Departures from the 2023 report
Where this 2026 tool differs from Muñoz Morán (2023), and how Rethink Priorities’ current model compares
A tool built years after a report should say exactly where it departs from the original. The first table lists every place this tool differs from the report as published (or from its companion notebook), including places where the tool picks one of several conventions the report left open. The second section compares the report’s model with Rethink Priorities’ current model of global catastrophic risk funds, which feeds its Donor Compass giving tool.
Departures from the report and notebook
| # | Aspect | Report (Oct 2023) | This tool (2026) | Why |
|---|---|---|---|---|
| 1 | Default horizon \(T\) | Text: “a cap of 120 thousand years” (fn. 27). Notebook: per-cell \(T\) from 5,000 to 120,000; Figure 8 uses 1,000 | Model page: user-selected, default 120,000 (the PDF’s stated cap). Replication page: the notebook’s per-cell values | Makes the horizon visible and changeable; the horizon table and tail bound show what any choice leaves out |
| 2 | Risk floor | No separate concept; the scenarios happen to have positive long-run risk (decay uses \(r_\infty = 10^{-9}\)) | Optional floor \(c\) with analytic truncation bounds (switch 1) | Evaluator’s point that a positive floor gives value a finite “approximate support” |
| 3 | Sign of \(f\) | “we could allow f ≤ 0”, but “For the rest of the report we focus on non-negative value” (p. 11); results are means | Optional distribution over \(f\) including negative values, optional delayed backfire, full distribution and \(\Pr[\mathbb{E}(M) < 0]\) (switch 2) | Evaluator’s point that practitioners choose among actions whose long-run sign is uncertain |
| 4 | Fixed risk path | Risk path fixed; actions never change later risk | Optional rebound: later risk scaled by \(1 + \beta f P / D\), or permanently by \(1 + \beta f\) (switch 3) | Simplest action-dependent risk; shows that a pure delay of risk can be net harmful when value grows |
| 5 | Value paths | Constant, linear, quadratic, cubic, logistic (all non-decreasing) | Adds polynomial of any order, uncapped exponential, hard-capped exponential, and a fluctuating (stochastic) path | Evaluator’s point that cubic isn’t special and that value need not rise steadily |
| 6 | Time aggregation | None beyond survival weighting; \(v_c\) normalized to 1 | Optional time-bucket factors (0-5, 5-10, 10-20, 20-100, 100-500, 500+ years) (switch 4) | Shows what share of \(\mathbb{E}(M)\) a Donor Compass style aggregation would discard |
| 7 | Mitigation timing | Starts at \(t = 1\) | Start year adjustable | Lets the RP 2026 preset start the effect in year 4; the report’s footnote 26 raises the contingency question but doesn’t model delay |
| 8 | Sensitivity analysis | Report lists Monte Carlo sensitivity as future work; the notebook has one-parameter sweeps | One-at-a-time sensitivity chart plus the Monte Carlo above | Report section 8, “explicit uncertainty analyses with Monte Carlo simulations” |
| 9 | Persistence uncertainty | Fixed \(P\) | Still fixed \(P\) (the persistence curve varies it deterministically) | Not implemented; a natural extension |
| 10 | Logistic functional form | Two forms appear: eq. 6 (\(v_c(c(1 - e^{-(t/a)^b}) + 1)\), used for Figure 2) and eq. 7 / Table 1 (standard logistic, used for all numbers) | Only the standard logistic (eq. 7), with the notebook’s \(c = 10^6\), \(\gamma = 0.03\) as defaults | Eq. 6 produces none of the reported numbers |
| 11 | Exponential decay parameters | Figure 3 draws a decay about 50 times faster than the one used in Table 3 and Figure 4 | Uses the notebook’s \(\lambda = 10^{-4}\), \(r_\infty = 10^{-9}\) everywhere; both adjustable | The tables are the quantitative claims |
| 12 | Printed precision | Three significant figures, truncated not rounded | Full double precision; the replication page shows both | Transparency |
| 13 | Concrete example risk | \(r = 0.002229\) (4 s.f.) rather than 0.0022289477 | Reproduced with the example’s own constant on the replication page; the model page uses the notebook constant | Fidelity to each number’s own convention |
| 14 | Risk input units | Per-year constants derived from per-century probabilities in footnotes | Sliders in percent per century, converted to per-year; the defaults match the report’s constants to within \(10^{-5}\) relative | Readability |
| 15 | Text claims not reproduced | “11 million times” (p. 30); “two billion \(v_c\) … assuming f is one basis point” (pp. 31-32); “up to 30 times” (p. 31) | The tables give about 6.2 million; about 546,000 at \(f = 0.0001\) and \(P = 50\) (2.8 billion at \(f = 0.5\), the efficacy used for the figures); and roughly 300-fold for Figure 8 | Flagged on the replication page rather than silently corrected; we may be misreading them |
| 16 | Code | Python notebook, per-period loops | Pure JavaScript module, same arithmetic, with node --test checks against the notebook’s printed output |
Runs in the browser with no server |
Rethink Priorities’ current model and Donor Compass
Donor Compass recommends how to split a donation across funds. Its allocation model takes fund-by-fund cost-effectiveness estimates as inputs. For its three global catastrophic risk funds (Sentinel Bio, Longview’s Nuclear Weapons Policy Fund and Longview’s Frontier AI Fund), those estimates come from a separate model described in RP’s fund write-up, marked “a working draft”. That write-up says:
“This framework was based on work by Tarsney (2020), and some parameter settings were taken from that paper. It was also influenced by Thorstad (2022) and Munoz Moran (2023).”
It also says RP plans “extensive internal and external review of model structure and parameter choices for the three Global Catastrophic Risk funds.”
| Feature | This report | RP’s GCR write-up (2026 working draft) | Here |
|---|---|---|---|
| Risk path | Time of Perils: \(r_{high}\) for 100 years, then \(r_{low}\) | Risk “peaks some time in the near future (in the next few decades) and remains elevated for some time”. Central values: 20% chance of existential catastrophe in the next 100 years; peak in year 17; Time of Perils lasting 90 years; long-run annual risk that “Implies an average civilizational lifespan of about 120,000 more years” | The RP preset keeps the report’s step shape (20% per century for 100 years) but uses RP’s long-run risk, about 12 times lower than the report’s |
| Mitigation | \(r'_t = (1-f)r_t\); “f ∈ (0, 1) is the fraction of the risk that is successfully mitigated” (p. 11) | “relative reductions in their cause’s peak risk levels during the time of perils”; risk split 90% AI, 3% biological, 3% nuclear, 4% other | Same multiplicative form. RP applies it to cause-specific risk; the preset keeps the report’s \(f = 0.0001\) of total risk as a common yardstick |
| Persistence | \(P\) free; report shows 1 to 2,000 years and suggests above 50 is unlikely | “finite counterfactual persistence”, 11 years (2.5 to 30), starting after about 4 years | Preset: \(P = 11\), starting in year 4 |
| Sign of the effect | Non-negative \(f\) in all analyses | “funds may have zero impact on existential risk levels, or inadvertently increase them”; “Probability of raising risk” 5% for Sentinel Bio and Longview nuclear, 15% for Longview AI | Switch 2 adds a comparable spread |
| Value | Constant, linear, quadratic, cubic or logistic growth | On Earth, growth “represented using a logistic growth function” (1% a year to 1.6 times today’s value); in a space-settlement branch with 6% probability, “The total value of civilization grows cubically” | The preset includes only the Earth-bound branch. RP says the average is “dominated by the extreme scenarios in which space settlement occurs”, so the preset understates RP’s mean |
| Uncertainty | Fixed parameters, means only | Monte Carlo over all parameters | Not reproduced, except through switch 2 |
| Time aggregation | None | Donor Compass applies six time-window factors chosen by the user’s worldview | Switch 4 |
| Where the value lands | Not reported | “The overwhelming majority of these impacts occur in the 500+ year time window, which model users might discount.” | The horizon table and switch 4 show the same pattern in the report’s model |
The RP preset is our approximation, using this report’s simpler structure. It is not RP’s model, and we’d welcome corrections from RP.
One thing we couldn’t reconcile: the Donor Compass walkthrough says GCR funds have “effects in the 500+ year column on the order of \(10^{32}\) human life-years”, while the GCR write-up’s table gives about \(6 \times 10^{19}\) to \(1.6 \times 10^{20}\) life-years per $1M in that window. This may be a difference in units or versions.
The 2023 predecessor
RP’s Cross-Cause Cost-Effectiveness Model overview (November 2023), which thanks the report’s author for “conceptual guidance”, described an earlier x-risk module with the same structure: “We measure mitigation effectiveness by the proportional reduction of yearly risk”; “existential risk mitigation work may lower (or accidentally raise) the chance of risk of catastrophic or existential events over a few decades, but has no perpetual impact on the level of risk”; and, on value, “we default to a cutoff point in a thousand years, where the population is limited by the Earth’s capacity”, with “a constant and equal rate of colonization in each direction” beyond that.
Sources checked on 12-13 October 2026. Quotations are verbatim; everything else is our summary and may be out of date. Corrections welcome as a Hypothes.is annotation on this page.
What the tool could not reproduce or did not attempt
- The report’s Figures 4, 5 and 7 (the 4 by 5 grids) are reproduced numerically through the scenario grid and the persistence curve on the model page, not as figures.
- The notebook’s “many filters” experiments (\(F\) up to 100) are reachable through the filters slider only up to \(F = 10\).
- The evaluator’s lim inf / lim sup suggestion on convergence is not formalized; the tool works with finite horizons and explicit tail bounds instead.
- Population growth, partial catastrophes, non-extinction existential risks, negative value paths and option value (all listed as future work in the report’s section 8) are not modeled.