Elapsed time, in years, from AI being able to mostly perform frontier AI research-engineer work to AI surpassing humans at all cognitive tasks
- Measures
- The interval, in years, between (a) the first time an AI system can do most of the work of a research engineer at a leading frontier AI lab (roughly the "superhuman coder" / automated-research-engineer milestone) and (b) the first time an AI system outperforms the best humans at essentially all cognitive tasks (artificial superintelligence in the AI 2027 sense).
- Population
- The actual trajectory of frontier AI development worldwide.
- Conditions
- Conditional on milestone (a) being reached. Milestones are dated by capability of the best internal system, whether or not publicly deployed. If (b) never follows (a), the value is undefined/infinite and that probability is reported as lying outside every value.
- Unit
- years
Most likely about 3 years; 90% interval 0.4 to 25 years, with a 15% chance that AI that mostly does frontier research-engineer work never goes on to surpass humans at essentially all cognitive tasks (plateau, deliberate halt, or catastrophe), so the interval is undefined.
Stated as a distribution, the precision the evidence supports. right-skewed, roughly log-normal
Assessment
The graph's estimate, all things considered: the figures and arguments it rests on, and how they were combined.
This quantity measures takeoff speed: how long after AI can do most of a frontier lab research engineer's work it comes to beat the best humans at essentially all cognitive tasks. No one has observed it, so every figure comes from a model or from judgment. The best-known figure comes from the AI 2027 scenario (April 2025). It puts the median at about one year, says five times slower or faster is plausible, and assumes training compute stays flat. Its authors' revised AI Futures Model (December 2025) moved full coding automation 3 to 5 years later than AI 2027 had it. That change came mainly from more cautious modelling of how much AI speeds up research. Davidson and Houlden (Forethought, 2025) model a software intelligence explosion and estimate about a 60% chance that it compresses more than three years of AI progress into under one year, and about a 20% chance that it compresses more than ten years into one.
On the other side, there are strong reasons to expect a slower takeoff. Experiments need compute, and compute could become the bottleneck. Research taste, physical-world tasks and the last hard-to-automate cognitive domains may lag well behind coding. Each further gain may also get harder to find. All of these would stretch the interval to many years. Overall the graph puts the median at about 3 years, with a 90% interval of roughly 0.4 to 25 years. That median is slower than AI 2027's headline figure, because the slower-takeoff considerations and the forecasters' own later revisions get real weight. The estimate also reserves about a 15% share for the second milestone never following the first.
The work: which figures, how they were combined, and the arithmetic
Inputs: (1) AI 2027 states a median of about 1 year and a plausible range of 0.2 to 5 years (ai-2027.com/; takeoff forecast at ai-2027.com/research/takeoff-forecast). It assumes no increase in training compute. (2) The AI Futures Model Dec 2025 update (blog.aifutures.org/p/ai-futures-model-dec-2025-update) moves the automated-coder milestone later by 3 to 5 years because it is less bullish about AI R&D speedups. This pass did not extract the update's exact takeoff figure. (3) Davidson and Houlden (www.forethought.org/research/how-quick-and-big-would-a-software-intelligence-explosion-be) give about 60% for compressing more than 3 years of progress into less than 1 year, and about 20% for compressing more than 10 years into 1. That suggests a fast takeoff is likely but not assured. Combination: I give about 40% weight to fast-takeoff model outputs, with a median of about 1 year. I give about 60% weight to bottleneck-dominated scenarios, with medians of 3 to 10 years and a long tail beyond 20 years because the gap from AI research to all cognitive tasks is wide. The mixture's median is about 3 years. The 5th percentile is 0.4 years and the 95th is 25. Separately, about 15% is held back for the case where milestone (b) never follows (a). These are judgment-weighted, not computed. The main gap is the exact takeoff distribution in the Dec 2025 model.
Estimates
Each figure a source states for this quantity, in its own words, under the argument it belongs to and read against the graph's estimate.
Grey band: the graph's 90% interval, with its central value as a line. Dark marks: each source's own figure. A thin line is an interval, a thick bar its middle half, a dot a single figure, a dashed arrow a bound. An open box is where an argument places the value.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this quantity's evidence comes from. The figures sources state are under Estimates above.
Over the course of 2027, the AIs improve from being able to mostly do the job of an OpenBrain research engineer to eclipsing all humans at all tasks. This represents roughly our median guess, but we think it’s plausible that this happens up to ~5x slower or faster.
Expandable on why uncertainty increases beyond 2026.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
Created by extractor · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.