The 10³² puzzle

RP’s walkthrough puts the GCR funds’ far-future effects 12 orders of magnitude above RP’s own data. Where could that come from?

Unofficial reconstruction from RP's public documentation, not RP's model. Draft for review: exploratory and premature. It was largely generated by AI and hasn't been verified by us or reviewed by Rethink Priorities. Don't cite these numbers as RP's, or use them for giving decisions.

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.

How to use Hypothes.is Read comments on this page

Model What drives the result How close is it The 10³² puzzle Questions for RP 2023 report vs 2026 model Technical notes

The puzzle

RP’s Donor Compass walkthrough says the GCR funds have effects in the 500+ year window “on the order of 10^32 human life-years”, glossed as the value of all future generations if extinction is averted. Its example scores, after a 0.10 weight on that window, imply 500+ values of about \(1.1 \times 10^{32}\) per $1M for Sentinel Bio and the nuclear fund and \(1.1 \times 10^{31}\) for the AI fund.

But every dataset in RP’s data viewer (17 April to 4 June 2026), and RP’s write-up, put the 500+ value at \(1.3 \times 10^{19}\) to \(1.6 \times 10^{20}\) per $1M. The walkthrough says it uses a 31 March 2026 dataset, which isn’t in the viewer. And in the walkthrough the AI fund is ten times below Sentinel Bio, while in every later dataset it is about 2.5 times above.

What the reconstruction says

Plot = import("https://cdn.jsdelivr.net/npm/@observablehq/plot@0.6.17/+esm")
D = import(new URL("./engine/dashboard.mjs", window.location.href).href)
chartW = Math.max(300, Math.min(860, width))
// Values from rp_gcr/results/walkthrough_1e32.md (1 million draws per case, seeds 1-3) and RP's walkthrough.
cases = [
  {label: "RP's data, all datasets: 500+ per $1M", group: "RP's data viewer and write-up (all datasets)", what: "500+ per $1M", lo: 1.3e19, hi: 1.6e20, ours: false},
  {label: "Ours, RP's values: 500+ per $1M", group: "Reconstruction, RP's values (3 seeds)", what: "500+ per $1M", lo: 5.2e19, hi: 1.4e20, ours: true},
  {label: "Ours, lognormal risk: 500+ per $1M", group: "Reconstruction, long-run risk lognormal (3 seeds)", what: "500+ per $1M", lo: 1.2e25, hi: 2.9e29, ours: true, note: "bio and nuclear funds; the AI fund ranges from -1.2e27 to 1.0e30, so its sign flips between seeds"},
  {label: "RP's walkthrough: 500+ per $1M", group: "RP's walkthrough (31 March dataset)", what: "500+ per $1M", lo: 1.1e31, hi: 1.1e32, ours: false},
  {label: "Ours, RP's values: whole future", group: "Reconstruction, RP's values (3 seeds)", what: "Value of the whole future", lo: 6.0e26, hi: 6.1e26, ours: true},
  {label: "Ours, lognormal risk: whole future", group: "Reconstruction, long-run risk lognormal (3 seeds)", what: "Value of the whole future", lo: 5.5e31, hi: 3.8e32, ours: true}
]
Plot.plot({
  width: chartW, height: 300, marginLeft: chartW < 600 ? 190 : 230, marginRight: 20, marginBottom: 40, marginTop: 10,
  ariaLabel: "Orders of magnitude: RP's data, the walkthrough and the reconstruction",
  ariaDescription: cases.map(c => `${c.group}, ${c.what}: ${D.fmt(c.lo, 2)} to ${D.fmt(c.hi, 2)}`).join("; "),
  x: {domain: [17, 34], ticks: chartW < 600 ? [18, 22, 26, 30, 34] : [18, 20, 22, 24, 26, 28, 30, 32, 34], tickFormat: k => `10${D.sup(k)}`, grid: true, label: "Life-years (log scale) →"},
  y: {domain: cases.map(c => c.label), label: null},
  marks: [
    Plot.ruleY(cases, {y: "label", x1: c => Math.log10(c.lo), x2: c => Math.log10(c.hi), stroke: c => c.ours ? "#2b6cb0" : "#1a202c", strokeWidth: 8, strokeLinecap: "round"}),
    Plot.ruleX([32], {stroke: "#c05621", strokeDasharray: "4,3"}),
    Plot.tip(cases, Plot.pointerY({y: "label", x: c => Math.log10(Math.sqrt(c.lo * c.hi)), title: c => `${c.group}\n${c.what}: ${D.fmt(c.lo, 2)} to ${D.fmt(c.hi, 2)}${c.note ? `\n(${c.note})` : ""}`}))
  ]
})

Bars span the range across seeds or datasets; the dashed line is \(10^{32}\). Blue: our reconstruction; black: RP’s published numbers. For the lognormal case, values are the lower-noise estimates for the bio and nuclear funds; the AI fund’s comes out negative in one seed of three.

1. With RP’s current inputs, \(10^{32}\) per $1M is out of reach. The expected value of the whole future is about \(6 \times 10^{26}\) life-years, and $1M shifts the chance of surviving by around \(10^{-7}\). That gives \(5 \times 10^{19}\) to \(1.4 \times 10^{20}\), RP’s order of magnitude.
2. One small change gets the walkthrough’s number, for the wrong quantity. RP gives long-run risk a mean and a 90% range but not the shape of the distribution. If it is lognormal rather than log-uniform (same 5th and 95th percentiles), very low long-run risk becomes much more likely in the far tail. The value of the whole future then averages \(5 \times 10^{31}\) to \(4 \times 10^{32}\) across seeds: the walkthrough’s order of magnitude, and its own description (“the value of all future generations”).
3. But then nothing is stable. The 500+ values per $1M swing from \(10^{25}\) to \(10^{30}\) between seeds of a million draws, and the AI fund’s sign flips. The lower-noise estimate of the value of the future ranges from \(10^{33}\) to \(2 \times 10^{36}\): there is no stable average.

Our best guess

Either (a) the walkthrough quotes the value of the future rather than a per-$1M effect, from a model version with a heavier-tailed long-run risk; or (b) the 31 March dataset had a much heavier tail and the walkthrough reports its per-$1M numbers. We can’t tell which. Neither explains why the AI fund is ten times below Sentinel Bio in the walkthrough; that looks like a different model version, and it is one of our questions for RP.

Why it matters

A choice that looks innocuous, log-uniform versus lognormal through the same two percentiles, moves the far-future number by five or more orders of magnitude. The 500+ window is, in effect, a statement about the lower tail of long-run risk. You can try it on the model page: under “Reconstruction choices”, set “Shape of the long-run risk distribution” to lognormal and change the random seed a few times.

Source: rp_gcr/results/walkthrough_1e32.md and rp_gcr/scripts/walkthrough_1e32.mjs (three seeds of a million draws per case).