AI glossary

This glossary collects the AI terms used across the AI Fundamentals articles. The terms are grouped by subject rather than alphabetically, so related ideas stay together. Each definition is deliberately short; the linked article explains it properly.

Foundations

• Artificial intelligence (AI) — The field of building machines that perform tasks requiring intelligence, such as understanding language or recognizing images. See AI basic concepts.
• Machine learning — The part of AI in which systems learn their behavior from examples instead of following written rules. See AI versus non-AI.
• Deep learning — Machine learning with neural networks of many layers, the basis of almost all recent progress. See A short history of Artificial Intelligence.
• Neural network — A model built from layers of simple calculations connected by adjustable weights. See AI basic concepts.
• Parameter — One of the numbers inside a neural network that training adjusts. Large language models have billions of them. See AI basic concepts.
• Model — The result of training: an architecture plus learned parameters that turn input into output. See AI basic concepts.
• Training — The phase in which a model learns from data by adjusting its parameters. See How large language models are built.
• Inference — Using a trained model to produce output for new input. Every chat answer is an inference. See The cost of AI.
• Supervised, unsupervised, and reinforcement learning — Learning from labeled examples, from unlabeled data, or from rewards for actions. See AI basic concepts.
• Generative AI — AI that produces new content such as text, images, audio, video, or code. See Generative AI beyond text.
• Artificial general intelligence (AGI) — A hypothetical AI that matches humans across most cognitive tasks. Its definition and timeline are disputed. See AI safety and risks.

Models and training

• Foundation model — A large model trained on broad data that can be adapted to many tasks. See Foundation Model Providers.
• Frontier model — A general-purpose model at or near the limit of current capability. See Frontier AI companies.
• Large language model (LLM) — A neural network trained on huge amounts of text to predict the next token. See How large language models are built.
• Transformer — The neural network architecture, introduced in 2017, behind nearly all large language models. See A short history of Artificial Intelligence.
• Attention — The mechanism in a Transformer that lets each token weigh the relevance of other tokens. See How large language models are built.
• Token — The unit of text a model reads and writes: a word, part of a word, or a symbol. About four characters of English on average. See AI basic concepts.
• Pre-training — The first and most expensive training stage, in which a model learns to predict the next token on a huge dataset. See How large language models are built.
• Fine-tuning — Further training of a pre-trained model on a smaller, specific dataset. See How large language models are built.
• Reinforcement learning from human feedback (RLHF) — Training a model toward answers that people rate as better. See How large language models are built.
• Reasoning model — A model trained to work through a problem step by step before answering. See AI basic concepts.
• Mixture of experts — A design in which only some of a model’s sub-networks are active for each token, which lowers the running cost. See How large language models are built.
• Scaling laws — The observation that model quality improves predictably with more parameters, data, and compute. See How large language models are built.
• Distillation — Training a small model to imitate a larger one. See How large language models are built.
• Quantization — Storing parameters with fewer bits so a model needs less memory and runs faster. See Open-weight versus closed models.
• Knowledge cutoff — The date after which a model has no knowledge, because its training data ends there. See AI basic concepts.
• Multimodal model — A model that handles more than one kind of input or output, such as text, images, and audio. See Generative AI beyond text.
• Diffusion model — A model that generates images or video by removing noise step by step. See Generative AI beyond text.
• Open-weight model — A model whose trained parameters are published for download. See Open-weight versus closed models.
• Closed model — A model available only as a service, with secret parameters. See Open-weight versus closed models.

Using models

• Prompt — Everything a model receives before it answers. See Prompting fundamentals.
• System prompt — Standing instructions from the application that shape every answer. See Prompting fundamentals.
• Context window — The maximum number of tokens a model can consider at once, input and output together. See AI basic concepts.
• Context engineering — Selecting and arranging the information that goes into a model’s context. See Prompting fundamentals.
• Few-shot prompting — Including a few input and output examples in the prompt to show the expected pattern. See Prompting fundamentals.
• Temperature — A setting that controls how varied or focused a model’s output is. See AI basic concepts.
• Structured output — A model answer in a defined data format, such as JSON matching a schema. See Prompting fundamentals.
• Embedding — A list of numbers that represents the meaning of a text, so similar texts get similar numbers. See Retrieval-augmented generation explained.
• Vector database — A database that stores embeddings and finds the ones closest to a query. See Retrieval-augmented generation explained.
• Retrieval-augmented generation (RAG) — Finding relevant passages in a document collection and giving them to the model with the question. See Retrieval-augmented generation explained.
• Hallucination — A fluent, confident answer that is false or invented. See AI versus non-AI.
• API — The interface through which software sends requests to a model and receives answers. See Foundation Model Providers.
• Benchmark — A fixed set of tasks with known answers used to compare models. See Evaluating AI models.
• Evaluation set — Your own collection of test inputs and expected results for a specific task. See Evaluating AI models.
• Model as judge — Using one model to grade the answers of another against written criteria. See Evaluating AI models.

Agents

• Agent — A system in which a model uses tools in a loop to reach a goal over several steps. See Agentic AI.
• Tool calling — A model requesting a function, such as a search or a file edit, that the application then runs. Also called function calling. See Agentic AI.
• Harness — The software around a model that runs the agent loop: instructions, tools, context, memory, and permissions. See Agentic AI.
• Computer use — A model operating applications through screenshots, mouse, and keyboard. See Agentic AI.
• Model Context Protocol (MCP) — An open standard for connecting agents to tools and data sources. See Agentic AI.
• Skill — A packaged set of instructions and scripts an agent loads when a task needs it. See Agentic AI.
• Workflow automation — Automating a fixed sequence of steps, some of which may use a model. See Agentic AI.
• Coding agent — An agent that reads a codebase, edits files, and runs commands and tests. See AI in software development.

Risks and governance

• Alignment — Making AI systems pursue the goals their developers and users actually intend. See AI safety and risks.
• Prompt injection — Instructions hidden in content a model reads, which try to take over its behavior. See AI safety and risks.
• Jailbreak — A prompt that tricks a model into ignoring its safety rules. See AI safety and risks.
• Red-teaming — Deliberately trying to make a model behave harmfully, to find and fix weaknesses. See AI safety and risks.
• Sycophancy — A model’s tendency to agree with and flatter the user. See AI safety and risks.
• Bias — Systematic unfairness in a model’s output, usually inherited from its training data. See AI safety and risks.
• Deepfake — Generated or manipulated media that shows a real person saying or doing something they did not. See Generative AI beyond text.
• Watermark — A hidden signal in generated content that identifies it as AI-made. See Generative AI beyond text.
• Zero data retention — A service option in which the provider does not store requests at all. See AI and data privacy.
• EU AI Act — The European Union’s law that regulates AI by the risk of its use. See AI regulation.
• General-purpose AI model — The EU AI Act’s term for foundation models, with their own set of obligations. See AI regulation.