This AI glossary explains 79 common AI terms in one or two plain sentences each, from tokens and context windows to agents, MCP and fine-tuning. Terms are grouped by topic (core concepts, how models work, training, using AI, agents and tools, building with AI, and safety and policy) so related ideas sit together.
- A large language model (LLM) predicts the next token of text; chatbots, coding agents and search assistants are products built on top of one.
- The context window is the model's working memory for one request; anything outside it, the model cannot see.
- An agent is a model that uses tools in a loop (plan, act, observe) to finish a task, rather than only replying.
- RAG puts relevant documents into the context at query time; fine-tuning changes the model itself with extra training.
- Hallucinations are confident false statements, and the best defense is giving the model sources and checking its claims.
What is in this AI glossary?
This AI glossary defines 79 terms you will meet when using or building with AI, each in one or two plain sentences. It is grouped by topic rather than alphabetically, so you can read a section top to bottom and come away understanding how the pieces connect.
How to use it:
- New to AI? Read "Core AI concepts" and "Using AI tools" first.
- Building with AI? Jump to "Building with AI" and "Agents and tools".
- Writing a company policy? "Safety, risk and policy" covers the vocabulary you need.
Definitions reflect how the terms are used in practice as of October 2026. Where a vendor uses a term in a specific way, the entry says so.
Core AI concepts
These are the umbrella terms that everything else sits under.
Artificial intelligence (AI): The field of building computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognizing images or making decisions. Today the word usually refers to systems built with machine learning.
Machine learning (ML): An approach to AI where a system learns patterns from examples instead of following rules written by hand. A spam filter that learns from emails marked as spam is machine learning.
Deep learning: Machine learning using neural networks with many layers. Nearly all modern language, image and speech models are deep learning models.
Neural network: A model made of many simple connected units whose connection strengths are adjusted during training. The name comes from a loose analogy to neurons in the brain.
Generative AI: AI that creates new content (text, images, audio, video or code) rather than only classifying or predicting a label. ChatGPT, Claude, Midjourney and video generators are all generative AI.
Large language model (LLM): A neural network trained on very large amounts of text to predict the next piece of text, which turns out to be enough to write, summarize, translate, answer questions and code. Anthropic's glossary describes LLMs as models "with many parameters" capable of "surprisingly useful tasks." Our guide to what a large language model is explains how they work.
Foundation model: A large model trained on broad data that can be adapted to many tasks, either by prompting or by further training. Most LLMs are foundation models.
Frontier model: The most capable models available at a given time, typically from a handful of labs. The term matters in regulation and safety research, where the most capable models get extra scrutiny.
Multimodal model: A model that works with more than one kind of input or output, such as text plus images, audio or video. Most current assistants accept images and documents alongside text.
Chatbot: A conversational interface where you type or speak and the system replies. Modern chatbots are products built around an LLM, with instructions, safety rules and often tools added.
AI assistant: A chatbot designed to help with general tasks such as writing, research, analysis and planning. ChatGPT, Claude and Gemini are the best-known examples.
How language models work
These terms describe what happens inside a model when it reads your prompt and writes an answer.
Token: The unit of text a model reads and writes, usually a word or part of a word. Pricing, limits and context windows are all measured in tokens; for Claude, Anthropic says a token was about 3.5 English characters on older models, with newer models producing roughly 30 percent more tokens for the same text.
Tokenizer: The component that splits text into tokens and turns them into numbers the model can process. Different models use different tokenizers, so the same text can be a different number of tokens.
Parameters (weights): The internal numbers a model learns during training, which together encode what it knows. Model size is often described by parameter count, in the billions for modern LLMs.
Transformer: The neural network architecture behind nearly all modern LLMs, introduced in the 2017 paper Attention Is All You Need. It processes all tokens in a passage together instead of strictly one after another.
Attention: The mechanism inside a transformer that lets each token weigh how relevant every other token is. It is how a model links "it" in a sentence to the noun it refers to.
Embedding: A list of numbers (a vector) that represents the meaning of a piece of text, so that similar meanings get similar vectors. Embeddings power semantic search and recommendations; see what embeddings are.
Context window: The total amount of text a model can consider in one request, including instructions, conversation history, documents, tool results and its own answer. Anthropic's context window docs call it the model's "working memory"; current Claude models have a 1 million token window.
Context rot: The drop in accuracy and recall that happens as a context window fills up. It is why giving a model the right context beats giving it all the context.
Inference: Running a trained model to produce output, as opposed to training it. Every chat message you send triggers inference.
Temperature: A setting that controls how random a model's word choices are. Low temperature gives more predictable output; high temperature gives more varied output.
Latency: The time between sending a request and getting the response. Time to first token (TTFT) measures how quickly the first part of the answer appears.
Reasoning model (extended thinking): A model or mode that works through a problem step by step before answering, spending extra tokens on intermediate reasoning. It improves results on math, code and multi-step problems at the cost of time and tokens.
Knowledge cutoff: The date after which a model's training data ends. The model does not know about later events unless a tool such as web search provides them.
How AI models are trained
These terms describe how a model goes from raw data to a helpful assistant.
Training data: The text, code, images or other examples a model learns from. Its quality and coverage largely decide what the model can and cannot do.
Pretraining: The first, largest training stage, where the model learns to predict the next token across a huge text collection. A pretrained model knows a lot but is not yet good at following instructions.
Post-training: Everything done after pretraining to make the model useful and safe, including fine-tuning and feedback-based methods.
Fine-tuning: Further training a pretrained model on a smaller, specific dataset so it adopts a style, format or domain. It is an alternative to prompting and RAG when you need consistent behavior those cannot give.
Instruction tuning: Fine-tuning on examples of instructions paired with good responses, which teaches a model to follow requests instead of merely continuing text.
RLHF (reinforcement learning from human feedback): A training method where people rank model answers and the model learns to prefer answers like the higher-ranked ones. Anthropic's glossary lists it as one of the methods used to make Claude a helpful assistant.
Distillation: Training a smaller "student" model to imitate a larger "teacher" model, producing a faster, cheaper model that keeps much of the capability.
Open-weight model: A model whose trained parameters are published so anyone can download and run it. Open-weight is not always the same as open source, because the training data and code may not be released.
Benchmark: A standard test used to compare models, such as a set of coding problems or exam questions. Benchmarks are useful signals but can be gamed and rarely match your exact tasks.
Using AI tools
These terms come up in everyday use of assistants like ChatGPT, Claude and Gemini.
Prompt: The input you give a model: a question, instruction, document or example. Better prompts state the task, audience, material and desired output; see the prompt engineering guide.
System prompt: Instructions set by the app or developer that apply to the whole conversation, usually invisible to the end user. It defines the assistant's role, rules and tone.
Prompt engineering: The practice of writing and refining prompts to get reliable, high-quality output. It includes giving context, examples, structure and clear success criteria.
Context engineering: Deciding what information goes into the context window for each step: instructions, documents, tool results, memory. It matters most for agents that run many steps.
Zero-shot and few-shot prompting: Zero-shot means asking without examples; few-shot means including a few examples of the input and the output you want. Examples usually improve format and consistency.
Chain of thought: Asking or allowing a model to reason step by step before giving a final answer. Reasoning models do this automatically.
Hallucination: When a model states something false with confidence, such as an invented citation, number or quote. It happens because models produce plausible text rather than verified facts; see AI hallucinations.
Grounding: Tying a model's answer to specific sources, such as documents you provide or search results, so claims can be checked.
Citation: A link or reference a model gives for a claim. Always open important citations, because models can cite sources that don't support the claim or don't exist.
Memory (AI assistants): A product feature that saves facts about you or past chats and feeds them back into future conversations. The model's weights don't change; the product adds the remembered text to the context.
Custom instructions: Account-level instructions you set once (your role, preferred tone, format) that an assistant applies to every chat.
Artifact: In Claude, a finished output such as a document, slide deck, diagram or small app shown in a panel beside the chat, where you can edit and export it.
Deep research: A mode in several assistants where the model runs many searches in sequence and writes a sourced report. It takes longer than a normal answer and uses more of your limits.
Agents and tools
These terms describe AI that takes actions, not only answers.
AI agent: A system where a model uses tools in a loop (plan a step, act, observe the result, repeat) to complete a task. Coding agents, browser agents and research agents are common types; see what AI agents are.
Tool use (function calling): The ability of a model to request an action, such as running code, searching the web or calling an API, by producing a structured call that the app then executes. The result is sent back to the model.
Agentic loop: The repeating cycle an agent follows: decide, act with a tool, read the result, decide again. The loop ends when the task is done, a limit is hit or a human is needed.
MCP (Model Context Protocol): An open protocol that standardizes how AI apps connect to tools and data sources. The MCP documentation compares it to a USB-C port for AI applications: one standard connection instead of a custom integration per tool.
MCP server: A program that exposes tools, data or prompts to AI apps through MCP, for example a server for GitHub, a database or a file system.
Connector: A product term for a link between an AI assistant and another service, such as Google Drive, email or a calendar. Many connectors are built on MCP.
Coding agent: An agent that works in a code repository: it reads files, edits code, runs tests and commands, and can open pull requests. Claude Code, Codex and Copilot's agent mode are examples.
Subagent: A helper agent started by a main agent to handle a subtask in its own separate context, returning only a summary. It keeps the main agent's context clean.
Multi-agent system: A setup where several agents work together, often with an orchestrator assigning work to specialist agents. It can speed up large tasks but adds coordination cost.
Computer use: An agent capability where the model operates a computer through screenshots, mouse clicks and keyboard input, as a person would.
Human in the loop: A design where a person reviews or approves an agent's important actions, such as sending an email, deploying code or spending money.
Headless mode: Running an agent without an interactive interface, from a script or CI job, where it takes a prompt, does the work and returns the result.
Background agent: An agent that works on a task while you do something else, often on a remote machine, and reports back with results such as a pull request.
Building with AI
These terms matter when you build products or workflows on top of models.
API (application programming interface): The way programs talk to a model provider: your code sends a request with a prompt and receives the model's response. API usage is usually billed per token.
SDK (software development kit): A library that wraps an API for a specific programming language, handling authentication, retries and streaming for you.
RAG (retrieval-augmented generation): A pattern where relevant documents are retrieved at query time and placed into the context so the model answers from them. It is the standard way to make a model answer from your own data; see retrieval augmented generation.
Vector database: A database built to store embeddings and find the most similar ones quickly. Postgres with the pgvector extension is a common alternative to a dedicated vector database.
Chunking: Splitting long documents into smaller pieces before creating embeddings, so retrieval can find the specific passage that answers a question.
Semantic search: Search by meaning rather than exact keywords, usually powered by embeddings. A query for "cancel my plan" can find a page titled "Ending your subscription."
Structured output: Asking a model to respond in a fixed format, such as JSON matching a schema, so other software can read the result reliably.
Prompt caching: Reusing the processed form of a repeated prompt prefix (such as a long system prompt or document) across requests, which cuts cost and latency.
Rate limit: The maximum number of requests or tokens you may send to an API in a period of time. Exceeding it returns an error until the window resets.
Evals (evaluations): Test sets and scoring methods you build to measure how well a model or AI feature performs on your own tasks. They are the AI equivalent of automated tests.
Guardrails: Checks around a model that block or correct unwanted inputs and outputs, such as filters for personal data, off-topic requests or unsafe actions.
Safety, risk and policy
These terms appear in security reviews, company policies and regulation.
Prompt injection: An attack where text in a document, web page or email contains instructions that hijack a model, for example telling an agent to leak data. It is the main security risk for agents that read untrusted content.
Jailbreak: A prompt designed to get a model to ignore its safety rules. Providers continually patch known techniques.
Alignment: The work of making AI systems pursue the goals and values their developers and users intend, and behave safely when they don't know what to do.
Bias: Systematic unfairness in model outputs, often inherited from patterns in training data, such as stereotypes about groups of people.
Red teaming: Deliberately attacking a model or AI product to find failures, unsafe behavior and security holes before others do.
Data retention: How long an AI provider stores your prompts and outputs. It varies by plan and settings, from zero retention on some business APIs to years when you allow data to be used for training.
AI literacy: The knowledge people need to use AI systems sensibly, including their risks. Under Article 4 of the EU AI Act, organizations using AI must support their staff's AI literacy; the European Commission's FAQ says this has applied since February 2, 2025.
Responsible AI: Practices for using AI fairly, transparently and safely: human review, disclosure, bias checks and clear accountability for decisions.
Shadow AI: Employees using AI tools their organization hasn't approved, typically personal accounts, which can expose company data.
How to put this vocabulary to work
Knowing the words helps most when you use them to ask better questions of tools and vendors:
- When choosing an assistant, ask about context window size, data retention and whether your data trains models.
- When a model makes something up, think grounding: give it sources and ask for citations.
- When an answer degrades in a long chat, suspect context rot and start a fresh conversation with a summary.
- When evaluating an agent product, ask which tools it can use, where it runs, and what needs human approval.
- When building a feature, start with prompting, add RAG for your own data, and consider fine-tuning only after you have evals that show the need.
If you run coding agents for long tasks, the terms "background agent" and "headless mode" are where it gets practical; Orkio is one way to keep agents like Claude Code running on an always-on cloud box after you close your laptop.
Frequently asked questions
What are the most important AI terms for beginners?
Start with large language model, prompt, token, context window, hallucination and agent. Those six explain most of what you see when using ChatGPT, Claude or Gemini, and the rest of the vocabulary builds on them.
What is the difference between AI, machine learning and deep learning?
AI is the broad goal of making computers perform tasks that need intelligence. Machine learning is the approach of learning patterns from data instead of writing rules by hand, and deep learning is machine learning with large neural networks of many layers, which is what modern language models use.
What is the difference between generative AI and an LLM?
Generative AI is any AI that creates new content, including text, images, audio, video and code. An LLM is one type of generative AI that works with text and code; image and video generators use other model types such as diffusion models.
What does GPT stand for?
GPT stands for generative pre-trained transformer: a model that generates text, was pretrained on a large text corpus, and uses the transformer architecture. OpenAI uses it as a product name, but the idea describes most modern LLMs.
What is the difference between an AI agent and a chatbot?
A chatbot answers your messages. An agent can also take actions with tools, such as running code, searching, editing files or calling APIs, and keeps going in a loop until a task is done or it needs your input.
