LongtermWiki Impact Model

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

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:

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:

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:

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.

Coefficient Giving Decision-Making

Coefficient Giving, the largest AI safety funder, describes its decision-making process:

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.

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

Pathway 2: Funder Information Improvement (Secondary)

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

Pathway 3: Field Coordination & Vocabulary

Confidence: Very low. Real but hard to measure.

Pathway 4: Policy/Government Influence

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.

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

Aggregate Conservative Estimate

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

Impact Pathway 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

Error 2: Ignoring Counterfactuals

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

Error 3: Optimistic Probability Stacking

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

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

Model Limitations

1. Self-assessment bias: This model is produced by LongtermWiki, creating incentive to underestimate (for credibility) or overestimate (for motivation)

2. Base rate generalization: GiveWell/80k Hours may not transfer to AI safety funding context

3. Unmeasurable pathways: "Inspiration" and "coordination" benefits are real but hard to quantify

4. Temporal dynamics: Value may be front-loaded (early field benefits most) or back-loaded (compounding effects)

5. Reference class selection: Different reference classes (encyclopedia, think tank, community wiki) yield different estimates

What Would Change This Estimate

Toward Higher Impact

Toward Lower Impact

Recommendations

Given this analysis:

1. Primary focus should be researcher onboarding — this has clearest counterfactual impact

2. Funder influence claims should be modest — base rates suggest limited behavioral change

3. Policy pathway should be de-prioritized — unless strong relationships exist

4. "Inspiration" pathway is worth trying — but shouldn't drive resource allocation

5. Track actual behavior change, not just usage — pageviews don't equal impact