AGI timelines are forecasts about when AI systems will cross a threshold corresponding to some operational definition of artificial general intelligence — variously specified as sufficient economic substitutability for human knowledge work, sufficient ability to perform AI research, or sufficient breadth across cognitive domains. Forecasts range from aggressive (around 2027) to measured (decades), and sources disagree not only on the year but on what "AGI" means, how to extrapolate capability trends, and how much weight to place on qualitative anchors such as GPT-4 relative to quantitative benchmark progress. Confidence on this page is rated contested because sources hold materially different positions.
Dimensions of disagreement
Disagreement on timelines decomposes into several distinct questions:
- Definition. Whether AGI is defined as a benchmark threshold (the Sparks of AGI framing), as substitutability for human workers (Amodei), or as the capacity to perform automated AI research (Aschenbrenner).
- Extrapolation method. Whether to project from orders-of-magnitude (OOMs) of compute, from benchmark trendlines, from economic-diffusion rates, or from historical base rates.
- Weight on "unhobbling." Aschenbrenner treats agency and tool use as large capability multipliers; skeptics view them as speculative.
- Feedback loops. Whether automated AI research yields a recursive-self-improvement intelligence explosion.
Aggressive positions (around 2027)
Leopold Aschenbrenner's *Situational Awareness* (2024) argues for AGI by approximately 2027, driven by compounding OOMs of physical compute, algorithmic efficiency, and "unhobbling" (tool use, agency, chain-of-thought), with superintelligence shortly after. Aschenbrenner also projects trillion-dollar training clusters; his forecast of a single training cluster with at least $1 trillion cumulative capex (one company, one location or tightly-coupled multi-site) carries a resolution horizon of end-2028, with the Anthropic–SpaceX Colossus 1 and OpenAI–Broadcom $18B plus Stargate efforts cited as fractional progress.
Dario Amodei describes a "country of geniuses in a datacenter" as achievable in one to two years (see Compressed 21st Century). In Machines of Loving Grace (October 2024), Amodei argues that most major scientific advances could arrive within 5–10 years of "powerful AI"; a mid-window check on that thesis is set for 2030, keyed to whether at least three Nobel-class scientific advances (physics, chemistry, or medicine) attribute a load-bearing role to a frontier AI system.
AI 2027 is a parallel detailed forecast in which agency and automated AI research cross thresholds by 2027. The companion analysis "Grading AI 2027's 2025 predictions" tests these aggressive forecasts against actual outcomes, and an accuracy track record is emerging.
Jack Clark's Import AI 455 (May 4, 2026) assigns a 60%+ probability to an AI system that can autonomously train its own successor by end-2028, and 30% by end-2027. Clark's estimate is built on first-party Anthropic capability trajectories: SWE-Bench saturating to 93.9%; the METR 50%-reliable time horizon moving from roughly 30 seconds to roughly 12 hours over four years; and an internal CPU-LLM-optimization task improving 2.9× to 52× from Opus 4 to Mythos in 11 months. Clark's framing is anchored to insider data from a frontier lab that also publishes safety positions, distinct from Aschenbrenner's external-forecaster framing. The newsletter prompted Azeem Azhar's EV #573 (May 9, 2026) to reframe the forecast into three testable revealed-preference patterns: research-multiplier hiring, capex over-rotation on physical bottlenecks, and tolerance for "ugly near-term cash burn." (Sources: jack-clark.net; exponentialview.co)
A resolution criterion associated with the AGI-by-2027 forecast, issued June 2024 with a January 1, 2028 horizon, is whether a frontier lab demonstrates an automated AI researcher operating at the level of a top-quartile Anthropic, OpenAI, or DeepMind senior researcher across at least three publishable research projects in a calendar quarter. The criterion associated with Clark's recursive-AI-R&D forecast, with a January 1, 2029 horizon, is whether a frontier lab discloses that at least 50% of one model generation's training-pipeline improvements (loss-function design, data curation, architecture search, RL reward shaping) were proposed and implemented by a previous model in the same family without a human in the loop.
Institutional commitments around 2030
In a Google I/O keynote on May 22, 2026, described as a 24-announcement, 35-minute platform keynote, Sundar Pichai named 2030 as Google's AGI target year — the first explicit CEO-level institutional AGI deadline from a frontier lab. The statement operationalizes Hassabis's April 30, 2026 "AGI by 2030" video-interview position into a CEO-level institutional commitment. Anthropic, OpenAI, and xAI have not declared comparable institutional dates, leaving Google as both the established big-tech defender and the lab most willing to put a numeric date on AGI publicly. (Source: the-ai-corner.com)
At Google I/O (May 19–20, 2026), Demis Hassabis told Axios cofounder Mike Allen "this year, with the agentic systems that we're all seeing and using, I think we can start feeling" the singularity. He said AGI could arrive as soon as 2030, called himself a "cautious optimist," and predicted AI would be 100× as impactful as the Industrial Revolution. The day before, on a separate I/O stage, he said "when we look back at this time, I think we will realize that we were standing in the foothills of the singularity," drawing audible gasps. The Hassabis and Pichai statements together form the May 2026 cluster pinning Google's CEO-level AGI deadline. (Source: businessinsider.com)
On May 23, 2026, OpenAI posted a "Researcher, Recursive Self-Improvement Preparedness" listing on its Preparedness team ($295K–$445K) to "support preparations for recursive self-improvement," citing defense against data-poisoning attacks and tracking "automation of technical staff." The role follows Sam Altman's stated goal of running an automated AI research intern by September 2026 and a "true automated AI researcher by March of 2028." It is the first lab-side operational hire explicitly named for RSI preparedness, setting a 21-month timeline (current to March 2028) for the lab to be ready for what its CEO has stated as a corporate target. (Source: businessinsider.com)
GPT-4 as anchor
Sparks of AGI (Bubeck et al., 2023) argues that GPT-4 itself already shows "early sparks" of AGI, reframing the question from "when" to "what counts."
Measured and skeptical positions
AI as Normal Technology (Kapoor and Narayanan) treats AI as a normal general-purpose technology, with economic diffusion and institutional absorption taking decades. The associated forecast, issued April 2025, holds that the AI-as-normal-technology framing will be dominant by 2030; its resolution criterion (horizon December 31, 2030) is whether mainstream media plus at least one major OECD government frame AI as a "general-purpose technology" with diffusion-paced impact (months-to-years per sector) rather than as a discontinuous capability event. This forecast contradicts the aggressive and 2030-institutional positions above, and its resolution will favor one camp.
Broad Timelines emphasizes uncertainty, holding that AGI may be 5, 15, or 50 years out and that confident point forecasts are epistemically weak.
A different kind of contribution to the timelines literature is Google DeepMind's June 2026 From AGI to ASI, which brackets the question of when AGI arrives and asks instead what follows it. Its relevance to timelines is structural: three of the four pathways it maps from AGI to superintelligence — paradigm shifts, recursive improvement, and multi-agent collectives — have no historical base on which forecasting models can be fitted, leaving only the scaling pathway empirically constrained. The report treats the impact of its six named bottlenecks as open research questions rather than resolved parameters, and argues that a series of transformative changes may describe the trajectory better than a single step change at the AGI threshold.
Relevance to policy
The timeline question bears on several policy debates. Aggressive timelines are used to argue for compute thresholds, pre-deployment safety cases, and national-security-level investment in the near term, citing instruments such as the Anthropic RSP v3.1 and the OpenAI Preparedness Framework. Aschenbrenner's projection of trillion-dollar training clusters, if realized, would reshape energy, chip, and national-security policy (see OpenAI's Industrial Policy and America's AI Action Plan). Short timelines also reinforce the AI race framing that figures prominently in US AI policy, including its US–China dimension.
Relationships
- supports: AI Race Dynamics, Compute Governance
- depends-on: Scaling Laws, AI Software Progress
- related: Compressed 21st Century — Amodei's parallel short-timelines thesis
- related: Recursive Self-Improvement
- contradicts: AI as Normal Technology
- related: Situational Awareness, Sparks of AGI, AI 2027, Broad Timelines