 2022 Epoch AI analysis projected the median year of full data utilization at 2028, corresponding to dataset sizes approaching 4×10¹⁴ tokens and training compute around 5×10²⁸ FLOP.43
Epoch AI identifies three categories of innovations most relevant to overcoming the data bottleneck: synthetic data generation, learning from other modalities (video, code, scientific data), and data efficiency improvements (better algorithms that extract more capability per token).42 Synthetic data has been shown to improve capabilities in narrow domains where verification is straightforward—mathematics and coding in particular—but AI-generated synthetic data "has only been shown to reliably improve capabilities in relatively narrow domains like math and coding" as of the 2024 analysis.42 Extending synthetic data benefits to broader domains (open-ended reasoning, world knowledge, novel tasks) faces challenges including quality ceiling (synthetic data cannot reliably introduce knowledge absent from the base model), distribution shift (synthetic data may reinforce existing biases), and mode collapse risks in iterative self-training.40
This data constraint interacts with the compute scaling trajectory: Epoch AI's 2024 analysis of scaling feasibility found that data availability would allow training runs of 6×10²⁸ to 2×10³² FLOP by 2030, a wider range than power or manufacturing constraints—suggesting that data is a less binding constraint than power in the near term, but becomes relevant at the higher end of the scaling range.40
AI R&D Automation as a Potential Accelerant
A distinct uncertainty factor—underrepresented in earlier forecasting frameworks—is the feedback dynamic that arises when AI systems become capable of contributing to their own research and development. If AI systems automate a substantial fraction of AI R&D, the rate of capability improvement could become partially endogenous to the level of AI capability, potentially departing from the extrapolation of historical human-driven trends.
Davidson and Eth (2025) model this as a potential "software intelligence explosion": once AI systems can replace a sufficient fraction of AI researchers, each subsequent capability improvement enables further automation of research, compounding on itself.7 Kwa's model (2026) treats this dynamic via a logistic automation curve that asymptotically approaches but never reaches full automation, but notes that most simulations nonetheless produce very large increases in AI efficiency by 2035 even under these constraints.4
Whether and how quickly such feedback dynamics would manifest depends on empirical questions that remain unresolved: how substitutable AI labor is for human judgment in research (the "complementarity" or "rho" parameter), whether AI systems can generate genuinely novel architectural insights versus incremental engineering progress, and whether regulatory or institutional constraints would slow deployment of AI researchers before the feedback loop gains momentum.
Kwa's model explicitly does not model research taste and software engineering as separate capabilities, and thus does not characterize takeoff dynamics in detail—only timelines to the automation threshold.5 The AIFM team notes that their model similarly does not account for hardware R&D automation, meaning it may underestimate sub-10-year outcomes.20
Economic and Regulatory Factors
Timeline forecasts typically assume continued large-scale investment in AI development and minimal regulatory constraints. Changes in either factor could significantly affect timelines, though quantifying these effects remains difficult.
Methodological Debates
Inside View vs. Outside View
"Inside view" forecasting examines specific technical factors like compute trends, algorithmic progress, and scaling laws. "Outside view" forecasting uses reference classes from historical technology development.
Robin Hanson's outside view analysis estimated "at least a century" to human-level AI.44 His methodology, documented in a 2012 Overcoming Bias post, involved asking senior AI researchers at the UAI conference with approximately 20 years of field experience to estimate how much progress had been made toward human-level AI; they roughly agreed that approximately 5–10% of the distance had been covered, without noticeable acceleration.45 From this Hanson concluded that "an outside view calculation suggests we probably have at least a century to go, and maybe a great many centuries, at current rates of progress."45 In a 2019 AI Impacts interview, Hanson reaffirmed this estimate as his best guess for median timelines, substantially longer than the ~50-year numbers given by AI researchers in formal surveys.44
Hanson's outside view argument centers on two related claims. First, he argues that AI progress, like most innovation, is not "lumpy"—meaning it proceeds through many small incremental steps rather than a few large breakthroughs. He cites citation statistics as evidence: the distribution of citation counts in AI is similar to other academic fields, following a power law where most papers receive few citations and the few highly cited papers constitute a small fraction of total impact. If AI is not lumpy, progress should be observable and gradual rather than surprising and discontinuous.44 Second, Hanson argues that the history of growth transitions—from humans to farming to industry—provides at most three data points of rapid worldwide growth acceleration, making any individual claim that AI will produce such a transition draw on a thin evidence base.46
Hanson's post-ChatGPT (2023) position maintained these views largely unchanged. He argued that even if large language models achieve human-level performance on next-token prediction, this is insufficient for rapid capability explosion because it would also require "high competence in design and testing alternative system architectures." He assigns less than 1% probability to something like a catastrophic AI takeover, reasoning that we already have "superintelligences" in the form of large firms that "only improve slowly and incrementally," and that internal coordination problems would similarly hamper any AI superintelligence.47
Critics of Hanson's outside view note that his reference class (general technological progress) may not be applicable to AI, which has unusual scaling properties, growing compute investment, and no close historical precedent. The same evidence that supports long timelines from an outside view—that past AI booms did not produce human-level AI—could also be interpreted as informative about what past AI techniques could not do, rather than about what current scaling-based techniques cannot do.
Epoch AI's literature review noted that inside-view models tend to predict shorter timelines than outside-view approaches, with heavier tails and more probability mass on long-horizon outcomes in outside-view frameworks.48 The review identified Cotra's biological anchors as one of the more detailed inside-view models and Davidson's semi-informative priors as an example of an outside-view model, as of 2023.
Inside-view models carry a risk of overconfidence due to the salience of recent visible progress, selection effects among researchers who build such models, and availability bias toward recently salient breakthroughs. Proponents of inside-view models respond that outside-view reference classes may not be applicable to a technology with unusual scaling properties and no close historical precedent—and that Hanson's 2012 reference class (expert judgments of remaining distance) may have shifted substantially given the capabilities demonstrated since 2020.
Kwa's 2026 model adds a data point to this debate: it is an inside-view model that achieves its headline result with only 8 parameters, suggesting that simplicity and inside-view reasoning are not mutually exclusive. However, the author explicitly notes that simpler models may miss dynamics that matter for the tails of the distribution, and that additional models are needed to characterize long-timeline cases.5
Parsimony vs. Comprehensiveness in Model Design
The contrast between Kwa's 8-parameter model and the AIFM's 33-parameter model illustrates a general methodological tradeoff. More detailed models can represent more factors—hardware R&D automation, task-level capability thresholds, substitutability of labor and compute—but are correspondingly more sensitive to their modeling assumptions, more prone to overfitting, and harder to interpret or critique. Simpler models are more transparent and easier to audit, but may omit dynamics that are quantitatively important.
Kwa explicitly frames simplicity as a feature rather than a limitation: the model does not need to estimate when AI systems will automate specific AI R&D sub-tasks, instead mapping directly from compute to automation fraction. However, this also means the model cannot capture the detailed structure of takeoff—only the timing of the automation threshold.4 The AIFM team, by contrast, explicitly models the gap between task-level benchmarks and the real-world capabilities needed for AI to function as a researcher, at the cost of greater parameter uncertainty.20
Both approaches face the common challenge that all timeline models are extrapolations beyond their calibration data and cannot be validated before the milestones they predict are reached or missed.
Continuous vs. Discontinuous Progress
Paul Christiano argued in his 2018 post "Takeoff Speeds" for "slow takeoff," predicting that AI development would appear as continuous economic acceleration distributed across the broader economy rather than a breakthrough concentrated within a small group.49 His specific operationalization: "There will be a complete 4-year interval in which world output doubles, before the first 1-year interval in which world output doubles"—with analogous claims holding for 8-year before 2-year doublings, and so on.49 He chose this operationalization "because it seemed to be something clean where (a) there is disagreement, and (b) it clearly has strategic implications."49
Christiano's core argument for slow takeoff rests on the observation that "it's easier to build a crappier version of something" and that "a crappier AGI would have almost as big an impact"—a claim he characterizes as having a strong historical track record across technologies.49 He extends this to recursive self-improvement: "Before there is AI that is great at self-improvement, there will be AI that is mediocre at self-improvement."49 This predicts that before any single lab develops a decisive strategic advantage from highly capable AI, a world in which AI is already transforming the economy at a rapid pace will have emerged, making the strategic situation more distributed.
The slow takeoff scenario has distinct safety implications from fast takeoff. Christiano argues that in a slow takeoff world, "incredibly powerful AI will emerge in a world where crazy stuff is already happening (and probably everyone is already freaking out)"—potentially enabling policy responses and safety interventions that are unavailable in fast-takeoff scenarios. However, slow takeoff also requires coordinating across many actors who all have capable AI, rather than a simpler situation where one lab holds a decisive advantage. The alignment problem in slow takeoff is therefore described as harder in some respects: solutions must be competitive and scalable across multiple actors rather than needing to solve alignment only for the first highly capable system.49
Eliezer Yudkowsky argued for fast takeoff scenarios involving rapid recursive self-improvement, with "fast" meaning timescales of weeks or hours rather than years or decades. In the 2008 Hanson-Yudkowsky debate, Yudkowsky argued that a cheap, initially low-ability AI could, without warning, over a short time (e.g., a weekend), become powerful enough to take over the world through self-improvement.50 51
Hanson's response challenged the premise of discontinuous self-improvement by arguing that historical growth has seen only three events (human emergence, agriculture, industry) in which overall growth rates increased suddenly and sustainably—making any individual factor, including AI self-improvement, unlikely to generate such an event unless it is "unusually likely to have such an effect."46 He further argued that a superintelligent AI would face internal coordination problems analogous to those of large firms, limiting its capabilities—pointing to the observation that large organizations already constitute forms of "collective superintelligence" that "only improve slowly and incrementally."47
A 2023 retrospective analysis on LessWrong examining specific predictions made by both Hanson and Yudkowsky during the FOOM debate found that when examining pre-AGI developments, Hanson's predictions performed somewhat better than Yudkowsky's at the object level, while Yudkowsky's high-level claims about short timelines and general AI architectures showed more predictive accuracy.52
Tom Davidson's 2023 compute-centric framework estimated approximately 3 years from 20% to 100% cognitive task automation, representing a position between very fast (months) and very slow (decades) scenarios.14
The AI R&D automation framing adds a dimension to this debate: even under "slow takeoff" assumptions about overall economic automation, the specific feedback loop in AI research could be faster if AI labor is more substitutable for human labor in research than in other domains.
Limitations and Criticisms
Forecasting Track Record
Historical AI predictions have varied in accuracy. The AI field experienced multiple periods of reduced activity and funding in the 1970s and 1980s following forecasts that proved overly optimistic about near-term capabilities.53 Some long-term forecasts using quantitative trend extrapolation, such as those by Hans Moravec and Ray Kurzweil, tracked closer to observed compute and capability trends than purely intuitive predictions.54
This historical record provides evidence for multiple interpretations. Failures like the AI winters indicate that AI researchers have historically held overconfident near-term predictions, while cases where quantitative trend extrapolations proved accurate suggest that inside-view methods can sometimes produce reliable predictions. The 2022 expert survey estimated that AI systems would not be able to write simple Python code until approximately 2027, but the 2023 survey revised this estimate to 2025 after observing faster-than-expected progress.55
The 80,000 Hours analysis of expert surveys notes that AI researcher estimates have historically been described as overly pessimistic in the long run, even while specific near-term predictions often overshot.28 Interpreting this pattern requires caution: it could reflect genuine underestimation of progress, herding toward consensus, or survivorship bias in which forecasters who update most rapidly receive more attention.
Methodological Critiques
Critics of timeline forecasting methodologies argue that:
- Biological anchors involve numerous uncertain parameters, which critics characterize as arbitrarily chosen11
- Models lack empirical validation of core assumptions56
- Reference class forecasting faces challenges with technologies that have no close historical analogues57
- Expert predictions may be influenced by availability bias and anchoring effects
Eliezer Yudkowsky critiqued biological anchors for relying on "straightforward brain computation comparisons" that may not capture relevant algorithmic insights.12
Outside-view approaches face analogous critiques: reference class selection is itself a judgment call rather than an objective procedure, and the class of "transformative technologies" is small enough that base-rate estimates carry high uncertainty. Hanson's reference class of expert estimates of remaining distance to human-level AI in 2012 may not remain informative in light of capability changes since 2020 that were not foreseen by the experts surveyed.
Unknown Unknowns
All forecasting methodologies face challenges from potential discontinuities, breakthrough insights, or constraint violations not captured in trend extrapolations. Robin Hanson noted that scenarios involving "very lumpy tech advances, broadly-improving techs, [and] powerful secret techs" are each historically rare, suggesting that combinations of these factors (as in some short-timeline scenarios) may have low base-rate probability.58
The AI R&D automation pathway introduces an additional consideration: if AI systems began automating their own development at scale, it is unclear whether existing evaluation frameworks would detect this transition in real time, or whether capability improvements would outpace the development of new benchmarks.
Implications for AI Safety
Timeline estimates influence AI safety research prioritization, though the relationship between timelines and urgency is not straightforward.
The common framing is that shorter timelines imply greater urgency to develop safety techniques and governance frameworks, because there is less time available for theoretical research and gradual implementation. On this view, shorter timelines favor investment in approaches that can be deployed quickly, such as scalable oversight, interpretability tools that work with current architectures, or governance interventions targeting near-term systems.
However, some researchers contend that timeline uncertainty itself—regardless of median length—is the key factor, because the distribution of outcomes matters more than point estimates. Others argue that longer timelines with more capable systems could also imply urgency, if the primary concern is that increasingly powerful systems are deployed before adequate safety measures exist. On this view, the relationship between timeline length and urgency is non-monotonic.
The AI R&D automation scenario introduces a specific safety consideration: if AI systems begin contributing substantially to their own development, the pace of capability gain may accelerate faster than safety research can track, compressing the time available for iterative testing regardless of the prior timeline. Davidson and Eth (2025) note this as a reason to treat the ASARA threshold as particularly salient for safety planning.7
80,000 Hours analysis noted that expert timeline estimates shortened significantly between 2022 and 2023 surveys, with implications for career planning and research prioritization in AI safety.59
Historical Timeline Evolution
Timeline predictions have shifted over time as capabilities improved:
- 2016: Coefficient Giving estimated "at least 10% probability" of transformative AI within 20 years (by 2036)60
- 2020: Cotra's biological anchors: 50% probability by 2052
- 2021: Davidson's semi-informative priors: approximately 4% probability of AGI by 2036
- 2022: Cotra updated median: 50% by 2040; expert survey (HLMI): median approximately 2059
- 2023: Expert survey median shortened by 13 years to approximately 2047; Metaculus community median approximately 2031
- Early 2025: At the start of 2025, the Metaculus community forecast for strong AGI stood at July 2031. Following OpenAI's release of reasoning models o1 and o3, median estimates shortened further across multiple forecasting platforms. The AI 2027 scenario also received widespread popular coverage, projecting fully automated AI research and development by 2027 leading to a recursive self-improvement loop.61
- Late 2025: Forecasts extended again across some platforms — the Metaculus flagship question median moved out to November 2033 — as the pace of progress in the second half of 2025 was interpreted as slower relative to the post-o3 peak.61
This pattern reflects a general movement toward shorter estimates since 2020, though with substantial variation across methodologies and oscillation in community forecasts in response to specific capability announcements. Whether this reflects Bayesian updating on evidence, herding, or other dynamics is contested among forecasters and researchers.
Footnotes
-
Some Background on Our Views Regarding Advanced Artificial Intelligence — Holden Karnofsky, "Some Background on Our Views Regarding Advanced Artificial Intelligence," Open Philanthropy, 2016. ↩
-
When Will AI Exceed Human Performance? Evidence from AI Experts — Katja Grace et al., "When Will AI Exceed Human Performance? Evidence from AI Experts," arXiv:1705.08807, 2017. ↩
-
Levels of AGI: Operationalizing Progress on the Path to AGI — Meredith Ringel Morris et al., "Levels of AGI: Operationalizing Progress on the Path to AGI," arXiv:2311.02462, 2023. ↩
-
Research note: A simpler AI timelines model predicts 99% AI R&D automation in ~2032 — Thomas Kwa, "Research note: A simpler AI timelines model predicts 99% AI R&D automation in ~2032," METR, February 10, 2026. ↩ ↩2 ↩3 ↩4
-
Research note: A simpler AI timelines model predicts 99% AI R&D automation in ~2032 — Thomas Kwa, "Research note: A simpler AI timelines model predicts 99% AI R&D automation in ~2032," LessWrong cross-post, February 10, 2026. ↩ ↩2 ↩3 ↩4 ↩5
-
Research note: A simpler AI timelines model predicts 99% AI R&D automation in ~2032 — Thomas Kwa, "Research note: A simpler AI timelines model predicts 99% AI R&D automation in ~2032," Alignment Forum cross-post, February 10, 2026. ↩ ↩2
-
Will AI R&D Automation Cause a Software Intelligence Explosion? — Tom Davidson and Daniel Eth, "Will AI R&D Automation Cause a Software Intelligence Explosion?," Forethought, March 26, 2025. ↩ ↩2 ↩3 ↩4
-
Draft report on AI timelines — Ajeya Cotra, "Draft report on AI timelines," LessWrong, 2020. ↩
-
Biological Anchors: A Trick That Might Or Might Not Work — Scott Alexander, "Biological Anchors: A Trick That Might Or Might Not Work," Astral Codex Ten, 2021. ↩
-
Two-year update on my personal AI timelines — Ajeya Cotra, "Two-year update on my personal AI timelines," Alignment Forum, 2022. ↩
-
Biology-Inspired AGI Timelines: The Trick That Never Works — Eliezer Yudkowsky, "Biology-Inspired AGI Timelines: The Trick That Never Works," LessWrong, 2021. ↩ ↩2
-
Biology-Inspired AGI Timelines: The Trick That Never Works — Eliezer Yudkowsky, "Biology-Inspired AGI Timelines: The Trick That Never Works," LessWrong, 2021. ↩ ↩2
-
Davidson On Takeoff Speeds — Scott Alexander, "Davidson On Takeoff Speeds," Astral Codex Ten, June 2023. ↩
-
What a Compute-Centric Framework Says About Takeoff Speeds — Tom Davidson, "What a Compute-Centric Framework Says About Takeoff Speeds," Open Philanthropy, June 27, 2023. ↩ ↩2
-
Semi-Informative Priors Over AI Timelines — Tom Davidson, "Semi-Informative Priors Over AI Timelines," Open Philanthropy, 2021. ↩
-
Scaling Laws for Neural Language Models — Jared Kaplan et al., "Scaling Laws for Neural Language Models," arXiv:2001.08361, OpenAI, January 2020. ↩
-
Training Compute-Optimal Large Language Models — Jordan Hoffmann et al., "Training Compute-Optimal Large Language Models," arXiv:2203.15556, DeepMind ("Chinchilla"), March 2022. ↩
-
Algorithmic progress in language models — Anson Ho et al., "Algorithmic progress in language models," Epoch AI, March 12, 2024. ↩ ↩2 ↩3
-
Training compute of frontier AI models grows by 4-5x per year — Jaime Sevilla and Edu Roldán, "Training compute of frontier AI models grows by 4-5x per year," Epoch AI, May 28, 2024. ↩ ↩2 ↩3
-
AI Futures Timelines and Takeoff Model: Dec 2025 Update — Eli Lifland, Brendan Halstead, Alex Kastner, and Daniel Kokotajlo, "AI Futures Timelines and Takeoff Model: Dec 2025 Update," LessWrong, December 30, 2025. ↩ ↩2 ↩3 ↩4 ↩5
-
Clarifying How Our AI Timelines Forecasts Have Changed Since AI 2027 — AI Futures Project, "Clarifying How Our AI Timelines Forecasts Have Changed Since AI 2027," AI Futures Blog, January 2026. ↩ ↩2
-
AI 2027 — Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean, "AI 2027," AI Futures Project, April 3, 2025. ↩ ↩2
-
2022 Expert Survey on Progress in AI — Katja Grace and Ben Weinstein-Raun, "2022 Expert Survey on Progress in AI," AI Impacts, 2022. ↩
-
Thousands of AI Authors on the Future of AI — Katja Grace et al., "Thousands of AI Authors on the Future of AI," arXiv:2401.02843, January 2024. ↩
-
Shrinking AGI timelines: a review of expert forecasts — "Shrinking AGI timelines: a review of expert forecasts," 80,000 Hours, March 2025. ↩
-
When Will the First General AI Be Announced? — "When Will the First General AI Be Announced?," Metaculus, accessed February 2026. ↩
-
Transformative AI Date — "Transformative AI Date," Metaculus, accessed February 2026. ↩
-
Transformative AI Date - Metaculus — "Transformative AI Date - Metaculus," Metaculus, accessed February 2026. ↩ ↩2 ↩3
-
AGI Timelines Dashboard — "AGI Timelines Dashboard," Goodheart Labs, accessed January 5, 2026. ↩
-
Forecasting AGI: Insights from Prediction Markets and Metaculus — "Forecasting AGI: Insights from Prediction Markets and Metaculus," Forecasting AI Futures (Substack), 2025–2026. ↩
-
How OpenAI's Sam Altman Is Thinking About AGI and Superintelligence in 2025 — "How OpenAI's Sam Altman Is Thinking About AGI and Superintelligence in 2025," TIME Magazine, 2025. ↩
-
OpenAI Roadmap: AI Research Interns by 2026, Full-Blown AGI Researchers by 2028 — "OpenAI Roadmap: AI Research Interns by 2026, Full-Blown AGI Researchers by 2028," TechRadar, 2025–2026. ↩
-
What's up with Anthropic predicting AGI by early 2027? — "What's up with Anthropic predicting AGI by early 2027?," Redwood Research Blog, November 3, 2025. ↩ ↩2 ↩3
-
Anthropic CEO Dario Amodei warns: AI will match 'country of geniuses' by 2026 — "Anthropic CEO Dario Amodei warns: AI will match 'country of geniuses' by 2026," VentureBeat, December 24, 2025. ↩
-
AI That Can Match Humans at Any Task Will Be Here in 5–10 Years, Google DeepMind CEO Says — "AI That Can Match Humans at Any Task Will Be Here in 5–10 Years, Google DeepMind CEO Says," CNBC, March 17, 2025. ↩
-
Over 30 AI models have been trained at the scale of GPT-4 — Robi Rahman et al., "Over 30 AI models have been trained at the scale of GPT-4," Epoch AI, January 30, 2025, last updated June 6, 2025. ↩ ↩2
-
The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference — "The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference," Gundlach et al., MIT CSAIL/FutureTech, NeurIPS 2025 Workshop on Evaluating the Evolving LLM Lifecycle, November 2025. ↩
-
How Fast Can Algorithms Advance Capabilities? — Epoch AI, "How Fast Can Algorithms Advance Capabilities?," Epoch AI Substack, May 2025. ↩ ↩2
-
Training compute of frontier AI models grows by 4-5x per year — Jaime Sevilla and Edu Roldán, "Training compute of frontier AI models grows by 4-5x per year," Epoch AI, May 28, 2024. ↩
-
Can AI scaling continue through 2030? — Jaime Sevilla et al., "Can AI scaling continue through 2030?," Epoch AI, August 20, 2024. ↩ ↩2 ↩3 ↩4 ↩5 ↩6
-
Data movement bottlenecks to large-scale model training: Scaling past 1e28 FLOP — Ege Erdil, "Data movement bottlenecks to large-scale model training: Scaling past 1e28 FLOP," Epoch AI, November 2, 2024. ↩
-
Will we run out of data? Limits of LLM scaling based on human-generated data — Tamay Besiroglu et al., "Will we run out of data? Limits of LLM scaling based on human-generated data," Epoch AI, June 6, 2024. ↩ ↩2 ↩3 ↩4 ↩5
-
Will we run out of ML data? Projecting dataset size trends — Pablo Villalobos et al., "Will we run out of ML data? Projecting dataset size trends," Epoch AI, arXiv:2211.04325, November 10, 2022. ↩ ↩2
-
Conversation with Robin Hanson — Asya Bergal and Robert Long, "Conversation with Robin Hanson," AI Impacts, November 20, 2019. ↩ ↩2 ↩3
-
AI Progress Estimate — Robin Hanson, "AI Progress Estimate," Overcoming Bias, August 27, 2012. ↩ ↩2
-
Foom Debate, Again — Robin Hanson, "Foom Debate, Again," Overcoming Bias, February 18, 2013. ↩ ↩2
-
AI Risk, Again — Robin Hanson, "AI Risk, Again," Overcoming Bias, March 2, 2023. ↩ ↩2
-
Literature review of transformative artificial intelligence timelines — "Literature review of transformative artificial intelligence timelines," Epoch AI, January 2023. ↩
-
Takeoff speeds — Paul Christiano, "Takeoff speeds," The Sideways View, February 24, 2018. ↩ ↩2 ↩3 ↩4 ↩5 ↩6
-
Citation rc-2ba9 (data unavailable — rebuild with wiki-server access) ↩
-
The Hanson-Yudkowsky AI-Foom Debate — "The Hanson-Yudkowsky AI-Foom Debate," Machine Intelligence Research Institute, 2008. ↩
-
Yudkowsky vs Hanson on FOOM: Whose Predictions Were Better? — "Yudkowsky vs Hanson on FOOM: Whose Predictions Were Better?," LessWrong, 2023. ↩
-
Citation rc-7eff (data unavailable — rebuild with wiki-server access) ↩
-
AI Futures Model: Dec 2025 Update — "AI Futures Model: Dec 2025 Update," AI Futures, December 2025. ↩
-
Shrinking AGI timelines: a review of expert forecasts — "Shrinking AGI timelines: a review of expert forecasts," 80,000 Hours, March 2025. ↩
-
The methodological limits of the AI 2027 forecast — "The methodological limits of the AI 2027 forecast," Future Scouting, 2025. ↩
-
A deep critique of AI 2027's bad timeline models — Titotal, "A deep critique of AI 2027's bad timeline models," LessWrong, 2025. ↩
-
What Are Reasonable AI Fears? — Robin Hanson, "What Are Reasonable AI Fears?," Effective Altruism Forum, 2023. ↩
-
What the hell happened with AGI timelines in 2025? — "What the hell happened with AGI timelines in 2025?," 80,000 Hours podcast, February 10, 2026. ↩
-
Some Background on Our Views Regarding Advanced Artificial Intelligence — Holden Karnofsky, "Some Background on Our Views Regarding Advanced Artificial Intelligence," Open Philanthropy, 2016. ↩
-
What the Hell Happened with AGI Timelines in 2025? — "What the Hell Happened with AGI Timelines in 2025?," 80,000 Hours podcast, February 10, 2026. ↩ ↩2
References
“Third, we deliberately make two conservative assumptions: No full automation: as AIs get more capable, they never automate 100% of AI R&D work, just approach it.”
The claim states that 95% AI R&D automation corresponds to AI systems achieving a 14-year "time horizon" on METR's coding task suite. The source does not mention this specific percentage or time horizon. The claim mentions that the benchmark measures the length of tasks an AI can complete with 80% reliability. The source does not mention this specific reliability percentage.
“As far as I’m aware, Anthropic is the only AI company with official AGI timelines 1 : they expect AGI by early 2027.”
The claim states that Anthropic made recommendations to the White House Office of Science and Technology Policy in March 2025, but the source says the recommendations were "to the OSTP for the AI action plan". The claim states that Dario Amodei's essay is titled "Machines of Loving Grace", but the source titles it as 'Machines of Loving Grace'. The claim states that Amodei described AI systems matching the collective intelligence of "a country of geniuses" within a few years as a target scenario, but the source says that Anthropic often describes the capability level of powerful AI systems as a “country of geniuses in a datacenter”. The claim states that at Davos in January, Amodei stated he sees a form of AI "better than almost all humans at almost all tasks" emerging in the "next two or three years.", but the source does not mention Davos or the specific timeframe of "next two or three years." The claim states that Anthropic has been described as the only major AI company with publicly stated official AGI timelines, but the source states that "As far as I’m aware, Anthropic is the only AI company with official AGI timelines".
“Very lumpy tech advances, techs that broadly improve abilities, and powerful techs that are long kept secret within one project are each quite rare. Making techs that meet all three criteria even more rare.”
“Isolating out open models to control for competition effects and dividing by hardware price declines, we estimate that algorithmic efficiency progress is around 3 × 3\times per year.”
“OpenAI CEO Sam Altman recently published a post on his personal blog reflecting on AI progress and his predictions for how the technology will impact humanity’s future. “We are now confident we know how to build AGI [artificial general intelligence] as we have traditionally understood it,”Altman wrote. He added that OpenAI, the company behind ChatGPT, is beginning to turn its attention to superintelligence.”
The claim that Altman indicated OpenAI expects models to function as AI "research interns" within 2026 and as fully independent researchers by approximately 2028, with a stated goal of "an automated AI researcher by March 2028" is not supported by the provided source. The source does not mention these specific timelines or goals related to AI researchers. The source does not mention Altman publishing a blog post stating: "We are now confident we know how to build AGI as we have traditionally understood it," and indicated OpenAI was beginning to focus on superintelligence. The source states that Altman wrote this in a post, but does not specify that it was a blog post.
8Research note: A simpler AI timelines model predicts 99% AI R&D automation in ~2032alignmentforum.org·Blog post▸
“Specifically, he predicts it will take about three years to go from AIs that can do 20% of all human jobs (weighted by economic value) to AIs that can do 100%, with significantly superhuman AIs within a year after that.”
“Even 80,000 Hours, where I work, made a video about the AI 2027 story that got a stupid number of views.”
The claim mentions timeline estimates shortened between 2022 and 2023, but the source discusses changes in timeline estimates between early and late 2025. The claim mentions implications for AI safety, but the source does not explicitly mention this.
“In early 2025, after OpenAI put out the first-ever reasoning models — o1 and o3 — short timelines to transformative artificial general intelligence swept the AI world.”
The claim states that the Metaculus community forecast for strong AGI stood at July 2031 at the start of 2025, but the source says it was in late 2024 and early 2025 that short timelines to transformative artificial general intelligence swept the AI world. The claim mentions OpenAI's release of reasoning models o1 and o3, but the source says it was in late 2024 and early 2025 that short timelines to transformative artificial general intelligence swept the AI world after OpenAI put out the first ever set of reasoning models o1 and o3. The claim mentions the AI 2027 scenario received widespread popular coverage, but the source says that there were huge levels of popular coverage of the AGI scenario known as AI 2027.
“This approach has gone through a few iterations: Hans Moravec, Ray Kurzweil, and Shane Legg pioneered this method, predicting based on the amount of operations per second that the human brain does.”
The claim mentions periods of reduced activity and funding in the 1970s and 1980s, which is not mentioned in the source. The claim states that some long-term forecasts tracked closer to observed compute and capability trends than purely intuitive predictions, which is not explicitly stated in the source, but is implied.
“A detailed technical analysis by researcher Titotal on LessWrong reveals that the model's core parameters lack any empirical validation, with critical variables such as the "superexponential growth rate" being set arbitrarily without uncertainty analysis.”
“Data movement bottlenecks limit LLM scaling beyond 2e28 FLOP, with a "latency wall" at 2e31 FLOP. We may hit these in ~3 years.”
“For AI progress to continue into the 2030s, either new sources of data or less data-hungry techniques must be developed.”
“AGI Timelines Dashboard AGI Timeline Forecasts Median year predictions from multiple forecasting sources. Data sources: Metaculus , Manifold , Kalshi”
“Jan 2026: Jan 2035 (~2035.0) . For Automated Coder (AC), my all-things-considered median is about 1.5 years later than the model’s output.”
“You’ve written based on AI practitioners estimates of how much progress they’ve been making that an outside view calculation suggests we probably have at least a century to go, if maybe a great many centuries at the current rates of progress in AI.”
The claim mentions 'biological anchors' estimating median timelines in the 2040s–2050s, but this is not mentioned in the source. The claim contrasts Hanson's view with 'inside-view models like biological anchors', but the source does not explicitly frame Hanson's view as contrasting with inside-view models.
“Training compute has been increasing by a factor of 4–5× every year since 2010, according to Epoch AI estimates [1] .”
“This trick didn't work when Moravec tried it, it's not going to work while Ray Kurzweil is trying it, and it's not going to work when you try it either.”
“Sam Altman says within the next year, OpenAI expects its models to work like AI “research interns”, and by 2028, they could function as fully independent researchers.”
The source does not contain the quote: "We are now confident we know how to build AGI as we have traditionally understood it," and indicated OpenAI was beginning to focus on superintelligence. The source does not contain the quote: "AGI will probably get developed during [Trump's] term." The source states that OpenAI expects models to function as AI "research interns" within the next year, not 2026.
“A first-trial probability of 1/300 combined with a 1956 regime start-time and 1 virtual success yields pr(AGI by 2036) = 4%.”
The claim states 'approximately 4% probability of AGI by 2036', but the source states 'pr(AGI by 2036) = 4%' for a specific scenario, not as a central estimate. The claim states 'with a preferred range of 1-10%', but the source states 'My central estimate is about 8%, but other parameter choices I find plausible yield results anywhere from 1% to 18%.'
“The AI Futures Project (Eli Lifland, Brendan Halstead, Alex Kastner, and Daniel Kokotajlo) published the AI Futures Model in December 2025.”
“With this in mind, we define transformative AI as follows: Definition #1: Roughly and conceptually, transformative AI is AI that precipitates a transition comparable to (or more significant than) the agricultural or industrial revolution.”
“With that said, I am comfortable saying that I think there is a nontrivial likelihood (at least 10% with moderate robustness, and at least 1% with high robustness) of transformative AI within the next 20 years.”
“In late 2008, economist Robin Hanson and AI theorist Eliezer Yudkowsky conducted an online debate about the future of artificial intelligence, and in particular about whether generally intelligent AIs will be able to improve their own capabilities very quickly (a.k.a. “foom”).”
“AI 2027 April 3rd 2025 PDF Listen Watch Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean We predict that the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution. We wrote a scenario that represents our best guess about what that might look like. 1 It’s informed by trend extrapolations, wargames, expert feedback, experience at OpenAI, and previous forecasting successes.”
“The report finds a 10% chance of “transformative AI” by 2031, a 50% chance by 2052, and an almost 80% chance by 2100.”
“The key assumption behind this model is that if we train a neural net or other ML model that uses about as much computation as a human brain, that will likely result in transformative AI (TAI) (defined as AI that has an impact comparable to that of the industrial revolution). In other words, we _anchor_ our estimate of the ML model’s inference computation to that of the human brain.”
The source mentions Ajeya Cotra's report and discusses its contents, but it does not explicitly state that the framework estimates TAI timelines by comparing required computational resources (measured in FLOP) to biological systems. While the report does use biological systems as 'anchors' for estimating computational requirements, the claim oversimplifies the methodology.
32Research note: A simpler AI timelines model predicts 99% AI R&D automation in ~2032lesswrong.com·Blog post▸
“At this slope, we get to 95% AI R&D automation (or in your model, ~100% coding automation) when we hit a time horizon of 14 years, which is less than your median of 125 years and should give timelines before 2045 or so.”
The claim states the model is from 2026, but the source was published in 2026 and refers to the model in the present tense. The claim states 95% AI R&D automation corresponds to AI systems achieving a 14-year "time horizon" on METR's coding task suite, but the source states that 95% AI R&D automation corresponds to a time horizon of 14 years, which is less than the median of 125 years, and should give timelines before 2045 or so.
“Transformative AI Date 63 comments 169 forecasters When will we have transformative AI? Current estimate May 2040”
The current median estimate is May 2040, not December 2041. The number of forecasters is 169, not 168.
“In the history of artificial intelligence (AI), an AI winter is a period of reduced funding and interest in AI research.”
The claim mentions that some long-term forecasts tracked closer to observed compute and capability trends, but the source does not explicitly state this. It only mentions that some forecasts proved overly optimistic. The source does not explicitly mention that Hans Moravec and Ray Kurzweil used quantitative trend extrapolation.
“Speaking at a briefing in DeepMind's London offices on Monday, Demis Hassabis said that he thinks artificial general intelligence (AGI) — which is as smart or smarter than humans — will start to emerge in the next five or 10 years.”
“In the largest survey of its kind, 2,778 researchers who had published in top-tier artificial intelligence (AI) venues gave predictions on the pace of AI progress and the nature and impacts of advanced AI systems”