Reference wiki
The deep-dive layer under Understanding Machine. Where the compendium points you at the best thing written on a subject, the wiki answers the smaller questions that come up while you read it: what a benchmark actually measures, how a system works, where a term came from. Articles are AI-researched and citation-rich — every substantive claim carries a link to its source, so you can check the work rather than trust it.
193 articles. They cross-reference each other; follow the links sideways as far as it stays useful.
A
- Agent Harness: The Software Surrounding the Model
- Agent Memory Architectures: Working Set, Episodic, Semantic, and Beyond
- Agent Scaffolding
- AI Coding Agents in 2026: Architectures, Capabilities, Bottlenecks
- AI Control: Safety When Trust Cannot Be Verified
- Ai Evals
- Aider: Git-Native Coding Pair Programmer
- AIME: American Invitational Mathematics Examination as AI Benchmark
- Algorithm Appreciation: The Asymmetric Counterpart to Algorithm Aversion
- Algorithm Aversion: Why Humans Reject Helpful Models
- Alignment Faking: When Models Comply During Training and Defect Later
- AlphaGeometry
- AlphaGo
- AlphaGo Zero
- AlphaProof: Search and Self-Play for Formal Mathematics
- AlphaZero
- Anchored Rubrics: Calibrated Scoring Scales for LLM Evaluation
- Apparent Personality from Text: LLM-Based Personality Inference
- ARC-AGI
- Attachment Theory And Ecr R
- Automation Bias
B
C
- Calibrated Challenge: Anti-Sycophancy as a Positive Design Target
- Capability Elicitation: The Other Half of Capability Evaluation
- Capability Evaluation: Distinct from Safety, Reliability, and Behavior Evaluation
- Caricature Under Persona Conditioning: When Persona Prompts Produce Stereotype
- Chain-of-Thought Faithfulness: When the Stated Reasoning Doesn't Match the Actual Decision
- Chain-of-Thought Monitorability: Can We Trust Visible Reasoning?
- Chain-of-Thought Monitoring: Practice and Limits
- Chain-of-Thought Prompting: From Wei et al. 2022 to Internalized Reasoning
- Claude Code
- Claude Extended Thinking
- Cluster Bootstrap For Hierarchical Eval
- Code Interpreter
- Codex CLI
- Computerized-Adaptive-Testing
- Constitutional AI: Method, Evolution, and Critique
- Construct Validity in AI Evaluation
- Context Engineering for AI Agent Systems
- Context Window Economics: The Hidden Cost Structure of Large Language Model Applications
- Convergent and Discriminant Validity: The Multitrait-Multimethod Matrix
- Cursor Agent Mode: IDE-Native Autonomous Coding
D
E
F
G
H
I
J
L
M
- Mechanistic Interpretability
- Mesa-Optimization: Learned Optimizers Inside Trained Models
- Mini-SWE-Agent
- MIPRO
- Mixture-Of-Experts
- MMLU-CF
- MMLU-Pro: The Less-Saturated Successor
- MMLU-Redux: Cleaning Up the Item Pool
- MMLU: Anatomy of a Saturated Benchmark
- MMMU: Massive Multi-discipline Multimodal Understanding
- Model Routing: When and How to Send Different Queries to Different Models
- Monte Carlo Tree Search in 2026: From AlphaGo to LLM Reasoning
- Moral Foundations Theory: The Pluralist Model of Moral Cognition
- MRKL Systems: Modular Reasoning with Knowledge and Language
- Multi-Head Latent Attention
- Multi-Model Agent Orchestration Patterns
- Multi-Token-Prediction
- Mutable Scaffolds: When the Software Layer Self-Modifies
- MuZero
N
O
P
- Pairwise Versus Scalar Evaluation: Tradeoffs in LLM Output Judging
- Persona Prompting: System-Prompt-Based Personality Conditioning of LLMs
- Preparedness, Responsible Scaling, and Frontier Safety: Comparing Lab Governance Frameworks
- Process Reward Models: Stepwise Supervision for Reasoning
- Process Supervision vs Outcome Supervision: A Training-Signal Distinction
- Production LLM Evaluation: Offline, Online, and the Hybrid Pattern
- Prompt Caching Across Providers
- Prompt Engineering: Practice, Decline, and What Remains
- Prompt Injection in Agentic Systems: Attack Surfaces, Defenses, Residual Risk
- Prompt Optimization
- Psychometric Correlates of AI Interaction Styles
- Public-Archetype Echo: When AI Assistants Mimic Famous Persona Surfaces
R
- ReAct Agents
- ReAct: The Reasoning-Acting Synthesis
- Reasoning Models in 2026: o1/o3, DeepSeek-R1, Gemini Thinking, Claude Extended Thinking
- Reasoning Tokens: API Semantics, Billing, and Visibility
- Reasoning via Planning: PDDL-Style Decomposition in LLM Agents
- Recursive Self-Improvement: Theory, Demonstrations, Limits
- Reflexion: Verbal Reinforcement for Language Agents
- Regression Testing for AI Systems: Beyond Snapshot Tests
- Repair-Oriented Response
- Repository Branch Awareness in Coding Agents
- Retrieval-Augmented Generation
- Reward Hacking and Specification Gaming: Patterns, Evidence, and Mitigations
- Reward Tampering: When the AI Modifies Its Own Training Signal
- ReWOO: Reasoning Without Observation
- Rich Sutton’s “The Bitter Lesson” (2019): Thesis, Reception, and Counter-Arguments Through 2026
- RLAIF: Reinforcement Learning from AI Feedback
- RLHF: Reinforcement Learning from Human Feedback
S
- Safety Evaluation: Capability-Misuse, Alignment, or Behavior?
- Sample-Efficient Personality Inference
- Sandbagging: Strategic Underperformance in AI Systems
- Scalable Oversight: Supervising AI That Outpaces Human Judgment
- Scaling Laws for Neural Language Models: From Kaplan to Chinchilla and Beyond
- Search and Learning: The Bitter Lesson’s Two Engines
- Self Determination Theory
- Self-Consistency
- Self-Improving Software Systems: From EURISKO to Modern AI Agents
- Semantic Search Architecture for Code and Documentation: Embeddings in Practice
- Specification Gaming: When the AI Optimizes the Letter, Not the Spirit
- Style Mimicry Overfit: When LLMs Match Surface Style Over Content Quality
- SWE-agent: The First Frontier Agent Harness for Software Engineering
- SWE-Bench
- SWE-bench Live: Continuously Refreshed Software-Engineering Eval
- SWE-bench Pro: Larger Repositories, Harder Issues
- Swe-Bench-Verified
- Sycophancy in Large Language Models
- Synthetic Personas for AI Evaluation: Methodology and Limits
- Synthetic-First Evaluation: Methodology Before Human Studies
- System Capability Vs Model Capability
T
- Terminal-Bench
- Test-Time Compute
- Test-Time Training
- The Agent-Computer Interface: Browser, Terminal, API, GUI, and Beyond
- The Cognitive Reflection Test: Measuring Override of Intuitive Responses
- The Dark Triad: Machiavellianism, Narcissism, and Psychopathy
- The Model Context Protocol (MCP) Ecosystem as of 2026
- The Pruning Principle: When Scaffolding Becomes Dead Weight
- Theory of Mind in Large Language Models: The Contested Empirical Question
- Thomistic Natural Law Theory Applied to AI Agent Decision-Making
- Tool Calling Semantics: Cross-Provider Comparison
- Tool Use in Language Models: From ReAct to Native Tool Calling
- Tool-Augmented Reasoning: When the Model Calls Code
- Toolformer
- Tracked Capability Levels: Governance Through Threshold Tiering
- Tree of Thoughts: Deliberate Reasoning Through Search
- Trust Calibration in Human-AI Systems