Questions for RP
What would most reduce our uncertainty about RP’s GCR fund model, roughly in order
Help improve this draftComment with Hypothes.is; flag @daaronr for a response.
Select any text, chart or number and annotate it with Hypothes.is (the tab at the top right) to flag errors, missing sources, or a clearer way to show something. David plans to respond, especially where a comment mentions @daaronr. Public Hypothes.is annotations can be read by anyone who opens this page, so don't post anything confidential.
How this was made: an AI agent (Claude, working for The Unjournal) rebuilt Rethink Priorities' GCR fund model from RP's public write-up, update log and Donor Compass data, and built these pages. David Reinstein set the task. Neither he nor RP has checked the numbers yet, and RP hasn't seen this page.
We rebuilt RP’s model from its public write-up, update log and Donor Compass data. These are the points where we had to guess, or where our rebuild and RP’s published numbers disagree. Each says why it matters and which control on the model page lets you see the effect of our guess.
1. The shape of the risk curve
What functional form do you use for the time of perils: what is the risk now relative to the peak, and how does it fall after the peak?
Why it matters: it sets the level of every window. We had to fit it to your numbers, and our 10-500 year windows still come out 2-6% low; a plateau instead of a steady fall would cut them by about a quarter.
Controls: "Shape of the risk curve" and "Risk now, as a share of its peak" (Reconstruction choices); "Years until risk peaks" and "How long risk stays high after the peak".
2. How a fund's cut in risk applies
Is the relative reduction applied to the cause's risk in each year the effect lasts, or to the peak level? Does it apply to the long-run risk as well as the extra risk of the time of perils?
Why it matters: the two readings differ by about 15% in the 20-100 and 500+ windows, and the second question could matter much more for the 500+ window.
Control: "How a fund's cut applies" (Reconstruction choices).
3. The distributions behind the inputs
Can you share the sampling code, or the list of distribution families? We inferred Dirichlet distributions for the cause shares and for the outcome chances, and log-uniform distributions for long-run risk and settlement speed. Are those right? Is the growth rate's mean 1% or about 1.2%?
Why it matters: the 500+ window depends on the far tails of these distributions. Swapping log-uniform for lognormal long-run risk, with the same 5th and 95th percentiles, moves it by five or more orders of magnitude.
Controls: "Shape of the long-run risk distribution" (Reconstruction choices); the cause shares under "Who causes the risk"; each fund's outcome chances.
4. The value of settling the stars
What is the exact formula for the value of a future in which humanity settles the stars, after the June 2026 fix (densities, radii, end date, whether risk changes after settlement), and what was the error the fix corrected?
Why it matters: futures with settlement carry essentially all of the expected value of the future, and so all of the 500+ window. We used Tarsney's (2020) cubic growth model as written.
Controls: "Value of each settled star" and "How far settlement can go" (Reconstruction choices); "Chance humanity settles the stars"; "Settlement speed".
5. Damage from smaller catastrophes
How many life-years are lost per death, and how are deaths within each size band set? Are the 10-year chances spread evenly over the next 100 years, or do they follow the time-of-perils shape?
Why it matters: these set the first 20 years, where non-existential effects are most of the value. We fitted one number (25.6 life-years per death) to Sentinel Bio's mean; the AI fund then comes out 5% high.
Control: "Life-years lost per death in smaller catastrophes" (Reconstruction choices).
6. The value-of-the-future table
Is the table of percentiles for the value of the future from the same run as the June 2026 fund numbers?
Why it matters: our mean and every percentile up to the 99th match yours, but our 99.9th percentile is about 8 times yours. Either your tail is thinner there, or the table is from a different run.
See "Where the value sits" on the model page and How close is it.
7. The walkthrough's 1032
Where does "on the order of 10^32" in the Donor Compass walkthrough come from, and why is the AI fund ten times below Sentinel Bio there (1.1 × 1030 vs 1.1 × 1031 after discounting) when it is 2.5 times above in all four datasets in the data viewer? What changed between the 31 March dataset and 17 April?
Why it matters: donors reading the walkthrough get a far-future number 12 orders of magnitude above the data, and the opposite ranking of the AI and bio funds.
See The 10^32 puzzle; try "Shape of the long-run risk distribution" set to lognormal.
8. Monte Carlo stability
With 10 million draws, how much do your 500+ averages move between seeds?
Why it matters: ours move about 10-15% between seeds of a million draws even with a lower-noise estimator, and your 500+ values moved 3 to 4 times between the May and June datasets. If Donor Compass allocations depend on the 500+ window, that noise reaches the recommendations.
Controls: "Draws per run" and "Random seed" (Simulation), and "Run more draws" under the results.
If you work at RP and can answer any of these, or think we’ve misread something, please annotate this page with Hypothes.is or email contact@unjournal.org.