LongtermWiki Vision

LongtermWiki: Comprehensive AI Impact & Risk Navigator

Vision Document (2-Person-Year Scope)

Version: 0.1 Draft Last Updated: 2025-01-13 Team Size: 3-4 people over ~6-8 months

---

Executive Summary

LongtermWiki is a strategic intelligence platform for AI safety prioritization. Its core purpose is to surface the key uncertainties and cruxes that, if resolved, would most change how resources should be allocated across AI safety interventions.

This is not an encyclopedia. It’s a decision-support tool for funders, researchers, and policymakers asking: “Where should the next marginal dollar or researcher-hour go?”

---

The Problem

The AI safety field suffers from:

1. Fragmented knowledge — Insights are scattered across papers, blog posts, forum threads, and institutional knowledge
2. Unclear cruxes — People disagree but often don’t know why they disagree or what evidence would change their minds
3. Poor prioritization legibility — It’s hard to see which interventions depend on which assumptions
4. Slow information synthesis — New developments take months to propagate into strategic thinking

flowchart TD
    subgraph Problem["Current State"]
        A[Scattered Knowledge] --> D[Poor Prioritization]
        B[Hidden Cruxes] --> D
        C[Slow Synthesis] --> D
        D --> E[Suboptimal Resource Allocation]
    end

    subgraph Solution["LongtermWiki"]
        F[Structured Knowledge Graph] --> I[Clear Strategic Priorities]
        G[Explicit Crux Mapping] --> I
        H[Living Document System] --> I
        I --> J[Better Marginal Decisions]
    end

    E -.->|"LongtermWiki bridges this"| F

---

Core Value Proposition

LongtermWiki provides strategic clarity by answering:

Question	How LongtermWiki Helps
”What are the key uncertainties in AI safety?”	Structured crux taxonomy with explicit dependencies
”If I believe X, what should I prioritize?”	Worldview → intervention mapping
”What would change my mind about Y?”	Explicit operationalization of cruxes
”Where do experts disagree and why?”	Disagreement decomposition into factual claims
”What’s the current state of Z?”	Living knowledge base with staleness tracking

---

Architecture

Information Flow

flowchart LR
    subgraph Sources["Input Sources"]
        S1[Papers & Reports]
        S2[Expert Interviews]
        S3[Prediction Markets]
        S4[Field Developments]
    end

    subgraph LongtermWiki["LongtermWiki Core"]
        direction TB
        K[Knowledge Base]
        C[Crux Graph]
        M[Causal Models]
        I[Intervention Map]

        K <--> C
        C <--> M
        M <--> I
    end

    subgraph Outputs["Strategic Outputs"]
        O1[Priority Rankings]
        O2[Crux Reports]
        O3[Worldview Analyses]
        O4[Disagreement Maps]
    end

    Sources --> LongtermWiki
    LongtermWiki --> Outputs

Content Layers

The system has four interconnected layers:

flowchart TB
    subgraph L1["Layer 1: Factual Foundation"]
        F1[Risks & Threat Models]
        F2[Interventions & Responses]
        F3[Actors & Institutions]
        F4[Technical Concepts]
    end

    subgraph L2["Layer 2: Causal Models"]
        M1[Risk Pathway Models]
        M2[Intervention Effect Models]
        M3[Transition Dynamics Models]
    end

    subgraph L3["Layer 3: Uncertainty Structure"]
        U1[Key Cruxes]
        U2[Empirical Uncertainties]
        U3[Value Uncertainties]
        U4[Worldview Clusters]
    end

    subgraph L4["Layer 4: Strategic Implications"]
        S1[Priority Rankings by Worldview]
        S2[Robust Interventions]
        S3[High-VOI Research Questions]
    end

    L1 --> L2
    L2 --> L3
    L3 --> L4

    style L4 fill:#e8f5e9

---

Core Components (2-Person-Year Scope)

1. Knowledge Base (~30% of effort)

Goal: Comprehensive, structured coverage of AI safety-relevant concepts, risks, and interventions.

Scope:

• ~50 risk pages (technical, misuse, structural, epistemic)
• ~80 intervention/response pages
• ~40 causal model pages
• Cross-linked, consistently formatted

Quality bar: Each page should have:

• Clear definition and scope
• Key claims with uncertainty estimates
• Links to primary sources
• Cross-references to related concepts
• Last-reviewed date and staleness tracking

2. Crux Graph (~25% of effort)

Goal: Explicit mapping of the key uncertainties that drive disagreement and prioritization.

Scope:

• ~30-50 major cruxes identified and operationalized
• Dependency structure (which cruxes affect which)
• Links to evidence and expert positions
• “What would change my mind” for each

Example cruxes:

• P(deceptive alignment) given current training approaches
• Timelines to transformative AI
• Tractability of interpretability research
• Likelihood of warning shots before catastrophe
• Value of current governance interventions

flowchart TD
    subgraph Cruxes["Key Cruxes"]
        C1[AI Timelines]
        C2[Deceptive Alignment Risk]
        C3[Warning Shot Likelihood]
        C4[Governance Tractability]
        C5[Interpretability Tractability]
    end

    subgraph Priorities["Priority Implications"]
        P1[Technical Safety Research]
        P2[Governance Work]
        P3[Field Building]
        P4[Capabilities Work at Labs]
    end

    C1 -->|"Short → more urgent"| P1
    C1 -->|"Short → less time for"| P2
    C2 -->|"High → prioritize"| P1
    C2 -->|"High → prioritize"| C5
    C3 -->|"Unlikely → prepare now"| P2
    C4 -->|"High → prioritize"| P2
    C5 -->|"High → prioritize"| P1

    style C1 fill:#fff3e0
    style C2 fill:#fff3e0
    style C3 fill:#fff3e0
    style C4 fill:#fff3e0
    style C5 fill:#fff3e0

3. Worldview → Priority Mapping (~20% of effort)

Goal: Show how different assumptions lead to different prioritizations.

Approach:

1. Define 4-6 “worldview archetypes” based on crux positions
2. For each worldview, show implied priority rankings
3. Identify “robust” interventions that score well across worldviews
4. Identify “worldview-specific” bets

Example worldviews:

• Short-timelines technical doomer: P(doom) > 50%, TAI < 2030, deceptive alignment likely
• Governance optimist: Institutions can adapt, warning shots likely, coordination tractable
• Slow takeoff pragmatist: Long transition period, many opportunities to course-correct
• Multipolar risk-focused: Concentration of power is the main risk, not misalignment

4. Disagreement Decomposition (~15% of effort)

Goal: Turn fuzzy disagreements into structured, resolvable questions.

Process:

1. Identify high-stakes disagreements (e.g., “Is current safety research useful?”)
2. Decompose into component claims
3. Identify which claims are cruxes vs. downstream disagreements
4. Link to evidence for each claim

5. Living Document Infrastructure (~10% of effort)

Goal: Keep content fresh and trustworthy.

Features:

• Staleness tracking (days since review, triggered updates)
• Source freshness (flag when cited papers are superseded)
• Confidence decay (uncertainties widen over time without review)
• Contributor attribution

---

Non-Goals (Out of Scope for 2-Person-Year)

Feature	Why Excluded
Original research	We synthesize, not generate
Real-time monitoring	Quarterly update cadence is sufficient
Quantitative forecasting	Link to Metaculus/prediction markets instead
Community features	Focus on content, not social
Comprehensive AI news	Not a news aggregator
Deep technical tutorials	Link to AI Safety Fundamentals, etc.

---

Success Metrics

Primary Metrics

Metric	Target	Measurement
Crux coverage	80% of major cruxes in discourse	Expert survey
User utility	”Changed my prioritization”	User survey
Citation rate	Referenced in 10+ strategy docs/year	Manual tracking
Expert endorsement	5+ senior researchers recommend	Testimonials

Secondary Metrics

• Pages maintained at quality ≥4: >80%
• Average page staleness: under 60 days
• Cross-linking density: >5 links per page
• Coverage completeness: >90% of standard risk taxonomies

---

Team Structure (3-4 People)

flowchart LR
    subgraph Core["Core Team"]
        L[Lead / Editor<br/>1.0 FTE]
        R1[Research Analyst<br/>0.5-1.0 FTE]
        R2[Research Analyst<br/>0.5-1.0 FTE]
    end

    subgraph Support["Support"]
        T[Technical / Dev<br/>0.25 FTE]
    end

    L --> R1
    L --> R2
    T -.-> Core

Roles:

• Lead/Editor (1.0 FTE): Overall vision, quality control, crux identification, stakeholder relationships
• Research Analysts (1.0-1.5 FTE combined): Page writing, source synthesis, model building
• Technical (0.25 FTE): Site maintenance, tooling improvements, data pipeline

---

Milestones

Phase 1: Foundation (Months 1-2)

• Core knowledge base structure complete
• 30 high-priority pages at quality ≥4
• Initial crux taxonomy (15-20 cruxes)
• Basic worldview mapping

Phase 2: Depth (Months 3-4)

• 80+ pages at quality ≥4
• Full crux graph with dependencies
• 4-6 worldview archetypes defined
• First “disagreement decomposition” case studies

Phase 3: Polish & Launch (Months 5-6)

• All core pages at quality ≥4
• Interactive worldview → priority tool
• Expert review and feedback incorporated
• Public launch

Phase 4: Maintenance Mode (Months 7-8+)

• Quarterly review cycle established
• Community contribution guidelines
• Integration with other resources (AI Safety Fundamentals, etc.)

---

Key Risks & Mitigations

Risk	Likelihood	Impact	Mitigation
Scope creep	High	Medium	Strict non-goals, regular pruning
Staleness	Medium	High	Automated tracking, review calendar
Low adoption	Medium	High	Early stakeholder involvement, utility focus
Quality inconsistency	Medium	Medium	Style guide, editor review
Key person dependency	Medium	High	Documentation, cross-training

---

Open Questions

1. Governance structure: Who has editorial authority? How are disagreements resolved?
2. Funding model: Grant-funded? Part of existing org? Independent?
3. Expert involvement: Advisory board? Paid reviewers? Community contribution?
4. Update cadence: Quarterly? Event-driven? Continuous?
5. Quantitative integration: How tightly to integrate with forecasting platforms?

---

Appendix: Comparison to Existing Resources

Resource	LongtermWiki Differentiator
AI Safety Fundamentals	LongtermWiki is strategic, not educational
LessWrong/AF	LongtermWiki is curated synthesis, not discussion
80K Problem Profiles	LongtermWiki goes deeper on cruxes and uncertainties
GovAI/CAIS research	LongtermWiki synthesizes across orgs, not original research
Wikipedia	LongtermWiki is opinionated about importance and uncertainty

---

Next Steps

1. Circulate this document for feedback
2. Identify potential funders and home organizations
3. Recruit initial team
4. Develop detailed Phase 1 workplan