AI glossary

Agents Agent memory Ways for an AI system to keep information across conversations or long tasks beyond a single context window. Agents Agentic workflow A process where one or more AI models carry out multi-step work with some autonomy, as opposed to a single question and answer. Agents AI agent An AI system that pursues a goal over several steps, deciding which actions to take, using tools and checking its own results. Basics AI assistant An AI system that helps a user with tasks through conversation, such as writing, searching, analysing or coding. Safety AI regulation Laws governing how AI may be built and used, such as the EU AI Act, which sets obligations by risk level. Safety AI safety Work to prevent AI systems from causing harm, through misuse, accidents or behaviour their developers did not intend. Training Alignment Making AI systems pursue the goals their developers and users intend, and behave helpfully, honestly and safely. Basics API (application programming interface) A way for software to call a service programmatically; AI providers sell model access through APIs billed per use. Inference API key A secret string that identifies your account when calling a provider's API; whoever holds it can spend your credit. Evaluation Arena Elo A rating built from many people comparing two anonymous model answers side by side and picking the better one. Basics Artificial intelligence (AI) The field of building computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognising images or making decisions. Architecture Attention mechanism The part of a transformer that lets each token weigh how relevant every other token is when computing its meaning. Architecture Autoregressive generation Producing output one token at a time, each new token predicted from all the ones before it.

Training Backpropagation The algorithm that computes how much each weight contributed to an error, so gradient descent knows how to adjust it. Pricing Batch API A way to submit many requests to be processed within hours instead of instantly, usually at about half price. Basics Benchmark A standard test used to measure and compare model abilities, such as coding, maths or knowledge. Safety Bias Systematic unfairness in a model's outputs, often inherited from patterns in its training data. Pricing Blended price A single comparable price combining input and output rates at an assumed ratio; this site uses 3 input tokens per output token.

Pricing Cached input Input tokens served from the prompt cache, billed at a reduced rate. Prompting Chain of thought Having a model work through intermediate reasoning steps before giving its answer, which improves accuracy on complex problems. Basics Chatbot A program that converses with people in natural language, today usually powered by a large language model. Retrieval Chunking Splitting documents into smaller passages before embedding them, so search returns focused, relevant pieces. Retrieval Citations References a model attaches to its claims, pointing to the sources it used. Basics Closed (proprietary) model A model whose weights are not published and which can only be used through the provider's API or apps. Agents Code execution A tool that lets a model write and run code in a sandbox, for calculations, data analysis or charts. Agents Coding agent An agent that writes, edits, runs and tests code across a project, often from a terminal or editor. Basics Completion (output) The text a model generates in response to a prompt. Output tokens usually cost several times more than input tokens. Training Compute The processing power used to train or run AI models, usually measured in GPU hours or floating-point operations. Agents Computer use An agent capability where a model operates a computer like a person: reading the screen, moving the mouse and typing. Training Constitutional AI Anthropic's method of training models to follow a written set of principles, using AI feedback guided by those principles. Safety Content moderation Detecting and filtering harmful content such as hate, violence or sexual material, in text or images. Prompting Context engineering Deciding what information goes into a model's context window, and in what form, so it has what it needs and nothing that distracts it. Basics Context window The maximum number of tokens a model can consider at once, counting both your input and its output. Agents Copilot An AI assistant embedded in a tool you already use, suggesting and drafting as you work.

Evaluation Data contamination When benchmark questions or answers appear in a model's training data, inflating its score. Inference Data residency Keeping data processing within a specific region, such as the EU or US, often required by law or contract. Basics Dataset A collection of examples used to train or evaluate a model. Basics Deep learning Machine learning using neural networks with many layers, which learn increasingly abstract features from raw data. Agents Deep research An agent mode that searches many sources over several minutes and writes a cited report on a question. Safety Deepfake Realistic synthetic media, usually video or audio, that shows someone saying or doing something they did not. Architecture Diffusion model A generative model that creates images, audio or video by starting from random noise and removing it step by step. Training Distillation Training a smaller 'student' model to imitate a larger 'teacher' model, keeping much of its ability at a fraction of the cost. Retrieval Document AI Using AI to read, classify and extract data from documents such as invoices, contracts and forms.

Inference Effort (reasoning effort) A setting on reasoning models that controls how much they think before answering, trading cost and speed for quality. Architecture Embedding A list of numbers (a vector) representing the meaning of text, an image or audio, so similar things have nearby vectors. Evaluation Evaluation (evals) Systematically testing how well a model or AI application performs on the tasks you care about. Safety Explainability How well people can understand why an AI system produced a given output.

Prompting Few-shot prompting Including a few examples of the input and desired output in the prompt so the model copies the pattern. Training Fine-tuning Further training of an existing model on a smaller, specific dataset to specialise it for a task, style or domain. Basics Foundation model A large model trained on broad data that can be adapted to many different tasks. Basics Frontier model One of the most capable models available at a given time, at the edge of what AI can currently do. Agents Function calling The API mechanism through which a model requests that your application run a named function with specific arguments.

Basics Generative AI AI that creates new content, such as text, images, audio, video or code, rather than only classifying or predicting. Training GPU A graphics processing unit; the parallel processor that trains and runs most AI models. Training Gradient descent The optimisation method that nudges a model's parameters, step by step, in the direction that most reduces its loss. Retrieval Grounding Basing a model's answer on supplied sources or live data rather than its memory alone. Safety Guardrails Checks and rules around an AI system that block unsafe inputs or outputs and keep it within its intended scope.

Basics Hallucination When a model states something false or invented with apparent confidence. Retrieval Hybrid search Combining keyword search and semantic search, so both exact terms (names, codes) and meaning are matched.

Multimodal Image editing (AI) Changing an existing image with instructions: removing objects, changing backgrounds, restyling or combining images. Basics Inference Running a trained model to produce an output; what you pay for when you call an AI API. Pricing Input tokens The tokens you send to a model: instructions, conversation history, documents and images. Billed at the input rate. Training Instruction tuning Fine-tuning a pre-trained model on examples of instructions and good responses, so it follows requests instead of just continuing text.

Safety Jailbreak A prompt crafted to get a model to ignore its safety rules and produce content it would normally refuse. Prompting JSON mode An API option that makes a model output valid JSON.

Retrieval Knowledge base A curated collection of an organisation's documents that an AI assistant can search to answer questions. Basics Knowledge cutoff The date after which a model's training data ends; it knows little or nothing about later events unless given that information.

Basics Large language model (LLM) A neural network trained on very large amounts of text to predict the next token, which lets it write, answer questions, summarise, translate and code. Basics Latency How long a model takes to respond; often measured as time to first token. Evaluation Leaderboard A public ranking of models on one or more benchmarks. Evaluation LLM as a judge Using a language model to grade other models' outputs against criteria, to evaluate at scale. Inference Long-context pricing Higher per-token rates some providers charge when a single request exceeds a size threshold. Training LoRA (low-rank adaptation) A cheap fine-tuning technique that trains a small set of extra weights instead of updating the whole model. Training Loss function The formula that measures how wrong a model's predictions are during training; training tries to minimise it.

Basics Machine learning A branch of AI in which systems learn to perform a task from examples and data rather than from explicitly written rules. Inference Max tokens The limit you set on how many tokens a model may generate in one response. Architecture Mixture of experts (MoE) An architecture where only a subset of the model's 'expert' sub-networks is used for each token, giving a large model the running cost of a smaller one. Agents Model Context Protocol (MCP) An open standard for connecting AI applications to external tools and data sources through a common interface. Pricing Model deprecation A provider's announcement that a model will stop being available on a set retirement date, with a recommended replacement. Pricing Model lifecycle The stages a model goes through: preview, generally available, legacy, deprecated and retired. Training Model size How large a model is, usually stated in parameters (for example 8B or 120B), which roughly predicts its capability and running cost. Agents Multi-agent system Several AI agents working together, often with different roles, such as a planner, researchers and a reviewer. Basics Multimodal Able to handle more than one kind of data, such as text, images, audio and video, in its input or output. Multimodal Music generation Creating music, songs or sound effects from a text prompt.

Basics Neural network A computing system made of layers of connected units whose connection strengths are learned from data.

Retrieval OCR (optical character recognition) Turning images of text, such as scans, photos or PDFs, into machine-readable text. Inference On-device (edge) AI Running models directly on phones, laptops or other devices rather than in the cloud. Basics Open-weight model A model whose trained weights are published so anyone can download and run it, under the licence's terms. Agents Orchestration Coordinating models, tools and steps in an AI application: what runs when, in what order, and what happens on failure. Pricing Output tokens The tokens a model generates, including any reasoning it produces. Usually priced several times higher than input. Training Overfitting When a model memorises its training examples instead of learning general patterns, so it performs worse on new data.

Basics Parameter A number inside a model that is learned during training; a model's size is usually counted in parameters. Training Pre-training The first, largest stage of training, where a model learns language and world knowledge by predicting text across a huge dataset. Pricing Price per million tokens The standard unit for pricing language-model usage, quoted separately for input and output. Basics Prompt The input you give a model: instructions, questions, examples and any documents it should use. Pricing Prompt caching Reusing the processed start of a prompt across requests, so repeated content such as a long system prompt or document costs much less. Prompting Prompt engineering The practice of writing and refining prompts to get reliable, high-quality results from a model. Safety Prompt injection An attack where instructions hidden in content the model reads, such as a web page, email or document, try to hijack its behaviour.

Training Quantization Storing a model's weights with fewer bits (for example 4 or 8 instead of 16) to cut memory use and speed it up, with a small loss in quality.

Inference Rate limit The cap a provider sets on how many requests or tokens you can use per minute or per day. Multimodal Realtime voice model A model that listens and speaks in a live conversation with low latency, handling interruptions. Architecture Reasoning model A model trained to think step by step before answering, improving results on maths, coding and complex problems at the cost of time and tokens. Safety Red teaming Deliberately attacking an AI system to find its failures and harmful behaviours before release. Training Reinforcement learning Training by trial and reward: the model tries actions or answers and is updated toward those that score well. Retrieval Reranking A second pass that re-orders search results with a more precise model, so the best passages reach the language model first. Safety Responsible AI Developing and using AI in ways that are fair, transparent, accountable, private and safe. Retrieval Retrieval-augmented generation (RAG) Answering questions by first retrieving relevant passages from your documents and then giving them to the model as context. Training RLHF (reinforcement learning from human feedback) A training method where people rank model responses and the model is optimised to produce the kind of answers they prefer.

Inference Sampling How a model picks each next token from its predicted probabilities, rather than always taking the single most likely one. Training Scaling laws Empirical findings that model performance improves predictably as you increase parameters, data and compute. Inference Self-hosting Running an open-weight model on your own servers or cloud instances instead of calling a provider's API. Retrieval Semantic search Search by meaning rather than exact words, using embeddings, so 'car price' can match 'cost of a vehicle'. Basics Small language model (SLM) A compact language model, often a few billion parameters, that is cheap and fast enough to run on a laptop, phone or modest server. Multimodal Speaker diarisation Identifying who spoke when in a recording, so a transcript is split by speaker. Multimodal Speech-to-text (transcription) Converting spoken audio into written text. Inference Streaming Sending a model's output to the user token by token as it is generated, instead of waiting for the full answer. Prompting Structured output Making a model return data in a fixed format, such as JSON matching a schema, so software can read it reliably. Agents Subagent An agent started by another agent to handle one part of a task, returning only its result. Pricing Subscription vs API pricing Consumer plans charge a flat monthly fee for an app; API access charges per use for developers building their own products. Pricing Successor model The newer model a provider recommends moving to when an older one is deprecated. Evaluation SWE-bench A benchmark that measures whether a model can fix real GitHub issues in open-source projects, verified by the project's own tests. Training Synthetic data Training data generated by AI models rather than collected from people or the world. Safety System card (model card) A document a provider publishes with a model, describing its capabilities, evaluations, limitations and safety testing. Prompting System prompt Instructions given to a model before the conversation that set its role, rules, tone and constraints.

Inference Temperature A setting that controls how random a model's output is: low values give focused, repeatable answers, high values more varied ones. Multimodal Text-to-image Generating an image from a written description. Multimodal Text-to-speech (TTS) Turning written text into natural-sounding speech. Multimodal Text-to-video Generating video clips from a description or a starting image. Inference Thinking tokens Tokens a reasoning model generates while working through a problem before its final answer; billed as output. Basics Throughput How fast a model produces output, usually in tokens per second. Inference Time to first token (TTFT) How long it takes from sending a request until the first token of the response arrives. Basics Token The unit of text a language model reads and writes: a word, part of a word or a punctuation mark. Usage is billed per token. Basics Tokenizer The component that splits text into tokens before a model processes it, and turns tokens back into text. Inference Tokens per second The rate at which a model generates output, a common measure of speed. Agents Tool use A model's ability to call external tools, such as search, a calculator, a database or an API, and use the results in its answer. Inference Top-p (nucleus sampling) A sampling setting that limits the model to the smallest set of likely tokens whose probabilities add up to p. Training TPU Tensor processing unit; Google's custom AI accelerator, used to train and serve Gemini. Basics Training The process of adjusting a model's parameters on data so it learns to perform a task. Architecture Transformer The neural-network architecture behind nearly all modern language models, built around the attention mechanism.

Retrieval Vector database A database designed to store embeddings and quickly find the ones most similar to a query. Architecture Vision-language model A model that understands images together with text, answering questions about photos, charts, screenshots or documents. Multimodal Voice cloning Creating a synthetic voice that sounds like a specific person from a sample of their speech.

Safety Watermarking Embedding an invisible signal in AI-generated images, audio or text so it can later be identified as synthetic. Agents Web search tool A tool that lets a model search the internet during a conversation to get information newer than its training. Basics Weights The learned numerical values of a neural network's connections; 'downloading a model' means downloading its weights.

Prompting Zero-shot Asking a model to do a task with instructions only, without examples.