AI terms, made useful.

Agent Skills

SKILL.md

A format for packaging task instructions and optional supporting files so compatible agents can discover and use them. A skill describes a way of working; it is not an account or a subscription.

AI agent

A system that uses a model to choose actions towards a goal, often by calling tools and checking results over several steps. Products use the term loosely; inspect the actual behaviour.

Application programming interface

API

A defined way for software systems to exchange requests and responses. An AI API lets an application call a service; it is not the same thing as a consumer chatbot subscription.

Software designed to carry out tasks associated with human intelligence, such as recognising speech, making predictions or generating text. AI is an umbrella term, not one particular product or level of ability.

Images & audio

Automatic speech recognition

ASR · speech to text

Converting speech into text. Some systems also attempt speaker separation or timestamps, but those features have their own limits.

Benchmark

A standardised test or dataset used to compare systems. It measures performance on its chosen tasks under particular conditions, not every job a customer might need.

Chunking

Splitting source material into pieces for processing or retrieval. The size, boundaries and overlap of those pieces affect what an AI search tool can find and explain.

Context window

The amount of token-based context a model can handle in a request. It includes the material the system supplies, not just the sentence you type. Output capacity and product limits can be separate.

Deep learning

Machine learning using neural networks with multiple layers. Those layers learn representations of the input, such as features useful for recognising an image or predicting the next token.

Distillation

Training a model using outputs or signals from another model, often to transfer useful behaviour into a smaller or more efficient system. The result is not a lossless copy.

Embedding

A numerical representation of an item, such as a passage or image, designed so useful similarities can be measured. An embedding is not a human-readable summary.

Evaluation

eval · evals

A deliberate check of how well a system performs against defined tasks and criteria. Good evaluation covers failure cases and the complete workflow, not only polished example answers.

Few-shot prompting

few-shot

Giving a model a small set of examples in the prompt to demonstrate the pattern you want. It usually changes the context for that request, not the model’s underlying weights.

Fine-tuning

Further training a pre-trained model on selected data to adapt its behaviour or task performance. This changes model parameters, unlike simply adding examples to a prompt.

Foundation model

A model trained broadly enough to be adapted to different downstream tasks. It is a starting point for applications, not a complete finished service for every job.

Generative AI

GenAI

AI that produces new text, images, sound, video or other outputs from learned patterns and supplied input. New output is not automatically correct, unique or cleared for every commercial use.

Grounding

Connecting an answer to relevant external information or supplied evidence rather than relying only on learned patterns. The word is used differently across products, so ask what the connection actually does.

Guardrails

Controls intended to keep an AI system within acceptable boundaries. They can include input checks, output validation, restricted tools, approval steps and operational limits.

Hallucination

A plausible-looking output that is false, unsupported or not grounded in the supplied material. It can include invented sources, quotations, dates or explanations.

Human in the loop

HITL

A workflow in which a person reviews or participates at defined points. It is meaningful only when that person has enough information, time and authority to catch or stop problems.

Inference

Running a trained model on an input to produce a prediction or output. Inference is the operational use of the model, not the process of teaching its weights.

Large language model

LLM

A language model trained at large scale to predict and generate sequences of tokens. A chatbot may use an LLM, but the model and the complete application are different things.

Latency

The delay between a request and a response or milestone. In AI products, time to first token and time to the complete usable answer measure different things.

Low-rank adaptation

LoRA

A fine-tuning technique that learns small additional weight updates while keeping the original model weights fixed. It changes learned behaviour; it is not simply a longer prompt.

Machine learning

A way of building systems that learn patterns from data instead of relying only on rules written explicitly by a programmer. Learning usually happens during training; it does not mean every conversation updates the model.

Mixture of experts

MoE

A model architecture with multiple expert components and a routing mechanism that selects which components process an input. In sparse versions, only a subset is active for a particular input.

Model card

Documentation describing a model, its intended use, evaluation and known limitations. Read it as a starting point for understanding a model, not as an independent certificate of safety or quality.

Model Context Protocol

MCP

An open protocol for connecting AI applications to external capabilities such as tools, resources and prompts. It standardises parts of the connection, not the trustworthiness of everything connected.

Multimodal

Able to work with more than one type of input or output, such as text, images and audio. Support is specific: reading a picture does not necessarily mean a model can generate one.

Neural network

A model built from connected mathematical units whose adjustable weights transform inputs into outputs. The name borrows an analogy from brains; a software network is not a biological brain.

Open weights

Model weights made available under stated terms. Availability of weights does not by itself mean the training data, training code or all commercial rights are open.

Images & audio

Optical character recognition

OCR

Turning visible text in an image or scan into machine-readable text. It is a recognition step, not an assurance that the document’s meaning or figures have been understood correctly.

Overfitting

Learning training examples so specifically that performance does not generalise well to new data. An impressive training score can conceal weak real-world behaviour.

Parameters

weights

Learned numerical values inside a model that shape how inputs become outputs. Parameter count describes one aspect of model size, not a direct score for usefulness.

Prompt

The input that tells a generative system what to do. In a product, the full prompt may combine your message with instructions, retrieved material and previous conversation supplied by the application.

Prompt injection

An attempt to make an AI system follow unwanted instructions supplied through input or material it reads. Indirect injection can sit inside a web page, document or tool response.

Quantization

Representing model weights or other values with lower numerical precision. It can reduce memory requirements and sometimes improve speed, depending on hardware and implementation.

Reranking

A second pass that reorders candidate search results using a more detailed relevance assessment. It operates after an initial retrieval step.

Retrieval-augmented generation

RAG

A workflow that retrieves relevant material and supplies it to a generative model when answering. The retrieval system and the model have separate jobs, and either can fail.

Semantic search

Search that aims to match meaning or intent, rather than only exact words. It often uses embeddings, and can be combined with keyword search and filters.

Structured output

JSON output

Output constrained to a defined shape, such as named fields in a JSON object. A valid shape makes software integration easier but does not establish factual correctness.

System prompt

Application-level instructions intended to shape a model’s behaviour, role or boundaries. The exact instruction hierarchy and names depend on the provider.

Temperature

A sampling setting that changes how strongly token choices favour higher-probability options. Where supported, lower values tend to make outputs less varied and higher values more varied.

Images & audio

Text to speech

TTS

Generating spoken audio from written text. Voice style, pronunciation controls and permitted uses vary between products and licences.

Throughput

The amount of work a system completes in a period, such as requests or tokens per second. It describes capacity and is not the same as the delay experienced by one user.

Token

A unit a model uses to process text or other supported input. In text, a token can be a word, part of a word or punctuation. It is not a fixed number of characters.

Tool use

function calling

A model proposing calls to functions or services that an application makes available. The surrounding application normally executes the call and returns the result.

Training

Adjusting a model’s parameters using data and an optimisation process. It is different from running the trained model to get an answer, which is inference.

Vector database

A database or database capability for storing vectors and finding similar ones. In AI applications those vectors are often embeddings, with links back to original documents and metadata.

zero-shot · direct prompting

Zero-shot prompting

Asking a model to complete a task without supplying a worked example of the desired response in that prompt. You can still provide detailed instructions and source material.