AI Glossary for Professionals

A hypothetical AI system with human-level or above-human-level cognitive abilities across all domains. No AGI currently exists. Most researchers consider it a long-term research goal, not an imminent development.

The European Union's Artificial Intelligence Act — the world's first comprehensive legal framework for AI. It classifies AI systems by risk level (minimal, limited, high, unacceptable) and sets obligations for developers and deployers of AI operating in the EU.

A software system that autonomously perceives its environment, makes decisions, and takes actions to achieve a goal — often by calling tools, browsing the web, or writing and executing code. AI agents differ from standard chatbots in that they act, not just respond.

An AI system that can autonomously plan and execute multi-step tasks, using tools such as web search, code execution, or file management to complete a goal with minimal human intervention. Examples include AI coding agents and research agents.

The challenge of ensuring AI systems behave in accordance with human values and intentions. Misalignment occurs when an AI optimises for a proxy goal in unintended ways — a central concern in AI safety research.

A set of protocols that allows one software system to communicate with another. In AI contexts, it typically refers to the interface through which developers send prompts to AI models programmatically and receive responses — for example, the OpenAI API or the Anthropic API.

The mathematical mechanism inside transformer models that allows them to weigh the relevance of different parts of the input when generating each output token. The foundation of modern LLM architecture, introduced in the 2017 paper "Attention Is All You Need".

A software system that perceives its environment, makes decisions, and takes actions to achieve goals with limited or no human oversight. AI autonomous agents typically combine an LLM with tools and a feedback loop.

A technique that crafts inputs designed to fool an AI model into making incorrect predictions or producing harmful outputs. In text-based systems, adversarial attacks include jailbreaks, prompt injections, and inputs designed to trigger misbehavior. Understanding adversarial vulnerabilities is essential for anyone deploying AI in security-sensitive contexts.

An AI assistant embedded into a professional tool or workflow that provides real-time suggestions, completions, or automation while the user remains in control. Microsoft Copilot and GitHub Copilot are the most widely adopted examples. The term emphasises human direction — the AI assists rather than acts independently.

The practice of using AI to produce a first draft of a document, email, or report, which a human then reviews, edits, and takes responsibility for before sending or publishing. An AI draft is a starting point, not a finished product — the professional remains accountable for the final output.

The branch of applied ethics concerned with the development and deployment of artificial intelligence. Key questions include: how to build fair and unbiased systems, who bears responsibility when AI causes harm, how to balance automation with human dignity, and what limits should be placed on AI autonomy. Increasingly embedded in corporate governance, regulation, and professional codes of conduct.

The policies, frameworks, and oversight mechanisms organisations use to manage AI development and deployment responsibly. AI governance covers risk assessment, model documentation, approval workflows, usage policies, and regulatory compliance. It is rapidly becoming a board-level concern for large organisations as AI adoption scales.

An AI safety company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei. Anthropic develops the Claude family of models and has published significant research on AI safety, constitutional AI, and model interpretability. One of the leading frontier AI labs alongside OpenAI and Google DeepMind.

The tendency of humans to over-rely on automated or AI systems and accept their outputs without sufficient critical evaluation. Automation bias is a documented cognitive risk in contexts where AI assistance is used for high-stakes decisions — such as medical diagnosis, legal research, or financial analysis. Awareness of automation bias is a core professional AI literacy skill.

The legal question of who is responsible when an AI system causes harm, error, or loss. In most jurisdictions, liability currently falls on the human or organisation that deployed or used the AI — not the AI itself. As AI systems become more autonomous, liability frameworks are evolving rapidly, particularly under the EU AI Act.

The ability to understand, critically evaluate, and effectively use AI tools in professional and everyday contexts. AI literacy encompasses knowing how AI systems work at a conceptual level, understanding their limitations, and applying them productively. It is increasingly listed as a required competency in job descriptions across all industries.

The ability of an AI system to retain information from previous conversations or sessions and use it in future interactions. Without explicit memory, LLMs treat every conversation as fresh. Persistent memory — built via external storage, summarisation, or retrieval — allows AI assistants to personalise responses and maintain context across sessions.

An organisational document governing how employees may use AI tools at work. A well-formed AI policy covers permitted uses, prohibited inputs (such as confidential data or personal information), quality standards for AI-assisted outputs, disclosure requirements, and escalation procedures for AI errors.

The capacity of an AI model to break down complex problems into steps, check intermediate conclusions, and arrive at a well-supported answer — as distinct from producing a fluent but potentially incorrect response. Reasoning capability varies significantly between models and is measurable via mathematical, logical, and scientific benchmarks.

A standardised test or dataset used to evaluate and compare AI model performance. Common benchmarks include MMLU (knowledge), HumanEval (coding), and HellaSwag (commonsense reasoning). Often cited in model announcements to justify capability claims.

Bidirectional Encoder Representations from Transformers — a foundational transformer-based language model from Google (2018) that revolutionised natural language understanding tasks. Trained by predicting masked words from both left and right context, unlike earlier left-to-right models.

Systematic errors in AI outputs that reflect skewed patterns in training data, model design, or objective functions. Can manifest as demographic bias, confirmation bias, or recency bias. A key concern in responsible AI deployment.

The algorithm used to train neural networks by computing gradients of the loss function with respect to model weights, then adjusting those weights to reduce error. Backpropagation works backwards through the network layers using the chain rule of calculus. Understanding it helps professionals grasp why LLMs require massive compute resources to train.

A pre-trained language model that has learned from large-scale data but has not yet been fine-tuned for specific tasks or instruction-following. Base models produce text completions rather than assistant-style responses. Most consumer AI products are instruction-tuned versions of base models, not the base models themselves.

A prompting technique where you instruct the model to reason through a problem step by step before giving a final answer. Significantly improves accuracy on complex reasoning tasks.

An AI assistant developed by Anthropic, designed with a focus on safety, helpfulness, and honesty. One of the leading frontier AI models alongside GPT-4 and Gemini.

The capability of an AI model to control a computer interface — clicking buttons, typing text, navigating applications, and reading the screen — to complete tasks on behalf of a user. Pioneered by Anthropic's Claude computer use feature (2024), this capability extends AI agents beyond text into direct interaction with any software.

An alternative term for hallucination preferred by some researchers, borrowed from neuroscience. Refers specifically to the generation of plausible but unsupported outputs, emphasising that the model is not deliberately deceptive — it simply fills gaps in its knowledge with confident-sounding content.

In LLM usage, the full text of the conversation or document that the model has access to, including prior messages and any system prompt. What is "in context" determines what the model can reference when generating a response.

The maximum amount of text an LLM can process in a single conversation or request. Content outside the context window is not visible to the model.

Microsoft's AI assistant brand, embedded across Office 365, Windows, GitHub, and other Microsoft products. Powered by OpenAI models. Also used generically to describe AI tools that assist professionals with tasks rather than replacing them entirely.

A training methodology developed by Anthropic in which AI models are trained to follow a set of principles (a 'constitution') — using AI feedback rather than exclusively human feedback to evaluate and revise outputs. Constitutional AI is designed to produce models that are more consistently helpful, harmless, and honest at scale.

AI systems designed to conduct natural language conversations with users, including chatbots, virtual assistants, and LLM-powered interfaces. Conversational AI ranges from simple rule-based chatbots to sophisticated systems like ChatGPT and Claude. Effective conversational AI design requires careful attention to turn management, memory, and fallback handling.

A numerical value assigned by an AI model to indicate how certain it is in a particular prediction or classification. Confidence scores are commonly used in classification models and structured AI outputs. Standard LLMs do not natively expose calibrated confidence scores for generated text, which is why hallucination detection remains challenging.

A large, structured collection of text used to train or evaluate a language model. Pre-training corpora for frontier models include hundreds of billions to trillions of tokens drawn from the web, books, academic papers, and code repositories. The composition and quality of a training corpus heavily influence model capability and bias.

A unique identifier assigned to a professional certificate that enables independent verification of its authenticity by employers or third parties. Credential IDs are typically included on certificate documents and can be entered into a verification page to confirm the holder's name, course, and issue date without contacting the issuing organisation directly.

Storing the results of expensive AI computations — such as prompt embeddings or repeated inference outputs — to avoid recomputing them on identical future requests. Caching reduces inference cost and latency in production AI systems. Prompt caching is increasingly supported by major LLM API providers for frequently repeated system prompts.

The process of making an AI model available for use in production — via an API, embedded in an application, or run locally. Deployment decisions affect latency, cost, data privacy, and how frequently the model can be updated.

A class of generative AI model that creates outputs (typically images, audio, or video) by learning to reverse a process of gradually adding noise to training data. Stable Diffusion, DALL-E 3, and Midjourney use diffusion or hybrid approaches.

Techniques used to artificially expand training datasets by creating modified versions of existing data — for example, rotating images, paraphrasing sentences, or adding noise. Data augmentation improves model robustness and reduces overfitting without requiring the collection of entirely new training data.

A form of adversarial attack in which malicious data is deliberately introduced into a model's training set to manipulate its outputs. Data poisoning is a significant concern for models trained on web-scraped data or open contribution datasets. It is one reason model provenance and data governance matter in enterprise AI deployments.

A verifiable, portable credential issued in a standardised digital format — typically following the Open Badges standard — that can be displayed on LinkedIn profiles, personal websites, and digital CVs. Digital badges embed metadata including the issuer, earner, criteria, and issue date, making them independently verifiable.

A technique in which a smaller 'student' model is trained to replicate the behavior of a larger 'teacher' model. Distillation produces compact, efficient models that retain much of the teacher model's capability at a fraction of the compute cost. Many smaller open-source models are produced via distillation from larger frontier models.

A subfield of machine learning that uses neural networks with many layers (deep networks) to learn representations of data directly from raw inputs. Deep learning enables models to learn hierarchical features without manual feature engineering. It underpins virtually all modern AI capabilities — from image recognition to language generation.

The requirement that data be stored and processed within a specific geographic jurisdiction. Data residency rules are imposed by laws such as the EU's GDPR and various national data localisation statutes. When using cloud-based AI tools, professionals in regulated industries must verify whether data is processed in compliant regions.

The process by which an LLM generates output tokens one at a time, selecting each from a probability distribution over its vocabulary. Decoding strategies — greedy, beam search, sampling with temperature — determine the trade-off between output quality, diversity, and computational cost. Choosing the right decoding strategy is a key consideration in production LLM applications.

Numerical representations of text (or other data) that capture semantic meaning. Used in semantic search, RAG systems, and clustering applications.

Capabilities that appear in large AI models that were not explicitly trained for and were not present in smaller versions of the same model. Examples include multi-step reasoning, code generation, and language translation emerging as model scale increases. Emergence is one reason AI capabilities are difficult to predict before training completes.

The part of a neural network architecture that converts raw input (text, image, audio) into a dense numerical representation — a vector — that captures semantic meaning. Encoders are used in retrieval systems, semantic search, and classification tasks. The encoder in a transformer processes input bidirectionally, unlike a decoder.

A machine learning technique that combines predictions from multiple models to improve accuracy, robustness, or coverage. Ensemble methods reduce the risk of any single model's failure modes dominating the output. They are commonly used in production AI systems where reliability matters more than inference speed.

The systematic process of measuring an AI model's performance, accuracy, fairness, and behavior. Evaluation methods include automated benchmarks, human preference ratings, red-teaming, and domain-specific accuracy tests. Rigorous evaluation is essential before deploying AI in professional or high-stakes contexts.

AI processing that occurs on local devices — smartphones, laptops, embedded systems — rather than in the cloud. Edge AI enables private, low-latency inference without sending data to external servers. It is particularly relevant for applications requiring data privacy or offline operation, and is enabled by model compression techniques like quantisation and distillation.

Providing 2–5 examples of the desired input/output format in your prompt so the model learns the pattern before attempting the real task.

Additional training of a pre-trained model on a specific dataset to specialise its behavior. Distinct from prompting — requires training compute and data.

A large AI model trained on vast amounts of data that can be adapted to a wide range of downstream tasks through fine-tuning or prompting. GPT-4, Claude, Gemini, and Llama are all foundation models. The term was coined by Stanford researchers in 2021.

A feature of some LLMs that allows the model to output structured arguments for predefined functions, enabling it to trigger real-world actions — such as querying a database, making an API call, or sending an email — within an AI application.

The process of selecting, transforming, and creating input variables (features) used to train a machine learning model. Good feature engineering was the primary source of model improvement in classical ML before deep learning. It remains important in structured data domains where domain expertise can significantly improve model performance.

Google DeepMind's family of large language models. Gemini Ultra, Pro, and Nano target different use cases and device types. Deeply integrated into Google Search, Google Workspace, and Android.

Generative Pre-trained Transformer — a family of large language models developed by OpenAI. GPT-1 was released in 2018; GPT-4 (2023) is one of the most capable publicly available models. The abbreviation is also used informally to refer to ChatGPT.

The practice of anchoring AI outputs to verified external sources or real-time data, reducing reliance on the model's internal — and potentially outdated or incorrect — knowledge. RAG is the most common grounding technique.

Safety or policy constraints applied to AI systems to prevent harmful, misleading, or off-topic outputs. Can be implemented via system prompts, model fine-tuning, output filters, or external classifiers.

AI systems that can produce new content — text, images, audio, video, code, or other outputs — rather than simply classifying or predicting. Generative AI includes large language models, image diffusion models, and multimodal systems. The term became mainstream following the release of ChatGPT in late 2022.

The optimisation algorithm used to train neural networks by iteratively adjusting model weights in the direction that reduces the loss function. Gradient descent comes in several variants — batch, stochastic, and mini-batch — each with different computational trade-offs. Understanding it contextually helps professionals evaluate why AI training is compute-intensive.

The architecture and model family developed by OpenAI. 'Generative' refers to the model's ability to produce new text; 'pre-trained' means trained on a large corpus before fine-tuning; 'transformer' is the underlying neural network architecture. GPT models (GPT-3, GPT-4, GPT-4o) are the foundation for ChatGPT and the OpenAI API.

Specialised hardware originally designed for rendering graphics, now the primary compute platform for training and running AI models. GPUs excel at the parallel matrix operations that neural network training requires. The availability and cost of GPU compute is one of the primary barriers to training large AI models independently.

When an AI model generates plausible-sounding but factually incorrect or fabricated information. A known limitation of current LLMs that requires human verification.

A design approach that keeps human oversight at critical decision points in an AI workflow. Required by many regulatory frameworks for high-stakes AI decisions in healthcare, finance, and legal contexts. Abbreviated HITL.

A design pattern in which humans are incorporated into the AI decision-making process — either to provide training labels, to review AI outputs before acting on them, or to handle cases the AI cannot resolve confidently. HITL is the recommended approach for high-stakes AI applications in legal, medical, financial, and HR contexts.

A configuration variable set before training begins that controls the learning process — distinct from model parameters, which are learned during training. Examples include learning rate, batch size, and number of training epochs. Hyperparameter tuning significantly affects model performance and is one of the primary activities in applied machine learning.

The ability of large language models to adapt to new tasks based on examples provided directly in the prompt, without updating the model's parameters. Few-shot and zero-shot prompting both rely on in-context learning.

The process of running a trained model to generate outputs. When you send a prompt to ChatGPT and receive a response, the model is performing inference.

A fine-tuning technique that trains a base language model on instruction-response pairs, making it better at following user directions. Used to create chat-optimised models from raw pre-trained models. ChatGPT and Claude are instruction-tuned models.

The degree to which the internal mechanisms of an AI model can be understood by humans. Modern large neural networks are largely 'black boxes' — their internal representations are difficult to interpret directly. Interpretability research aims to understand why models produce specific outputs, which is critical for debugging failures and building trustworthy AI systems.

The process of cleaning and validating user inputs before they are processed by an AI system — removing or neutralising content that could trigger prompt injection, jailbreaks, or unintended model behavior. Input sanitisation is a standard defensive measure in production AI applications that handle untrusted external content.

A technique used to bypass an AI model's safety guardrails by crafting inputs that manipulate the model into generating content it would normally refuse. AI developers actively work to patch jailbreaks through model updates and improved safety training.

An output mode available in some LLM APIs that constrains the model to produce valid JSON rather than free-form text. JSON mode makes AI outputs easier to parse programmatically and reduces the need for post-processing. It is commonly used in production AI applications that need to feed LLM outputs into downstream systems.

A structured representation of information as entities (nodes) and relationships (edges), used to represent factual knowledge in a machine-readable form. Knowledge graphs can be used to supplement LLMs with reliable structured facts — reducing hallucination risk in fact-sensitive applications. Notable examples include Google's Knowledge Graph and Wikidata.

The process of training a smaller model to mimic the outputs of a larger model, often used interchangeably with 'model distillation'. Knowledge distillation enables organisations to deploy compact, fast models that retain much of the larger model's capability at a fraction of the inference cost.

The date after which an AI model has no training data, meaning it lacks knowledge of events, publications, or developments that occurred after that date. Knowledge cutoffs are a significant practical limitation for professionals who need current information. Retrieval-augmented generation (RAG) and web-search tools are the primary mitigations.

The time delay between sending a prompt to an AI model and receiving the first token of its response. Critical for real-time applications. Measured as time-to-first-token (TTFT) or total generation time. Larger models tend to have higher latency.

A family of open-weights large language models developed by Meta. Llama 2 and Llama 3 are widely used by researchers and businesses who want to run models locally or build custom fine-tunes without API access restrictions.

A type of AI model trained on large amounts of text data to understand and generate human language. Examples include GPT-4, Claude, and Gemini.

The set of practices, tools, and workflows for operating and maintaining LLM-based applications in production. Includes prompt versioning, model evaluation, monitoring for output quality drift, and cost management.

The compressed, abstract representation of data that a neural network learns internally. In image generation, the latent space encodes visual concepts in a form the model can manipulate — allowing techniques like interpolation between images. Understanding latent spaces helps explain why AI models can combine concepts in ways that are surprising but structurally coherent.

A mathematical function that measures the difference between a model's predictions and the correct answers during training. The loss function guides the optimisation process — the model is trained to minimise it. Language models typically use cross-entropy loss over predicted next-token probabilities.

A parameter-efficient fine-tuning technique that inserts small trainable matrices into a model's layers rather than updating all weights. LoRA dramatically reduces the memory and compute cost of fine-tuning large models. It has become the dominant method for producing custom domain-specific model variants, particularly in the open-source community.

An LLM with an extended context window — typically 100,000 tokens or more — capable of processing entire books, large codebases, or lengthy document sets in a single pass. Claude 3 and Gemini 1.5 Pro are prominent long-context models. Long-context capability is valuable for document analysis, contract review, and knowledge synthesis tasks.

A subset of AI in which systems learn patterns from data rather than being explicitly programmed with rules. Modern LLMs are built on machine learning. Deep learning — using multi-layer neural networks — is a further subset and the dominant approach for frontier AI.

In AI, a mathematical structure — typically a neural network — that has been trained on data to perform a task. Colloquially, "the model" refers to the AI system you are interacting with, such as GPT-4o or Claude 3.5 Sonnet.

A theoretical failure mode where AI models trained on AI-generated content progressively degrade in quality and diversity over successive generations. A growing concern as AI-generated text and images flood the training corpora used for future models.

AI systems that can process and generate more than one type of data — for example, text, images, audio, and video within a single model. GPT-4o and Gemini are multimodal models. Multimodal capability enables use cases such as analysing a chart uploaded as an image, or transcribing and summarising a meeting recording.

AI models that can process and generate multiple types of data — typically text, images, audio, and video in combination. GPT-4o and Gemini Ultra are multimodal. Multimodal models significantly expand AI capabilities beyond text-only interactions.

A widely used benchmark that tests language model performance across 57 academic subjects including law, medicine, history, and mathematics. MMLU scores are commonly cited in model comparisons and press releases. Like all benchmarks, high MMLU scores correlate with but do not guarantee real-world task performance.

A standardised document that describes a machine learning model's intended use, performance characteristics, evaluation results, ethical considerations, and known limitations. Model cards were proposed by Google researchers in 2019 and are now a common form of AI transparency documentation. Reviewing a model card before deployment helps organisations understand what they are working with.

A financial services framework for identifying, assessing, and controlling risks arising from the use of AI and quantitative models in decision-making. Rooted in US banking guidance (SR 11-7) and increasingly applied to AI systems globally. MRM requires model validation, performance monitoring, and clear governance over who can approve model deployment.

An AI architecture in which multiple AI agents interact, collaborate, or compete to complete tasks too complex for a single model to handle alone. Multi-agent systems can divide a large task into specialised subtasks — one agent plans, another executes, a third reviews. They are increasingly common in agentic AI workflows.

A short, focused certification that demonstrates competency in a specific skill or topic, as distinct from a full degree or multi-year qualification. Micro-credentials are designed to be earned quickly, stacked over time, and displayed alongside traditional qualifications. They are increasingly accepted by employers as evidence of current, practical skill.

A neural network architecture in which different parts of the model ('experts') specialise in different types of input, with a routing mechanism selecting which experts to activate for each token. MoE architectures allow very large effective model capacity while keeping per-token compute cost manageable. GPT-4 is widely believed to use a MoE architecture.

An open standard developed by Anthropic for connecting AI models to external tools, data sources, and systems in a standardised way. MCP defines how AI assistants discover and invoke tools, and how tools return results. It is emerging as an industry-standard interface layer for agentic AI applications.

A prompt that instructs an AI model to generate or improve another prompt, rather than directly completing a task. Meta-prompting is used to systematically improve prompt quality and is a technique in automated prompt optimisation. It illustrates the flexibility of LLMs as general-purpose instruction followers.

AI designed to perform a specific task — in contrast to Artificial General Intelligence (AGI). All current commercial AI is narrow AI.

A computational model loosely inspired by biological neurons, consisting of layers of interconnected nodes that process and transform inputs. The foundational architecture of modern machine learning. Deep neural networks — with many layers — power virtually all current frontier AI systems.

The property of AI models whereby the same prompt can produce different outputs on different runs. LLMs are non-deterministic by design — a temperature setting above zero introduces randomness in token selection. Non-determinism requires careful handling in production applications where consistency matters.

A technique used within neural networks to stabilise and accelerate training by normalising the activations of each layer. Layer normalisation is a key component of the transformer architecture. Without normalisation, deep neural networks suffer from training instability — gradients either vanish or explode.

The field of AI focused on enabling computers to understand, interpret, and generate human language. NLP underpins search engines, translation tools, chatbots, sentiment analysis, and large language models. While older NLP relied on hand-crafted linguistic rules, modern NLP is dominated by transformer-based deep learning models.

The AI research company behind the GPT model family, ChatGPT, and DALL-E. Founded in 2015 as a non-profit, later restructured to a capped-profit company with Microsoft as a major investor. One of the most influential organisations in modern AI development.

When an AI model learns training data too well — including its noise and errors — resulting in poor performance on new, unseen data. A fundamental machine learning problem addressed through techniques like regularisation, dropout, and more diverse training datasets.

AI models whose weights and, in some cases, training code are made publicly available for inspection, fine-tuning, and deployment. Notable open-source or open-weight models include Meta's Llama, Mistral, and Falcon. Open-source AI enables organisations to run models on their own infrastructure — important for data privacy and cost control in enterprise contexts.

The practice of instructing an AI model to produce outputs in a specific structure — bullet points, JSON, markdown tables, numbered steps. Consistent output formatting is one of the most practically valuable prompt engineering techniques — it makes AI outputs easier to process, integrate, and act on.

The process of coordinating multiple AI models, tools, and data sources to complete complex workflows. AI orchestration frameworks like LangChain and LlamaIndex manage prompt routing, memory, tool calling, and response chaining. Orchestration is the engineering layer that enables sophisticated multi-step AI applications.

The numerical values (weights and biases) inside a neural network that are adjusted during training. A model's parameter count is often used as a rough measure of its scale — GPT-3 had 175 billion parameters; larger models have hundreds of billions or more.

The initial large-scale training phase of a foundation model, where it learns general language patterns from vast amounts of text. Computationally very expensive — costing millions of dollars for frontier models. Pre-training creates the base capability; fine-tuning and prompting refine it.

The input you give to an AI model — the instructions, question, or context that tells it what to generate. Prompt quality directly affects output quality.

A security vulnerability in AI applications where malicious instructions embedded in user input, documents, or retrieved data hijack the AI's behavior. A critical concern when building agents that process untrusted external inputs.

A curated, organised collection of tested prompts designed for specific professional tasks. Used by teams to standardise AI outputs, preserve effective prompt patterns, and onboard new team members to AI workflows quickly.

A family of techniques that fine-tune a small subset of a model's parameters (or add small adapter layers) rather than updating all weights. PEFT methods like LoRA and Prefix Tuning dramatically reduce the compute cost of fine-tuning large models. They have made domain-specific model adaptation practical for organisations without large GPU clusters.

A metric used to evaluate language model quality — roughly measuring how surprised the model is by a given text. Lower perplexity indicates the model assigns higher probability to the observed text, suggesting better language modeling. Perplexity is a useful intrinsic metric but does not directly measure real-world task performance.

A technique in which the output of one LLM prompt is used as the input to a subsequent prompt, breaking complex tasks into a sequence of smaller, more manageable steps. Prompt chaining improves reliability on tasks that are too complex for a single prompt to handle well. It is a foundational pattern in agentic AI workflows.

A reusable prompt structure with variable placeholders that can be filled with specific context for each use. Prompt templates standardise AI interactions for repeatable tasks — such as email drafts, report summaries, or data extraction — ensuring consistent output quality across different users and inputs.

The practice of tracking changes to prompts over time — storing previous versions, documenting what changed and why, and measuring the impact of changes on output quality. Prompt versioning is part of good LLMOps hygiene: a prompt that works well today may behave differently after a model update, making version history valuable for debugging.

A technique that reduces the numerical precision of a model's parameters (e.g. from 32-bit to 4-bit floats) to decrease file size and memory requirements. Enables large models to run on consumer hardware with some tradeoff in output quality.

A technique that combines a language model with a search step, allowing the model to retrieve relevant documents before generating a response. Reduces hallucinations for factual queries.

An LLM variant designed for slow, deliberate, step-by-step problem solving rather than fast text generation. Examples include OpenAI o1 and o3. Generally more accurate on complex logic, mathematics, and coding tasks but slower and more expensive per query than standard chat models.

A framework and set of practices for developing and deploying AI in ways that are fair, transparent, accountable, and aligned with human values and legal requirements. Increasingly codified in corporate governance policies and regulations such as the EU AI Act.

A training technique where human raters evaluate model outputs and their preferences are used to reward-train the model to produce more helpful, accurate, and less harmful responses. Used to align ChatGPT, Claude, and similar models with human values.

A structured process in which a team attempts to find failure modes, safety violations, and unexpected behaviors in an AI system — analogous to penetration testing in cybersecurity. Red-teaming is used by AI labs to identify and mitigate harmful outputs before public deployment. It is increasingly recommended as a standard step in enterprise AI deployment reviews.

An architecture that enhances LLM outputs by retrieving relevant documents from an external knowledge base before generating a response. RAG reduces hallucination by grounding the model's output in retrieved factual content. It is the standard approach for building AI systems that need access to proprietary or up-to-date knowledge without full model fine-tuning.

A prompting technique where you instruct the model to adopt a specific persona, expertise level, or professional role before responding. Role prompting can significantly improve output quality for domain-specific tasks by framing the model's response from an appropriate professional perspective. Example: 'You are an experienced employment lawyer...' before asking a legal question.

Search that finds conceptually relevant results based on meaning rather than keyword matching. Uses embeddings to compare query meaning to document meaning, enabling retrieval of relevant content even when exact words do not match.

A feature of LLMs that constrains the model to produce outputs in a defined format — such as JSON, XML, or a custom schema. Critical for integrating AI into software systems where outputs must be reliably parsed by downstream code.

A type of machine learning where the model is trained on labelled examples (input-output pairs). Instruction tuning is a form of supervised learning. Contrasts with unsupervised learning (finding patterns in unlabelled data) and reinforcement learning (optimising through trial and reward).

Instructions provided to an AI model before the user conversation begins. Used to set the model's role, tone, constraints, and behavior for a specific application.

The process of selecting the next token in an LLM's output based on probability distributions. Sampling strategies include greedy (always pick the highest-probability token), top-k, top-p (nucleus), and temperature scaling. The right sampling strategy depends on the use case — high temperature for creative tasks, low temperature for factual or consistent outputs.

The use of AI tools by employees without official company approval, IT knowledge, or security review. Shadow AI is the AI equivalent of shadow IT — it typically arises when approved tools are absent or slow to procure, but carries significant compliance, data privacy, and liability risks when employees input confidential information into unapproved systems.

A mathematical function applied in neural networks to convert raw output scores (logits) into a probability distribution. In language models, softmax is applied to the vocabulary at each generation step to determine token selection probabilities. It is a core component of the transformer architecture.

Training data that is artificially generated rather than collected from real-world sources. Synthetic data is used to augment training sets, protect privacy, address class imbalance, or generate rare-case examples. It is increasingly used to train AI models where real data is scarce, expensive, or sensitive.

The core mechanism in transformer models that allows every position in a sequence to attend to every other position when computing its representation. Self-attention enables transformers to capture long-range dependencies that earlier architectures missed. It is computationally expensive — scaling quadratically with sequence length — which is why extending context windows is technically challenging.

A critical term coined by AI researchers Emily Bender and colleagues in a 2021 paper, arguing that large language models 'stochastically parrot' statistically likely word sequences without true understanding. The term sparked significant debate about whether LLMs 'understand' language or merely pattern-match. Understanding this debate is useful context for professionals assessing LLM reliability.

A neural network architecture in which only a subset of parameters are active for any given input — as opposed to dense models where all parameters participate in every computation. Mixture of Experts (MoE) models are the most prominent form of sparse models. Sparsity enables large model capacity without proportionally increasing inference cost.

One of the most widely used AI capabilities in professional contexts — condensing long documents, meeting transcripts, research papers, or email threads into concise summaries. LLMs perform extractive summarization (selecting key passages) and abstractive summarization (rephrasing and synthesising). Summary accuracy should always be verified for high-stakes use cases.

A parameter that controls the randomness of an LLM's outputs. Low temperature (near 0) makes responses more deterministic and focused; high temperature (near 1 or above) increases variety and creativity. Should be lowered for factual or structured tasks and can be raised for creative writing.

The basic unit of text that LLMs process. Roughly 1 token ≈ 0.75 words in English. Models have token limits per request (context window) and are often priced per token.

The dataset used to train an AI model. The quality, scale, and diversity of training data heavily influence model capability and biases. Modern frontier models are trained on trillions of tokens drawn from the web, books, code, scientific papers, and other sources.

A machine learning technique where a model trained on one task is adapted for a related task. Foundation models are trained once at great cost, then transfer-learned (fine-tuned) cheaply for specific applications — the basis of modern commercial AI deployment.

The neural network architecture that underlies virtually all modern LLMs. Introduced in the 2017 paper "Attention Is All You Need". Processes entire input sequences in parallel using attention mechanisms, enabling training on far larger datasets than previous sequential architectures.

The process of breaking text into smaller units (tokens) that an LLM can process. Tokens can be words, subword units, or individual characters depending on the tokeniser. Most English words map to 1–2 tokens; context window limits are measured in tokens, not words — typically around 0.75 words per token.

The field within AI development focused on identifying, measuring, and mitigating harmful model behaviors including bias, toxicity, misinformation generation, and misuse. Trust and safety teams at AI labs run red-teaming exercises, evaluate outputs for policy violations, and design guardrails. It is closely related to AI alignment and AI ethics.

AI models that generate images from text descriptions. Leading text-to-image systems include DALL·E 3, Midjourney, Stable Diffusion, and Adobe Firefly. Text-to-image AI has practical applications in marketing, design prototyping, and content production — and raises significant questions about copyright and creative authorship.

The ability of AI models to invoke external tools — web search, code execution, calculators, file systems, APIs — during a conversation to complete tasks that require information or actions beyond the model's intrinsic capabilities. Tool use is a key enabler of agentic AI systems. It is sometimes called 'function calling' in API contexts.

The volume of requests or tokens an AI system can process per unit of time. Throughput is a key operational metric for production AI deployments alongside latency. High-throughput AI infrastructure requires careful engineering of batching, caching, and model serving — a concern primarily for engineering teams building AI-powered products.

Methods for estimating how confident an AI model is in its outputs — and communicating that confidence to users. LLMs do not natively produce calibrated confidence scores, making uncertainty quantification an active research area. In professional AI deployments, understanding model uncertainty is critical for deciding when to trust AI outputs and when to require human review.

A machine learning approach in which a model learns patterns from data without labelled examples. Unsupervised learning techniques include clustering, dimensionality reduction, and generative modeling. Pre-training large language models on unlabelled text is a form of unsupervised learning — the model learns language structure without explicit right-or-wrong labels.

A database optimised for storing and querying high-dimensional numerical vectors (embeddings). Used in RAG systems and semantic search to quickly find data points similar to a query. Examples include Pinecone, Weaviate, Qdrant, and pgvector (PostgreSQL extension).

Video Random Access Memory — the memory on a GPU used to store model weights and intermediate computations during inference. A model's VRAM requirement is the primary hardware constraint that determines what consumer or enterprise hardware can run it. Quantisation and model distillation are the primary techniques for reducing VRAM requirements.

A practice where developers describe what they want in natural language and let AI generate the code, iterating on the output rather than writing code directly. Popularised in 2025, vibe coding lowers the barrier to software creation but requires the practitioner to verify that AI-generated code is correct, secure, and maintainable.

A multimodal AI model that can process both images and text, enabling tasks like image captioning, visual question answering, and document analysis. GPT-4o and Claude 3 are examples of vision language models. VLMs are increasingly used in business applications that involve processing invoices, diagrams, photographs, and mixed-format documents.

The challenge of ensuring that an AI system's goals, behaviors, and outputs are consistent with human values — particularly the values of those affected by its decisions. Value alignment is a central problem in AI safety research. It is distinct from capability: a highly capable AI that pursues misaligned goals is considered more dangerous, not less.

The numerical parameters of a neural network that encode the knowledge learned during training. 'Model weights' and 'model parameters' are often used interchangeably. Releasing model weights publicly (open-weights models) allows others to run and fine-tune the model locally.

Techniques for embedding detectable patterns into AI-generated content to identify its origin. AI watermarking is an active area of research and policy development — proposed as one mechanism for distinguishing AI-generated text, images, and audio from human-created content. No widely deployed watermarking standard exists at the time of publication.

A neural network design technique in which multiple parts of a model share the same parameter values — reducing total parameter count without reducing network capacity for certain tasks. Weight sharing is used in convolutional neural networks and some efficient transformer variants. It is part of the broader toolkit for building smaller, more efficient models.

The use of AI to automate sequences of business tasks that previously required human intervention at each step. Modern AI workflow automation goes beyond rule-based automation — using LLMs to handle variable inputs, generate outputs, and make routing decisions. It is one of the highest-ROI applications of AI for operational teams.

A set of methods and techniques designed to make AI model decisions understandable to human users. XAI is particularly important in regulated industries — finance, healthcare, legal — where decision-makers must be able to justify AI-assisted outcomes. Key XAI approaches include feature attribution, attention visualisation, and counterfactual explanations.

A human-readable data serialisation format widely used for AI system configuration — defining model parameters, prompts, pipelines, and agent behaviors in a structured, version-controllable format. Many AI orchestration frameworks and LLMOps platforms use YAML for pipeline definitions. Familiarity with YAML is increasingly useful for professionals configuring AI workflows.

Asking an AI model to perform a task without providing any examples, relying entirely on the knowledge and instruction-following ability the model acquired during training. Often works well for simple, clearly defined tasks. Contrasts with few-shot prompting, which includes 2–5 examples.

The ability of an AI model to perform a task it has never been explicitly trained on, based solely on the task description. Zero-shot learning is a hallmark capability of large language models — enabling them to write code, translate languages, or analyse documents without any task-specific training examples. It is distinct from few-shot learning, which provides a small number of examples in the prompt.