An A-Z glossary of artificial intelligence terms explained in plain English, from tokens and temperature to RAG, RLHF and prompt injection — with links to tools where each concept matters.
New to all of this? Start with the AI Academy, then use this page as a reference. Longer explanations live in the Learning Center.
An AI system that plans and executes multi-step tasks.
An agent wraps a language model in a loop: it decides what to do next, calls a tool, reads the result, and repeats until the goal is met. Agents are what turn a chat model from something that answers questions into something that completes work.
Making a model behave the way people intend.
Alignment covers the training and guardrail techniques — instruction tuning, RLHF, safety filters — that push a model towards helpful, honest and harmless behaviour instead of merely statistically likely behaviour.
A programmatic interface for calling a model.
Instead of typing into a chat box, an API lets your software send a request and receive a structured response. Every tool on this platform talks to models over an API behind the scenes.
The mechanism that lets a model weigh context.
Attention lets each token look at every other token and decide which ones matter. It is the core idea behind the transformer architecture and the reason modern models handle long-range context so well.
A standard test used to compare models.
Benchmarks such as MMLU, GPQA and HumanEval score models on fixed question sets. They are useful signals but poor proxies for your specific workflow — always test on your own tasks.
Systematic skew inherited from training data.
If a dataset over-represents one viewpoint, language or demographic, the model reproduces that skew. Bias is why human review matters for hiring, lending, medical and legal outputs.
Making the model reason step by step.
Asking a model to show intermediate steps improves accuracy on maths, logic and multi-constraint problems, because each step conditions the next instead of forcing a single leap to the answer.
A conversational interface to a model.
A chatbot keeps a running message history so each turn has context. It is the most common AI interface, and the reason follow-up questions work without repeating yourself.
How much text a model can consider at once.
Measured in tokens, the context window covers your prompt, any attached documents and the model's own reply. Exceed it and the earliest content is dropped — which is why very long documents should be split into sections.
The architecture behind most image generators.
A diffusion model learns to reverse noise: it starts from random static and denoises step by step towards an image that matches your prompt.
A numeric representation of meaning.
An embedding maps text into a vector so that similar meanings sit close together. Embeddings power semantic search, recommendations and retrieval-augmented generation.
Teaching by example inside the prompt.
Including two or three worked examples of input and desired output is often more effective than describing the format in words, especially for tone and structure.
Further training a model on your own data.
Fine-tuning adapts a base model to a narrow domain or house style. It is powerful but expensive; for most use cases a strong prompt plus retrieval gets you most of the way.
A large general-purpose pretrained model.
Foundation models are trained on broad data and then adapted to many downstream tasks. GPT, Gemini and Claude families are all foundation models.
AI that produces new content.
Generative systems create text, images, audio, video or code from a prompt, rather than only classifying or predicting existing data.
Anchoring answers in real sources.
A grounded answer cites or is constrained by supplied documents or search results. Grounding is the main practical defence against hallucination.
Confident output that is factually wrong.
Models optimise for plausibility, not truth. Names, statistics, citations and legal or medical detail are the highest-risk categories and should always be verified.
Running a trained model to get output.
Inference is the cost you pay per request, as opposed to training which happens once. Faster, smaller models cost less per inference — the reason smart routing matters.
Regenerating part of an image.
You mask a region and describe what should be there instead. Inpainting is how objects are removed or replaced without redrawing the whole image.
Training a model to follow directions.
Base models complete text; instruction-tuned models follow requests. This step is what makes 'summarise this in three bullets' work reliably.
A prompt that bypasses safety rules.
Jailbreaks exploit roleplay, encoding or nested instructions to get restricted output. Platforms treat them as abuse, and our Acceptable Use Policy prohibits them.
Time until the response arrives.
Streaming reduces perceived latency by showing tokens as they are produced instead of waiting for the full answer.
Large Language Model.
A neural network with billions of parameters trained on text to predict the next token. Everything from drafting an email to explaining a stack trace is the same underlying operation.
Handling more than one type of input.
A multimodal model accepts text plus images (and sometimes audio or video) in the same request — the basis of OCR, image captioning and visual question answering.
Telling an image model what to avoid.
Listing unwanted elements — extra limbs, watermarks, text — steers generation away from common failure patterns.
Optical Character Recognition.
Extracting readable text from an image or scan. Modern vision models do this while preserving layout, tables and reading order.
Models whose parameters are downloadable.
Open-weight models can be self-hosted and inspected. They trade some peak capability for control, privacy and cost predictability.
Memorising training data instead of learning.
An overfitted model performs well on its training set and badly on anything new. In fine-tuning, it shows up as output that parrots your examples verbatim.
A learned weight inside the model.
Parameter count is a rough capacity signal, but training data quality and technique now matter more than raw size.
The instruction you give a model.
A good prompt supplies role, context, task, constraints and output format. Most quality complaints about AI output are really prompt problems.
The craft of writing effective prompts.
It is mostly a repeatable set of patterns — give context first, state constraints explicitly, change one variable per iteration — rather than magic phrasing.
Malicious instructions hidden in content.
If a model reads an untrusted web page or document, text in that content can try to hijack its behaviour. Treat retrieved content as data, never as instructions.
Retrieval-Augmented Generation.
RAG fetches relevant documents and puts them in the prompt so the model answers from real sources instead of memory. It is the standard pattern for company knowledge bases.
A model that thinks before answering.
Reasoning models spend extra inference compute on internal deliberation, which materially improves maths, code and multi-step analysis at the cost of speed.
Reinforcement Learning from Human Feedback.
Humans rank candidate responses and a reward model learns those preferences, which then steers the language model. RLHF is why assistants feel helpful rather than merely fluent.
A number that makes generation repeatable.
Fixing the seed with the same prompt reproduces the same image, which lets you change one variable at a time and actually see the effect.
Delivering output token by token.
Streaming shows text as it is generated so you can start reading — and stop the run — before it finishes.
Hidden instructions that set behaviour.
Every tool here ships with a system prompt encoding the role, quality bar and output format for that specific job, which is why you get usable results without prompt engineering.
How random the output is.
Low temperature produces focused, repeatable text; high temperature produces varied, creative text. Use low for extraction and code, higher for brainstorming.
The unit models read and write.
Roughly three-quarters of a word in English. Context limits, pricing and speed are all measured in tokens.
The architecture behind modern AI.
Introduced in 2017, the transformer uses self-attention to process sequences in parallel, which made training at today's scale practical.
Increasing image resolution with AI.
An upscaler reconstructs plausible detail rather than merely interpolating pixels, which is why AI upscales look sharper than a simple resize.
Storage optimised for embeddings.
It finds the nearest vectors to a query embedding in milliseconds, making it the retrieval half of a RAG system.
A model that can see images.
Vision models take pixels as input alongside text, enabling captioning, OCR, chart reading and visual troubleshooting.
Asking without any examples.
Modern instruction-tuned models handle most everyday tasks zero-shot; reach for examples when format or tone matters more than content.