AI glossary

A controlled comparison between two versions of something to see which performs better. In AI products, an A/B test might compare two prompts, models, interfaces, or workflows with similar users and tasks.

A measure of how often a system gets the answer right. Accuracy can be useful, but it can hide important problems when some kinds of mistakes matter much more than others.

An AI-driven system that can pursue a goal by deciding what to do next, using tools, observing results, and continuing until it reaches a stopping point. A chatbot mainly responds; an agent can take actions.

The repeating cycle behind many agents: understand the goal, decide on an action, use a tool or perform a step, inspect what happened, and decide what to do next.

Describes AI behavior that involves taking multiple steps or actions toward a goal rather than producing a single response. The more decisions the system can make on its own, the more agentic the workflow is.

A workflow in which AI decides at least some of the steps, tools, or sequence needed to reach an outcome. This differs from a fixed automation where every step is predetermined.

A debated term for a hypothetical AI system with broad, flexible capability across many kinds of intellectual work rather than strength in only narrow tasks. There is no universally accepted definition or test for AGI.

A broad term for computer systems that perform tasks associated with human intelligence, such as understanding language, recognizing patterns, generating media, making predictions, planning, or solving problems.

An AI product designed to help a person with tasks through conversation or other interfaces. An assistant may answer questions, create content, search information, use tools, or perform actions depending on its capabilities.

Using AI as part of a process that runs with little or no manual work. AI automations are useful when some steps require interpretation or generation rather than simple fixed rules.

The policies, responsibilities, controls, and oversight an organization uses to manage how AI is selected, used, monitored, and evaluated.

The practices used to reduce the chance that AI systems cause harm, behave unpredictably, expose sensitive information, or are misused.

A WorkingLab term for someone who is exploring or occasionally using AI but has not yet made it a systematic part of how they accomplish work.

A WorkingLab term for someone who can naturally choose and combine AI models, context, resources, tools, workflows, and human judgment to accomplish work. AI-native does not mean letting AI do everything; it means understanding how to work effectively with it.

A defined method or set of rules for solving a problem. Machine-learning systems use algorithms to learn patterns from data and to make predictions or generate outputs.

The effort to make an AI system behave in ways that match intended goals, instructions, safety requirements, and human preferences.

Information added to data to label or describe it. For example, humans might label emails as spam or not spam so a model can learn from those examples.

An AI company that develops the Claude family of models and products, including Claude, Claude Code, and Claude Cowork.

A structured way for one piece of software to communicate with another. An AI API lets a website, app, script, or agent send information to an AI service and receive a result without using the normal chat interface.

A specific address or function exposed by an API for a particular action, such as generating text, creating an image, retrieving data, or updating a record.

A secret credential used by software to authenticate with an API. Treat API keys like passwords; exposing one can let someone else use the service or incur charges on your account.

A user-facing product built around one or more AI models plus instructions, tools, memory, data, and an interface. The app is not the same thing as the underlying model.

A point in an AI workflow where a human must explicitly allow an action before it occurs. Approvals are commonly used before sending messages, deleting data, spending money, publishing content, or making consequential changes.

A file, document, image, spreadsheet, piece of code, presentation, or other useful output produced during AI-assisted work. Some AI products also use “artifact” as a specific product feature name.

A hypothetical form of AI that would exceed human capability across most or all intellectual domains. It is a speculative concept, not a current product category.

A mechanism that helps a neural network decide which parts of the input are most relevant when processing information. In language models, attention helps the model relate words and ideas across the context.

A record of actions that occurred in a system: who or what did something, when it happened, and sometimes what data or tools were involved. Audit logs are important when AI is taking actions in business systems.

An agent allowed to pursue a goal with relatively little human intervention. “Autonomous” is a matter of degree; most practical systems still operate within permissions, limits, and approval rules.

A model that generates output step by step, using what has already been generated as part of the basis for predicting what comes next. Most modern text-generating LLMs work this way.

A broadly trained model before it has been customized for a specific application or task. A base model can be further instruction-tuned, fine-tuned, or wrapped inside an application.

A standardized test or collection of tasks used to compare AI systems. Benchmarks are useful indicators, but strong benchmark scores do not guarantee that a model will be best for your real use case.

A systematic tendency in data, a model, or a process that favors some outcomes over others. Bias can come from training data, design choices, prompts, evaluation methods, or the surrounding human system.

In Git, a separate line of development that lets you change code without immediately changing the main version. AI coding agents often create or work on branches so changes can be reviewed safely.

The ability of an AI system to navigate websites, read pages, click controls, enter information, and sometimes complete tasks through a browser.

In software development, the process of turning source code and related assets into a runnable or deployable application. AI coding tools may run builds to check whether changes work.

Reusing previously processed prompt or context data so the system does not have to process the same information from scratch every time. Caching can reduce latency and cost.

How well a system’s stated confidence matches its actual likelihood of being correct. A well-calibrated system that says “80% confident” should be right about 80% of the time in comparable cases.

A prompting technique that asks an AI to break a problem into intermediate steps or explain its reasoning process. Some AI systems may use internal reasoning that is not exposed directly to users.

A conversational interface for interacting with an AI system through a sequence of messages. A chat is an interface; the underlying model may also be accessible through APIs, agents, coding tools, or other environments.

A conversational software system that responds to messages. A chatbot may use AI, fixed rules, or both; modern AI chatbots are usually built around language models.

OpenAI’s user-facing AI product for conversation, research, creation, analysis, and other tasks. ChatGPT combines AI models with product features such as files, projects, memory, apps, tools, and specialized work experiences.

A ChatGPT workspace that groups related chats, files, and instructions so work can stay organized around a topic or objective. Project memory behavior depends on the project and account settings.

An OpenAI work environment designed for longer research, analysis, and deliverable creation such as documents, spreadsheets, and presentations.

A saved state of a model during or after training. A checkpoint contains model parameters that can be loaded later for inference, evaluation, or additional training.

A smaller piece of a larger document or dataset. Retrieval systems often split long material into chunks so they can find and return only the most relevant parts.

The process of splitting documents or data into smaller pieces for retrieval, indexing, or processing. Good chunking helps an AI retrieve useful context without flooding the context window.

A reference showing where a claim or piece of information came from. Citations make AI-generated research easier to inspect and verify; the cited source still needs to be checked.

Anthropic’s AI assistant and model platform for language, reasoning, analysis, coding, and other tasks.

Anthropic’s agentic coding tool. It can read a codebase, edit files, run commands, use development tools, and work across multiple steps from the terminal, IDE, desktop app, or browser.

Anthropic’s environment for longer-running agentic work beyond ordinary chat. Cowork can work with files, apps, connectors, and computer access depending on the user’s setup and permissions.

A text-based way to control software by typing commands in a terminal. Tools such as Codex and Claude Code can operate through a CLI even if the user does not know the commands beforehand.

Computing resources that run on remote servers rather than on your own device. Cloud AI tasks can continue independently of your laptop, depending on the product.

OpenAI’s coding agent for writing, reviewing, testing, and shipping code. Codex can work in local development tools or complete delegated work in isolated cloud environments.

An AI agent specialized for software work. It may inspect a codebase, plan changes, edit multiple files, run commands and tests, find bugs, review code, and prepare changes for deployment.

In Git, a saved set of changes with a record of what changed. Commits create checkpoints in a project’s history that can be reviewed or reverted.

The ability of an AI system to interact with a computer interface using actions such as viewing screenshots, moving a pointer, clicking, and typing. This gives AI access to software that may not expose a direct API.

AI techniques for understanding or analyzing visual information such as photographs, diagrams, screenshots, documents, and video.

A numerical estimate of how confident a system is in a prediction. Not every generative AI response has a meaningful confidence score, and fluent wording should never be mistaken for confidence.

A connection that lets an AI product access an external service or data source such as cloud storage, email, a database, CRM, project tracker, or knowledge system. Different platforms use “connector,” “app,” and “integration” somewhat differently.

An isolated software environment packaged with the code and dependencies needed to run something consistently. AI coding systems often use containers or similar environments to execute code safely.

A system that detects, blocks, limits, or flags content based on safety or policy rules.

A product name used by Microsoft and others for AI assistants embedded in software. Because “copilot” is now used broadly, check which product and underlying model someone means.

An AI-focused code editor built around software-development agents and model-assisted coding. It is one of several environments where developers can use AI directly inside a codebase.

A customized ChatGPT experience configured with specific instructions, knowledge, capabilities, and sometimes apps or actions. A custom GPT is an application configuration, not a newly trained foundation model.

The information available to the model for the current task. Context can include the conversation, instructions, uploaded files, retrieved documents, tool results, memories, and other information inserted by the application.

Reducing a long conversation or working state into a shorter representation so important information can remain available without using as much context space.

Designing what information, instructions, tools, memory, and retrieved resources an AI receives so it can perform a task well. It is broader than prompt engineering because the prompt is only one part of the working context.

The maximum size of a model’s context window, usually measured in tokens. Inputs, tool results, retrieved content, and sometimes output allowances all compete for finite space.

An informal term for the decline in usefulness that can happen when a context becomes too long, cluttered, repetitive, contradictory, or full of irrelevant information.

The maximum amount of information a model can consider at one time during a request. The application decides what to place inside that window, which may include more than the visible chat.

A mathematical method commonly used to measure how similar two embedding vectors are. In practical terms, it helps systems find items that are close in meaning.

A common way API providers price language-model usage: a price for processing input tokens and another price for generating output tokens. Some tools or media models use different pricing units.

A simple text format for tables and datasets. AI tools can often read, analyze, transform, and generate CSV files.

The unauthorized transfer of data out of a system. In AI workflows, prompt injection or overly broad tool permissions can sometimes create a path for sensitive information to be exposed.

Sensitive, private, proprietary, or evaluation information appearing where it should not. Leakage can happen through prompts, logs, training data, tool access, outputs, or poorly separated datasets.

A record of where data came from, how it was changed, and where it moved. Lineage helps organizations understand and audit AI inputs and outputs.

Information about the origin and history of data or content. Provenance helps answer questions such as: Where did this fact, image, dataset, or document come from?

The geographic region or jurisdiction where data is stored or processed. This can matter for enterprise security, privacy, and regulatory requirements.

How long a provider or organization keeps prompts, files, outputs, logs, or other data after use.

An organized system for storing and retrieving information. AI tools may query databases directly, generate database queries, or use databases as sources of truth.

A collection of examples or records used for training, evaluation, analysis, or retrieval.

The part of some neural-network architectures that generates output from an internal representation. Many modern LLMs use decoder-focused Transformer architectures for text generation.

Machine learning based on neural networks with many layers. Modern language, vision, audio, and generative AI systems are largely built with deep-learning techniques.

A general name used by several AI products for longer, multi-step research that searches or retrieves many sources, analyzes them, and produces a synthesized result. Features and depth vary by platform.

Synthetic or manipulated media designed to convincingly imitate a real person, event, voice, image, or video.

Software or a library that another piece of software relies on. AI coding agents often install, update, or debug dependencies while working on a project.

Putting software into an environment where people or systems can actually use it. Deployment might mean publishing a website, releasing an application, or moving a service into production.

A process that produces the same result from the same inputs and conditions. Many normal software operations are deterministic; generative AI is usually at least partly probabilistic.

A view showing what changed between two versions of a file or codebase. Reviewing diffs is one of the safest ways to inspect changes made by an AI coding agent.

A kind of generative model commonly used for image generation. It learns to create images by reversing a process that gradually adds noise to training images.

A digital representation of a real object, system, place, or process that is updated with data and may be used for monitoring, analysis, or simulation. A digital twin is not automatically an AI system.

Training a smaller or more efficient model to imitate capabilities or outputs from a larger model. The goal is usually to retain useful performance while reducing cost or computation.

A change between the kind of data a model learned or was tested on and the data it encounters in real use. Performance can fall when the real world differs from the model’s training or evaluation conditions.

A numerical representation of meaning. Text, images, or other data are converted into lists of numbers so a computer can compare how similar or related they are.

The part of some neural-network architectures that converts input into an internal representation. Encoder models are commonly used for understanding, classification, and embeddings.

A method of encoding data so it can be read only by someone or something with the correct key. Encryption protects data in storage or transit but does not eliminate every AI security risk.

The software, configuration, files, permissions, and services available while a program or agent is running. Differences between environments are a common source of software problems.

A configuration value stored outside the program itself, often used for API keys, database addresses, or deployment settings.

A structured test used to measure whether an AI system performs the behavior you care about. Good evals use realistic tasks and clear success criteria rather than relying only on general benchmarks.

A set of examples reserved for measuring how well a model or AI workflow performs. It should not be used to train or tune the system in a way that makes the test meaningless.

Techniques for making an AI system’s behavior or decisions easier for people to understand. Exact explanations can be difficult for complex models, so explainability often relies on evidence, tests, feature analysis, or observable behavior rather than a literal account of every internal computation.

A metric that balances precision and recall into a single score. It is useful when both missing true cases and incorrectly flagging false cases matter.

The goal of ensuring an AI system does not produce unjustified or harmful differences in treatment across people or groups.

A case where a system says something is absent or negative when it is actually present or positive. Example: a fraud detector misses a fraudulent transaction.

A case where a system says something is present or positive when it is actually absent or negative. Example: a spam filter incorrectly labels a legitimate email as spam.

Giving an AI several examples of the kind of input and output you want before asking it to perform the task.

Searching uploaded or connected files to find information relevant to a request. File search is often one component of retrieval-augmented generation.

Additional training performed on an existing model to change how it behaves on particular tasks, styles, or domains. Fine-tuning changes model parameters; simply giving the model documents at runtime does not.

A large, broadly trained model that can be adapted or used for many different tasks. Modern language, vision, and multimodal models are often foundation models.

An informal term for a highly capable, leading-edge AI model near the current state of the art. There is no single technical threshold that defines “frontier.”

A structured way for a model to request that software run a particular function with specified arguments. It is a common mechanism behind tool use.

Google’s family of AI models and user-facing AI products. Capabilities and product packaging vary across Google services.

AI that creates new content such as text, images, audio, video, code, or structured data based on patterns learned from training data and information provided at runtime.

A version-control system that records changes to files over time. It is widely used for software projects and makes it possible to compare, branch, merge, and revert changes.

A cloud platform for hosting Git repositories and collaborating on software. AI coding agents often connect to GitHub to read projects, create branches, open pull requests, or review code.

A carefully curated set of trusted examples used to test an AI system. A strong gold set reflects the kinds of real tasks and edge cases that matter to your use case.

The trusted answer, label, measurement, or outcome used as the reference for evaluating a prediction or model output.

A name for a family of Transformer-based generative models. OpenAI uses GPT as part of many model and product names; in casual conversation, people sometimes incorrectly use “GPT” to mean any chatbot or LLM.

Connecting an AI response to specific trusted information—such as documents, databases, search results, or tool outputs—rather than relying only on what the model learned during training.

A rule, filter, approval step, validation check, permission limit, or other control designed to constrain AI behavior and reduce risk.

An AI output that sounds plausible but is false, unsupported, invented, or inconsistent with the available evidence. Hallucinations are a normal failure mode of generative models and must be managed with grounding, verification, and good workflow design.

Running a website, application, database, or service on infrastructure that makes it accessible to users or other systems.

The broader practice of keeping people responsible for monitoring, reviewing, and governing AI behavior rather than assuming the AI is self-validating.

A workflow where a person reviews, approves, corrects, or guides AI at important points. Human involvement is especially important when errors have meaningful consequences.

Combining multiple search methods—commonly keyword search and semantic/vector search—to improve retrieval.

Software used to write, navigate, test, and debug code. AI coding tools often integrate directly into IDEs.

Using AI to modify an existing image—for example, removing an object, changing a background, adjusting style, or adding new content.

Generating or transforming an image using another image as part of the input.

A prompt-injection attack hidden inside content the AI reads, such as a webpage, document, email, or tool result. The malicious instruction is not typed directly by the user.

Using a trained model to produce a prediction or output. Training teaches the model; inference is what happens when you actually use it.

AI image editing that replaces or generates content inside a selected region of an existing image.

The information sent into an AI system for a particular request: prompts, files, images, audio, tool results, retrieved information, or other context.

A token contained in the information sent to a language model. API pricing and context limits often distinguish between input and output tokens.

A direction telling an AI what to do, how to behave, what constraints to follow, or what output to produce. Instructions can come from the user, system, project, skill, tool, or application.

Additional model training designed to make a model better at following natural-language instructions and behaving like an assistant.

A connection between software systems that allows them to exchange data or trigger actions. An integration may use an API, MCP, a connector, a plugin, a webhook, or another mechanism.

An attempt to persuade or manipulate an AI system into bypassing its safety rules or restrictions.

A widely used text format for structured data. AI systems often use JSON when exchanging information with software or returning structured results.

A product feature or prompting approach that constrains a model to return JSON rather than ordinary prose. Exact behavior varies by platform.

Search based mainly on matching the words or phrases in a query with words in the source material. It can be precise for known terms but may miss conceptually similar material that uses different wording.

An organized collection of information used by people or AI systems as a reference. It may contain documents, database records, help articles, policies, notes, or other sources.

The date where a model's training data ends. Built-in coverage thins on the approach, so treat the published date as a warning line rather than a guarantee on either side; anything later can only reach the model through the conversation — pasted by you, or fetched in by a tool such as web search. Each model has its own cutoff, published in its documentation.

A structured representation of entities and the relationships between them. Knowledge graphs can help systems reason about how people, organizations, concepts, places, or events are connected.

A WorkingLab term for a short, outcome-first experiment in which a member uses AI to accomplish a practical use case, sees how the system is assembled, and learns the underlying practice through use.

A language model trained at large scale to predict and generate sequences of tokens. LLMs can perform many language-related tasks because they learn statistical patterns across enormous amounts of text and other training data.

The time between making a request and receiving a response or result. Lower latency means the system feels faster.

An internal mathematical space where a model represents learned features or concepts. Similar ideas or visual characteristics may be located near one another in this representation.

A security principle: give a user, application, or AI agent only the permissions it actually needs. This limits the damage if the system makes a mistake or is manipulated.

A tool that checks code for likely errors, style problems, or suspicious patterns without necessarily running the program. AI coding agents often use linters as part of verification.

A computing environment running on your own machine rather than in a remote cloud service.

A model designed to accept unusually large amounts of information in a single context window. A large context window helps, but it does not guarantee perfect recall or reasoning over everything inside it.

Information an AI application can retain or retrieve across sessions. Long-term memory is usually an application feature built around storage and retrieval; it is not the model continuously thinking between conversations.

A branch of AI in which systems learn patterns from data rather than relying only on rules explicitly programmed by humans.

A lightweight plain-text format for headings, lists, links, tables, and simple formatting. AI systems work well with Markdown because it is easy for both humans and machines to read.

The maximum number of tokens a model is allowed to generate in a response. A model may stop earlier if the answer is complete or another stopping condition is reached.

The AI application or host that connects to MCP servers and uses the capabilities they expose.

A user-interface extension delivered through the MCP ecosystem that can let connected tools present interactive experiences inside an AI application. Support depends on the MCP client and specification version.

A reusable prompt or interaction template exposed through an MCP server. MCP capabilities evolve over time, so availability depends on the specification and client.

Data or content exposed by an MCP server for an AI application to read, such as documents, records, or other contextual information.

A service that exposes tools, data, or other capabilities to AI applications through MCP.

An action exposed by an MCP server that an AI application can invoke, such as searching a database, updating a ticket, or running a calculation.

Information an AI application saves and later brings back when it is useful. Memory is separate from a model’s fixed training and separate from the temporary context window, although remembered information may be inserted into context.

A model architecture made of multiple specialized internal components, or “experts.” For a given token or task, the system activates only some of them rather than using every parameter at once.

A trained mathematical system that transforms inputs into outputs based on patterns learned from data. The model is the underlying intelligence engine; the chat app or product around it may add memory, tools, files, search, safety controls, and an interface.

Documentation describing a model’s intended uses, capabilities, limitations, evaluation results, and sometimes safety information.

A product control that lets a user choose which AI model or reasoning mode handles a task. Different models may trade off speed, cost, context size, reasoning ability, media support, and tool access.

Automatically choosing which model should handle a request based on the task, cost, speed, difficulty, availability, or other rules. Some AI products route requests without requiring the user to pick a model manually.

An open standard for connecting AI applications to external tools, data sources, and workflows in a consistent way. MCP reduces the need for every AI product to invent a different custom integration for the same service.

A change in a model’s behavior or performance over time, often because the model, surrounding product, data, prompts, or real-world conditions have changed.

Automated or human review used to detect or manage content that violates safety, legal, platform, or community rules.

A system where multiple AI agents or subagents perform different roles and coordinate their work. Multiple agents can add specialization, but they also add complexity, cost, and more ways for errors to propagate.

Able to work with more than one type of information, such as text, images, audio, video, or files. A multimodal model may accept multiple kinds of input, produce multiple kinds of output, or both.

AI designed or trained to perform particular classes of tasks rather than possessing broad human-like intelligence across all domains.

Human language such as English, Spanish, Japanese, or conversational speech, as opposed to programming languages or formal machine instructions.

An interface that lets people control software or access information using ordinary human language instead of menus, code, or specialized commands.

A test of whether a model can find and use a small but important piece of information buried inside a large amount of context.

In some generative-media tools, instructions describing what should not appear in the generated result.

A machine-learning model made of connected layers of mathematical units that learn patterns by adjusting numerical parameters during training.

A widely used authorization standard that lets one service grant another limited access without sharing the user’s password. AI connectors and MCP integrations often use OAuth to request specific permissions.

Technology that converts text visible in images or scanned documents into machine-readable text.

An AI company that develops ChatGPT, Codex, GPT models, and related AI platforms and developer tools.

Giving the AI one example of the desired task or output before asking it to perform a new instance.

A WorkingLab term for a public or shared live experiment that demonstrates a practical AI use case from outcome through execution, inspection, and learning.

A model released under terms intended to permit broad inspection, modification, or redistribution. Exactly what counts as “open source” for AI is debated because model code, weights, training data, and training methods may have different access rights.

A model whose trained weights are available for others to download or run. Open weights do not necessarily mean the training data, source code, or full development process is open.

The component or agent that coordinates a larger workflow, deciding which tools, agents, or steps should handle different parts of the job.

AI image generation that extends an existing image beyond its original boundaries.

The result produced by an AI system: text, code, an image, audio, a tool request, structured data, or another artifact.

A token generated by a language model as part of its response.

A numerical value inside a machine-learning model that is adjusted during training. Modern models can contain billions or more parameters; these values encode learned patterns rather than readable facts or rules.

An authorization that determines what data or actions an AI system can access. Permissions should be limited to what the task actually requires.

An AI search and research product that combines language models with web retrieval and citations. Like other AI research systems, its outputs still require source inspection for consequential work.

A set of instructions that asks an AI to adopt a particular role, perspective, expertise, or communication style. A persona can shape output but does not turn the model into the real person or expert being simulated.

Information that can identify or reasonably be linked to a specific person, such as a name combined with contact, financial, government, health, or account information.

The part of an agentic system responsible for breaking a goal into steps or deciding what should happen next.

An installable extension that adds capabilities, integrations, workflows, or tools to an AI product. What a plugin contains and how it works varies by platform.

Of all the items a system marked positive, the percentage that really were positive. High precision means relatively few false positives.

The large-scale training stage where a model learns general patterns from broad datasets before being adapted for specific assistant behavior or tasks.

A model whose weights, training process, or implementation are controlled by an organization and not fully available to the public.

Based on probabilities rather than fixed rules. Generative AI chooses likely outputs from many possibilities, so the same prompt can produce different answers and a fluent answer is not guaranteed to be true.

The live environment where a system is used for real work or by real users, as opposed to a test or development environment.

A product feature or organizational structure that groups related chats, files, instructions, tools, or work around a shared goal. The exact meaning varies by AI platform.

The information or instruction given to an AI to guide what it should do. A prompt may be a short question or a complex package of instructions, examples, files, and constraints.

Designing and refining instructions, examples, constraints, and output requirements to help an AI produce better results. It is one part of the broader practice of context engineering.

An attack or failure mode where instructions inside user input or external content try to override the system’s intended rules or manipulate the AI into unsafe or unauthorized actions.

A reusable prompt structure with placeholders for changing information, such as a customer name, document, task, or desired format.

A proposed set of code changes submitted for review before being merged into a shared codebase. AI coding agents can create or review pull requests.

Reducing the numerical precision used to store or run a model so it requires less memory and computation. Quantization can make models cheaper or faster, sometimes with a small loss in quality.

A request for information sent to a search system, database, or retrieval system. In RAG, the user’s request is often transformed into one or more retrieval queries.

A technique where an AI retrieves relevant information from external sources at runtime and uses that information as context when generating an answer. RAG changes what the model can see for the request; it does not retrain the model.

A limit on how many requests, tokens, tool calls, or other units can be used during a particular period. Rate limits protect service capacity and often vary by plan.

A permission system that assigns access according to a person’s role in an organization. Enterprise AI products use RBAC to control who can use particular models, tools, data sources, or administrative features.

A setting on some AI systems that controls how much computational effort the model is allowed to spend before answering. More effort can help on difficult tasks but usually increases time and cost.

A model optimized to spend additional computation on planning, problem-solving, or verification before producing its final answer. Reasoning models are useful for complex tasks but are still fallible.

Of all the truly positive cases, the percentage the system successfully found. High recall means relatively few false negatives.

Deliberately trying to make an AI system fail, behave unsafely, expose weaknesses, or produce unacceptable results so those problems can be found before they matter.

An image supplied to a generative model to guide composition, subject, identity, style, color, or other visual characteristics.

WorkingLab’s supporting collection of concise explanations, glossary entries, platform guides, and foundational material that members can consult when a Lab introduces an unfamiliar concept.

A test that checks whether a change broke something that previously worked. AI workflows need regression tests because model, prompt, or software updates can improve one behavior while harming another.

A learning approach in which a system improves behavior using rewards or feedback tied to outcomes or actions.

A version-controlled project folder, usually containing code, files, configuration, and history. AI coding agents often work directly inside repositories.

A unit of communication sent to an AI service or API. A request can include input, instructions, settings, tools, and other context.

A second pass that reorders retrieved results so the most useful items are placed first before being given to the AI.

The result returned by an AI service after a request. A response may contain generated content, structured data, citations, reasoning summaries, or requests to use tools.

The practice of designing, deploying, and governing AI with attention to safety, fairness, privacy, security, transparency, accountability, and real-world consequences.

Finding relevant information from an external source so it can be used for a task. Retrieval may use keyword search, semantic search, databases, knowledge graphs, or other methods.

A family of techniques where feedback generated or assisted by AI is used in reinforcement-learning-style training, often alongside human-designed principles or evaluation.

A family of techniques that use human preferences or ratings to help train AI behavior toward outputs people judge more useful, safe, or appropriate.

An AI task configured to run later or repeatedly according to a schedule or trigger. The exact feature name varies by product.

An isolated environment where code or agent actions can run with limited access to the rest of the system. Sandboxes reduce risk but do not make every action automatically safe.

A formal description of how data is structured: which fields exist, their types, and sometimes which values are allowed. Schemas help AI systems produce or consume reliable structured data.

A package of code and tools that makes it easier for developers to use a platform or API without building every connection from scratch.

A starting value used by some generative systems to make randomness more repeatable. Using the same seed may produce similar results, but it does not guarantee identical output across every model or system.

The Transformer mechanism that lets each token weigh how relevant other tokens in the same sequence are to understanding or generating the current representation.

Training where the data provides its own learning signal—for example, hiding or predicting parts of text rather than requiring humans to label every example.

Search based on meaning rather than exact word matching. It usually relies on embeddings to find content that is conceptually related to the query.

Information that could cause harm, violate privacy, or create legal or security risk if exposed. Examples can include health, financial, identity, authentication, confidential business, or regulated information.

A computer or software service that provides data or capabilities to other software. In MCP, an MCP server exposes tools or resources to AI applications.

A cloud-computing model where developers run code or services without managing the underlying servers directly. The servers still exist; the provider handles most of their operation.

Information kept temporarily for the current interaction or task, typically through context, working state, or application-managed storage.

A reusable package of instructions, reference material, workflows, and sometimes scripts that helps an AI perform a task consistently. Implementations vary by platform, but the core idea is reusable procedural knowledge.

A language model with substantially fewer parameters or lower computational requirements than the largest LLMs. SLMs can be faster, cheaper, easier to run locally, and strong on focused tasks.

OpenAI’s generative video product and model family. Product capabilities and access can change over time.

The authoritative location a workflow should treat as the current, trusted version of information—for example, a database, project repository, policy document, or financial model.

A model designed so only part of its capacity is activated for a given input. Mixture-of-Experts models are a common form of sparse architecture.

AI that converts spoken audio into written text. Transcription is a common speech-to-text task.

A language used to query and modify relational databases. AI systems can often generate SQL from natural-language questions, but the queries should be reviewed before running against important data.

An enterprise authentication system that lets users sign into multiple services through a central identity provider.

Another word for probabilistic or involving randomness. Generative AI output is often stochastic.

A specified sequence of characters or tokens that tells a model or API to stop generating when that sequence appears.

Output constrained to a known structure—such as fields in JSON—rather than free-form prose. Structured outputs make AI results easier for software to validate and use.

Using AI to transform the visual or expressive style of content while retaining some of its underlying subject or structure.

A specialized agent delegated a narrower task by a primary agent or orchestrator. A subagent may have its own context, tools, instructions, and permissions.

Training a model on examples where humans or another process provide the desired answer or label.

Data generated artificially rather than collected directly from real-world events or people. AI can create synthetic data for training or testing, but generated data can also reproduce errors or biases from the model that created it.

High-priority instructions supplied by the application or platform that shape how an AI should behave. Users may not always be able to see or change these instructions.

A sampling setting that can influence how varied or predictable generated output is. Higher temperature generally increases variation; lower temperature generally makes choices more conservative. Not every modern model exposes or uses temperature in the same way.

A text-based interface for interacting with a computer through commands. AI coding agents can operate in a terminal and can often explain or execute commands on the user’s behalf.

Data held aside to measure how well a trained model or system performs on examples it did not learn from directly.

A collection of automated tests used to check whether software behaves correctly. AI coding agents often run test suites after making changes.

Generating an image from a text description.

AI that converts written text into spoken audio.

Generating or transforming video based on text instructions, sometimes with images or video supplied as additional context.

How much work a system can process in a given amount of time—for example, requests per second or tokens per second.

A small unit of data that a language model processes. In text, a token may be a whole word, part of a word, punctuation, or another text fragment. Tokens are the practical unit behind context size, usage, and much API pricing.

The amount of context or generation capacity available for a task, usually expressed in tokens. Good workflows spend that limited space on information that actually helps the model.

The process of breaking text into tokens before the model processes it.

A capability an AI can invoke to do something beyond generating text from its model alone—for example, search the web, query a database, run code, send email, calculate, or modify a file.

A structured request from an AI model to use a tool with particular arguments.

The information returned after a tool has been used. The result is often added back into context so the model can decide what to do next.

A sampling control that restricts generation to a set of likely next tokens whose combined probability reaches a chosen threshold. It is another way to control variation in generation.

The process of adjusting a model’s parameters by exposing it to data and optimizing it to perform better on a learning objective.

The information used to train a model. Training data influences what patterns the model learns, but the model does not behave like a searchable copy of every training document.

The subset of a dataset used directly to train a model.

Turning spoken audio into written text, usually with speech-to-text AI.

The neural-network architecture behind most modern LLMs. Transformers use attention mechanisms to process relationships among tokens efficiently and at scale.

Removing part of an input or context because it exceeds a system limit. Truncation can cause important information to disappear, so long-running workflows need deliberate context management.

When a model or system is too simple or insufficiently trained to capture the important patterns in the task, leading to poor performance even on familiar data.

Learning patterns from data without explicit human-provided labels for the desired answer.

Increasing the resolution of an image or video, often using AI to reconstruct or generate additional detail.

A specific real-world job or outcome someone wants a system to accomplish. WorkingLab starts with practical use cases rather than teaching tools in isolation.

The instruction or information supplied by the user in a conversation or request.

Data used during model or system development to tune choices and compare approaches without using the final test set.

A list of numbers used to represent data mathematically. Embeddings are vectors designed so items with similar meaning tend to have similar numerical representations.

A database optimized for storing and searching vectors such as embeddings. It is commonly used in semantic search and RAG systems.

A system for recording changes to files so teams can compare versions, collaborate, and restore earlier states. Git is the most widely used version-control system for software.

An AI model designed to analyze or generate visual information.

A model designed to work with both visual information and language, such as reading a chart, describing an image, interpreting a screenshot, or answering questions about a photograph.

A way for one system to automatically notify another system when an event happens, usually by sending an HTTP request. Webhooks can trigger automations or agent workflows.

The learned numerical values inside a neural network that determine how strongly different signals affect its output. Model training changes these weights.

A sequence of steps used to accomplish a task. A workflow can be fully manual, fixed automation, AI-assisted, or agentic.

A shared or organized environment containing users, permissions, data, projects, tools, or settings. The exact meaning differs across AI products.

A model that represents aspects of an environment and how that environment changes, allowing the system to predict or simulate possible future states or consequences. World models may be learned from text, images, video, sensor data, actions, simulations, or combinations of these.

A human-readable text format often used for software configuration. AI tools may create or edit YAML files when configuring applications, deployments, automations, or workflows.

A provider configuration or contractual arrangement in which qualifying customer content is not retained after processing, subject to the provider’s exact terms and exceptions.

Asking an AI to perform a task without first giving it an example of the desired input and output.