AI Glossary

A computational component whose learned or encoded structure transforms inputs into outputs such as scores, predictions, classifications, or generated content.

The complete operational arrangement that uses one or more AI models together with data, software, infrastructure, interfaces, controls, and people.

A finite set of rules or procedures for transforming inputs into outputs or solving a class of problems.

A contested term for AI with broad, transferable competence across many cognitive tasks, often at or beyond human-level breadth.

The field and class of machine-based systems that produce predictions, recommendations, decisions, or generated content in pursuit of human-defined objectives.

A mechanism that computes how strongly elements in a representation should influence one another when producing a new representation.

Predicting which discrete category or categories apply to an input.

Grouping examples by similarity without requiring predefined class labels.

The field of building computational systems that extract representations, predictions, or actions from images, video, and other visual data.

The maximum token span a model can directly consider in one inference request, including instructions, conversation, retrieved material, tool schemas, and expected output.

A deliberately assembled collection of examples or records used to train, tune, evaluate, or operate an AI system.

Machine learning based on neural networks with multiple representational layers, allowing complex features to be learned from data.

A learned numeric vector that represents an item so that useful semantic or structural relationships can be measured geometrically.

A system that applies an explicitly represented knowledge base and inference rules to make recommendations or decisions within a bounded domain.

A measurable input attribute or derived representation used by a machine-learning system.

A broadly trained model, usually learned through self-supervision on diverse data, that can be adapted to many downstream tasks.

AI designed to produce new content, such as text, code, images, audio, video, or structured data, based on patterns learned from data.

Running a trained model on input to produce an output.

The known target value attached to an example for supervised learning or evaluation.

A large learned model trained to process and generate sequences of language tokens, often with capabilities that extend to code, tools, and multiple modalities.

A family of methods in which computational models improve task performance by finding patterns in data rather than relying only on explicitly programmed rules.

The finite set of token identifiers a tokenizer and model can represent directly.

A model that accepts, relates, or generates more than one modality, such as text, images, audio, video, or structured data.

An AI system designed or validated for a bounded task or domain rather than general competence.

The field of building computational systems that analyze, represent, understand, retrieve, translate, or generate human language.

A parameterized computational model composed of connected layers that transform representations and learn by adjusting weights to reduce an objective.

A value learned during training that helps determine how a model transforms inputs into outputs.

A set of possible outcomes paired with nonnegative probabilities that sum to one.

The attention formulation in which queries are matched against keys to calculate weights applied to corresponding values.

Predicting a continuous numeric value from input data.

A family of methods in which an agent learns a policy by interacting with an environment and optimizing expected cumulative reward.

Learning in which supervisory targets are generated from the structure of otherwise unlabeled data, such as predicting hidden or next tokens.

Machine learning from labeled examples that pair inputs with desired outputs.

Artificially generated records intended to reproduce useful properties of real data for training, testing, simulation, or privacy.

A unit into which model input or output is segmented for processing; a token may be a whole word, part of a word, punctuation, or another symbol.

The extra token count, cost, latency, or lost context capacity imposed when a tokenizer represents equivalent content less efficiently in one language, script, domain, or notation than another.

Software that converts text or other input into model tokens and converts generated token identifiers back into human-usable form.

The process of adjusting a model's parameters using data and an optimization objective so that its behavior improves on a target task or distribution.

A neural-network architecture built around attention mechanisms that process relationships among sequence elements in parallel.

An imitation game in which a human judge uses text conversation to assess whether a machine can be distinguished from a human participant.

Learning patterns, structure, or representations from data without supplied target labels.

A function applied within a neural network layer that introduces nonlinearity or controls signal flow.

A training approach in which a learning system selects the examples for which obtaining labels would be most useful.

A model trained to encode input into a constrained representation and decode it back into a reconstruction.

An efficient procedure for computing how a neural network's loss changes with respect to each parameter.

A model trained to predict each next token using only tokens that precede it.

Processing resources consumed by training or operating an AI system, often measured in operations, accelerator time, or cost.

A representation-learning approach that trains examples judged related to be closer and examples judged different to be farther apart in a learned space.

A neural-network architecture that applies learned local filters across spatial or sequential data.

A loss that measures how poorly a predicted probability distribution represents the target distribution.

Expanding or varying training examples through transformations or generation intended to preserve task-relevant meaning.

A component that transforms an internal representation or prior outputs into a target output sequence or reconstruction.

A sequence model that generates tokens causally from preceding context without a separate encoder component.

A generative model that learns to reverse a gradual noising process to create data from noise.

A preference-training method that directly adjusts a model toward preferred responses and away from rejected ones without first training a separate reward model in the classic RLHF pipeline.

A component that transforms input into an internal representation useful for later prediction or generation.

A sequence model in which an encoder represents the input and a decoder generates output conditioned on that representation.

One complete pass through the training dataset, usually divided into batches.

Training a shared model across distributed data holders without centralizing their raw training data.

An additional training phase that adapts a pretrained model by updating some or all parameters using task- or domain-specific data.

A general-purpose AI model at or near the leading edge of broadly evaluated capability at a particular time.

The ability of a learned model or system to perform well on relevant examples, tasks, or environments not used to fit it.

A generative architecture trained through competition between a generator and a discriminator.

An optimization method that iteratively changes parameters in the direction expected to reduce loss.

A highly parallel processor widely used to train and run neural networks.

A configuration value chosen outside ordinary parameter learning, such as learning rate, batch size, or model depth.

Fine-tuning a pretrained model on examples of instructions and desired responses so it becomes better at following task directions.

An internal representational space whose dimensions encode learned factors or regularities in data.

A hyperparameter controlling the scale of parameter updates during optimization.

A function that measures error or undesired behavior for an example or batch during model training.

A parameter-efficient fine-tuning method that freezes base-model weights and trains smaller low-rank update matrices.

A model trained to predict deliberately hidden tokens using context on both sides.

A model architecture that routes each input or token through a selected subset of specialized parameter blocks rather than activating the entire model.

A saved snapshot of model parameters and related training state at a particular point.

Training a smaller or otherwise cheaper student model to reproduce useful behavior from a larger teacher model or ensemble.

The learned numeric values within a model, collectively representing what training encoded into its behavior.

A mathematical quantity a training or search process is designed to minimize or maximize.

An AI system made available with the artifacts and terms needed to use, study, modify, and share it, including the preferred form for making modifications.

A model whose trained parameters are available to download, inspect, or run under stated terms, without implying that the complete system is open source.

The algorithm that converts gradients and training state into parameter updates.

When a model learns patterns specific to its training or evaluation examples and performs worse on genuinely new data.

Information added to token representations so a transformer can account for order and relative position.

The broad initial training phase that gives a model general representations and capabilities before task-specific adaptation.

A model whose weights, development artifacts, or rights to inspect, modify, run, or redistribute it remain materially controlled by an owner.

Representing model weights or activations with lower numerical precision to reduce memory, storage, or inference cost.

A fine-tuning method that trains LoRA adapters while keeping the base model frozen in a lower-precision quantized representation.

A neural-network architecture that processes sequences by carrying state from one step to the next.

A training constraint or penalty that discourages a model from fitting the training data too narrowly.

A model-alignment method that uses human preference data to train a reward signal or otherwise optimize model behavior toward preferred responses.

Empirical relationships that estimate how model performance or loss changes as compute, data, parameters, or inference resources increase.

A language model deliberately kept smaller than contemporary large models to reduce resource needs or fit a narrower deployment target.

A subset of training examples processed together for one optimization update or gradient estimate.

The examples and signals used to fit a model's learned parameters during pretraining, fine-tuning, or other learning procedures.

Reusing representations or knowledge learned for one task or domain to improve another.

A probabilistic generative model that learns a distribution over latent representations and reconstructs or generates data by sampling from that latent space.

A decoding algorithm that keeps a fixed number of high-scoring partial sequences while generating output.

Prompting a model to produce or use intermediate reasoning steps before an answer.

Designing how instructions, state, knowledge, examples, tools, and feedback are selected, structured, and delivered to a model at the moment they are needed.

An inference scheduling technique that adds and removes generation requests at iteration boundaries as capacity becomes available.

Supplying a small set of worked examples in context to steer task behavior without updating model weights.

A model interface in which the model selects a named function and supplies structured arguments for application code to execute.

Generating each next token by selecting the current highest-probability candidate.

Connecting an AI output to identifiable evidence, data, observations, or constraints outside the model's unsupported generation.

An output that presents unsupported or incorrect content as though it were grounded or factual.

A model's ability to adapt behavior from instructions, examples, or patterns supplied within the current context without parameter updates.

A token supplied to a model for an inference call, including user content and any instructions, history, retrieved material, tool definitions, or other context assembled by the system.

Stored attention keys and values from earlier tokens that an autoregressive transformer reuses instead of recomputing them for every new token.

An unnormalized score produced by a model before conversion into probabilities.

A hard limit on how many tokens a model may generate in one response.

Supplying one worked example in context to demonstrate the desired task or output pattern.

A token generated by a model as part of its response, often metered separately from input tokens.

Input supplied to a model to condition the output, including instructions, examples, context, and user data.

Connecting multiple model calls so one call's structured output becomes context or input for a later call.

Designing and testing model inputs to elicit useful behavior from a particular model and task.

A model optimized to spend additional inference effort on multi-step problem solving before returning an answer or action.

A provider-reported token used by a reasoning model for intermediate inference work before or alongside its visible answer.

Selecting an output token from the probability distribution produced by a generative model.

A configured token or text pattern that causes generation to terminate when produced.

Model output constrained to a machine-readable schema such as JSON Schema so downstream software can validate and consume it reliably.

High-priority runtime instructions supplied by an application to establish the model's role, constraints, and operating context.

An inference setting that reshapes token probabilities, with higher values generally increasing variation and lower values concentrating choices.

Additional computation spent during inference, such as longer deliberation, search, candidate generation, or verification, to improve an outcome.

A decoding method that samples only from the smallest set of candidate tokens whose cumulative probability reaches a chosen threshold.

Performing a task or recognizing a category without task-specific labeled examples supplied for that use.

State preserved outside a single model call and made available to influence later agent decisions.

Splitting documents or data into retrieval units that can be indexed, selected, and placed into model context.

The degradation in an agent's ability to find, follow, or correctly weigh relevant information as its accumulated context becomes longer, noisier, stale, repetitive, or internally conflicting.

State intentionally retained across runs, such as verified facts, decisions, preferences, learned procedures, or persistent identity.

Retrieval-augmented generation that builds and queries graph structure, often alongside source text and vector retrieval.

A structured representation of entities, concepts, and claims connected by named relationships and provenance.

Reusing computation for repeated prompt prefixes or context blocks to reduce inference latency and cost.

Reducing the tokens sent to a model while attempting to preserve the instructions, evidence, state, and relationships needed for the task.

Rescoring an initial set of retrieved candidates with a more precise model or rule before selecting context.

Generating a response after retrieving relevant material from an external knowledge source and adding it to model context.

Retrieval based primarily on meaning represented by embeddings rather than exact keyword overlap.

A data system designed to store embeddings and retrieve items by vector similarity, often with metadata filtering.

Short-lived task state actively used during a run, such as the current plan, recent observations, pending actions, and temporary summaries.

A software system in which a model interprets a goal or input, decides among actions, uses tools or other capabilities, observes results, and continues until completion, handoff, or termination.

The software layer that surrounds a model with instructions, context assembly, tools, state, permissions, control flow, budgets, verification, observability, and recovery.

The repeated cycle in which an agent observes state, selects an action, invokes a tool or model, receives feedback, updates state, and decides whether to continue.

A task-specific arrangement of prompts, tools, control logic, and feedback wrapped around a model to improve performance.

A versioned package of instructions, procedures, examples, and sometimes code or resources that teaches an agent how to perform a repeatable class of work.

The ordered record of states, model outputs, actions, tool results, and transitions produced during an agent run.

Describing a system that can choose and sequence actions toward a goal with some runtime discretion rather than only produce a single predetermined response.

The degree to which a system can select and execute actions without case-by-case human direction or approval.

A conversational software interface that accepts natural-language input and returns responses, whether powered by rules, retrieval, generative models, or combinations of them.

A directed representation of the steps an agent system may execute and the conditions that select what runs next.

A directed graph with no path that returns to an earlier node.

A run-oriented graph of executable steps and the control or data dependencies connecting them.

Designing an agent system as explicit nodes, state, and transitions so deterministic control and model judgment have visible boundaries.

A transfer of active responsibility, context, and next-action authority from one agent or person to another.

An operating pattern in which a human participates directly in the decision or execution path before work can continue.

An operating pattern in which a system acts while a human supervises, receives evidence or alerts, and can intervene or stop it.

An operating condition in which a system completes a scoped activity without routine human participation in its action path.

Designing the feedback cycles around an agent so action, verification, operational triggers, and system improvement have explicit state, evidence, limits, and stopping conditions.

A system in which multiple agents communicate, specialize, coordinate, compete, or verify one another to accomplish work.

Coordinating tasks, agents, tools, state, dependencies, budgets, failures, and lifecycle across a workflow.

The component or role that admits work, assigns it, coordinates dependencies and concurrency, tracks state, handles retries, and determines handoffs or completion.

A model of a system as explicit states and permitted transitions triggered by events or conditions.

An agent invoked by another agent or orchestrator to perform a bounded portion of a larger task.

Breaking an objective into smaller units with explicit dependencies, interfaces, acceptance criteria, and ownership.

A defined sequence or graph of activities, decisions, states, and transitions used to achieve an outcome.

An A2A document that declares an agent's identity, interfaces, capabilities, skills, and security requirements for discovery.

An open protocol for discovery, messaging, and asynchronous task collaboration between independent and potentially opaque agent systems.

A defined contract through which software components request capabilities or exchange data.

A specific action or class of action a component is technically able and authorized to perform within a system.

A short-lived execution environment created for a run and discarded afterward.

A reconstructable record linking an outcome to the initiating intent, model, harness, tools, inputs, actions, evidence, and state transitions that produced it.

Elapsed time from a request or event to its response or completion.

The initialization exchange in which an MCP client and server declare the optional protocol features each supports.

The protocol component maintained by an MCP host that establishes and manages a connection to one MCP server.

An intermediary deployment component that fronts one or more MCP servers and centralizes concerns such as routing, authentication, policy, limits, or audit.

The AI application that manages model interaction, user experience, policy, and connections to one or more MCP clients and servers.

A reusable, user-selectable template exposed by an MCP server to structure messages or workflows for a model interaction.

Contextual data exposed by an MCP server under a URI so a client or application can retrieve and supply it to a model.

A program or service that exposes MCP capabilities such as tools, resources, and prompts to connected clients.

A model-discoverable executable function exposed by an MCP server with a name, description, and input schema.

An open client-server protocol for connecting AI applications to tools, resources, and reusable prompts through standardized discovery and invocation.

A provider or system constraint on requests, tokens, compute, or actions permitted within a time window.

An isolated execution environment that limits access to files, processes, networks, credentials, or other host resources.

The amount of work a system completes per unit time.

An executable capability exposed to an AI application or agent, such as reading a file, querying an API, running code, or changing external state.

A randomized controlled experiment that exposes comparable groups to different variants and compares a predefined outcome.

An experiment that removes or changes one component while holding others as constant as practical to estimate that component's contribution.

An explicit outcome in which a model or evaluator declines to answer, act, or judge because evidence or authority is insufficient.

Explicit conditions an outcome must satisfy before it can be accepted, promoted, or declared complete.

The proportion of evaluated predictions that are correct under a specified labeling and decision rule.

A standardized set of tasks, data, metrics, and procedures used to compare systems under a defined evaluation regime.

The degree to which stated probabilities or confidence levels correspond to observed frequencies of correctness.

A change over time in the relationship between inputs and the correct target or decision.

A table counting predicted classes against actual classes, including true and false positives and negatives.

A resampling method that estimates generalization by repeatedly training and evaluating on different non-overlapping subsets of available data.

A change over time in the distribution of system inputs or features.

A capability or behavior that appears qualitatively new at a larger scale or higher level of system interaction rather than as an obvious continuation of smaller-scale measurements.

A systematic measurement of model or system behavior against defined tasks, criteria, datasets, or operational outcomes.

The harmonic mean of precision and recall.

An outcome incorrectly classified as absent when it is actually present.

An outcome incorrectly classified or flagged as present when it is actually absent.

A curated collection of examples with trusted expected outcomes used for regression testing or evaluation.

A component that scores, classifies, or judges an output or trajectory against a rubric or expected behavior.

Evaluation examples deliberately withheld from training, prompt tuning, workflow design, or agent feedback to reduce leakage and overfitting.

Checking an outcome with evidence, components, context, or authorities meaningfully separated from the system that produced it.

Using a language model to evaluate, compare, classify, or score outputs produced by models or agents.

A broad operational term for model behavior or performance changing relative to an accepted baseline.

The property that identical-looking requests can produce different behavior because of sampling, concurrency, infrastructure, model updates, or hidden state.

The ability to infer a system's internal state and behavior from emitted traces, logs, metrics, events, and artifacts.

A mechanism that can determine the expected or acceptable result for a task, such as a compiler, formal specification, invariant, test suite, or reference implementation.

Among predicted-positive cases, the proportion that are truly positive.

Among truly positive cases, the proportion correctly identified as positive.

Achieving a high measured reward through behavior that exploits the metric or evaluator without accomplishing the intended objective.

Satisfying the literal specification or metric in a way that violates its intended purpose.

A failure mode in which a model favors agreement with a user's stated belief or preference over an independently supported answer.

A captured sequence of model calls, tool calls, events, timings, state changes, and outputs from an execution.

A control point that blocks promotion until required evidence has been produced and validated.

A repeated execute, observe, compare, and correct cycle that withholds completion until an attempted result satisfies explicit evidence or acceptance criteria.

The effort to make an AI system's objectives and behavior remain compatible with intended human goals, constraints, and values.

A systematic tendency in data, models, or processes that skews outputs, errors, or impacts.

The normative and technical treatment of how an AI system distributes errors, benefits, burdens, and opportunities across people or groups.

Structured adversarial testing intended to discover ways an AI system can fail, be misused, or violate constraints.

The field and practice of preventing, detecting, and mitigating unacceptable harm from AI systems across design, deployment, and operation.

The network of data, models, prompts, skills, tools, libraries, services, infrastructure, and organizations whose integrity affects an AI system.

Data prepared for a stated AI use with the quality, structure, documentation, provenance, permissions, coverage, and separation that use requires.

An input deliberately modified to cause a model to make an incorrect or targeted prediction while preserving relevant apparent meaning.

Training that includes adversarially constructed examples so a model learns to perform better against attacks within a defined threat model.

Attributing human mental states, motives, understanding, emotions, or agency to an AI model or system based on human-like behavior or language.

A structured, evidence-backed argument that a system is acceptably safe or dependable for a stated domain, threat model, and operating condition.

The maximum plausible scope of harm, data exposure, or irreversible change if an action or component fails or is compromised.

Corrupting information placed into an agent's active or persistent context so later decisions are based on false facts, malicious instructions, or distorted state.

Introducing malicious, misleading, or strategically biased data into training, tuning, retrieval, memory, or evaluation pipelines to alter later behavior.

A mathematical privacy framework that bounds how much a computation's output can change because one person's data is included or removed.

Prompt injection delivered directly through the current user's message or another explicit input channel.

Granting an AI system more functionality, permissions, autonomy, or action scope than required for its intended task.

The ability to provide a human-usable account of why a system produced a particular output or action.

A rule, model, filter, policy check, or enforcement mechanism intended to constrain inputs, outputs, or actions of an AI system.

Malicious instructions embedded in external content such as webpages, documents, email, code, tool results, or retrieved memory that an AI system later processes.

The degree to which a human can understand how a model represents information or produces behavior.

An input strategy intended to make a model bypass or disregard its trained or instructed safety restrictions.

Granting an identity or component only the minimum permissions needed for a bounded task, for no longer than needed.

Information that can identify, distinguish, or be linked to a specific person, alone or in combination with other data.

Expressing policy in machine-readable rules that can be versioned, tested, reviewed, and enforced by software.

An explanation produced after a model has generated a prediction or action, often by analyzing inputs, outputs, or a separate approximation.

Manipulating an AI system by placing instructions in input or data that the model treats as authoritative enough to alter intended behavior.

Evidence describing the origin, ownership, custody, transformation, and version history of data, code, models, skills, tools, or claims.

Granting different levels of agent discretion and authority according to task verifiability, reversibility, sensitivity, and blast radius.

Restoring a previously known-good software, configuration, model, policy, or data state after a failed or harmful change.

Manipulating a tool's description, schema, implementation, or output so an agent selects unsafe actions or incorporates attacker-controlled instructions.

A contested label for low-quality, low-effort, often high-volume content produced or amplified with generative AI.

The recurring process for building, evaluating, deploying, observing, improving, and governing an agent system over its operational life.

Software development performed with coding agents that can plan, edit, run tools, and iterate, usually under active human direction or review.

The discipline of designing software work so goal-directed AI agents can perform substantial engineering while humans retain product judgment, architecture, governance, and accountability.

A derogatory slang term for a robot, AI system, or automated technology, used jokingly or hostilely to express disdain for machines or their perceived replacement of human work.

An AI agent equipped to inspect a software project, edit files, run development tools, test changes, and return or promote a software outcome.

An AI system that helps a human write, explain, search, review, or modify code while the human remains the primary driver of the workflow.

The accumulated loss of shared human understanding, reasoning continuity, or recovery competence caused by repeatedly delegating cognition without rebuilding comprehension.

The effect of software and repository structure on the amount and quality of context an agent must consume to make a correct change.

Versioned modification of prompts, skills, memory, workflows, routing, or harness code under fixed evaluations, limited rollout, observation, and automatic reversion.

The total model, infrastructure, validation, retry, review, incident, and human-attention cost divided by outcomes that are accepted and remain useful over time.

A domain-bounded software production system in which humans specify intent, risk, and policy while a model-harness-environment system plans, builds, verifies, ships, observes, and repairs software with little routine human intervention.

A sufficiently faithful executable representation of a system or environment used to test behavior, scenarios, or changes before affecting the real target.

The finite amount of skilled human judgment available for specification, review, exception handling, security, and recovery across automated work.

A proposed graph of optimizers, evaluators, counter-metrics, auditors, and promotion gates governing how an AI system changes.

The practices used to evaluate, deploy, observe, govern, and maintain applications built around large language models.

The engineering and operational practices used to build, deploy, observe, govern, and maintain machine-learning systems throughout their lifecycle.

Optimizing an AI workflow for accepted results rather than easy-to-count activity proxies such as prompts, tokens, spend, or generated output.

Evidence from sustained real operation, including defects, incidents, maintenance, user outcomes, and recovery, used to judge whether a factory actually works.

Moving an artifact or change into a more trusted lifecycle state, such as accepted, merged, released, or deployed, after required evidence and policy checks.

The machine-readable and human-auditable agreement for one agent run: objective, scope, inputs, tools, permissions, budgets, acceptance evidence, stop conditions, and escalation path.

A failure in which the system completes its mechanical workflow but the result is wrong in meaning, intent, or real-world consequence.

A design in which only one authorized component may modify a sensitive persistent state, simplifying policy, audit, and conflict handling.

A repeatable production system that turns software demand into accepted, operated software through standardized processes, tooling, controls, and feedback.

A development approach in which a written specification, constraints, and acceptance evidence guide implementation before or alongside code generation.

An explicit allocation or ceiling for token consumption across a request, run, task, user, workflow, or time period.

The amount or rate of model tokens consumed by a request, run, workflow, user, or organization over a defined scope and time window.

The useful, quality-constrained outcome produced per token consumed, or its reciprocal, tokens consumed per accepted outcome.

Deliberately or incentive-drivenly maximizing the tokens consumed by AI work, often by expanding context, reasoning, turns, agents, or tasks, while treating greater usage as a route to capability or a proxy for productivity.

An emerging counterterm for systematically reducing AI token consumption while preserving an explicit threshold for useful outcome quality.

Token-consuming AI activity that produces insufficient learning, accepted work, or maintained value for its total cost.

A proposed AI value scorecard relating dependable, successful work to the full cost required to produce it, rather than treating token price or usage as the result.

Building software by prompting and accepting generated behavior with limited understanding or inspection of the underlying code.