Glossary.
A living glossary of AI terms, linked to the tools that use them.
Agent
an AI system that doesn't just answer, but acts: given a goal, it plans and carries out steps (calling tools, writing files, running code) with some independence.
Updated 2026-06-03
Agentic
describes AI that takes initiative and completes multi-step work on its own, rather than waiting for the next instruction.
Updated 2026-06-03
AGI (Artificial General Intelligence)
a hypothetical AI that matches or exceeds humans across essentially all cognitive tasks.
Updated 2026-06-03
API
the interface that lets one piece of software talk to another.
Updated 2026-06-03
Benchmark
a standard test used to compare models (e.g.
Updated 2026-06-03
Chain-of-thought
prompting (or training) a model to work through its reasoning step by step before answering, which improves results on harder problems.
Updated 2026-06-03
Closed weights
a model whose parameters are kept private and accessed only through a provider's API (e.g.
Updated 2026-06-03
Compute
the raw processing power (mostly GPUs) used to train and run models.
Updated 2026-06-03
Context window
how much text a model can consider at once (prompt plus its reply), measured in tokens.
Updated 2026-06-03
Distillation
training a smaller, cheaper model to mimic a larger one, capturing much of its ability at a fraction of the running cost
Updated 2026-06-03
Embedding
a numerical representation of text (or images) that captures meaning, so software can measure how similar two things are.
Updated 2026-06-03
Fine-tuning
further training a pre-trained model on specific data to specialise it for a task, tone or domain.
Updated 2026-06-03
Foundation model
a large model trained broadly on huge data, then adapted to many downstream uses.
Updated 2026-06-03
Frontier model
the most capable models available at any given moment (currently the top Claude, GPT, Gemini tiers).
Updated 2026-06-03
Function calling (tool use)
a model's ability to invoke external tools — search, code, a calculator, an API — rather than relying only on what it learned in training.
Updated 2026-06-03
Guardrails
the limits and safety measures placed around a model to keep its outputs within bounds.
Updated 2026-06-03
Hallucination
when a model states something false with full confidence.
Updated 2026-06-03
Inference
running a trained model to produce an output (as opposed to training it).
Updated 2026-06-03
Jailbreak
a prompt crafted to make a model bypass its safety guardrails.
Updated 2026-06-03
Latency
the delay between asking a model and getting a response.
Updated 2026-06-03
LLM (Large Language Model)
a model trained on vast text to predict and generate language.
Updated 2026-06-03
MCP (Model Context Protocol)
an open standard for connecting AI models to external tools and data sources in a consistent way.
Updated 2026-06-03
Multimodal
a model that handles more than text — images, audio, video — in and out.
Updated 2026-06-03
Open weights
a model whose parameters are released publicly, so anyone can run, inspect or adapt it (e.g.
Updated 2026-06-03
Parameters
the internal values a model learns during training; loosely, a measure of its size and capacity (often in billions).
Updated 2026-06-03
Pre-training
the first, largest training phase, where a model learns general patterns from enormous data before any specialisation.
Updated 2026-06-03
Prompt
the input you give a model.
Updated 2026-06-03
RAG (Retrieval-Augmented Generation)
giving a model relevant documents to draw on at answer time, so it responds from your data rather than memory alone.
Updated 2026-06-03
Reasoning model
a model trained to "think" longer before answering — generating intermediate steps — to do better on hard maths, code and logic.
Updated 2026-06-03
RLHF (Reinforcement Learning from Human Feedback)
training a model using human ratings of its outputs to make it more helpful and better-behaved.
Updated 2026-06-03
System prompt
hidden instructions that set a model's role, rules and tone before the user's conversation begins.
Updated 2026-06-03
Token
the unit models read and write in — roughly a word-piece.
Updated 2026-06-03
Vector database
a store optimised for embeddings, so software can find the most semantically similar items fast.
Updated 2026-06-03