AGI Timelines and Forecasting

AGI Forecasting & Timeline Methodology is the analytical discipline of predicting when Artificial General Intelligence (AGI) will be developed. It employs a combination of expert elicitation (surveys), trend extrapolation (scaling laws), biological anchor models, and economic modeling to estimate probability distributions for the arrival of human-level machine intelligence. Such timelines are argued to bear on the urgency of AI safety and governance efforts (see Grace et al., 2024, for a survey of researcher opinion). All such estimates are highly uncertain.

The timeline to AGI is argued by many to be an important strategic variable. If AGI is decades away, there is more time to adapt institutions and study alignment empirically; if it is a few years away, the transition period would be compressed. Whether AGI is achievable at all on any given timescale is itself uncertain, and the associated risk claims are contested.

Forecasting methodology is significant because it shifts AGI from science fiction into quantitative risk management. By decomposing the problem into measurable inputs—compute availability, algorithmic efficiency, and data constraints—analysts can state explicit but highly uncertain probability distributions, which makes assumptions inspectable even though it does not make the estimates reliable.

Historically, AI pioneers consistently underestimated the difficulty of AGI. Following the 1956 Dartmouth Workshop, Herbert Simon is widely quoted as predicting in 1965 that machines would within twenty years be capable of doing any work a man can do [VERIFICATION NEEDED: exact wording and source, see AGI-001]. Some historians attribute the AI winters partly to such unmet expectations, alongside funding decisions and critical reports; the causal account is contested [VERIFICATION NEEDED: cross-check with the history volume].

Systematic surveys of researcher opinion on AI timelines date at least to the early 2010s [VERIFICATION NEEDED: cite Müller and Bostrom (2013) and Grace et al. (2018); not checked here]. Later methods add compute-based approaches such as biological anchors (Cotra, 2020) and empirical scaling laws (Kaplan et al., 2020), alongside, not instead of, expert elicitation. In successive AI Impacts surveys, the aggregate 50% date for high-level machine intelligence (HLMI) moved about 13 years earlier between 2022 and 2023, to 2047 (Grace et al., 2024). Some forecasting communities and AI-lab leaders have stated shorter timelines [VERIFICATION NEEDED: cite specific sources, dates and definitions].

#4.1 Biological Anchors Framework

The Biological Anchors framework (Cotra, 2020) models the compute required to train an AGI by anchoring it to the computational capacity of the human brain. Cotra (2020) is a draft report about transformative AI (TAI), not AGI or HLMI as such. It combines six biological anchor hypotheses, giving 90% weight to the anchors and 10% to the possibility that all of them underestimate the compute required; the resulting distribution spans more than 20 orders of magnitude of compute (as summarized in secondary sources; the report itself was not re-read here).

1. Estimate brain compute: Anchors estimate the human brain's computation, expressed as a rate in FLOP/s (floating-point operations per second). [VERIFICATION NEEDED: the report's specific brain-compute value and range; the figure of $10^{15}$ FLOP/s in an earlier draft is unconfirmed.]
2. Distinct anchors for total compute: The evolution anchor (compute used by evolution) and the lifetime anchor (compute used by a brain from birth to adulthood) differ by many orders of magnitude and must not be merged. Total compute is a quantity in FLOP, not FLOP/s. [VERIFICATION NEEDED: the report's values for each anchor; the earlier single figure of $10^{24}$ FLOP is unconfirmed.]
3. Hardware extrapolation: Using projections of hardware price-performance and spending, analysts estimate when training with the required compute becomes financially feasible. This framework is hypothetical (tag: Hypothetical); the framework has been critiqued and the author may have updated her views since 2020 [VERIFICATION NEEDED: later updates and critiques].

#4.2 Compute-Centric Scaling Extrapolation

This method relies on empirical scaling laws. If a benchmark (e.g., SWE-bench or MMLU) requires a specific cross-entropy loss to be solved at a human level, and loss scales as a power-law with training compute, forecasters can derive an explicit but highly uncertain compute estimate for reaching a given loss. The step from a loss value to human parity on a broad suite of tasks is not an established empirical relation; scaling laws describe loss on held-out text, not capabilities.

#4.3 Expert Elicitation

Surveys ask thousands of published machine learning researchers to provide subjective probability distributions. These surveys typically frame the question operationally. In Grace et al. (2024), HLMI refers to machines that outperform humans at every task; the exact survey wording is not quoted here [VERIFICATION NEEDED: quote from the survey appendix]. Results are sensitive to framing, and respondents may not be representative of all experts.

Forecasting models generally fall into two epistemological categories:

• Inside View (Mechanistic): Builds a model from the ground up, estimating the exact FLOPs, data, and algorithms required.
• Outside View (Reference Class): Looks at historical trends in similar technological breakthroughs (e.g., nuclear fission, human flight) to estimate how long a field typically takes to transition from theoretical proofs to practical engineering.

• Epoch AI Models: Villalobos et al. (2024) project that the stock of public human-generated text could be fully used between 2026 and 2032 under current trends. They say nothing directly about AGI timelines; any implication for timelines (for example, a data constraint slowing scaling unless synthetic data proves viable) is this article's inference. [VERIFICATION NEEDED: claims about Epoch AI database maintenance are uncited and have been removed.]
• Grace et al. (2024) Survey: In the 2023 AI Impacts survey (n=2,778 researchers who published at top AI venues), the aggregate forecast gave a 10% chance of machines outperforming humans on every task by 2027 and a 50% chance by 2047, assuming science continues undisrupted; the 50% date was 13 years earlier than in a similar survey a year before. The same survey gave a 10% chance of full automation of all human occupations by 2037 and 50% by 2116. Results are expert opinion from a Tier 3 arXiv preprint, with large differences across question framings and different samples across years (Grace et al., 2024).

#Limitations

• The "Unknown Unknowns": Compute models assume that current architectures can reach the target by scaling. If a fundamental algorithmic breakthrough is required, compute extrapolations would not apply.
• Expert Bias: Forecasts by AI researchers may be affected by biases such as recency effects, and historical forecasts have been wrong in both directions [VERIFICATION NEEDED: cite empirical evidence on the accuracy of AI forecasters].

#Unresolved Scientific Questions

1. Hardware-Algorithm Substitutability: How quickly does algorithmic improvement reduce the compute required for AGI? Faster algorithmic efficiency gains would, other things equal, lower the compute needed and bring forecasts earlier; no specific doubling time is asserted here [VERIFICATION NEEDED: cite measured rates before restoring one].
2. The Definition of the Target: AGI forecasting is notoriously difficult because the definition of AGI keeps shifting (the "moving goalpost" phenomenon).

Estimates are widely dispersed. The expert survey cited here (Grace et al., 2024) gave an aggregate 10% chance of HLMI by 2027 and 50% by 2047; the 2027 figure is the 10% point, not a median. Aggregate dates in that survey have shortened in recent years, but they remain contested and framing-sensitive. Community forecasting platforms such as Metaculus have at times given earlier median dates [VERIFICATION NEEDED: cite question ID, snapshot date and resolution criteria]; no source is cited for the previous "2025 to 2070" interval, which has been removed. For current figures see the Living Reference collection FCST-REG-001.

Unattributed predictions about forecaster behavior and an "early 2030s" date were removed because no source supports them. One registered forecast follows; confidence is an editorial judgement.

• Foundational Concepts:
  • AGI-001 Artificial General Intelligence: The target being forecasted.
  • ECON-COMP-001 The Economics of Compute & Scaling: The physical infrastructure constraints dictating the timelines.
• Advanced Implications & Future Trajectories:
  • ASI-EXP-001 The Intelligence Explosion Hypothesis: What happens immediately after the timeline resolves.