Missing a term? This page grows along with the web of AI terminology. Some definitions are still in flux, because the technology itself is still in flux. If you want to dig deeper, just click the linked chapter name.
Agent
An AI that doesn't just answer, it independently uses tools, plans multiple steps, and works through tasks in a loop. More on this in the chapter AI Agents.
Alignment
Getting an AI model to match what people actually want, not just the literal instruction. A central topic in Safety & Alignment.
API
An interface through which one program communicates with another. For AI: the way your application talks to a model like Claude or GPT. See API & SDK.
Bias
Prejudices a model picks up from its training data, for example gender or skin-color stereotypes. More in the chapter Safety.
Chain of Thought
A prompting technique where you ask the model to "think out loud," writing out intermediate steps before giving the answer. Noticeably improves quality on complex tasks.
Chunk
A small piece of text that a document is broken into before making it accessible to an AI via RAG. Getting the size and boundaries right is a small art in itself.
Constitutional AI
A training method where a model is given explicit principles, a kind of "constitution," that it aligns its answers to. Developed primarily at Anthropic. More in Safety.
Context Window
The maximum amount of text (measured in tokens) a model can process at once. Input and output combined. Typical windows currently range from 8,000 to over a million tokens.
Deep Learning
Machine learning with especially deep neural networks, meaning many layers. Practically every modern AI uses deep learning. See Neural Networks.
Embedding
A numerical representation of a text's meaning. Texts with similar content get similar lists of numbers. The foundation for RAG and semantic search.
Few-Shot
A prompting technique where you show the model a few examples of the desired input-output pairs before giving it the actual task. See Prompts.
Fine-Tuning
Further training an already-trained model on specific data or tasks. Makes the model better at a given area, but takes effort. More in the chapter Training.
GPT
"Generative Pre-trained Transformer," the architecture and model family behind ChatGPT. The name has since become a brand for OpenAI's models. Almost all modern LLMs, though, use the underlying Transformer architecture.
Guardrails
Additional safety layers that filter problematic inputs or outputs of an AI. More in Safety.
Hallucination
When an AI invents facts, sources, or numbers and presents them with full confidence. Its own chapter: Hallucinations.
Harness
The "harness" around an AI model, the code that gives it tools, processes inputs, and drives the loop. Turns a model into a usable product. See Harness.
Inference
The process by which a fully trained model produces an output from an input, the opposite of training. Every answer you get from ChatGPT is an inference.
AI
Artificial Intelligence. Systems that solve tasks that used to seem to require human intelligence. More in the introductory chapter What is AI?.
LLM
"Large Language Model," a large language model trained on huge amounts of text that generates text by predicting the next token. See Large Language Models.
MCP
"Model Context Protocol," an open format that lets AIs connect to external tools and data sources. A kind of "USB for AIs." See MCP.
Multimodal
A model that can handle more than just text, meaning also images, audio, or video, both as input and as output.
Neural Network
The mathematical structure behind modern AI, loosely inspired by the brain, made up of layers of "neurons" with weights. More in the chapter Neural Networks.
Parameter
A model's "dials," the weights that get adjusted during training. Modern LLMs have billions to trillions of them. More parameters doesn't automatically mean better, but usually more capable.
Prompt
The input given to an AI model, the question, the task, the context. Good prompts are half the battle. See Prompts.
Prompt Injection
An attack where someone hides instructions inside text that the AI later reads, and then unintentionally executes. More in the chapter Hallucinations.
RAG
"Retrieval Augmented Generation," the AI fetches the matching documents from a collection before every answer. Its own chapter: RAG.
Red Teaming
Systematically testing an AI for weaknesses. Specialists try to make the model misbehave so it can be hardened. More in Safety.
RLHF
"Reinforcement Learning from Human Feedback." Humans rate the model's answers, and the model learns from those ratings. The main tool for friendly, helpful behavior. More in the chapter Training.
SDK
"Software Development Kit," a bundle of tools, libraries, and examples that makes it easier for developers to work with an API. See API & SDK.
System Prompt
A special prompt that sets the AI's behavior for an entire session, for example "You are a friendly tutor for middle-school students." Sent along before every user question. More in Prompts.
Temperature
A dial that determines how "creative" or "random" an AI's answers are. Low temperature (e.g. 0) means: always the most likely word. Higher values: more variation, more surprise.
Token
The smallest unit in which an LLM processes text, usually part of a word, sometimes a whole word. "Artificial Intelligence," for example, gets broken into several tokens. More in How does AI work?.
Tool Use
A model's ability to use external tools, for example a web search, a calculator, or a database. The foundation for Agents.
Training
The process by which a model learns from data, meaning its parameters are adjusted so that it produces the desired predictions. Effortful and expensive. Its own chapter: Training & Data.
Transformer
The architecture that has shaped nearly all modern language models since 2017, based on a mechanism called "attention." Makes it possible to process long texts efficiently.
Vector Database
A special database for embeddings, it searches not by words but by similarity of meaning. Well-known examples: Pinecone, Chroma, Qdrant. More in RAG.
Zero-Shot
A task is given without examples, the model has to understand what to do from the description alone. Works surprisingly often with good models. The counterpart to Few-Shot.
With that, you have the vocabulary together to join in on almost any AI conversation, whether over lunch, in a meeting, or on a forum. You won't run into every term right away, and some definitions will still shift over the next few years. That's normal in a field that's still inventing itself. What matters is: you have the map. When the next hype term crosses your path, you'll know where to place it, and you'll spot empty buzzwords the moment someone throws them around without explaining them.