AI glossary

Plain-English definitions of the AI terms we use. 19 and growing.

A

AI Agent

An AI agent is a system that does not just answer questions but takes actions to get things done. Given a goal, it can plan a series of steps, use tools such as a web browser, a calendar, or another piece of software, check its own progress, and carry on until the task is complete. The shift is from an assistant that tells you how to do something to one that goes and does it for you.

AI Alignment

AI alignment is the effort to make AI systems pursue what people actually intend, rather than something subtly or dangerously different. As models become more capable, the worry is not that they turn evil, but that they follow their instructions too literally or optimise for the wrong goal. Alignment research tries to ensure that an AI's behaviour reliably matches human values and intentions, and it is the foundation of most AI safety work.

C

Computer Vision

Computer vision is the branch of AI that interprets images and video. It lets machines identify objects, read text in a photograph, detect faces, spot defects on a production line, or guide a self-driving car. Where a camera simply records light, computer vision works out what that light actually shows, turning raw pixels into useful understanding of a scene.

D

Deep Learning

Deep learning is a kind of machine learning that uses neural networks with many layers stacked on top of one another. The word "deep" refers to that depth, not to any special cleverness. Each layer learns to spot slightly more complex patterns than the one below it, so the network can build up from simple features to rich understanding. It is the approach behind most of the AI breakthroughs of the past decade, from image recognition to chatbots.

Diffusion Model

A diffusion model is a type of AI that generates images and video by learning to remove noise. During training it is shown pictures that have been gradually smothered with random visual static, and it learns to reverse that process step by step. To create something new, it starts from pure static and cleans it up until a coherent image emerges. This is the technology behind popular tools such as Stable Diffusion.

E

Embedding

An embedding is a way of representing something, such as a word, sentence, or image, as a list of numbers that captures its meaning. Things with similar meanings end up with similar numbers, so a computer can measure how related two items are by comparing their embeddings. This quiet piece of machinery is what lets modern search understand intent rather than just matching keywords, and it is central to how systems like RAG find relevant information.

F

Fine-tuning

Fine-tuning means taking a model that has already been trained on a broad range of data and training it a bit further on a narrower, more specific set of examples. This specialises the model for a particular job, tone, or field without the enormous cost of building one from scratch. A general language model might be fine-tuned to sound like a legal assistant, to speak a company's house style, or to read medical notes.

Foundation Model

A large, general-purpose AI model trained on a huge amount of data, built to be adapted to many different tasks rather than one. Think of it as a broadly educated graduate you can then train up for a specific job.

G

Generative AI

Generative AI describes systems that create brand new content rather than only analysing existing content. Give it a prompt and it can produce written text, images, music, speech, or video that did not exist before. It is the technology behind chatbots that write essays, tools that turn a sentence into a picture, and apps that clone a voice. The results are original in the sense that they are freshly assembled, though they are learned from huge amounts of existing human work.

H

Hallucination

In AI, a hallucination is when a model states something false as though it were plainly true. It might invent a quotation, cite a study that does not exist, or confidently give the wrong date. This happens because language models are built to produce plausible sounding text, not to check facts, and they have no built-in sense of what is true. Hallucination is one of the main reasons AI answers should be verified before they are trusted.

I

Inference

Inference is the act of using a trained AI model to produce an answer. Training is the one-off, expensive process of teaching a model; inference is what happens every time afterwards when you actually ask it something and it responds. Because a popular model runs inference millions of times a day, it is the ongoing, recurring cost of operating an AI system, as opposed to the upfront cost of building it.

L

Large Language Model (LLM)

An AI system trained on enormous amounts of text to predict what words come next, which turns out to be enough to hold a conversation, draft an email, or write code. The technology behind tools like ChatGPT and Gemini.

M

Machine Learning (ML)

A way of building software that learns patterns from examples rather than following rules a programmer wrote by hand. Show it enough labelled photos of cats and it works out for itself what a cat tends to look like.

N

Natural Language Processing (NLP)

Natural Language Processing, or NLP, is the branch of AI concerned with human language: reading it, understanding it, and producing it. It covers everything from spam filters and search engines to translation apps, voice assistants, and chatbots. The goal is to let computers work with the messy, ambiguous, endlessly varied way that people actually write and speak, rather than the rigid commands computers traditionally understand.

Neural Network (NN)

A neural network is a computer system loosely inspired by the way brain cells connect. It is built from layers of simple units, called nodes, that pass signals to one another. By adjusting the strength of those connections while learning from examples, the network gradually gets better at a task, whether that is recognising a face, translating a sentence, or predicting the next word. Almost all modern AI is built on neural networks of one kind or another.

P

Parameters

Parameters are the adjustable internal values that a model learns during training. Each one is a small number that the model tunes as it works through examples, and together they store everything the model has learned. When you read that a system has "billions of parameters", that figure is a rough measure of its size and capacity, in the same spirit that engine size hints at the power of a car.

R

Retrieval-Augmented Generation (RAG)

Retrieval-Augmented Generation, usually shortened to RAG, is a technique that lets an AI look things up before it answers. Instead of relying only on what it memorised during training, the system first searches a chosen knowledge base, such as a company's documents, for relevant information, then uses what it finds to write its reply. This makes answers more accurate, more current, and far less likely to be made up.

T

Training Data

Training data is the collection of examples a model learns from. For a language model it might be billions of pages of text; for an image model, millions of pictures. The model studies this material to pick up patterns, so the quality, variety, and fairness of the data shape everything the model can do. As the old computing saying goes, garbage in, garbage out: a model is only ever as good as what it was taught with.

Transformer

A transformer is the neural-network design that powers almost every modern language model, including the systems behind popular chatbots. Introduced by Google researchers in a 2017 paper titled "Attention Is All You Need", its key trick is called attention: the ability to weigh up how much each word in a sentence relates to every other word. This lets the model understand context and long-range meaning far better than earlier designs could.