LongtermWiki Impact Model

📋Page Status

Page Type:ContentStyle Guide →Standard knowledge base article

Quality:55 (Adequate)

Importance:50 (Useful)

Last edited:2026-02-04 (2 days ago)

Words:2.2k

Structure:

📊 18📈 1🔗 5📚 0•6%Score: 11/15

LLM Summary:Fermi estimation of LongtermWiki value grounded in base rates. GiveWell shows 3% of donors choose based on effectiveness research; 80k Hours achieves ~107 significant plan changes/year; think tanks rarely demonstrate causal policy impact. Conservative central estimate: $100-500K/yr in effective value through researcher onboarding (primary), funder information improvement (secondary), with high-variance 'inspiration' pathway that's hard to quantify. Much lower than naive estimates due to: limited counterfactual impact of information on decisions, small target audience, and low probability of behavioral change.

InfoBox requires type prop or entityId/expertId/orgId for data lookup

Overview

LongtermWiki is an AI-assisted knowledge base covering AI safety, longtermist prioritization, and related topics. This model attempts to estimate its potential value creation using base rates from comparable interventions rather than inside-view reasoning.

Core Question: How much value does LongtermWiki create, and how does this compare to alternative uses of resources?

Key Finding: Naive estimates that assume “better information → better decisions” dramatically overstate impact. Base rates from GiveWell, 80k Hours, and think tank research suggest information interventions change behavior far less than expected.

Quick Assessment

Dimension	Assessment	Evidence
Central Value Estimate	$100-500K/yr	Base-rate-grounded Fermi model
Primary Pathway	Researcher onboarding	Clearest counterfactual impact
Secondary Pathway	Funder information improvement	Low confidence in behavioral change
High-Variance Pathway	”Inspiration” for epistemic infrastructure	Too speculative to quantify
Cost-Effectiveness	$0.25-1.25/$ vs typical interventions	Uncertain but plausibly positive

Base Rates: What We Know About Information Interventions

GiveWell Data on Donor Behavior

GiveWell provides the clearest data on whether effectiveness research changes funder behavior:

Metric	Value	Source
Donors who “choose based on effectiveness research”	3%	GiveWell surveys
Donors aware of GiveWell recommendations	10.1%	GiveWell awareness studies
Conversion rate: awareness → action	≈30%	Implied
Total giving influenced by GiveWell	≈$500M/yr	GiveWell reports
GiveWell operating budget	≈$30M/yr	990 filings

Key insight: Even the most successful effectiveness research organization achieves only 3% behavioral change among its target audience, and this after 15+ years of operation with significant resources.

Application to LongtermWiki: If LongtermWiki achieved GiveWell-level penetration in the AI safety funding space (≈$300M/yr), it might influence 3% × $300M = $9M/yr in decisions. But LongtermWiki is not GiveWell-level quality or reach, so realistic penetration is likely 0.1-1%, suggesting $300K-3M/yr in influenced decisions (not improved decisions).

80,000 Hours Data on Plan Changes

80,000 Hours tracks “significant plan changes” attributable to their advice:

Metric	Value	Source
Significant plan changes/year	≈107	80k Hours impact reports
Definition of “significant”	≥20% career shift attributable to 80k	Self-reported
Users engaging with content	≈100,000/yr	Traffic estimates
Conversion rate: engagement → plan change	≈0.1%	Implied
Operating budget	≈$4M/yr	990 filings
Cost per plan change	≈$40K	Budget / changes

Key insight: Even excellent career advice changes behavior in only 0.1% of readers. Self-reported attribution is likely inflated.

Application to LongtermWiki: If LongtermWiki had 10,000 engaged users/year and achieved 80k-level conversion, that’s ~10 “significant” decision changes. At $50K average impact per decision change, that’s $500K/yr. But LongtermWiki likely has fewer users and lower conversion.

Think Tank Policy Influence

Research on think tank policy influence is sobering:

Finding	Source
77% of think tanks claim policy influence, but causal evidence is weak	McGann (2019)
Policy changes attributed to specific research are rare and hard to verify	Rich (2004)
Think tank influence is mediated by relationships, not publications	Abelson (2009)
Congressional staff report using think tank research for “ammunition” not “education”	Weiss (1991)

Key insight: Think tanks provide legitimacy and talking points, not decision-driving analysis. Policymakers who cite research were usually already inclined toward that position.

Application to LongtermWiki: Policy influence pathway should be discounted heavily. LongtermWiki might provide “ammunition” for already-aligned actors but is unlikely to change minds.

Open Philanthropy Decision-Making

Open Philanthropy, the largest AI safety funder, describes its decision-making process:

Aspect	Reality
Primary input	Program officer judgment and relationships
Role of external research	Supplementary, not determinative
Decision style	”Hits-based” with heavy reliance on internal worldviews
Response to external analysis	May inform but rarely drives decisions

Key insight: Major funders rely on internal expertise and relationships, not external knowledge bases. Even excellent external analysis is filtered through existing worldviews.

Application to LongtermWiki: Direct influence on major funder decisions is likely small. Value more likely comes from indirect channels: improving researcher quality, providing shared vocabulary, etc.

Fermi Model: Conservative Estimates

Pathway 1: Researcher Onboarding (Primary)

This is likely LongtermWiki’s clearest counterfactual impact.

Parameter	Estimate	Reasoning
New AI safety researchers/year	≈200	Field growth estimates
Current onboarding time	6-12 months to productivity	Researcher interviews
LongtermWiki’s reach	20-40% of new researchers	Optimistic given competition with AI Safety Fundamentals, Alignment Forum
Time reduction if used	1-2 months	Modest improvement, not transformation
Researchers actually affected	40-80	200 × 30% reach
Time saved per researcher	1.5 months	Midpoint
Value of researcher time	$8K/month	Junior researcher cost
Total time value	$480K-960K/yr	40-80 × 1.5 × $8K
Counterfactual adjustment	50%	Would partially upskill anyway
Net value	$240K-480K/yr	After counterfactual

Confidence: Medium. This pathway has clearer counterfactual than others.

Pathway 2: Funder Information Improvement (Secondary)

Parameter	Estimate	Reasoning
AI safety funding/year	$300M	Field estimates
Funders who might use LongtermWiki	10-20%	Optimistic
Funding “influenced”	$30-60M/yr	300M × 15%
Base rate: information → behavior change	3%	GiveWell data
Decisions actually changed	$1-2M/yr	30-60M × 3%
Quality of change	20% improvement	Modest, not dramatic
Net value	$200K-400K/yr	Decisions changed × improvement

Confidence: Low. The “information → behavior” chain is weak based on base rates.

Pathway 3: Field Coordination & Vocabulary

Parameter	Estimate	Reasoning
Value of shared vocabulary	Hard to quantify	Enables better disagreement
Reduced duplication	$50-100K/yr equivalent	Some analyst time saved
Better gap identification	$50-100K/yr equivalent	Maybe one project better targeted
Net value	$100-200K/yr	Highly uncertain

Confidence: Very low. Real but hard to measure.

Pathway 4: Policy/Government Influence

Parameter	Estimate	Reasoning
P(LongtermWiki cited in policy)	5-15%	Low penetration expected
P(citation influences decision)	5-10%	Base rates suggest low
P(influenced decision is good)	60-70%	Some net improvement if any
Expected policy value	$0-100K/yr	Very low base rates
Net value	$0-100K/yr	Essentially speculative

Confidence: Very low. Think tank research suggests minimal causal impact.

Pathway 5: “Inspiration” for Epistemic Infrastructure

This is the highest-variance pathway but hardest to estimate.

Parameter	Estimate	Reasoning
P(relevant person sees LongtermWiki)	20-40%	If we actively promote
P(they find it compelling)	10-30%	Quality-dependent
P(it influences their decisions)	5-15%	Idea may already exist independently
P(resulting action is valuable)	50-70%	Uncertain what “inspired” action looks like
Conditional value if chain completes	$5-50M+	Wide range
Expected value	Highly uncertain	Too many conjunctive probabilities

Confidence: Cannot reliably estimate. Include in sensitivity analysis but not central estimate.

Aggregate Conservative Estimate

Pathway	Low	Central	High	Confidence
Researcher onboarding	$100K	$300K	$600K	Medium
Funder information	$50K	$200K	$500K	Low
Field coordination	$25K	$100K	$300K	Very Low
Policy influence	$0	$25K	$100K	Very Low
Subtotal (quantifiable)	$175K	$625K	$1.5M	Low-Medium
Inspiration pathway	???	???	???	Unquantifiable

Central estimate: $100-500K/yr after accounting for uncertainty and optimism bias.

Impact Pathway Diagram

Loading diagram...

Why Naive Estimates Are Wrong

The value proposition document suggested $15-40M/yr central estimate. Why is this Fermi model 50-100x lower?

Error 1: Assuming Information Changes Behavior

Naive assumption	Reality (base rates)
“If funders have better info, they’ll make better decisions”	3% of donors change behavior based on effectiveness research (GiveWell)
“Policy staff will use our analysis”	Think tanks rarely demonstrate causal policy impact
”Researchers will use our onboarding materials”	0.1% conversion rate for career advice (80k Hours)

Error 2: Ignoring Counterfactuals

Claim	Counterfactual question
”LongtermWiki saves researcher time”	Would they not learn this from other sources?
”LongtermWiki enables better funder decisions”	Are funders actually information-constrained?
”LongtermWiki creates shared vocabulary”	Does the Alignment Forum already serve this function?

Most benefits are partially counterfactual—the impact would occur through other channels without LongtermWiki.

Error 3: Optimistic Probability Stacking

Pathway	Probability chain
Anthropic “inspiration”	P(sees) × P(compelled) × P(acts) × P(valuable) ≈ 0.5-5%
Policy influence	P(reaches policy) × P(influences decision) × P(good decision) ≈ 0.2-1%

Multiplying many uncertain probabilities yields very low expected values, even with high conditional values.

Error 4: Conflating “Influenced” with “Improved”

Even if LongtermWiki influences $30M in decisions:

Most “influenced” decisions were already trending that direction
Influence doesn’t mean improvement (could be neutral or negative)
Measurement is confounded (users seek confirmation, not education)

Comparison to Alternative Interventions

Intervention	Annual Cost	E[Impact]	Impact/$
LongtermWiki (2 FTE)	$400K	$100-500K	$0.25-1.25
GiveWell operations	$30M	$500M influenced	≈$17
80,000 Hours	$4M	≈$5M value of plan changes	≈$1.25
Direct AI safety research (per researcher)	$200K	$0.5-2M	$2.50-10
Grantmaking (per $ moved)	$0.05-0.10	$1 moved	$10-20 leverage

Interpretation: LongtermWiki’s cost-effectiveness is uncertain but plausibly competitive with other information interventions. It is likely less cost-effective than direct research or grantmaking if those options are available.

Key Cruxes

Crux	If True → Impact	If False → Impact	Current Belief
Information changes funder behavior	$500K+/yr from funder pathway	$100K/yr (mostly onboarding)	20% true
LongtermWiki is unique resource	Higher counterfactual value	Lower (other sources substitute)	40% true
”Inspiration” pathway is real	Could be $1M+	Negligible	10-20% real
Quality can be maintained	Sustained value	Value decays over 2-3 years	50% maintainable
AI safety is information-constrained	Information interventions valuable	Resources better spent elsewhere	30% constrained

Model Limitations

Self-assessment bias: This model is produced by LongtermWiki, creating incentive to underestimate (for credibility) or overestimate (for motivation)
Base rate generalization: GiveWell/80k Hours may not transfer to AI safety funding context
Unmeasurable pathways: “Inspiration” and “coordination” benefits are real but hard to quantify
Temporal dynamics: Value may be front-loaded (early field benefits most) or back-loaded (compounding effects)
Reference class selection: Different reference classes (encyclopedia, think tank, community wiki) yield different estimates

What Would Change This Estimate

Toward Higher Impact

Evidence	Implication
Funders report using LongtermWiki in actual decisions	Direct behavioral change
Significant user growth beyond AI safety community	Broader reach
Demonstrated policy citations	Policy pathway becomes real
Anthropic or similar org builds on the concept	”Inspiration” pathway validates

Toward Lower Impact

Evidence	Implication
User research shows researchers prefer other resources	Onboarding pathway weakens
Funder interviews show no behavior change	Funder pathway essentially zero
Content quality degrades	All pathways weaken
Better alternatives emerge	Counterfactual value drops

Recommendations

Given this analysis:

Primary focus should be researcher onboarding — this has clearest counterfactual impact
Funder influence claims should be modest — base rates suggest limited behavioral change
Policy pathway should be de-prioritized — unless strong relationships exist
“Inspiration” pathway is worth trying — but shouldn’t drive resource allocation
Track actual behavior change, not just usage — pageviews don’t equal impact