AI glossary
AI terms, explained in plain English
Every term starts with one sentence you can quote, then an example and the detail. Kept current as the field changes.
A
- Agent harnessharness, agent scaffold, agent scaffolding, agent runtimeAn agent harness is the software around a language model that turns it into a working agent, running the loop of model calls and tool calls, managing context and memory, enforcing permissions and recording what happens.
- Agentic AIagentic systems, agentic workflows, autonomous AIAgentic AI describes AI systems that act with a degree of autonomy toward a goal, planning steps, using tools and adapting to results with limited human direction, rather than only producing a single response to a single prompt.
- AI agentLLM agent, autonomous agent, agentAn AI agent is software that uses a language model to pursue a goal by deciding its own steps, calling tools such as search, code or business APIs, observing the results and repeating until the task is done or it needs a human.
- AI hallucinationhallucination, LLM hallucination, confabulationAn AI hallucination is output from a language model that sounds confident and plausible but is false, made up or unsupported by its sources, such as an invented citation, a wrong figure or a function that does not exist.
- AI OverviewsAI Overview, Google AI Overviews, SGE, Search Generative ExperienceAI Overviews are AI-generated summaries that Google Search shows above or among the regular results for some queries, answering the question directly with Gemini and linking to the web pages that support the answer.
- Artificial general intelligenceAGI, General AI, Strong AIArtificial general intelligence (AGI) is a hypothetical AI system that can understand, learn and perform the wide range of intellectual tasks a human can, at human level or beyond, rather than excelling only at narrow, specific tasks.
C
- Chain of thoughtCoT, Chain-of-thought prompting, Step-by-step reasoningChain of thought (CoT) is a technique where a language model writes out intermediate reasoning steps before its final answer, which improves accuracy on maths, logic and multi-step problems compared with answering in one jump.
- Context engineeringContext managementContext engineering is the practice of deciding what information goes into a language model's context window at each step, including instructions, retrieved documents, tool definitions, memory and conversation history, so the model has exactly what it needs and nothing that distracts it.
- Context windowcontext length, context size, context limitA context window is the maximum amount of text, measured in tokens, that a language model can take into account at once, including the system prompt, conversation history, documents, tool results and the response it is writing.
D
E
F
- Fine-tuningmodel fine-tuning, LLM fine-tuning, supervised fine-tuning, SFTFine-tuning is further training of an already pre-trained model on a smaller, task-specific dataset, so it adapts its behaviour, style or skills to a particular job without being trained from scratch.
- Function callingTool calling, Tool useFunction calling (also called tool calling or tool use) is the ability of a language model to return a structured request to run a function you defined, with arguments in JSON, so your code can execute it and pass the result back.
G
- Generative AIGenAI, Gen AI, GAIGenerative AI is a class of artificial intelligence models that create new content, such as text, images, audio, video or code, by learning patterns from large amounts of training data and producing original outputs in response to a prompt.
- GuardrailsAI guardrails, LLM guardrails, Safety filtersAI guardrails are the checks and controls placed around an AI model, such as input filters, output validation, topic limits, permission rules and human approval steps, that keep its behaviour safe, on-task and within policy.
I
L
M
- Mixture of expertsMoEA mixture of experts (MoE) is a model design that splits a neural network into many smaller "expert" sub-networks and activates only a few of them for each token, so a model can have a huge total parameter count while using a fraction of it per step.
- Model Context ProtocolMCPThe Model Context Protocol (MCP) is an open standard that lets AI applications connect to external tools, data sources and workflows through one common interface, so a tool built once as an MCP server works in any MCP-compatible client.
- Multimodal AImultimodal model, multimodal LLM, MLLM, vision-language model, VLMMultimodal AI is AI that can take in or produce more than one type of data, such as text, images, audio and video, within one model, so it can describe a photo, answer questions about a chart or respond in speech.
O
P
Q
R
- Reasoning modelThinking model, Large reasoning model, LRMA reasoning model is a language model trained to work through a problem in an internal chain of thought before giving its final answer, spending extra compute at inference time to do better on maths, coding, planning and other multi-step tasks.
- Reinforcement learning from human feedbackRLHFReinforcement learning from human feedback (RLHF) is a training method where people rank a model's answers, a reward model learns those preferences, and the language model is then optimised to produce answers the reward model scores highly.
- Retrieval-augmented generationRAGRetrieval-augmented generation (RAG) is a technique where an AI system first retrieves relevant passages from an external source, such as your documents or a search index, and gives them to a language model so its answer is grounded in that material.
S
- Small language modelSLM, Small LLM, Compact language modelA small language model (SLM) is a language model with relatively few parameters, typically from a few hundred million to around ten billion, built to run cheaply and quickly on laptops, phones or a single GPU while handling focused tasks well.
- System promptSystem message, Developer message, System instructionsA system prompt is the set of instructions an application gives a language model before the user's messages, defining its role, rules, tone, tools and output format for the whole conversation, usually hidden from the end user.
T
- TemperatureSampling temperatureTemperature is a setting that controls how random a language model's output is: a low temperature makes it pick the most likely next tokens for focused, repeatable answers, and a higher temperature spreads the choice for more varied, creative text.
- Tokentokens, tokenization, tokenisationA token is the unit of text a language model reads and writes, typically a word, part of a word or a punctuation mark; models count context, speed and price in tokens, and one token averages about four characters of English.
- Transformertransformer architecture, transformer modelA transformer is the neural network architecture behind most modern AI models; it uses a mechanism called self-attention to weigh how every token in an input relates to every other token, and it processes whole sequences in parallel.