AI Glossary
AI terms, explained in plain English
46 definitions of the AI concepts behind today's headlines — each linked to a deeper guide when you want to go further.
A
- A2A (Agent2Agent)
- An open protocol that lets AI agents from different vendors discover each other, communicate and delegate tasks. It complements MCP, which connects agents to tools and data.AI agents for business
- Agentic AI
- AI systems that pursue a goal by planning steps, using tools and checking results, rather than only answering a single prompt.AI agents for business
- AI agent
- Software that uses an AI model to decide what to do next in pursuit of a goal — calling tools such as search, databases, APIs or a browser, and completing multi-step tasks, often with human approval for important actions.AI agents for business
- AI literacy
- The practical understanding of what AI tools can and cannot do, how to use them well and how to check their output. The EU AI Act requires organisations using AI to take steps to ensure staff AI literacy.AI skills to learn in 2026
- AIGC
- AI-Generated Content: text, images, video, audio, music and avatars created wholly or partly by AI. The term is especially common in China and across Asia.What is AIGC?
B
- Benchmark
- A standard test used to compare AI models, such as SWE-bench for coding or OmniDocBench for document parsing. Useful as a guide, but results on your own data matter more.Best open-source LLMs
C
- C2PA (Content Credentials)
- An open standard for signed metadata that records how a piece of media was created and edited — including whether AI was involved.AIGC and labelling rules
- Chunking
- Splitting documents into smaller passages before indexing them for retrieval. Splitting by headings and sections usually works better than splitting by a fixed number of characters.RAG vs fine-tuning
- Computer use
- The ability of an AI agent to operate a computer or browser the way a person does — reading the screen, clicking, typing and navigating applications.What's new in AI in 2026
- Context engineering
- Deciding exactly what information an AI model sees for a task — instructions, retrieved documents, tool results and history — so it has what it needs and nothing that distracts it.AI skills to learn
- Context window
- The maximum amount of text (measured in tokens) a model can consider at once, including the prompt and its answer. Several 2026 models support one million tokens.What's new in AI in 2026
D
- Deepfake
- Realistic AI-generated or manipulated images, audio or video of real people or events. The EU AI Act requires deepfakes to be labelled.AIGC and labelling rules
E
- Embedding
- A list of numbers (a vector) that represents the meaning of a piece of text or an image, so that similar meanings sit close together. Embeddings power semantic search and RAG.Open-source embedding models
- EU AI Act
- The European Union's risk-based AI regulation. Bans and AI-literacy duties apply since 2025, transparency duties since August 2026, and the main high-risk rules from December 2027.EU AI Act guide
- Evals (evaluations)
- Test sets and metrics used to measure whether an AI feature works — and to catch regressions when prompts, data or models change.RAG vs fine-tuning
- Excessive agency
- A security risk where an AI system has more permissions or autonomy than its task requires, so a mistake or attack can cause real damage.AI security guide
F
- Fine-tuning
- Continuing to train an existing model on your own examples to change its default behaviour — format, tone or a narrow task. It is a poor way to add facts.RAG vs fine-tuning
- Foundation model
- A large model trained on broad data that can be adapted to many tasks, such as GPT, Claude, Gemini or Llama-family models.
- Frontier model
- The most capable AI models available at a given time, typically from labs such as OpenAI, Anthropic and Google.What's new in AI in 2026
G
- GEO (generative engine optimisation)
- Optimising content to be used and cited in AI-generated answers from Google AI Mode, ChatGPT and Perplexity. Google says it is still SEO: expert, original, well-structured content.How to rank in AI search
- GGUF
- A file format for storing (often quantised) model weights, used by llama.cpp and Ollama to run models efficiently on CPUs and consumer GPUs.Self-hosting LLMs
- Guardrails
- Rules and checks around an AI system — input filtering, output validation, permission limits and human approval — that keep it safe and on task.AI security guide
H
- Hallucination
- When an AI model states something false or invented with confidence. Retrieval with citations, evals and human review reduce it.RAG vs fine-tuning
- Hybrid search
- Combining semantic (embedding) search with keyword search, so retrieval catches both meaning and exact terms such as product codes or names.RAG vs fine-tuning
I
- Inference
- Running a trained model to produce an output. Inference cost and speed are what you pay for when you use an AI API or host a model.Self-hosting LLMs
K
- KV cache
- Memory a model uses to store intermediate results for the text it has already processed. It grows with context length and concurrent users, and often dominates GPU memory needs.Self-hosting LLMs
L
- LLM (large language model)
- An AI model trained on large amounts of text to understand and generate language, such as GPT, Claude, Gemini, Qwen or DeepSeek.Best open-source LLMs
M
- MCP (Model Context Protocol)
- An open standard for connecting AI applications to tools and data through one common interface, supported by ChatGPT, Claude, Gemini, Copilot and VS Code.What is MCP?
- Mixture of experts (MoE)
- A model architecture where only part of the network (the relevant 'experts') is active for each token, so a very large model can run with much less compute per token.Best open-source LLMs
- Multimodal AI
- AI that works across several types of data — text, images, audio and video — in a single model.Vision-language models
O
- OCR (optical character recognition)
- Converting text in images and scanned documents into machine-readable text. Modern OCR uses vision-language models that also understand layout and tables.Best open-source OCR models
- Open-weight model
- A model whose trained weights can be downloaded and run on your own infrastructure. Licences vary — not every open-weight model is open source.Open-source LLM licences
P
- Prompt engineering
- Writing instructions, examples and constraints that get reliable results from an AI model.AI skills to learn
- Prompt injection
- An attack where malicious instructions — typed by a user or hidden in a web page, email or document — override what an AI system was told to do.AI security guide
Q
- Quantisation
- Storing model weights at lower precision (such as 8-bit or 4-bit) to cut memory and cost, with modest quality loss.Self-hosting LLMs
R
- RAG (retrieval-augmented generation)
- Adding a search step before a model answers, so it responds using relevant passages from your own documents or data — ideally with citations.RAG vs fine-tuning
- Reasoning model
- A model that works through a problem step by step before answering, trading extra time and tokens for better results on complex tasks.
- Reranker
- A model that re-scores retrieved passages by true relevance before they are passed to the LLM — often the cheapest way to improve RAG quality.Embeddings and rerankers
S
- Self-hosting
- Running AI models on your own servers or cloud account instead of calling a provider's API — for privacy, control or cost at scale.Self-hosting LLMs
- Speech-to-text (ASR)
- Automatic speech recognition: converting spoken audio into text, as done by models such as Whisper and Parakeet.Open-source speech models
- SynthID
- Google's invisible watermarking technology for AI-generated images, video, audio and text, used to identify content made with Google's AI tools.AIGC and provenance
T
- Text-to-speech (TTS)
- Generating natural-sounding speech from text. Modern TTS models can clone a voice from seconds of audio — which requires the speaker's consent.Open-source TTS models
- Token
- The unit of text a model reads and writes — roughly three-quarters of an English word on average. AI pricing and context windows are measured in tokens.
V
- Vector database
- A database optimised for storing embeddings and finding the most similar ones quickly — a core component of semantic search and RAG.RAG vs fine-tuning
- Vibe coding
- Building software mainly by describing what you want to an AI coding agent and accepting its output, with little manual coding. Fast for prototypes; risky without review and tests.Best AI coding tools
- VLM (vision-language model)
- A model that understands images and text together — describing photos, reading documents and screens, and in many cases understanding video.Open-source VLMs
From terminology to working AI
We build AI agents, RAG assistants and AI features into real products.