Glossary · 63 terms
AI terms, one page each.
From agents and context rot to self-attention and Elo ratings. Each term has its definition in plain language and the chapters that explain it in depth.
Looking for the full agent glossary on one page? Read it here.
A
- AgentAn AI agent is a system that uses a large language model (LLM) as its reasoning engine to decide what to do…
- Agent harnessThe harness is the software wrapper around the model. The LLM provides the thinking and produces a "tool…
- Agentic searchA way for an agent to find information by using ordinary computer tools like grep (search text in files)…
B
C
- Chain-of-thought (and extended / interleaved thinking)Chain-of-thought means prompting a model to write out its reasoning step by step before giving a final…
- CompactionA reversible way to shrink the context window. Compaction removes information from the running history when…
- Constrained decoding (structured output / schema)Techniques that force a model's output to follow a fixed format, such as a specific JSON shape.
- Context confusionA failure mode where the model cannot tell apart instructions, data, and structural markers, or is given…
- Context engineeringThe discipline of designing a system that puts the right information and tools, in the right format, at the…
- Context rotThe observation that a model's quality drops as its context window fills up, even when the total tokens are…
- Context windowThe fixed-size working memory of a language model, measured in tokens. Everything the model can "see" at once…
D
E
- Effort scalingGiving an agent explicit rules about how much work a task deserves, so it does not over-invest in easy…
- Elo ratingEach model has a numeric rating. When two models compete and humans vote for the winner, the winner's rating…
- EmbeddingA list of numbers (a vector) that represents the meaning of a piece of text, produced by a model.
- Emergent abilityAn emergent ability is a skill that a model does not have when small but suddenly displays once it crosses a…
F
G
H
I
K
L
M
- MAST (Multi-Agent System Failure Taxonomy)A structured catalog of 14 ways multi-agent systems fail, grouped into three stages: specification issues…
- MCP (Model Context Protocol)An open standard that lets an agent connect to external tools and services (like Slack, GitHub, or Google…
- Memory consolidationThe step where new information is merged into existing memory in a clean way: combining related facts…
- Memory poisoning (and prompt injection)A security risk where malicious or misleading instructions get written into an agent's memory or context, and…
- MetricA metric is the specific rule you use to turn a model's output into a number. Accuracy (the fraction of…
- Mixture of Experts (MoE)A small router picks which experts handle each word. The benefit is that the model can hold a very large…
- Multi-hop reasoningAnswering a question that requires chaining several pieces of information together across multiple steps…
O
P
- Parameters (weights)Parameters, also called weights, are the adjustable numbers inside a neural network. They are the heart of…
- PPOPPO, short for Proximal Policy Optimization, is the specific reinforcement learning algorithm used in the…
- Progressive disclosureA design principle where an agent loads information only as it becomes relevant, rather than all at once.
R
- RAG (Retrieval-Augmented Generation)A technique where, instead of relying only on what a model learned during training, the system retrieves…
- Reasoning (in language models)Reasoning means working through a problem in multiple connected steps to reach a conclusion, rather than…
- RecursionRecursion is when a process solves a problem by calling a fresh copy of itself on a smaller piece of the same…
- Reflection (and Reflexion)Reflection is an agent pattern where the model critiques its own output and uses that critique to improve on…
- Reinforcement learning (RL), PPO, GRPOReinforcement learning is a training approach where a system learns by trial and error, guided by rewards for…
- REPLREPL stands for Read-Eval-Print Loop, a small interactive programming environment where you can type a piece…
- Reward modelA reward model is a separate model trained to predict how much a human would like a given answer.
- RLHFRLHF stands for Reinforcement Learning from Human Feedback. RLHF was the breakthrough that made ChatGPT…
- Router (gating network)The router, sometimes called the gating network, is the small, fast component inside a Mixture of Experts…
S
- SaturationSaturation is what happens when a benchmark becomes too easy: nearly every model scores near the top, so the…
- Scaling lawA scaling law is a predictable mathematical relationship showing how a model's performance improves as you…
- Self-attentionSelf-attention lets every word in a sentence look at every other word and decide how much each one matters…
- Semantic driftThe slow corruption of stored knowledge over time, usually caused by repeated summarization.
- Staleness (temporal obsolescence)When stored information was correct once but is now out of date because the world changed (an address, a…
- Sub-agentA secondary agent that a main agent spins up to handle a specific piece of work. A sub-agent runs in its own…
- Supervised fine-tuning (SFT)Humans write high-quality example answers to many prompts, and the model is trained to imitate them.
T
- TokenThe unit a language model reads and writes. A token is roughly a word fragment; a word can be one or several…
- Tool calling (function calling)The mechanism by which an agent takes action in the world. The model outputs a structured request to run a…
- Tool useTool use is a model's ability to call external functions or services to get something done: running a web…
- TransformerTransformer processes all words in parallel and lets any word attend to any other, which made it fast to…
U
V
- VectorA vector is simply an ordered list of numbers, like [0.2, -1.3, 0.8]. That is the whole definition.
- Vector databaseA specialized store for embeddings that can quickly find the vectors closest to a query vector.
- VerifierA verifier is a model whose job is to check work rather than produce it. A good verifier is hard to fool…