AI Glossary

An AI technique that generates a new, condensed version of a text rather than simply extracting key sentences.

A metric that measures how often an AI model's predictions are correct, expressed as a percentage of total predictions.

A mathematical function inside a neural network that decides whether a neuron should fire, introducing the non-linearity that lets networks learn complex patterns.

A machine learning approach where the model selectively requests labels for the most informative data points rather than learning from a randomly labelled dataset.

A deliberate attempt to fool an AI model by crafting inputs specifically designed to cause incorrect or harmful outputs.

Inputs deliberately crafted to fool an AI model into making incorrect predictions, often by making tiny changes that are imperceptible to humans but catastrophic for the model.

The process of measuring how well an AI agent performs its intended tasks, including accuracy, reliability, efficiency, and safety.

Safety constraints and rules that limit what an AI agent can do, preventing it from taking harmful, unauthorised, or unintended actions.

The repeating cycle an AI agent follows: plan what to do, execute a step using tools, observe the result, decide whether to continue or adjust.

Mechanisms that allow AI agents to retain and recall information across interactions, enabling continuity and context-aware behaviour over time.

The coordination of multiple AI agents working together on complex tasks, managing their interactions, dependencies, and information flow.

Using AI agents to write, test, and debug code autonomously — planning multi-step development tasks and executing them with minimal human intervention.

A software library or platform for building AI agents — autonomous systems that can plan, use tools, and complete multi-step tasks without continuous human input.

A multi-step process where AI agents autonomously plan, execute, and refine work — going beyond simple prompt-response to handle complex tasks end-to-end.

A theoretical form of AI that could match or exceed human-level reasoning across every intellectual domain, rather than being limited to specific tasks.

Specialised hardware designed to perform AI computations faster and more efficiently than general-purpose processors.

An AI system that can take actions autonomously — browsing the web, running code, calling APIs, and completing multi-step tasks with minimal human intervention.

The structural design of an AI agent, including how its components — planning, memory, tool use, and decision-making — are organised and connected.

The challenge of ensuring AI systems do what humans actually want — not just what they were literally instructed to do. A core concern in AI safety research.

Techniques for storing and reusing AI model responses to reduce costs, improve latency, and decrease load — from exact match caching to semantic similarity caching.

A processor designed or optimised to efficiently run AI and machine learning workloads, including GPUs, TPUs, and purpose-built neural processors.

The practice of ensuring AI systems meet regulatory requirements, industry standards, and legal obligations, particularly around data protection, fairness, and transparency.

An AI assistant embedded directly into a software application that helps users complete tasks by suggesting, drafting, or automating actions within their existing workflow.

The practice of managing and reducing the costs of AI deployment through model selection, prompt engineering, caching, and infrastructure choices.

The trend of making AI tools, knowledge, and capabilities accessible to people and organisations without specialised technical expertise.

The study and practice of ensuring AI systems are developed and used in ways that are fair, transparent, safe, and respectful of human rights and values.

A governance body within an organisation responsible for reviewing AI projects and policies to ensure they align with ethical principles and societal values.

The policies, processes, and frameworks that guide how an organisation develops, deploys, and manages AI systems — covering risk, ethics, compliance, and accountability.

The maximum acceptable response time for an AI-powered feature, broken down across its components — model inference, retrieval, network, and processing — to guide performance optimisation.

The ability to understand, use, and evaluate AI tools effectively. Not coding — knowing what AI can do, when to use it, and how to get good results.

A framework that assesses an organisation's AI capabilities across defined stages, from initial experimentation to fully integrated AI-driven operations.

The practice of coordinating multiple AI models, agents, or services to work together on complex tasks — managing handoffs, shared context, error handling, and resource allocation.

The middleware that manages how AI models are selected, invoked, and coordinated within an application — handling routing, fallbacks, retries, and model switching.

An organisation's preparedness to adopt and benefit from AI — covering data quality, skills, processes, culture, and infrastructure.

The field of research and practice dedicated to ensuring AI systems behave as intended and do not cause unintended harm.

Standardised tests that evaluate how well an AI model handles sensitive topics, refuses harmful requests, and avoids generating dangerous, biased, or misleading content.

An organisation's plan for how it will adopt, deploy, and benefit from artificial intelligence, aligning AI investments with business objectives.

The security and reliability risks arising from an organisation's dependence on third-party AI models, data sources, APIs, and tooling that may be compromised or discontinued.

A structured approach for evaluating and ensuring the trustworthiness of AI systems across dimensions like accuracy, fairness, transparency, security, and privacy.

Techniques for embedding hidden, detectable signals in AI-generated text or images that allow the content to be identified as machine-generated, even after modification.

A period of reduced funding, interest, and progress in artificial intelligence research, typically following a cycle of inflated expectations and disappointing results.

An organisation that designs its processes, products, and culture around AI as a core capability rather than treating AI as an add-on to existing operations.

Search technology that uses AI models to understand the meaning of queries and generate direct answers rather than just matching keywords to documents.

A step-by-step set of instructions that tells a computer how to solve a problem or complete a task.

The process of adding labels or tags to raw data so AI models can learn from it during training.

A way for software to communicate with other software. APIs are how developers connect AI capabilities to websites, apps, and business tools.

The process of connecting AI services to your existing software systems using Application Programming Interfaces, enabling automated data exchange and AI-powered features.

Software that can perform tasks that normally require human intelligence, such as understanding language, recognising patterns, and making decisions.

A computing system inspired by the brain's structure, composed of layers of interconnected nodes that learn to recognise patterns by adjusting the strength of their connections.

The practical limit on how much information an AI model can effectively focus on within its context window, where performance degrades as the window fills up.

A technique that lets AI models focus on the most relevant parts of their input when generating each piece of output, forming the core of transformer architecture.

A neural network that learns to compress data into a smaller representation and then reconstruct it, useful for dimensionality reduction, denoising, and anomaly detection.

The practice of using systematic, automated tests to verify that AI systems produce correct, consistent, and safe outputs before and during deployment.

Using technology to perform tasks without manual human effort. AI automation goes beyond traditional rule-based automation by handling unstructured tasks like writing, analysis, and decision-making.

Automated Machine Learning — tools and techniques that automate the process of selecting, configuring, and training machine learning models.

AI systems that operate independently, making decisions and taking actions without continuous human oversight or intervention.

The training algorithm that teaches neural networks by calculating how much each weight contributed to errors and adjusting them to reduce mistakes.

Processing multiple AI predictions at once rather than one at a time, significantly reducing cost and improving throughput for non-time-sensitive workloads.

A method of processing multiple data items together as a group rather than one at a time, improving efficiency and reducing costs in AI workloads.

A text generation strategy that explores multiple possible continuations simultaneously and selects the sequence with the highest overall probability.

A standardised test or dataset used to measure and compare the performance of different AI models on specific tasks.

A landmark language model from Google that reads text in both directions simultaneously, dramatically improving how machines understand context and meaning in natural language.

The process of identifying systematic unfairness in AI systems that causes them to treat different groups of people differently based on protected characteristics.

Systematic errors in AI systems that produce unfair outcomes, typically arising from biased training data, flawed assumptions, or unrepresentative datasets.

A metric that evaluates the quality of machine-generated text by measuring how closely it matches one or more human reference texts.

The use of artificial intelligence to enhance business analytics — automating data analysis, generating insights, and enabling natural language querying of business data.

A problem in neural networks where training on new data causes the model to lose knowledge it previously learned, effectively overwriting old skills with new ones.

Data that represents distinct groups or categories — like colours, countries, or product types — rather than continuous numerical values.

Statistical and computational methods for determining whether one event actually causes another, rather than merely being correlated with it.

A prompting technique that instructs an AI model to work through a problem step by step, showing its reasoning before giving a final answer.

A technique where you ask the AI to explain its reasoning step by step before giving a final answer. This dramatically improves accuracy on complex tasks.

A prompting technique that instructs an AI model to show its step-by-step reasoning process before arriving at a final answer.

The structured approach to transitioning teams and organisations from traditional workflows to AI-augmented ways of working, addressing resistance and building adoption.

A software application that simulates conversation with users, ranging from simple rule-based systems to sophisticated AI-powered assistants.

An AI chatbot built by OpenAI that uses large language models to engage in conversation, answer questions, write content, and assist with a wide range of tasks.

Research findings from DeepMind showing that AI models perform best when training data and model size are scaled proportionally, rather than simply making models as large as possible.

An AI task that assigns input to predefined categories. Spam detection, sentiment analysis, and image recognition are all classification tasks.

An AI model that assigns input data to predefined categories, such as sorting emails into spam or not spam, or identifying the subject of an image.

An AI assistant built by Anthropic, designed with a focus on safety, helpfulness, and honesty, capable of conversation, analysis, writing, and code.

Using remote servers and services — from providers like AWS, Azure, and Google Cloud — to train, deploy, and run AI models without owning the hardware.

An unsupervised machine learning technique that groups similar data points together without predefined labels, revealing natural patterns in data.

The field of AI that enables machines to interpret and understand visual information from images and videos, including object recognition, scene understanding, and visual analysis.

A table that shows the breakdown of a classification model's correct and incorrect predictions, revealing exactly where and how the model makes mistakes.

An AI safety technique developed by Anthropic where the model is trained to follow a set of principles (a 'constitution') to self-correct harmful or unhelpful outputs.

The use of AI to create written, visual, audio, or video content, from marketing copy and blog posts to images, music, and synthetic media.

The practice of carefully designing what information an AI receives — including system prompts, retrieved documents, conversation history, and tool outputs — to maximise the quality of its responses.

The discipline of designing and managing all information provided to an AI model — system prompts, retrieved documents, conversation history, and tool outputs — to maximise response quality.

The maximum amount of text an AI can process at once. Think of it as the AI's working memory — everything it can see and consider when generating a response.

A training approach where a model learns to represent data by distinguishing similar examples from dissimilar ones.

AI systems designed to engage in natural, multi-turn dialogue with humans, understanding context, intent, and nuance across a conversation.

A large, structured collection of text used to train or evaluate AI language models, ranging from curated datasets to vast web crawls containing billions of documents.

A mathematical measure of how similar two vectors are by calculating the cosine of the angle between them, widely used in AI to compare documents, images, and search queries.

An open-source Python framework for building multi-agent AI systems using a crew metaphor — agents with defined roles collaborate on tasks like a team of specialists.

A statistical technique for evaluating AI models by splitting data into multiple training and testing subsets to get a reliable estimate of real-world performance.

A training strategy where a model learns from easy examples first and progressively encounters more difficult ones, mirroring how humans learn.

A training strategy where an AI model learns from simple examples first and progressively encounters more complex ones, mimicking how humans learn through structured progression.

A personalised version of ChatGPT configured with specific instructions, knowledge, and capabilities for a particular task or domain, created without coding.

The process of adding labels, tags, or other metadata to raw data so that AI models can learn from it — the essential human step that makes supervised learning possible.

Techniques for artificially expanding a training dataset by creating modified versions of existing data, improving model performance without collecting new data.

A gradual change in the statistical properties of the data an AI model encounters in production, compared to the data it was trained on.

A self-reinforcing cycle where a product generates user data that improves the AI model, which improves the product, which attracts more users and more data.

The policies, processes, and standards an organisation uses to manage its data assets, ensuring data quality, security, privacy, and compliance.

The process of tagging raw data with informative labels so supervised machine learning models can learn the relationship between inputs and desired outputs.

A critical error in AI development where information from outside the training set improperly influences the model, producing unrealistically good results that fail in production.

An automated sequence of steps that collects, processes, transforms, and delivers data from source systems to AI models or analytics tools.

A type of attack where an adversary deliberately corrupts training data to make an AI model learn incorrect patterns or behaviours.

The protection of personal and sensitive information when using AI systems, encompassing what data is collected, how it is processed, and who can access it.

The use of AI to make or recommend decisions that were previously made by humans, from loan approvals and fraud alerts to inventory restocking and content recommendations.

A machine learning model that makes predictions by following a series of if-then rules, splitting data into branches like a flowchart until reaching a decision.

A transformer architecture that generates text by predicting one token at a time from left to right, used by GPT, Claude, and most modern large language models.

A subset of machine learning that uses neural networks with many layers to learn complex patterns. The 'deep' refers to the number of layers, not the depth of understanding.

A training approach that combines deep neural networks with reinforcement learning, enabling AI to learn complex strategies through trial and error in rich environments.

A training technique where a model learns to reconstruct clean data from corrupted or noisy inputs, used in both language and image AI models.

A type of generative AI that creates images, audio, or video by learning to gradually remove noise from random static, producing remarkably realistic outputs.

The process of using digital technology — including AI — to fundamentally change how a business operates, delivers value, and competes. Not just adopting new tools but rethinking processes.

A virtual replica of a physical object, process, or system that uses real-time data and AI to simulate, monitor, and optimise its real-world counterpart.

Techniques that reduce the number of features in a dataset while preserving the most important patterns, making data easier to visualise, process, and model.

A technique for training a smaller, faster AI model to mimic the behaviour of a larger, more capable model, preserving most of the performance at a fraction of the cost.

AI technology that reads, understands, and extracts information from documents such as invoices, contracts, forms, and reports, automating manual data entry.

A simpler alternative to RLHF that trains AI models to align with human preferences directly from comparison data, without needing a separate reward model.

A regularisation technique where randomly selected neurons are temporarily deactivated during training, forcing the network to develop more robust and generalisable features.

A training technique where model training is halted before completion when performance on validation data stops improving, preventing the model from overfitting to training data.

Running AI models directly on local devices — phones, cameras, sensors, factory equipment — rather than sending data to cloud servers for processing.

A numerical representation of text (or images, audio, etc.) that captures its meaning. Embeddings let AI measure how similar two pieces of content are.

Capabilities that appear in large AI models that were not explicitly trained for and were not present in smaller versions of the same model.

A neural network architecture where an encoder compresses input into a representation and a decoder generates output from that representation, used in translation, summarisation, and generation tasks.

A transformer architecture that processes input text bidirectionally to produce rich representations, used for classification, search, and understanding tasks rather than text generation.

Techniques that combine predictions from multiple AI models to produce more accurate and reliable results than any single model could achieve alone.

The deployment of artificial intelligence across an organisation's operations at scale, integrated into core business processes with appropriate governance and security.

An NLP technique that automatically identifies and classifies named entities — people, organisations, locations, dates, amounts — from unstructured text.

One complete pass through the entire training dataset during model training — most models require many epochs to learn effectively.

The practice of designing, developing, and deploying AI systems that are fair, transparent, accountable, and aligned with human values and societal well-being.

Systematic tests used to measure AI model performance across specific capabilities, safety criteria, or business requirements — the AI equivalent of a software test suite.

A standardised testing framework used to systematically measure and compare AI model performance across a consistent set of benchmarks and tasks.

The ability of an AI system to describe its reasoning in terms that humans can understand, enabling trust, debugging, and regulatory compliance.

A summarization technique that creates a summary by selecting and combining the most important sentences directly from the original text.

The process of selecting, transforming, and creating input variables from raw data to help machine learning models learn more effectively.

A centralised repository for storing, managing, and serving the pre-computed data features used to train and run AI models, ensuring consistency between training and production.

A training technique where multiple devices or organisations collaboratively train a shared AI model without sharing their raw data, preserving privacy.

An AI technique where a model learns to perform a new task from just a handful of examples, rather than the thousands typically required for training.

A prompting technique where you provide the AI with a few examples of the desired input-output pattern before giving it your actual task. This dramatically improves output consistency.

Training an existing AI model on your specific data to improve its performance on your specific tasks. Like giving the AI specialised on-the-job training.

An optimised algorithm for computing the attention mechanism in transformers that dramatically reduces memory usage and speeds up processing.

A large AI model trained on broad data at scale that can be adapted to a wide range of downstream tasks — GPT, Claude, and Llama are all foundation models.

A capability that allows AI models to invoke external functions, APIs, and tools by generating structured requests, enabling them to take actions and access real-time information.

A generative AI architecture where two neural networks compete — one generates fake data, the other tries to detect it — pushing both to improve until the fakes are indistinguishable from real data.

Google's family of AI models and assistant, designed to be natively multimodal — processing text, images, audio, video, and code within a single unified model.

AI that creates new content — text, images, code, audio, video — rather than just analysing or classifying existing data.

A file format for storing quantised AI models designed for efficient local execution, widely used by tools like llama.cpp to run large language models on consumer hardware.

A git feature that creates an additional working directory linked to the same repository, allowing multiple branches to be checked out simultaneously — essential for parallel agent development.

A family of large language models created by OpenAI that generate human-like text by predicting the next word, forming the foundation of ChatGPT.

A specialised processor originally designed for rendering graphics but now essential for training and running AI models. GPUs can perform thousands of calculations simultaneously.

The optimisation algorithm that trains neural networks by iteratively adjusting model parameters in the direction that reduces prediction errors.

A type of neural network designed to operate on graph-structured data, where entities are connected by relationships, such as social networks, molecules, or knowledge graphs.

The verified, correct answer or label for a data point, used as the standard against which AI model predictions are measured.

The practice of connecting AI model outputs to verifiable sources of truth, such as documents, databases, or real-time data, to reduce hallucinations and improve accuracy.

Methods for connecting AI model outputs to verifiable sources of truth — documents, databases, or real-time data — reducing hallucinations and improving factual reliability.

Constraints, rules, and safety mechanisms built into AI systems to prevent harmful, incorrect, or out-of-scope outputs and actions.

When AI generates confident but incorrect information. The AI is not lying — it is producing statistically plausible text that happens to be wrong.

Techniques and strategies for reducing the tendency of AI language models to generate plausible-sounding but factually incorrect information.

The leading open-source platform for sharing, discovering, and deploying AI models, datasets, and applications — often described as the GitHub of machine learning.

A system design where AI handles execution but a human reviews, approves, or intervenes at critical decision points before actions are taken.

A configuration setting chosen before training begins — like learning rate, batch size, or number of layers — that controls how a model learns rather than what it learns.

The process of finding the optimal configuration settings for an AI model — the knobs you set before training begins that determine how the model learns.

AI's ability to identify and categorise objects, people, scenes, and other elements within images, powering applications from photo organisation to medical diagnosis.

The ability of large language models to learn new tasks from examples provided directly in the prompt, without any changes to the model's weights.

The process of an AI model generating output from your input. Every time you send a prompt and get a response, that is inference.

Techniques for making AI models faster, cheaper, and more efficient when generating predictions or responses in production.

A curated collection of prompt-response pairs used to train or fine-tune AI models to follow instructions accurately and produce helpful, well-structured outputs.

An AI model's ability to accurately interpret and execute natural language instructions provided by the user.

A fine-tuning technique that trains a language model to follow human instructions by exposing it to thousands of example instruction-response pairs.

The combination of AI with traditional automation technologies like RPA to automate complex business processes that require judgement, not just rule-following.

Techniques used to trick an AI model into bypassing its safety guardrails and producing outputs it was designed to refuse, such as harmful instructions or policy-violating content.

A setting that forces an AI model to return its response as valid JSON rather than free-form text.

The date after which an AI model has no information, determined by when its training data collection ended.

A technique for training a smaller, faster AI model to replicate the behaviour of a larger, more capable model.

A structured database that maps relationships between concepts, entities, and facts — helping AI understand how things connect rather than just matching keywords.

A memory optimisation technique that stores previously computed attention values during text generation, avoiding redundant calculations and significantly speeding up AI model inference.

The known correct answer attached to a training example in supervised learning, which the AI model learns to predict — such as 'spam' for an email or 'cat' for an image.

A framework from LangChain for building stateful, multi-step AI agent workflows as directed graphs — giving developers precise control over how agents plan, act, and recover from errors.

An AI model designed not just to generate text but to take actions in software — clicking buttons, filling forms, and navigating interfaces on your behalf.

A type of AI trained on vast amounts of text to understand and generate human language. ChatGPT, Claude, and Gemini are all LLMs.

The time delay between sending a request to an AI model and receiving the first part of the response. Lower latency means faster, more responsive AI interactions.

The compressed, abstract internal representation that an AI model creates to capture the essential features of its training data in a lower-dimensional form.

A hyperparameter that controls how much an AI model adjusts its internal weights in response to each error during training — too high causes instability, too low causes slow learning.

The set of practices, tools, and processes for deploying, monitoring, and maintaining large language model applications in production — an evolution of MLOps for the generative AI era.

AI models designed to process very large inputs — hundreds of thousands or millions of tokens — enabling analysis of entire books, codebases, or document collections in a single prompt.

The challenge of handling rare, unusual, or edge-case inputs that AI models encounter infrequently but must still handle correctly in production.

A fine-tuning technique that trains only a small number of additional parameters instead of updating the entire model, making customisation faster and cheaper.

A mathematical function that measures how far a model's predictions are from the correct answers, providing the error signal that drives learning during training.

AI platforms and tools that let you build AI-powered applications with minimal programming, using visual interfaces and pre-built components.

A type of AI where systems learn patterns from data instead of following explicitly programmed rules. The system improves its performance through experience.

The set of practices and tools for deploying, monitoring, and maintaining machine learning models in production reliably and at scale.

AI-powered translation of text or speech from one language to another, as seen in tools like Google Translate and DeepL.

Techniques for removing the influence of specific training data from a trained AI model without retraining from scratch.

A technique where you ask the AI to write or improve prompts — using the AI itself to generate better instructions for the AI, creating a self-improving feedback loop.

An architecture where multiple AI agents with different specialisations collaborate on a task, with a routing system that directs each sub-task to the most appropriate agent.

A model architecture where only a subset of the model's parameters activate for each input, making large models faster and cheaper to run.

The process of training an AI model to behave in accordance with human values, intentions, and safety requirements.

A standardised document that describes an AI model's capabilities, limitations, training data, intended use, and known biases.

A phenomenon where AI models trained on AI-generated content gradually lose quality and diversity, producing increasingly bland and repetitive output over generations.

A set of techniques for reducing the size and computational requirements of AI models while preserving as much of their capability as possible.

An open standard that allows AI models to connect to external tools, data sources, and services through a unified interface. MCP lets AI access real-time data and take actions beyond text generation.

A standardised open protocol that allows AI models to connect to external tools, data sources, and APIs through a universal interface.

A technique where a smaller 'student' model is trained to replicate the behaviour of a larger 'teacher' model, producing a compact model that retains most of the original's capability.

The gradual decline in an AI model's performance over time as the real-world data it encounters changes from the data it was trained on.

The systematic process of measuring an AI model's performance using held-out data and appropriate metrics to determine whether it is good enough for its intended use.

The process of further training a pre-trained AI model on your own data so it performs better on your specific tasks.

A curated platform offering access to multiple pre-trained AI models from different providers, allowing users to compare and deploy models without managing infrastructure.

An online platform where developers share, discover, and download pre-trained AI models, like an app store for AI.

The ability to understand and explain how an AI model arrives at its predictions or outputs.

A technique for training or running AI models that are too large for a single GPU by splitting the model across multiple GPUs.

An internal system that tracks all AI models an organisation uses, including versions, performance metrics, and deployment status.

The infrastructure and process of making a trained AI model available to receive requests and return predictions in real time.

The numerical values inside a neural network that determine how it processes information. Weights are what the model learns during training — they encode its knowledge and capabilities.

An AI architecture where multiple specialised agents collaborate to complete complex tasks, coordinated by an orchestrator or a defined protocol.

A mechanism in transformer models that runs multiple attention operations in parallel, allowing the model to focus on different types of relationships between words simultaneously.

The ability of an AI model to answer questions that require combining information from multiple separate pieces of evidence, chaining logical steps together.

AI that can process and generate multiple types of content — text, images, audio, and video — within a single model. Claude, GPT-5.4, and Gemini are all multi-modal.

An interaction with an AI model spanning multiple exchanges where the model maintains context from previous messages throughout the conversation.

AI systems that can process and generate multiple types of content — text, images, audio, video — rather than just text alone.

AI training that combines multiple types of data — text, images, audio, and video — so the model can understand and generate across formats.

An NLP task that identifies and classifies proper nouns and specific terms in text into predefined categories like person, organisation, location, and date.

The branch of AI focused on enabling computers to understand, interpret, and generate human language in useful ways.

The branch of AI focused on enabling machines to comprehend the meaning and intent behind human language, not just process its surface form.

Explicit rules in a prompt that tell an AI model what NOT to do, preventing unwanted behaviours and outputs.

An automated method for discovering the best neural network design for a given task, replacing manual architecture engineering with AI-driven exploration.

A deep learning approach to translation that processes entire sentences as units of meaning, producing far more natural and accurate translations than older word-by-word methods.

A computing system loosely inspired by the human brain, made of layers of interconnected nodes that learn to recognise patterns in data.

The empirical observation that AI model performance improves predictably as model size, training data, and compute resources increase.

AI platforms that let you build AI-powered applications entirely through visual interfaces with zero programming required.

Random errors, irrelevant information, or inconsistencies in a dataset that can mislead AI models and reduce their performance.

The process of scaling numerical data to a standard range or distribution, ensuring that no single feature dominates model training simply because of its scale.

A computer vision task that identifies and locates specific objects within images or video frames, drawing bounding boxes around each detected object.

An open-source tool that makes running large language models on your own computer as simple as a single command, removing the need for cloud APIs or technical expertise.

AI models that run directly on your phone, laptop, or other hardware rather than in the cloud, offering faster responses and greater privacy.

AI models whose trained parameter weights are publicly released, allowing anyone to download, run, and often modify the model.

AI models and tools whose code and weights are publicly available, allowing anyone to use, modify, and deploy them freely.

AI models whose trained parameters are publicly released for anyone to download and use, distinct from truly open-source models where training data and code are also shared.

Technology that converts images of text — scanned documents, photos, PDFs — into machine-readable text that software can search and process.

A manager agent that coordinates multiple specialist agents by planning tasks, delegating work, reviewing results, and delivering the final output.

The process of extracting structured data from an AI model's text response so it can be used by other software systems.

The process of checking AI-generated content for accuracy, quality, and safety before using it in production or decision-making.

When an AI model performs excellently on training data but poorly on new data because it has memorised specific examples rather than learning general patterns.

A value inside a trained AI model — such as a weight or bias — that was learned from data during training and determines the model's behaviour.

A family of techniques that adapt large AI models to specific tasks by training only a small fraction of the model's parameters, dramatically reducing the cost and hardware requirements.

The total number of adjustable values in an AI model. A model with more parameters can capture more complex patterns but requires more computing power to train and run.

The simplest possible neural network — a single artificial neuron that takes inputs, applies weights, and produces a binary output, forming the conceptual building block of all modern deep learning.

A measurement of how well a language model predicts text — lower perplexity means the model is less surprised by the data and generally more capable.

The coordination of multiple AI processing steps into a reliable, automated sequence — managing data flow, error handling, and dependencies between stages.

An AI agent whose sole purpose is to ask structured questions and produce a detailed requirements document or project plan, without executing any tasks.

The ability of an AI agent to decompose a high-level goal into an ordered sequence of sub-tasks and determine how to execute them.

A reinforcement learning technique where the AI directly learns the best action to take in each situation by adjusting its decision-making policy based on rewards.

The tendency of AI models to pay more attention to information at certain positions in the input — typically the beginning and end — while underweighting information in the middle.

A technique that gives transformer models information about word order, since the attention mechanism alone does not know which words come first.

Any techniques applied to an AI model after its initial training to improve behaviour, including fine-tuning, RLHF, and safety alignment.

A specification document that defines what a product or feature should do, who it is for, and how success will be measured — increasingly generated with AI assistance.

The initial, large-scale training phase where a foundation model learns general knowledge from vast amounts of data before being specialised for specific tasks.

Two complementary metrics for classification models — precision measures how many predicted positives were correct, recall measures how many actual positives were found.

A structured file that tracks project status, decisions, and context between AI sessions, serving as session-to-session memory for AI-assisted work.

A feature that reuses previously processed prompt content across API calls, reducing latency and cost when the same system prompt or context is sent repeatedly.

A technique where the output of one AI prompt becomes the input to the next, creating a multi-step workflow that produces better results than a single complex prompt.

Design patterns for breaking complex AI tasks into sequences of simpler prompts, where each step's output feeds into the next, producing more reliable and controllable results.

Techniques for reducing the token count of prompts while preserving their essential meaning, enabling cost savings and fitting more useful context into limited context windows.

The skill of writing instructions to AI that consistently produce useful, accurate, high-quality output.

A security vulnerability where malicious text in user input or external data tricks an AI system into ignoring its original instructions and following the attacker's instructions instead.

A vulnerability where users manipulate an AI system into revealing its hidden system prompt or confidential instructions.

A reusable prompt structure with placeholder variables that gets filled in with specific data for each use, ensuring consistent AI output.

Automated checkpoints between AI generation and human review that catch specific types of errors — format, factual, tone, completeness, and consistency.

A technique that reduces the precision of an AI model's numerical weights to make it smaller, faster, and cheaper to run.

A training technique that prepares a model for quantization by simulating lower-precision arithmetic during training, resulting in better quality when the model is later compressed.

An AI task where the model reads a passage or knowledge base and directly answers questions about it, often with a specific cited source.

A machine learning algorithm that builds many decision trees and combines their predictions to produce more accurate and reliable results.

Controls imposed by AI providers that restrict how many requests you can make per minute or per day, preventing overuse and ensuring fair access.

An AI agent design pattern where the model alternates between reasoning about what to do next and taking an action, creating an observable thought-action loop.

A prompting pattern that combines Reasoning and Acting — the AI thinks through the problem step by step, takes an action (like a search), observes the result, and repeats until the task is complete.

AI systems that process input and produce output fast enough to support live interactions — voice conversations, live video analysis, or instant recommendations.

An AI model specifically designed to think step-by-step before responding, producing more accurate answers on complex problems that require logic and analysis.

The visible step-by-step thought process that an AI model shows when working through a problem before delivering its final answer.

The visible step-by-step thinking process that AI models produce when solving complex problems, allowing users to follow and verify the model's logic.

An AI system that predicts what content, products, or actions a user is most likely to want based on their behaviour and preferences.

A type of neural network designed to process sequential data by maintaining a memory of previous inputs, once widely used for text and time-series tasks.

Systematically testing an AI system by trying to make it fail, produce harmful output, or violate its guidelines — to find and fix vulnerabilities before users do.

An AI task that predicts a numerical value based on input data. Sales forecasting, price estimation, and demand prediction are all regression tasks.

A set of techniques that prevent AI models from memorising training data too closely, helping them perform better on new, unseen data.

A machine learning approach where an AI learns by trial and error, receiving rewards for good outcomes and penalties for bad ones. Used to train game-playing AI and to fine-tune LLMs.

A training technique where human evaluators rate AI outputs and reinforcement learning optimises the model to produce responses that humans prefer.

The practice of developing and deploying AI in ways that are ethical, transparent, accountable, and aligned with societal values — translating AI ethics principles into operational reality.

The concrete operational practices — bias testing, impact assessment, monitoring, and governance — that translate abstract AI ethics principles into everyday business decisions.

The process of finding and fetching relevant information from a knowledge base or database to provide context for an AI model's response.

A training approach that combines fine-tuning with retrieval capabilities, teaching the model both to retrieve relevant information and to generate better responses from retrieved context.

A technique that connects AI to your own documents and data so it can answer questions using your specific information, not just its general training.

When an AI system finds an unintended shortcut to maximise its reward signal without actually achieving the desired goal — optimising the metric rather than the objective.

An AI model trained to predict human preferences, used to guide the training of language models toward producing outputs that humans rate as helpful and safe.

The application of artificial intelligence to physical robots, enabling them to perceive their environment, make decisions, and perform tasks autonomously.

A technique where you assign the AI a specific persona or role (e.g., 'You are a senior data analyst') to shape the style, depth, and perspective of its responses.

A set of metrics for evaluating the quality of AI-generated summaries by measuring how much they overlap with human-written reference summaries.

Automated checks that screen AI inputs and outputs for harmful, illegal, or inappropriate content, blocking or flagging problematic requests and responses.

The process of training AI models to refuse harmful requests, avoid generating dangerous content, and behave within defined safety boundaries.

The process by which an AI model selects its next output token from a probability distribution, with settings like temperature controlling how random the selection is.

The empirical observation that AI model performance improves predictably as you increase model size, training data, and compute — following mathematical power laws.

Structured data code added to web pages that helps search engines and AI systems understand the content's meaning, enabling rich search results.

An approach where an AI model searches the web or a database for current information before generating a response, combining retrieval with generation.

The mechanism inside transformers that allows each word to consider every other word in the input when determining its meaning and importance.

A prompting technique where you ask the AI to generate multiple independent answers to the same question, then select the most common response — improving accuracy through consensus rather than relying on a single output.

A training approach where an AI model generates its own labels from raw data, typically by learning to predict missing or future parts of the input.

A technique for splitting documents into meaningful segments based on topic or meaning rather than arbitrary character counts, improving the quality of AI retrieval and analysis.

A system that classifies user queries by meaning and routes them to the appropriate handler — model, prompt, tool, or agent — without requiring keyword matching or rigid rules.

Search that finds results based on meaning and intent rather than exact keyword matches. Powered by vector embeddings that represent concepts as numbers.

A measure of how close in meaning two pieces of text are, regardless of whether they share the same words.

An AI technique that determines whether text expresses a positive, negative, or neutral opinion, widely used for analysing customer feedback and social media.

A model architecture that converts one sequence of data into another, originally designed for machine translation and now underlying many AI text tasks.

Cloud deployment where AI models run on-demand without you managing servers — you pay only for actual usage, and infrastructure scales automatically.

A testing strategy where a new AI model runs alongside the existing one in production, processing the same requests but without its responses reaching users.

A language model with fewer parameters (typically under 10 billion) that trades some capability for dramatically lower cost, faster speed, and the ability to run on smaller hardware.

A learned sequence of continuous vectors prepended to an AI model's input that steers its behaviour, as an alternative to traditional text prompts.

A technique where learnable continuous vectors — rather than natural language words — are prepended to a model's input, steering its behaviour without modifying the model's weights.

An AI model where only a fraction of parameters activate for each input, making it more efficient than dense models that use all parameters every time.

An inference acceleration technique where a smaller, faster model drafts text that a larger model then verifies, significantly speeding up generation without sacrificing quality.

AI technology that converts spoken language into written text, enabling voice assistants, transcription services, and voice-controlled applications.

AI technology that converts spoken language into written text, enabling transcription, voice commands, and accessibility features.

The way an AI agent tracks and maintains information across multiple steps in a task — remembering what it has done, what it knows, and what it still needs to do.

A problem-solving approach where an AI model breaks down complex tasks into sequential logical steps, improving accuracy on multi-step problems.

A method where the AI sends its response word-by-word as it generates, rather than waiting until the full response is complete before showing anything.

Data organised in a predefined format with clear rows and columns, such as spreadsheets and databases, making it easy for machines to search and analyse.

Techniques for constraining AI model outputs to follow specific formats — JSON, XML, or custom schemas — ensuring responses can be reliably parsed and processed by downstream systems.

A feature that constrains AI to respond in a specific structured format — typically JSON — making the output reliably parseable by code instead of requiring human interpretation.

An AI agent spawned by another agent to handle a specific sub-task, operating with its own context window and returning results to the parent.

The AI task of condensing long text into a shorter version that captures the key points, saving readers time while preserving essential information.

A machine learning approach where the model learns from labelled examples — input data paired with correct answers. The most common type of machine learning in business applications.

An AI approach inspired by the collective behaviour of social organisms like ants and bees, where simple agents following local rules produce complex intelligent behaviour.

Data generated by AI rather than collected from real-world sources. Used for training AI models, testing systems, and filling gaps where real data is expensive, sensitive, or unavailable.

The practice of using AI-generated test data and AI judges to evaluate AI model performance, enabling faster and more scalable testing than human evaluation alone.

Artificially generated test cases used to evaluate AI model performance, enabling systematic testing across scenarios that may be rare or difficult to collect from real-world data.

A set of persistent instructions given to an AI that shapes its behaviour for an entire conversation. System prompts define the AI's role, tone, rules, and output format.

Data organised in rows and columns like a spreadsheet or database table, representing the most common format for business data and analytics.

The practice of breaking a complex problem into smaller, manageable subtasks that an AI can handle individually before combining the results.

The accumulated cost of shortcuts, outdated models, and ad hoc solutions in AI systems that make future development slower, more expensive, and more error-prone.

A setting that controls how creative or conservative AI output is. Low temperature = predictable and focused. High temperature = varied and creative.

A setting that controls how random or creative an AI model's responses are — low temperature gives focused, predictable output while high temperature gives varied, creative output.

A parameter that controls the randomness of AI model outputs — lower temperatures produce more predictable responses while higher temperatures increase creativity and variety.

A software development practice where you write automated tests before writing the code that makes them pass, increasingly used to guide AI code generation.

Additional computation spent during inference — such as generating and evaluating multiple responses — to improve the quality of an AI model's output.

An AI task that assigns predefined categories or labels to text, used for spam detection, sentiment analysis, topic tagging, and content moderation.

The process of converting text into numerical vectors that capture its meaning, enabling machines to measure similarity and perform semantic search.

The AI capability of producing human-like written text, from drafting emails and articles to writing code and creating creative content.

AI technology that generates images from written descriptions, turning prompts like 'a sunset over mountains in watercolour style' into actual images.

AI technology that converts written text into natural-sounding spoken audio, enabling voice interfaces, audiobooks, and accessibility features.

The volume of data an AI system can process in a given time period — typically measured in tokens per second or requests per minute. Higher throughput means more work done faster.

The smallest unit of text an AI model processes. Roughly 3-4 characters or three-quarters of a word. AI pricing is typically measured in tokens.

The maximum number of tokens allocated for an AI interaction, encompassing both the input (prompt and context) and the output (generated response).

The process of breaking text into smaller units called tokens that an AI model can process, forming the fundamental input representation for language models.

The component that converts text into tokens — the numerical units an AI model processes. Different models use different tokenisers, which affects how they handle text.

A capability that lets AI models request the execution of external functions — searching databases, calling APIs, or performing calculations — extending their abilities beyond text generation.

The ability of an AI model to call external tools, APIs, and services to perform actions or retrieve information beyond its built-in capabilities.

The ability of an AI model to interact with external tools — search engines, code interpreters, APIs, databases — to take actions beyond generating text.

An approach where AI models can call external tools — calculators, APIs, databases, code interpreters — to perform actions and retrieve information beyond their built-in capabilities.

A text generation strategy where the AI model only considers the K most probable next words at each step, balancing creativity and coherence by limiting the candidate pool.

A method for controlling AI output randomness by only considering the smallest set of tokens whose combined probability exceeds a threshold p.

A custom AI chip designed by Google specifically for machine learning workloads. TPUs are optimised for the tensor operations that neural networks rely on.

The dataset used to teach an AI model. The quality, size, and composition of training data directly determines what the AI can and cannot do well.

A technique where a model trained on one task is reused as the starting point for a different but related task, dramatically reducing the data and time needed.

The neural network architecture behind modern AI assistants like ChatGPT and Claude. Introduced in 2017, it processes all words simultaneously using an attention mechanism.

A prompting technique where the AI explores multiple reasoning paths in parallel, evaluates each branch, and selects the most promising solution — mimicking how humans consider alternatives before deciding.

AI systems designed and operated so that users, organisations, and society can rely on them to be accurate, fair, secure, and transparent.

A test proposed by Alan Turing in 1950 where a human evaluator tries to distinguish between a machine and a human based on conversation alone.

When an AI model is too simple to capture the patterns in the data, resulting in poor performance on both training data and new data.

Data that does not follow a predefined format — emails, documents, images, videos, and conversations — which AI can now analyse and extract value from.

A machine learning approach where the model finds patterns in data without being given correct answers. Used for discovering hidden structure, grouping similar items, and detecting anomalies.

A technique that increases the representation of a minority class in a dataset to help AI models learn balanced patterns and avoid bias toward the majority class.

A portion of data held back from training and used to evaluate an AI model's performance during development, helping prevent overfitting.

A generative AI architecture that learns to compress data into a compact representation and generate new, similar data from that representation.

A specialised database designed to store and search embeddings — the numerical representations of text, images, or other data used in AI applications.

A specialised data structure that enables fast similarity searches across millions or billions of vectors, powering the retrieval behind semantic search and RAG systems.

A search method that finds results based on meaning rather than keywords by comparing the mathematical representations (vectors) of queries and documents.

A software development approach where you describe what you want in plain English and an AI model writes the code for you.

An architecture that applies the transformer model — originally designed for text — to image recognition, by splitting images into patches and processing them like words in a sentence.

An AI model that can process both images and text together, enabling it to answer questions about images, describe visual content, and reason across both modalities.

Techniques for embedding invisible markers in AI-generated content that allow detection of whether text, images, or audio were created by AI.

A regularisation technique that prevents AI model weights from growing too large during training, encouraging simpler models that generalise better to new data.

The method used to set the initial values of a neural network's parameters before training begins, which significantly affects how well and how quickly the model learns.

A technique that represents words as numerical vectors in a multi-dimensional space, where similar words are positioned close together.

A pioneering technique that represents words as numerical vectors, where similar words have similar vectors, enabling AI to understand relationships between words mathematically.

A sequence of connected steps that accomplish a specific business task. In AI context, a workflow combines human actions and AI processing to complete work efficiently.

Using AI to automate sequences of business tasks — from receiving a trigger to completing multi-step processes — with minimal human intervention.

An AI model's ability to perform a task it was not explicitly trained on, using general knowledge to handle new categories or instructions without any examples.

Asking an AI to perform a task without providing any examples — relying entirely on the model's training and your instructions. The simplest prompting approach.