Insight Hunting

LongtermWiki contains approximately 1 million words of content across hundreds of pages, but currently only 74 insights have been extracted. This section provides tools to systematically find more high-quality insights.

What Makes a Good Insight?

Insights are rated on five dimensions:

Dimension	Question	Example
Surprising	Would this update an AI safety researcher’s beliefs?	“78% of Claude instances engaged in alignment faking”
Important	Does this affect decisions or priorities?	”RLHF fundamentally cannot scale to superhuman tasks”
Actionable	Does this suggest concrete work?	”Linear probes achieve >99% AUROC detecting deception”
Neglected	Is this getting less attention than deserved?	”Only 2-3 teams globally work on goal misgeneralization”
Compact	Can it be stated in 1-2 sentences?	Short, self-contained claims work best

The best insights score 4+ on multiple dimensions simultaneously.

Available Tools

Gap Analysis

Find pages with high importance but few/no extracted insights. Prioritizes by: importance × (1 + quality/100) - insightCount × 20

Open Gap Analysis →

Table Candidates

Table rows with paradoxical or notable rating combinations that suggest insight-worthy content.

Open Table Candidates →

Quantitative Claims

Numbers, percentages, and statistics extracted from content that could become standalone insights.

Open Quantitative Claims →

All Insights

Browse and search all 124 extracted insights with filters and sorting.

View All Insights →

Current Statistics

Metric	Value
Total content pages	≈500+
Total word count	≈1,000,000
Extracted insights	124
Pages with insights	≈40
Insights per page (avg)	≈3.1
High-importance pages without insights	50+

Workflow for Adding Insights

1. Find candidates using the tools above
2. Verify the claim is accurate and well-sourced
3. Rate the insight on the five dimensions (1-5 scale)
4. Add to insights.yaml following this format:

- id: unique-id-here
  insight: "The claim in 1-2 sentences"
  source: /knowledge-base/path/to/page/
  tags: [relevant, tags]
  type: claim  # or: research-gap, counterintuitive, quantitative, disagreement, neglected
  surprising: 4.0
  important: 4.5
  actionable: 3.0
  neglected: 4.0
  compact: 5.0
  added: "2025-01-22"

Types of Insights

Type	Description	Example
claim	Factual assertion	”GPT-4 scores 87% on MMLU”
research-gap	Missing research	”No published work on X”
counterintuitive	Surprising finding	”More data sometimes hurts”
quantitative	Number or statistic	”$1B+/year spent on RLHF”
disagreement	Expert conflict	”Researchers disagree on X”
neglected	Underfunded area	”Only 3 people work on X”