Open Philanthropy (OP) is a major funder of work which aims to reduce AI risk. Consequently, studying OPās past AI grants may be helpful to inform efforts to decrease AI risk.
OPās AI grants were considered as those in OPās grants database whose focus area is Potential Risks from Advanced AI[2]. The resulting 107 grants are in tab āGrantsā of this Sheets. The grant amounts were adjusted for inflation based on data from in2013dollars (see tab āInflationā), and are expressed in 2020-$.
The following linear regressions were studied for OPās AI grants:
The results were determined with this Colab. The linear regression parameters, data points, linear trends and respective 90 % confidence intervals are presented in the tables and figures below.
Linear regression of grant size on date | |
| Slope (M$/grant/year) | -0.257 |
| Intercept (M$) | 522 |
| Correlation coefficient | -0.0601 |
| Coefficient of determination | 0.00361 |
| P-value[3] | 0.539 |
| Standard error of the slope (M$/grant/year) | 0.417 |

Linear regression of annual amount granted on year (2015-2021) | |
| Slope (M$/year) | 9.34 |
| Intercept (M$) | -18.8 k |
| Correlation coefficient | 0.650 |
| Coefficient of determination | 0.423 |
| P-value[3] | 0.114 |
| Standard error of the slope (M$/year) | 4.88 |

The above analysis of OPās AI grants indicates that:
The trend of the grant size cannot be meaningfully extrapolated forward, given the low coefficient of determination of 0.4 %.
The trend of the annual amount granted can be projected forward more meaningfully, given the higher coefficient of determination of 40 %. Assuming an annual inflation rate of 1.75 %[4], the annual amount granted is predicted to be:
This analysis was performed as part of a paid work trial at Epoch.
The focus areas of the downloadable CSV file of the database are not in agreement with those of the website. The latter were considered correct, as there were only 3 grants in the CSV with focus area āPotential Risks from Advanced AIā.
Considering null slope as the null hypothesis.
Annual inflation rate between 2010 and 2020. Computed in cell B12 of tab āInflationā.
It seems like normality is violated on the first graph, have you tried taking a log transform or something?
Thanks for commenting!
I had not tested it, but your are kind of right! Applying the Shapiro-Wilk normality test to:
So it looks like we can confidently reject the grant size following a normal distribution, but it could well follow a lognormal distribution. This aligns well with my expectations.