AI Glossary — every concept, explained clearly

1067 AI and machine-learning concepts with TL;DR, visual diagrams, math, and worked examples at three difficulty levels.

• A* search — A* search is a pathfinding algorithm that finds the shortest path between two points in a graph.
• A/B testing — A statistical method comparing two versions (A and B) of something to see which performs better.
• Abductive logic programming (ALP) — Abductive logic programming (ALP) is a framework for solving problems by finding the simplest explanations for observations.
• Abductive reasoning — Abductive reasoning infers the most likely explanation for a set of observations, unlike deductive reasoning which proves.
• Ablation — Ablation involves removing a component of an AI system to study its impact on overall performance.
• Ablation Study — An experiment that removes one component of a system at a time to measure its individual contribution.
• Abstract data type — An abstract data type (ADT) defines data behavior by specifying possible values and operations, independent of implementation.
• Abstraction — Abstraction is the process of simplifying complex systems by focusing on essential features and ignoring details.
• Accelerating change — Accelerating change refers to the perceived historical trend of technology advancing at an ever-increasing rate.
• Accelerator — Any specialized hardware designed to run AI workloads faster or more efficiently than a general CPU.
• Accelerator chip — Specialized hardware that speeds up computationally intensive AI tasks like deep learning.
• Accuracy — The fraction of predictions a model gets correct on a labeled dataset.
• Act — The stage in an agent's loop where it executes a chosen action, like sending an API request.
• Action — The mechanism by which an agent in reinforcement learning changes the state of its environment.
• Action language — An action language formally describes how actions affect the state of a system over time, used in AI and robotics.
• Action model learning — Action model learning focuses on agents learning about the effects and prerequisites of actions in their environment.
• Action selection — Action selection is the fundamental problem of deciding "what to do next" for intelligent systems.
• Action space — The set of all possible actions an agent can take to interact with its environment.
• Activation Function — A non-linear function applied to a neuron's output, enabling the network to learn complex, non-linear relationships.
• Activation Patching — An interpretability technique that swaps activations from one forward pass into another to causally localize where a behavior lives in a model.
• Active Learning — An ML approach where the model selects which examples it most wants labeled next.
• AdaGrad — A gradient descent optimization algorithm that adapts the learning rate for each parameter individually.
• Adaptation — A synonym for tuning or fine-tuning a pre-trained AI model on new data.
• Adaptive algorithm — An adaptive algorithm modifies its behavior during execution based on predefined criteria or rewards.
• Adaptive Learning Systems — Educational technologies powered by AI that personalize learning paths and content based on an individual's progress and needs.
• Adaptive neuro fuzzy inference system (ANFIS) — ANFIS integrates neural networks and fuzzy logic to approximate nonlinear functions and learn from data.
• Admissible heuristic — An admissible heuristic never overestimates the cost to reach a goal, ensuring optimal pathfinding.
• Adversarial Attack — An input crafted to fool a model into making a wrong prediction.
• Adversarial Examples — Inputs crafted by an attacker to intentionally cause a machine learning model to make an incorrect prediction.
• Affective computing — Affective computing develops systems that can recognize, interpret, and simulate human emotions.
• Affective Robots — Robots designed with the ability to perceive, express, and potentially experience emotions, enabling more natural human interaction.
• Agent architecture — An agent architecture is a blueprint detailing the components and arrangement of a software or intelligent agent.
• Agent Assist — AI tools that support human customer service agents with real-time information and suggestions.
• Agent Framework — A software library or toolkit that simplifies the development of AI agents capable of planning, executing, and adapting tasks.
• Agent orchestration — Agent orchestration is the management of multiple AI agents or LLM calls to efficiently handle complex tasks.
• Agent Sandbox — A controlled, isolated environment used to safely develop, test, and evaluate the behavior of AI agents without impacting external systems.
• Agent-Based Model — A computational model simulating the actions and interactions of autonomous agents to observe system-level properties.
• Agentic — Agentic describes qualities associated with agents, such as autonomy and the ability to act independently.
• Agentic AI — AI systems designed to independently plan, execute, and adapt actions to achieve a given goal, often involving multiple steps and external tools.
• Agentic Commerce — Commerce driven by AI agents that compare, negotiate, and purchase on behalf of buyers or sellers.
• Agentic loop — An agentic loop is a cycle where an agent observes, reasons, acts, and receives feedback until a set goal is met.
• Agentic workflow — An agentic workflow is a dynamic process where an AI autonomously plans and executes actions to reach a goal, potentially self-correcting.
• Agglomerative clustering — Agglomerative clustering is a bottom-up hierarchical clustering method that iteratively merges the closest clusters.
• AI accelerator — Specialized hardware designed to speed up AI computations, especially for neural networks.
• AI Agent — An AI system that autonomously plans, uses tools, and takes actions to accomplish goals through iterative reasoning.
• AI Alignment — The research field aimed at making AI systems pursue the goals their developers and users actually intend.
• AI as a Service (AIaaS) — The provision of AI capabilities and tools, such as pre-trained models or API access, by third-party providers via cloud platforms.
• AI Assistant — A software program designed to help users with tasks, information retrieval, and productivity through conversational interfaces.
• AI bubble — A theorised stock market bubble focused on rapid investment increases in AI, impacting the broader economy.
• AI Compliance — Adhering to legal requirements — notably the EU AI Act — for how AI systems are built and operated.
• AI data center — A specialized data center optimized for the intensive computational needs of AI and machine learning model training and deployment.
• AI Ethics — A field studying the moral principles and values that should guide the design, development, and use of artificial intelligence.
• AI Ethics Principles — A set of guiding rules or values (e.g., fairness, transparency, accountability) used to steer responsible AI development and deployment.
• AI Factory — An integrated system or process for rapidly developing, training, deploying, and managing multiple AI models at scale within an organization.
• AI for Accessibility — The application of artificial intelligence technologies to assist individuals with disabilities, making technology and information more accessible.
• AI for Social Good — The application of AI technologies to address global challenges and contribute to positive societal impact.
• AI Governance — The framework of policies, regulations, and principles designed to guide the development and deployment of AI technologies.
• AI in Education — The application of artificial intelligence technologies to enhance teaching, learning, assessment, and administrative tasks in educational settings.
• AI Lifecycle Management — The holistic process of planning, developing, deploying, monitoring, and maintaining AI models throughout their entire operational lifespan.
• AI Red Teaming — A structured process of rigorously testing an AI system, often by adversarial attack simulations, to identify vulnerabilities, biases, and potential for misuse.
• AI Safety & Alignment — The field ensuring AI systems behave as intended, remain under human control, and avoid unintended harm.
• AI Skills Gap — The mismatch between the AI-related skills organizations need and those their workforce actually has.
• AI slop — AI slop is low-quality, high-volume output generated by AI systems, prioritizing quantity over usefulness.
• AI Treadmill — The continuous cycle of increasing AI model size, complexity, and resource demands to achieve incremental performance gains.
• AI Workflow Automation — The use of AI technologies to automate and optimize sequences of tasks or processes, reducing human intervention and error.
• AI-complete — AI-complete problems are computational challenges as difficult as achieving human-level artificial intelligence.
• Algorithm — A finite, well-defined sequence of steps used to learn from data or solve a computational problem.
• Algorithmic efficiency — Algorithmic efficiency measures how well an algorithm uses computational resources like time and memory.
• Algorithmic Justice — A framework addressing fairness, accountability, and transparency in algorithmic decision-making, especially concerning social and legal ramifications.
• Algorithmic probability — A method from algorithmic information theory for assigning prior probabilities to observations based on their complexity.
• Algorithmic Transparency — The ability to understand how and why an algorithm makes specific decisions or predictions.
• AlphaGo — AlphaGo is a computer program developed by DeepMind that famously defeated human Go champions.
• Analytics — Analytics is the process of discovering, interpreting, and communicating meaningful patterns found in data.
• Annotation — The act of attaching labels, tags, or structured information to raw data.
• Anomaly detection — Anomaly detection identifies unusual data points that deviate significantly from the expected pattern.
• Answer set programming (ASP) — A declarative programming paradigm for solving difficult search problems using logic programming principles.
• Ant colony optimization (ACO) — An optimization algorithm inspired by the foraging behavior of ants to find shortest paths in graphs.
• Anytime algorithm — An anytime algorithm can return a valid solution at any point during its execution, even if interrupted.
• API — An application programming interface — the contract by which software components communicate.
• Application programming interface (API) — An API is a set of rules and protocols that allows different software applications to communicate with each other.
• Approximate string matching — Finding strings that are approximately similar to a given pattern, allowing for errors or variations.
• Approximation error — The difference or discrepancy between an exact value and its approximated value.
• AR — AR is an abbreviation for augmented reality, which overlays digital information onto the real world.
• Area under the PR curve — Area under the PR curve (PR AUC) measures a classifier's performance, especially for imbalanced datasets.
• Area under the ROC curve — Area under the ROC curve (AUC) evaluates a classifier's ability to distinguish between classes across all possible thresholds.
• Argumentation framework — A formal system for representing and reasoning about arguments and their conflicts.
• Artificial general intelligence (AGI) — AI that can perform any intellectual task that a human being can.
• Artificial immune system (AIS) — AI systems inspired by biological immune systems to solve problems.
• Artificial intelligence (AI) — Machines that can perform tasks typically requiring human intelligence.
• Artificial intelligence arms race — Competition between states to develop and deploy advanced AI.
• Artificial Intelligence Markup Language — An XML format for creating AI that understands natural language.
• Asymptotic computational complexity — Analyzing algorithm efficiency as input size grows infinitely large.
• Attention — Attention in neural networks highlights important parts of the input to improve predictions, like focusing on key words.
• Attention Head — One of multiple parallel attention computations, each learning to focus on different types of relationships in the data.
• Attention Mechanism — A technique that lets models dynamically focus on the most relevant parts of the input when producing each output element.
• Attribute — An attribute is a characteristic or feature of a data point, often used in discussions of fairness and data.
• Attribute sampling — Attribute sampling trains decision trees using only random subsets of features at each node, improving efficiency and diversity.
• Attributional calculus — A logic system for natural induction and rule-based learning.
• AUC — Area Under the ROC Curve — a threshold-independent measure of classifier quality.
• Augmented Intelligence — A human-centered AI approach focused on assisting and enhancing human capabilities rather than replacing them.
• Augmented reality — Augmented reality (AR) enhances the real world by overlaying digital information, like images or sounds.
• Auto-regressive model — A model that predicts future outputs based on its own past predictions, commonly used in language models to generate sequences.
• Autoencoder — A neural network that learns compressed representations by training to reconstruct its own input through a bottleneck layer.
• Automata theory — The study of abstract machines and their computational capabilities.
• Automated machine learning (AutoML) — Automating the process of building and optimizing machine learning models.
• Automated planning and scheduling — AI that figures out sequences of actions to achieve goals.
• Automated reasoning — Developing computer programs that can perform logical reasoning automatically.
• Automatic evaluation — Automatic evaluation uses software to assess model output quality, often by comparing it to a known correct answer.
• Automatic Speech Recognition (ASR) — The technology that enables computers to convert spoken language into written text.
• Automation bias — Occurs when humans overly trust automated systems, even when they make mistakes, favoring their outputs over non-automated information.
• AutoML — Automated processes that build machine learning models, handling tasks like model selection, hyperparameter tuning, and data preparation.
• Autonomic computing (AC) — Computer systems that can self-manage and adapt to changes.
• Autonomous agent — An agent that operates independently to achieve goals by planning, acting, and adapting without constant human oversight.
• Autonomous Agents — AI systems that pursue goals over many steps with minimal human intervention.
• Autonomous car — A vehicle that can drive itself without human input using sensors and AI.
• Autonomous robot — A robot that can perform tasks independently with minimal human control.
• Autorater evaluation — A hybrid evaluation method combining human judgment with an ML model trained to mimic human evaluators for generative AI quality.
• Autoregressive Model — A statistical model that predicts future values based on past values, where each prediction contributes to the context for the next.
• Auxiliary loss — An additional loss function used during neural network training to speed up convergence and combat the vanishing gradient problem.
• Average precision at k — A metric evaluating ranked results by averaging precision scores at different recall levels up to k, measuring relevance in ordered lists.
• AWQ (Activation-aware Weight Quantization) — A 4-bit LLM quantization method that protects the most salient weight channels identified by activation magnitudes, beating GPTQ on many models.
• Axis-aligned condition — In decision trees, a condition that splits data based on a single feature's value, like 'age > 30'.
• Backpropagation — An algorithm that efficiently computes gradients by propagating errors backward through the network using the chain rule.
• Backpropagation through structure (BPTS) — A training technique for recurrent neural networks that calculates gradients by backpropagating errors through the network's structure.
• Backpropagation through time (BPTT) — A core algorithm for training recurrent neural networks by unfolding them over time and applying backpropagation.
• Backward chaining — An inference method that starts from a potential outcome and works backward to find supporting evidence or conditions.
• Bag of words — A text representation that ignores grammar and word order, focusing only on word frequency within a document.
• Bag-of-words model — A text representation that counts word occurrences, ignoring grammar and word order, to capture content.
• Bag-of-words model in computer vision — Adapts the bag-of-words concept for images by treating visual features as words.
• Bagging — An ensemble technique where models are trained on random subsets of data sampled with replacement to reduce variance.
• Base model — A pre-trained model used as a starting point for fine-tuning on specific downstream tasks.
• Baseline — A simple model or metric used as a reference point to evaluate the performance of more complex models.
• Baseline Model — A simple reference model used to put more sophisticated approaches in context.
• Batch — A group of training examples processed together in a single forward and backward pass.
• Batch inference — Processing multiple data examples at once to generate predictions, leveraging parallelism for efficiency.
• Batch Normalization — A technique that normalizes layer inputs across a mini-batch, stabilizing and accelerating deep network training.
• Batch size — The number of training examples processed in one iteration of a machine learning algorithm's update cycle.
• Batching — Grouping multiple inference requests so they run together on the GPU.
• Bayesian Network — A probabilistic graphical model representing a set of random variables and their conditional dependencies via a directed acyclic graph.
• Bayesian neural network — A probabilistic neural network that quantifies uncertainty in its weights and predictions by modeling them as probability distributions.
• Bayesian Optimization — An adaptive hyperparameter search that uses a probabilistic model to choose the next configuration to try.
• Bayesian programming — A methodology for specifying and solving problems using probabilistic models when information is incomplete.
• Bees algorithm — A population-based optimization algorithm inspired by the foraging behavior of honey bees.
• Behavior informatics (BI) — An interdisciplinary field focused on extracting intelligence and insights from behavioral data.
• Behavior tree (BT) — A tree structure used for defining complex agent behaviors through modular composition of tasks.
• Belief–desire–intention software model (BDI) — A cognitive agent architecture based on an agent's beliefs, desires, and intentions to guide its actions.
• Bellman equation — A formula in reinforcement learning that defines the value of an action based on immediate rewards and future expected rewards, used in Q-learning.
• Benchmark — A standardized dataset and evaluation protocol used to compare models.
• BERT — A bidirectional Transformer model pre-trained on masked language modeling, revolutionizing NLP benchmarks across the board.
• Bias (ethical) — Systematic, unfair model behavior that disadvantages particular groups of people.
• Bias (ethics/fairness) — Unfair prejudice or favoritism towards certain groups or things, which can influence data, system design, and user interactions.
• Bias (math) or bias term — A parameter in machine learning models, often representing an intercept, that shifts the output prediction.
• Bias (model parameter) — A learnable constant added to a neuron's weighted input that shifts the activation function.
• Bias-Variance Trade-off — The conflict in simultaneously minimizing two sources of error that prevent models from generalizing beyond their training data.
• Bias-Variance Tradeoff — The tension between a model's ability to fit training data (low bias) and its ability to generalize to new data (low variance).
• Bidirectional — Describes a system that processes information by looking at context both before and after a specific point.
• Bidirectional language model — A language model that considers both preceding and succeeding text to determine the probability of a token.
• Big data — Extremely large and complex datasets that require specialized tools for processing and analysis.
• Big O notation — A mathematical notation used to describe the performance or complexity of an algorithm as input size grows.
• Binary classification — A machine learning task that categorizes input into one of two mutually exclusive classes, like 'spam' or 'not spam'.
• Binary condition — A condition in a decision tree that results in one of two possible outcomes, typically 'yes' or 'no'.
• Binary tree — A tree data structure where each node has at most two children, known as the left and right child.
• Black box model — A model whose internal workings are opaque and difficult for humans to understand, despite knowing its inputs and outputs.
• Blackboard system — An AI architecture where multiple specialized agents cooperate by updating a shared knowledge base.
• BLEU (Bilingual Evaluation Understudy) — A metric to evaluate machine translation quality by comparing generated text to human references based on N-gram overlap.
• BLEU Score — A metric that measures how close machine-generated text is to human reference translations.
• BLEURT (Bilingual Evaluation Understudy from Transformers) — A metric that uses pre-trained language models to evaluate machine translation quality, focusing on semantic similarity.
• Boltzmann machine — A type of stochastic recurrent neural network used for generative learning and optimization.
• Boolean Questions (BoolQ) — A dataset used to evaluate an LLM's ability to answer yes/no questions based on a given text passage.
• Boolean satisfiability problem — The problem of determining if a given Boolean formula can be made true by assigning TRUE or FALSE values to its variables.
• BoolQ — Abbreviation for Boolean Questions, a dataset for evaluating yes/no question answering capabilities of LLMs.
• Boosting — A machine learning technique that trains models sequentially, with each new model correcting the errors of the previous ones, primarily to reduce bias.
• Bootstrap aggregating — An ensemble technique that trains multiple models independently and averages their predictions to reduce variance, commonly known as bagging.
• Bounding box — A rectangle defined by coordinates surrounding a region of interest in an image, often used in object detection.
• Brain technology — Technology inspired by neuroscience, aiming to create self-learning systems and 'know-how maps' for applications like robotics.
• Branching factor — The number of children or possible next moves from a given node in a tree structure, commonly used in search algorithms and game theory.
• Broadcasting — An operation that expands operand shapes to match for arithmetic operations between arrays of different sizes.
• Browser AI — Artificial intelligence functionalities integrated directly into web browsers to enhance user experience, productivity, and content interaction.
• Brute-force search — A general problem-solving technique that checks every possible solution systematically until the correct one is found, often computationally expensive.
• Bucketing — Converting a continuous feature into discrete categories or 'bins' based on value ranges.
• Business AI Applications — AI technologies and solutions specifically designed and implemented to solve business problems and improve organizational efficiency.
• Caching — Storing the results of expensive computations so they can be reused without recomputing.
• Calibration layer — An adjustment made after a model's predictions to correct for bias and align probabilities with observed outcomes.
• Candidate generation — The initial step in a recommendation system that narrows down a large set of items to a smaller, more manageable list of potential recommendations.
• Candidate sampling — A training optimization technique that reduces computational cost by only calculating probabilities for positive labels and a random sample of negative labels.
• Capsule Network — A type of neural network designed to preserve hierarchical spatial relationships between features.
• Capsule neural network (CapsNet) — A type of neural network that models hierarchical relationships more effectively by using 'capsules' to represent entities and their properties.
• Case-based reasoning (CBR) — A problem-solving approach that relies on finding and adapting solutions from similar past problems.
• Catastrophic Forgetting — The phenomenon where a neural network rapidly loses previously learned information upon learning new, distinct tasks.
• Categorical data — Data that represents categories or labels, where each data point belongs to one specific group from a defined set of possibilities.
• Causal Inference — The process of determining whether a cause-and-effect relationship exists between variables, not just correlation.
• Causal language model — A language model that predicts the next token based only on the preceding tokens, meaning it processes text in one direction.
• CB — An abbreviation for CommitmentBank, likely used as a metric in specific research contexts.
• Centroid — The central point of a cluster, typically calculated as the mean of all data points within that cluster in algorithms like k-means.
• Centroid-based clustering — A type of clustering algorithm that groups data points around a central point or centroid, aiming for non-hierarchical structures.
• Chain-of-Thought (CoT) Prompting — Asking models to show step-by-step reasoning before giving a final answer, improving accuracy on complex tasks.
• Chain-of-thought prompting — An advanced prompting technique that guides LLMs to break down complex problems into intermediate reasoning steps before providing a final answer.
• Character N-gram F-score (ChrF) — A metric for evaluating machine translation by comparing character n-grams between generated and reference text.
• Chat — An interactive dialogue between a user and an AI system, where previous turns influence the context of the ongoing conversation.
• Chatbot — A computer program designed to simulate conversation with human users, typically via text or voice.
• Chatbot Framework — A software library or platform providing tools and components to build, deploy, and manage conversational AI agents.
• Checkpoint — A saved state of a machine learning model's parameters, allowing training to be paused and resumed or enabling model deployment.
• Chief AI Officer (CAIO) — A C-suite executive responsible for a company's AI strategy, budget, and ethical/legal compliance.
• Choice of Plausible Alternatives (COPA) — A dataset and evaluation task designed to test an LLM's ability to choose the more plausible cause or consequence for a given premise.
• Citation precision — A metric measuring the proportion of citations in an LLM's response that accurately support the claims made within that response.
• Citation recall — A metric assessing what proportion of the source documents an LLM used to generate a response are actually cited within that response.
• Class — A category that a label can belong to, such as 'spam' or 'not spam' in an email classification task.
• Class-balanced dataset — A dataset where each category has roughly an equal number of examples, ensuring fair representation for all classes.
• Class-imbalanced dataset — A dataset where the number of examples for each class is significantly different, potentially biasing model performance.
• Classification — A supervised learning task where the model assigns inputs to discrete categories.
• Classification model — A predictive model that assigns an input to one of several predefined categories or classes.
• Classification threshold — A cutoff value used in binary classification to convert a model's raw output score into a final class prediction.
• Classifier — A common, informal term for a classification model.
• CLIP — A multi-modal model connecting images and text in a shared embedding space for zero-shot visual classification.
• Clipping — A method to control outliers by capping feature values at predefined minimum and maximum thresholds.
• Cloud robotics — A field integrating robotics with cloud computing for enhanced processing, storage, and shared intelligence among robots.
• Cloud TPU — Google's specialized hardware accelerator designed to significantly speed up machine learning computations on its cloud platform.
• Cluster analysis — The process of grouping similar data points together into clusters, used for exploratory data mining and pattern recognition.
• Clustering — An unsupervised technique that groups similar data points without using labels.
• Co-adaptation — An undesirable phenomenon where neurons become overly reliant on specific other neurons, leading to overfitting during training.
• Co-training — Co-training is a semi-supervised learning method using two independent model views to train on large unlabeled datasets.
• Cobweb — An incremental algorithm for hierarchical conceptual clustering that organizes data into a classification tree based on attribute distributions.
• Cognitive AI — AI systems designed to mimic human cognitive functions like perception, reasoning, learning, and problem-solving.
• Cognitive architecture — A theoretical framework specifying the fixed structures and mechanisms of a mind (natural or artificial) that enable intelligent behavior.
• Cognitive Architectures — Integrated computational frameworks designed to model an entire range of human-like cognitive capabilities, such as perception, reasoning, and learning.
• Cognitive computing — Computing systems that mimic human thought processes for enhanced decision-making, sensing, reasoning, and response.
• Cognitive Robotics — A field integrating AI, cognitive science, and robotics to create robots capable of human-like understanding, learning, and reasoning.
• Cognitive science — The interdisciplinary study of the mind, intelligence, and how information is processed, learned, and remembered.
• Collaborative filtering — A recommendation technique that predicts user preferences by analyzing the behavior and preferences of similar users.
• Collective Intelligence — The shared intelligence exhibited by groups of individuals or decentralized systems, whether human, animal, or artificial.
• Combinatorial optimization — Finding the best option from a finite set of possibilities, often used in fields like operations research and computer science.
• CommitmentBank (CB) — A dataset used to evaluate an LLM's ability to discern if a passage's author believes a specific stated clause.
• Common Sense Reasoning — The ability of AI systems to make inferences and judgments based on a broad, implicit understanding of the world, similar to humans.
• Commonsense knowledge — The vast store of implicit, everyday facts about the world that humans naturally understand and use.
• Commonsense reasoning — AI's attempt to replicate the human ability to make educated guesses about ordinary, everyday situations.
• Compact model — A small, efficient machine learning model designed for deployment on devices with limited resources.
• Completion — The text a language model generates in response to a prompt.
• Computational chemistry — Using computer simulations to solve problems and understand chemical phenomena.
• Computational complexity theory — Classifying computational problems by difficulty and understanding the relationships between different problem classes.
• Computational creativity — An interdisciplinary field exploring how computers can be used to generate novel and creative outputs.
• Computational cybernetics — Merging cybernetics principles with computational intelligence techniques for complex system control.
• Computational humor — AI research focused on using computers to understand, generate, or analyze humor.
• Computational intelligence (CI) — AI methods inspired by nature, focusing on learning from data and experimental observation.
• Computational learning theory — The theoretical study of machine learning algorithms, analyzing their performance and limitations.
• Computational linguistics — An interdisciplinary field modeling natural language using computational methods for analysis or processing.
• Computational mathematics — A field where computer simulation and computation are essential tools for solving mathematical problems.
• Computational neuroscience — Using mathematical models and computer simulations to understand the structure and function of the brain.
• Computational number theory — The study and development of algorithms for number theoretic computations.
• Computational Photography — The use of computational techniques to create new photographic capabilities that go beyond traditional optics.
• Computational problem — A computational problem is a collection of questions that computers can potentially answer, forming a core concept in theoretical computer science.
• Computational Social Science — The use of computational methods, including AI, to analyze large-scale social data and understand human behavior and societal dynamics.
• Computational statistics — Computational statistics is an interdisciplinary field that bridges statistics and computer science, focusing on the use of computational methods for statistical analysis.
• Compute — The processing power and resources required to train or run machine learning models.
• Computer science — Computer science is the study of computation, including the theory, design, and application of algorithms and computational systems.
• Computer Use — An agent capability where an LLM controls a computer GUI by viewing screenshots and emitting mouse/keyboard actions, automating arbitrary software.
• Computer vision — Computer vision enables computers to interpret and understand information from digital images and videos, automating tasks typically done by human sight.
• Computer-automated design (CAutoD) — Computer-automated design (CAutoD) extends CAD by using computers to automate aspects of the design process across various engineering disciplines.
• Concept Drift — When the relationship between inputs and the target outcome shifts over time, even if inputs look the same.
• Condition — In decision trees, a test performed at a node to split the data based on a feature's value.
• Confabulation — An alternative and more technically precise term for the phenomenon of AI hallucination, where an AI generates false or nonsensical information.
• Confidence Score — A numerical value indicating the AI's certainty in its prediction, such as an intent classification.
• Configuration — Setting initial property values for model training, including layers, data location, and hyperparameters like learning rate.
• Confirmation bias — The tendency to favor information that confirms existing beliefs, potentially influencing data collection and interpretation in ML.
• Confusion Matrix — A table showing true vs predicted labels for every class.
• Connectionism — Connectionism is an approach in cognitive science that explains mental processes using artificial neural networks.
• Consistency Model — A diffusion-derived model trained so any point on a noise trajectory maps to the same clean sample, enabling 1–4 step generation.
• Consistent heuristic — A consistent heuristic in AI pathfinding ensures that its estimated cost to the goal never overestimate the actual cost from any neighboring state.
• Constituency parsing — Breaking down a sentence into its grammatical components or constituents, like noun phrases and verb phrases, for easier analysis.
• Constitutional AI — An approach to aligning AI models, particularly LLMs, by providing them with a 'constitution' of principles to guide their behavior without human feedback.
• Constrained conditional model (CCM) — A constrained conditional model (CCM) is a machine learning framework that combines probabilistic or discriminative models with declarative constraints.
• Constraint logic programming — Constraint logic programming integrates logic programming with constraint satisfaction techniques to solve problems with complex variable relationships.
• Constraint programming — Constraint programming is a paradigm where problems are solved by stating relationships among variables in the form of constraints.
• Constructed language — A constructed language (conlang) is a language intentionally created with a specific phonology, grammar, and vocabulary, rather than evolving naturally.
• Contact Center AI — AI technologies deployed in customer service environments to automate tasks and assist human agents.
• Containerization — Packaging an application with all its dependencies so it runs consistently across environments.
• Content Moderation AI — AI systems designed to identify, filter, and manage inappropriate or harmful content on online platforms.
• Context Awareness — An AI system's ability to take situational signals — user, time, history, environment — into account.
• Context Window — The maximum number of tokens a model can process in a single input, determining information capacity.
• Contextual AI — AI systems that can understand and respond to the situational context, including user intent, environment, and temporal factors.
• Contextual Understanding — The ability of an AI system to interpret information based on its surrounding situation or details.
• Contextualized language embedding — Word representations that capture meaning based on surrounding text, going beyond static embeddings to understand nuance and context.
• Continual Learning — Training models that keep learning new tasks over time without forgetting old ones.
• Continuous feature — A numerical feature that can take any value within a given range, such as temperature or weight, rather than just specific discrete values.
• Contrastive Learning — A self-supervised technique that learns representations by pulling similar samples together and pushing dissimilar ones apart.
• Control theory — Control theory is a field of engineering and mathematics focused on managing the behavior of dynamic systems through control actions.
• ControlNet — A neural add-on that lets diffusion models be steered by extra conditioning images such as edges, depth maps, poses, or scribbles.
• Convenience sampling — Selecting data based on ease of access rather than rigorous methodology, suitable for quick initial tests but not final analysis.
• Convergence — The point in model training where improvements in performance (like reduced loss) become negligible with further iterations.
• Conversational AI — AI systems designed to simulate human conversation through text or voice, enabling natural language communication.
• Conversational coding — An interactive process where a user and an AI model collaborate through dialogue to generate and refine software code.
• Convex function — A mathematical function shaped like a 'U', where the line segment connecting any two points on the graph lies above or on the graph itself.
• Convex optimization — Finding the minimum value of a convex function using mathematical methods like gradient descent, ensuring a global minimum is found.
• Convex set — A geometric set where the line segment connecting any two points within the set is entirely contained within that set.
• Convolution — A mathematical operation that applies a filter to an input matrix, used in deep learning to extract features efficiently.
• Convolutional filter — A small matrix used in convolutional operations to detect specific features like edges or textures within an input matrix.
• Convolutional layer — A layer in a neural network that applies convolutional filters across the input to create feature maps, commonly used in image processing.
• Convolutional Neural Network (CNN) — A neural network that uses learnable filters to detect spatial patterns like edges, textures, and objects in images.
• Convolutional operation — The core calculation where a convolutional filter's values are element-wise multiplied with a slice of the input matrix, and the results are summed.
• COPA — COPA is a metric measuring the Choice of Plausible Alternatives in NLP tasks.
• Coreference Resolution — The NLP task of identifying all expressions that refer to the same entity in a text.
• Corpus — A large and structured collection of texts, often used to train NLP models.
• Cost Function — Another name for the loss function, typically averaged over the entire training set.
• Counterfactual fairness — Counterfactual fairness ensures a model's output is consistent across individuals identical except for sensitive attributes.
• Coverage bias — Coverage bias occurs when a model is trained on data that does not accurately represent the full population it will be used on.
• CPU — Central Processing Unit — the general-purpose chip that runs most software.
• Crash blossom — A crash blossom is an ambiguous phrase or sentence, creating multiple interpretations problematic for language understanding.
• Creative AI — AI systems designed to generate novel and aesthetically valuable content, including art, music, text, and designs.
• Cross-entropy — Cross-entropy measures the difference between two probability distributions, commonly used for multi-class classification loss.
• Cross-Validation — A technique that evaluates model performance by training and testing on different subsets of the data in rotation.
• Crossover — Crossover, or recombination, is a genetic operator in evolutionary computation that combines genetic information from two parent solutions to create new offspring solutions.
• Cumulative distribution function (CDF) — A CDF describes the probability that a random variable will take a value less than or equal to a specific point.
• Curiosity Reinforcement Learning — A reinforcement learning technique where an agent is intrinsically motivated to explore its environment and learn new information.
• Curse of Dimensionality — Various phenomena that arise when analyzing and organizing data in high-dimensional spaces, becoming sparse and difficult to generalize.
• Custom Instructions — User-defined persistent directives given to an AI chatbot to influence its behavior, tone, response style, or default persona across conversations.
• Customer Effort Score (CES) — A metric measuring how easy it is for customers to resolve an issue or complete a task with an AI system.
• Customer Satisfaction Score (CSAT) — A commonly used metric to gauge customer satisfaction with a specific interaction or service.
• Cyber-Physical Systems (CPS) — Integrated systems of computation, networking, and physical processes that interact with and control the physical world.
• Cyborg Intelligence — The integration of human biological intelligence with artificial intelligence and technological enhancements.
• Darkforest — Darkforest is a computer Go program developed by Facebook, using deep learning and convolutional neural networks in its design.
• Dartmouth workshop — The 1956 Dartmouth workshop is widely considered the seminal event that formally established artificial intelligence as a distinct field of research.
• Data analysis — Data analysis involves examining, cleaning, transforming, and visualizing data to discover useful information and patterns.
• Data Augmentation — Techniques that artificially expand training datasets by applying transformations to existing samples.
• Data Cleaning — Identifying and fixing or removing errors, inconsistencies, and duplicates in a dataset.
• Data Drift — When the statistical distribution of inputs to a deployed model changes over time.
• Data fusion — Combining data from multiple sources to create more accurate and useful information than any single source provides.
• Data Imbalance — When some classes or groups are vastly more frequent in the dataset than others.
• Data integration — Combining data from different sources to provide a unified view, crucial for handling large and diverse datasets.
• Data Leakage — When information that wouldn't be available at inference time accidentally influences training.
• Data Lineage — The lifecycle of data, including its origin, transformations, and where it moves over time.
• Data mining — Discovering patterns and insights in large datasets using methods from machine learning, statistics, and databases.
• Data Moat — A sustainable competitive advantage derived from exclusive access to or control over a unique and valuable dataset.
• Data parallelism — Data parallelism scales training by distributing data subsets across replicated models on multiple devices.
• Data Pipeline — An automated workflow that ingests, transforms, and delivers data to models.
• Data Preprocessing — Transforming raw data into a format suitable for model training.
• Data science — An interdisciplinary field that extracts knowledge and insights from data using scientific methods, algorithms, and systems.
• Data set — A collection of data, typically organized as rows and columns, where each row represents an item and columns represent its attributes or variables.
• Data set or dataset — A dataset is a collection of raw data, often organized in formats like spreadsheets or CSV files.
• Data Versioning — Tracking changes to datasets the way Git tracks changes to code.
• Data warehouse (DW or DWH) — A central system for reporting and analysis that integrates historical and current data from various sources into one repository.
• Data-Centric AI — An approach to AI development that prioritizes improving the quality, consistency, and quantity of data over model architecture.
• DataFrame — A DataFrame is a 2D labeled data structure, similar to a spreadsheet or SQL table, commonly used in data manipulation.
• Datalog — A declarative logic programming language, often used for deductive databases and in areas like data integration and program analysis.
• Dataset — An organized collection of examples used to train, validate, or test a model.
• Dataset API (tf.data) — The Dataset API (tf.data) is a TensorFlow tool for building efficient input pipelines for machine learning.
• Dataset Shift — A change in the distribution of the data between the training and test (or deployment) phases.
• Decision boundary — The line or surface that separates different classes in a classification problem, defining where a model makes its predictions.
• Decision forest — A decision forest is an ensemble of multiple decision trees, combining their predictions for improved accuracy and robustness.
• Decision Making — Choosing actions or recommendations under uncertainty, often using probabilistic or utility-based reasoning.
• Decision support system (DSS) — An information system that helps organizations make better decisions by analyzing data, especially for unstructured or semi-structured problems.
• Decision theory — The study of how agents make choices, encompassing both how they ideally should make decisions (normative) and how they actually do (descriptive).
• Decision threshold — A decision threshold is a value used in classification to convert probabilities into discrete class labels.
• Decision tree — A decision tree is a supervised learning model that uses a tree-like structure of decisions and their possible consequences.
• Decision tree learning — A machine learning method that uses a tree-like structure to make predictions by recursively splitting data based on feature values.
• Declarative programming — A programming paradigm that expresses the logic of a computation without explicitly describing its control flow.
• Decoder — A model component that generates output sequences, typically one token at a time.
• Decoder-only Transformer — A type of transformer model that consists solely of a stack of decoder layers, primarily used for generative tasks like text generation.
• Deductive classifier — An AI inference engine that uses domain knowledge expressed as declarations to classify information.
• Deep Blue — IBM's pioneering chess-playing computer that famously defeated world champion Garry Kasparov in 1997.
• Deep Learning — A subset of machine learning using neural networks with many layers to learn hierarchical representations from large datasets.
• Deep model — A neural network with more than one hidden layer, also known as a deep neural network.
• Deep Q-Network (DQN) — A type of Q-learning algorithm that uses a deep neural network to approximate the Q-value function.
• Deep Reinforcement Learning — The combination of deep learning and reinforcement learning, allowing agents to learn complex tasks directly from raw sensory input.
• Deepfake — Synthetic media in which a person in an existing image or video is replaced with someone else's likeness.
• Default logic — A non-monotonic logic that handles reasoning based on default assumptions, allowing conclusions to be retracted if new information contradicts them.
• Demographic parity — A fairness metric ensuring prediction rates are the same across different demographic groups.
• Denoising — A self-supervised learning technique where a model learns to remove artificial noise added to data.
• Dense feature — A feature where most values are non-zero, typically represented as a floating-point tensor.
• Dense layer — Another name for a fully connected layer in a neural network.
• Density-based spatial clustering of applications with noise (DBSCAN) — A clustering algorithm that groups together points that are closely packed, marking points in low-density regions as outliers.
• Deployment — The process of moving a trained model into a production environment where it serves real users.
• Depth — The total number of hidden, output, and embedding layers in a neural network.
• Depthwise separable convolutional neural network (sepCNN) — A CNN that uses depthwise separable convolutions for efficiency.
• Description logic (DL) — A family of formal knowledge representation languages used for AI, offering more expressivity than propositional logic but with decidable reasoning.
• Deterministic — A system or function that produces the exact same output for a given input every time.
• Developmental robotics (DevRob) — A field studying how embodied machines can learn lifelong and open-ended skills through developmental mechanisms, inspired by human development.
• Device — Hardware, like a CPU or GPU, that runs computations for a machine learning framework.
• Diagnosis — The process of identifying system malfunctions by analyzing observed behavior to pinpoint the cause of incorrect functioning.
• Dialogue Management — The process of controlling the flow and state of a conversation between a user and an AI system.
• Dialogue system — A computer system designed to converse with humans using various modes of communication and maintaining a coherent interaction.
• Differential Privacy — A mathematical framework that bounds how much any single individual's data can affect a model or query result.
• Diffusion Model — A generative model that learns to create data by reversing a gradual noising process, producing high-quality images and audio.
• Diffusion Transformer (DiT) — A diffusion model that replaces the U-Net denoiser with a Transformer operating on image or video patches, scaling cleanly with compute.
• Digital Twin — A high-fidelity virtual replica of a physical system that updates in real time from sensor data.
• Dijkstra's algorithm — An algorithm that finds the shortest paths from a single source node to all other nodes in a graph with non-negative edge weights.
• Dimension reduction — Techniques to reduce the number of features or variables in a dataset while retaining important information.
• Dimensionality reduction — The process of reducing the number of features or variables in a dataset while retaining essential information.
• Dimensions — Refers to the number of elements or coordinates needed to specify an item in a data structure like a tensor.
• Direct Preference Optimization (DPO) — A simpler RLHF alternative that fine-tunes an LLM directly on preference pairs using a closed-form loss, with no reward model or RL loop.
• Direct prompting — Another term for zero-shot prompting, where a model answers a prompt without prior examples.
• Disambiguation — The process of resolving ambiguities in natural language, choosing the correct meaning or interpretation.
• Discrete feature — A feature that can only take on a finite, countable number of distinct values.
• Discrete system — A system characterized by a finite or countable number of distinct states, often modeled using graphs.
• Discriminative model — A model that learns to distinguish between different classes or predict a value based on input features.
• Discriminator — A system that identifies whether input data is real or fake, often used within generative adversarial networks.
• Disparate impact — When an algorithm's decisions disproportionately affect different population subgroups, even if not intentionally biased.
• Disparate treatment — When an algorithm explicitly uses sensitive attributes to make different decisions for different population subgroups.
• Distillation — Compressing a large AI model into a smaller one that mimics the original's performance, improving efficiency.
• Distributed artificial intelligence (DAI) — A subfield of AI focused on developing decentralized solutions to problems using multiple intelligent agents.
• Distributed Training — Training a model across multiple machines or accelerators in parallel.
• Distribution — The pattern or frequency of different values within a dataset or for a specific feature.
• Divisive clustering — A hierarchical clustering method that starts with all data points in one cluster and progressively splits them.
• Document Classification — Assigning predefined categories or labels to entire documents based on their content.
• Document Understanding — The ability of AI systems to process, interpret, and extract meaningful information from unstructured and semi-structured documents.
• Domain Adaptation — Techniques for applying a model trained on a source domain to a different but related target domain with limited labeled data.
• Domain-Specific Language — Vocabulary and phrasing unique to a particular industry or area of expertise.
• Double descent — A phenomenon where model test error is low for both very simple and very complex models, but high for models with intermediate complexity.
• Downsampling — Reducing the number of data points, either by lowering resolution or selecting fewer examples from an over-represented class.
• DQN — Deep Q-Network, a reinforcement learning algorithm combining deep learning with Q-learning for complex tasks.
• Dropout — A regularization technique that randomly deactivates neurons during training, forcing the network to build robust, redundant features.
• Dropout regularization — A technique that randomly deactivates neurons during training to prevent overfitting.
• Dynamic — Describes processes or models that change or adapt frequently over time, often in response to new data.
• Dynamic epistemic logic (DEL) — A logical framework that models how agents' knowledge and beliefs change over time due to events or new information.
• Dynamic model — A model that is continually retrained or updated to adapt to new data and changing patterns.
• Eager execution — A TensorFlow execution mode where operations are performed immediately as they are called.
• Eager learning — A machine learning approach where the model builds a general hypothesis during training, applying it to all future predictions.
• Early stopping — A regularization technique used in iterative training to stop model learning when performance on a validation set starts to degrade.
• Earth mover's distance (EMD) — A metric measuring the minimum cost to transform one probability distribution into another.
• Ebert test — A proposed test to evaluate a synthesized voice's ability to tell a joke effectively, requiring human-like delivery and timing.
• Echo state network (ESN) — A type of recurrent neural network with a fixed, randomly connected hidden layer, where only output weights are learned to reproduce temporal patterns.
• Edge AI — Running AI models directly on user devices instead of cloud servers.
• Edit distance — A metric that calculates the minimum number of single-character edits required to change one string into another.
• Einsum notation — A concise notation for expressing tensor operations like multiplication and summation across specified axes.
• Embedding — A dense vector representation that captures semantic meaning, mapping discrete items like words into continuous mathematical space.
• Embedding layer — A hidden neural network layer that learns lower-dimensional representations (embeddings) for high-dimensional categorical data.
• Embedding space — A multi-dimensional space where learned representations (embeddings) of data are mapped to, capturing meaningful relationships.
• Embedding vector — A dense array of numbers learned by a model that represents an input item in a lower-dimensional space.
• Embeddings — Dense vector representations of discrete items, capturing their semantic relationships and meanings.
• Embodied agent — An intelligent agent that perceives and acts within an environment through a physical or virtual body.
• Embodied AI — AI systems that exist within a physical body, allowing them to interact with and learn from the real world through physical experience.
• Embodied Cognition — A theory suggesting that intelligence is shaped by the physical body and its interactions with the environment.
• Embodied cognitive science — An interdisciplinary field studying intelligence by modeling mind and body as one entity and using robotic agents.
• Emergent Abilities — New capabilities that appear in large models only after they reach a certain scale, not predictable from smaller versions of the same model.
• Emergent Abilities — Capabilities like multi-step arithmetic or instruction following that appear suddenly above a scale threshold rather than improving smoothly.
• Emergent behavior — Capabilities or behaviors in large language models that were not explicitly programmed or trained for.
• Empirical cumulative distribution function (eCDF or EDF) — A function showing the proportion of data points at or below a specific value, based on observed data.
• Empirical risk minimization (ERM) — A machine learning principle that aims to find a model that performs best on the observed training data.
• Encoder — A model component that compresses input data into a dense, information-rich representation.
• Encoder-Decoder Transformer — A transformer model composed of both an encoder (processing input) and a decoder (generating output), ideal for sequence-to-sequence tasks.
• Endpoints — Network addresses, typically URLs, where a service can be accessed and communicated with.
• Ensemble — Combining multiple independently trained models to improve overall prediction accuracy and robustness.
• Ensemble learning — Combining multiple machine learning models to improve overall prediction accuracy and robustness.
• Ensemble Methods — Techniques that combine multiple models to produce better predictions than any single model alone.
• Enterprise AI — AI strategies, technologies, and applications deployed across an entire organization to solve business challenges.
• Entity Recognition (Named Entity Recognition) — Identifying and classifying named entities in text into predefined categories like names, dates, or locations.
• Entropy — A measure of the unpredictability or randomness in a probability distribution.
• Environment — The external world or system with which an agent in reinforcement learning interacts and receives feedback.
• Environment grounding — The feedback and contextual data an agent receives from its environment after taking an action.
• Episode — A complete sequence of interactions between an agent and its environment in reinforcement learning, from start to finish.
• Episodic memory — In LLMs, information learned or retained by the model after its initial training period.
• Epoch — One full pass through the entire training dataset during model training.
• Epsilon greedy policy — A strategy where an agent randomly explores the environment most of the time, but sometimes exploits known good actions.
• Equality of opportunity — A fairness metric ensuring a model's positive predictions are equally accurate across different sensitive groups.
• Equalized odds — A fairness metric where true positive rate and false negative rate are equal across all groups.
• Error-driven learning — A machine learning approach where an agent learns by minimizing errors caused by its actions in an environment.
• Escalation Path — A defined route for transferring a customer's query to a higher level of support, often a human agent.
• Estimator — A deprecated TensorFlow API for building and training machine learning models.
• Ethical AI — The study and practice of designing, developing, and deploying AI systems in a responsible and morally sound manner.
• Ethics of artificial intelligence — The study of moral principles and values concerning the design, development, and deployment of AI.
• Evals — Abbreviation for LLM evaluations or any form of evaluation.
• Evaluation — The process of measuring a model's quality or comparing different models.
• Evaluation Metrics — Quantitative measures used to assess the performance, accuracy, and quality of AI models for specific tasks.
• Evaluator agent — An agent that checks and validates the results produced by another agent.
• Evolutionary algorithm (EA) — An optimization algorithm inspired by biological evolution, using processes like selection and mutation to find solutions.
• Evolutionary computation — A family of optimization algorithms that mimic biological evolution to solve problems.
• Evolving classification function (ECF) — Classification methods, often for data streams, that adapt and change dynamically over time.
• Exact match — A metric that requires model output to perfectly match the ground truth.
• Example — A single instance of data, consisting of features and possibly a label.
• Existential risk from artificial general intelligence — The potential risk that advanced artificial general intelligence (AGI) could lead to human extinction or catastrophe.
• Experience replay — A technique in reinforcement learning that stores and samples past transitions to train agents.
• Experiment Tracking — Systematically recording the inputs, code, parameters, and outputs of every ML experiment.
• Experimenter's bias — Bias introduced by the experimenter's expectations influencing the results.
• Expert system — AI programs that mimic the decision-making ability of a human expert using a knowledge base and inference rules.
• Explainability (XAI) — Techniques making AI decisions understandable to humans, crucial for trust and regulatory compliance.
• Explainable AI (XAI) — A set of techniques that allow humans to understand the output of AI models, especially deep learning models.
• Explainable Reinforcement Learning (XRL) — Methods that provide insights into the decision-making processes of reinforcement learning agents.
• Exploding gradient problem — Gradients become excessively large during neural network training, hindering learning.
• Extreme Summarization (xsum) — A dataset evaluating LLMs on single-document summarization to a single sentence.
• F1 — A metric balancing precision and recall for binary classification tasks.
• F1 Score — The harmonic mean of precision and recall, balancing both into a single number.
• Factual Accuracy — The degree to which an AI model's generated or retrieved information aligns with verifiable facts and real-world knowledge.
• Factuality — A property of a model whose output is grounded in factual reality.
• Fairness — The principle that AI systems should treat individuals and groups equitably.
• Fairness constraint — An imposed restriction on an algorithm to ensure specific fairness criteria are met.
• Fairness metric — A quantifiable measure used to assess the fairness of a model's predictions across different groups.
• Fall-back Strategies — Pre-defined actions an AI takes when it cannot understand a user's query or fulfill a request.
• False negative (FN) — A false negative occurs when a model incorrectly predicts the negative class for a data point that actually belongs to the positive class.
• False negative rate — The false negative rate measures the proportion of actual positive instances that were incorrectly predicted as negative.
• False positive (FP) — A false positive occurs when a model incorrectly predicts the positive class for a data point that actually belongs to the negative class.
• False positive rate (FPR) — The false positive rate measures the proportion of actual negative instances that were incorrectly predicted as positive.
• Fast decay — Fast decay is a training technique for LLMs where the learning rate is rapidly decreased to improve generalization and prevent overfitting.
• Fast-and-frugal trees — A simple decision-making tool that uses a series of simple, sequential steps (like a tree) to make classifications.
• Feature — An individual, measurable characteristic or property used as input for machine learning models.
• Feature cross — A feature cross combines two or more categorical features to create a new synthetic feature, capturing interaction effects.
• Feature Engineering — The process of creating, selecting, and transforming input variables to improve a machine learning model's performance.
• Feature Extraction — Deriving informative numerical signals from raw data for use as model inputs.
• Feature importances — Feature importances indicate the relative contribution of each feature to a model's predictions, showing which features are most influential.
• Feature learning — Techniques that allow AI systems to automatically discover and create relevant features from raw data.
• Feature Selection — Choosing the most useful subset of features to improve performance and interpretability.
• Feature set — A feature set is the collection of all input features used by a machine learning model during training and inference.
• Feature spec — A feature spec defines how to extract and interpret features from data, specifying their type and shape for model consumption.
• Feature vector — A feature vector is a numerical representation of an example, composed of its feature values, used as input for machine learning models.
• Featurization — Featurization is the process of transforming raw input data into a feature vector that a machine learning model can process.
• Federated Learning — Distributed training where models learn from data on many devices without the data ever leaving those devices.
• Feedback — Feedback is the evaluation of an agent's action, providing information about its success or failure to guide future behavior.
• Feedback loop — A feedback loop occurs when a model's output influences its future input or training data, potentially leading to system drift or reinforcement of biases.
• Feedforward neural network (FFN) — A feedforward neural network is a type of artificial neural network where connections between nodes do not form cycles; information moves in only one direction.
• Few-Shot Learning — Training a model to recognize new patterns from just a handful of labeled examples.
• Few-shot prompting — Few-shot prompting provides a large language model with a small number of examples within the prompt to guide its response to a query.
• Fiddle — A Python-first configuration library for managing ML model and training hyperparameters without invasive code changes.
• Fine-grained Control — Steering model outputs precisely along specific dimensions like style, length, tone, or factuality.
• Fine-Tuning — Adapting a pre-trained model to a specific task by continuing training on a smaller, task-specific dataset.
• First-order logic — A formal logic system using variables and quantifiers to express propositions more complexly than propositional logic.
• Flash model — A family of fast and low-latency Gemini models optimized for applications requiring quick responses and high throughput.
• FlashAttention — An IO-aware exact attention algorithm that tiles Q, K, V into SRAM blocks, cutting memory from O(n²) to O(n) and giving 2–4× speedups.
• Flax — A high-performance neural network library built on JAX, providing tools for training and evaluating deep learning models.
• Flaxformer — An open-source Transformer library based on Flax, primarily used for NLP and multimodal research.
• Flow Matching — A generative-modeling framework that learns a velocity field transporting samples from noise to data along a chosen probability path.
• Fluent — A condition or property that can change over time, often represented in logic by predicates depending on a time argument.
• Forget gate — A component within an LSTM cell that selectively discards information from the cell state to maintain relevant context.
• Formal language — A precisely defined set of strings, built from an alphabet and adhering to specific formation rules.
• Forward chaining — A reasoning strategy that starts with data and applies rules to reach a goal, often used in expert systems.
• Fraction of successes — A metric that measures the proportion of generated outputs that meet a specific success criterion.
• Frame — An AI data structure representing stereotyped situations by dividing knowledge into structured units.
• Frame language — A knowledge representation technology using 'frames' to model concepts and their relationships as ontologies.
• Frame problem — The challenge in AI of specifying which aspects of an environment *do not* change when an action occurs.
• Friendly artificial intelligence — A hypothetical AGI designed to have a beneficial impact on humanity and its values.
• Full softmax — A synonym for the standard softmax function used in multi-class classification.
• Fully connected layer — A layer in a neural network where each neuron is connected to every neuron in the preceding layer.
• Function Calling — Provider-defined schemas that let an LLM invoke developer-supplied functions in a structured way.
• Function transformation — A process that takes a function and returns a modified version of it, often used for optimization or differentiation.
• Futures studies — The interdisciplinary study of possible, probable, and preferable future scenarios and their underlying worldviews.
• Fuzzy control system — A system that uses fuzzy logic to manage complex processes based on imprecise or continuous input values.
• Fuzzy logic — A logic system that uses 'degrees of truth' between 0 and 1, allowing for imprecise reasoning.
• Fuzzy rule — An 'IF-THEN' statement used in fuzzy logic systems to map fuzzy inputs to fuzzy outputs.
• Fuzzy set — A set where elements can have a degree of membership between 0 and 1, unlike traditional binary sets.
• Game theory — The mathematical study of strategic interactions between rational decision-makers.
• GAN — Abbreviation for Generative Adversarial Network, a class of machine learning frameworks.
• Gemini — Google's advanced AI ecosystem, encompassing various sophisticated multimodal models and related tools.
• Gemini models — Google's state-of-the-art Transformer-based multimodal models designed for integration with AI agents.
• Gemma — A family of lightweight, open AI models derived from the same technology as Google's Gemini models.
• GenAI or genAI — Abbreviation for Generative Artificial Intelligence, focusing on AI systems that create new content.
• General game playing (GGP) — An AI approach focused on creating agents capable of playing multiple games without game-specific programming.
• Generalization — A model's ability to perform well on new, unseen data — not just its training set.
• Generalization curve — A plot showing training and validation loss over time to help detect overfitting in ML models.
• Generalization error — A measure of how poorly a machine learning model performs on new, unseen data compared to its training data.
• Generalized linear model — A flexible regression model that generalizes standard linear regression to accommodate different error distributions and link functions, allowing it to model a wider range of data types.
• Generated text — Text produced by a machine learning model, often used in evaluating the performance of language models against human-created reference texts.
• Generative Adversarial Network (GAN) — Two neural networks — a generator and discriminator — compete against each other to produce increasingly realistic synthetic data.
• Generative Adversarial Networks (GANs) — A class of deep learning frameworks composed of two neural networks, a generator and a discriminator, competing against each other.
• Generative agents (simulacra) — AI agents equipped with personas, memories, and routines designed to simulate realistic and emergent human behavior in interactive environments.
• Generative AI — An AI field focused on creating models that can generate novel and complex content, such as text, images, audio, and video, that is coherent and original.
• Generative artificial intelligence — AI that can create new content like text, images, or music based on patterns learned from existing data.
• Generative Design — An iterative design exploration process that uses algorithms to rapidly generate numerous design options based on specified goals and constraints.
• Generative Engine Optimization (GEO) — Optimizing content so it is preferentially cited and recommended by generative AI assistants.
• Generative Fill — An AI-powered feature, often in image editing software, that allows users to add or remove content from images by generating new pixels based on context.
• Generative model — A type of machine learning model that can create new data instances similar to the data it was trained on, or estimate the likelihood of a given data point originating from the training distribution.
• Generative Pre-trained Transformer (GPT) — A type of large language model (LLM) that utilizes the transformer architecture and is pre-trained on a massive amount of text data, capable of generating human-like text.
• Generative pretrained transformer (GPT) — A type of large language model that generates human-like text by predicting words in a sequence after being trained on vast amounts of text data.
• Generator — The component of a Generative Adversarial Network (GAN) responsible for creating new, synthetic data samples.
• Genetic algorithm (GA) — An optimization technique inspired by natural selection, using processes like mutation and crossover to find solutions to complex problems.
• Genetic operator — A function used in genetic algorithms to modify potential solutions, mimicking biological processes like mutation and crossover.
• GGUF — A single-file binary format from the llama.cpp project for distributing quantized LLM weights together with tokenizer and metadata.
• Gini impurity — A measure of how well a function separates data into distinct classes, often used in decision trees to evaluate the quality of splits.
• Glowworm swarm optimization — An optimization algorithm inspired by the cooperative behavior of glowworms, using principles of swarm intelligence.
• Golden dataset — A curated dataset representing the ground truth, used to evaluate the performance and accuracy of machine learning models.
• Golden response — A manually created, high-quality output used as a benchmark to evaluate the quality of responses generated by AI models.
• GPT (Generative Pre-trained Transformer) — A family of powerful Transformer-based language models developed by OpenAI, known for their ability to generate human-like text and perform various NLP tasks.
• GPTQ — A one-shot post-training quantization method that compresses LLM weights to 3–4 bits using approximate second-order information.
• GPU — Graphics Processing Unit — a massively parallel processor that powers most modern AI workloads.
• Gradient — The vector of partial derivatives of the loss with respect to each model parameter.
• Gradient accumulation — A technique to simulate larger batch sizes during training by accumulating gradients over multiple mini-batches before performing a single parameter update.
• Gradient boosted (decision) trees (GBT) — An ensemble machine learning technique that combines multiple decision trees sequentially, where each new tree corrects the errors of the previous ones.
• Gradient boosting — A machine learning method that builds models sequentially, with each new model correcting the errors of the previous ones.
• Gradient clipping — A technique to prevent exploding gradients during neural network training by capping the magnitude of gradients that exceed a certain threshold.
• Gradient Descent — An optimization algorithm that iteratively adjusts model parameters by moving in the direction of steepest decrease of the loss function.
• Graph (abstract data type) — An abstract data type representing a network of nodes connected by edges, used to model relationships between objects.
• Graph database (GDB) — A database that uses graph structures with nodes and edges to store and query data based on relationships.
• Graph execution — A mode of operation in TensorFlow where a computation graph is first defined and then executed, enabling optimizations and portability.
• Graph Neural Network (GNN) — A class of deep learning methods designed to perform inference on data structured in graph form.
• Graph theory — The study of graphs, which are mathematical structures used to model relationships between objects.
• Graph traversal — The process of systematically visiting each node in a graph, often used in algorithms for searching or mapping networks.
• GraphRAG — A retrieval-augmented generation variant that builds a knowledge graph from the corpus and retrieves entity- and community-level summaries.
• Greedy policy — A reinforcement learning policy that always selects the action expected to yield the highest reward.
• Greeting Management — The AI's ability to appropriately initiate and respond to conversational greetings.
• Grid Search — Exhaustively trying every combination of a predefined set of hyperparameter values.
• Grokking — A training phenomenon where a network memorizes the training set quickly but only generalizes much later, after many additional optimization steps.
• Ground truth — The actual reality or outcome of a situation, used to assess model performance.
• Groundedness — A model's output is grounded if it directly reflects information from its provided source material.
• Grounding — Tying a model's outputs to verifiable, external sources of truth.
• Group attribution bias — Assuming an individual possesses characteristics solely because they belong to a certain group.
• Grouped-Query Attention (GQA) — An attention variant where multiple query heads share each key/value head, shrinking the KV-cache while keeping near-multi-head quality.
• GRPO (Group Relative Policy Optimization) — A PPO variant from DeepSeek that drops the value network and estimates advantages from the relative reward of a group of sampled completions.
• GRU — A gated recurrent unit — a streamlined alternative to the LSTM with fewer parameters.
• Guardrails — Runtime checks that constrain model inputs and outputs to stay within safe, useful bounds.
• Hallucination — When an AI model generates confident, plausible-sounding content that is factually incorrect or fabricated.
• Hashing — A technique to group many categorical features into fewer, more manageable buckets.
• Heuristic — A practical method or rule of thumb used to find a solution quickly, often sacrificing optimality for speed.
• Heuristics — Practical rules of thumb that produce good-enough solutions when optimal ones are too expensive.
• Hidden Layer — Any layer in a neural network that is neither the input nor the output.
• Hierarchical clustering — An algorithm that creates a tree-like structure of nested clusters.
• High-Risk AI Classification — EU AI Act designation for AI systems that materially affect people's rights, safety, or opportunities.
• Hill climbing — An iterative optimization technique that moves towards better solutions until no further improvement is possible.
• Hinge loss — A loss function that penalizes incorrect classifications and aims to maximize the margin around the decision boundary.
• Historical bias — Bias present in data reflecting past societal prejudices, inequalities, or stereotypes.
• Holdout data — Data intentionally set aside and not used during model training for evaluation.
• Host — The part of a machine learning system, often a CPU, that manages training on accelerators like GPUs.
• Human evaluation — Assessing model performance by using human judgment to rate its outputs.
• Human Handoff — Seamlessly transferring a user's conversation from an AI system to a live human agent.
• Human-AI Interaction (HAI) — The study of how people communicate with and relate to artificial intelligence systems, focusing on design, usability, and user experience.
• Human-AI Teaming — The collaborative interaction between humans and AI systems to achieve shared goals more effectively than either could alone.
• Human-Centered AI (HCAI) — An approach to AI development prioritizing human needs, values, and well-being throughout the design and deployment process.
• Human-in-the-loop — System designs where humans review, correct, or guide AI outputs as part of the workflow.
• Human-in-the-Loop AI — A system design approach where human intelligence is integrated into an AI workflow to review, validate, or enhance AI decisions.
• Human-Robot Interaction (HRI) — The field dedicated to understanding, designing, and evaluating interactions between humans and robots.
• HumanEval — OpenAI's 164-problem Python coding benchmark where models must write a function passing hidden unit tests — pass@k is the standard metric.
• HyDE (Hypothetical Document Embeddings) — A retrieval technique that asks an LLM to generate a hypothetical answer first, then embeds and retrieves real documents similar to it.
• Hyper-automation — The application of advanced technologies like AI and RPA to automate as many business processes as possible.
• Hyper-heuristic — An automated method that combines or adapts simpler heuristics to solve complex search and optimization problems.
• Hyperautomation — Combining AI, RPA, and process mining to automate entire end-to-end business processes.
• Hyperparameter — A configuration variable set before the training process begins, controlling aspects of the learning algorithm itself.
• Hyperparameter optimization — The process of finding the best set of hyperparameters for a machine learning model to achieve optimal performance.
• Hyperparameter Tuning — The process of finding optimal configuration values that control model training, such as learning rate, batch size, and architecture choices.
• Hyperplane — A decision boundary used in classification algorithms to separate data points into different classes in a high-dimensional space.
• I.i.d. — Independently and identically distributed; a common assumption in machine learning that data points are unrelated and follow the same distribution.
• Image Classification — The task of categorizing images into one of several predefined classes.
• Image recognition — The process of identifying and classifying objects, patterns, or concepts within an image.
• Image Segmentation — Classifying every pixel in an image to delineate objects and boundaries with detailed spatial understanding.
• Imbalanced dataset — A dataset where the classes are not represented equally, posing challenges for model training.
• Imitation Learning — Training an agent to mimic expert behavior by learning directly from demonstrations.
• Implicit bias — An unconscious association or assumption that influences perceptions and decisions, potentially affecting data collection and model design.
• Imputation — The process of replacing missing data points with substituted values.
• In-context learning — A technique where a language model learns to perform a task by conditioning its output on examples provided within the input prompt.
• In-group bias — Favoring one's own group or characteristics, potentially leading to biased evaluations or data in machine learning.
• In-set condition — A condition in a decision tree that checks if a feature's value is present within a specified set of acceptable values.
• Incompatibility of fairness metrics — The mathematical reality that some definitions of fairness cannot be simultaneously satisfied, requiring context-specific choices.
• Incremental learning — A machine learning approach where a model continuously learns from new data without forgetting previous knowledge.
• Independently and identically distributed (i.i.d) — Data points are independent if one does not affect another, and identically distributed if they come from the same underlying probability distribution.
• Indexing — Organizing data so that lookups, searches, or similarity queries run quickly.
• Individual fairness — A fairness principle stating that similar individuals should be treated similarly by a machine learning model.
• Inference — Using a trained model to make predictions on new data — the deployment phase of machine learning.
• Inference Endpoint — A network address that accepts inputs and returns model predictions.
• Inference engine — A component of an AI system that uses logical rules to deduce new information from a knowledge base.
• Inference path — The sequence of decisions or conditions a data point follows through a decision tree to reach a final prediction.
• Information Extraction (IE) — Automatically identifying and extracting structured information from unstructured or semi-structured text.
• Information gain — A metric used in decision trees to measure the expected reduction in entropy achieved by splitting data on a particular feature.
• Information integration (II) — Merging data from various sources that may have different formats or meanings into a unified view.
• Information Processing Language (IPL) — An early programming language that introduced foundational concepts for list processing and dynamic memory allocation.
• Input generator — A component that processes raw data into the tensor format suitable for input into a neural network during training or inference.
• Input Layer — The first layer of a neural network, where raw features enter the model.
• Instruction Tuning — Fine-tuning a pretrained language model on examples of instructions paired with desired responses.
• Intelligence amplification (IA) — Using technology to enhance human intellectual capabilities and problem-solving skills.
• Intelligence explosion — A hypothetical rapid increase in artificial intelligence capabilities, potentially leading to superintelligence.
• Intelligent agent (IA) — An autonomous entity that perceives its environment, makes decisions, and acts to achieve goals.
• Intelligent Agents — Autonomous entities that perceive their environment and take actions to maximize their chances of achieving their goals.
• Intelligent control — Control systems that utilize AI techniques like neural networks and fuzzy logic for improved performance.
• Intelligent Document Processing (IDP) — AI-powered technology that automates the extraction, understanding, and processing of data from diverse document types.
• Intelligent personal assistant — Software that performs tasks or services for a user based on voice commands or text input.
• Intent Recognition — Identifying the underlying purpose or goal of a user's natural language input.
• Inter-rater agreement — A measure of the consistency or agreement between two or more human raters performing the same task.
• Interpretability — Understanding why a model makes the predictions it does, by inspecting its internals.
• Interpretation — Assigning meaning to the symbols of a formal language, making it understandable.
• Intersection over union (IoU) — A metric measuring the overlap between two bounding boxes, calculated as the intersection area divided by the union area.
• Intrinsic motivation — A drive in agents to act based on the inherent 'interestingness' or novelty of experiences, akin to curiosity.
• IoU — Abbreviation for Intersection over Union, a metric used in computer vision to measure overlap between bounding boxes.
• Issue tree — A graphical tool that breaks down a complex problem into smaller, manageable components for analysis.
• Item matrix — In recommender systems, a matrix containing latent feature representations (embeddings) for each item.
• Iteration — One update step of the model's parameters, processing a single batch.
• Jailbreak — Adversarial prompts that bypass an aligned model's safety training, eliciting outputs the model would normally refuse.
• JEPA (Joint Embedding Predictive Architecture) — Yann LeCun's self-supervised architecture that predicts abstract representations of masked regions instead of reconstructing raw pixels or tokens.
• Junction tree algorithm — An algorithm used in machine learning for efficient inference on graphical models by transforming them into trees.
• K-means clustering — K-means clustering is an algorithm that partitions data into k distinct clusters based on feature similarity.
• K-nearest neighbors — K-Nearest Neighbors (KNN) is a supervised learning algorithm used for classification and regression that predicts a data point's class based on its nearest neighbors.
• Kernel method — A class of machine learning algorithms, like SVMs, that use a kernel function to transform data for pattern analysis.
• KL-ONE — KL-ONE is a knowledge representation system that uses frames and structured inheritance networks to explicitly represent conceptual information and avoid ambiguity.
• Knowledge acquisition — Knowledge acquisition is the process of gathering and formalizing knowledge from domain experts to build knowledge-based systems.
• Knowledge Cutoff — The specific date or point in time beyond which an LLM's training data does not contain information, limiting its understanding of recent events.
• Knowledge Distillation — Compressing a large teacher model into a smaller student model by training the student to mimic the teacher's outputs.
• Knowledge engineering (KE) — Knowledge engineering (KE) encompasses all aspects of building, maintaining, and using knowledge-based systems.
• Knowledge extraction — Knowledge extraction creates machine-readable knowledge from structured or unstructured data, going beyond simple information retrieval.
• Knowledge Graph — A structured representation of knowledge as a network of entities and their relationships.
• Knowledge Graph Construction — The process of automatically building structured representations of information as a graph, connecting entities and their relationships.
• Knowledge Graphs — Structured knowledge bases representing real-world entities, their properties, and the relationships between them in a graph format.
• Knowledge Interchange Format (KIF) — Knowledge Interchange Format (KIF) is a language designed for sharing knowledge between different knowledge-based systems.
• Knowledge Management System (KMS) — A system used to store, organize, and retrieve information, often integrated with AI for quick access.
• Knowledge Representation — How facts and relationships are encoded so a system can reason over them.
• Knowledge representation and reasoning (KR² or KR&R) — Knowledge Representation and Reasoning (KR&R) is the AI field focused on how computers can store and use knowledge to solve problems.
• Knowledge-based system (KBS) — A knowledge-based system (KBS) is a program that uses a knowledge base and an inference engine to solve complex problems by reasoning.
• Kubernetes — An open-source platform for orchestrating containerized applications at scale.
• KV-Cache — A memory optimization storing previously computed key-value pairs during autoregressive generation to avoid redundant computation.
• Label — The correct answer attached to a training example in supervised learning.
• Language model — A language model is a statistical tool used to predict the probability of a sequence of words occurring in a natural language.
• Large Language Model (LLM) — A massive neural network trained on vast text corpora to understand and generate human language with remarkable fluency.
• Latency — The time between sending a request and receiving the first (or final) response.
• Latent Space — A compressed, lower-dimensional representation of data discovered by an unsupervised learning model.
• Layer — A group of neurons that perform the same kind of transformation in parallel.
• Layer Normalization — Standardizes activations across features within each sample, independent of batch size.
• Lazy learning — Lazy learning is a machine learning approach where model generalization is deferred until a query is made, contrasting with eager learning.
• Learning Rate — A hyperparameter that controls how large each parameter update step is during gradient descent optimization.
• Lemer — The base form of a word, often obtained through morphological analysis, ignoring inflections.
• Lethal autonomous weapon (LAW) — A Lethal Autonomous Weapon (LAW) is a weapon system capable of independently identifying and engaging targets without direct human control.
• Lexical Analysis — The initial stage of text processing where a stream of characters is converted into a sequence of tokens (words or meaningful units).
• Linguistic Rules — Explicit, human-defined rules that govern grammar, syntax, and semantics of language, used in some NLP systems.
• Lisp (programming language) (LISP) — Lisp is a family of high-level programming languages known for their distinctive prefix notation and extensive use in AI research.
• LLM Tooling — A suite of software tools and libraries designed to facilitate the development, deployment, and management of Large Language Models.
• Load Balancing — Distributing incoming requests across multiple servers to maximize utilization and reliability.
• Logic programming — Logic programming is a programming paradigm based on formal logic, where programs express facts and rules to solve problems.
• Logits — The raw, unnormalized scores a model outputs before they are converted to probabilities.
• Long Context — The ability of an LLM to process and maintain coherence over very long input sequences or conversations, beyond typical token limits.
• Long Short-Term Memory (LSTM) — An RNN variant with gating mechanisms that can learn long-range dependencies without suffering from vanishing gradients.
• LoRA (Low-Rank Adaptation) — A parameter-efficient fine-tuning method injecting small trainable matrices into frozen pre-trained layers.
• Loss Function — A mathematical function that measures how far the model's predictions are from the actual values, guiding the learning process.
• Low-Rank Adaptation (LoRA) — A parameter-efficient fine-tuning technique that reduces computational cost by only training a small number of new parameters.
• Machine Ethics — The field concerned with conferring moral agency or ethical behavior upon autonomous machines.
• Machine Learning — A field of AI where systems learn patterns from data instead of following hard-coded rules.
• Machine listening — The field focused on developing algorithms and systems for enabling machines to understand and interpret audio data.
• Machine perception — A computer system's ability to interpret data in a way that mimics human sensory perception and understanding of the world.
• Machine Reading Comprehension (MRC) — The ability of a system to read a text and answer questions about it, often without prior knowledge.
• Machine Translation — The automatic translation of text or speech from one natural language to another using AI.
• Machine vision (MV) — Technology for automatic inspection and analysis using imaging, often applied in industrial settings like manufacturing and robotics.
• Mamba — A selective state-space model (SSM) architecture that processes sequences in linear time with content-aware gating, rivaling Transformers on language tasks.
• Markov chain — A model where future states depend only on the current state, not on past events, used for sequences of events with probabilistic transitions.
• Markov decision process (MDP) — A framework for modeling decision-making in random environments where outcomes depend on both controllable actions and chance.
• Mathematical optimization — The process of finding the best solution from a set of available alternatives according to a specific criterion.
• Mean Absolute Error (MAE) — The average absolute difference between predictions and true values.
• Mean Squared Error (MSE) — The average squared difference between predictions and true values.
• Mechanism design — An approach to designing economic incentives and systems by working backward from desired outcomes, considering rational participants.
• Mechanistic Interpretability — A research program that reverse-engineers neural networks into human-understandable circuits, features, and algorithms.
• Mechatronics — An interdisciplinary engineering field combining mechanical, electrical, and computer systems for sophisticated product design.
• Memory (AI agents) — Persistent state that lets an AI agent retain information across turns or sessions.
• Metabolic network reconstruction and simulation — Modeling biological metabolic pathways to understand and simulate how organisms function at a molecular level.
• Metaheuristic — A high-level strategy that guides a search algorithm to find good solutions for complex optimization problems, especially with limited resources.
• Microservices — An architectural style where applications are built as small, independently deployable services.
• Mini-batch — A small subset of the training data used to compute one gradient update.
• Mixture of Experts (MoE) — An architecture routing inputs to specialized sub-networks, activating only a subset for each input to scale efficiently.
• MLOps — A set of practices combining Machine Learning, DevOps, and Data Engineering to reliably and efficiently deploy and maintain ML systems.
• MMLU — Massive Multitask Language Understanding — a 57-subject multiple-choice benchmark spanning humanities, STEM, law, and medicine.
• Moat — A strategic advantage or barrier that protects a company's AI technology or data from competitors, often involving proprietary data, models, or talent.
• Mode Collapse — A GAN failure where the generator produces only limited variety, ignoring the full diversity of training data.
• Model — A mathematical representation learned from data that maps inputs to outputs.
• Model Alignment — The process of training AI models, especially LLMs, to behave in accordance with human values, intentions, and ethical principles.
• Model Card — A document providing transparent information about an AI model, including its performance, limitations, ethical considerations, and intended use cases.
• Model checking — An automated technique for verifying if a system model meets specific requirements or properties, like absence of critical errors.
• Model Collapse — A hypothesized phenomenon where AI models continually trained on data generated by other AI models degrade in quality over generations.
• Model Compression — Techniques for shrinking models while preserving most of their performance.
• Model Context Protocol (MCP) — An open standard from Anthropic that lets LLM apps connect to external tools, data sources, and prompts through a uniform client–server interface.
• Model Deployment — The process of making a trained machine learning model available for use in a production environment.
• Model Drift — The degradation of a model's performance over time due to changes in the underlying data distribution.
• Model explainability — The ability to understand and interpret how an AI model arrives at its decisions or predictions, making its operations transparent.
• Model Grafting — The technique of attaching a specialized smaller model or 'head' to a larger foundational model to adapt it for a specific task.
• Model Hallucinations — A more specific term for the phenomenon where AI models generate plausible but incorrect or fabricated information.
• Model Interpretability — The degree to which a human can understand the causality of an AI model's prediction.
• Model Merging — Combining the weights of several fine-tuned checkpoints into a single model with no extra training, often via averaging in parameter space.
• Model Monitoring — Continuously tracking the performance and behavior of deployed AI models in production.
• Model Observability — The ability to monitor, understand, and troubleshoot the internal workings and predictions of an AI model in production.
• Model Quantization — A technique to reduce the memory footprint and computational cost of AI models by representing their weights and activations with lower precision numbers.
• Model Scaling Laws — Empirical observations indicating how AI model performance improves predictably with increases in model size, dataset size, and computational budget.
• Model Serving — Hosting a trained model behind an interface so applications can request predictions in real time.
• Modus ponens — A basic rule of logic stating that if a conditional statement is true, and its antecedent (hypothesis) is true, then its consequent (conclusion) must also be true.
• Modus tollens — A logical rule where if a statement implies another, and the second statement is false, then the first statement must also be false.
• Monte Carlo tree search — A heuristic search algorithm used in decision processes, particularly in games, that balances exploration and exploitation using random sampling.
• Morphological Analysis — The process of analyzing the internal structure of words and their constituent morphemes.
• Multi-agent system (MAS) — A system composed of multiple autonomous agents that interact to solve problems or achieve goals that individual agents cannot.
• Multi-Agent Systems (MAS) — Systems composed of multiple interacting intelligent agents that cooperate or compete to achieve individual or collective goals.
• Multi-Modal AI — AI systems that process and reason across multiple data types — text, images, audio, video — in a unified model.
• Multi-swarm optimization — An optimization technique using multiple subgroups (sub-swarms) to search different regions of a problem space.
• Multilayer perceptron (MLP) — A feedforward neural network with multiple layers of fully connected neurons and nonlinear activation functions.
• Multimodal AI — AI systems capable of processing and understanding information from multiple data types, such as text, images, audio, and video, simultaneously.
• Multimodal LLM — An LLM extended to process and understand multiple types of data, such as images, audio, and text, beyond just language.
• Multimodal Model — A model that processes and/or generates more than one modality — for example text, images, audio, and video — within a single architecture.
• Multivariate Analysis — Statistical techniques used to analyze data involving more than one dependent variable at a time.
• Mutation — A genetic algorithm operator that introduces random changes to chromosomes to maintain diversity and explore new solutions.
• Mycin — An early AI expert system for diagnosing infections and recommending antibiotic treatments.
• N-gram Model (for Text) — A statistical language model that predicts the next word in a sequence based on the preceding N-1 words.
• Naive Bayes classifier — A simple probabilistic classifier based on Bayes' theorem with a strong assumption of feature independence.
• Naive semantics — An AI approach representing domain knowledge using a limited set of generally understood concepts.
• Name binding — In programming, the association of an identifier (name) with a specific piece of data or code.
• Named Entity Recognition (NER) — An NLP task that identifies and classifies named entities like people, organizations, and locations in text.
• Named graph — A Semantic Web concept where a set of RDF statements (a graph) is uniquely identified by a URI.
• Natural language generation (NLG) — The process of transforming structured data into human-readable text.
• Natural language processing (NLP) — A field of AI enabling computers to understand, interpret, and generate human language.
• Natural language programming — Programming a computer using natural language sentences instead of traditional code.
• Natural Language Understanding (NLU) — The ability of a computer program to understand human language as it is spoken or written.
• Network motif — Recurring and statistically significant subgraph patterns found in complex networks.
• Neural Architecture Search (NAS) — An automated technique for designing optimal neural network architectures.
• Neural machine translation (NMT) — Using neural networks to translate entire sentences as a single model.
• Neural Network — A computing system inspired by biological neural networks that learns patterns from data through interconnected layers of nodes.
• Neural Turing machine (NTM) — A neural network combined with external memory, capable of learning algorithms.
• Neuro-fuzzy — An AI approach that combines artificial neural networks with fuzzy logic.
• Neuro-Symbolic AI — A hybrid AI approach combining connectionist (neural networks) and symbolic (rule-based) methods for enhanced reasoning and learning.
• Neuro-Symbolic AI — Hybrid systems that combine neural networks for perception/learning with symbolic methods for reasoning, planning, and verification.
• NeuroAI — An interdisciplinary field combining neuroscience and artificial intelligence to build more brain-like AI and understand biological intelligence.
• Neurocybernetics — A direct communication pathway between a brain and an external device, allowing for bidirectional information flow.
• Neuromorphic Computing — A computing paradigm that mimics the brain's neurobiological structure and operation to process information.
• Neuromorphic engineering — Engineering that mimics the structure and function of biological neural networks using electronic circuits.
• Node — A fundamental component of data structures like trees, containing data and links to other nodes.
• Non-monotonic Logic — A logic system where adding new information can reduce the set of derivable conclusions, in contrast to classical logic.
• Nondeterministic algorithm — An algorithm that can produce different outputs for the same input on different runs.
• Nouvelle AI — An AI approach aiming for insect-level intelligence by emergent behaviors from simple interactions.
• NP — A complexity class for decision problems where 'yes' answers can be verified quickly.
• NP-completeness — The hardest problems in NP, where checking a solution is fast, but finding it is generally hard.
• NP-hardness — Problems that are at least as hard as the most difficult problems in the NP complexity class.
• Nucleus Sampling (top-p) — A decoding method that samples from the smallest set of tokens whose cumulative probability exceeds p, balancing diversity and coherence.
• Object Detection — A computer vision task that identifies and localizes multiple objects in an image with bounding boxes.
• Occam's razor — A principle favoring simpler explanations when multiple hypotheses explain the same phenomenon.
• Offline learning — Training a model on a fixed dataset without further updates during the learning process.
• Offline Processing — Computing predictions or transforming data in batches, without strict latency requirements.
• On-device AI — AI models that run directly on edge devices (e.g., smartphones, IoT) rather than relying on cloud servers for inference.
• One-Shot Learning — A machine learning scenario where a model learns to recognize new categories after observing only one training example.
• Online machine learning — A machine learning method that updates a model incrementally as new data arrives sequentially.
• Ontology — A formal specification of the concepts, properties, and relationships in a domain.
• Ontology learning — Automatically extracting and structuring knowledge, concepts, and relationships from text.
• Open Mind Common Sense — A project collecting commonsense knowledge from the public to build a large knowledge base.
• Open-source Model — An AI model whose code, weights, and sometimes training data are publicly accessible, allowing for inspection, modification, and redistribution.
• Open-source software (OSS) — Software with source code released under a license granting rights to study, change, and distribute it freely for any purpose.
• OpenCog — An open-source project aiming to build an AI framework for achieving human-equivalent artificial general intelligence.
• Optical Character Recognition — A technology that enables the conversion of different types of documents, such as scanned paper documents, into editable and searchable data.
• Optimization — The mathematical process of finding parameter values that minimize a loss function.
• Optimizer — An algorithm that updates model weights during training to minimize the loss function, with strategies beyond basic gradient descent.
• Orchestration — The process of coordinating and managing multiple AI models, tools, and data flows to achieve a larger, more complex AI application or workflow.
• Output Layer — The final layer of a neural network that produces the prediction.
• Overfitting — When a model learns noise and specific patterns in training data too well, causing it to perform poorly on new, unseen data.
• PagedAttention — A memory-management technique used in vLLM that stores the KV-cache in non-contiguous fixed-size pages, eliminating fragmentation and enabling sharing.
• Parallelism (Data / Model) — Strategies for splitting computation across devices to train or serve large models.
• Parameter — A configuration variable of the model that is learned from the training data.
• Parameter-Efficient Fine-Tuning (PEFT) — A family of techniques that adapt large pre-trained models to new tasks by updating only a small subset of their parameters, saving computation.
• Parameters — The internal variables — weights and biases — that a model learns from data.
• Paraphrase Detection — Identifying if two different phrases or sentences convey the same meaning.
• Partial order reduction — A technique to reduce the search space in model checking or planning by exploiting the commutativity of concurrent transitions.
• Partially observable Markov decision process (POMDP) — An extension of MDPs where the agent cannot directly observe the state, but must infer it from observations.
• Particle swarm optimization (PSO) — An optimization algorithm where particles move through a search space, influenced by their own best-known positions and the swarm's best.
• Pathfinding — The process of automatically determining the shortest or most efficient route between two points, often in a graph or map.
• Pattern Matching — The process of comparing inputs to a stored set of patterns to identify correlations or similarities.
• Pattern recognition — The automatic identification of regularities and patterns in data using algorithms for tasks like classification.
• PEFT (Parameter-Efficient Fine-Tuning) — A family of methods that adapt large pretrained models by training only a small fraction of parameters, drastically cutting compute and storage.
• Perceptive AI — AI systems endowed with advanced sensory capabilities to interpret and understand the physical world through various inputs.
• Perceptron — A simple algorithm for supervised learning of binary classifiers, forming a basic type of artificial neuron.
• Perplexity — A metric measuring how well a language model predicts text — lower means less 'surprised' by the data.
• Personalized AI — AI systems that adapt and customize their behavior, recommendations, or content to individual user preferences and needs.
• Personalized Learning — An educational approach using AI to tailor learning experiences, content, and pace to individual student needs and preferences.
• Planning — Deciding on a sequence of actions to achieve a goal under given constraints.
• Policy — In reinforcement learning, the strategy an agent follows to choose actions given a state.
• Policy Gradient — RL algorithms that directly optimize the policy function by gradient ascent on expected reward.
• Pooling — An operation that reduces spatial dimensions of feature maps by summarizing local regions.
• Positional Encoding — A mechanism that injects sequence order information into Transformers, which otherwise have no notion of position.
• Pre-trained Model — An AI model that has already been trained on a very large dataset for a general task.
• Pre-training — The initial phase of training a large model on a massive dataset, often in an unsupervised manner, to learn general representations.
• Precision — Of the items the model predicted positive, the fraction that are actually positive.
• Precision & Recall — Complementary metrics measuring classifier accuracy on positive predictions (precision) and ability to find all positives (recall).
• Predicate logic — A formal system that uses quantified variables and predicates to express logical propositions about objects and their properties.
• Prediction — The output a trained model produces when given new input data.
• Predictive analytics — Uses current and historical data with data mining and machine learning techniques to forecast future events or outcomes.
• Pretraining — The first, expensive training stage where a model learns general patterns from massive unlabeled data.
• Principal component analysis (PCA) — A dimensionality reduction technique that transforms correlated variables into uncorrelated principal components, prioritizing variance.
• Principle of rationality — Posits that agents act in the most adequate way possible given the objective circumstances of their situation.
• Privacy — Protecting personal information that flows into or out of an AI system.
• Privacy-Preserving AI — Techniques and methodologies that enable AI development and deployment while safeguarding sensitive personal data.
• Probabilistic programming (PP) — A programming paradigm that unifies probabilistic modeling with general-purpose programming for easier specification and inference.
• Production system — An AI program consisting of rules and an inference engine to execute them, enabling the system to respond to world states.
• Productivity AI — AI tools and features designed to enhance user efficiency, automate tasks, and streamline workflows across various applications and platforms.
• Programming language — A formal language consisting of instructions used to create programs that produce output on a computer.
• Prolog — A declarative logic programming language widely used in AI and computational linguistics, based on facts and rules.
• Prompt — The input text — instructions, context, examples — given to a language model to elicit a response.
• Prompt Chaining — A technique where multiple prompts are linked sequentially, with the output of one prompt serving as the input for the next, to achieve a complex goal.
• Prompt Compression — Techniques used to shorten or condense lengthy input prompts for LLMs while retaining essential information and instructions.
• Prompt Engineering — The art of crafting effective input instructions to guide LLM behavior without changing model weights.
• Prompt Injection — An attack where untrusted input (a webpage, email, tool output) contains instructions that hijack an LLM's behavior in an agentic context.
• Prompt Templates — Pre-defined structures or patterns for prompts that guide LLMs to perform specific tasks or generate particular output formats.
• Prompt Tuning — A parameter-efficient fine-tuning technique that learns a small, task-specific soft prompt (a sequence of learnable tokens) to steer a frozen LLM.
• Prompt Versioning — The practice of tracking and managing different versions of prompts used to interact with LLMs, similar to code version control.
• Propositional calculus — A branch of logic dealing with true/false propositions and how they connect, without focusing on objects or quantifiers.
• Proprietary Data — Data owned by a company or individual, not publicly available, and often a key competitive asset.
• Proprietary Model — An AI model, often highly performant, developed and owned by a specific entity, with its internal workings and weights kept confidential.
• Proximal policy optimization (PPO) — A reinforcement learning algorithm that trains agents by making small, consistent updates to their decision policy.
• Pruning — Removing redundant weights or neurons from a network to reduce size and improve inference speed.
• Python — A versatile, high-level programming language known for its readability and clear code structure.
• PyTorch — An open-source machine learning library, originally from Meta AI, widely used for deep learning tasks like computer vision and NLP.
• Q-Learning — A model-free RL algorithm that learns action values without knowing environment dynamics.
• QLoRA — Fine-tuning method that adds LoRA adapters on top of a 4-bit quantized base model, letting 65B LLMs be trained on a single 48 GB GPU.
• Qualification problem — The AI challenge of listing all necessary preconditions for an action to succeed in complex real-world scenarios.
• Quantifier — A logical symbol indicating the quantity of elements in a set that satisfy a condition (e.g., 'for all', 'there exists').
• Quantization — Reducing numerical precision of model weights (e.g., 32-bit to 4-bit) to shrink size and speed up inference.
• Quantum computing — A computation method using quantum mechanics principles like superposition and entanglement to solve complex problems.
• Query language — A specialized computer language designed to retrieve information from databases or information systems.
• Question Answering (QA) — A system's ability to automatically answer questions posed in natural language.
• R programming language — A free software environment and programming language popular for statistical computing and data visualization.
• Radial basis function network — A type of artificial neural network that uses radial basis functions as its activation functions.
• Random forest — An ensemble machine learning method that builds multiple decision trees to improve prediction accuracy and control overfitting.
• Random Search — Sampling hyperparameter configurations at random instead of testing every combination.
• Real-time Processing — Handling data and producing results within strict latency bounds, often milliseconds.
• Reasoning — An AI system's ability to draw logical, multi-step conclusions from given information.
• Reasoning engine — A component of a software system that infers conclusions from existing knowledge using logical methods.
• Reasoning model — A specialized large language model trained to perform multi-step logical reasoning for complex tasks.
• Reasoning system — A software system that uses logical techniques to draw conclusions from available knowledge.
• Recall — Of all the actually positive items, the fraction the model successfully found.
• Recurrent Neural Network (RNN) — A neural network with loops that maintain hidden state, designed to process sequential data like text and time series.
• Recursive self-improvement — The process where an AI system enhances its own capabilities by rewriting its code, potentially leading to superintelligence.
• Red Teaming — Systematically probing an AI system for failures, exploits, and harmful behavior before deployment.
• Reflexion — An agent technique where an LLM reflects on failed attempts and writes verbal self-feedback into memory to guide future trials.
• Regression — A supervised learning task that predicts a continuous numeric value.
• Regression analysis — Statistical method to estimate relationships between a dependent variable and independent variables.
• Regularization — Techniques that constrain a model's complexity to prevent overfitting and improve generalization to unseen data.
• Reinforced Prompting — An iterative prompting technique where an LLM's response is used as feedback to refine subsequent prompts, guiding it towards better outcomes.
• Reinforcement Learning — A paradigm where an agent learns to make decisions by receiving rewards or penalties from its environment through trial and error.
• Reinforcement Learning Agent — The entity that makes decisions and performs actions in an environment to maximize cumulative rewards.
• Reinforcement learning from human feedback (RLHF) — Training AI using human feedback to guide model behavior and improve output quality.
• Relation Extraction — Identifying and classifying semantic relationships between named entities in text.
• ReLU — A simple activation function that outputs the input if positive and zero otherwise.
• Representation Learning — The process of automatically discovering meaningful representations of data from raw inputs.
• Reproducibility — The ability for others to re-run an experiment and obtain the same results.
• Reranker — A second-stage model (often a cross-encoder) that re-orders an initial set of retrieved candidates for higher precision before they reach an LLM.
• Reservoir computing — A computing framework using fixed, complex dynamical systems to process input signals.
• Residual Connection (Skip Connection) — A shortcut that adds a layer's input directly to its output, enabling training of very deep networks.
• ResNet (Residual Network) — A type of convolutional neural network that uses 'skip connections' to allow for training much deeper models.
• Resource Description Framework (RDF) — A standard model for describing resources and their relationships on the web.
• Response Time — The duration an AI system takes to process a user's input and generate a reply.
• Responsible AI (RAI) — A holistic framework encompassing the ethical, legal, and societal implications of AI, promoting trustworthy and beneficial systems.
• Restricted Boltzmann machine (RBM) — A generative neural network that learns probability distributions over input data.
• Rete algorithm — An efficient pattern-matching algorithm for rule-based systems and AI applications.
• Retrieval augmented generation (RAG) — LLMs that can access and use external information to improve their responses.
• Retrieval Model — An AI model that answers queries by finding and presenting existing relevant information from a knowledge base or dataset.
• Retrieval Pipeline — A sequence of steps and components used to efficiently retrieve relevant information from a knowledge source, often for RAG systems.
• Retrieval-Augmented Generation (RAG) — A technique that enhances LLM responses by retrieving relevant documents from an external knowledge base before generating an answer.
• Reward — A feedback signal in reinforcement learning, indicating the desirability of an agent's action in a given state.
• Reward Function — A scalar signal that tells a reinforcement-learning agent how good its action was.
• Reward Model — A specialized AI model, often used in RLHF, that learns to predict human preferences or scores for different LLM outputs.
• Reward Shaping — Designing or modifying the reward signal to guide an RL agent toward desired behavior more efficiently.
• RLAIF (Reinforcement Learning from AI Feedback) — A technique similar to RLHF, but where the reward signal for training the language model comes from another AI model rather than humans.
• RLHF (Reinforcement Learning from Human Feedback) — A technique that aligns LLM outputs with human preferences by training a reward model from human comparisons.
• RMSNorm — A normalization layer that rescales activations by their root-mean-square, dropping the mean-centering step of LayerNorm for ~10% speed-up.
• Robot Learning — The field dedicated to enabling robots to acquire new skills and adapt their behavior through experience and instruction.
• Robotics — The interdisciplinary field focused on designing, building, and operating robots.
• Robotics Process Automation (RPA) — Software technology that uses 'software robots' to automate repetitive, rule-based digital tasks and streamline business processes.
• Robustness — The ability of an AI model to maintain its performance and reliability even when faced with variations or perturbations in input data.
• ROC Curve — A plot of true-positive rate against false-positive rate across all classification thresholds.
• RoPE (Rotary Position Embedding) — A positional encoding that rotates query and key vectors by an angle proportional to position, giving relative-position behavior with no extra parameters.
• ROUGE Score — A family of metrics that measure overlap between generated and reference text, used mainly for summarization.
• Rule-based system — AI systems that use human-defined rules to store and manipulate knowledge.
• Safety — Designing AI systems to behave reliably and avoid causing harm under realistic conditions.
• Sampling — The process of choosing the next token from the probability distribution a model outputs.
• Sampling (data) — Selecting a subset of records from a larger dataset for training, evaluation, or analysis.
• Satisfiability — A concept in logic determining if a formula can be made true by assigning values to its variables.
• Scalability — A system's ability to handle increasing load by adding resources.
• Scaling Laws — Empirical power-law relationships predicting how LLM loss decreases as model size, dataset size, and compute increase together.
• Search algorithm — A process for finding information or solutions within data or a problem space.
• Selection — Choosing individuals from a population for reproduction in genetic algorithms.
• Self-Attention — A mechanism where every token in a sequence attends to every other token to compute context-aware representations.
• Self-Consistency — A decoding strategy that samples multiple chain-of-thought reasoning paths and returns the majority-voted answer for improved accuracy.
• Self-correction — The ability of an AI model to evaluate its own outputs, identify errors or inefficiencies, and adjust its subsequent actions or reasoning.
• Self-management — Systems that autonomously monitor and control their own operations.
• Self-Supervised Learning — A training paradigm that generates supervisory signals from the data itself, eliminating the need for human labels.
• Semantic Analysis — The process of interpreting the meaning of words, phrases, and sentences in natural language.
• Semantic Caching — A caching strategy for LLM interactions that stores and reuses responses for semantically similar prompts, reducing latency and API costs.
• Semantic network — A graph-based knowledge representation showing concepts and their relationships.
• Semantic reasoner — Software that infers logical conclusions from asserted facts and axioms.
• Semantic Search — A search technology that understands the intent and contextual meaning of queries, rather than just matching keywords.
• Semantic Similarity — Measuring how closely two words, phrases, or documents are related in meaning.
• Semantic Understanding — Grasping the meaning of language, not just its surface form.
• Semantics — The study of meaning in language, focusing on how words, phrases, and sentences convey sense.
• Semi-supervised Learning — A hybrid approach that uses a small amount of labeled data alongside a large amount of unlabeled data.
• Sentiment Analysis — The task of determining the emotional tone or opinion expressed in text — positive, negative, or neutral.
• Sequence-to-Sequence (Seq2Seq) — An architecture that maps an input sequence to an output sequence of potentially different length.
• Serverless — A deployment model where the cloud provider runs your code on demand without managing servers.
• Sharding — Splitting a large dataset or model across multiple machines or storage devices.
• Shared Autonomy — A control paradigm where both a human operator and an autonomous system contribute to task execution, often blending control.
• Sigmoid — An S-shaped activation function that squashes any input into the range (0, 1).
• Similarity Search — Finding the items in a dataset most similar to a given query, usually by vector distance.
• Simulation — Synthetic environments that let agents or models train and be evaluated without real-world risk.
• Singularity — A hypothetical future point when technological growth becomes uncontrollable and irreversible, resulting in unfathomable changes to human civilization.
• Slot Filling — The process of extracting specific pieces of information (slots) from a user's utterance needed to fulfill an intent.
• Small Language Models (SLMs) — Compact language models that run efficiently on-premise, on edge devices, or on a laptop.
• Small Talk — Conversational AI's ability to engage in casual, non-goal-oriented dialogue, mimicking human social interaction.
• Social Robotics — The study and development of robots that can interact and communicate with humans in a natural, socially intelligent manner.
• Softmax Function — A function that converts a vector of raw scores into a probability distribution where all values sum to one.
• Sora — OpenAI's text-to-video diffusion-transformer model that generates minute-long high-resolution clips with strong temporal coherence and 3D consistency.
• Sovereign AI — Building and operating in-house AI infrastructure and models to keep full control of data and reduce dependence on hyperscalers.
• Sparse Attention — Attention mechanisms attending to only a subset of positions, reducing quadratic complexity.
• Sparse Autoencoder (SAE) — An overcomplete autoencoder with an L1 penalty used in interpretability to decompose neural activations into monosemantic features.
• Sparse Mixture of Experts (SMoE) — A neural network architecture where a 'router' network activates only a few specialized 'expert' subnetworks for each input, improving efficiency.
• Speculative Decoding — An inference technique where a small draft model proposes several tokens which a large target model verifies in parallel, accelerating generation.
• Speech Recognition — The ability of a machine to identify spoken words and convert them into human-readable text.
• Speech-to-Text — Automatic transcription of spoken audio into written text.
• State / Action — The two core elements of every reinforcement-learning step: what the world looks like and what the agent does.
• State Tracking — The process of monitoring and updating the current state of a conversation and user goals.
• Streaming — Returning model output incrementally — token by token — as it is generated.
• Structured Output — Constraining an LLM's generation to a defined schema (JSON, regex, grammar) so downstream code can parse it reliably.
• Supervised Learning — Learning from input–output pairs where each training example carries a correct label.
• Supervised Machine Learning — A subset of ML where models learn from labeled data to predict outcomes.
• Swarm Intelligence — An AI technique inspired by the collective behavior of decentralized, self-organized systems in nature, such as ant colonies.
• SwiGLU — A gated linear-unit activation, x · Swish(W₁x) ⊙ (W₂x), that consistently outperforms ReLU and GELU in Transformer FFN blocks.
• Symbolic AI — An approach to AI that attempts to represent human knowledge explicitly using symbols and rules.
• Syntactic Analysis (Parsing) — The process of analyzing natural language to determine its grammatical structure (syntax).
• Syntactic Dependency Parsing — Analyzing the grammatical relationships between words in a sentence, represented as a dependency tree.
• Synthetic Agents — AI programs designed to mimic human or other intelligent entities in simulations or interactive environments for testing and research.
• Synthetic Data — Artificially generated data mimicking real-world properties, used for training augmentation or privacy protection.
• Synthetic Data Generation — The process of computationally creating artificial data that mimics the statistical properties and patterns of real-world data.
• Synthetic Speech — Speech audio that is artificially generated by AI models, rather than recorded from a human voice.
• Synthetic Users — Automated agents or AI models designed to simulate user interactions and behaviors for testing, validating, or iterating on AI systems.
• System 2 Reasoning — A concept from cognitive science applied to AI, referring to slow, deliberate, logical, and effortful thought processes.
• System Prompt — A high-priority instruction that sets a model's role, tone, and constraints for an entire conversation.
• Tanh — An activation function that maps inputs to the range (-1, 1), centered at zero.
• Taxonomy — A hierarchical classification system used to organize information or objects into categories and subcategories.
• Temperature (Sampling) — A parameter controlling output randomness — lower values are more focused, higher values more creative.
• Tensor — A multidimensional array used to represent data in many dimensions, fundamental to deep learning.
• Term Extraction — Automatically identifying relevant single or multi-word terms from a corpus of text.
• Test Data — A separate, untouched dataset used only at the end to estimate real-world performance.
• Test-Time Compute — Spending more inference compute (longer chains of thought, sampling, search) to improve answers, often outperforming a larger model.
• Text Analytics — The process of deriving high-quality information from text to discover patterns and insights.
• Text Classification — The task of categorizing text documents into predefined classes or labels.
• Text Mining — The process of extracting high-quality information from text by identifying patterns and trends.
• Text Summarization — The process of condensing a larger text into a shorter, coherent summary that retains the main points.
• Text-to-Image — A generative task that produces images from natural-language descriptions.
• Text-to-Speech — A model that converts written text into natural-sounding spoken audio.
• Threshold (AI) — A predefined level that an AI model's confidence score must meet to trigger a specific action.
• Throughput — The number of requests or tokens a system can process per unit of time.
• Time Series Forecasting — Using historical data points to predict future values in a sequence over time.
• Token — The atomic unit of text a language model processes — typically a word piece or subword.
• Tokenization — The process of breaking text into smaller units (tokens) that language models can process as numerical inputs.
• Tokenizer — A component in NLP that breaks down raw text into smaller units called tokens, such as words or subwords.
• Tool Calling — An LLM capability where the model emits structured calls to external functions or APIs.
• Tool Use — The ability of an AI model, especially an LLM agent, to autonomously identify, select, and utilize external tools and APIs to accomplish tasks.
• Top-k Sampling — A decoding strategy that restricts the next-token choice to the k most likely tokens.
• Top-p (Nucleus) Sampling — A decoding strategy that samples from the smallest set of tokens whose cumulative probability exceeds p.
• Topic Modeling — Statistical models for discovering the abstract 'topics' that occur in a collection of documents.
• TPU — Tensor Processing Unit — Google's custom-designed AI accelerator chip.
• Training — The process of adjusting a model's parameters so it learns patterns from labeled or unlabeled data.
• Training Data — The portion of a dataset used to fit a model's parameters.
• Transfer Learning — Leveraging knowledge from a model trained on one task to improve performance on a different but related task.
• Transformer — An architecture that uses self-attention to process sequences in parallel, powering modern language models like GPT and BERT.
• Transformer Architecture — A neural network architecture, predominantly used in NLP, that relies heavily on self-attention mechanisms to process sequential data.
• Transparency — Openly disclosing how a model was built, what data it used, and how it behaves.
• Tree of Thought (ToT) — A prompting technique that enables LLMs to explore multiple reasoning paths by maintaining and evaluating intermediate 'thought' steps in a tree structure.
• Tree of Thoughts (ToT) — A prompting strategy that explores multiple reasoning branches as a search tree, evaluating intermediate thoughts and backtracking when needed.
• Trusted AI — A quality standard — often certified — proving that an AI system is secure, transparent, and robust.
• Trustworthy AI — A broad concept encompassing AI systems that are reliable, secure, transparent, fair, and accountable, adhering to ethical principles.
• Turing Test — A test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.
• Type system — Rules that assign a property called 'type' to programming constructs like variables and functions to prevent bugs.
• Underfitting — A phenomenon where a model is too simple to capture the underlying patterns in the training data, resulting in poor performance on both training and new data.
• Unstructured Data — Information that does not have a predefined data model or is not organized in a predefined manner.
• Unsupervised Learning — Learning patterns from data that has no labels — only the inputs.
• Unsupervised Machine Learning — A type of ML where models find patterns in unlabeled data without explicit guidance.
• Utterance — A complete unit of spoken or written language from a user, representing a single turn in a conversation.
• Validation Data — A held-out subset used to tune hyperparameters and detect overfitting during training.
• Value-alignment complete — A problem analogous to AI-complete, requiring full AI control solutions for resolution.
• Variance — The sensitivity of a model to small fluctuations in the training data, leading to inconsistent predictions on new data.
• Variational Autoencoder (VAE) — A generative model that learns a smooth, probabilistic latent space enabling meaningful interpolation and data generation.
• Vector Database — A specialized database optimized for storing, indexing, and querying high-dimensional embedding vectors using similarity search.
• Vector Embeddings — Numerical representations of text, images, or other data that capture semantic meaning, allowing for mathematical operations like similarity comparisons.
• Vector Representation — Encoding objects (words, images, users) as numerical vectors so machines can compute similarity and patterns.
• Vector Space Model — A mathematical model for representing text documents as vectors of identifiers, such as index terms.
• Vertical AI — AI solutions tailored and optimized for specific industries or business domains.
• Virtual Agent — An AI-powered digital assistant that interacts with users to answer questions and complete tasks.
• Virtual Assistant — An AI-powered software agent that can perform tasks or services for an individual based on commands or questions.
• Vision processing unit (VPU) — Specialized hardware designed to speed up computer vision tasks in AI.
• Vision Transformer (ViT) — An architecture that applies the Transformer to images by splitting them into patches and processing them as sequences.
• Voice Agent — An AI system that interacts with users through spoken language, often used in customer service.
• Voice Cloning — Synthesizing speech in a target speaker's voice from a short reference sample, often just a few seconds long.
• Watson — IBM's natural language question-answering system, developed to understand and respond to complex queries.
• Weak AI — AI designed and trained for a specific, narrow task, often called narrow AI.
• Weights — The learnable numerical parameters that determine how a neural network transforms its inputs.
• Whisper — OpenAI's open-source speech recognition model trained on 680k hours of multilingual audio, handling transcription, translation, and language ID.
• Word embedding — Vector representations of words where similar meanings correspond to nearby vectors.
• Word Sense Disambiguation (WSD) — The computational process of determining the correct meaning of a word when it has multiple possible meanings based on context.
• Word2Vec — A pioneering technique that learns dense word embeddings by predicting surrounding words from large text corpora.
• Work-as-a-Service — A business model where customers pay for delivered outcomes instead of software seats or subscriptions.
• World model — A neural network simulating real-world dynamics, including physics, to generate realistic environments.
• XGBoost — An optimized distributed gradient boosting library designed for speed and performance.
• Zero-Shot Learning — The ability of a model to perform tasks or classify categories it has never explicitly been trained on.