Model comparison · evidence snapshot: 2 October 2026
Compare overall AI-linked philanthropy, AI safety, EA-aligned giving, and global health & development (GH&D). The original AIS/EA model, GH&D fork, Lewars’s sheet, Radiant, and other estimates cover different donors, causes, and deadlines. Choose an outcome below; distinguish native predictions from calculations that add a cause-allocation assumption.
A 3.44× central gap. For the broader horizon, Radiant’s all-cause mean is $30.7B through mid-2031, with P1–P99 $1.98–150B. Compare scope and uncertainty.
Our original central output versus an extension of Lewars, not his own EA forecast: a 4.24× gap. Our native median is $0.83B; P10–P90 $0.16–2.98B. Inspect allocation assumptions.
A native safety benchmark with no quantified interval. The core IPO models need added cause shares. Explore conditional allocations or select AI safety in the prediction table.
A 7.35× gap on the closest native cause match. Our full fork’s median is $0.54B; P10–P90 $0.16–2.78B. The bridge now covers all three matched targets.
1 · Outcomes before numbers
Choose an outcome. Dollar amounts are in billions of US dollars as stated by each source, without a common inflation adjustment. Rows retain their donor set, cause, time window, and stock/flow definition. All-cause and cause-specific views receive equal treatment; native, recomputed, and assumed-allocation outputs are labeled separately.
| Model / snapshot | Central output | Range / uncertainty | Exactly what is counted | Basis / comparability |
|---|
P10–P90 contains 80% of a model’s simulated outcomes; P1–P99 contains 98%. Neither is an empirically validated confidence interval. More draws improve numerical stability, not the truth of assumptions. A bear/base/bull spread has no coverage probability unless its author supplies one.
A plug-in estimate evaluates a formula at the central inputs. A median is the halfway point of the output distribution; a mean averages all outcomes, including large tails. Our AIS/EA figures are $0.95B plug-in, $0.83B median, and $1.29B mean. Our GH&D figures are $0.625B, $0.545B, and $1.207B respectively. For the matched Anthropic-only subset, rerunning the same GH&D simulation with the two program-dollar inputs set to zero gives median $0.412B and an 80% band of $0.059–2.603B. The unchanged random seed and other defaults are retained; this is a rescoping diagnostic, not an adopted forecast. Those differences arise within the same models. Reproduction details.
The separate all-cause rescoping diagnostic has central $2.951B, median $2.494B, mean $4.337B, and P10–P90 $0.503–10.191B. An in-memory copy fixes cause allocation at 100% and returns the uncapped, epistemically adjusted total. It retains all original upstream/timing draws, 4,000 draws, and seed 73519. The original cause-specific cap is not a general charitable-sector cap. This diagnostic is not an adopted forecast.
Radiant’s cards state simulated means. Its histogram “Midpoint” is the middle of the axis, not a median. We read the source’s P1/P99 labels and checked the documentation; we did not fabricate an 80% output interval from the visible 98% band.
Overall giving counts every charitable cause covered by that model. AI safety here means work explicitly aimed at mitigating serious AI risks; funding AI research or public-benefit AI is not automatically safety funding. EA-aligned is a broader, partly judgmental category that can include AI safety, cost-effective global health, animal welfare, and other impartial-welfare priorities. Lab affiliation alone does not establish EA alignment.
Our original first fork has one combined AIS/EA-relevant bucket, not separate AI-safety-only and EA-only forecasts. GH&D can overlap with that bucket. These views are not mutually exclusive slices to add together.
Our original calculators and simulations use an end-2027 deadline; separate judgment rows also discuss 2028. Extending the calculators to 2030 would require a new timing model. Radiant has no cause-allocation node; Ransohoff and Giving Atlas do not quantify cause shares. Their native cause-specific results stay unavailable, with conditional allocations explored below.
A four-year total divided by four is an average over that window, not a prediction for each year. Calendar deadlines differ from years after an IPO or lockup. Native definitions of regrantor transfers, operating-charity grants, and cash versus in-kind support are not perfectly matched.
2 · Overall, AI safety, and EA-aligned outcomes
Most external models predict all-cause giving. They cannot supply an AI-safety or EA-aligned forecast without another assumption. This lens applies the same editable share to each model’s all-cause grants. It makes that additional assumption visible and retains each source’s deadline and donor coverage.
| Model / native horizon | All-cause central output | Conditional cause output | Conditional range | Meaning / added assumptions |
|---|
For a fixed share s, the conditional mean, median, and source percentiles scale by s: cause grants = s × all-cause grants. Radiant at 10% therefore gives mean $3.07B and P1–P99 $0.198–15B through mid-2031. This follows from that assumption; it is not Radiant’s published AI-safety forecast. Scenario bounds remain scenarios after scaling.
The displayed interval includes only source-model uncertainty conditional on a fixed share. Uncertainty in allocation, donor-specific preferences, and cause-specific capacity can widen or change it. If giving and cause share are dependent, E[s × X] need not equal E[s] × E[X]. Treating one constant share as universal is a simplification for comparison.
The 30% EA starting lens loosely rounds our original combined fork’s central weighted share (about 32% of its uncapped all-cause diagnostic); 20% GH&D loosely rounds its corresponding share (about 17%). Neither establishes the share in a different donor pool. The narrower 10% AI-safety lens is an illustrative subset, with no observed allocation estimate behind it. These choices are editable assumptions, not fitted forecasts. AI safety, EA-aligned giving, and GH&D can overlap; do not add their totals.
Our first page also contains the following unchanged judgment ranges. These are probabilities of thresholds, not dollar uncertainty intervals or outputs of the calculator.
| Outcome and deadline | Judgment probability | Interpretation |
|---|---|---|
| At least $1B newly disbursed to AIS/EA by end-2027 | 10–30% | Separate from the original simulation’s 43.25% |
| At least $5B newly committed or disbursed to AIS/EA by end-2028 | 15–35% | Mixed commitment / grant stage |
| At least $20B/year steady-state AIS/EA disbursement by 2028 | 5–20% | Requires fast cause-specific deployment |
| At least $20B/year broad AI-era philanthropy by 2028 | 25–50% | Broader public-benefit spending; includes possible OpenAI pathways |
Original forecast table. A mean, central scenario, or interval at a different horizon does not supply a comparable probability for these thresholds.
3 · Reconcile the matched outcomes
Use the same donor set and deadline: Anthropic founders and employee charitable holdings, through end-2027. Lewars has zero flows in 2026, so his calendar-2027 flow is also his cumulative flow through end-2027. Choose overall giving, our combined AIS/EA category, or GH&D. Spreadsheet inputs and formula.
Turn on substitutions from Lewars’s sheet. These are diagnostic copies of the arithmetic; no changes are saved to either original model. The timing factors combine annual payout with the sheet’s four-year founder / two-year employee availability ramps.
| Assumption group | Contribution ($B) | Our value | Lewars value |
|---|
Overall: $2.951B versus $10.151B, a 3.44× ratio. Combined AIS/EA with the same cause shares: $0.950B versus $4.030B, a 4.24× ratio. GH&D with each model’s native cause shares: $0.503B versus $3.697B, a 7.35× ratio. The larger GH&D disagreement partly reflects different allocation shares; the other comparisons still differ substantially without that disagreement.
Holding cause shares equal, the employee pathway rises 5.56× and the founder pathway 1.99×. In the native GH&D comparison, additional allocation differences raise these to 12.5× and 2.65×. Different pathway weights explain why the total ratios differ. Multiplying all pathway ratios together would be wrong.
Lewars’s first-year effective deployment fractions, 2.5% and 5%, are below our 7% and 10%, although the definitions differ. His availability ramp and our realization gate overlap in meaning. Removing the latter to reproduce his formula does not establish that he believes “perfect liquidity.”
The default Shapley decomposition averages marginal changes over all 720 orders of six assumption groups and sums to the selected gap. In the overall and combined AIS/EA views, the cause-share group makes no change and contributes zero. This is an accounting convention; different groupings change attribution. The sequence switch exposes order dependence.
Our employee pool is an absolute $20B input; Lewars scales $60B at $900B to $133.33B at $2T. Our simulation co-moves employee wealth with a latent market factor, which differs from automatic central-slider rescaling. Both matched models omit Foundation grants, so Foundation coverage cannot explain these matched gaps.
4 · What the models actually do
Filter the models or search for a mechanism. Tooltips identify the relevant source or formula; expand the source ledger for provenance. “Not explicit” means we did not find that mechanism in the inspected implementation.
| Model | Wealth / donor coverage | Conversion and cause | Timing and balances | Uncertainty / evidence |
|---|
Ransohoff cites LongtermWiki for employee charitable equity. Lewars cites Ransohoff; Giving Atlas adopts her stock assumptions and adds a separate Good Ventures pathway. Our pages also discuss this same evidence. These are partly different transformations of shared estimates, rather than independent observations of donated holdings.
Radiant adds live community forecasts for valuations, IPOs and lockups, but its giving and payout assumptions still require judgment. Forecasting an IPO well does not automatically calibrate a downstream donation forecast. An arithmetic average over all the headlines would also give extra weight to repeatedly reused upstream assumptions.
5 · What the exercise teaches
The models make their authors’ questions visible. Our original exercise asks how a funding expectation could fail to become usable near-term grants. Lewars asks how much reaches development. Ransohoff asks what institutional capacity a large intended wave would require. These purposes help explain the variables emphasized; they do not establish that any author selected numbers to reach a preferred answer.
For one uncapped pathway, grants = wealth × ownership × giving × realization × cause share × deployment. Every multiplicative gate has elasticity one: increasing any one by 10% increases that pathway by 10%, holding the rest fixed. Six gates each 20% lower give 0.8⁶ = 0.262 of the original output. Six each 20% higher give 1.2⁶ = 2.986. Small choices can jointly move a result by more than an order of magnitude between those cases.
Dependence. In a product, uncertainty is additive in logs: Var(log grants) = Σ Var(log gates) + 2Σ Cov(log gates) for positive, uncapped quantities. Shared valuation / liquidity shocks can widen tails. Urgency may speed grants while lowering selection quality. Our models encode latent factors; those correlations are themselves assumptions. A causal-looking diagram does not prove statistical dependence: Radiant connections explain relationships; formulas do the calculation.
Caps and tails. A capacity ceiling can flatten upper sensitivity. Our default $5B / $10B caps do not bind the central calculators, so they do not explain the central gap. Our simulations also impose valuation bounds and a collapse branch. Radiant’s map clamps valuation tails. Such choices change the bandwidth even when central inputs look similar.
Growth. Giving Atlas’s base assumptions yield $121.45B annually in 2030 on our reproduction; keeping its other assumptions but setting all future growth to zero yields . That is a counterfactual calculation, not an alternative published forecast. The difference shows how a long-run curve can mostly reflect compounding assumptions rather than new donation evidence. Formula source.
Ransohoff explicitly discusses using “taste and good judgement.” That concerns allocation decisions. It also raises a useful modeling question: what qualifies as a fundable opportunity? A cap on responsible, cause-appropriate grants can be much lower than the capacity of the entire charitable sector. A preference for building new institutions rather than funding existing ones can therefore change an effective deployment constraint. This is a hypothesis to quantify, not evidence of hidden bias.
The safety sketch judges which giving qualifies as useful safety. That decision changes the target relative to an unrestricted charitable reservoir. Such definitional choices visibly change a model; they do not identify a personality effect.
A useful audit changes the question, the gate values, the correlations and the growth assumptions separately. Ideally, modelers would agree on a target before seeing one another’s answers, record defaults and revisions, and score fixed-date predictions against later observed cash flows. We have too few independent models and no matched realized outcomes here to estimate how much disagreement is caused by personality, anchoring, or optimism.
There is a concrete internal warning: our AIS/EA simulation gives a 43.25% chance of exceeding $1B by end-2027, while the separately stated judgment is 10–30%. Both remain as originally published. They are different objects and should not be presented as one calibrated belief. Labeling that distinction is more informative than treating a polished uncertainty dashboard as validated.
6 · A synthesis with explicit limits
A potentially large reservoir is the most consistent structural conclusion. The disputed step is conversion into grants, with deadlines and causes attached. Charitable designation, liquidation, a transfer to a regrantor, a grant to an implementer, and measured impact are successive outcomes; agreement at the first does not settle the later ones.
Matching donors and time exposes different gaps for different causes. Through 2027, the Anthropic-only central ratios are 3.44× for overall giving, 4.24× for combined AIS/EA with shared cause shares, and 7.35× for native GH&D shares. The first is an all-cause rescoping diagnostic; the second adds an allocation to Lewars. These are ranges of central assumptions, not probability intervals or bounds. They show that the GH&D gap includes extra allocation disagreement, while substantial upstream disagreement survives without it.
Broad annual giving in the tens of billions is conditional on a wider donor base and/or later deployment. Including the OpenAI Foundation creates an additional pathway absent from the live Radiant total and Lewars’s development sheet. Starting grants earlier, increasing cause fit, or compounding remaining assets can matter as much as an IPO valuation. It would be misleading to use that broad annual headline as the expected unrestricted funding available to an AIS or GH&D organization next year.
AI-safety and EA-aligned predictions require a further decision about what counts and how donors allocate. Our combined AIS/EA output is not safety alone. The allocation lens makes conditional estimates inspectable, but the missing cause share and its dependence on total giving remain real uncertainties. The safety-specific benchmark also counts existing funding, so it cannot be added to new IPO giving without a donor/flow overlap audit.
Numerical pooling would become more defensible with common-horizon distributions, non-overlapping donor cohorts, source-specific cash-versus-in-kind definitions, explicit dependence, and evidence about out-of-sample accuracy. Today, “average the models” would hide these missing decisions.
A verified cap table / employee DAF total would narrow the largest upstream disagreement. Actual lockup terms and sales would narrow availability. Cause-tagged grants to operating organizations would test allocation and payout assumptions. Auditable grants made, funds remaining and capacity growth would distinguish a delayed reservoir from a genuinely fast flow. These are more informative than another repetition of a high paper valuation.
7 · A separate experimental model
This small calculator combines separate donor cohorts, calendar liquidity dates, cause shares, a diminishing balance, and a Foundation ramp. It borrows mechanisms from the models above. Its defaults are our explicit illustrative choices, not fitted estimates, an update to either original fork, or a demonstrated “best” forecast.
| Calendar year | Anthropic people | OpenAI people / Foundation | Selected cause / all causes |
|---|
The Anthropic founder pool is valuation × 13% ownership × 80% pledge × follow-through × availability. The employee pool is a separately specified charitable pool at a $2T valuation, rescaled linearly by valuation and multiplied by availability. Availability means the fraction of paper wealth that can fund grants in the projection; lockup dates then control when that usable fraction starts paying out. Follow-through applies only to the non-binding founder pledge.
For each person cohort, years start when liquidity arrives. Grants equal remaining balance × [1 − (1 − payout rate)^fraction of year]. No new inflows, returns, tax-law calculations or changing cause shares are assumed. A $20B OpenAI-person charitable pool at a $1.85T reference valuation is an illustrative pathway informed by the live map’s scale, not a verified balance.
The Foundation starts with OpenAI valuation × 26% × availability; its annual all-cause grants are the smaller of balance-based payout and the first-year capacity × ramp^calendar years since opening, both prorated for a partial opening year. Pool accounting separates the Foundation stake from employee charitable holdings; employee matching must be net of overlaps. The calculator assumes those specified pools are disjoint, an assumption the public estimates do not verify.
Middle choices: $2T Anthropic; $1.85T OpenAI; $80B Anthropic employee charitable holdings at $2T; 60% founder follow-through; 70% availability; 10% annual payout; Anthropic liquidity July 2027, OpenAI November 2027; Foundation first-year capacity $3B and 2× yearly ramp. GH&D shares: 20% founders, 35% Anthropic employees, 20% OpenAI people, 10% Foundation. AIS/EA shares: 20%, 50%, 25%, 5%. The narrower AI-safety-only illustration uses 10%, 20%, 10%, and 2%. These shares are additional judgments, not observed allocations. OpenAI shares are also explicit added assumptions because the native map has no cause forecast.
The displayed scenario bandwidth spans the cautious, middle and faster presets at the selected cause, deadline and Foundation inclusion. Cautious: valuations $1T/$1.2T, employee pool $20B at $2T, follow-through 35%, availability 40%, payout 5%, later dates, $1B first-year Foundation capacity and 1.5× ramp. Faster: valuations $3T/$2.5T, employee pool $140B at $2T, follow-through 85%, availability 90%, payout 15%, earlier dates, $5B capacity and 3× ramp. Cause shares are held fixed across presets unless the reader changes them. Current slider edits are also included in the envelope. This is a coherent scenario envelope, not a percentile interval; the extrema do not cover all possible futures.
This synthesis retains Lewars’s cause decomposition, Radiant’s explicit dates and separate people cohorts, our realization / intent distinctions, Atlas’s capacity ramp, and the DAF model’s remaining-balance accounting. We set future returns to zero so a donation comparison is not silently dominated by speculative compound growth. The scenario choices are transparent judgment; no probabilistic weights are claimed.
8 · Verification and caveats
Our two calculators and 4,000-draw seeded simulations were rerun from their source files. A separate, in-memory all-cause copy retains the main fork’s upstream/timing draws while fixing cause allocation at 100% and removing its cause-specific output cap. No original file is edited by that diagnostic. Lewars’s v1.3 sheet was read in view-only mode, including its 2027 formula. Radiant’s public map was read node by node, together with its distribution documentation. Giving Atlas’s defaults and yearly calculation were inspected and reproduced. Webpages and spreadsheets can change; the numbers here refer to the dated snapshots in the ledger.
No author is assumed to have calibrated downstream philanthropy uncertainty merely because a model is interactive or uses Monte Carlo. None of these estimates measures the causal effect of AI on total giving, displacement of other donations, or ultimate welfare impact. Transfers through regrantors can count as “disbursed” before implementers receive them. Native definitions are retained; the available records do not support a perfect common operating-charity / cash-only adjustment.
9 · Sources and audit trail
Open the primary source from each entry. Hover, focus or tap an information marker elsewhere on the page for the relevant definition or source pointer.
Defaults and original code in src/content/modelDefaults.ts, src/utils/model.ts, globalDevelopmentModel.ts, and the two correlated-model files. AIS seed 73519; GH&D seed 91357; 4,000 draws each. No model inputs or tracker data changed for this comparison.
All-cause rescoping is reproduced by scripts/reproduce_comparison_scope.mjs, which transforms a diagnostic bundle in memory only. Central $2.951B; median $2.494B; P10–P90 $0.503–10.191B. This retains the original main model’s upstream/timing, collapse, valuation-bound, and local-confidence assumptions; it is not a validated general-philanthropy distribution.
Model tab: rows 8–20 inputs, 23–27 stocks, 31–38 annual / cumulative flows. Central C32 uses stock × availability ramp × cause share × payout. Bear/central/bull 2030 flows are $1.2/$8.8/$38.8B; they are scenarios without assigned probabilities.
Explains the donation gates and GH&D focus. The sheet, not rounded article prose, supplies the reproduced comparison.
Article says $25.7B over four years after lockups (and $24.5B later in the text). Live preview: $132B committed, $30.7B disbursed; P1/P99 $19.9–410B and $1.98–150B. The live formula instead ends mid-2031 and opens after the later company’s lockup. Both snapshots are distinguished.
total = OAI people + Anthropic employees + Anthropic founders; grants = total × clamp(annual rate × grant months / 12, 0, 1); grant months = max(60 − later lockup month since July 2026, 0). Payout prior P10/P90 2.2%/12.5%; employee giving prior 15%/32%; founder giving prior 8%/50%. Linked valuation-question cards displayed $2.09T Anthropic and $1.84T OpenAI, versus our $965B Anthropic default and Lewars’s $2T central liquidity valuation.
The founder-rate node says there is no public pledge to anchor it to, whereas the linked public literature reports an 80% pledge. Its effective giving-rate prior should not be equated with the pledge’s face percentage. Likewise, the map describes “committed” as irrevocable but includes promised sale proceeds; that wording does not itself verify a legally binding transfer.
The map uses a common opening date after both lockups, rather than starting each cohort separately. Its disbursement formula uses a linear fraction of the initial pool, capped at 100%, instead of annual depletion of the remaining balance. Those are observable model choices.
Directional $370B reservoir: approximately $220B Foundation, $90B Anthropic founders and $60B employees. At 10% intended annual spending, $37B/year; larger valuations / payout lead toward $100B. Explicitly distinguishes target spending from actual deployment. These are not fixed-horizon probability bounds.
Stock $641B; 2027 flow $32.4B; reproduced 2030 flow $121.45B; 2027–46 cumulative $4,328B. Valuations $2T/$850B; individuals spend down over 20 years; Foundation starts at $3B, ramps 3× to a 10% payout ceiling. Growth starts at 40%, decays toward 7% with a three-year half-life. Source labels it illustrative.
Headline method: roughly 16% of assumed $75–200B IPO proceeds becomes $12–32B in DAF contributions; rough annual grant capacity $3–8B. Includes SpaceX, not just AI labs. Its embedded employee calculator is a separate method; it should not be confused with the headline. Historical aggregate ratios do not identify a causal IPO effect.
Judgment estimates: $1.6B/year in 2026 and lifetime present value above $100B. Includes existing funding and investors; assumes roughly 7% of Anthropic value reaches qualifying safety work. No quantified interval or matched grant deadline.
Employee estimate $20–40B, combined pre-IPO DAF estimate $22–48B. Useful for tracing the lineage of later calculations. These are estimated cohort / participation inputs, not verified audited DAF balances or a dated grant forecast. Possible overlap with founder equity is acknowledged.
We followed both supplied posts, their calculators and upstream references, and searched for other public AI-windfall philanthropy models. Giving Atlas, Ransohoff’s widely discussed estimate, and the Whole Whale DAF model add distinct mechanisms or prominent headline estimates. LongtermWiki is included as upstream context. A targeted safety-funding search added the LessWrong sketch as a different-scope benchmark. Financing proposals mixing taxes, public budgets, and charitable capital are not treated as IPO-donation forecasts. This is not an exhaustive census. A quoted “GWWC $15B/year” claim was not used as a separate model: the inspected references did not establish a primary annual forecast with that definition.
For a standalone LessWrong post, use the title Why AI-philanthropy models disagree: matching the outcome before averaging forecasts. Add: “The most useful part for me was distinguishing uncertainty within a model from disagreement over its structure. The central gap changes with the target: roughly threefold for overall giving and sevenfold for native GH&D, while the uncertainty bands have different meanings and the models share much of their evidence.”