Skip to content
Multigrid

AGI Forecast Explorer

Enter the AGI forecasts you want to compare as three quantile years each, and read them all off as probabilities at any year you choose.

The three lines above are illustrative placeholders, shaped to show what the tool does. They are not anyone’s published forecast and are attributed to nobody. Replace them with the forecasts you want to compare, and cite each one to its source and date when you share the link — a forecast without a date on it is not a forecast.

Pooled probability of AGI by 2040
33.5%

Equal weight across 3 forecasts, which individually run from 0.7% to 74.6%. The pool crosses 50% in 2054.

Cumulative probability of AGI by year, per forecast25%50%75%204020602080204020262100 · dashed line is the equal-weight pool
Each curve is one forecast's cumulative probability, fitted through the three years you gave it. The vertical line is the cursor. A curve that flattens below the top of the chart is one with weight on 'never'.
  • Illustrative: fast view
  • Illustrative: middle view
  • Illustrative: sceptical view
1. Illustrative: fast view — P(by 2040)
74.6%
2. Illustrative: middle view — P(by 2040)
25.2%
3. Illustrative: sceptical view — P(by 2040)
0.7%
1. Illustrative: fast view — fit vs target at its own 10/50/90
7%/48%/83% vs 10%/48%/86%
2. Illustrative: middle view — fit vs target at its own 10/50/90
6%/43%/74% vs 9%/43%/77%
3. Illustrative: sceptical view — fit vs target at its own 10/50/90
5%/30%/53% vs 6%/30%/54%
Equal-weight pool at the cursor
33.5%
Year the pool crosses 50%
2054
What this assumes: each forecast is turned into a lognormal distribution over years-from-2026: the median year sets its centre and the gap between the 10% and 90% years sets its width. The three years are read as conditional on it happening at all, and the “never” mass then scales the whole curve down — so a forecast with 40% on never tops out at 60% however far right you drag the cursor. That is an interpolation, not something anybody said, and it is symmetric in log-time, so if your three years are not, the curve will not pass exactly through them. The breakdown prints what the fit gives at each of your years next to what it should give, so the error is visible rather than taken on trust. The pool is a plain equal-weight average; it is not a considered aggregate, and weighting three forecasts equally is itself a strong claim.

The hardest thing about comparing predictions of this kind is that they are almost never stated in the same form. One is a median year, another a probability by a fixed date, a third a range with no distribution attached, and a fourth is a sentence with no numbers in it at all. Converted to a curve, they become comparable — and the disagreement usually turns out to be smaller in some places and much larger in others than the headlines suggest. Drag the cursor to 2035 and then to 2075 and watch the gap between the highest and lowest curve open and close.

The “never” column is the one people leave out and should not. A forecast with 40% on never can have an early median among the remaining 60% and still sit below a more patient forecast at every year on the chart. If a source does not give you that number, the honest thing is to enter your reading of it and say so, not to set it to zero because zero is the default.

Two limits worth stating plainly. First, none of this is evidence about when AGI arrives; it is bookkeeping on what other people have said, and the arithmetic is exactly as good as the inputs you type. Second, almost every disagreement about timelines is a disagreement about the definition — the same person can hold two very different curves for “matches a skilled human at most economically valuable work” and “can do novel science unaided”. Put the definition in the label. It is doing more work than the years are.

The reasoning behind this

Related tools

AGI Forecast Explorer · Multigrid