Skip to content

Quick Reference

Fresh Updated January 2026

Fast-access reference materials for LLM technical phenomena and GEO/AEO strategy.

Available References

Reference Index

ReferenceDescriptionUse When
Data PointsAll key statistics and metricsNeed specific numbers
TroubleshootingCommon issues and solutionsSomething isn't working

Key Numbers at a Glance

Variance & Non-Determinism

  • 11.67 unique completions from GPT-4 at temp=0 (30 runs)
  • 30 unique completions from MoE models (30 runs)
  • 42% citation overlap between platforms

Performance Degradation

  • 20-60% context rot past 100k tokens
  • 30% worse retrieval for middle-positioned content
  • 5-15% NLP performance drop post-RLHF alignment

Research Status

  • 100+ interpretability papers (2024-2025)
  • 0 consensus solutions for interpretability
  • 80-90% variance explained by SAEs (incomplete)

Platform Quick Reference

PlatformPrimary SourceKey Strategy
ChatGPTTraining dataHistorical depth, Wikipedia alignment
PerplexityReal-time searchFreshness, multi-source presence
ClaudeTraining + conservativeThird-party validation, neutral framing
GeminiGoogle indexSEO, Google ecosystem integration

Content Position Quick Reference

PositionAttention LevelBest Use
First 20%HighestPrimary claims, key differentiators
Middle 60%LowestSupporting details, background
Last 20%Elevated (recency)Summary, repeated key points

Based on research from Thinking Machines Lab, Chroma Research, and ACL 2024-2025