AISATAI Glossary
316 plain-English definitions across 10 core domains, from transformer internals, RAG and the Model Context Protocol (MCP) to evaluation harnesses, UK data protection and the EU AI Act. Plus 30 in-depth practitioner guides on the concepts that come up most in real projects.
Assistant General Counsel, Product, Privacy & AI at Sprinklr
This AISAT AI Glossary was inspired by Lucy, who first started sharing AI terms on LinkedIn and gave many people an easy, accessible way to learn.
Listen to the Glossary · Read aloud in British English (Aoede)
Play the quick overview, or press Listen on any term below.
Showing 316 of 316 terms across 26 letters
A/B Testing is The online experimentation method of splitting live production traffic between two prompt variants, model versions or retrieval pipelines to measure statistically significant differences in task completion, cost and user satisfaction.
A2A (Agent2Agent Protocol) is An open interoperability standard launched by Google in April 2025 and donated to the Linux Foundation in June 2025 that enables AI agents across different vendors and frameworks to discover capabilities, delegate tasks and exchange artefacts.
An Ablation Study is A controlled engineering experiment in which a single architectural layer, retrieval step or prompt component is systematically removed to measure its exact causal contribution to overall system accuracy and latency.
Accountability in AI governance is the allocation of explicit legal and operational responsibility for an AI system's outcomes, risk controls and regulatory compliance to named human executives and system owners.
Accuracy is An evaluation metric calculated as the ratio of correct predictions to total predictions made across a dataset, suitable for balanced classification benchmarks but misleading on skewed classes such as rare fraud events.
An Activation Function is A mathematical operation inside each neural network node, such as ReLU, GELU or SwiGLU, that determines how strongly a signal propagates forward and enables the network to learn complex non-linear relationships.
An Adversarial Attack is A security exploit in which an attacker crafts subtly perturbed images, audio waves or text sequences designed to cause a machine learning model to misclassify inputs or bypass safety controls.
An Agent Card is A machine-readable JSON metadata manifest published by an A2A-compatible agent, typically at a well-known HTTPS endpoint, declaring its identity, supported skills, input-output modalities and authentication schemes.
Agent Skills are Modular, versioned packages of Markdown instructions, helper scripts and domain reference files that an AI agent discovers and loads into its context window on demand to perform specialised workflows.
Agentic AI refers to Artificial intelligence systems engineered to pursue multi-step objectives with bounded autonomy by decomposing goals, invoking external tools, inspecting intermediate results and adapting plans under defined governance controls.
The Agentic AI Foundation (AAIF) is A Linux Foundation directed fund launched in December 2025 and co-founded by Anthropic, Block and OpenAI to provide neutral open-source governance for agent standards including MCP, goose and AGENTS.md.
An Agentic Loop is The iterative runtime execution cycle in which an AI agent observes current context, reasons about the next action, invokes an external tool or API, inspects the returned observation and repeats until its goal or stop condition is met.
Agentic RAG is A retrieval-augmented generation architecture in which an AI agent iteratively plans search queries, grades whether fetched passages answer the question, rewrites queries when evidence is missing and synthesises verified multi-source citations.
An Agentic Workflow is A multi-step operational or engineering process in which a language model dynamically plans steps, calls external tools, critiques its own output and branches based on intermediate results within an orchestrated pipeline.
AGENTS.md is An open convention stewarded under the Agentic AI Foundation for placing a plain-Markdown instruction file at the root of a software repository to give AI coding agents build commands, test rules and architectural conventions.
AI Adoption is The measurable transition of an organisation from isolated individual chatbot experiments to governed, repeatable artificial intelligence workflows embedded into daily engineering, customer service and back-office business operations.
An AI Agent is A software system powered by a large language model that perceives environment context, maintains state, formulates multi-step plans and executes actions across external tools, APIs and databases to complete a delegated goal.
An AI Assistant is A user-directed conversational software tool that drafts content, answers questions and summarises documents in direct response to human prompts, operating reactively rather than pursuing autonomous background goals.
An AI Champion is A domain practitioner embedded inside an operational business unit who tests approved AI tools against real team workflows, documents reusable prompt patterns and coaches colleagues during rollout.
AI Enablement is The organisational programme of combining enterprise-licensed AI tools, role-specific practitioner training, shared prompt libraries and clear usage policies so employees turn model access into measurable output gains.
AI Ethics is The normative framework of moral principles and practical design standards guiding organisations to build and deploy artificial intelligence systems that respect human autonomy, fairness, privacy and societal wellbeing.
An AI Gateway is A centralised infrastructure proxy positioned between internal applications and model providers that enforces unified authentication, model routing, automatic fallbacks, rate limiting, PII redaction and per-team spend attribution.
AI Governance is The operational framework of internal policies, model inventories, risk assessments, approval gates, human oversight rules and audit trails that an organisation uses to deploy artificial intelligence safely, ethically and legally.
AI Literacy is The documented competence of staff and contractors to understand how AI systems function, interpret their outputs critically and recognise risks; legally mandated for providers and deployers under Article 4 of the EU AI Act since 2 February 2025.
An AI Maturity Model is A structured diagnostic benchmark used by executive leadership teams to assess an organisation's readiness across data foundations, engineering tooling, evaluation rigour, governance and workforce adoption.
AI Overviews are Google Search's generative AI summary panels rendered above traditional organic web results that synthesise direct answers and clickable source citations from multiple indexed web pages in real time.
An AI Risk Assessment is A structured pre-deployment audit that identifies an AI use case's potential failure modes across data privacy, algorithmic bias, hallucination, cybersecurity and regulatory duties, alongside required mitigating controls.
The AI Security Institute (UK AISI) is The UK government directorate within the Department for Science, Innovation and Technology (DSIT), renamed from the AI Safety Institute in February 2025, that conducts empirical evaluations of frontier AI capabilities and security risks.
An AI Strategy is An executive commercial and technical roadmap that defines which high-value business problems an organisation will solve with AI, alongside build-versus-buy rules, data readiness investments, governance and target ROI.
AI Washing is The deceptive marketing practice of overstating the presence or sophistication of artificial intelligence in a vendor product or corporate strategy, such as rebranding deterministic if-then rules as autonomous AI agents.
AI-First is An operating and product design strategy in which teams re-architect processes, data schemas and user interfaces around language models and autonomous agents from first principles rather than bolting chat widgets onto legacy software.
An Algorithm is An explicit, step-by-step set of mathematical or logical instructions that a computer executes to solve a defined problem, process data or make a calculation, ranging from simple sorting rules to complex machine learning training routines.
Algorithmic Bias is Systematic, repeatable disparity in an AI system's predictions, scores or allocations that unlawfully or unfairly disadvantages individuals on protected characteristics such as sex, race, age or disability.
Alignment is The post-training and technical governance discipline of steering a machine learning model so its runtime behaviour consistently follows human instructions, factual truthfulness standards, organisational policies and ethical safety boundaries.
Answer Engine Optimisation (AEO) is The technical content discipline of structuring web pages with direct declarative definitions, question-led headings and Schema.org JSON-LD markup so search answer boxes and AI assistants extract and quote them verbatim.
Anthropomorphism is The cognitive bias of attributing human emotions, consciousness, beliefs or personal intent to an artificial intelligence system, such as assuming a large language model genuinely feels remorse when it generates an apology.
An API (Application Programming Interface) is A formal programmatic contract over HTTPS that enables software applications to send prompts, context and tool schemas to a hosted AI model and receive structured completions.
An API Key is A secret cryptographic bearer token passed in request headers to authenticate an application against an AI provider endpoint, enforce rate limits and attribute token consumption for billing.
Artificial General Intelligence (AGI) is A hypothetical artificial intelligence system capable of matching or exceeding human cognitive performance across almost any intellectual, scientific or operational task, rather than being restricted to a single narrow domain.
Artificial Intelligence (AI) is The discipline of building computer systems that perform tasks typically requiring human intelligence, including understanding natural language, recognising visual patterns, reasoning over complex data and generating text, code or media.
Artificial Superintelligence (ASI) is A theoretical stage of artificial intelligence that far surpasses the brightest human minds across every field of endeavour, including scientific discovery, strategic planning, engineering design and social reasoning.
An Attention Mechanism is A neural network component that dynamically calculates weightings across all tokens in an input sequence, allowing a model to focus on the most relevant context regardless of distance.
An Audit Trail is A tamper-evident, timestamped operational log recording the exact input prompt, retrieved context documents, model version, tool invocations, output and human reviewer identity for every high-stakes AI transaction.
An Autonomous System is A software agent, robotic platform or industrial control system capable of sensing its environment, reasoning over objectives and executing operational actions independently without step-by-step human intervention.
The Autonomy Spectrum is The architectural design scale for enterprise AI systems, ranging from human-driven copilot suggestions and deterministic workflow chains through human-in-the-loop approvals to fully autonomous background agent execution.
An Autoregressive Model is A sequence generation model that produces output one token at a time, conditioning the probability distribution of each next token on the prompt and all tokens generated so far.
Backpropagation is The foundational training algorithm for neural networks that computes the gradient of the loss function by propagating prediction errors backwards from the output layer to update every internal weight.
A Base Model is A foundation language model immediately after large-scale self-supervised pre-training and prior to instruction tuning or preference alignment, capable of continuing text patterns but not reliably following user instructions.
Batch (Batch Size) refers to The fixed number of training examples processed together in a single forward and backward pass through a neural network before the optimiser updates the model's parameter weights.
Batch Processing is An asynchronous inference deployment mode in which large volumes of non-interactive requests, such as overnight document classification or eval runs, are queued for completion within 24 hours at roughly half the token cost.
A Benchmark is A standardised, reproducible test suite, such as SWE-bench Verified, GPQA Diamond or MMLU-Pro, used to compare the coding, scientific reasoning or domain accuracy of competing foundation models.
Benchmark Contamination is A data leakage failure in which evaluation questions and ground-truth answers from a public test suite are inadvertently included in a model's pre-training corpus, artificially inflating its reported scores.
Bias (Data Bias) is A systematic skew, gap or historical prejudice present in a training dataset that causes a machine learning model to produce inaccurate, unrepresentative or unfair predictions for specific groups or edge cases.
Big Data refers to High-volume, high-velocity and high-variety datasets that exceed the capacity of traditional relational databases, serving as the raw foundational corpus for training modern deep learning and analytical models.
A Black Box is A complex machine learning or deep neural network model whose observable inputs and outputs can be inspected, but whose internal decision-making pathway across billions of parameters cannot be traced step by step.
BM25 is A probabilistic lexical ranking algorithm used in search engines to score document relevance based on exact term frequency, inverse document frequency and document length normalisation, routinely paired with vector search in hybrid RAG pipelines.
Bring Your Own Key (BYOK) is A software deployment model in which an enterprise connects its own direct OpenAI, Anthropic, Azure or AWS Bedrock API key to a third-party SaaS tool, retaining control over billing and zero-data-retention terms.
Browser Use (Computer Use) is An agent capability in which a vision-language model inspects live desktop screenshots or browser DOM trees and executes mouse clicks, scrolling and keyboard input to operate legacy software without APIs.
Build vs Buy is The strategic architectural and commercial trade-off analysis between engineering a custom AI workflow on foundation model APIs in-house versus licensing an off-the-shelf vertical SaaS software application.
Catastrophic Forgetting is The tendency of a neural network to abruptly lose previously learned general reasoning and language capabilities when its weights are fine-tuned too aggressively on a narrow downstream dataset.
A Centre of Excellence (CoE) is A cross-functional internal engineering, data and compliance team that establishes AI reference architectures, manages model gateways, runs shared evaluation harnesses and helps business units ship production use cases.
Chain-of-Thought (CoT) Prompting is A technique that instructs or demonstrates to a language model how to articulate intermediate reasoning steps sequentially before stating a final conclusion, significantly improving accuracy on logic and mathematical tasks.
Change Management in AI transformation is the structured leadership work of redesigning job workflows, performance incentives, output verification habits and internal communications so employees trust and adopt new AI systems.
A Chatbot is A conversational software interface that exchanges natural-language text or voice messages with human users to answer questions, triage support tickets or assist with tasks, today typically powered by a large language model.
Chunking is The preprocessing step in a retrieval pipeline that splits long documents into smaller, semantically coherent passages with controlled token overlap before embedding and indexing them inside a vector database.
Citation (Source Attribution) is The mechanism by which a retrieval-augmented AI system links specific claims in its generated answer back to the exact source document, page number or URL from which the fact was drawn.
Classification is A supervised machine learning task in which a trained model assigns each incoming record, document or image to one or more predefined discrete categories, such as flagging a payment as legitimate or fraudulent.
A Closed Model (Proprietary Model) is A commercial artificial intelligence model whose trained weights, training dataset and architecture details are kept private by its provider and accessed exclusively through hosted applications or paid APIs.
Cloud AI refers to Running model training, fine-tuning, vector indexing and production inference workloads on managed GPU clusters hosted by hyperscalers or frontier model laboratories rather than on local on-premises hardware.
Clustering is An unsupervised machine learning technique that automatically groups data points with shared statistical attributes into distinct segments without relying on predefined category labels, widely used for customer segmentation and anomaly discovery.
A Coding Agent is An autonomous software engineering system, such as Claude Code, Cursor Agent or GitHub Copilot Coding Agent, that navigates codebases, edits multiple files, executes test suites and resolves bugs in a terminal loop.
A Completion is The output sequence of natural-language text, source code or structured JSON tokens generated by a large language model in response to an input prompt and system context.
Compute refers to The aggregate hardware processing capacity, quantified in floating-point operations per second (FLOPs) or accelerator GPU hours, consumed to pre-train, fine-tune, evaluate or serve inference on artificial intelligence models.
Computer Vision (CV) is The field of artificial intelligence that enables software systems to inspect, interpret and extract structured information from digital images and video streams, including defect inspection, optical character recognition and object tracking.
A Confusion Matrix is A contingency table used in classification evaluation that breaks model predictions down into true positives, false positives, true negatives and false negatives across every target category.
Constitutional AI is A post-training alignment method introduced by Anthropic in which a language model critiques and revises its own candidate responses against an explicit charter of human-authored safety and helpfulness principles.
Content Provenance (C2PA) is An open technical standard developed by the Coalition for Content Provenance and Authenticity that binds cryptographically signed metadata manifests (Content Credentials) to digital media to verify its origin and AI editing history.
Context Engineering is The engineering discipline of dynamically selecting, structuring, compressing and ordering the system instructions, retrieved knowledge, tool definitions, memory and examples placed inside a language model's context window at inference time.
Context Rot is The progressive degradation in a language model's instruction following, fact recall and reasoning precision as its context window accumulates long, repetitive, outdated or contradictory conversation history and tool outputs.
A Context Window is The maximum number of tokens a language model can process in working memory within a single request, encompassing system instructions, retrieved passages, tool schemas, conversation history and the generated reply.
Continued Pre-Training is The process of training an existing foundation model further on billions of tokens of unlabelled domain-specific text, such as UK statute law, biomedical literature or proprietary codebases, before instruction tuning.
Conversational AI is The category of enterprise software systems, spanning customer support chatbots, internal knowledge assistants and real-time voice agents, that enable natural written or spoken dialogue with business systems.
A Convolutional Neural Network (CNN) is A deep learning architecture that uses sliding mathematical filters to detect spatial hierarchies of edges, textures and shapes across grid-structured data such as images and medical scans.
A Copilot is An interactive AI assistant embedded directly inside a professional software environment, such as a code editor, productivity suite or CRM, that drafts, summarises and suggests actions while the human user retains control.
A Corpus is The structured body of documents, knowledge base articles, transcripts or web pages indexed by a retrieval system for runtime grounding or assembled to pre-train a language model.
Cosine Similarity is A geometric distance metric used in vector search that measures how semantically similar two embedding vectors are by calculating the cosine of the angle between them in multi-dimensional space.
Custom Instructions are Persistent user- or workspace-level profile rules saved inside an AI assistant that automatically prepend role context, British English spelling preferences and formatting constraints to every new conversation.
Data Augmentation is A training technique that expands a dataset by generating transformed copies of existing examples, such as rotating images, adding background audio noise or paraphrasing sentences, so models generalise better to unseen inputs.
Data Labelling (Annotation) is The human or automated process of attaching verified target tags, bounding boxes or ground-truth categories to raw text, audio or images so a supervised machine learning model can learn accurate patterns.
Data Poisoning is An integrity attack in which an adversary injects maliciously crafted or mislabelled records into a model's training corpus or RAG knowledge base to corrupt behaviour or plant hidden backdoor triggers.
Data Privacy in AI engineering is the enforcement of technical and legal safeguards, under UK GDPR, EU GDPR and the UK Data (Use and Access) Act, ensuring personal data is minimised, redacted and never leaked into model training.
Data Residency refers to The contractual, architectural and regulatory guarantees specifying the exact geographical country or sovereign cloud region, such as the UK or European Economic Area, where prompts, embeddings and logs are processed and stored.
A Dataset is A curated collection of structured records or unstructured documents, images and audio samples assembled to train, validate or benchmark the accuracy and safety of a machine learning model.
A Decoder is The generative component of a sequence-to-sequence or transformer model that produces output tokens autoregressively one step at a time, forming the core architecture of modern large language models.
Deep Learning (DL) is A sub-discipline of machine learning that uses artificial neural networks with many stacked hidden layers to automatically learn hierarchical representations directly from vast volumes of text, audio or visual data.
A Deepfake is Synthetic audio, video or photography generated by deep neural networks that convincingly fabricates a real person's likeness or voice saying or doing something they never actually did.
A Deployer, as defined by Article 3 of the EU AI Act, is any natural or legal person, public authority or enterprise that uses an AI system under its own authority in professional operations, distinct from the upstream model provider.
A Diffusion Model is A generative deep learning architecture that synthesises high-resolution images, video or audio by training a neural network to iteratively reverse a gradual Gaussian noise-corruption process step by step.
A Digital Worker is An enterprise AI agent provisioned with role-specific identity permissions, system integrations, standard operating procedures and audit logging to execute a bounded business function alongside human teams.
Direct Preference Optimisation (DPO) is A stable post-training alignment algorithm that updates a language model's weights directly on pairs of preferred and rejected completions using a closed-form classification loss without training a separate reward model.
Distillation (Knowledge Distillation) is A model compression technique in which a compact student model is trained to reproduce the token probability distributions, reasoning traces or structured outputs of a much larger teacher model.
Dropout is A neural network regularisation technique that randomly zeroes out a subset of neuron activations during each training step, preventing co-adaptation and improving how well the model generalises to production data.
Edge AI (On-Device AI) is The deployment of quantised small language, vision or audio models directly on local hardware such as smartphones, laptops or industrial sensors, eliminating network latency and cloud data egress.
An Embedding Model is A specialised neural network trained to convert variable-length text passages, images or audio clips into fixed-dimensional numerical vectors that preserve semantic similarity for search, clustering and retrieval.
Embeddings (Vector Embeddings) are Dense arrays of floating-point numbers generated by an embedding model to represent the semantic meaning of text, code, images or audio so conceptually related items sit close together in vector space.
Emergent Behaviour refers to Complex capabilities, such as multi-step arithmetic, code translation or analogical reasoning, that arise in foundation models once they cross a threshold of parameter scale and training compute without explicit programming.
An Encoder is The neural network component that reads an entire input sequence in parallel and compresses it into dense contextual vector representations, powering embedding models, semantic search and classification pipelines.
An Endpoint is The specific HTTPS network URL exposed by a hosted model API, self-hosted inference server or Model Context Protocol service to which client software sends authenticated inference requests.
An Epoch is One complete pass of the entire training dataset forward and backward through a neural network during model training or fine-tuning, often comprising thousands of individual mini-batch updates.
The EU AI Act (Regulation (EU) 2024/1689) is the European Union's risk-based AI law: prohibited practices and AI literacy applied on 2 February 2025 and GPAI rules on 2 August 2025, while under the May 2026 Digital Omnibus agreement standalone high-risk duties apply from 2 December 2027.
Evals (Evaluations) are Automated, repeatable testing pipelines that score an AI application's outputs against a golden dataset using deterministic assertions, reference metrics and LLM-as-a-judge rubrics before any prompt or model change ships.
An Expert System is An early form of symbolic artificial intelligence that emulates human specialist decision-making by applying an inference engine to a hand-coded knowledge base of explicit if-then rules.
Explainability (Explainable AI / XAI) refers to Interpretability methods, such as SHAP values, feature attribution and reasoning traces, that allow engineers, auditors and affected customers to understand why a model produced a specific decision.
The F1 Score is A classification evaluation metric defined as the harmonic mean of precision and recall, providing a balanced single measure of model performance on datasets with severe class imbalance.
Fairness in machine learning is the discipline of measuring and constraining model error rates, false-positive ratios and outcome distributions across demographic cohorts so automated systems do not perpetuate unlawful discrimination.
A Feature is An individual measurable input variable, attribute or data column passed into a machine learning model to make a prediction, such as customer tenure, transaction value or account login frequency.
Feature Engineering is The discipline of selecting, normalising, combining and transforming raw operational data into informative input variables that enable a machine learning algorithm to learn predictive patterns with higher accuracy.
Federated Learning is A decentralised machine learning approach in which multiple client devices or hospitals train a shared global model on local data, exchanging encrypted weight updates rather than centralising sensitive raw records.
Few-Shot Prompting is The practice of embedding two to ten curated input-output demonstration pairs directly inside a prompt so the language model learns the exact output structure, classification boundary or tone in context.
Fine-Tuning is The post-training process of updating a pre-trained model's weights on a curated dataset of task-specific input-output examples to lock in domain vocabulary, strict formatting or lower inference cost.
A Foundation Model is A large neural network pre-trained on broad, web-scale datasets that can be adapted to a wide range of downstream tasks, classified as a general-purpose AI model under the EU AI Act.
A Frontier Model is A general-purpose foundation model at the leading edge of compute scale, reasoning capability and multimodal performance at any given time, developed by major AI research laboratories.
Function Calling (Tool Use) is A model capability in which a language model inspects declared tool schemas and emits a structured JSON invocation naming an external function and arguments for the host runtime to execute.
A General-Purpose AI (GPAI) Model is The legal classification under Chapter V of the EU AI Act for foundation models trained at scale to perform a wide variety of downstream tasks, subject to transparency and copyright obligations since 2 August 2025.
A Generative Adversarial Network (GAN) is A deep learning architecture in which two neural networks, a generator that creates synthetic samples and a discriminator that attempts to spot fakes, compete during training.
Generative AI (GenAI) refers to Deep learning systems capable of synthesising original text, source code, images, audio or video in response to prompts, rather than solely classifying or scoring existing data records.
Generative Engine Optimisation (GEO) is The practice of engineering authoritative primary content, high entity density, third-party citations and machine-readable endpoints such as llms.txt and Markdown URLs so generative answer engines cite and recommend your brand.
Generative Pre-trained Transformer (GPT) is The architectural family of decoder-only transformer models pre-trained via next-token prediction on large text corpora, as well as the brand name of OpenAI's flagship language models.
A Golden Dataset is A versioned, domain-expert-verified benchmark of representative production prompts, edge cases, adversarial traps and ideal ground-truth responses used as the primary reference standard in automated model evaluations.
The GPAI Code of Practice is The voluntary compliance framework published by the European Commission in July 2025 setting out concrete commitments on transparency, copyright and safety for providers of general-purpose AI models under the EU AI Act.
A GPU (Graphics Processing Unit) is A massively parallel processor containing thousands of tensor cores engineered to execute matrix multiplications simultaneously, serving as the primary hardware accelerator for deep learning training and inference.
Gradient Descent is The iterative first-order optimisation algorithm used to train machine learning models by computing the slope of the loss function and stepping parameter weights in the direction that reduces error.
GraphRAG is An advanced retrieval-augmented generation architecture that combines vector similarity search with a structured entity-relationship knowledge graph to answer complex multi-hop and thematic questions spanning entire enterprise document collections.
Ground Truth is The verified, real-world factual answer or human-audited reference label for a given input, against which a machine learning model's predictions or generated completions are scored during training and evaluation.
Grounding is The practice of tethering a language model's generation to authoritative, verifiable runtime evidence, such as retrieved enterprise documents, SQL query results or live web search, to suppress hallucination.
Guardrails are Programmable validation layers placed on the input and output paths of an AI application to detect prompt injections, redact personally identifiable information, verify citations and block policy-violating completions in real time.
Guardrails (Agent Guardrails) are Runtime policy controls wrapped around an autonomous agent that restrict permitted tool calls, enforce token and spend budgets, block indirect prompt injections and gate irreversible actions behind human approval.
Hallucination is A failure mode in which a generative AI model outputs confident, fluent statements that are factually false, fabricated or unsupported by source documents because it predicts statistically plausible tokens rather than verifying truth.
A Handoff is The structured transfer of task control, conversation history and validated state variables from one specialised AI agent to another subagent or to a human operator when an escalation rule triggers.
A Heuristic is A practical, rule-of-thumb shortcut or deterministic approximation used in software engineering and search algorithms to deliver a fast, good-enough decision when computing an exact mathematical optimum is too slow or costly.
A High-Risk AI System is An AI application regulated under Annex I or Annex III of the EU AI Act, including recruitment, creditworthiness, education and critical infrastructure systems, requiring conformity assessments, risk management and human oversight.
Human Evaluation is The quality assurance process in which subject-matter experts or target users blind-review and score model completions for clinical accuracy, legal nuance, tone and helpfulness to calibrate automated evals.
Human-AI Collaboration is The deliberate design of operational business workflows that combine machine speed in synthesis, retrieval and first-draft generation with human domain judgement, stakeholder empathy and final decision accountability.
Human-in-the-Loop (HITL) is An operational governance pattern in which an AI system or agent prepares a recommended decision, communication or tool call, but pauses execution until a qualified human reviewer approves or edits it.
Human-on-the-Loop (HOTL) is A supervision pattern in which an AI system executes high-volume actions autonomously within policy guardrails while human operators monitor real-time telemetry, audit exceptions and retain an immediate kill switch.
Hybrid Search is A retrieval pipeline that executes sparse lexical keyword search alongside dense semantic vector search and merges the candidate rankings using reciprocal rank fusion or a cross-encoder reranker.
A Hyperparameter is An external configuration variable set by engineers before a training run begins, such as learning rate, batch size, weight decay or network depth, as distinct from learned model weights.
A Hyperscaler is A global cloud infrastructure provider, principally Amazon Web Services, Microsoft Azure and Google Cloud Platform, that operates the massive GPU data centres and managed model platforms underpinning enterprise AI.
Image Generation (Text-to-Image) is The synthesis of original digital illustrations, marketing assets, product renders or photorealistic visuals from natural-language text descriptions using latent diffusion transformers or multimodal autoregressive foundation models.
In-Context Learning is The emergent capability of transformer language models to adapt to a novel task at inference time solely from instructions and examples placed in the prompt, without updating any model weights.
Indexing is The data ingestion pipeline that parses, cleans, chunks, embeds and organises source documents inside inverted keyword indexes and approximate nearest-neighbour vector graphs so runtime retrieval queries resolve in milliseconds.
Inference is The runtime operational phase in which a trained machine learning or language model processes live input data through its frozen weights to generate predictions, classifications or token completions.
Input and Output Tokens are The two metered billing units of language model inference: input tokens cover the prompt and retrieved context processed in parallel, while output tokens cover the reply decoded sequentially at higher cost.
An Instruction-Tuned Model is A pre-trained base language model that has undergone post-training on curated instruction-response pairs and human preference data so it reliably follows user commands and formatting constraints.
Intelligent Process Automation (IPA) is The combination of deterministic workflow automation and robotic process automation with large language models and document extraction so enterprise pipelines can handle unstructured inputs and exception routing.
ISO/IEC 42001 is The international certifiable management system standard published by ISO and IEC that specifies requirements for establishing, implementing, auditing and continually improving an organisational Artificial Intelligence Management System (AIMS).
A Jailbreak is An adversarial prompt engineering technique that uses roleplay, cipher encoding or hypothetical framing to trick a language model into bypassing its post-training safety alignment and content filters.
A Knowledge Base is A governed, versioned repository of enterprise policies, technical documentation, product manuals and structured records that an AI assistant or RAG pipeline queries to ground its answers.
A Knowledge Cutoff is The final timestamp of events and publications included in a language model's static pre-training corpus, beyond which the model has no intrinsic awareness unless grounded via live search or RAG.
A Knowledge Graph is A structured semantic network that models real-world domain entities, such as organisations, people, products, contracts and regulations, as graph nodes connected by explicit, typed relationship edges.
A KV Cache (Key-Value Cache) is A GPU memory buffer used during autoregressive transformer inference to store the attention key and value tensors of preceding tokens so they are not recomputed at every generation step.
A Label is The target ground-truth output attached to a training example in supervised machine learning, such as a churn indicator on a customer record or a diagnostic category on a medical scan.
A Large Language Model (LLM) is A deep transformer neural network containing billions of parameters trained on massive text and code corpora to comprehend, summarise, translate, reason over and generate human and programming languages.
Latency is The elapsed wall-clock delay between dispatching an API inference request and receiving the model's response, tracked in production via Time to First Token (TTFT) and total completion time.
Latent Space is The compressed, multi-dimensional mathematical representation learned inside a deep neural network in which concepts, documents or images with similar semantic properties sit close to one another geometrically.
A Layer is A structured functional block of neurons or mathematical operations within a neural network, such as a self-attention or feed-forward block, that transforms incoming tensors and passes them forward.
The Learning Rate is A critical training hyperparameter that controls the step size by which an optimiser adjusts a model's weights in response to the calculated error gradient after each batch.
LLM Visibility (AI Share of Voice) is A commercial search metric that tracks how frequently, accurately and favourably an organisation or product is cited when prospective buyers ask ChatGPT, Claude, Perplexity or Google AI Overviews category questions.
LLM-as-a-Judge is An automated evaluation pattern in which a capable language model grades another AI system's responses against a structured rubric for faithfulness, completeness or tone, calibrated against human annotations.
LLMOps is The operational engineering discipline for production language model applications, encompassing prompt version control, automated evaluation gates, semantic caching, runtime guardrail enforcement, distributed observability tracing and token cost management.
llms.txt is A proposed web convention placed at the root of a domain (/llms.txt, paired with /llms-full.txt and .md URLs) that provides language models and answer engines with a concise, Markdown-formatted index of a site's canonical content.
Long-Term Memory is Persistent external vector or relational storage that enables an AI assistant or autonomous agent to store, index and retrieve user preferences, episodic summaries and domain facts across separate sessions.
LoRA (Low-Rank Adaptation) is A parameter-efficient fine-tuning method that freezes a pre-trained model's weight matrices and injects small, trainable rank-decomposition matrices into transformer layers, cutting GPU memory and training cost by over 90%.
A Loss Function is The mathematical objective formula, such as cross-entropy or mean squared error, that quantifies the discrepancy between a model's predicted output and the ground-truth target during training.
Machine Learning (ML) is The core branch of artificial intelligence in which algorithms learn statistical patterns and predictive rules directly from historical data and experience rather than following hand-written procedural code.
Max Tokens is An inference API parameter that sets a hard upper ceiling on the number of output tokens a language model is permitted to generate in a single completion before stopping.
An MCP Client is The protocol connector embedded inside an AI host application, such as an IDE, desktop assistant or agent orchestrator, that negotiates capabilities and maintains isolated 1:1 sessions with MCP servers.
An MCP Server is A lightweight adapter service that wraps an external system, such as PostgreSQL, GitHub, Slack or an internal API, and exposes its tools, resources and prompt templates over the Model Context Protocol.
Meta-Prompting is The engineering workflow of using a frontier language model alongside automated evaluation feedback to systematically draft, critique, expand and optimise system prompts or templates intended for downstream production models.
Metadata Filtering is The retrieval practice of constraining vector similarity queries using structured document attributes, such as publication date, business department, legal jurisdiction or role-based access permissions, before ranking candidate passages.
Mixture of Experts (MoE) is A sparse neural network architecture comprising multiple specialised feed-forward sub-networks in which a learned gating router activates only a small subset of parameters for each token.
MLOps is The engineering discipline that applies DevOps automation principles to machine learning systems, standardising feature pipelines, reproducible training runs, model registries, CI/CD deployment gates and production data drift monitoring.
A Model is A trained mathematical artefact combining a specific neural or statistical architecture with learned parameter weights that maps incoming data inputs to predictions, classifications or generated content during runtime inference.
A Model Card (System Card) is A standardised transparency document published alongside a trained model that discloses its architecture, intended use cases, evaluation benchmark scores, red-teaming findings and known safety limitations.
Model Collapse is A degenerative training failure in which generative models trained recursively on synthetic data produced by earlier models progressively lose tail-distribution knowledge, lexical diversity and factual accuracy over generations.
Model Context Protocol (MCP) is An open JSON-RPC standard introduced by Anthropic in November 2024 and donated to the Linux Foundation's Agentic AI Foundation in December 2025 that standardises how AI models and agents connect to external data and tools.
Model Drift is The progressive degradation of a deployed model's predictive accuracy caused by shifts in real-world input data distributions (data drift) or changes in the underlying business relationship (concept drift).
Model Routing is An inference architecture that classifies each incoming request by complexity and directs routine tasks to fast, low-cost small language models while escalating complex reasoning to frontier models.
A Multi-Agent System is An orchestration architecture in which multiple specialised AI agents with distinct system prompts, toolsets and context windows collaborate through structured handoffs or a supervisor to solve complex workflows.
Multimodal AI refers to Foundation models engineered to natively ingest, reason across and generate multiple distinct data modalities, including text, still images, spoken audio and video streams, within a unified representation space.
Narrow AI (Weak AI) refers to Artificial intelligence systems engineered and trained to perform a single, bounded operational task, such as credit risk scoring, speech transcription or warehouse route optimisation.
Natural Language Processing (NLP) is The branch of artificial intelligence and computational linguistics concerned with enabling computers to parse, interpret, translate, summarise and generate written or spoken human language across enterprise workflows.
Natural Language Understanding (NLU) is A specialised sub-field of natural language processing focused on extracting semantic meaning, user intent, sentiment and structured named entities from unstructured human text or speech transcripts.
A Neural Network is A computational model composed of interconnected layers of artificial nodes that learn to approximate complex functions by adjusting the numerical weights of connections across training examples.
Neuro-Symbolic AI is A hybrid artificial intelligence architecture that pairs the statistical pattern recognition of deep neural networks with the explicit logical rules, knowledge graphs and formal verification of symbolic AI.
A Neuron (Node) is The fundamental mathematical unit of an artificial neural network, which computes a weighted sum of its incoming signals, adds a learnable bias term and applies a non-linear activation function.
Next-Token Prediction is The self-supervised pre-training objective of autoregressive language models, in which the network learns grammar, world knowledge and reasoning by estimating the probability distribution of the token immediately following a given context.
The NIST AI Risk Management Framework (AI RMF) is A voluntary governance standard published by the US National Institute of Standards and Technology that organises operational AI risk controls across four core functions: Govern, Map, Measure and Manage.
An NPU (Neural Processing Unit) is A dedicated, low-wattage silicon accelerator integrated into modern consumer laptops and smartphones to execute on-device neural network inference without draining battery or GPU resources.
Observability in AI engineering is the capture of structured end-to-end telemetry traces recording prompts, retrieved RAG chunks, tool calls, model versions, latency and token spend so teams can diagnose production regressions.
One-Shot Prompting is The in-context learning technique of supplying a single canonical input-output demonstration inside the prompt alongside task instructions, giving the language model a concrete template for formatting and granularity.
An Open-Weight Model is A trained AI model whose parameter weights are publicly released for download, self-hosting, quantisation and fine-tuning under an open or permissive licence, even when its training corpus remains proprietary.
Orchestration is The software control layer that coordinates model routing, state persistence, context assembly, parallel tool execution, error retries and human approval gates across multi-step AI applications and agent systems.
An Orchestrator Agent (Supervisor) is A central coordinating agent in a hierarchical multi-agent system that decomposes a user objective into sub-tasks, dispatches them to worker subagents and synthesises their verified outputs.
Overfitting is A modelling failure in which a machine learning model memorises the specific noise and coincidental patterns of its training data so closely that its accuracy degrades sharply on unseen production inputs.
Parameter-Efficient Fine-Tuning (PEFT) is A family of adaptation methods, including LoRA and QLoRA, that fine-tune only a tiny fraction of trainable adapter parameters while keeping the base model weights frozen.
Parameters are The internal, learnable numerical variables, primarily connection weights and bias terms, that a neural network adjusts during training and stores to encode its knowledge, often measured in billions for modern language models.
Perplexity is An intrinsic statistical evaluation metric that measures how uncertain an autoregressive language model is when predicting the next token across a held-out text sample, where lower perplexity indicates stronger fit.
Persona (Role Prompting) is The practice of instructing a language model in its system prompt to adopt a specific professional identity, such as a UK commercial solicitor or staff site reliability engineer, to anchor its analytical lens.
Personally Identifiable Information (PII) is Any data element, such as a full name, email address, National Insurance number, postcode or biometric record, that can identify a living individual and requires strict protection in AI pipelines.
A Pilot (Proof of Concept / PoC) is A time-boxed, production-representative trial of an AI workflow with a defined cohort of users, designed to validate eval accuracy, latency, unit economics and adoption before full rollout.
Planning is The cognitive capability of an AI agent or reasoning model to break a high-level goal into a dependency-ordered sequence of verifiable sub-tasks and revise that plan when runtime obstacles arise.
Post-Training encompasses all training stages applied after self-supervised pre-training, including supervised fine-tuning, RLHF, direct preference optimisation and reinforcement learning with verifiable rewards, to turn a base model into a safe, capable assistant.
Pre-Training is The foundational, compute-intensive stage of building a language or multimodal model in which it learns syntax, factual world knowledge and reasoning representations across trillions of tokens of unlabelled data.
Precision is A classification and retrieval metric measuring the proportion of items flagged as positive or relevant by a model that are genuinely positive, quantifying how effectively the system avoids false alarms.
A Prediction is The specific output value, probability score or category label produced by a trained machine learning model when it evaluates a new, unseen data input during production inference.
Productivity Uplift is The empirically measured improvement in task cycle time, output throughput or defect reduction achieved when practitioners integrate a governed AI workflow, benchmarked against a pre-AI control baseline.
Prohibited AI Practices are Unacceptable-risk AI applications banned outright across the European Union under Article 5 of the EU AI Act since 2 February 2025, including social scoring, workplace emotion recognition and untargeted facial-image scraping.
A Prompt is The complete runtime payload of natural-language instructions, retrieved context documents, few-shot examples and media inputs sent to a generative AI model to condition and elicit a completion.
Prompt Caching is An inference optimisation that persists pre-computed KV cache states for recurring prompt prefixes, such as long system instructions, tool schemas and reference documents, cutting time-to-first-token and input token cost by up to 90%.
Prompt Chaining is An orchestration pattern that decomposes a complex task into a sequence of focused model calls in which the validated output of one prompt is passed as the structured input to the next.
Prompt Engineering is The systematic engineering discipline of designing, versioning and evaluating model instructions, contextual constraints, examples and output schemas to maximise accuracy, determinism and safety in production AI applications.
Prompt Injection is A security vulnerability in which adversarial instructions submitted directly by a user or hidden indirectly inside retrieved web pages, emails or documents override a language model's system prompt to hijack tool calls or exfiltrate data.
A Prompt Template is A version-controlled, reusable prompt blueprint containing named variable placeholders that application code populates dynamically with validated user inputs, system rules and retrieved database records at runtime.
A Prototype is An early, working software build of an AI feature created rapidly to test user interaction patterns, prompt behaviour and edge-case failure modes with stakeholders prior to production engineering.
A Provider, under the EU AI Act, is the natural or legal person or organisation that develops an AI system or general-purpose AI model, or has it developed, and places it on the market under its own name or trademark.
QLoRA is A memory-efficient fine-tuning technique that backpropagates gradients through a frozen, 4-bit NormalFloat quantised base language model into trainable 16-bit Low-Rank Adaptation matrices, enabling 70B-parameter fine-tuning on a single GPU.
Quantisation is A model compression technique that converts neural network weights and activations from high-precision 16-bit or 32-bit floating-point numbers into lower-bit representations such as INT8, FP8 or INT4, reducing VRAM footprint and accelerating inference.
Query Rewriting is A retrieval preprocessing step in which a fast language model reformulates, disambiguates or expands a user's conversational prompt into one or more self-contained search queries to improve recall.
A Rate Limit is An infrastructure quota enforced by an AI model provider or API gateway that caps the maximum number of requests per minute (RPM) or tokens per minute (TPM) an organisation can consume.
ReAct (Reason + Act) is A foundational agent execution paradigm in which a language model interleaves explicit verbal reasoning traces with external tool invocations and environment observations until it can answer accurately.
A Reasoning Model is A large language model trained with reinforcement learning to expend additional inference-time compute generating, critiquing and verifying multi-step internal thought chains before emitting its final response.
Recall is A classification and information retrieval metric measuring the proportion of all genuine positive records or relevant passages in a dataset that a model or search pipeline successfully retrieves.
A Recurrent Neural Network (RNN) is An earlier sequential deep learning architecture that processes inputs token by token using a recurrent hidden state, largely superseded by parallel self-attention transformers for language tasks.
Red Teaming is Structured adversarial security testing in which specialist engineers and domain experts probe an AI model or agent workflow for prompt injections, jailbreaks, data exfiltration paths and harmful outputs before production release.
Regression is A supervised machine learning task in which a model learns from historical data to predict a continuous numerical quantity, such as property value, energy demand or expected delivery time.
Regularisation refers to Mathematical constraints and training techniques, including weight decay, L1 and L2 penalties and dropout, applied during model optimisation to penalise excessive parameter complexity and prevent overfitting on training data.
Reinforcement Learning (RL) is A machine learning paradigm in which an agent learns an optimal decision policy by interacting with an environment through trial and error, maximising cumulative numerical rewards over time.
Reinforcement Learning from AI Feedback (RLAIF) is A post-training alignment technique in which a frontier language model evaluates and ranks candidate completions against a written constitution in place of human preference annotators.
Reinforcement Learning from Human Feedback (RLHF) is A post-training alignment pipeline in which human raters rank candidate model responses to train a reward model that then optimises the language model via reinforcement learning.
Reinforcement Learning with Verifiable Rewards (RLVR) is The post-training method behind modern reasoning models, rewarding the model against deterministic ground-truth checks such as passing software unit tests or solving mathematical equations.
Reranking is A second-stage retrieval pass that scores candidate chunks returned by fast vector or hybrid search using a cross-encoder model to place the highest-relevance passages at the top of the prompt context.
Responsible AI is The operational practice of designing, procuring, governing and monitoring enterprise artificial intelligence systems with verifiable engineering controls for safety, demographic fairness, data privacy, explainability and human accountability.
Retrieval is The runtime search operation of querying an inverted index, relational database or vector store to fetch the specific document passages or structured records most relevant to a user's prompt.
Retrieval-Augmented Generation (RAG) is An AI architecture that retrieves relevant passages from external documents or databases at query time and injects them into a language model's context window to generate cited, grounded answers.
Return on Investment (ROI) is The financial performance ratio comparing the net commercial value generated by an AI deployment, through labour capacity freed, revenue conversion or error reduction, against its total build and run cost.
A Reward Model is A specialised neural network trained on human or AI preference comparisons to output a scalar quality score for any prompt-completion pair, guiding policy optimisation during RLHF.
Robotic Process Automation (RPA) refers to Deterministic software scripts and UI bots that automate repetitive clerical tasks across legacy enterprise interfaces by replaying fixed keystrokes, mouse clicks and rule-based field mappings.
A Sandbox is A hardened, isolated compute environment, such as an ephemeral container or microVM with restricted network and filesystem access, in which an AI agent can safely execute untrusted generated code or shell commands.
Scaling Laws are Empirical power-law relationships, such as the Chinchilla compute-optimal frontier, that predict how a foundation model's cross-entropy loss improves as parameter count, dataset tokens and training compute scale.
A Scheduled Task is An automated background agent job triggered by a recurring cron expression or timer to execute autonomous research, pipeline monitoring, data reconciliation or reporting without manual user prompting.
An SDK (Software Development Kit) is A typed client library, such as the Anthropic SDK, OpenAI SDK or Vercel AI SDK, that provides application developers with idiomatic abstractions for streaming, structured outputs and tool calling.
Self-Attention is The core mechanism within a transformer block that compares every token in a sequence against every other token in the same sequence, capturing long-range grammatical and semantic dependencies.
Self-Consistency is A decoding and prompting strategy that samples multiple independent chain-of-thought reasoning paths for the same question at non-zero temperature and selects the most frequent final answer by majority vote.
Self-Hosting is The deployment of open-weight language or vision models on an organisation's own virtual private cloud or on-premises GPU servers using inference engines such as vLLM or TensorRT-LLM for strict data sovereignty.
Self-Supervised Learning is A training paradigm in which a model learns rich representations from raw, unlabelled data by generating its own supervisory targets, such as predicting masked or subsequent words in text.
Semantic Search is An information retrieval method that matches queries to documents based on conceptual meaning and user intent using dense vector embeddings rather than relying solely on exact keyword overlap.
Semi-Supervised Learning is A machine learning approach that combines a small, high-quality set of human-labelled examples with a much larger corpus of unlabelled data to train accurate models at lower annotation cost.
Shadow AI is The unsanctioned use of consumer chatbots, unvetted SaaS AI features, browser extensions or personal API keys by employees to process proprietary company data or customer records outside IT and information-security governance.
Short-Term Memory refers to The active conversational turns, intermediate reasoning traces and recent tool outputs held directly inside a language model's context window for the duration of a single session.
A Small Language Model (SLM) is A compact neural language model, typically between one billion and ten billion parameters, optimised for low-latency, low-cost inference on a single GPU or on-device NPU for domain-specific workflows.
Speculative Decoding is A lossless inference acceleration technique in which a compact draft model rapidly proposes several candidate tokens ahead and the target large language model verifies them in parallel within a single forward pass.
Speech Recognition (Speech-to-Text) is The acoustic and language modelling technology that converts spoken human audio streams into accurate written text transcripts for voice assistants, call centre analytics and clinical dictation.
Speech Synthesis (Text-to-Speech) is The generative audio technology that converts written text into natural-sounding human speech with controllable prosody, emotional tone, pacing and vocal timbre using neural acoustic foundation models.
A State Space Model (SSM) is A sequence modelling architecture, exemplified by Mamba, engineered to process long token sequences with linear compute and memory scaling rather than the quadratic cost of full attention.
Stochastic describes A process governed by probability distributions and controlled random sampling; because language models sample tokens stochastically at temperatures above zero, identical prompts can yield varying completions across runs.
Streaming is The transport pattern of sending generated tokens from an inference server to the client UI incrementally via Server-Sent Events or WebSockets as each token is decoded, reducing perceived user latency.
Structured Data is Information organised within a strict, predefined schema of tables, rows, columns and typed fields, such as relational SQL databases, financial ledgers and enterprise CRM records, ready for programmatic querying.
Structured Output (JSON Mode) is A constrained decoding feature that forces a language model to emit responses strictly conforming to a supplied JSON Schema or Pydantic type definition so downstream software can parse them reliably.
A Subagent is A subordinate AI agent invoked by a parent orchestrator to execute a narrowly scoped task inside its own clean, isolated context window before returning a concise structured summary.
Supervised Fine-Tuning (SFT) is The post-training stage in which a pre-trained base model is trained on curated, high-quality demonstrations of user instructions and ideal assistant completions using next-token cross-entropy loss.
Supervised Learning is The foundational machine learning paradigm in which a model is trained on paired input features and verified ground-truth labels so it learns to map new inputs to accurate predictions.
Symbolic AI is The classical approach to artificial intelligence built on explicit, human-readable symbols, formal domain ontologies and deterministic logic rules rather than statistical pattern learning across multi-layered neural networks.
Synthetic Data is Artificially generated information created by language models, statistical simulations or rendering engines to mirror the statistical properties of real-world production records when real data is scarce, private or imbalanced.
A System Prompt is The privileged, developer-authored instruction block placed at the start of a model's context window that establishes its role, tone, tool-use rules and safety boundaries before any user message is processed.
Systemic Risk, under Article 51 of the EU AI Act, denotes high-impact societal, cybersecurity or CBRN risks associated with frontier general-purpose AI models trained above 10^25 FLOPs, triggering mandatory adversarial evaluations and incident reporting.
Temperature is An inference sampling hyperparameter, typically scaled between 0.0 and 2.0, that rescales token logits before selection; values near zero favour deterministic, high-probability replies whereas higher values increase lexical diversity.
A Test Set is A strictly held-out partition of a dataset kept completely separate from model training and hyperparameter tuning, used only at the end of development to measure unbiased real-world performance.
Test-Time Compute (Inference-Time Scaling) is The allocation of additional GPU compute during inference so a model can generate extended internal reasoning chains, sample parallel candidate solutions and verify answers before replying.
Text-to-Video is The generative AI capability of synthesising temporally coherent video clips, camera movements and physical motion sequences directly from natural-language prompts, storyboards or reference keyframes using spacetime diffusion transformers.
Throughput is An infrastructure capacity metric measuring the aggregate volume of output tokens per second or concurrent requests per second that an inference server cluster can sustain under production load.
A Token is The atomic unit of text, code or media that a language model ingests and generates, representing a whole word, sub-word fragment or punctuation character, where 100 English tokens equal roughly 75 words.
Tokenisation is The deterministic preprocessing step that segments raw input text into discrete sub-word tokens using algorithms such as Byte-Pair Encoding (BPE) and maps each token to an integer ID in the model's vocabulary.
Top-k Sampling is A decoding filter that restricts a language model at each generation step to sampling only from the k single highest-probability next tokens, discarding the remaining vocabulary tail.
Top-p (Nucleus Sampling) is A dynamic decoding filter that selects the next token from the smallest set of candidate tokens whose cumulative probability reaches a threshold p, adapting vocabulary breadth to model certainty.
Total Cost of Ownership (TCO) is The full lifecycle cost of an enterprise AI system, combining model API tokens or GPU hosting, vector database storage, integration engineering, ongoing evaluation suites, governance audits and maintenance.
A TPU (Tensor Processing Unit) is Google's custom-designed application-specific integrated circuit (ASIC) engineered around high-density systolic array matrix units to accelerate large-scale neural network training and high-throughput cloud inference workloads.
Training Data is The primary corpus of labelled or unlabelled examples fed into a machine learning algorithm during training so the model can iteratively adjust its internal weights and learn domain patterns.
Transfer Learning is The machine learning paradigm of reusing feature representations learned by a model trained on a large general dataset as the initialisation point for a specialised downstream task.
The Transformer is The deep learning architecture introduced in the 2017 paper Attention Is All You Need that processes sequences in parallel using multi-head self-attention, forming the backbone of modern large language models.
Transparency in AI governance is the legal and ethical obligation to inform users clearly when they are interacting with an AI system, label synthetically generated media and document the data sources behind automated decisions.
Tree of Thoughts (ToT) is A deliberate reasoning framework that generalises chain-of-thought by enabling a language model to branch across multiple candidate reasoning paths, score intermediate states and backtrack from dead ends.
The Turing Test is A behavioural benchmark proposed by British mathematician Alan Turing in 1950 to evaluate whether a computer conversing via text can fool a human interrogator into believing it is also human.
Underfitting occurs when a machine learning model has insufficient capacity, features or training iterations to capture the underlying structure of the data, resulting in high error rates on both training and test sets.
Unstructured Data refers to Information that does not conform to a fixed relational database schema, including commercial contracts, emails, PDF reports, support tickets, call centre audio recordings, images and video files.
Unsupervised Learning is A machine learning paradigm in which algorithms analyse raw, unlabelled datasets to discover hidden customer clusters, latent feature representations or statistical anomalies without requiring human-annotated target answers.
Upskilling is Structured, role-specific practitioner training that equips existing employees to engineer context, verify model outputs, build agentic automations and apply AI governance rules within their daily professional business workflows.
A Use Case is A specific, bounded operational workflow or customer problem, such as automating procurement contract review or triaging first-line support tickets, selected for AI implementation with a defined success metric.
A User Prompt is The runtime message, query or uploaded document submitted by the end user within an interactive session, processed alongside but kept logically distinct from the application's system prompt.
A Validation Set is An intermediate split of data held separate from the training set and used repeatedly during model development to tune hyperparameters, compare architectures and detect overfitting early.
The Vanishing Gradient is An optimisation pathology in deep neural networks in which backpropagated error gradients shrink exponentially towards zero across layers, preventing early layers from updating their weights effectively.
A Vector is An ordered, fixed-length array of floating-point numbers that represents a point and direction in high-dimensional mathematical space, used in AI to encode the semantic features of text or media.
A Vector Database is A specialised storage engine, such as pgvector, Qdrant, Pinecone or Weaviate, built to index high-dimensional vector embeddings using approximate nearest-neighbour graphs and execute low-latency similarity queries alongside metadata filters.
Vibe Coding is An AI-native software development workflow, coined by Andrej Karpathy in 2025, in which builders steer coding agents through natural-language prompts and runtime feedback rather than authoring source code line by line.
A Vision-Language Model (VLM) is A multimodal foundation model combining a visual encoder with a language model backbone to jointly interpret images, scanned PDFs, charts or video frames alongside natural-language instructions.
A Voice Agent (Speech-to-Speech) is A real-time conversational AI system that processes raw spoken audio inputs and streams back generated speech with sub-second latency, supporting natural prosody and user barge-in interruptions.
Watermarking (AI Watermarking) is The embedding of imperceptible statistical or cryptographic signals, such as Google SynthID, into AI-generated text, images, audio or video so provenance tools can detect synthetic media in line with EU AI Act Article 50.
A Webhook is An event-driven HTTP POST callback dispatched automatically by a software system when a state change occurs, commonly used to trigger asynchronous AI agent workflows or notify applications when batch jobs finish.
Weights are The learnable floating-point or quantised numerical coefficients inside an artificial neural network that determine the strength and sign of the signal passed between connected neurons across consecutive layers.
Workflow Automation is The engineering of event-driven, multi-step business processes across SaaS APIs, internal databases and language models using visual or code-first orchestration platforms such as n8n, Make, Zapier or Temporal.
A World Model is An AI architecture trained to build an internal predictive simulation of 3D physical space, object permanence and causal dynamics, enabling robots and autonomous agents to plan actions before executing them.
XGBoost (Extreme Gradient Boosting) is A high-performance, open-source gradient-boosted decision tree library that remains a benchmark machine learning method for structured, tabular datasets across finance, retail pricing and operational risk.
YOLO (You Only Look Once) is A family of real-time computer vision models that frame object detection as a single neural network pass, locating and classifying multiple objects across live video streams in milliseconds.
Zero-Click Search is An information discovery journey in which a user receives a complete, synthesised answer directly inside an AI Overview, featured snippet or conversational answer engine without clicking through to an external website.
Zero-Shot Prompting is The practice of instructing a language model to perform a task using only direct task rules and its pre-trained capabilities, without providing any worked input-output examples in the prompt.
Put the terminology to work
Get the weekly Tuesday AI news roundup in your inbox, or apply for free membership to join our London, Leeds, Online and Women that Use AI meetups and WhatsApp community.