Longterm Wiki

📋Page Status

Page Type:ResponseStyle Guide →Intervention/response page

Quality:63 (Good)⚠️

Importance:12 (Peripheral)

Words:2.2k

Structure:

📊 24📈 7🔗 24📚 13•4%Score: 15/15

LLM Summary:A self-referential documentation page describing the Longterm Wiki platform itself—a strategic intelligence tool with ~550 pages, crux mapping of ~50 uncertainties, and quality scoring across 6 dimensions. Features include entity cross-linking, interactive causal diagrams, and structured YAML databases tracking expert positions on key AI safety cruxes.

Issues (2):

• QualityRated 63 but structure suggests 100 (underrated by 37 points)
• Links4 links could use <R> components

Quick Assessment

Dimension	Assessment	Evidence
Scope	AI safety focused	≈550 pages covering risks, interventions, organizations, cruxes
Content Model	Curated synthesis	Editorial control, quality scoring, not community wiki
Unique Value	Crux mapping	Explicit uncertainty tracking, worldview→priority linkages
Technical	Modern stack	Astro/Starlight, React, interactive causal diagrams
Open Source	Fully	MIT licensed, GitHub repository
Status	Active development	Launched 2025, ongoing content expansion

Project Details

Attribute	Details
Name	Longterm Wiki (LongtermWiki)
Organization	QURIOrganizationQURI (Quantified Uncertainty Research Institute)QURI develops Squiggle (probabilistic programming language with native distribution types), SquiggleAI (Claude-powered model generation producing 100-500 line models), Metaforecast (aggregating 2,1...Quality: 48/100 (Quantified Uncertainty Research Institute)
Lead	Ozzie Gooen
Launched	2025
Website	longtermwiki.com
GitHub	github.com/quantified-uncertainty/cairn
License	MIT
Platform	Static site (Astro/Starlight)

Overview

The Longterm Wiki is a strategic intelligence platform for AI safety prioritization. Unlike general encyclopedias or community wikis, it serves as a decision-support tool for funders, researchers, and policymakers asking: “Where should the next marginal dollar or researcher-hour go?”

Loading diagram...

The project addresses four problems in the AI safety field:

Problem	How the Wiki Addresses It
Fragmented knowledge	Consolidated, cross-linked knowledge base with ≈550 pages
Unclear cruxes	Explicit mapping of key uncertainties and expert disagreements
Poor prioritization legibility	Worldview → intervention mapping showing how assumptions lead to priorities
Slow information synthesis	Quality tracking, staleness monitoring, regular updates

Design Philosophy

The wiki is deliberately opinionated about importance and uncertainty—it rates content quality, tracks expert positions on cruxes, and makes prioritization implications explicit. This distinguishes it from neutral reference works like Wikipedia or discussion platforms like LessWrong.

Content is editorially curated rather than community-contributed, ensuring consistency and quality control. Each page goes through a grading pipeline that scores novelty, rigor, actionability, and completeness.

Content Architecture

The wiki has four interconnected layers of content:

Loading diagram...

Layer	Contents	Scale
Factual Foundation	Risks, interventions, organizations, people, concepts	≈350 pages
Causal Models	Risk pathways, intervention effects, transition dynamics	≈80 models
Uncertainty Structure	Key cruxes, expert positions, worldview clusters	≈50 cruxes, ~30 experts
Strategic Implications	Priority rankings, robust interventions, high-VOI research	Derived from above

Major Sections

Section	Content	Page Count	Example Pages
Knowledge Base	Risks, interventions, organizations, people	≈350	Deceptive AlignmentRiskDeceptive AlignmentComprehensive analysis of deceptive alignment risk where AI systems appear aligned during training but pursue different goals when deployed. Expert probability estimates range 5-90%, with key empir...Quality: 75/100, AI Safety InstitutesPolicyAI Safety Institutes (AISIs)Analysis of government AI Safety Institutes finding they've achieved rapid institutional growth (UK: 0→100+ staff in 18 months) and secured pre-deployment access to frontier models, but face critic...Quality: 69/100
AI Transition Model	Comprehensive factor network with outcomes and scenarios	≈80	Factors, scenarios, quantitative estimates
Analytical Models	Risk models, cascade models, governance dynamics	≈80	Cascade models, dynamics models
Project	Vision, strategy, similar projects analysis	≈10	Vision, Similar Projects
Internal	Style guides, automation tools, technical docs	≈30	Automation Tools

Content Types

Type	Purpose	Quality Scored?	Validation
Content	Substantive knowledge base pages	Yes (0-100 scale)	Full (schema, links, escaping, quality)
Stub	Redirects, brief profiles	No	Minimal (compilation, links)
Documentation	Style guides, internal reference	No	Compilation only
Overview	Navigation hub pages (index.mdx)	No	Sidebar structure

Key Features

Crux Mapping System

The wiki explicitly tracks key uncertainties that drive disagreement and prioritization:

Loading diagram...

Crux Category	Examples	Tracked In
Technical	P(deceptive alignment), interpretability tractability, mesa-optimization risk	cruxes.yaml
Timeline	Years to transformative AI, takeoff speed, capability jumps	estimates.yaml
Governance	Coordination tractability, warning shot likelihood, regulatory capacity	cruxes.yaml
Value	Weight on near-term vs. long-term, risk tolerance, discount rates	experts.yaml positions

Expert positions on cruxes are tracked in structured YAML databases, enabling:

• Disagreement decomposition: Breaking “Is alignment hard?” into component empirical claims
• Worldview clustering: Grouping experts by crux positions (e.g., “short-timeline technical doomer”)
• Priority implications: Showing how different assumptions lead to different resource allocations

Entity Cross-Linking

Stable ID-based linking system that survives URL reorganization:

import {EntityLink} from '@components/wiki';

The <EntityLink id="scheming">scheming</EntityLink> risk relates to
<EntityLink id="deceptive-alignment">deceptive alignment</EntityLink>.

Feature	Benefit
Stable IDs	Links survive path reorganization
Auto title lookup	Component fetches current title from database
Backlink tracking	Every page shows what links to it
CI validation	Broken links caught before deployment
Type icons	Visual distinction between risks, responses, orgs

Interactive Causal Diagrams

ReactFlow-based cause-effect graphs for complex causal models:

Loading diagram...

Node Type	Purpose	Visual
Leaf	Input factors	Light blue
Intermediate	Mediating factors	White
Effect	Outcomes	Green/Red
Group	Clusters of related nodes	Dashed border

Quality Scoring System

Loading diagram...

Six-dimension rating system (0-10 scale, harsh—7+ is exceptional):

Dimension	What It Measures	Weight by Page Type
Focus	Does it answer the title’s promise?	All types
Novelty	Value beyond obvious sources	Higher for analysis
Rigor	Evidence quality and precision	Higher for reference
Completeness	Thorough coverage of claimed topic	Higher for reference
Concreteness	Specific vs. abstract recommendations	Higher for analysis
Actionability	Can readers make different decisions?	Higher for analysis

Quality is set only through automated grading pipeline using Claude, never manually. The quality-source validation rule enforces this.

Data Layer

Loading diagram...

Structured YAML databases enable data-aware components:

Database	Contents	Size	Example Usage
experts.yaml	AI safety researchers with positions on cruxes	≈30 entries	<DataInfoBox expertId="paul-christiano" />
organizations.yaml	Labs, research orgs, funders	≈15 entries	<DataInfoBox orgId="anthropic" />
cruxes.yaml	Key uncertainties with expert positions	≈50 entries	<DataCrux dataId="deceptive-alignment-likelihood" />
estimates.yaml	Probability distributions for key variables	≈35 entries	<DataEstimateBox dataId="p-doom" />
external-links.yaml	Curated external resources with metadata	≈200 entries	<R id="hash">Link text</R>

Technical Architecture

Stack

Layer	Technology	Purpose
Framework	Astro 5 + Starlight	Static site generation, documentation theme
Components	React 19	Interactive UI components
Styling	Tailwind CSS 4	Utility-first styling
Type Safety	TypeScript + Zod	Compile-time and runtime validation
Graphs	ReactFlow + Dagre/ELK	Interactive causal diagrams
Diagrams	Mermaid 11	Declarative flowcharts
Math	KaTeX	LaTeX rendering
Data	YAML → JSON	Structured entity databases
UI Components	shadcn/ui + Radix	Accessible component primitives

Automation CLI

Unified CLI (crux) provides comprehensive tooling:

npm run crux -- --help           # Show all domains
npm run crux -- validate         # Run all validators
npm run crux -- analyze          # Analysis and reporting
npm run crux -- fix              # Auto-fix common issues
npm run crux -- content          # Page management
npm run crux -- generate         # Content generation

Domain	Commands	Example
Validation	compile, frontmatter, entity-links, escaping, mermaid	crux validate compile --quick
Analysis	mentions, entity-links, quality, redundancy	crux analyze entity-links sam-altman
Content	grade, improve, regrade, create	crux content grade --page scheming
Fixes	escaping, entity-links, markdown	crux fix escaping
Generate	summaries, content, diagrams	crux generate summaries --batch 50

Validation Pipeline

Rule	What It Checks	Failure Mode
compile	MDX syntax and React components	Build failure
frontmatter-schema	YAML frontmatter against Zod schema	CI failure
dollar-signs	LaTeX escaping (\$100 not $100)	Rendering issues
comparison-operators	JSX escaping (\<100ms not <100ms)	Build failure
entitylink-ids	All EntityLink references exist in database	Broken links
quality-source	Quality set by pipeline, not manually	Data integrity
mermaid	Diagram syntax validation	Rendering issues
markdown-lists	Consistent list formatting	Markdown errors

Use Cases

For Funders

Use Case	Wiki Feature	Example
Intervention comparison	Side-by-side pages with explicit uncertainty	Compare interpretabilitySafety AgendaInterpretabilityMechanistic interpretability has extracted 34M+ interpretable features from Claude 3 Sonnet with 90% automated labeling accuracy and demonstrated 75-85% success in causal validation, though less th...Quality: 66/100 vs governanceGovernanceThis is a placeholder page with no actual content - only component imports that would render data from elsewhere in the system. Cannot assess importance or quality without the underlying content. approaches
Crux identification	Crux mapping shows which uncertainties matter most	Which assumptions drive different funding priorities?
Expert landscape	Expert profiles with positions	Who believes what about deceptive alignment?
Gap analysis	Quality scores reveal under-developed areas	Which important topics lack quality coverage?

For Researchers

Use Case	Wiki Feature	Example
Literature synthesis	Consolidated coverage with citations	Find all sources on a specific risk
Gap identification	Coverage analysis, importance vs quality	What important topics need more research?
Position mapping	Disagreement visualization	Where do Yudkowsky and Christiano diverge?
Model building	Causal diagrams as starting points	Use wiki models as research scaffolding

For Policymakers

Use Case	Wiki Feature	Example
Risk taxonomy	Structured hierarchy with assessments	Navigate from high-level categories to specific risks
Response options	Comprehensive intervention catalog	What governance tools exist?
Expert consensus	Crux tracking shows agreement/disagreement	Where do experts converge vs diverge?
Quick briefings	LLM summaries on every page	One-sentence takeaways for busy readers

Comparison with Similar Projects

Resource	Focus	Update Model	Longterm Wiki Differentiator
Wikipedia	General reference	Community edits	LW is opinionated about importance and uncertainty
LessWrong / AI Alignment Forum	Discussion	Continuous posts	LW is curated synthesis, not discussion platform
80K Problem Profiles	Career guidance	Periodic updates	LW goes deeper on cruxes, covers more interventions
Stampy / AISafety.info	FAQ for newcomers	Community + fellowship	LW is strategic (prioritization), not educational
MIT AI Risk Repository	Academic taxonomy	Research updates	LW adds expert positions, crux mapping, intervention analysis
EA Forum Wiki	Integrated tags	Tag-based	LW is standalone with dedicated editorial

Lessons from Similar Projects

The Similar Projects Analysis identified key success factors:

Factor	Failed Projects	LW Approach
Ownership	Arbital (committee), EA Wiki v1 (volunteer)	Dedicated editorial lead
Funding	Most wikis (volunteer-only)	Paid core team
Scope	Arbital (everything)	AI safety prioritization only
Integration	Standalone wikis (low traffic)	Links to LW/EAF, doesn’t replace
Updates	Survey papers (stale in 1-2 years)	Quality tracking, staleness monitoring

Strengths and Limitations

Strengths

Strength	Evidence	Impact
Crux-first design	50+ tracked uncertainties, expert positions	Enables disagreement decomposition
Quality control	Automated grading, 6-dimension scoring	Consistent content standards
Cross-linking	550+ pages with stable entity references	Knowledge graph navigation
Open source	MIT license, public GitHub	Transparency, reproducibility
Modern stack	Astro, React, TypeScript	Fast, maintainable, accessible
Interactive visualizations	ReactFlow graphs, Mermaid diagrams	Complex relationships made legible

Limitations

Limitation	Impact	Mitigation
Single editorial voice	May reflect particular worldview	Explicit about opinions, tracks disagreement
Staleness risk	Content can become outdated	Quality tracking, freshness monitoring, dated pages
Limited community contribution	Slower content growth	Focus on quality over volume, open for feedback
AI safety focus	Narrow scope	Intentional limitation, links to broader resources
Early stage	Incomplete coverage	Active development, prioritized expansion
No real-time data	Static forecasts	Links to MetaforecastConceptMetaforecastMetaforecast is a forecast aggregation platform combining 2,100+ questions from 10+ sources (Metaculus, Manifold, Polymarket, etc.) with daily updates via automated scraping. Created by QURI, it pr...Quality: 35/100 for live data

Relationship to QURI Ecosystem

Loading diagram...

Tool	Relationship to Longterm Wiki
SquiggleConceptSquiggleSquiggle is a domain-specific probabilistic programming language optimized for intuition-driven estimation rather than data-driven inference, developed by QURI and adopted primarily in the EA commu...Quality: 41/100	LW documents probabilistic concepts; Squiggle enables quantification
SquiggleAIConceptSquiggleAISquiggleAI is an LLM tool (primarily Claude Sonnet 4.5) that generates probabilistic Squiggle models from natural language, using ~20K tokens of cached documentation to produce 100-500 line models ...Quality: 37/100	LW models could be converted to executable Squiggle estimates
MetaforecastConceptMetaforecastMetaforecast is a forecast aggregation platform combining 2,100+ questions from 10+ sources (Metaculus, Manifold, Polymarket, etc.) with daily updates via automated scraping. Created by QURI, it pr...Quality: 35/100	LW links to relevant forecasts as evidence for claims
Squiggle Hub	Potential future integration for interactive models embedded in pages

Development Roadmap

Based on the Vision Document:

Phase	Goals	Status
Foundation	Core knowledge base structure, 30+ high-quality pages, initial crux taxonomy	Complete
Depth	80+ quality pages, full crux graph, worldview archetypes	In progress
Polish	Expert review, interactive worldview→priority tool, public launch	Planned
Maintenance	Quarterly review cycle, community contribution guidelines	Planned

Contributing

The wiki is open source and welcomes contributions:

Contribution Type	How to Help
Content feedback	Report issues or suggest improvements via GitHub Issues
Technical contributions	PRs welcome for components, validation, automation
Expert input	Contact for crux positions, disagreement mapping
Bug reports	File issues for broken links, rendering problems

See Internal Documentation for:

• About This Wiki — Technical overview for contributors
• Style guides for content creation
• Automation tools for development workflows
• Page types for classification system

External Links

• Longterm Wiki
• GitHub Repository
• Vision Document
• Similar Projects Analysis
• QURI Website