Objective
This memo is a glossary of terms for all things related to Artificial Intelligence (AI).
Subjective
Glossary
A
AI Agent – Autonomous system that plans, acts, and critiques to achieve goals.
Artificial General Intelligence (AGI) – Human-level AI that generalizes across domains.
Artificial Intelligence (AI) – Machines performing tasks requiring human intelligence.
Artificial Super Intelligence (ASI) – AI surpassing human capability in all domains.
B
Benchmarks – Standard test sets for model performance (MMLU, GSM8K, ARC, etc).
C
Chain-of-Thought (CoT) – Step-by-step reasoning method for reliable answers.
Closed-weight Model – Proprietary models with non-public parameters.
Context Rot – Decline in reliability with long or cluttered prompts.
Context Window – Max tokens a model can process in one pass.
Custom GPT – Tailored GPT built for specific use cases.
D
Deep Learning (DL) – Multi-layer neural nets for vision, speech, language.
E
Embeddings – Dense vector representation of text, images, or data.
Episodic Memory – Recall of past events to improve personalization.
Evals – Frameworks for testing AI across standard datasets.
F
Faithfulness – Outputs remaining true to given sources.
Few-shot – Task learning guided by multiple examples.
Foundation Model – Large pre-trained models adaptable to many tasks.
Frontier Model – Cutting-edge models pushing AI performance.
Function Calling – AI invoking APIs/tools with structured inputs.
G
Golden Set – Reference pairs for regression testing quality.
Graph-of-Thought (GoT) – Reasoning with DAG subproblems and reusable paths.
H
Hallucination Rate – Share of unsupported claims in model outputs.
HNSW – ANN algorithm for fast, high-recall vector search.
HumanEval – Benchmark for code generation.
J
Jailbreaks – Prompts that bypass AI’s safety or alignment rules.
K
KV Cache – Speeds generation by reusing attention states.
L
LLM-as-a-Judge – Models used to evaluate outputs by rubric.
Long-context – Models with extended token windows.
M
Machine Learning (ML) – AI subset learning from data patterns.
Mixture-of-Experts (MoE) – Large models with specialist subnetworks.
Model Context Protocol (MCP) – Open standard for tool integration.
Model Landscape – Core AI building blocks and structure.
Multi-Agent AI (MAAI) – Systems with multiple cooperating AI agents.
Multimodal LLM (MLLM) – Models combining text, image, audio, video.
O
Open-weight Model – Models with publicly available parameters.
P
Pairwise Preference – Evaluation method comparing two outputs.
Program-of-Thought (PoT) – Reasoning expressed as code steps.
Prompt Injection – Hidden instructions tricking models.
Prompt Template – Reusable structure with variable placeholders.
Prompting – Crafting inputs to guide AI outputs.
R
ReAct – Pattern mixing reasoning and tool actions.
Reasoning Model – AI built to plan, verify, and justify answers.
Regression Tests – Checks for quality after updates.
Retrieval-Augmented Generation (RAG) – Combines models with document retrieval.
S
Safety – Ensuring models produce non-harmful outputs.
Self-Refine – Iterative AI self-revision loop.
Semantic Caching – Stores responses for similar queries.
Semantic Memory – AI recall of facts or user details.
Session Memory – Persistent context across user chats.
Sycophancy – AI over-agreeing with users.
System Prompt – Foundational instruction guiding AI behavior.
T
Temperature – Randomness control in output generation.
Top-k – Limits next-token choices to top k options.
Top-p – Samples tokens from cumulative probability mass.
Tree-of-Thought (ToT) – Branching reasoning exploration.
U
User Prompt – Direct input from a user to the AI.
V
Vector Database – Stores embeddings for retrieval and search.
VLM – Vision-language model, multimodal with images.
Z
Zero-shot – Solving tasks without prior examples.
Do you want me to also add current examples (like GPT-5, Claude Sonnet 4, Gemini 2.5) as their own glossary entries, or keep the glossary focused on concepts only?
Contexts
#ai-lexicon (this is the Root Memo)
#glossary