AI Glossary

AGI is a hypothetical AI able to handle any intellectual task at human level, not only the ones it was trained for. No such system exists, and there is no agreed definition by which one would be recognised.

An AI agent is a model given tools and the right to act in a loop: it decides what step to take, takes it, looks at the result and carries on until the task is done.

A hallucination is a confidently delivered but invented answer: a quotation, link, paper or figure that does not exist. It is not a malfunction but a direct consequence of the design: the model predicts a plausible continuation rather than verifying a fact.

Chain-of-thought is the technique of having a model work through intermediate steps before answering. On counting and logic tasks it measurably improves accuracy compared with answering straight away.

The context window is how much text a model holds at once: your request, the whole prior conversation and the answer. It is measured in tokens, and anything outside the window stops existing for the model.

A diffusion model creates an image by gradually removing noise: it starts from random pixels and over dozens of steps turns them into a picture matching the request. Most image generators work this way.

An embedding represents text, an image or audio as a list of numbers in which things close in meaning end up close together. Semantic search, recommendations and the whole RAG pattern rest on it.

Few-shot means putting a handful of example answers into the request so the model copies their format. Zero-shot is the same request with instructions only and no examples.

Fine-tuning adapts an existing model on your own examples so it takes on a required style, format or narrow task. It changes the weights, unlike prompting and RAG, which work only with context.

A foundation model is a large model trained on broad data and intended not for one task but as a base for many. It is adapted with prompts, fine-tuning or tools instead of training something new from scratch.

Inference is a trained model doing its job: it takes a request and produces an answer. Unlike training, the weights do not change, and inference is what you pay for when using an AI service.

An LLM is a neural network trained on vast amounts of text that predicts the next fragment of a word. Everything else grows out of that single mechanism: answers, translation, code, reasoning.

MCP is an open protocol defining one standard way to connect models to external data and tools. Created by Anthropic and released openly: an integration is written once and works with any application that supports the protocol.

Model weights are the billions of numbers that training produced. They are the model: an architecture without weights is useless, and weights without it are just a file.

A multimodal model works not only with text but with images, audio or video, inside the same mechanism. You can show it a chart, ask about it in words and get a written answer.

A neural network is a computational model built from layers of simple units connected by numeric weights. Nothing in it is programmed by hand: the behaviour comes from tuning those weights on examples.

Prompt engineering is writing requests so a model does the right thing. Not magic phrasings but ordinary precision: a clear task, the context it needs, a defined output format, and a criterion for judging the result.

RAG is an approach where the model first retrieves relevant fragments from your own data and answers from them. It is how a model works with fresh or private material it was never trained on, and it invents markedly less.

RLHF trains a model on human preferences: people compare candidate answers, a separate judge model learns from those comparisons, and the main model is tuned against its scores. It is how a model is taught to be useful rather than merely plausible.

A token in AI is a chunk of text the model works with instead of letters or words. Usually it is part of a word, about four characters in English. Tokens are the unit for both context length and price.

The transformer is a neural network architecture introduced in the 2017 paper "Attention Is All You Need". Nearly every modern language model is built on it, and its central idea is the attention mechanism.

A vector database stores embeddings and finds the closest ones by meaning, fast. It is the technical basis of RAG: it answers the question of which fragments to place into the model's context.