M = import(new URL("./xrisk-model.mjs", window.location.href).href)
manifest = fetch(new URL("./model-manifest.json", window.location.href).href).then(r => r.ok ? r.json() : null).catch(() => null)
md`Model engine version: **${M.MODEL_VERSION}**; settings schema ${M.SETTINGS_SCHEMA_VERSION}.${manifest ? ` This build: commit \`${String(manifest.git_commit).slice(0, 7)}\`, built ${String(manifest.built).slice(0, 10)}${manifest.uncommitted_changes ? " (with uncommitted changes)" : ""}.` : ""}`About
What this is
An interactive version of the model in:
Arvo Muñoz Morán, Technical Report: The Value of Existential Risk Mitigation, Rethink Priorities, Worldview Investigations Team, 23 October 2023 (part of the CURVE sequence). Companion notebook: github.com/arvomm/verm (VERM 3.0.ipynb).
The Unjournal evaluated the report in 2026. The evaluation raised three main points: the report’s worked analyses treat the risk path as fixed and the mitigation’s effect as known and positive, so they don’t address actions whose long-run effect on risk is uncertain in sign; the value-growth paths (up to cubic, with no ceiling) do much of the work; and with any floor on annual risk, a finite horizon matters more than whether an infinite series converges. Each of these is a switch on the model page.
The report’s concluding remarks (section 8) list “a friendlier online platform with sliders and buttons to select and tweak the scenarios users want to visualise” as possible follow-up work, along with “explicit uncertainty analyses with Monte Carlo simulations”. This is our attempt at both, plus the evaluator’s alternatives.
What it’s for, and what it isn’t
It’s a research model for seeing how the modeled value of reducing extinction risk depends on assumptions about risk, the value of the future, how long an effect lasts, and whether it could backfire. It isn’t a forecast, an estimate of what any real intervention achieves, or a recommendation about where to give. Like the report, it models extinction only: catastrophes that permanently curtail humanity’s potential without ending it, and partial catastrophes, are outside it (the report’s footnote 1, p. 7).
Uses we have in mind:
- Checking the report’s numbers. They reproduce exactly, once the notebook’s conventions are used.
- Seeing how much of each number depends on the far future, the growth path and the persistence assumption.
- Trying the evaluator’s alternatives one at a time, starting from the report’s assumptions.
- Comparing with Rethink Priorities’ current model of global catastrophic risk funds, which feeds its Donor Compass giving tool and lists this report as an influence (Departures). The Donor Compass page shows, with RP’s published data, how that model’s far-future values play out in Donor Compass. The Context page places the tool among related models and research.
What has and hasn’t been checked
The checks so far are about whether the code does what the report does. None of them shows that any input is right.
| Check | Status |
|---|---|
| Reproduces the report’s arithmetic (Table 3, Table 2 closed forms, the section 2.2.1 example, Figure 8) | Done: matches the notebook’s output to 1e-8 or better, in automated tests |
| Closed forms agree with the numerical sums | Done, in tests |
| Invariants (no effect when f = 0; survival falls and stays between 0 and 1; value rises with f and P in the report’s model; tail bounds cover what a horizon leaves out) | Done, in tests over many random scenarios |
| Monte Carlo draws agree with direct recomputation; seeded runs repeat exactly | Done, in tests |
| Links and shared scenarios reopen the same model; bad values are reported, not silently replaced | Done, in tests |
| Donor Compass page: our scoring, the six aggregation methods RP mixes for the Cross-Cause Fund, and the mix itself give the same splits as the code in RP’s Donor Compass app, on RP’s data and default worldviews | Done, in tests; at $200M the mix is within a quarter of a point of RP’s July 2026 distribution for every fund |
| Review by an outside expert | Not yet |
| Review or endorsement by the report’s author or Rethink Priorities | Not yet |
| Evidence for the risk paths, value paths and persistence | Outside this tool: these are the report’s assumptions, RP’s, or yours |
| Evidence for the backfire and rebound settings (switches 2 and 3) | None: the defaults are illustrative |
| Whether the value of the future is positive, and how to value it | A moral question the model doesn’t settle |
Each setting in use is tagged on the model page by where its value comes from (the 2023 report, Rethink Priorities’ 2026 model, an illustrative choice of ours, or you), so it’s clear which of these categories a result rests on.
How it was built
The Unjournal built this with an AI coding assistant (Claude), working from the PDF, the companion notebook and the evaluation. Automated tests check the engine against the notebook’s printed output, and we spot-checked ten Table 3 cells by hand against the PDF. An AI-generated audit of the live site in October 2026 suggested several changes on the model page (the source tags, validation of link values, the quick start, chart descriptions for screen readers); its larger suggestions are listed as future work. No outside expert has reviewed the tool yet. The report’s author hasn’t reviewed it, and it isn’t endorsed by the author or by Rethink Priorities.
The site is built with Quarto and Observable JS. The calculations run in your browser, in a single JavaScript file (xrisk-model.mjs) shared by the pages and the tests. It’s hosted on Netlify. Charts use Observable Plot 0.6.17, loaded from jsDelivr; comments use Hypothes.is.
Privacy
- There are no analytics or tracking scripts on this site.
- Your settings stay in your browser unless you share them. A share link or summary carries all your settings, so anyone you send it to sees your assumptions. The page removes the settings from the address bar after loading.
- If you submit a scenario, what you type in the form is stored by Netlify (our host) and read by The Unjournal. We publish a scenario on the Shared scenarios page only if you tick the box, and we never publish email addresses.
- Hypothes.is annotations are public by default and are stored by Hypothes.is.
- The page remembers in your browser whether you hid the settings panel and which scenario you pinned for comparison. Nothing else is stored.
Version
- 2026-10-19 (later): the Donor Compass page now covers all six aggregation methods RP mixes for the Cross-Cause Fund, and the mix, at the Fund’s $200M budget. Under every method, scaling the GCR funds’ far-future values up tenfold or down a billionfold doesn’t change the split; whether that window counts, and the risk profiles, do.
- 2026-10-19: new Donor Compass page: how RP’s Donor Compass and Cross-Cause Fund draw on the report, with dashboards built on RP’s published fund data (June 4, 2026 version) and default worldviews. The model’s arithmetic is unchanged.
- 2026-10-14: changes from our own Hypothes.is comments. The intro says plainly what E(M) measures, with the caveats in explanations; the quick start offers a wider range of long-run risks (or your own number), states the size of any backfire, and links each question to its exact control; setting names under the results jump to their controls; under constant risk the risk slider is labeled as such; with switch 2 on, the headline box shows the average (net) effect and the chance of net harm.
- 2026-10-13 (later): readable switch prompts (they had been painted over in gray), explanations on hover or tap for every setting and result, a “Hide settings” button, a share section (link, summary, downloads) and a form for submitting scenarios to a shared list; source tags on every setting, notices when a link’s values can’t be used as given, a quick start, a comparison with a pinned scenario, and a Context page. Headline numbers now show three significant figures. The model’s arithmetic is unchanged.
- 2026-10-13: first hosted version. Added presets (including Rethink Priorities’ 2026 central values), a comparison with RP’s current GCR model, and a how-to-comment page; controls now show only what’s relevant to the chosen scenario; corrected our reading of the report’s “two billion” sentence.
- 2026-09-17: first build; reproduces Table 3, Table 2, the section 2.2.1 example and Figure 8.
License
Code and text on this site are released under the MIT license. The report and notebook belong to their author and Rethink Priorities; the tool reproduces their equations and parameter values, not their text or figures.
Comments and corrections
See how to comment. In short: Hypothes.is on any page (the
<tab at the top right), or email contact@unjournal.org.