Reference

Glossary

Plain-language definitions of common AI terms, for readers and answer engines.

AI agent: A system that uses a model to pursue a goal across several steps, deciding what to do next and often calling external tools such as search, code execution or an internal API before it reports a result. Example: an agent that reads a bug report, opens the repository, edits a file and runs the tests. Do not confuse this with a chatbot or a single-turn classifier, which replies once without planning or acting on other systems.
Alignment: The effort to make AI systems behave according to human intentions and values.
Attention: A mechanism that lets a model weigh which parts of the input matter most for each output.
Chain-of-thought: A prompting or training approach in which a model works through intermediate reasoning steps before answering.
Citation: A citation is a reference attached to an output that points to where a statement, a quotation or a figure came from, usually naming a document, a page or a source so that a reader can check it. Example: an answer that gives a figure and links the paragraph of the report it was taken from. Do not confuse a citation with grounding or with accuracy: a citation shows only which source a system attached, and the same system can attach a plausible-looking citation to a claim that the source does not actually support.
Commercial use: Commercial use is any use of a model, a tool or its output that is meant to earn revenue or support a business, such as selling a product built with it, embedding it in a paid service or applying it to billable client work. Example: a studio generates images with a model and delivers them to paying clients. Do not confuse commercial use with paid access: buying a subscription does not by itself grant commercial rights, and some terms that cost nothing still allow commercial work.
Context window: The maximum amount of text, measured in tokens, that a model can take into account in a single request, covering the instructions, the conversation so far and any documents supplied with it; material beyond that limit is left out or truncated. Example: a long contract that exceeds the window has to be split or summarised before questions can be asked about all of it. Do not confuse this with a memory or with the training data: the window is a per-request limit and does not store everything the model has ever seen.
Diffusion model: A generative model that produces images or other media by iteratively removing noise from a random signal.
Distillation: Training a smaller model to imitate a larger one, keeping much of its capability at lower cost.
Embedding: A numeric vector that represents a piece of text, an image or another item, so that items with similar meaning end up close together in a high-dimensional space, which makes semantic search and clustering possible. Example: a help centre stores one vector per article so that a user question can be matched to the nearest articles. Do not confuse this with a prompt or a summary: an embedding is not readable text, and its numbers cannot be interpreted by reading them.
Few-shot learning: Guiding a model with a handful of examples in the prompt instead of retraining it.
Fine-tuning: Additional training of an already pretrained model on a narrower dataset so that it follows a particular task, tone or domain more closely. Example: a company trains a base model on its own curated question-and-answer pairs so that replies match its house style. Do not confuse this with retrieval-augmented generation, which leaves the weights untouched and adds documents when answering, or with prompt engineering, which only rewrites the instructions.
Foundation model: A large model trained on broad data that can be adapted to many downstream tasks.
Free tier: A free tier is a no-charge level of access that a service offers alongside its paid plans, usually limited by usage quotas, available features, seats or time so that people can try the product or keep light use at no cost. Example: an interface that allows a set number of requests each month before a paid plan is needed. Do not confuse a free tier with free software or with open source: it is a commercial offer that can be changed or withdrawn, and it says nothing about the license of the model behind it.
Function calling: A model capability that returns structured calls to developer-defined functions instead of plain text.
Grounding: Connecting a model's responses to verifiable external sources such as documents or search results.
Guardrails: Rules and filters that constrain an AI system's outputs to keep them safe, accurate, and on-topic.
Hallucination: An output in which a model states something false or unsupported, or invents a detail such as a citation, a figure or a function name, while writing as fluently and confidently as when it is correct. Example: a summary that cites a section of a document that does not contain the quoted claim. Do not confuse this with a lie or with bias: hallucination comes from how the model generates text, not from an intention to deceive or from a particular viewpoint.
Inference: Running an already trained model on new input to produce an output, such as a prediction, a translation or a generated image. This is the step where a service consumes compute for each request, and it is separate from training, which happened earlier. Example: an app sends an uploaded photo to a model and receives a caption back within a second. Do not confuse this with training or fine-tuning, which adjust the model's parameters over a dataset and are typically far more expensive per run.
Large language model (LLM): A neural network trained on large amounts of text to predict and generate language.
License: A license is the legal permission that a rights holder grants for how a model, its weights or a piece of software may be used, copied, modified and redistributed, and it sets the conditions attached to those rights instead of describing the technology. Example: one model may be released for research use only, while another allows commercial products but requires modified versions to keep the same terms. Do not confuse a license with open weights or with open source: weights can be downloadable under terms that forbid commercial use, and a permissive code license says nothing about the training data.
Mixture of experts (MoE): A model architecture that routes each input to a subset of specialized subnetworks to save computation.
Model Context Protocol (MCP): An open protocol that standardises how assistants and other model-driven applications connect to external data sources and tools, so that one connector can be reused by several clients instead of being rewritten for every combination of application and service. Example: a server that exposes a company's ticket system can be attached to a desktop assistant and to a coding tool without changes. Do not confuse this with a model or with a single vendor's API: it describes how tools are offered and called, and it says nothing about the quality of the model on the other side.
Multimodal: Describing a model that can process or generate more than one type of data, such as text and images.
Open weights: Model parameters that are published so that others can download, run and often modify them, in contrast to a model that is reachable only through a vendor's hosted service. Example: a team downloads published weights and runs them on its own machines to keep data inside its own network. Do not confuse this with open source: weights alone say nothing about the licence, the training data or the training code, so a model with open weights is not automatically open source.
Prompt: The text or other input given to a model to instruct or elicit a response.
Prompt engineering: The practice of designing and refining prompts to get more reliable or useful model outputs.
Quantization: Reducing the numerical precision of a model's weights so it runs faster and uses less memory.
Reasoning model: A model trained to spend extra computation thinking through a problem step by step before answering.
Red-teaming: Deliberately probing an AI system for harmful, unsafe, or unintended behavior before and after release.
Reinforcement learning from human feedback (RLHF): A training method that uses human preference judgments to steer a model toward more helpful behavior.
Retrieval-augmented generation (RAG): A technique that searches an external corpus for passages relevant to a question and passes those passages to a model, so the answer rests on retrieved sources instead of the model's stored knowledge alone. Example: a support assistant that finds the matching paragraph in a product manual and quotes it in its reply. Do not confuse this with fine-tuning, which changes the model's weights; retrieval-augmented generation leaves the weights untouched and changes only the text the model sees while answering.
Self-hosting: Self-hosting means running a model or an application on infrastructure you control, such as your own servers, a private cloud or a laptop, instead of reaching it through a vendor's hosted service. Example: a company installs open-weight models inside its own network so that prompts and documents never leave it. Do not confuse self-hosting with open weights: whether you may run a model yourself depends on its license, and a model whose weights are downloadable may still be offered mainly as a hosted service.
Small language model (SLM): A compact language model designed to run efficiently on devices or at low cost with competitive quality.
Speech recognition (ASR): Automatically converting spoken audio into written text.
Synthetic data: Artificially generated data used to train or evaluate models when real data is scarce or sensitive.
System prompt: A hidden instruction that sets a model's role, rules, and behavior before user input.
Temperature: A sampling setting that controls how random or predictable a model's outputs are.
Text-to-image: Generating images from written descriptions using a generative model.
Text-to-video: Generating video clips from natural-language descriptions using generative models.
Token: A unit of text, such as a word or word-piece, that a language model reads and generates.
Tool use (function calling): A model's ability to call external functions, APIs, or apps to complete a task.
Transformer: A neural network architecture based on attention that underlies most modern language and multimodal models.
Vector database: A database that stores embeddings and finds items by similarity, often used for retrieval in AI apps.
Vibe coding: A style of software development where developers describe goals in natural language and let AI generate the code.
Voice cloning: Voice cloning builds a synthetic voice from recordings of a person speaking, so that new text can be read aloud in a voice resembling theirs; methods range from a short sample to hours of audio. Example: a narrator records a few minutes once, then has the same voice read a corrected sentence without booking another session. Do not confuse this with speech recognition: recognition turns audio into text, while cloning turns text into audio that imitates a particular speaker.
World model: An AI system that learns an internal representation of an environment to predict how it changes over time.
Zero-shot learning: A model's ability to perform a task without having seen any examples of it during prompting.