Share of human cognitive work AI can do
Vertical axis is breadth at human level. Horizontal axis is the year.
You are off the baseline. The faint line is the most likely timeline.
The drivers
Drag these. The milestones move on their own.
The assumptions, printed
What is actually under the curve.
The whole thing is a logistic curve. Each driver moves the year the curve passes 50 percent by a fixed number of years, and compute and algorithmic progress also sharpen the takeoff slightly. Those coupling numbers are below, and they are read straight out of the model, so what you see here cannot drift from what the page is actually computing.
| Driver | Years it moves AGI on its own | Slider runs from |
|---|
The part to be suspicious of. The coupling numbers are the weakest link in this model, and they are judgement, not measurement. Compute and algorithmic progress are ranked hardest because those are the levers with the clearest historical track record. Quantum is deliberately tiny because on current roadmaps it does close to nothing for AI capability this decade, and inflating it would be the easiest way to make this dishonest. The baseline date for AGI sits between aggregated forecaster estimates, which run earlier, and large surveys of AI researchers, which run later.
The single most contested number is the gap from AGI to ASI. The baseline here is 12 years, roughly the median of the surveys that ask it directly. People who expect recursive self-improvement to bite hard think it is closer to months. People who expect it to grind think it is decades. That disagreement is worth more than a footnote, so it is a dial rather than a hidden constant, and moving it changes the answer more than almost anything else on this page.
The single most contested number is the gap from AGI to ASI. The baseline here is 12 years, roughly the median of the surveys that ask it directly. People who expect recursive self-improvement to bite hard think it is closer to months. People who expect it to grind think it is decades. That disagreement is worth more than a footnote, so it is a dial rather than a hidden constant, and moving it changes the answer more than almost anything else on this page.
Where the baseline comes from
- Epoch AI for frontier training-compute growth and algorithmic-efficiency trends, which set the shape and the two strongest couplings.
- Large surveys of published AI researchers for the late end of the AGI range.
- Aggregated public forecasting for the early end of the AGI range.
- Robotics industry shipment and cost data for the mass-production milestone.
Live source links are not wired up in this first build. That is the next thing to fix, and until it is done, treat the baseline as a defensible starting point rather than a cited one.