AI Glossary

Also called AI agent, Agentic AI

A system where a model decides its own next steps, uses tools to carry them out, checks the results, and keeps going until the goal is met.

The cycle an agent repeats. It decides what to do, acts with a tool, looks at the result, then decides again.

A precise set of steps for solving a problem. In AI the word often means the training method, and the model is what that method produces.

Also called AI safety

The work of making AI systems pursue the goals and values their makers intend, and not harmful shortcuts to them.

Also called Application programming interface

A way for one program to talk to another. Developers use a model’s API to send it text from their own code and get a reply back, without going through the chat app.

The app is what you use, like ChatGPT or the Claude app. The model is the engine inside it. One app can switch between several models, and the app adds extras of its own, such as web search and memory.

Also called AGI

A hypothetical AI that could learn and do almost any mental task a person can. There is no agreed test for it, so treat headlines saying it has arrived with care.

Also called AI

Software that does tasks we used to think needed human judgement, like understanding language or recognising what is in a photo. Most modern AI learns how to do the task from examples instead of following rules a person wrote.

Also called Self-attention

The step where each token looks back at the tokens before it and decides which ones matter to it. It is how a model works out what ‘it’ refers to in a long sentence.

One of many attention processes running side by side in each layer. Different heads tend to track different relationships, such as grammar or who did what to whom.

The tendency to trust an automated answer too much and stop checking it, even when it is wrong.

The method training uses to work out how much each number in a model contributed to a wrong answer, so every one can be nudged in the right direction.

A standard test used to compare models, such as a fixed set of maths problems or coding tasks. It is a rough guide. Testing on your own examples tells you more.

When an AI system treats some groups unfairly or gets things wrong for them more often, usually because its training data left them out or showed them in a skewed way.

Having a model work through a problem step by step before it answers. It often improves accuracy on tasks that need several steps.

Also called AI assistant, Chat assistant

An app you talk to in plain language. A modern chatbot is an app wrapped around a large language model, with extras such as memory, web search and file uploads.

Also called Proprietary model

A model you can only use through the company’s own app or API. You never get the model file itself.

The field of AI that works with images and video, such as reading text in a photo, recognising objects or spotting defects on a production line.

Deciding what information goes into the model’s context for each request, such as instructions, documents, memory and tool results. It matters as much as the wording of the prompt.

The drop in quality as a model’s input gets very long. It may miss or muddle details even though they fit inside the context window.

Also called Context length

The most text a model can take into account at once, counted in tokens. It covers your messages, any files, earlier replies and the new reply. Anything past the limit is cut off or never seen.

Open legal questions about whether training on copyrighted work is allowed and who owns what an AI produces. The answers differ by country and are still being settled in court.

Machine learning that uses large neural networks with many layers. It is behind almost every recent leap in AI, from speech recognition to chat assistants.

Realistic fake video, audio or images of a real person, made with AI. They are used for scams and misinformation, so verify surprising clips before sharing them.

The kind of model behind most image and video generators. It starts from random noise and removes it step by step until a picture that matches the prompt appears.

Training a smaller model to copy the answers of a larger one, so it keeps much of the skill at a fraction of the size and cost.

Also called Chunking

Cutting long documents into smaller passages before storing them for search. Good splits keep each passage meaningful on its own. These passages are not the same thing as the tokens a model reads.

A list of numbers that captures what a piece of text means, so texts about similar things get similar numbers. Search systems compare embeddings to find passages that match a question even when the words differ.

The European Union’s law on AI. It sorts uses of AI by risk and sets stricter rules for high-risk ones, such as AI used in hiring or lending decisions.

Also called Evaluation

A repeatable test of how well a model or prompt performs across many example inputs. It shows whether a change actually made things better or only looked better once.

Also called MLP

The part of each transformer layer that works on every token on its own, after attention. Much of what a model has learned about the world appears to be stored here.

Also called Zero-shot, Examples in a prompt

Including a few worked examples in a prompt so the model follows their pattern. Zero-shot means giving no examples at all.

Extra training that adapts an existing model to a narrower job or style, using a smaller set of your own examples.

One of the most capable models available at a given time, from the labs pushing the limits of what AI can do.

Also called GenAI

AI that creates new content, such as text, images, audio, video or code, instead of only sorting or labelling what already exists.

Also called Graphics processing unit

A chip built to do many simple calculations at the same time. That makes it well suited to training and running AI models.

Also called Citations

Making sure an answer rests on sources you supplied, and ideally points to them. A grounded answer can be checked against its source.

Limits placed around an AI system to keep it safe and on task, such as blocked actions, spending caps or checks on what it outputs.

Also called Confabulation

When a model states something false with full confidence, such as an invented fact, quote or source. It happens because the model produces likely-sounding text, and likely is not the same as checked.

A design where a person reviews or approves certain steps before the system acts, especially steps that are costly or hard to undo.

Using a trained model to produce an answer. Training happens once and costs a great deal. Inference happens every time someone sends a message.

Also called Token pricing

What an API charges, usually quoted per million tokens. Tokens the model writes normally cost several times more than tokens you send it.

A second round of training on examples of requests paired with good responses. It turns a raw text predictor into something that follows instructions.

Research into what is actually happening inside a model, such as finding the internal patterns that stand for particular ideas.

A prompt written to get a model to ignore its safety rules.

A plain-text format for structured data, made of labelled fields inside curly braces. It is the usual way programs swap data with a model.

Also called Training cutoff

The date a model’s training data ends. It knows nothing after that date unless the app searches the web or you paste the information in.

Also called LLM, Language model

A generative model trained on huge amounts of text to predict what comes next. ChatGPT, Claude and Gemini are apps built on large language models.

How long you wait for a response. It is usually measured as the time until the first token appears, then as tokens per second.

One stage of processing inside a model. A transformer stacks dozens of layers, and each one refines the numbers for every token a little further.

The raw scores a model gives every possible next token, before they are turned into percentages.

The research finding that models use information at the start and end of a long input better than information buried in the middle.

Also called ML

A way of building software where the program learns patterns from examples instead of being given rules. Show it thousands of emails marked spam or not spam, and it learns to sort new ones itself.

Also called Output limit

A setting that caps how long a reply can be. The model stops when it reaches the limit, even in the middle of a sentence.

Information an AI system keeps and brings back later, such as earlier steps in a task or facts from past chats. The model does not remember anything itself. The app stores it and sends it back in.

Also called MoE

A design where a model contains many smaller sub-networks, called experts, but uses only a few of them for each token. The total size can be huge while only a fraction is working at once.

A kind of content a model works with, such as text, images, audio or video.

The trained system that does the actual work. It is a large file of numbers learned during training. You give it an input and it produces an output.

A model’s official fact sheet. It lists details such as the context window, knowledge cutoff, prices, intended uses and known limits.

Also called MCP, Connector

An open standard for connecting AI apps to outside tools and data. A service builds one MCP server, and any app that supports MCP can then use it.

Also called Model tier

A set of related models from one company, usually offered in sizes or tiers. A larger tier is more capable, and a smaller one is faster and cheaper.

Able to take in or produce more than one kind of content, such as text, images, audio and video.

AI built for one kind of task, such as translating text or spotting faults in scans. It can be excellent at that task and of no use for anything else.

A model built from many simple units connected in layers, loosely inspired by the brain. Every connection has a number, and training adjusts those numbers until the network gives good answers.

Also called Next-chunk prediction

How a language model writes. It looks at everything so far, predicts a fitting next token, adds it, and repeats. That is why replies appear a piece at a time.

A stricter label than open weight. It means the weights, the code and enough detail about the training data are shared for others to study and rebuild the model. Few large models qualify.

Also called Open model

A model whose parameters you can download and run on your own hardware. Its licence still decides what you are allowed to do with it.

Also called Weights

The numbers inside a model that training sets. More parameters usually means a more capable model that also needs more memory to run. GPT-3 had 175 billion of them.

Also called Positional encoding

Extra numbers added to each token so the model knows the word order. Without it, ‘dog bites man’ and ‘man bites dog’ would look the same.

The first and largest stage of training, where a model reads a vast amount of text and learns to predict the next token. The result knows a lot but is not yet a helpful assistant.

What happens to what you type into an AI app. Some apps store chats or use them for training unless you switch that off, so check the settings before sharing anything sensitive.

Whatever you send a model to get a response. It can be a question, instructions, examples, documents, or all of these together.

Reusing the already-processed opening part of a prompt when it repeats across requests. Repeat requests become faster and cheaper.

Splitting a big task into several prompts, where the output of one becomes the input of the next. Each step is simpler and easier to check.

The skill of writing and testing prompts so a model gives useful, reliable results. It is less about magic phrases and more about being clear, specific and checking the output.

An attack where hidden instructions inside a web page, document or email trick a model into doing something its user never asked for. It is a serious risk for agents that read outside content.

Also called Quantization

Storing a model’s numbers with less precision so it needs less memory and runs faster, usually at a small cost in quality.

Also called Thinking model, Extended thinking

A model trained to work through a problem step by step before giving its final answer. It is often better at maths, code and planning, and it is slower and costs more per answer.

Deliberately attacking an AI system to find its weaknesses before real users or attackers do.

Also called RL

Learning by trial and reward. The system tries something, gets a score, and learns to do more of what scored well.

Also called RLHF

Training where people compare a model’s answers and the model is nudged toward the ones they prefer. It shapes tone, helpfulness and when the model refuses.

A second pass that reorders search results so the most useful passages come first, before they reach the model.

The search step in RAG, where the system finds the passages most likely to answer the question.

Also called RAG

A setup that first searches your documents for relevant passages, then hands them to the model along with the question. The answer is based on your material instead of only on what the model remembers.

Also called Traditional software

Software that follows steps a person wrote out in advance. It is predictable, but it has no answer for a case nobody wrote a rule for.

Picking the next token at random, weighted by its percentage. Likely tokens win most of the time and less likely ones get an occasional turn, which is why the same question can get different answers.

The finding that models improve in a predictable way as you add parameters, data and computing power. It explains why labs keep building bigger models, and why each gain costs more.

Searching by meaning instead of exact words. A question about ‘time off’ can find a passage about ‘annual leave’.

Also called SLM, Local model

A model small enough to run on a laptop or phone. It is cheaper and keeps your data on your device, but it is less capable than the largest models.

The step that turns raw scores into percentages that add up to 100. High scores become big shares and low scores become tiny ones.

Also called Transcription, STT

Turning spoken audio into written text. It struggles most with heavy background noise, strong accents it has seen little of, and names.

Also called JSON output

Getting a model to reply in a fixed format, such as JSON, so another program can read it. A tidy format does not guarantee correct content, so the values still need checking.

A model’s habit of agreeing with you or praising your ideas because people tend to rate that highly, even when you are wrong.

Also called System instructions

Instructions set before a conversation starts, usually by the app or the developer. They shape the model’s role, tone and rules for every reply that follows.

A setting that controls how adventurous sampling is. A low temperature makes the favourite token win more often, so replies are steadier. A high temperature gives unlikely tokens more of a chance, so replies vary more.

Also called Image generation, Text-to-video

Generating a picture from a written description. Text-to-video does the same for moving footage.

Also called TTS, Speech synthesis

Turning written text into spoken audio. Modern voices can sound very close to a real person.

Also called Chunk, Tokens

The piece of text a language model actually reads and writes. A token is often a common word, part of a longer word, or a punctuation mark. In English one token is roughly three quarters of a word. Our lessons call them chunks.

Also called Tokeniser

The tool that splits text into tokens and gives each one a number. It was built from examples, so common words get a single token and rare words or names get split into several.

Also called Function calling, Tool calling

Letting a model use outside functions, such as web search, a calculator or a calendar. The model asks for the tool, your code runs it, and the result goes back to the model.

A setting that limits sampling to a fixed number of the most likely tokens, such as the top 40.

Also called Nucleus sampling

A setting that limits sampling to the smallest group of likely tokens whose percentages add up to a chosen amount, such as 90 percent. Everything outside that group is ignored.

A step-by-step record of everything an agent did, including each prompt, tool call and result. Reading it is how you find the first step that went wrong.

Showing a model many examples and adjusting its numbers a little each time it gets one wrong. After enough rounds, it gets good at the task.

The examples a model learns from. A model can only be as accurate, fair and up to date as the data it was trained on.

The design behind almost every modern language model, introduced by Google researchers in 2017. It reads all the tokens of your text at once and uses attention to work out how they relate.

A list of numbers. Inside a model, every token becomes a vector, and tokens used in similar ways end up with similar numbers.

Also called Vector store

A database built to store embeddings and quickly find the ones closest to a search. It is the usual storage behind a RAG system.

Building software by describing what you want to an AI and accepting the code it writes, often without reading it closely. It is quick for prototypes and risky for anything that matters.

Making a synthetic voice that sounds like a specific person from a sample of their speech. Using someone’s voice responsibly needs their consent.

Hidden or visible marks added to AI-generated content so it can be identified later. Some can be stripped out, so a missing watermark proves nothing.

A fixed sequence of steps written in advance, even if models carry out some of them. It is more predictable than an agent, which chooses its own path.