Artificial Intelligence Glossary
More than 300 AI terms explained in plain English: models, agents, AI-assisted coding, infrastructure and business. Each term has a short definition, a practical example and related reading.
325 terms
A
- A2A (Agent2Agent)AgentsOpen standard that lets independent AI agents discover capabilities, exchange messages and track tasks across different systems.
- AI AgencyBusinessA company that creates and delivers artificial intelligence solutions for clients. From chatbots and automations to code agents and custom multi-agent systems.
- AI AgentAgentsAn AI agent is a system that uses a model to interpret a goal, choose actions and interact with tools or data. It observes the results and can repeat steps until the task is complete or a defined limit is reached.
- AI AlignmentSafety and ethicsAI alignment is the effort to ensure an artificial intelligence system's behavior matches the intentions, goals and boundaries defined by people. Alignment must be assessed in the context of use and does not eliminate all risks.
- AI AssistantAgentsAn AI system that takes requests and helps a person get information or complete tasks. It can just chat or also use tools, depending on its permissions.
- AI AutomationAgentsUsing artificial intelligence to perform repetitive or complex tasks without human intervention.
- AI AvatarBusinessIt is a digital representation created or animated with artificial intelligence, often used to speak, present content or interact. It can be fictional or represent a real person with proper authorization.
- AI BusinessBusinessA business with AI uses artificial intelligence in products, services or operations to solve a need for customers or for the company itself. Adoption only creates value when it improves a relevant outcome and can be maintained safely.
- AI ConsultingBusinessAI consulting helps companies identify, plan and implement artificial intelligence solutions, from strategy to operations.
- AI Content CreationBusinessIt is the use of artificial intelligence tools to support research, writing, editing or media production, with human review before publication.
- AI CopywritingBusinessIt is the use of artificial intelligence tools to create or review persuasive texts, such as headlines and product pages. The responsible person must check facts, tone and promises before publishing.
- AI Data CenterInfrastructureAn AI data center is a facility that combines servers, accelerators, storage, networking, power and cooling to run artificial intelligence workloads.
- AI EthicsSafety and ethicsAI ethics is the analysis of the values, risks and responsibilities involved in developing and using artificial intelligence systems. It guides decisions about fairness, privacy, transparency, security and impact on people.
- AI FreelancerBusinessAn independent professional who offers development, integration or consulting services in artificial intelligence for different clients. Works on one-off projects such as fine-tuning, agent creation or automations.
- AI GatewayInfrastructureAn AI gateway is a layer that sits between applications and models or providers, centralizing routing, authentication, rate limits and observability.
- AI GovernanceSafety and ethicsIt is the set of responsibilities, policies, processes and controls that guide how an organization selects, develops, uses and monitors artificial intelligence systems.
- AI ImagesBusinessThey are images created or modified using artificial intelligence models, based on prompts, reference images or other inputs supported by the tool.
- AI Legal FrameworkSafety and ethicsA legislative proposal that seeks to set rules for the development and use of artificial intelligence in Brazil. The bill under consideration should not be confused with an already in force law.
- AI MisuseSafety and ethicsUsing artificial intelligence systems in a deceptive, harmful, unauthorized, or incompatible way with their intended purpose. It can be intentional or result from insufficient controls and review.
- AI MonetizationBusinessAI monetization is how to generate revenue or economic value from products, services or processes that use artificial intelligence, aligning the offering with customer benefit and operating costs.
- AI OverviewsBusinessA Google Search feature that shows an AI-generated summary for some queries, pulling information from web pages and displaying links to related sources.
- AI ProductBusinessAn AI product is a service or tool that uses artificial intelligence to solve a user need. The model is only one component: data, interface, controls, evaluation and support also define the experience and reliability.
- AI ProviderInfrastructureAn AI provider is a company or organization that offers access to artificial intelligence models, services or infrastructure via API or other integration methods.
- AI ROIBusinessAI ROI measures whether the benefits of an AI initiative justify the resources invested, considering costs, observable results and a defined comparison period.
- AI Regulation in BrazilSafety and ethicsAI regulation in Brazil brings together laws, sectoral rules, policies and proposals that guide how artificial intelligence systems can be developed and used in the country.
- AI SafetySafety and ethicsAI safety brings together practices to prevent and reduce harm caused by artificial intelligence systems, evaluating their behaviors, limits, permissions and impacts during development and use.
- AI SearchBusinessSearch systems that use AI to understand queries, retrieve relevant information or synthesize answers, and can show links to the sources consulted.
- AI StartupBusinessAn AI startup is a young company that uses artificial intelligence as a core part of its product, service or operations. The label does not replace validation of a real problem, customers and a sustainable business model.
- AI SubscriptionBusinessAn AI subscription is a recurring access model for a tool or service that offers artificial intelligence features. Features, limits and terms vary by provider.
- AI SupercomputerInfrastructureAn AI supercomputer is an integrated high-performance infrastructure used for training, inference, or intensive processing. Its performance depends on accelerators, memory, networking, storage, and software working together.
- AI Use PolicySafety and ethicsIt is the set of rules that guides how an organization can use AI tools, including data limits, responsibilities, output review, and approval of new uses.
- AI VideosBusinessAudiovisual content created or altered with artificial intelligence systems from text, images, videos or other reference materials.
- AI WatermarkSafety and ethicsAn AI watermark is a signal embedded in or associated with content to help indicate its origin or identify involvement of generative systems. On its own, it does not prove that the content is true.
- AI WorkflowAgentsA sequence of automated steps that an artificial intelligence executes to complete a complex task.
- AI code reviewAI-assisted codingAI code review is the analysis of software changes by models or tools that suggest possible defects, risks and improvements. Findings should be validated by tests and human review.
- AI in AgribusinessBusinessAI in agribusiness uses artificial intelligence to analyze data and support farming, livestock and operational activities, with results that need to be validated in the local context.
- AI in EducationBusinessAI in education is the use of artificial intelligence systems to support teaching, learning, research or administrative tasks, with pedagogical oversight and data protection.
- AI in HRBusinessUsing artificial intelligence systems to support HR tasks, like answering internal questions, organizing information and preparing drafts. Decisions about people require clear criteria and human oversight.
- AI in HealthcareBusinessAI in healthcare is the use of computer systems to support clinical, administrative, research or public health tasks, with validation, data protection and human oversight.
- AI in LawBusinessUse of artificial intelligence systems to support law practice, legal departments and justice institutions, such as document research and draft preparation. Outputs require professional validation and data protection.
- AI in MarketingBusinessArtificial intelligence systems support marketing activities such as research, creation, personalization, results analysis and task automation. AI helps produce and interpret information, but it does not replace strategy or human review.
- AI in RetailBusinessAI in retail is the use of artificial intelligence to support store and e-commerce tasks, such as customer service, sales analysis, recommendations and inventory organization.
- AI in SalesBusinessUse of artificial intelligence tools and models to support commercial tasks, such as organizing contacts, preparing conversations, summarizing information and analyzing opportunities. The team remains responsible for validating data and decisions.
- AI in customer supportBusinessAI in customer support uses intelligent systems to respond to, organize or route requests. It can support agents or automate steps with defined limits and supervision.
- AI in the Financial SectorBusinessAI in the financial sector is the use of intelligent models and systems to analyze data and support tasks in banks, payments, insurance and other financial activities, under appropriate controls.
- AI-Assisted ProgrammingAI-assisted codingAI-assisted programming is the use of artificial intelligence models and tools to support development activities, such as completing, explaining, generating, testing or reviewing code. The developer remains responsible for validating changes.
- AI-Generated TestsAI-assisted codingAI-generated tests are test cases and checks suggested by AI tools based on code, requirements or examples, and they should be reviewed before use.
- AI-Generated Text DetectionSafety and ethicsAI-generated text detection is an attempt to estimate whether content was produced by a generative model. The result is probabilistic and does not prove authorship of the text.
- AI-Powered ServicesBusinessAI-powered service delivery is a business model where professionals and agencies use artificial intelligence to deliver client work faster and more scalably.
- AIO (AI Overviews optimization)BusinessAcronym used for SEO practices aimed at visibility in Google AI Overviews responses. It is not an official technique and does not guarantee that a page will appear in the feature.
- API CreditsInfrastructureThey are usage units or balance that a provider associates with API consumption. How usage is counted varies between services, so a credit does not necessarily mean a fixed number of tokens or calls.
- AWSInfrastructureAmazon Web Services, or AWS, is Amazon's cloud computing platform, with services to run applications, store and analyze data, connect networks and build AI solutions.
- ActivationFundamentalsActivation is the transformation of a signal inside a neural network, usually performed by a mathematical function. It influences how information flows between the model's layers.
- Activation functionFundamentalsA mathematical mechanism that decides whether an artificial neuron should fire, introducing non-linearity into neural networks.
- Adversarial AttackSafety and ethicsAn attempt to manipulate an AI system to change its behavior, compromise security or privacy, or exploit a vulnerability. It can target data, the model, training, or integrations.
- Agent PlanningAgentsAgent planning is the organization of steps, tools and completion criteria so an AI agent can try to accomplish a complex task. The plan can be updated as new information emerges.
- Agent SquadAgentsAn agent squad is a coordinated group of AI agents that splits work and combines outputs to reach a shared goal, with defined roles and boundaries.
- Agent handoffAgentsAgent handoff is the transfer of a task or execution control to another agent, along with the necessary context to continue the work.
- Agent memoryAgentsA feature that records and retrieves relevant information so an AI agent can maintain continuity across steps or sessions. It can use context, files or external databases and requires clear rules for access, updates and retention.
- Agent orchestrationAgentsAgent orchestration is the coordination of AI agents' work through tasks, rules, context and tools. It defines who handles each step, how results flow through the system, and when the workflow ends or requests human review.
- Agentic AIAgentsAgentic AI describes AI systems that execute multi-step tasks, using context and tools to move toward a goal. The level of autonomy ranges from predetermined flows to more dynamic model decisions.
- Agentic LoopAgentsAn execution cycle where an agent queries the model, uses tools, and evaluates results until the task is complete, approval is requested, or a defined limit is reached.
- Agentic RAGAgentsIt's an architecture that combines RAG with an agent that can plan queries, choose search tools, and repeat retrieval to gather context before answering.
- AiderAI-assisted codingAider is an AI-assisted programming tool that runs in the terminal and helps you understand, create and edit files in a code repository using language models.
- AlgorithmFundamentalsAn algorithm is an organized set of instructions for solving a problem or performing a task. It takes in data, applies rules, and produces a result.
- Algorithmic JusticeSafety and ethicsAlgorithmic justice is the assessment of how automated systems distribute decisions, opportunities and errors across people and groups, with a focus on reducing harm and enabling review.
- Algorithmic biasSafety and ethicsAlgorithmic bias is a systematic tendency for an algorithm to produce distorted or unequal outcomes for people or groups. It can arise from the data, the system's objective, or the context of use.
- AmpAI-assisted codingProgramming agent and development environment that helps you execute coding tasks through web and terminal interfaces, in local or remote sessions.
- AnonymizationSafety and ethicsIt is the use of reasonable and available technical means at the time of processing so that a piece of data loses any possibility of direct or indirect association with a person. The assessment depends on the context and the risk of re-identification.
- AntigravityAI-assisted codingA collection of agent development interfaces documented by Google, including App, CLI, IDE integrations and SDK. The interface you choose defines the workflow.
- Antigravity CLIAI-assisted codingAntigravity's command line interface, documented by Google with the `agy` command. It is not the IDE interface nor the Gemini CLI.
- Artificial IntelligenceFundamentalsArtificial intelligence is the field that builds systems able to perform tasks linked to capabilities such as perceiving, learning, predicting, generating content or supporting decisions. The term covers distinct methods and does not imply human consciousness.
- Artificial neuronFundamentalsAn artificial neuron is a mathematical unit in a neural network that combines inputs with weights, applies an activation function, and produces an output. Its parameters can be adjusted during training.
- Attention mechanismFundamentalsIt is a neural network technique that calculates which parts of the input are most relevant for producing a representation or output. In language models, it helps relate different tokens within the context.
- Audio modelModels and LLMsAn audio model is an AI system that receives, generates or transforms sounds. It can transcribe speech, produce voice, classify sounds or perform other tasks, depending on its capabilities.
- AuditSafety and ethicsStructured assessment of a system's controls, processes and evidence against defined criteria, which can include security, data, models and AI operations.
- AuthenticationSafety and ethicsThe process of verifying the identity of a user or system before granting access.
- AuthorizationSafety and ethicsA mechanism that determines which resources and actions an authenticated user, service, or agent can access. It is not the same as authentication, which verifies identity.
- Autonomous AgentAgentsAn autonomous agent is an AI system that can carry out steps of a task without human instruction for every action, within defined objectives, tools and limits.
B
- BPE (Byte Pair Encoding)Models and LLMsBPE (Byte Pair Encoding) is a tokenization method that merges frequent parts of text into subwords to build the vocabulary used by a model.
- BackpropagationFundamentalsA procedure that computes gradients in a neural network by propagating backwards the effect of parameters on the loss function. These gradients can be used by an optimizer to adjust the model.
- Base modelModels and LLMsA base model is a version of a model trained to learn general patterns before a specific adaptation, such as fine tuning or instruction following training. The exact meaning can vary depending on the project's documentation.
- BatchFundamentalsBatch is a set of items processed as a single unit, for example a group of data used to train or run inference on an AI model.
- Batch APIInfrastructureIt is an interface that receives batches of requests for asynchronous processing, letting you track the batch and retrieve the responses later.
- Beam searchModels and LLMsIt is a decoding strategy that keeps several promising continuations of a sequence and compares them before choosing the final output. It explores more options than greedy selection, but can require more computation.
- Bias and VarianceFundamentalsBias and variance are two sources of error in machine learning models. Bias reflects excessive simplifications, while variance is an exaggerated sensitivity to the data used in training.
- Black BoxSafety and ethicsBlack box describes an AI system whose workings or path to a response are hard to understand. This opacity can limit auditing and trust, but approximate explanations do not guarantee that the result is correct.
- Bolt.newAI-assisted codingBolt.new is an AI powered site and app builder from StackBlitz. You describe what you want in natural language and can iterate on the generated project.
- BranchAI-assisted codingAn independent line of development in a Git repository. It lets you work on features, fixes or experiments without affecting the main codebase until merge.
- Browser AgentAgentsA browser agent is an AI system that interacts with web pages to perform tasks, such as searching for information or filling in fields. Actions must respect permissions and require confirmation when they have external impact.
C
- CLI (command line interface)AI-assisted codingIt is a program controlled by commands typed in a terminal. A CLI lets you run tasks, check options and automate operations without relying on a graphical interface.
- CUDAInfrastructureCUDA is NVIDIA's parallel computing platform and programming model that lets you run workloads on compatible GPUs. Developers use it to speed up computations, such as matrix operations in AI applications.
- CerebrasInfrastructureCompany that builds wafer-scale AI chips, the world's largest processors.
- Chat modelModels and LLMsA chat model is an AI model prepared or configured to interact via messages in turns. It uses the provided context to respond and continue a conversation, but it does not guarantee correct answers.
- ChatGPT PlusBusinessChatGPT Plus is a subscription plan for using ChatGPT, with access conditions that can differ from other plans. It is not a model and by itself does not include use of the OpenAI API.
- ChatGPT ProBusinessChatGPT Pro is an individual OpenAI subscription that expands access to advanced features and models in ChatGPT. It is aimed at people who use the service frequently for complex tasks, but it does not include API usage, which is billed separately.
- ChatbotAgentsSoftware that talks with people via text or voice to answer questions, guide workflows or trigger functions, using rules, AI models or both.
- ChunkingAgentsA technique for breaking texts or documents into smaller parts to make indexing, search and use as context in AI systems easier.
- Civil Liability for AISafety and ethicsThe duty to repair damages caused by artificial intelligence systems, whether due to design error, operational failure or unpredictable model output. It defines who pays when AI causes harm to third parties.
- ClassificationFundamentalsClassification is a machine learning task that assigns data to categories, such as sorting messages by topic or identifying whether a transaction looks suspicious.
- Claude CodeAI-assisted codingAnthropic's programming tool that uses Claude to understand repositories and support development tasks, such as analyzing or editing code based on configured permissions.
- Claude MaxBusinessAnthropic's individual plan that offers higher Claude usage limits than Pro. Offerings and specific limits can change, so check the current official documentation.
- Claude ProBusinessClaude Pro is a paid individual plan to use the Claude assistant with expanded access and additional features. Use of the Anthropic API is separate and requires its own setup.
- ClineAI-assisted codingCline is an open source coding agent available as an editor extension, terminal tool, and desktop app. It helps you work with code using tools and configurable approval controls.
- Cloud ComputingInfrastructureOn-demand access model for processing, storage, networks and applications over the internet, using infrastructure operated by a provider or shared among customers.
- ClusteringFundamentalsA machine learning technique that organizes examples into groups based on similarities in the data, without requiring predefined categories.
- Code AutocompleteAI-assisted codingCode autocomplete is the feature that suggests how to continue or complete code as you program. It can use simple rules or AI models to consider the editor and project context.
- Code ModelModels and LLMsAn AI model focused on programming tasks, such as generating, explaining or modifying code. Its responses need to be reviewed and tested.
- Code agentAgentsA code agent is an AI system that can analyze a project and carry out development steps, such as editing files and running tests, using authorized tools. Its changes need to be reviewed by a person.
- Code generationAI-assisted codingCode generation is creating instructions in a programming language from requirements, examples or rules. AI tools can suggest code, which still needs to be reviewed and tested.
- Code reviewAI-assisted codingCode review is the review of code changes before they are integrated into a project. It evaluates requirements, risks, tests and maintainability, complementing automated checks without guaranteeing the absence of bugs.
- Codex (OpenAI)AI-assisted codingCodex is OpenAI's programming agent that helps with software tasks, such as understanding a codebase, proposing changes and reviewing work. It can be used across different interfaces, depending on OpenAI's current offering.
- CommitAI-assisted codingA commit is a permanent record of changes in a Git repository. It works like a snapshot of the code at a specific moment, with a message that explains what changed.
- ComplianceSafety and ethicsIt refers to meeting applicable requirements, such as laws, contracts, standards and internal policies, for an organization, product or process.
- Computer useAgentsA capability that lets an AI system interact with graphical interfaces through tools, observing the screen and requesting actions like clicking or typing.
- ConsentSafety and ethicsIt is an expression of agreement given by a person for a specific purpose. In the processing of personal data, it must be understood in the context of the applicable law, the information provided, and the options available to the data subject.
- Content Marketing with AIBusinessAI content marketing uses intelligent tools to support research, planning, writing, adaptation and analysis of materials, while keeping human review and verification of information.
- Context WindowModels and LLMsThe context window is the limit of information a model can consider in one step, including instructions, history and data sent in to produce a response.
- ContinueAI-assisted codingAI programming assistant and open source project, available as IDE extensions and a command line interface. It connects your development workflow to configured models and resources.
- Convolutional Neural Network (CNN)FundamentalsIt is a neural network architecture that uses convolutions to recognize local patterns in data, especially images. Layers combine simple features, such as edges, into more complex representations.
- CopilotAgentsAn AI feature that supports a person on a task with suggestions, content generation or guided actions. The degree of autonomy depends on the product and its permissions.
- Copilot CLIAI-assisted codingCopilot CLI is the terminal interface for GitHub Copilot, which lets you chat with a programming agent to explore projects, suggest changes and run authorized tasks.
- Copyright and AISafety and ethicsQuestions about rights related to works used in AI development, generated content and usage permissions. Analysis depends on the circumstances and applicable rules.
- CoreWeaveInfrastructureCoreWeave is a cloud infrastructure provider focused on artificial intelligence workloads, with compute, storage and networking services.
- Cost per tokenInfrastructureIt is a unit of measurement used to estimate how much an AI API charges for processed text. Billing can separate input and output tokens, and the rules vary by provider and the resource used.
- CursorAI-assisted codingCursor is a code editor with AI features and an agent that can help you understand repositories, plan changes, edit files and verify results, based on the available tools and permissions.
D
- Data ExfiltrationSafety and ethicsData exfiltration is the unauthorized transfer of information outside an environment or organization, through technical or human means.
- Data PoisoningSafety and ethicsData poisoning is the intentional or accidental contamination of the data used to train, fine-tune or evaluate a model, which can compromise its behavior or its evaluation.
- Data PrivacySafety and ethicsPractices to ensure that the collection, use, sharing and retention of information respect the purpose and expectations of the people the data is about.
- Data leakSafety and ethicsA data leak is the exposure of information to people, systems or services without authorization, due to error, misconfiguration or malicious action.
- DatasetFundamentalsA dataset is an organized collection of examples used for analysis, training or evaluation of machine learning systems. Its quality depends on the relevance, representativeness, source and consistency of the data.
- Debugging with AIAI-assisted codingDebugging with AI is the use of artificial intelligence tools to investigate software errors and suggest hypotheses or fixes. Proposals need to be verified with problem reproduction, tests and human review.
- DecoderFundamentalsA model component that transforms a representation into an output or generates a sequence from the available context.
- Deep learningFundamentalsDeep learning is an area of machine learning that uses neural networks with multiple layers to learn patterns from data. It is used in tasks like language, image and audio, but it requires proper evaluation for the problem.
- DeepfakeSafety and ethicsVisual, audio or multimodal media generated or altered artificially. The term describes the content, but does not by itself determine whether the intent is to deceive.
- Dense modelModels and LLMsAn AI model whose architecture processes each input using its parameters broadly, without relying on sparse routing to a subset of experts. The term describes the computation, not the model's quality.
- DevinAI-assisted codingAI agent developed by Cognition for autonomous programming, capable of planning and executing development tasks.
- Diffusion ModelModels and LLMsA generative model that learns to create data by gradually removing noise. It is used, among other tasks, to generate images from text or from an initial image.
- Digital transformationBusinessA planned change in how an organization works or delivers value, combining digital technology with a review of processes and capabilities.
- DisinformationSafety and ethicsFalse or misleading information deliberately spread to confuse, manipulate or gain an advantage. Intent is what sets it apart from misinformation shared without intent to deceive.
- Distributed TrainingInfrastructureDistributed training is the technique of splitting model training across multiple computers or GPUs, speeding up the process and enabling larger models.
E
- EU AI ActSafety and ethicsThe EU AI Act is the European Union regulation that sets rules for developing, making available and using artificial intelligence systems, with obligations that vary based on risk, purpose and the organization's role.
- Edge AIInfrastructureUsing artificial intelligence on devices or servers close to where data is produced. Processing at the edge can reduce reliance on cloud communication, depending on available resources and connectivity.
- EmbeddingFundamentalsA numerical representation of text, images or other data as a vector of numbers. It allows AI models to compare meaning and find similar items by mathematical proximity.
- Embedding modelModels and LLMsAn embedding model converts text, images, code or other data into numeric vectors that represent learned features, enabling items to be compared for tasks like semantic search and recommendation.
- EncoderFundamentalsA component that transforms an input, such as text or an image, into a representation that other stages of the model can process.
- EncryptionSafety and ethicsA technique that transforms readable data into encoded content that can be recovered by anyone with the right key. It protects data in transit or at rest, but it does not replace other security controls.
- EpochFundamentalsA full pass through the training set, where each example is processed once. An epoch groups the iterations needed to go through the data, but it does not by itself measure model quality.
- Existential Risks of AISafety and ethicsExistential risks from AI are scenarios in which artificial intelligence systems cause irreversible harm to humanity, such as loss of control over systems much smarter than us.
- Explainability (XAI)Safety and ethicsExplainability in AI refers to methods that make it clearer how a model produces predictions or recommendations, and what limits exist in that interpretation.
F
- FLOPSInfrastructureFLOPS measures how many floating point operations a processor can perform per second. It is the most used unit for comparing the compute power of GPUs, TPUs and AI supercomputers.
- Fine-tuningModels and LLMsThe process of continuing training on an existing AI model with a specific dataset to adapt it to a task, style or domain. It reduces the need for long prompts and improves response consistency.
- Firebase StudioAI-assisted codingGoogle's browser-based development environment for building apps with AI assistance.
- Fireworks AIInfrastructureAI platform that offers model inference, dedicated GPU deployments, and training or fine-tuning resources for applications.
- Foundation modelModels and LLMsA foundation model is an AI model trained to serve as a base for different tasks and applications. It can be used directly or adapted with instructions, data and additional components, depending on the goal.
- Free AI PlanInfrastructureIt is a way to access an artificial intelligence product without direct charges to the user, within limits set by the provider. Features, availability, data handling and restrictions vary by service and can change.
- Function callingAgentsFunction calling is a mechanism where a model signals that an application function should be called and supplies structured arguments. The integrated software validates the request and executes the authorized action.
G
- GANModels and LLMsGenerative architecture that trains two neural networks in competition: one generates samples and the other tries to distinguish synthetic samples from real ones.
- GDPRSafety and ethicsGDPR is the European Union's General Data Protection Regulation. It sets rules for processing personal data and protects individuals' rights, and can apply to certain organizations outside the EU based on their activities.
- GEO (Generative Engine Optimization)BusinessContent optimization to be cited and referenced by generative AI search engines.
- GPUInfrastructureA GPU is a processor designed to run many operations in parallel. Beyond graphics, it is used in artificial intelligence tasks that can be broken down into simultaneous calculations.
- GPU ClusterInfrastructureA GPU cluster connects multiple graphics accelerators across one or more servers to split compute workloads. In AI, it is used to scale training or inference capacity when a single GPU is not enough.
- Gemini CLIAI-assisted codingCommand line interface to interact with Gemini models. Since June 18, 2026, the individual experience was moved to Antigravity CLI, but certain business and API access remained available.
- GitAI-assisted codingDistributed version control system that tracks changes to files and helps you compare, organize, and collaborate on projects.
- Git worktreeAI-assisted codingA Git feature that lets you have multiple working copies of the same repository in different directories, each on a separate branch. Useful for running code agents in parallel without file conflicts.
- GitHub CopilotAI-assisted codingGitHub Copilot is an AI assistant for software development that suggests code, answers questions about projects, and can help review or execute programming tasks.
- Google AI ProBusinessGoogle One subscription plan that bundles access to select Google AI features and other benefits. The package contents vary by region and can change.
- Google AI UltraBusinessGoogle AI Ultra is a Google One subscription tier that provides expanded access to select Google AI features. Benefits, limits and availability vary by plan and region.
- Google CloudInfrastructureGoogle Cloud is Google's cloud computing platform, with infrastructure, storage, networking, data and AI services to build and run applications without managing all the physical infrastructure on premises.
- Gradient descentFundamentalsAn optimization method that adjusts a model's parameters by taking repeated steps in the direction that reduces a loss function. It is one of the core techniques used for training many machine learning models.
- GraphRAGAgentsIt's a RAG approach that uses graphs to represent entities and relationships between information, helping to retrieve connected context before a model generates a response.
- Greedy decodingModels and LLMsA generation strategy that picks the most likely token at each step, without exploring other complete sequences.
- GroqInfrastructureSemiconductor company that develops specialized chips for low-latency AI inference.
- GroundingModels and LLMsIt is the practice of anchoring an AI response to information provided or retrieved from identifiable sources. The goal is to provide context and evidence for the generation, without guaranteeing the response is correct on its own.
- GuardrailsAgentsGuardrails are controls that limit inputs, responses, permissions and actions of AI systems to reduce risk. They can combine validation, technical isolation and human approval.
H
- HallucinationModels and LLMsIt is a response generated by an AI system that sounds plausible but contains incorrect, invented, or unsupported information. Convincing text alone does not prove the content is true.
- Hidden layerFundamentalsAn intermediate layer in a neural network that transforms the incoming data before passing it to the output layer, helping the model learn useful representations.
- Hugging FaceInfrastructureThe Hugging Face platform for exploring, hosting and sharing repositories of models, datasets and machine learning Spaces.
- Human in the loopAgentsHuman in the loop is an approach where a person takes part in key steps of an automated system, such as reviewing outputs, approving actions or correcting AI decisions.
- Hybrid searchAgentsThe combination of keyword search (lexical) with semantic similarity search (vector) in a single ranked result. It increases accuracy in RAG systems and knowledge base search.
- HyperparameterFundamentalsA hyperparameter is a setting that controls the training, architecture, or generation of a model, such as learning rate, batch size, or number of epochs.
I
- IDEAI-assisted codingAn integrated development environment brings together a code editor and programming tools, such as a terminal, debugger and project management, in a single application.
- Image modelModels and LLMsAn AI system that processes visual content. Depending on its capabilities, it can generate, edit or analyze images from text and other images.
- Impact of AI on EmploymentSafety and ethicsThe set of changes in the labor market driven by automation and artificial intelligence capabilities. It includes both task replacement and the creation of new roles and the transformation of existing professions.
- Inference serverInfrastructureSoftware that makes a trained model available to receive requests and run inferences, usually via an API. It can coordinate model loading, requests and response delivery.
- Input and Output TokensInfrastructureInput tokens are the units sent to the model, output tokens are the units it generates. The count depends on the tokenizer and the content.
- Instruct ModelModels and LLMsAn instruct model is a language model fine-tuned to respond to natural language instructions. This tuning helps with tasks like summarizing, explaining, extracting data, or generating code.
J
K
- Kilo CodeAI-assisted codingAI programming agent that assists with repository tasks and can be used in environments like VS Code, JetBrains and terminal. It offers work modes to understand, plan, edit and debug code.
- KiroAI-assisted codingAgentic development environment from AWS that offers IDE, CLI and Web interfaces to collaborate with AI agents on software tasks, with shared settings and capabilities.
- Knowledge baseAgentsA structured repository of information that an AI system consults to answer questions or make decisions. It acts as an external source of truth for the model, reducing hallucinations and keeping responses up to date.
L
- LGPDSafety and ethicsBrazilian Law No. 13.709/2018 that establishes rules for the processing of personal data and the rights of data subjects within its scope of application.
- LLM (large language model)Models and LLMsAn LLM is an artificial intelligence model trained to process and generate language, predicting token sequences from the received context. Some also work with images, audio or other types of data.
- LM StudioInfrastructureAn app that helps you download and run language models on your local machine, chat with them and expose APIs for other programs.
- LSTMFundamentalsIt is a recurrent neural network architecture that uses states and internal gates to learn which information to preserve when processing sequences.
- Learning rateFundamentalsA hyperparameter that defines the size of the adjustments made to a model's parameters during training. An inappropriate rate can make learning slow or unstable.
- Local inferenceInfrastructureRunning an AI model on a computer, server or device you control, without relying on a remote service for every request. Hardware requirements and configuration influence performance and privacy.
- LogitsModels and LLMsThey are unnormalized numerical scores that a model calculates for possible tokens or classes before converting them into probabilities or choosing an output.
- Long ContextModels and LLMsLong context is using a large amount of information in an AI task. A larger window expands the available content, but does not guarantee that all details will be used correctly.
- Long-term memoryAgentsA mechanism in an AI application to store relevant information and retrieve it in future sessions or tasks. It requires storage, selection criteria, and controls for updating and deletion.
- Loss FunctionFundamentalsA loss function is a mathematical rule that measures the error between a model's predictions and the expected results. During training, this value guides adjustments to the parameters.
- LovableAI-assisted codingLovable is a platform that uses natural language instructions to help create, edit and publish applications, with development capabilities described by the company itself.
- llama.cppInfrastructurellama.cpp is an open source project for running language model inference on different hardware, including CPU and compatible GPUs. It also provides an HTTP server for applications.
M
- MCP (Model Context Protocol)AgentsMCP, or Model Context Protocol, is an open protocol that standardizes how AI applications connect to tools, data and other resources offered by compatible servers.
- MCP ClientAgentsA component that keeps communication between a host and an MCP server, letting the application access the capabilities advertised by the server.
- MCP ServerAgentsA program that provides data or functionality to an application that supports the Model Context Protocol, via an MCP client.
- Machine learningFundamentalsMachine learning is a field of artificial intelligence where models learn patterns from data to make predictions, classifications, or other tasks, instead of relying only on manually written rules.
- Make Money with AIBusinessMaking money with AI means using artificial intelligence systems to create or improve a product, service or process that customers are willing to pay for. The technology alone does not guarantee demand or revenue.
- Maximum output lengthModels and LLMsIt is the token ceiling a model can generate in a response. The limit controls the maximum output size, but does not guarantee the response will use all of that space.
- Mechanistic InterpretabilitySafety and ethicsMechanistic interpretability is research that seeks to understand how neural networks work internally, mapping the circuits and representations inside the model.
- Microsoft AzureInfrastructureMicrosoft Azure is Microsoft's cloud computing platform, with services to build, deploy and run applications, infrastructure, data and AI solutions.
- Mixture of Experts (MoE)Models and LLMsMixture of Experts (MoE) is a neural network architecture with multiple experts and a router that selects which parts participate in processing each input or token.
- ModalInfrastructureCloud computing platform that runs code and applications in containers, with scalable functions for AI inference, batch processing and other intensive workloads.
- Model APIInfrastructureA model API is an interface that connects applications to AI models, letting you send inputs and receive responses through structured requests.
- Multi-agent systemAgentsA multi-agent system coordinates two or more AI agents to solve a task. Each agent can have its own role, tools and context, while the system defines how they exchange results and finish the work.
- Multimodal ModelModels and LLMsAn AI multimodal model processes or combines different types of information, such as text, image, audio or video, according to the capabilities available in the system.
N
- NPUInfrastructureAn NPU is a processor or specialized hardware block designed to accelerate common artificial intelligence operations, such as those used in neural networks. Its architecture and features vary by manufacturer and device.
- NVIDIA B200 (Blackwell)InfrastructureThe NVIDIA B200 is a GPU based on the Blackwell architecture, designed for accelerated computing and AI workloads. It is a hardware component, not a complete server.
- NVIDIA H100InfrastructureThe NVIDIA H100 is a Tensor Core GPU for accelerated computing in data centers, built on the Hopper architecture. It handles AI workloads and high performance computing, with configurations that vary by product.
- NVIDIA NIMInfrastructureNVIDIA NIM is a family of microservices for deploying AI models for inference, packaged in containers and integrated into applications via standardized APIs.
- Neural networkFundamentalsA neural network is a computational model composed of connected units that transform inputs into outputs through adjustable parameters. During training, these parameters are adjusted to learn patterns from examples.
O
- OllamaInfrastructureA tool for downloading, running and serving language models on your own computer or remote infrastructure. It includes a terminal interface and an API that applications can use.
- On-Device AI (on-device)InfrastructureRunning artificial intelligence tasks directly on a user's phone, computer or other device. It can reduce reliance on remote services, but it depends on hardware, the model and the app's connections.
- Open Source ModelModels and LLMsAn open source model is an AI system made available with sufficient rights and information to study, use, modify and share its components, according to the applicable license.
- Open model (open weights)Models and LLMsOpen model is a term used for AI systems whose weights or other components are made available. The term alone does not guarantee unrestricted licensing or access to all training data and methods.
- OpenCodeAI-assisted codingA programming agent that helps you work in repositories using models and tools configured by the user, with an interface and features described in the official documentation.
- OpenRouterInfrastructureOpenRouter is an API platform that connects applications to models from different providers and offers routing options between them.
- Oracle CloudInfrastructureOracle Cloud is Oracle's cloud computing platform, with services to run applications, store and process data, configure networks and use databases.
- OrchestratorAgentsAn orchestrator coordinates steps, agents and tools in an AI workflow. It routes tasks, tracks results and decides when to continue, finish or request human intervention.
- OverfittingFundamentalsOverfitting happens when a model learns the quirks of the training data and struggles to generalize to new examples.
P
- Pair programming with AIAI-assisted codingPair programming with AI is a way of programming where a person works with an AI assistant or agent to plan, write, review or test code. The person remains responsible for validating changes.
- ParameterFundamentalsA parameter is a value used by a system. In machine learning, it is usually an internal value learned during training and applied by the model when generating predictions or responses.
- PerceptronFundamentalsA classification model that combines inputs with weights and applies a decision rule to separate examples into classes, also serving as an educational foundation for neural networks.
- Permission ScopeSafety and ethicsPermission scope defines which resources and actions a person, application or agent can access. A well defined scope grants only the access needed for the task.
- PerplexityModels and LLMsIt is a metric that estimates how surprised a language model is when predicting the next tokens in a text. In general, lower perplexity indicates better prediction on the evaluated set, but it does not guarantee more useful answers.
- PerplexityBusinessIt's an AI search and answers service that looks up information on the web and presents summaries with citations for the sources consulted. The references help you verify the answer, but they don't replace critical reading.
- Perplexity ProBusinessPerplexity Pro is a subscription plan for Perplexity's AI search service, with expanded access to research and tools. Features and limits can change based on the current offering.
- Post-trainingModels and LLMsSteps applied to a model after initial training to adapt its behavior to specific instructions, tasks, or response criteria.
- Pre-trainingModels and LLMsIt is an initial stage in which a model learns general patterns from large datasets before being adapted to specific tasks. The result is usually a base model that may still require adjustment or additional instructions.
- Price per million tokensInfrastructureIt is a reference unit used to express how much an API charges for one million processed tokens. The actual calculation depends on volume and the billing categories defined by the provider.
- Principle of Least PrivilegeSafety and ethicsA security practice that grants each user or system only the minimum permissions needed for its role.
- Process AutomationBusinessUsing software, rules and integrations to run repetitive work steps, route tasks and handle exceptions with less manual intervention.
- Productivity with AIBusinessUsing artificial intelligence to support task steps and workflows. Its usefulness depends on the goal, the quality of the output, and how it is reviewed.
- PromptAgentsA prompt is the text or instruction you send to an AI model to get a response. It is how humans communicate with generative AI systems.
- Prompt EngineeringAgentsPrompt engineering is the practice of writing and refining instructions for AI models to get more accurate, useful and consistent responses.
- Proprietary modelModels and LLMsAn AI model whose access, use or distribution is controlled by an organization under its own terms. The label does not determine its quality or what has been disclosed.
- Pull requestAI-assisted codingA pull request is a proposed change to a Git repository, submitted for review before being integrated into the main codebase. It is the primary collaboration mechanism in software projects.
Q
R
- RAG (Retrieval-Augmented Generation)AgentsAn architecture that retrieves passages from an external source and provides them to a language model to generate contextualized answers based on the content found.
- RLAIFModels and LLMsRLAIF is an alignment technique where an AI model evaluates and ranks responses instead of humans, making training more scalable.
- RLHFModels and LLMsRLHF is a fine-tuning technique that uses human feedback to align a language model with human preferences, making its responses more useful and safer.
- RPABusinessAutomation of repetitive tasks by software robots that interact with applications, often mimicking keyboard and mouse actions in the interface.
- ReActAgentsReAct is an approach that interleaves a language model's reasoning with actions in tools or environments. The observations from those actions help the system decide the next steps.
- ReLUFundamentalsReLU, or Rectified Linear Unit, is an activation function that maps negative inputs to zero and keeps positive values unchanged, and it is widely used in neural networks.
- Reasoning modelModels and LLMsA reasoning model is an AI system designed for tasks that require multiple steps, such as planning and problem solving. The term does not guarantee that every answer is correct.
- Recurrent Neural Network (RNN)FundamentalsIt is a neural network architecture that processes sequential data in order and keeps an internal state, so earlier information can influence how later elements are processed.
- Red teamingSafety and ethicsAuthorized assessment that simulates adversarial scenarios to identify vulnerabilities and unwanted behaviors in systems, models and AI applications. Findings guide fixes and security controls.
- Refactoring with AIAI-assisted codingAI refactoring is the use of artificial intelligence tools to reorganize code structure while preserving the software's expected behavior.
- RegressionFundamentalsRegression is a machine learning task that estimates numerical values from data, such as forecasting demand, price, or delivery time.
- Reinforcement LearningFundamentalsA machine learning method where an agent learns to choose actions by interacting with an environment and receiving reward signals for the consequences.
- Repetition PenaltyModels and LLMsA generation setting that reduces the chance of repeating tokens already used. Its effect varies by model, provider, and the configuration applied.
- ReplicateInfrastructurePlatform that lets you run machine learning models via API and integrate them into applications without building your own inference infrastructure.
- ReplitAI-assisted codingReplit is a web based software development platform that brings together programming tools and AI resources to help you build and run applications.
- RerankingAgentsIt is a step that reorders search results, evaluating which are most relevant to a specific query. In RAG systems, it helps prioritize documents before sending them to the generative model.
- Responsible AISafety and ethicsResponsible AI is the design and use of artificial intelligence systems with attention to security, privacy, transparency, impacts and human oversight. It is a continuous practice, not a badge that guarantees risk free operation.
- Risk assessmentSafety and ethicsProcess to identify potential harms associated with an AI system, analyze their severity and likelihood, and define controls, owners and monitoring throughout the lifecycle.
- Roo CodeAI-assisted codingAn AI programming agent that worked as a VS Code extension and organized tasks into specialized modes. The original project was shut down and archived on May 15, 2026.
S
- SEO with AIBusinessUsing artificial intelligence to optimize websites and content for search engines.
- SLM (small language model)Models and LLMsAn SLM is a relatively small language model compared to larger models. It can be tuned for specific tasks and run with fewer resources in some scenarios, but a smaller size does not guarantee better cost or quality.
- SamplingModels and LLMsA method that selects an option from a probability distribution. In generative models, it is used to choose the next tokens and can vary the responses.
- SandboxAgentsSandbox is an isolated and limited execution environment used to control which files, tools, resources and connections a program can access.
- Secrets in CodeSafety and ethicsSecrets in code are private credentials, such as API keys, tokens and passwords, included in project files. If they are exposed, someone else can use them to access services or data.
- Self-attentionFundamentalsIt is a type of attention where each element in a sequence calculates relationships with other elements from the same sequence. In Transformers, it helps build representations that incorporate the context around each token.
- Self-supervised learningFundamentalsSelf-supervised learning is an approach in which the model creates training signals from the structure of the data itself, reducing the need for manual labels to learn useful representations.
- Semantic SearchAgentsA search method that retrieves information related to the meaning of the query, even when documents use different words.
- Semi-supervised LearningFundamentalsSemi-supervised learning is a machine learning approach that trains models by combining labeled data with unlabeled data, making use of unlabelled examples when obtaining ground truth is costly or labor intensive.
- Sensitive dataSafety and ethicsInformation whose exposure or misuse can cause harm to a person or organization, such as credentials, financial records or sensitive personal data.
- Shadow AISafety and ethicsIt is the use of AI tools within an organization without approval, visibility or proper institutional governance, often to handle day to day tasks.
- Shared ContextAgentsInformation that multiple agents or components can consult during a workflow, such as objectives, requirements and intermediate results. The application defines how this context is stored, accessed and updated.
- Short-term memoryAgentsTemporary information that an AI system keeps available during a conversation, session or task, such as recent instructions and work state. It can use history, summaries or session storage.
- SigmoidFundamentalsActivation function that maps a real value to an output between 0 and 1. It is common in binary classification and neural networks.
- SoftmaxFundamentalsA function that turns a list of scores into normalized probabilities that sum to 1. It is used in classification and in selecting next tokens for language models.
- SubagentAgentsA subagent is an AI agent that receives a bounded subtask from another agent or a coordinating system. It executes the authorized work and returns the result for integration or review.
- SuperGrokBusinessSuperGrok is a subscription tier to access Grok features, xAI's AI assistant. Features, limits and availability vary by plan and may change.
- Supervised Fine-Tuning (SFT)Models and LLMsA training stage in which the model learns from human-provided input and correct response pairs. It is the most common technique for turning a base model into an instruction-following assistant.
- Supervised LearningFundamentalsSupervised learning is a machine learning method in which the model learns to predict outputs from examples that already have known labels. Label quality and data splitting affect evaluation.
T
- TPUInfrastructureTPU, or Tensor Processing Unit, is an application-specific integrated circuit developed by Google to accelerate machine learning workloads. Access and usage options depend on compatible services and infrastructure.
- TUIAI-assisted codingIt is a user interface presented inside a terminal, with interactive text elements like menus, panels and lists. Unlike a simple CLI, it can keep an interactive screen open.
- TemperatureModels and LLMsTemperature is a generation parameter that adjusts token selection and, in general, the variation in a model's responses. The effect depends on the implementation and the task.
- TerminalAI-assisted codingText interface used to run commands and interact directly with the operating system or development tools.
- Together AIInfrastructureAI infrastructure provider that offers access to models, hosted inference, and resources to run or adapt models in applications.
- TokenModels and LLMsThe unit a language model uses to represent and process text. A token can be a whole word, part of a word, punctuation, or another fragment, depending on the model's tokenizer.
- Token ProbabilityModels and LLMsIt is the chance the model calculates that a specific token will be chosen as the next element in the response, given the context available so far.
- TokenizationModels and LLMsThe process of splitting text into tokens and converting them into identifiers that a language model can process. Rules and counts vary depending on the tokenizer.
- TokenizerModels and LLMsA tokenizer converts text into tokens, the units a language model processes, and converts the generated tokens back into readable text.
- ToolAgentsA tool is an operation that an AI model or agent can call to query data or perform an action in another system, such as looking up an order or creating a record.
- Tool approvalAgentsTool approval is a control that requires authorization before an AI agent executes an action through a tool, such as changing files or calling an API.
- Tool calling (tool calls)AgentsA pattern where a model asks an application to run an external function or service and return the result so the model can continue the task.
- Top-kModels and LLMsText generation technique that restricts sampling to the k most likely words.
- Top-p (nucleus sampling)Models and LLMsA text generation technique that selects words until their cumulative probability reaches a threshold p.
- TraeAI-assisted codingIt is an AI-assisted programming tool. The current official page uses the name TraeCode and presents it as a programming engineer for developers and engineering teams.
- Training DataFundamentalsTraining data are examples used to adjust a machine learning model and help it learn patterns relevant to a task.
- TransformerFundamentalsIt's a neural network architecture that uses attention mechanisms to process relationships between elements in a sequence, such as text tokens.
- Transparency in AISafety and ethicsTransparency in AI is clear communication about a system's purpose, limits, data, risks and those responsible for it. It helps people assess how to use it without promising that every model decision will be explainable.
U
- UnderfittingFundamentalsA situation where a model learns too little from the data and performs poorly even on training examples, making reliable predictions on new cases difficult.
- Unified MemoryInfrastructureA memory architecture where the CPU and GPU share the same physical address space. It eliminates data copying between them and lets you run larger models on hardware with limited memory.
- Unsupervised learningFundamentalsMachine learning methods that find patterns or structures in data without labeled responses for each example.
- Usage LimitInfrastructureA rule that restricts the volume, frequency, or resources an account, application, or team can use in a service.
V
- VAEModels and LLMsA VAE, or variational autoencoder, is a generative neural network that learns distributions over latent representations to reconstruct data and sample new examples similar to those seen during training.
- VRAMInfrastructureMemory tied to the GPU that stores data needed for graphics processing and parallel workloads, including AI operations.
- VS Code ExtensionAI-assisted codingIt is a package that extends Visual Studio Code with features for languages, tools and development workflows, integrated into the editor interface.
- VectorFundamentalsAn ordered list of numbers that represents a point in a multidimensional space. In AI, vectors encode meaning: text, images or sounds become vectors so the model can compare similarity.
- Vector SpaceFundamentalsA vector space is a set of vectors with addition and scalar multiplication operations. It is the mathematical structure behind embeddings and semantic search in AI.
- Vector searchAgentsIt is a retrieval technique that compares numerical vectors for queries and content to find semantically similar items, even when there are no exact word matches.
- Vibe codingAI-assisted codingVibe coding is a way to create or modify software by talking to an AI in natural language. You describe the desired outcome, evaluate what was generated, and guide further changes, without skipping testing and code review.
- Video ModelModels and LLMsAn AI system that processes or generates sequences of images over time, able to create, transform, edit or analyze videos according to its capabilities.
- Virtual InfluencerBusinessA digital character that publishes content and interacts with an audience as a public figure, managed by a person or team.
- Vision Language Model (VLM)Models and LLMsA vision language model (VLM) is an AI that combines visual processing and language to interpret images and answer questions about them. Its answers can contain errors and should be checked on critical tasks.
- Voice CloningBusinessA technique for synthesizing a voice that reproduces characteristics of a reference voice, such as timbre, pronunciation and rhythm, from recordings.
- v0AI-assisted codingv0 is Vercel's AI development agent that helps turn natural language instructions into code, interfaces and applications you can iterate on.
- vLLMInfrastructurevLLM is an open source library for serving language models and handling inference requests. It provides tools to expose models through an HTTP server.
W
- WeightFundamentalsIn machine learning, a weight is a learned numeric parameter that controls the influence of an input on a model's calculations, such as in a neural network.
- WindsurfAI-assisted codingWindsurf was a development environment with an editor and AI assistance. The company now introduces Devin Desktop as the new name for the product and outlines features for tracking agents.
Z
- ZedAI-assisted codingZed is a code editor with collaboration features and integration with AI models and agents that can help you search, edit and test projects.
- Zero trustSafety and ethicsA security approach that evaluates every access request based on identity, context and permissions, without automatically trusting connections just because they are inside the network.