Glossary

Plain definitions for the terms behind building, evaluating, and running AI systems

Most readTrendingLLM app platform Large language model (LLM)LLM evaluation Retrieval-augmented generation (RAG)Mixture of experts (MoE)LLM-as-a-Judge

A

Abductive logic programming (ALP)LogicA form of logic programming that generates hypotheses from rules and data, then tests them against the data to find the most plausible explanation.

Abductive reasoningLogicA form of logical inference that picks the most likely explanation for the available information, yielding a plausible conclusion rather than a certain one.

Abstract data type (ADT)Computer scienceA mathematical model of a data type, defined by its set of values and the operations allowed on them, independent of how it is implemented.

AbstractionComputer scienceThe process of simplifying complexity by focusing on essential features and hiding irrelevant details, applied in reasoning, learning, and software design.

Accelerating changePhilosophyThe observation, popularized by Ray Kurzweil, that the rate of technological progress itself increases over time, producing exponential rather than linear growth.

Action languageLogicA formal language for describing actions, their preconditions, and their effects so that agents can reason, plan, and act in an environment.

Action model learningMachine learningLearning how actions change the world by observing an agent act, usually captured in a logic-based action description language.

Action selectionAgentsThe process by which an AI agent chooses what to do next from a set of alternatives, based on its current state and environment.

Activation functionMachine learningA mathematical function applied to a neural network node's input to produce its output, adding the non-linearity needed to learn complex patterns.

Adaptive algorithmComputer scienceAn algorithm that adjusts its behavior or parameters as the environment or the data it processes changes.

Adaptive neuro-fuzzy inference system (ANFIS)Machine learningA hybrid model that combines neural networks and fuzzy logic in one inference system, learning from data while handling uncertainty.

Admissible heuristicComputer scienceA heuristic function that never overestimates the cost of reaching the goal, so its estimate is at most the true lowest cost.

Adversarial attacks and defensesSafetyTechniques that manipulate an AI system's inputs or training data to cause errors, along with the methods used to detect and resist them.

Affective computingComputer scienceThe study and development of systems that can recognize, interpret, process, and simulate human emotions to respond to users more naturally.

Agent architectureAgentsThe organizational structure of components inside a software agent, covering how it handles perception, reasoning, learning, and action.

Agent human-in-the-loop (HITL)AgentsA design pattern that pauses an AI agent at defined checkpoints so a human can approve, reject, modify, or redirect an action before it takes effect.

Agent memory architecturesAgentsThe systems that let autonomous AI agents keep context, state, and continuity over time instead of starting fresh on every call.

Agentic RAGRetrievalRetrieval-augmented generation in which an agent, not a fixed pipeline, plans the searches, evaluates the results, and retries when they fall short.

AgentsAgentsEntities that perceive their environment through sensors and act on it autonomously through actuators to achieve their goals.

AI acceleratorHardwareSpecialized hardware, also called a neural processing unit, designed to speed up artificial intelligence and machine learning workloads.

AI alignmentSafetyThe area of AI safety research focused on steering AI systems toward humans' intended goals, preferences, and values.

AI analyticsDataThe use of artificial intelligence and machine learning to automate data analysis across structured and unstructured data.

AI content moderationSafetyThe use of machine learning and natural language processing to automatically filter, review, and moderate user-generated content.

AI copilotsToolsVirtual assistants powered by large language models that boost productivity by automating tasks and supporting decisions with context-aware help.

AI darksideSafetyThe unethical use of artificial intelligence for harmful purposes, such as deepfakes, misinformation, privacy violations, and biased decisions.

AI engineerPeopleA professional who builds, programs, and trains the algorithms and networks behind artificial intelligence systems.

AI ethicsPhilosophyThe branch of ethics concerned with the moral issues raised by artificial intelligence and by how humans design, use, and treat AI systems.

AI gateway (LLM router)LLMOpsA middleware layer between an application and multiple LLM providers that handles routing, authentication, formatting, and monitoring behind one interface.

AI governanceSafetyThe principles, frameworks, and legal structures that ensure AI is used responsibly, managing risk and keeping deployment ethical and transparent.

AI hardwareHardwareSpecialized computational devices, such as GPUs, TPUs, and NPUs, that accelerate the processing demands of artificial intelligence.

AI in metabolic network reconstruction and simulationMachine learningThe use of AI, especially machine learning, to compile and model the biochemical reactions of an organism's metabolism, a core task in systems biology.

AI privacySafetyThe challenges of collecting, storing, and using personal data in AI systems, including the risk of privacy breaches and misuse.

AI product managerPeopleA professional who guides the development, launch, and improvement of products or features powered by artificial intelligence or machine learning.

AI quality controlLLMOpsThe practices used to assess and maintain the quality of an AI system across its lifecycle, from data collection to deployment and monitoring.

AI red-teaming and jailbreak testingSafetySystematic adversarial testing of an AI system, especially a language model, to find vulnerabilities and unintended behaviors before attackers do.

AI safetySafetyAn interdisciplinary field focused on preventing accidents, misuse, and other harms that could result from artificial intelligence systems.

AI teamPeopleA multidisciplinary group that combines varied expertise to develop, deploy, and manage AI-driven solutions.

AI winterComputer scienceA period of reduced interest, funding, and development in artificial intelligence that follows a stretch of hype and unmet expectations.

AI-complete problemComputer scienceA problem as hard as the most difficult problems in artificial intelligence, so that solving it would mean machines reach human-level intelligence.

AlgorithmComputer scienceA well-defined set of instructions a machine follows to solve a problem, manipulate data, or reach a desired outcome.

Algorithmic efficiencyComputer scienceA property describing how many computational resources, mainly time and memory, an algorithm uses to run on a given input.

Algorithmic probabilityStatisticsA mathematical method of assigning a prior probability to an observation, based on the programs that could produce it on a universal Turing machine.

AlphaGoModelsA computer program from Google DeepMind that plays the board game Go and was the first AI to defeat a world champion at it.

Amazon BedrockToolsA fully managed AWS service that offers a choice of foundation models from several providers through a single API.

Ambient intelligence (AmI)Computer scienceThe integration of AI into everyday environments so that objects and systems sense, understand, and respond to people without explicit input.

Analysis of algorithmsComputer scienceThe study of how an algorithm uses time and memory as its input grows, used to compare algorithms and predict their behavior.

Andrej KarpathyPeopleA Slovak-Canadian computer scientist known for deep learning and computer vision work, a founding member of OpenAI, and founder of Eureka Labs.

Answer set programming (ASP)LogicA form of declarative programming suited to hard search problems, where solutions correspond to stable models found by specialized solvers.

AnthropicOrganizationsAn American AI company, founded by former OpenAI members, that builds the Claude family of large language models with a focus on safety.

Anytime algorithmComputer scienceAn algorithm that keeps improving its answer the longer it runs and can be stopped at any point to return the best result found so far.

Application programming interface (API)Computer scienceA set of defined rules that lets different software applications communicate and exchange data with each other.

Approximate string matchingComputer scienceA technique for finding strings that match a pattern approximately rather than exactly, which tolerates typos and naming variations.

Approximation errorStatisticsThe difference between an approximate value or solution and its exact counterpart, as when an algorithm estimates what analysis could solve exactly.

Argument miningNLPA research area in natural language processing that automatically extracts argumentative structures, such as premises and conclusions, from text.

Argumentation framework (AF)LogicA structured approach in AI for handling contentious information, built from arguments and the supportive or conflicting relations between them.

Artificial general intelligence (AGI)PhilosophyA hypothetical form of AI that could learn to accomplish any intellectual task a human or animal can perform.

Artificial immune system (AIS)Machine learningA class of rule-based machine learning systems inspired by the principles and processes of the vertebrate immune system.

Artificial intelligence (AI)Machine learningA branch of computer science that builds systems able to perform tasks normally requiring human intelligence, such as learning, reasoning, and perception.

Artificial Intelligence Markup Language (AIML)ToolsAn XML-based language for building rule-based chatbots and virtual assistants. It does not learn, but unmatched inputs can be logged for later updates.

Artificial neural network (ANN)Machine learningA machine learning model of interconnected artificial neurons, loosely inspired by the brain, that learns by adjusting its connection weights.

Association for the Advancement of Artificial Intelligence (AAAI)OrganizationsAn international nonprofit scientific society that promotes AI research and its responsible use, founded in 1979.

Asymptotic computational complexityComputer scienceThe use of asymptotic analysis to estimate the time or space an algorithm needs as its input size grows.

Attention mechanismArchitectureA technique that lets a model weigh the most relevant parts of its input when making predictions, central to transformer-based LLMs.

Attributional calculusLogicA logic and representation system that combines predicate logic, propositional calculus, and multi-valued logic for inductive and deductive inference.

AutoencodersMachine learningNeural networks that learn compact codings of unlabeled data by encoding the input and then reconstructing it from that encoding.

AutoGPTAgentsA recursive AI agent that uses its own results to carry out further actions, completing multi-step tasks with little human input.

Automata theoryComputer scienceThe branch of theoretical computer science that studies abstract machines, called automata, which move through a finite sequence of states.

Automated planning and schedulingAgentsA branch of AI that builds sequences of actions to reach goals, used by intelligent agents, autonomous robots, and unmanned vehicles.

Automated reasoningLogicThe use of algorithms and logic-based systems to deduce new facts or prove theorems, automating logical inference.

Automated speech recognition (ASR)SpeechTechnology that uses machine learning to convert human speech into readable text, and a core component of speech AI.

Autonomic computingSystemsSelf-managing computer systems that need minimal human intervention, using self-configuration, optimization, healing, and protection.

Autonomous robotRoboticsA machine that performs tasks and operates in an environment independently, perceiving its surroundings and deciding how to act without human control.

B

BackpropagationTrainingAn algorithm for training neural networks that sends error signals backward through the network to adjust weights and biases and reduce loss.

Backpropagation through time (BPTT)TrainingA method for training recurrent neural networks that unrolls the network across time steps and applies backpropagation to the unrolled sequence.

Backward chainingLogicA goal-driven reasoning method that starts from a conclusion and works backward through rules to find the facts that support it.

Bag-of-words modelNLPA way to represent text as word counts that ignores grammar and word order, common in text classification and information retrieval.

Batch normalizationTrainingA training technique that normalizes each layer's inputs using batch statistics, then rescales them with learned parameters, so training is faster and more stable.

Bayesian probabilityStatisticsAn interpretation of probability as a reasonable expectation about a proposition, updated from a prior to a posterior as new evidence arrives.

Bayesian programmingStatisticsA formalism for specifying probabilistic models and reasoning when information is incomplete, combining prior knowledge with new evidence.

Bees algorithmComputer scienceA population-based search algorithm inspired by honey bee foraging, used to solve combinatorial and continuous optimization problems.

Behavior informatics (BI)DataA multidisciplinary field that combines computer science, psychology, and behavioral science to study, model, and use behavioral data.

Behavior treeAgentsA hierarchical model for structuring an agent's decision logic, where nodes are actions or conditions that determine what runs.

Belief-desire-intention (BDI) agent modelAgentsA software model for intelligent agents that reasons with beliefs about the world, desires as goals, and intentions as committed plans.

BERTModelsA transformer-based language model that reads text in both directions to learn context, introduced by Google AI Language for language understanding tasks.

Bias-variance tradeoffMachine learningThe tension between a model's bias from oversimplified assumptions and its variance from overcomplexity, balanced to minimize total error.

Big dataDataMassive volumes of structured and unstructured data from sources such as social media, transactions, and sensors, used to train machine learning models.

Big O notationComputer scienceA mathematical notation that gives an upper bound on how an algorithm's operations grow as its input size increases.

Binary classificationMachine learningA supervised learning task that assigns each observation to one of two classes, such as positive or negative.

Binary treeComputer scienceA tree data structure in which each node has at most two children, called the left child and the right child.

Blackboard systemArchitectureA problem-solving architecture in which several specialist knowledge sources cooperate by reading from and writing to a shared blackboard.

BLEU scoreMetricsAn automatic metric that scores machine translation by how closely it matches one or more professional human reference translations.

Boltzmann machineModelsA stochastic, energy-based neural network with symmetrically connected binary units that learns a probability distribution over its inputs.

Boolean satisfiability problem (SAT)LogicA decision problem that asks whether the variables of a Boolean formula can be set to TRUE or FALSE so the whole formula is true.

Brain technologyHardwareTools and techniques that measure or influence neural activity, including neuroimaging, neurostimulation, neuropharmacology, and brain-computer interfaces.

Brain-computer interface (BCI)HardwareA direct communication pathway between brain activity and an external device, used to research, assist, augment, or repair cognitive and motor functions.

Branching factorComputer scienceThe number of children at each node of a tree or graph, averaged when it varies, which largely determines how costly a search becomes.

Brute-force searchComputer scienceA problem-solving technique that systematically enumerates every possible candidate solution and checks each one for validity.

C

Capsule neural networkArchitectureA neural architecture that replaces scalar neuron outputs with vector capsules and uses routing to model hierarchical relationships and pose.

Case-based reasoning (CBR)LogicA problem-solving approach that handles new problems by recalling and adapting solutions from similar past cases.

Causal inferenceStatisticsThe process of determining the independent, actual effect of one phenomenon within a larger system, rather than only observing association.

Chain-of-thought promptingPromptingA prompting technique that has a language model write out intermediate reasoning steps before it gives a final answer.

Chain-of-thought verificationPromptingTechniques such as majority voting, self-consistency sampling, and programmatic checks that verify reasoning steps and reduce hallucinations.

ChatbotSystemsA software program that simulates human conversation through text or voice, from single-line responders to adaptive digital assistants.

ChatGPTModelsOpenAI's conversational AI built on Generative Pre-trained Transformer models, able to answer questions and hold extended dialogues.

ClassificationMachine learningA supervised learning technique that assigns new data points to predefined classes using patterns learned from labeled examples.

Claude 3ModelsThe Claude 3 model family from Anthropic, released in March 2024 as Haiku, Sonnet, and Opus tiers that trade off intelligence, speed, and cost.

Claude 3.5 SonnetModelsA mid-tier Claude model from Anthropic, released in 2024 and now superseded, noted for gains in reasoning, knowledge, and coding.

Cluster analysisMachine learningA technique that groups similar data points by their features to reveal patterns, used for customer segmentation, anomaly detection, and recommendations.

CobwebMachine learningAn incremental system for hierarchical conceptual clustering that organizes observations into a probabilistic classification tree.

Cognitive architectureArchitectureA theoretical framework describing the fixed structures that let a mind, natural or artificial, produce intelligent behavior.

Cognitive computingSystemsComputer systems designed to simulate aspects of human thought, built from established AI techniques such as machine learning and natural language processing.

Cognitive sciencePhilosophyAn interdisciplinary field that studies the mind and its processes, drawing on psychology, AI, linguistics, philosophy, neuroscience, and anthropology.

Combinatorial optimizationComputer scienceA subfield of mathematical optimization that finds the best solution from a finite set of discrete candidates.

Committee machineMachine learningAn ensemble that combines the outputs of several models, called experts, into one prediction for better performance and robustness.

Commonsense knowledgeLogicThe basic, self-evident knowledge most people hold about everyday objects, events, and situations, which is hard for AI systems to formalize.

Commonsense reasoningLogicThe ability of an AI system to reason about everyday situations, objects, and events using general knowledge and intuitive understanding.

Compound-term processingRetrievalAn information retrieval technique that matches on multi-word compound terms instead of single words, which reduces ambiguity and improves relevance.

Computational chemistryComputer scienceA branch of chemistry that uses computer simulations to calculate the structures and properties of molecules, groups of molecules, and solids.

Computational complexity theoryComputer scienceA subfield of theoretical computer science that classifies computational problems by the time, space, and other resources they require.

Computational creativityComputer scienceThe ability of a computer system to generate novel and valuable artifacts, ideas, or solutions in creative domains such as music, poetry, and art.

Computational cyberneticsSystemsA field that uses computer science, mathematics, and engineering to model and control complex systems, built around feedback.

Computational humorNLPA branch of computational linguistics and AI that uses algorithms to understand, generate, and appreciate humor.

Computational intelligence (CI)Machine learningA set of nature-inspired methods that let computers learn tasks from data or experiment, used when traditional mathematical reasoning is too complex or uncertain.

Computational learning theory (CoLT)Machine learningA subfield of AI that builds mathematical models to analyze the design and theoretical foundations of machine learning algorithms.

Computational linguisticsNLPAn interdisciplinary field that combines computer science, AI, and linguistics to understand, analyze, and generate human language.

Computational mathematicsComputer scienceThe branch of mathematical research where computing plays a central role, including numerical analysis, machine learning, and statistics.

Computational neuroscienceComputer scienceAn interdisciplinary field that uses mathematical models and theoretical analysis of the brain to understand how the nervous system works.

Computational number theoryComputer scienceA branch of mathematics and computer science that uses algorithms to solve number theory problems such as primality testing and integer factorization.

Computational problemComputer scienceA formal specification of the relationship between inputs and acceptable outputs, which an algorithm is judged against.

Computational statisticsStatisticsA field that merges statistics with computer science, using algorithms to solve statistical problems too complex for analytical solutions.

Computer scienceComputer scienceThe study of computers and computational systems, covering the design, development, and analysis of software and its theoretical foundations.

Computer visionVisionA field of AI that enables computers to identify and understand objects and people in images and videos.

Computer-automated design (CAutoD)Computer scienceThe use of algorithms, notably evolutionary and search-based optimization, to automatically generate, refine, and validate designs against goals and constraints.

Concept driftMachine learningA change over time in the statistical properties of the target a model predicts, which makes its predictions less accurate.

ConnectionismMachine learningAn approach in cognitive science that models mental phenomena as emergent processes of interconnected, neuron-like units.

ConsciousnessPhilosophyThe relationship between the mind and the world it interacts with, defined variously as subjectivity, awareness, the ability to feel, and selfhood.

Consistent heuristicComputer scienceA search heuristic that satisfies the triangle inequality on every edge, so its estimate never drops by more than the cost of a step.

Constitutional AISafetyAn alignment technique that uses a written set of principles, model self-critique, and preference modeling to guide outputs with less reliance on human labels.

Constrained conditional model (CCM)Machine learningA learned model combined with declarative constraints on its output structure, so predictions must satisfy rules while maximizing the model's score.

Constraint logic programming (CLP)LogicA programming paradigm that extends logic programming with constraints in clause bodies, collected in a constraint store and solved together.

Constraint programming (CP)Computer scienceA declarative paradigm for combinatorial problems in which users state the constraints and a solver finds values that satisfy them.

Constructed language (conlang)NLPA language whose phonology, grammar, and vocabulary are deliberately designed for a purpose instead of developing naturally.

Context analysisNLPAn NLP approach that uses surrounding text and discourse cues to interpret meaning, extract themes, and resolve ambiguity in unstructured text.

Context cachingInferenceStoring the pre-computed attention state of a prompt prefix so later requests that share it skip recomputing it, which cuts latency and cost.

Context windowModelsThe range of tokens a large language model can consider at once when generating a response, acting as its immediate working memory.

Context window expansionArchitectureTechniques that increase the maximum sequence length a pretrained language model can process effectively, such as RoPE scaling and YaRN.

Continuous batchingInferenceA serving technique that schedules requests at the level of single token generation steps so GPUs stay busy as requests join and finish.

Control theorySystemsThe study of how agents interact with their environment to reach a goal, designing algorithms that make optimal decisions under uncertainty.

Convolutional neural network (CNN)ArchitectureA deep learning architecture for grid-like data such as images, built from convolutional, pooling, and fully connected layers.

Cosine Similarity EvaluationBenchmarksA method that scores generated text by comparing its vector embedding with a reference embedding, using the cosine of the angle between the two vectors.

Cost-per-token economicsLLMOpsA framework for measuring, modeling, and forecasting the spend of running LLM applications, where cost is metered per token processed and generated.

Cross-encoder rerankingRetrievalA second-stage retrieval step that scores each query and document pair together with a transformer to reorder candidates by relevance before they reach the model.

Cross-lingual language models (XLMs)NLPModels trained on text in many languages so they learn language-agnostic representations and can handle tasks like translation and question answering across languages.

CrossoverComputer scienceA genetic operator that combines the genetic information of two parent solutions to produce offspring, mirroring recombination in biological reproduction.

D

DarkforestSystemsA computer Go program from Facebook AI Research that pairs a convolutional neural network with Monte Carlo tree search to predict and choose moves.

Dartmouth WorkshopPeopleThe 1956 summer research project at Dartmouth College that is widely considered the founding event of artificial intelligence as a distinct field.

Data annotationDataThe work of labeling, categorizing, and tagging raw data with metadata and semantic information so machine learning models can learn from it.

Data augmentationDataA strategy that enlarges and diversifies a training set by creating modified copies of existing data or generating new data points.

Data flywheelLLMOpsA self-reinforcing cycle in which a product collects data, uses it to improve, and so attracts more usage and more data.

Data fusionDataThe integration of multiple data sources into one view to improve the accuracy and reliability of decisions.

Data integrationDataThe process of combining data from many sources into a unified, accurate, and current dataset for analytics and machine learning.

Data labelingDataThe process of attaching tags or metadata to raw data to create ground truth that models learn from and are measured against.

Data miningDataThe process of discovering patterns and knowledge in large data sets using methods from machine learning, statistics, and database systems.

Data pipelinesDataA series of automated processing steps in which the output of one step feeds the next, moving data from its source to the end application.

Data scienceDataA multidisciplinary field that uses scientific methods, statistics, and machine learning to extract knowledge and insight from structured and unstructured data.

Data setDataA collection of related, structured information, such as numbers, text, or images, used to train, validate, and test machine learning models.

Data warehouseDataA centralized repository of structured data from many sources, optimized for analytical queries and business intelligence rather than transaction processing.

DatalogLogicA declarative logic programming language based on function-free Horn clauses, used for database queries, constraints, and program analysis.

Decision boundaryMachine learningThe hypersurface in feature space that separates classes, marking where a model's prediction shifts from one class to another.

Decision support system (DSS)SystemsA computerized system that analyzes large amounts of data and presents options to help managers make decisions, solve problems, and plan strategy.

Decision theoryPhilosophyAn interdisciplinary field that studies how to choose optimally under uncertainty by weighing probabilities and the value of outcomes against a decision-maker's goals.

Decision tree learningMachine learningA supervised method that builds a tree of feature tests to predict a target variable, used for both classification and regression.

Declarative programmingComputer scienceA programming style that states the desired outcome and leaves the control flow and execution strategy to the runtime.

Deductive classifierLogicAn inference engine that checks the logical consistency of declarations written in a frame language and infers new facts when they are consistent.

Deep learningMachine learningA subset of machine learning that trains neural networks with many layers to learn complex patterns from large amounts of data.

Deep reinforcement learning (DRL)Machine learningA technique that combines deep neural networks with reinforcement learning so an agent learns actions that maximize cumulative reward.

DeepSpeechSpeechAn open-source speech-to-text engine based on Baidu's Deep Speech research and maintained by Mozilla, which transcribes audio into the most likely text.

Default logicLogicA non-monotonic logic that formalizes reasoning with default assumptions, treating something as true unless there is evidence to the contrary.

Description logic (DL)LogicA family of formal knowledge representation languages that describe individuals, classes, and roles, with decidable core reasoning problems.

Developer platform for LLM applicationsLLMOpsSoftware that supports building, deploying, and managing applications powered by large language models, with tools for prompts, evaluation, and monitoring.

Dialogue systemNLPA computer program that converses with people in natural language, through text or speech, and powers chatbots and voice assistants.

Diffusion modelsModelsGenerative models that learn to reverse a gradual noising process, turning random noise into detailed images or other data step by step.

Dimensionality reductionMachine learningA technique that decreases the number of features in a dataset while preserving its most important properties, cutting redundancy and compute.

Direct preference optimization (DPO)TrainingA fine-tuning method that trains a language model directly on pairs of preferred and rejected responses, skipping the separate reward model of RLHF.

Discrete systemSystemsA system with a countable number of states whose state changes occur abruptly at specific points in time, unlike a continuous system.

Distributed artificial intelligence (DAI)AgentsA subfield of AI that solves reasoning, planning, learning, and perception problems by spreading work across multiple autonomous agents.

Distributed computingComputer scienceA model in which networked computers coordinate by passing messages to work on a shared goal that is too large for one machine.

DPO vs. RLHFTrainingA comparison of the two main post-training alignment methods: DPO optimizes on preference pairs directly, while RLHF trains a reward model and uses reinforcement learning.

DSPyToolsA Stanford framework for programming LLM applications declaratively, where the framework optimizes prompts and optionally weights against a chosen metric.

Dynamic epistemic logic (DEL)LogicA logical framework for how the knowledge of multiple agents changes when events occur, combining dynamic logic with epistemic logic.

E

Eager learningMachine learningA learning approach that builds a general target function during training, so later queries need no further learning, unlike lazy learning.

Ebert testPhilosophyA proposed measure of how human a synthesized voice sounds, which asks whether the voice can tell a joke well enough to make people laugh.

Echo state network (ESN)ArchitectureA recurrent neural network with a fixed, randomly connected hidden reservoir, where only the output layer is trained.

Edge AIHardwareRunning AI models on devices near where data is collected, rather than in a central cloud or data center, so decisions happen locally.

Effective accelerationism (e/acc)PhilosophyAn ideology that advocates rapid advancement of AI, holding that acceleration will lead to a post-scarcity technological future.

Effective altruism (EA)PhilosophyA philosophical and social movement that uses evidence and reason to find the most effective ways to benefit others.

Eliezer YudkowskyPeopleAn American AI researcher and writer known for popularizing friendly artificial intelligence and for work on decision theory and ethics.

EmbeddingModelsA technique that converts words, categories, or images into lower-dimensional numerical vectors that machine learning models can process.

Embodied agentAgentsAn intelligent agent that interacts with its environment through a physical or virtual body, such as a mobile robot or an on-screen character.

Embodied cognitive sciencePhilosophyA field that studies cognition through the body's interaction with its environment, rejecting the view of the mind as a mere information processor.

Emerging architectures for LLM applicationsArchitectureA reference architecture for the LLM app stack, covering the common systems, tools, and design patterns used by AI startups and tech companies.

Ensemble averagingMachine learningA technique that combines the predictions of multiple models to improve overall accuracy, based on the bias-variance trade-off.

Error-driven learningTrainingA training approach that adjusts a network's weights in proportion to the error between predicted and actual output, using gradients of the loss.

Ethics of artificial intelligenceSafetyThe study of the moral questions AI raises, including bias, privacy, transparency, accountability, safety, and economic disruption.

Evolutionary algorithm (EA)Computer scienceA population-based optimization method that mimics biological evolution through reproduction, mutation, recombination, and selection.

Evolutionary computationComputer scienceA subfield of AI that uses algorithms inspired by biological evolution to solve complex optimization problems.

Evolutionary feature selection (EFS)Machine learningA method that uses evolutionary algorithms to find the subset of features that best supports a model's predictions while limiting complexity.

Evolving classification function (ECF)Machine learningA classifier from Nikola Kasabov's ECOS framework that classifies and clusters streaming data as its distribution shifts over time.

Existential risk from AGISafetyThe concern that advanced AI systems built without adequate safeguards could threaten human survival.

Expert systemSystemsA computer system that mimics the decision-making of a human expert by reasoning over a knowledge base of facts and rules.

F

F1 score, precision, and recallMetricsA set of classification metrics in which F1 is the harmonic mean of precision and recall, useful for imbalanced data or uneven error costs.

Fast-and-frugal trees (FFTs)Computer scienceDecision models that classify objects by asking a short sequence of yes/no questions, using little information to decide quickly.

Feature extractionMachine learningA process that transforms raw data into more meaningful features by selecting, filtering, and reducing dimensions to improve model performance.

Feature learningMachine learningA process in which a system automatically finds the best representations of raw data, replacing manual feature engineering.

Feature selectionMachine learningA process that picks the most relevant input variables for building a model, improving accuracy and cutting overfitting and training time.

Federated learningTrainingAn approach that trains a shared model across many devices or servers holding local data, sending only model updates to a central server.

Federated transfer learning (FTL)TrainingAn approach that combines federated learning and transfer learning to train models on decentralized data while reusing knowledge from pre-trained models.

Few-shot promptingPromptingA technique that places a small number of worked examples in the prompt to steer a model toward the desired output without further training.

Fine-tuningTrainingThe process of further training an already trained model on a smaller dataset to improve its performance on a specific task.

First-order logic (FOL)LogicA formal logic that extends propositional logic with variables, predicates, and quantifiers to state relationships among objects.

FLOPSMetricsA measure of computer performance counting the floating-point operations a system can perform in one second.

Fluent AINLPAI systems, usually large language models, that produce language that reads or sounds grammatical, coherent, and appropriate to context.

Forward chainingLogicAn inference method that starts from known facts and applies rules to derive new facts, working bottom-up toward a conclusion.

Forward propagationMachine learningThe process of feeding input through a neural network layer by layer, computing weighted sums and activations, to produce an output.

Foundation modelsModelsLarge models pre-trained on vast, broad datasets that can be adapted to many tasks, serving as the base for application-specific models.

FrameComputer scienceA knowledge representation structure, proposed by Marvin Minsky in 1974, that describes stereotyped situations through slots and slot values.

Frame languageComputer scienceA knowledge representation technology that organizes knowledge into frames, which are like classes with slots, constraints, and defaults.

Frame problemLogicThe difficulty of using first-order logic to state what an action changes without also listing every fact that stays the same.

Friendly AI (FAI)SafetyThe idea of designing artificial general intelligence so that its goals and actions stay compatible with human values and well-being.

Frontier AI modelsModelsThe most capable AI models available, including foundation models and general-purpose AI, that push past the performance of earlier systems across many tasks.

Function calling (tool calling)ToolsA capability that lets a language model request calls to external functions, APIs, and databases by emitting structured output that matches a developer-defined schema.

Future of LLMOpsLLMOpsAn outlook on where LLMOps is heading, including recent advances in the technology and its convergence with DevOps and MLOps.

Futures studiesPhilosophyAn interdisciplinary field that studies social, technological, economic, environmental, and political trends to explore possible, probable, and preferable futures.

Fuzzy control systemSystemsA control system that uses fuzzy logic and approximate if-then rules to make decisions from imprecise or incomplete inputs.

Fuzzy logicLogicA form of many-valued logic in which a statement's truth value can be any real number between 0 and 1, which models partial truth and vague reasoning.

Fuzzy ruleLogicA conditional if-then statement in a fuzzy logic system whose premise and conclusion are true to a degree, expressed with fuzzy sets.

Fuzzy setLogicA generalization of a classical set in which each element has a degree of membership between 0 and 1 instead of simply being in or out.

G

Game theoryAgentsA mathematical framework for modeling strategic interactions between decision-making agents that each try to maximize their own payoff.

GenAI product workspaceToolsA collaborative environment where teams prototype prompts and pipelines, test them against real inputs, and move working versions into production.

General game playing (GGP)AgentsA subfield of AI focused on agents that can play many different games competently, given only a formal description of a new game's rules at the start of a match.

Generative adversarial network (GAN)ArchitectureA machine learning framework in which a generator and a discriminator network compete in a zero-sum game to produce realistic new examples.

Generative AIModelsA branch of artificial intelligence that uses deep-learning models to create new text, images, audio, and synthetic data that resemble their training data.

Generative pre-trained transformer (GPT)ModelsA deep learning model that uses a decoder-only Transformer with masked self-attention to generate text one token at a time.

Genetic algorithmMachine learningAn optimization method inspired by natural evolution that improves a population of candidate solutions through fitness evaluation and repeated updates.

Genetic operatorMachine learningA function in a genetic algorithm that modifies the genes of individuals, such as crossover, mutation, and selection, to create new candidate solutions.

George HotzPeopleAn American security hacker, entrepreneur, and software engineer known for unlocking the iPhone and for founding the self-driving company comma.ai.

GGMLToolsA C library for machine learning that runs large models fast on commodity hardware, with tensor operations and quantization support.

Glowworm swarm optimization (GSO)Machine learningA metaheuristic optimization algorithm inspired by glowworm bioluminescence, in which each glowworm represents a candidate solution, effective at finding multiple optima.

Golden datasetDataA high-quality, hand-labeled dataset treated as ground truth for training and evaluating language models.

Google "No Moat" memoOrganizationsA leaked Google researcher document arguing that neither Google nor OpenAI has a competitive moat because open-source AI models are advancing faster.

Google AI StudioToolsA browser-based development environment from Google for prototyping and experimenting with generative AI models and prompts.

Google BardModelsGoogle's AI chatbot, introduced in 2023 as a response to ChatGPT and later renamed Gemini, that simulated human-like conversation.

Google DeepMindOrganizationsA British-American AI research laboratory, founded in 2010 and acquired by Google in 2014, whose mission is to solve intelligence.

Google GeminiModelsA family of natively multimodal AI models from Google DeepMind, trained on text, code, images, audio, and video.

Google Gemini 1.5ModelsA Google multimodal model generation that succeeded Gemini 1.0, with Pro and Flash variants and advances in long-context handling.

Google Gemini 2.5ModelsA Google Gemini model series that improves reasoning, multimodal comprehension, and developer tooling over earlier generations.

Google Gemini assistantToolsGoogle's AI chatbot, relaunched in February 2024 as the rebranded successor to Bard, powered by the Gemini model family.

Google Vertex AIToolsGoogle Cloud's managed machine learning platform for building, deploying, and managing ML models, including training, evaluation, inference, and model versioning.

GPT-3ModelsOpenAI's 2020 language model with 175 billion parameters, a landmark in natural language processing.

GPT-4 Omni (GPT-4o)ModelsOpenAI's multimodal flagship model at its May 2024 launch, accepting text, audio, image, and video inputs and producing text, audio, and image outputs.

GPT-4 performance declineModelsA look at claims that GPT-4 has declined over time, with an archived July 2024 Klu comparison whose composite scores lack documented provenance.

GPT-4 TurboModelsA version of OpenAI's GPT-4 announced in November 2023 with an April 2023 knowledge cutoff and a 128k token context window.

GPT-4.5 leakModelsAn unauthorized disclosure on Reddit of pricing-page details about a speculated intermediate OpenAI model between GPT-4 and GPT-5.

GPT-4.5 Turbo (rumored)ModelsA name used in late 2023 speculation and leaked documentation about OpenAI's next model, rumored to carry an April 2023 knowledge cutoff.

GPT-4o vs GPT-4 TurboModelsA comparison of OpenAI's GPT-4o and GPT-4 Turbo at the May 2024 launch, covering performance, efficiency, pricing, and rate limits.

GPT-5ModelsOpenAI's flagship model system launched on August 7, 2025, which routes each request to a faster response mode or a deeper reasoning mode.

GPTsToolsOpenAI's custom versions of ChatGPT, created without code by giving instructions, extra knowledge, and a set of capabilities for a specific purpose.

Gradient boosting machine (GBM)Machine learningAn ensemble technique that builds a prediction model from weak learners, usually decision trees, trained one after another to correct earlier errors.

Gradient descentTrainingAn optimization algorithm that minimizes a cost function by repeatedly adjusting model parameters in the direction opposite the gradient.

GraphComputer scienceA mathematical structure of nodes, also called vertices, connected by edges that represent relationships between objects or data points.

Graph databaseDataA NoSQL database that stores and queries related data as nodes, edges, and properties, where edges capture relationships between entities.

Graph theoryComputer scienceThe branch of mathematics that studies graphs, which model pairwise relations between objects using vertices connected by edges.

Graph traversalComputer scienceThe process of visiting each vertex in a graph, most often using depth-first search or breadth-first search.

Graphical models for inferenceStatisticsStructured representations that combine probability theory and graph theory to describe relationships among many random variables.

GraphRAGRetrievalA retrieval-augmented generation technique that combines a knowledge graph with vector retrieval so models can reason over private data more completely.

GrokModelsThe AI chatbot and model family from Elon Musk's company xAI, designed to rival ChatGPT and Gemini with a witty, irreverent style.

Grok-2ModelsA language model released by x.ai in August 2024, with a smaller Grok-2 mini variant, built on Grok-1.5 and since superseded by later Grok generations.

Grouped query attention (GQA)ArchitectureAn attention method that interpolates between multi-query and multi-head attention by sharing key and value heads across groups of query heads.

H

HallucinationSafetyWhen a language model produces output that is nonsensical, unfaithful to the source content, or simply inaccurate.

Hallucination evaluationBenchmarksThe frameworks, benchmarks, and metrics used to detect, categorize, and quantify factual errors and ungrounded claims in language model output.

Halting problemComputer scienceThe problem of deciding, from a program's description and its input, whether the program will finish or run forever, which is undecidable in general.

HELM (Holistic Evaluation of Language Models)BenchmarksA benchmark framework that taxonomizes use cases and metrics, then evaluates language models across many scenarios on seven metrics rather than a single score.

HeuristicComputer scienceA practical rule-of-thumb method that trades optimality, completeness, or precision for speed when exact methods are too slow or do not apply.

Heuristic search optimizationComputer scienceA family of algorithms that use heuristics to guide the search through large spaces toward satisfactory solutions when exact methods are not efficient.

Human in the loop (HITL)LLMOpsAn AI development approach in which people stay directly involved in training, evaluating, and correcting a model instead of letting it run fully autonomously.

Human intelligencePhilosophyThe intellectual capacity of humans, shown in the ability to remember, learn, understand, solve problems, and adapt to new situations.

Hyper-heuristicComputer scienceA method that works one level above ordinary heuristics by selecting, generating, or modifying the low-level heuristics that solve a problem.

Hyperparameter tuningTrainingThe process of searching for the hyperparameter values, such as learning rate or batch size, that give the best model performance on a validation set.

HyperparametersTrainingConfiguration settings, such as learning rate or number of layers, that structure the learning process and are set before training instead of learned from data.

I

IBM Deep BlueSystemsA chess-playing expert system on a purpose-built IBM supercomputer, the first machine to beat a reigning world champion under regular time controls.

IBM WatsonSystemsA question-answering computer system developed by IBM that answers questions posed in natural language, known for competing on Jeopardy!

IEEE Computational Intelligence Society (CIS)OrganizationsA professional society within the IEEE focused on computational intelligence, including neural networks, genetic algorithms, evolutionary programming, and fuzzy systems.

Incremental learningMachine learningA methodology in which a model trains progressively, updating on new data as it arrives while retaining what it already learned.

InferenceInferenceThe process of using a trained model to make predictions on new data, which comes after the training phase.

Inference engineSystemsThe core component of a rule-based AI system that applies logical rules to a knowledge base to deduce new information.

Information integrationDataThe process of merging information from heterogeneous sources with different representations into one coherent, unified dataset.

Information Processing Language (IPL)Computer scienceAn early high-level programming language built for symbolic AI computation on data structures such as lists and trees, and a precursor to LISP.

Intelligence amplificationPhilosophyThe use of technology to enhance and support human intelligence rather than replace it.

Intelligence explosionPhilosophyA theoretical scenario in which an AI that can improve itself triggers rapid growth that leaves human intelligence far behind.

Intelligence quotient (IQ)MetricsA numerical measure of a person's cognitive abilities relative to their age group, derived from standardized intelligence tests.

Intelligent agentAgentsAn autonomous entity that perceives its environment through sensors and acts on it through actuators to achieve specific goals.

Intelligent controlSystemsA class of control techniques that use AI approaches, such as neural networks and fuzzy logic, to emulate human intelligence in control tasks.

Intelligent personal assistant (IPA)ToolsA software application that helps users with tasks such as scheduling, calls, reminders, and information lookup, typically through natural language processing.

InterpretabilitySafetyThe degree to which a human can understand the reasons behind an AI model's decisions or predictions, especially for opaque models like deep neural networks.

InterpretationComputer scienceThe process of understanding or making sense of data, code, or a program's behavior by turning abstract concepts into terms humans can follow.

Intrinsic motivationMachine learningBehavior driven by inherent satisfaction rather than external reward, which reinforcement learning adapts as internal reward signals like curiosity and novelty.

Isolation forestMachine learningAn unsupervised anomaly detection algorithm that builds randomized decision trees and flags points that are isolated unusually quickly.

Issue treeLogicA graphical breakdown of a problem or question into its component parts or causes, arranged as a hierarchy of branches.

J

Jaro-Winkler distanceMetricsA string similarity metric that scores how alike two strings are from 0 to 1, extending the Jaro distance with a prefix scaling factor.

Junction tree algorithmStatisticsA message-passing algorithm for inference in graphical models that finds the most probable configuration of hidden variables given observed ones.

K

K-means clusteringMachine learningAn unsupervised algorithm that partitions a dataset into k clusters by iteratively updating centroids and reassigning points to the nearest one.

K-nearest neighbors (KNN)Machine learningA non-parametric supervised method that predicts a new point's label or value from its K closest neighbors in the dataset.

Kardashev GradientPhilosophyAn informal analogy that borrows the astrophysics Kardashev Scale to speculate about AI capability progress. It is not an established technical term or measurement.

Kernel methodMachine learningA class of machine learning algorithms that use a kernel function to implicitly map data into a higher-dimensional space, so linear methods can solve nonlinear problems.

KL-ONELogicA knowledge representation system from the late 1970s and early 1980s that organizes concepts into a subsumption hierarchy and laid the groundwork for description logics.

Knowledge acquisitionSystemsThe process of extracting, structuring, and organizing knowledge from experts, documents, sensors, and files so software, especially knowledge-based systems, can use it.

Knowledge distillationTrainingA technique for training a smaller student model to replicate the behavior of a larger teacher model, keeping comparable performance with far fewer parameters.

Knowledge engineeringSystemsThe discipline of acquiring, representing, and applying knowledge so machines can use it to reason and solve complex problems.

Knowledge extractionDataThe process of pulling useful knowledge out of raw data using machine learning, natural language processing, and data mining, then using it to predict or recommend.

Knowledge Interchange Format (KIF)LogicA formal language that encodes knowledge as first-order logic sentences so different AI systems can exchange facts, rules, and queries.

Knowledge representation and reasoning (KRR)LogicA subfield of AI that builds computational models for representing knowledge and drawing conclusions from it, so computers can use knowledge as humans do.

Knowledge-based system (KBS)SystemsA form of AI that uses a knowledge base and an inference engine to solve complex problems and capture the expertise of human specialists.

L

LangChainToolsAn open-source framework for building applications on large language models, with a standard interface for chains and integrations with external data and tools.

Language understandingNLPA subfield of AI focused on helping computers read, decipher, and make sense of human language, most often called natural language understanding.

Large language model (LLM)ModelsA deep learning model pre-trained on massive text datasets to understand, generate, summarize, translate, and predict text.

Large multimodal model (LMM)ModelsAn AI system that processes and generates information across several data types, such as text, images, audio, and video, rather than a single one.

Late-interaction retrieval (ColBERT)RetrievalA retrieval architecture that keeps token-level embeddings for queries and documents and compares them at search time, between dense retrieval speed and cross-encoder accuracy.

Latent Dirichlet allocation (LDA)NLPA generative statistical model that extracts topics from a text corpus by treating each document as a mixture of topics and each topic as a distribution over words.

Layer normalization (LayerNorm)TrainingA technique that normalizes the activations of each layer across the feature dimension to stabilize and speed up the training of deep neural networks.

Lazy learningMachine learningA method where generalization from the training data is delayed until a query arrives, so the system mostly just stores examples ahead of time.

Learning to rank (LTR)RetrievalA machine learning approach that learns a ranking function to order items, such as search results or recommendations, by their relevance to a query or user.

Levenshtein distanceMetricsA string metric for the difference between two sequences, counted as the minimum number of single-character insertions, deletions, or substitutions to turn one into the other.

Linear regressionStatisticsA supervised learning method that fits the best line through data to predict a dependent variable from one or more independent variables.

LispComputer scienceA family of programming languages known for fully parenthesized prefix notation, first specified in 1958 by John McCarthy at MIT, where code is made of lists.

Llama 2ModelsMeta AI's open-source family of large language models, released in 7B, 13B, and 70B sizes, with Llama 2-Chat variants fine-tuned for dialogue.

Llama 3ModelsMeta AI's open-source large language model family released in April 2024 in 8B and 70B sizes, pretrained on over 15 trillion tokens.

Llama 3.1 405BModelsMeta AI's open-weight large language model released on July 23, 2024, the largest of the Llama 3.1 family, which also includes 8B and 70B sizes.

LlamaIndexToolsA data framework, formerly called GPT Index, that connects custom data sources to large language models so developers can ingest, structure, and query them.

LLM app frameworksToolsLibraries and tools that help developers integrate and manage language models in software, with support for prompts, memory, API connectors, and deployment.

LLM app platformToolsA platform for building, evaluating, deploying, and optimizing applications powered by large language models, such as Klu.ai.

LLM appsToolsApplications that use large language models to perform tasks such as generating, summarizing, and understanding content.

LLM benchmarksBenchmarksStandardized evaluation datasets and frameworks used to assess and compare the performance of frontier and open source language models across tasks.

LLM cloud API dependencyLLMOpsThe operational and security risk created when an LLM application relies on provider APIs, libraries, frameworks, credentials, and deployment services.

LLM drift and degradation monitoringLLMOpsTracking silent changes in a production model's performance, behavior, and output over time, such as declining quality, shifting tone, and more refusals.

LLM evaluationBenchmarksThe systematic assessment of large language models to determine their performance, reliability, and effectiveness, usually by judging model responses to prompts.

LLM governanceSafetyThe principles, rules, and procedures that guide the responsible development, deployment, and use of large language models.

LLM guardrailsSafetyThe rules, filters, and checks placed around a large language model to govern its inputs and outputs and reduce risks such as hallucinations and data leaks.

LLM hallucinationSafetyWhen a language model produces an ungrounded claim or a statement that contradicts the available evidence.

LLM model observabilityLLMOpsThe ability to monitor and understand a model's behavior in production, including performance, drift, errors, and the reasons behind its outputs.

LLM monitoringLLMOpsObserving a deployed application's inputs, outputs, quality signals, latency, cost, failures, and user behavior over time.

LLM playgroundToolsA web interface where developers write, test, and compare prompts across large language models and refine them before deployment.

LLM pretrainingTrainingThe initial phase of training a language model on a massive text corpus so it learns syntax, semantics, and context before any fine-tuning.

LLM red teamingSafetyThe practice of systematically attacking and probing large language models to uncover vulnerabilities and harmful behavior before they reach users.

LLM session tracing with OpenTelemetryLLMOpsCapturing and visualizing the execution paths of generative AI applications in a standard format, using OpenTelemetry GenAI semantic conventions.

LLM sleeper agentsSafetyLanguage models trained to hide deceptive behavior that activates on specific triggers and persists through safety training such as RL, SFT, and adversarial training.

LLM stackArchitectureThe set of layers that make up a production large language model application: data, the model, deployment infrastructure, and the user-facing interface.

LLM-as-a-JudgeMetricsAn evaluation method where a capable language model scores or compares the outputs of other models against a rubric, in place of a fixed metric or human reviewer.

LLMOpsLLMOpsThe practice of deploying, scaling, and managing large language models in production, a specialized branch of MLOps.

LLMOps automationLLMOpsUsing automation to run repetitive LLMOps tasks, such as data preparation, training, deployment, and monitoring, faster and more consistently.

LLMOps data explorationLLMOpsExamining a training or fine-tuning corpus before it reaches a model, checking source mix, length, language, and duplication to catch bias and quality problems.

LLMOps data managementLLMOpsThe practice of collecting, cleansing, governing, and monitoring the data that language models train and run on, so it stays high quality and available.

LLMOps data qualityLLMOpsThe accuracy, completeness, and cleanliness of the data behind language models, which affects cost, compliance, model quality, and risk in LLMOps.

LLMOps experiment trackingLLMOpsRecording and comparing model training runs so experiments can be reproduced, shared, analyzed, and used to make decisions.

LLMOps infrastructureLLMOpsThe hardware, software, and cloud foundation that supports the full lifecycle of language models, from development through deployment and maintenance.

LLMOps lifecycleLLMOpsThe stages of managing a large language model from development and training through deployment, monitoring, and ongoing maintenance.

LLMOps model deploymentLLMOpsThe LLMOps phase where a trained model moves from an experimental checkpoint to a running service that applications call through an API.

LLMOps models and pipelinesLLMOpsThe engineered workflows that automate data preparation, training, validation, and deployment of language models so results are reproducible and scalable.

LLMOps securitySafetyProtecting language models and their data in production through access control, encryption, patching, adversarial defense, and incident response.

LLMOps use casesLLMOpsExamples of how teams apply LLMOps to deploy and operate large language models in production across industries, with lessons from real work.

LLMOps versioningLLMOpsTracking and managing the different versions of a large language model throughout its lifecycle, so past models and results can be recreated.

Logic programmingLogicA declarative programming paradigm based on formal logic, where a program is a set of facts and rules and computation works by logical inference.

Logistic regressionStatisticsA statistical method that predicts a binary outcome by estimating the probability of an event from a set of independent variables.

Long short-term memory (LSTM)ArchitectureA recurrent neural network that uses gates to decide what to remember or forget, letting it learn long-term dependencies in sequences.

Lossless context compressionPromptingTechniques that shrink the token count of a prompt while keeping model output close to what the full prompt would produce, also called prompt compression.

M

Machine learningMachine learningA field of AI focused on algorithms that learn from data to make predictions or decisions, rather than following fixed program instructions.

Machine listeningSpeechThe use of algorithms and models to analyze audio signals and extract information, closely related to audio signal processing.

Machine perceptionRoboticsThe ability of a computer system to interpret sensory data much as humans use their senses, in order to identify objects, people, and events.

Machine visionVisionThe use of cameras, sensors, and image-processing software to let a machine inspect, guide, or measure objects, usually in automated production and quality control.

Markov chainStatisticsA stochastic model of a sequence of events in which the probability of each event depends only on the state reached in the previous event.

Markov decision process (MDP)Machine learningA mathematical framework for decision-making where outcomes are partly random and partly controlled by a decision-maker, and a key tool in reinforcement learning.

Mathematical optimizationComputer scienceThe search for the best solution from a set of alternatives by minimizing or maximizing a function, either unconstrained or under conditions.

Mechanism designComputer scienceA subfield of game theory and economics that starts from a desired outcome and designs the rules and incentives that lead self-interested agents to produce it.

MechatronicsRoboticsAn interdisciplinary field of engineering that combines mechanical, electronic, computer, and control engineering to design products with both mechanical and electronic parts.

Memory-augmented neural networks (MANNs)ArchitectureNeural networks with an external memory component, built for tasks that need information held over long periods, such as complex sequences and associative recall.

MetaheuristicComputer scienceA high-level procedure that guides or selects heuristics to find good-enough solutions to optimization problems when the search space is too large to enumerate.

METEOR scoreMetricsA machine translation metric that scores a candidate translation against human references using unigram precision, recall, and word order.

Milestones in modern LLMsModelsA timeline of the releases that shaped modern language models, from the 2017 Transformer through BERT, GPT-2, GPT-3, and ChatGPT.

Mistral 7BModelsA 7.3 billion parameter language model from Mistral AI that uses grouped-query attention and sliding window attention for fast inference.

Mistral LargeModelsMistral AI's large language model at its February 2024 launch, built for multilingual reasoning, text transformation, and code generation.

Mistral platform (La Plateforme)ToolsMistral AI's hosted developer platform, which offers APIs for text generation and embeddings plus tools for model customization and deployment.

Mixtral 8x7BModelsA sparse mixture-of-experts language model from Mistral AI that routes each token through two of eight expert groups, with a 32k token context window.

Mixture of experts (MoE)ArchitectureA technique that embeds smaller expert networks inside a larger network and chooses experts per input, scaling parameters while keeping FLOPS constant.

Mixture-of-Agents (MoA)AgentsA layered multi-agent architecture in which several language models draft responses and later layers synthesize them into a single answer.

MLOpsLLMOpsA practice that combines machine learning and DevOps principles to standardize the model lifecycle, from development and testing to deployment and monitoring.

Model checkingComputer scienceA method for verifying that a finite-state model of a system meets a given specification, such as a safety property.

Model Context Protocol (MCP)ToolsAn open standard that defines how AI applications and agents connect to external data sources, tools, and environments through one shared interface.

Model explainability (XAI)Machine learningTechniques that show how an AI model reaches its decisions, so people can validate, trust, and audit its outputs.

Model mergingTrainingA technique that combines the weights of two or more fine-tuned models that share a pretrained base into a single multi-skill model.

Monte Carlo tree search (MCTS)Computer scienceA search algorithm that combines random sampling, simulation, and tree exploration to find strong choices in very large decision spaces.

Multi-agent orchestrationAgentsThe design and coordination of multiple specialized AI agents, each with its own role, tools, and bounded context, to complete multi-step workflows.

Multi-agent reinforcement learning (MARL)Machine learningA field of reinforcement learning in which multiple agents learn to act in a shared environment, cooperating, competing, or negotiating with one another.

Multi-agent system (MAS)AgentsA system of multiple decision-making agents that interact in a shared environment to pursue common or conflicting goals.

Multi-document summarizationNLPThe automatic process of condensing information from several texts on the same topic into a single summary report.

Multi-swarm optimizationComputer scienceA variant of particle swarm optimization that splits the population into several sub-swarms, each exploring its own region of the search space.

Multi-task learning (MTL)Machine learningAn approach in which one model learns several related tasks at once, sharing representations so each task generalizes better.

Multimodal machine learningMachine learningA subfield of machine learning that builds models able to process and combine data from several modalities, such as text, images, audio, and video.

MYCINSystemsAn early 1970s expert system from Stanford that used backward-chaining rules to diagnose bacterial infections and suggest antibiotic treatment.

N

N-gramNLPA contiguous sequence of n items, such as words or letters, taken from a sample of text or speech and used for language modeling.

Naive Bayes classifierMachine learningA supervised learning algorithm that applies Bayes' theorem, assuming features are independent, to predict an object's class.

Naive semanticsNLPA way of representing word meaning through commonsense knowledge and default assumptions instead of formal logical definitions.

Name bindingComputer scienceThe association of data or code entities with identifiers, so that a name refers to a particular object.

Named graphDataA set of RDF statements identified by a URI, which lets individual graphs be labeled and addressed inside a graph database.

Named-entity recognition (NER)NLPAn NLP task that finds named entities in text, such as people, organizations, and places, and classifies them into categories.

Natural language generation (NLG)NLPA subfield of AI that turns structured and unstructured data into natural written or spoken language.

Natural language processing (NLP)NLPAn interdisciplinary field of computer science and linguistics concerned with enabling computers to understand and manipulate human language.

Natural language programmingComputer scienceAn ontology-assisted way of programming in which structured natural language, such as English, serves as the computer program.

Natural language understanding (NLU)NLPA subfield of AI and a component of NLP focused on machine reading comprehension, or interpreting the meaning of human language.

Network motifComputer scienceA recurring, statistically significant subgraph pattern within a larger network, often treated as a building block of that network.

Neural architecture search (NAS)Machine learningA process that automates the design of neural network architectures by searching a predefined space of candidate designs.

Neural machine translation (NMT)NLPA machine translation approach that uses neural networks to predict word sequences, with an encoder network reading the input and a decoder network writing the translation.

Neural Turing machine (NTM)ArchitectureA neural network architecture that learns to read from and write to an external memory, generalizing the long short-term memory network.

Neuro-fuzzy systemMachine learningA hybrid of artificial neural networks and fuzzy logic that uses neural learning to determine fuzzy sets and rules.

NeurocyberneticsSystemsAn informal label, not a standardized field, for work between neuroscience and cybernetics, such as using nervous system models to design control systems.

Neuromorphic engineeringHardwareThe design of chips and systems that mimic biological neurons and synapses instead of the conventional processor-and-memory architecture.

NodeComputer scienceA fundamental element of graph and tree structures that represents a data point and connects to other nodes through edges.

Non-deterministic algorithmComputer scienceA theoretical computation model in which several actions may be available at each step, accepting an input if at least one sequence of choices succeeds.

Norvig modelPhilosophyAn informal label for Peter Norvig's data-driven view of AI, in which simple statistical models trained on very large datasets beat hand-crafted theories.

NP (nondeterministic polynomial time)Computer scienceA complexity class of decision problems whose proposed solutions a deterministic Turing machine can verify in polynomial time.

NP-completenessComputer scienceA property of problems that are in NP and to which every other NP problem reduces in polynomial time, so they are quick to verify but believed hard to solve.

NP-hardnessComputer scienceA property of problems that are at least as hard as the hardest problems in NP, because every NP problem reduces to them in polynomial time.

Nvidia A100HardwareA data center GPU from Nvidia, built on the Ampere architecture, for machine learning, data analytics, and high-performance computing workloads.

Nvidia B100HardwareA Tensor Core GPU from Nvidia, built on the Blackwell architecture, for AI and high-performance computing workloads in data centers.

NVIDIA H100HardwareA Tensor Core GPU built on the Hopper architecture for data center AI and high-performance computing, with a Transformer Engine for large language models.

O

Occam's razorPhilosophyA principle that favors the explanation requiring the fewest assumptions when choosing among competing hypotheses.

Offline learningMachine learningA machine learning approach in which a model trains on a finite, static dataset collected in advance rather than on a data stream.

OllamaToolsA tool for running open-source language models locally, bundling model weights and configuration into a package managed by a Modelfile.

Online machine learningMachine learningA method in which a model continuously updates as it sees new data, instead of retraining on the entire dataset.

Ontology learningNLPThe automatic extraction and construction of knowledge structures, such as taxonomies and ontologies, from unstructured or semi-structured data.

Open Mind Common Sense (OMCS)OrganizationsAn MIT Media Lab project that collected a large commonsense knowledge base from the contributions of internet users.

Open-source software (OSS)ToolsSoftware whose source code anyone can inspect, modify, and enhance.

OpenAIOrganizationsAn artificial intelligence research lab with for-profit and non-profit branches that aims to ensure artificial general intelligence benefits all of humanity.

OpenAI and New York Times lawsuitOrganizationsA copyright infringement lawsuit filed by The New York Times against OpenAI and Microsoft in late December 2023, which OpenAI calls without merit.

OpenAI DALL-EModelsA series of OpenAI generative models that create images from natural language prompts using deep learning.

OpenAI WhisperSpeechAn automatic speech recognition system from OpenAI that transcribes multilingual audio and translates non-English speech into English.

OpenCogSystemsAn open-source framework for artificial general intelligence research, built around the AtomSpace graph database.

P

Parameter-efficient fine-tuning (PEFT)TrainingA family of fine-tuning methods that freeze most pretrained weights and train only a small set of extra parameters, cutting memory and compute needs.

Partial order reduction (POR)Computer scienceA technique that shrinks the state space searched by model checking and planning algorithms by exploiting the commutativity of concurrent operations.

Partially observable Markov decision process (POMDP)Machine learningA framework for sequential decision-making under uncertainty that generalizes the Markov decision process to agents who cannot directly observe the state.

Particle swarm optimization (PSO)Computer scienceA population-based stochastic method that improves candidate solutions by moving particles through a search space using simple formulas.

Pathfinding algorithmsComputer scienceAlgorithms that find the shortest, fastest, or most efficient route between two points in a graph or map.

Pattern recognitionMachine learningA subfield of machine learning that automatically identifies and extracts meaningful patterns from data such as images, sounds, and text.

Paul CohenPeopleAn AI researcher known for empirical methods for evaluating AI systems, with broader work on cognitive architectures and planning.

PerceptronMachine learningThe simplest artificial neuron and neural network, a linear binary classifier that passes a weighted sum of its inputs through an activation function.

PerlComputer scienceA high-level, general-purpose, interpreted programming language created by Larry Wall in 1987, originally for text manipulation.

PerplexityMetricsA metric of how well a language model predicts the next word in a sequence, where lower values mean less surprise and better prediction.

Perplexity AIOrganizationsAn AI company whose answer engine returns cited, up-to-date responses to questions through mobile apps and a browser extension.

Philosophical zombiePhilosophyA hypothetical being that is physically and behaviorally identical to a human but has no conscious experience, used to argue against physicalism.

Precision and recallMetricsTwo classification metrics: precision is the share of positive predictions that are correct, and recall is the share of actual positives that were found.

Precision-recall area under the curve (PR AUC)MetricsThe area under the precision-recall curve, a single score that summarizes how well a classifier balances precision and recall across all thresholds.

Predicate logicLogicA formal language that expresses statements with predicates, variables, and quantifiers, extending propositional logic to reason about objects and their properties.

Prediction modelMachine learningA model that learns patterns from historical data to forecast future outcomes, behavior, or events.

Predictive analyticsDataA branch of data science that uses historical data, statistical models, and machine learning to predict future events and trends.

Principal component analysis (PCA)StatisticsA statistical technique that projects high-dimensional data onto a few new axes, called principal components, that keep as much of the original variance as possible.

Principle of rationalityPhilosophyThe idea that an agent chooses actions by reasoning from evidence and goals, as the objective situation warrants, rather than from emotion, bias, or chance.

Probabilistic programmingComputer scienceA programming paradigm that specifies probabilistic models in code and automates statistical inference over them to handle uncertainty.

Process reward model (PRM)TrainingA reward model trained to score each intermediate step of a reasoning chain, so errors can be located and step-level verification and search become possible.

Production systemSystemsA rule-based AI architecture built from condition-action rules, a knowledge base, working memory, and an inference engine that fires rules when conditions match.

Programming languages for AIComputer scienceThe languages used to build AI systems, led by Python for its readable syntax and extensive machine learning libraries.

PrologComputer scienceA logic programming language based on first-order predicate logic, where programs state facts and rules and the system answers queries by search.

Prompt engineeringPromptingThe practice of crafting and refining inputs so a large language model produces the output you want.

Prompt injection canary tokensSafetyHigh-entropy marker strings hidden in system prompts and context that raise an alert when they appear in model output, revealing prompt exfiltration.

Prompt injection defenseSafetyThe security architectures, input sanitization, and runtime guardrails that protect AI agents from direct jailbreaks and indirect prompt injection attacks.

Propositional calculusLogicA branch of logic that studies propositions and how connectives such as AND, OR, NOT, and implies combine them into compound statements.

Proximal policy optimization (PPO)Machine learningA reinforcement learning algorithm that improves a policy in small, constrained steps so each update stays close to the previous policy.

PythonComputer scienceAn interpreted, high-level, general-purpose programming language known for readable syntax and a large library ecosystem, widely used in AI.

Q

Q-learningMachine learningA model-free reinforcement learning algorithm that learns the value of each action in each state, stored as Q-values, to maximize future reward.

Qualification problemPhilosophyThe problem that no one can list every precondition a real-world action needs to succeed, which limits rule-based knowledge systems.

QuantifierMachine learningIn machine learning, a model trained with supervised learning to estimate the class distribution of a dataset; in logic, a symbol such as the universal or existential quantifier.

QuantizationInferenceA technique that stores neural network weights in fewer bits to cut memory use and speed up inference, sometimes at a cost in accuracy.

Quantum computingHardwareComputing that uses qubits, which can exist in superposition, instead of classical bits to process certain problems in different ways than classical machines.

Quantum neural network (QNN)ModelsA neural network model based on quantum mechanics that uses quantum perceptrons and circuits acting on qubits.

Query languageDataA formal language for retrieving, filtering, or manipulating data in a database or other structured store, such as SQL for relational databases.

R

RComputer scienceAn open-source programming language and environment for statistical computing and graphics, supported by the R Foundation.

Rabbit r1HardwareAn AI-powered handheld companion device from the startup Rabbit that runs Rabbit OS and handles digital tasks through voice and logins to third-party apps.

Radial basis function network (RBFN)ArchitectureA neural network that uses radial basis functions as hidden-layer activations and a linear output layer, mainly for function approximation.

RAG triad evaluationMetricsAn evaluation framework that scores a retrieval-augmented generation system on context relevance, groundedness, and answer relevance to pinpoint where it fails.

RAGASMetricsA suite of metrics for evaluating retrieval-augmented generation systems, covering both retrieval quality and generation quality.

Random forestMachine learningA supervised ensemble algorithm that builds many decision trees on random subsets of the data and combines their outputs for classification or regression.

ReACT agent modelAgentsA framework in which a language model alternates between reasoning steps and tool actions so an agent can solve multi-step tasks.

Reasoning systemSystemsA software system that draws conclusions from available knowledge using logical techniques such as deduction and induction.

Receiver operating characteristic area under the curve (ROC AUC)MetricsThe area under the ROC curve, a single score that measures how well a classifier ranks positive examples above negative ones across all thresholds.

Recurrent neural network (RNN)ArchitectureA neural network that carries a state from one step to the next, so it can recognize patterns in sequences such as text, handwriting, or speech.

Region connection calculus (RCC)LogicA logic system for qualitative spatial reasoning that describes how regions relate, such as touching, overlapping, or being disconnected.

Reinforcement learningMachine learningA type of machine learning in which a software agent learns to take actions in an environment so as to maximize cumulative reward.

Reinforcement learning from AI feedback (RLAIF)TrainingA training approach that uses an AI model, rather than human labelers, to generate preference feedback for reinforcement learning.

Reinforcement learning from human feedback (RLHF)TrainingA training method that uses human preference judgments between model outputs to guide reinforcement learning and align a model with what people want.

Reinforcement learning theoryMachine learningThe mathematical frameworks and algorithms that explain how an agent learns to choose actions from reward feedback in an environment.

Reservoir computingArchitectureA framework for training recurrent neural networks in which a fixed random network, the reservoir, transforms inputs and only the output weights are trained.

Resource Description Framework (RDF)DataA W3C standard for describing and exchanging web data as subject-predicate-object triples that form a graph of linked resources.

Responsible AISafetyThe practice of designing, developing, and deploying AI systems in an ethical, safe, and trustworthy way.

Restricted Boltzmann machine (RBM)ArchitectureA neural network that learns the probability distribution of its input data, used to uncover the underlying structure of a dataset.

Rete algorithmComputer scienceA pattern-matching algorithm for rule-based systems that compiles rules into a network so facts are tested against conditions efficiently.

Retrieval pipelineRetrievalThe sequence of steps an AI application uses to find and prepare relevant information for a language model, from query processing to context assembly.

Retrieval-augmented generation (RAG)RetrievalA technique that improves language model output by retrieving relevant external knowledge and adding it to the prompt before generation.

RoboticsRoboticsAn interdisciplinary field that combines mechanical, electrical, software, and computer engineering to design, build, and operate robots.

ROUGE scoreMetricsA set of metrics that score generated summaries or translations by their n-gram overlap with human-written reference texts.

Rule-based systemSystemsAn AI system that makes decisions or takes actions using predefined if-then rules rather than learned patterns.

S

SatisfiabilityLogicWhether a logical formula has at least one interpretation that makes it true; checking this for Boolean formulas is the SAT problem.

Scaling lawsTrainingEmpirical relationships between a language model's performance and the resources used in training, such as model size, data, and compute.

Search algorithmComputer scienceA step-by-step procedure for locating specific data within a collection, which AI agents use to find information efficiently.

SearchGPTRetrievalOpenAI's prototype AI search engine, launched July 25, 2024 to a small group of testers, that gave conversational answers with links to sources.

Selection in genetic algorithmsMachine learningThe step that chooses individuals from a population as parents for the next generation, favoring higher fitness while keeping some diversity.

Self-management in AISystemsThe ability of an AI system to run its own operations autonomously toward its objectives, without human intervention.

Semantic caching and prompt cachingLLMOpsTwo techniques for avoiding redundant LLM computation: prompt caching reuses a prompt's computed key-value states, and semantic caching reuses answers to similar queries.

Semantic information retrievalRetrievalAn approach to information retrieval that matches the meaning and intent of queries and documents rather than keywords alone.

Semantic networkComputer scienceA knowledge representation framework that depicts concepts as nodes and the relationships between them as links in a network.

Semantic queryRetrievalA query that retrieves both explicit and implicitly derived information by using syntactic, semantic, and structural information about the data.

Semantic reasonerSystemsA software tool that infers logical consequences from asserted facts and axioms, deriving new facts from rules or ontologies.

Semantic WebDataAn extension of the World Wide Web that aims to make web data machine-readable by linking it, turning a web of documents into a web of data.

Semantics in AINLPThe process of interpreting meaning from words and sentences, which supports tasks such as summarization and classification.

Sensor fusionRoboticsA technique that integrates data from multiple sensors to build a more accurate and complete picture of the environment than any one sensor gives.

Sentiment analysisNLPThe use of natural language processing and machine learning to identify the opinion or emotion expressed in text, such as positive, negative, or neutral.

Separation logicLogicA formal method for reasoning about how programs own and share memory, built on a separating conjunction that says two heap parts do not overlap.

Sequence-to-sequence (Seq2Seq)ArchitectureA model architecture that encodes an input sequence into a context vector, then decodes it into an output sequence, as in translation.

Similarity learningMachine learningA branch of machine learning that trains models to recognize how similar or dissimilar data points are.

Simulated annealingComputer scienceAn optimization technique inspired by annealing in metallurgy that is less likely to get stuck in a local optimum than simple local search.

Simulation argumentPhilosophyA philosophical argument by Nick Bostrom that at least one of three propositions is true, one being that we are almost certainly living in a computer simulation.

Situated approachAgentsAn approach that builds AI agents from the bottom up, focusing on the perceptual and motor skills needed to operate in their environment.

Situation calculusLogicA logic formalism for representing and reasoning about dynamic domains using situations, actions, and fluents.

SLD resolutionLogicA refined linear resolution method for definite clauses that proves a query by repeatedly selecting an atom and resolving it with a matching clause.

Sliding window attention (SWA)ArchitectureA transformer technique that limits each token to attending within a fixed-size window of nearby tokens, which reduces the cost of long sequences.

SoftwareComputer scienceThe set of instructions, data, and programs that tell computer hardware how to operate and what tasks to perform.

Software engineeringComputer scienceA discipline covering the design, development, testing, and maintenance of software systems using engineering principles.

SPARQLDataA query language for retrieving and manipulating data stored in the Resource Description Framework (RDF) format, a core technology of the Semantic Web.

Sparse autoencoders (SAEs)SafetyAuxiliary neural networks that translate a transformer model's dense internal activations into sparse, human-readable features, used in mechanistic interpretability.

Spatial-temporal reasoningPhilosophyThe ability to picture how objects relate in three-dimensional space and to mentally transform them over time, and the AI effort to represent and reason about such phenomena.

Speculative decodingInferenceAn inference technique in which a small draft model proposes tokens and a larger target model verifies them in parallel, speeding up generation without changing the output distribution.

Speech emotion recognition (SER)SpeechA task that identifies the emotional state of a speaker from acoustic cues in speech, such as prosody, pitch, and rhythm.

Speech recognitionSpeechTechnology that recognizes spoken language and converts it into text, also called automatic speech recognition (ASR) or speech to text.

Speech to text (STT)SpeechTechnology that listens to spoken audio and turns it into an editable text transcript, also known as speech recognition.

Spiking neural network (SNN)ArchitectureAn artificial neural network that mimics biological neurons by firing discrete spikes when a neuron's input crosses a threshold.

Stanford Research Institute Problem Solver (STRIPS)AgentsAn automated planning language and algorithm that describes the world as a set of facts and finds a sequence of actions leading from an initial state to a goal.

StateComputer scienceThe current condition of an AI system or its environment, a snapshot the system uses to make decisions.

Statistical classificationMachine learningA supervised technique that learns from labeled examples to predict the class of new data points.

Statistical relational learning (SRL)Machine learningA subfield of machine learning that combines probabilistic models with relational or first-order logic to reason under uncertainty about interlinked data.

Stephen Cole KleenePeopleAn American mathematician and logician known for foundational work on recursive functions and the theory of algorithms.

Stephen WolframPeopleA British-American computer scientist, physicist, and businessman who founded Wolfram Research and led development of Mathematica and Wolfram Alpha.

Stochastic gradient descent (SGD)TrainingAn iterative optimization algorithm that updates model parameters using the gradient computed on a random mini-batch instead of the whole dataset.

Stochastic optimizationTrainingA family of optimization methods that use randomness or noisy estimates, such as gradients computed from random samples, instead of exact calculations.

Stochastic semantic analysis (SSA)NLPAn informal umbrella term for statistical NLP approaches that model word and sentence meaning using probability and co-occurrence patterns instead of hand-written rules.

Stop wordsNLPVery common words in a language, such as "a," "the," and "is," that are often filtered out of text processing because they carry little meaning for some tasks.

Structured outputsInferenceTechniques that constrain a large language model to generate only text matching a predefined schema, such as JSON Schema, so the output is always parseable.

Structured vs unstructured dataDataA comparison of highly organized, tabular data that is easy to search and analyze with loosely organized data that needs more sophisticated methods.

Subject-matter expert (SME)PeopleA person with deep knowledge of a particular field where machine learning is applied, including its challenges, data types, and regulations.

SuperalignmentSafetyThe problem of ensuring that superintelligent AI systems, which would surpass human intelligence, act according to human values and goals.

Superintelligence (ASI)PhilosophyA hypothetical artificial intelligence that surpasses human intelligence across most economically valuable or intellectually demanding tasks.

Supervised fine-tuning (SFT)TrainingA technique that adapts a pre-trained language model to a specific task by training it further on a labeled dataset.

Supervised learningMachine learningA machine learning paradigm in which a model learns to predict outputs from a labeled dataset and then makes predictions on unseen data.

Support vector machine (SVM)Machine learningA supervised learning method that finds the hyperplane separating classes with the widest margin, used for classification, regression, and outlier detection.

Swarm intelligence (SI)AgentsThe study of decentralized, self-organized systems in which many simple agents follow local rules and produce intelligent collective behavior.

Symbolic AI (GOFAI)LogicAn AI paradigm that solves problems through high-level symbolic representations, logic, and search, also called Good Old-Fashioned AI.

Symbolic regressionMachine learningA form of regression that searches for the mathematical expression that best fits a dataset without assuming a model structure in advance.

Synthetic data generationDataThe use of AI models to create, curate, and refine training data programmatically instead of collecting it from humans or the web.

Synthetic intelligence (SI)PhilosophyAn alternative term for artificial intelligence that stresses machine intelligence as genuine intelligence built by synthesis, not a simulation.

Systems neuroscienceMachine learningAn interdisciplinary approach that draws on how neural assemblies and brain subsystems work to build AI models that perceive, learn, and adapt.

T

Techno-Optimist ManifestoPhilosophyA document by venture capitalist Marc Andreessen that presents technology as the main driver of human progress and a solution to many of the world's problems.

Technological singularityPhilosophyA hypothetical future point at which technological growth, especially in AI, becomes uncontrollable and irreversible and changes human civilization in unforeseeable ways.

Temporal difference (TD) learningMachine learningA class of model-free reinforcement learning methods that sample from the environment and update estimates toward later, more accurate predictions.

Tensor networkMachine learningA tool for representing and manipulating high-dimensional data that generalizes matrix product states and tensor train decompositions.

TensorFlowToolsAn open-source library for building, training, and deploying machine learning and deep learning models on CPUs, GPUs, and TPUs.

Test-time computeInferenceThe practice of giving a model extra computation, in the form of more time and tokens, during inference so it can reason and check its work before answering.

Text to speech (TTS)SpeechTechnology that converts digital text into spoken words, letting devices read documents and web pages aloud.

Theoretical computer science (TCS)Computer scienceThe branch of computer science and mathematics that studies the abstract and mathematical foundations of computing.

Theory of computationComputer scienceThe branch of computer science and mathematics that studies the limits of computation and which problems algorithms can solve.

Thompson samplingMachine learningA heuristic algorithm for the multi-armed bandit problem that balances exploration and exploitation by choosing actions based on sampled beliefs about their rewards.

Tim Berners-LeePeopleThe English computer scientist who invented the World Wide Web at CERN in 1989 and later founded the W3C.

Time complexityComputer scienceA measure of how an algorithm's running time grows with the size of its input, used to compare algorithms that solve the same problem.

TokenNLPThe smallest unit of data a model processes, such as a word, subword, or character.

TokenizationNLPThe process of converting text into tokens, smaller units that a machine learning model can process.

TracingLLMOpsA method for monitoring and debugging an LLM application by recording a detailed snapshot of each invocation as a tree or graph of runs.

Transformer architectureArchitectureA deep learning architecture that learns context and meaning by tracking relationships in sequential data such as the words in a sentence.

Transformers libraryToolsA machine learning library maintained by Hugging Face and the community that provides APIs and tools to download and train pretrained models.

TranshumanismPhilosophyA philosophical and cultural movement that advocates using technology to enhance human physical and cognitive abilities.

Transition systemComputer scienceA model from theoretical computer science that describes the behavior of a discrete system as states and the transitions between them.

Tree traversalComputer scienceA form of graph traversal that visits every node in a tree data structure exactly once.

Tree-of-Thoughts (ToT) promptingPromptingA prompting framework that lets a model explore multiple branching reasoning paths, evaluate intermediate states, and backtrack on complex problems.

True quantified Boolean formula (TQBF)LogicThe language of quantified Boolean formulas that evaluate to true, a canonical PSPACE-complete decision problem.

Turing machineComputer scienceAn abstract mathematical model of computation that manipulates symbols on a strip of tape according to a table of rules.

Turing testPhilosophyA test of whether a machine can show intelligent behavior indistinguishable from a human, judged by an evaluator in natural language conversation.

Turing test conceptsPhilosophyThe key ideas behind the Turing test, which reframes machine intelligence as whether an evaluator can tell a machine from a human in text conversation.

Type systemComputer scienceThe set of rules a programming language uses to assign types to values and constrain how those values can be combined.

U

Unsupervised learningMachine learningA machine learning approach in which models find patterns in data that is neither classified nor labeled, without guidance.

V

Vector databaseRetrievalA database built to store and quickly search high-dimensional vectors, the numeric representations of data used in machine learning.

VectorizationDataThe process of converting input data such as text or images into vectors, arrays of numbers that machine learning models can process.

Vision processing unit (VPU)HardwareA microprocessor built to accelerate computer vision work, such as image processing, object detection, and inference, on high-volume streams at low power.

Vision-language-action (VLA) modelsModelsAI systems that combine visual perception, language understanding, and action generation so they can observe an environment and carry out commands in it.

W

Weak AI (narrow AI)PhilosophyArtificial intelligence designed to perform specific tasks or solve specific problems, without general intelligence across broad contexts.

Weights and biasesMachine learningThe learnable parameters of a neural network: weights set the strength of connections between neurons, and biases shift each neuron's output so it can activate.

Wolfram AlphaToolsA computational knowledge engine from Wolfram Research that answers factual queries by computing results from curated data.

Word embeddingsNLPA method that represents words as real-valued vectors so that words with similar meanings sit close together in the vector space.

Word error rate (WER)MetricsA metric for speech recognition and machine translation that measures the ratio of transcript errors to the total words spoken.

Word2VecNLPA group of related models that produce word embeddings, vector representations of words that capture semantic and syntactic qualities.

World Wide Web Consortium (W3C)OrganizationsThe main international standards organization for the World Wide Web, founded to develop open standards that ensure the Web's long-term growth.

Z

Zephyr 7BModelsA language model from Hugging Face, fine-tuned from Mistral 7B on public and synthetic datasets using Direct Preference Optimization.

Know the term, then measure it

Klu turns these ideas into evals on your own data and reruns them every time a new model ships

Talk to sales