Semi-informative priors over AI timelines

This website lets you provide your own parameter values for some of the calculations in the Open Philanthropy report "Semi-informative priors over AI timelines". We recommend that you read about the framework info here before entering your own values.

See parameter values used in report info : Low, Central, High

Calendar year trial definition

A trial is a year of calendar time.

First trial probability info

Regime start year info

Probability of AGI by 2036: 4.2% info

Researcher-year trial definition

A trial is a small % increase in the number of AI researchers. If the number of AI researchers grows exponentially, there's a constant number of trials each year. (More)

Typical annual growth for STEM researchers (%) info

Annual growth of AI researchers until 2036 (%) info

Probability of AGI by 2036: 8.0% info

Computation trial definition: relative importance of research and computation

Under this trial definition, a trial is a small % increase in the maximum amount of computation used to develop an AI system. If computation grows exponentially, there’s a constant number of trials each year. (More)

The regime starts in 1956.

A 1% increase in the number of researchers is equivalent to a

% increase in computation info

Maximum computation spend by 2036 ($ billions) info

Probability of AGI by 2036: 9.0% info

Computation trial definition: lifetime anchor

Consider the computation inside a human brain during the first 30 years of life. We calculate the first trial probability by assuming that by the time we’ve used that amount of computation to develop one AI system, there’s a 50% chance that we’ll have developed AGI (this would have been a salient assumption to make in 1956). (More)

The regime starts in 1956.

Probability of AGI by 2036: 11% info

Computation trial definition: evolutionary anchor

Consider the total computation needed for the human brain to evolve. We calculate the first trial probability by assuming that by the time we’ve used that amount of computation to develop one AI system, there’s a 50% chance that we’ll have developed AGI (this would have been a salient assumption to make in 1956). (More)

The regime starts when the computation needed to run the human brain for two weeks is first used to develop an AI system. This means that many fewer failures have been observed than for the other computation rules.

Probability of AGI by 2036: 16% info

All trial definitions on one plot

Hyper-prior update info

Initial weights of update rules info	Initial weights info	Weights in 2020	Probability of AGI by 2036
Calendar-year trial definition	30%	32%	4.2%
Researcher-year trial definition	30%	25%	8.0%
Computation trial definition: relative importance of research and computation	5.0%	4.0%	9.0%
Computation trial definition: lifetime anchor	10%	6.5%	11%
Computation trial definition: evolutionary anchor	15%	19%	16%
AGI is impossible	10%	13%	0.0%

Weighted average probability of AGI by 2036: 7.5% info

Additional parameters

These parameters affect all trial definitions.

Virtual successes info

We can rule out AGI being developed before the end of

Note: some results may differ very slightly from those in the report, due to approximations we use on this website for the sake of speed.