AI glossary

The proportion of predictions that are correct under a defined scoring rule.

in: AI evaluation › Choose measures that fit the task

A function that transforms intermediate values, often adding nonlinearity.

in: Deep learning › How a neural network learns

A small trainable component added to adapt an existing model.

in: Fine-tuning › Choose an adaptation method

In the A2A documentation checked 2026-10-09, metadata describing an agent’s identity, capabilities, skills and interaction requirements.

in: Agent interoperability › Four roles, four integration needs

As documented on 2026-10-09, a protocol connecting code editors or IDEs with coding agents.

in: Agent interoperability › Four roles, four integration needs

Software around a model that manages the agent loop, tool routing, approvals, recovery, and run state.

in: AI agents › Agents, workflows, and harnesses

Information retained for reuse across agent steps or sessions.

in: Agent memory and orchestration › Choose what to retain and coordinate

As specified on 2026-10-09, an open format packaging instructions and resources around a required SKILL.md file.

in: Agent interoperability › Four roles, four integration needs

In the A2A documentation checked 2026-10-09, an open standard for communication and interoperability between independent agent systems.

in: Agent interoperability › Four roles, four integration needs

A predefined process that contains model-driven steps or actions.

in: AI agents › Agents, workflows, and harnesses

A repository-file convention for instructions intended for coding agents.

in: AI coding agents › Give the agent repository context

Software in which an AI model directs some of its own processes and tool use to accomplish a task.

in: AI agents › Agents, workflows, and harnesses

Examining values and responsibilities in AI development and use.

in: AI safety and governance › Name the harms and responsibilities

Rules, processes, and accountability for developing and using AI.

in: AI safety and governance › Name the harms and responsibilities

The ability to understand, use, and critically assess AI.

in: AI in education and research › Use AI to support learning

A computational representation used to make predictions, generate outputs, or choose actions.

in: AI foundations › What AI means

Work to identify and reduce harms caused by AI systems.

in: AI safety and governance › Name the harms and responsibilities

A period of reduced confidence, funding or activity following unmet AI expectations.

in: History of AI › Milestones and tests

Efforts to make AI behavior match intended goals, values and constraints.

in: AI alignment › Goals, constraints and behavior

Identifying observations that differ from expected patterns.

in: Predictive machine learning › Match the method to the output

A defined software interface through which one component can interact with another component’s capabilities.

in: AI integration standards › Choose the interface and control access

A method for efficiently finding vectors likely to be close without checking every candidate.

in: Embeddings and search › Representations and retrieval methods

Proposed AI intended to learn, reason, and apply knowledge across a wide range of tasks; there is no universally accepted test for it.

in: AI foundations › What AI means

Computer systems performing tasks associated with perception, reasoning, learning, or action.

in: AI foundations › What AI means

Hypothetical AI that surpasses human cognitive capabilities across most domains.

in: AI foundations › What AI means

A computation that weights information according to its relevance to another representation.

in: Neural-network architectures › Architecture building blocks

Using technology to assist people in performing their tasks.

in: AI at work › What AI can change in a job

Verifying credentials and the identity they assert.

in: AI integration standards › Choose the interface and control access

Deciding whether an authenticated party has permission to access a particular resource or service.

in: AI integration standards › Choose the interface and control access

Converting spoken audio into text.

in: Speech and audio AI › Choose the speech task and measure its output

Having a system perform tasks with reduced direct human execution.

in: AI at work › What AI can change in a job

A method for calculating parameter gradients through a model’s chain of computations.

in: Deep learning › How a neural network learns

A pretrained model before a particular instruction or task adaptation.

in: Foundation models › What kind of model is it?

A reference approach against which another approach is compared.

in: AI evaluation › Choose measures that fit the task

The number of examples processed in a batch for a training iteration.

in: Deep learning › How a neural network learns

A defined collection of tests and a procedure used to compare performance.

in: AI evaluation › Choose measures that fit the task

Prior exposure to evaluation material that can distort measured generalization.

in: AI benchmarks › What makes two scores comparable?

A 16-bit floating-point format with a wider exponent range than FP16.

in: Model quantization › What fewer bits change

The number of bits used to represent a value.

in: Model quantization › What fewer bits change

A keyword-ranking method that uses term-frequency and document-length normalization.

in: Embeddings and search › Representations and retrieval methods

An organization that develops technical standards for certifying the source and history of media content.

in: AI media and information › What provenance can tell you

How well predicted probabilities match observed outcome frequencies.

in: AI evaluation › Choose measures that fit the task

Greenhouse-gas emissions attributed to an activity within stated boundaries.

in: AI environmental impact › What does a footprint include?

Estimating effects of interventions under explicit assumptions and evidence.

in: Causal AI › What makes a claim causal?

Generated intermediate reasoning text that need not reveal the model's actual causes.

in: LLM inference and reasoning › Choose how the next token is selected

As documented on 2026-10-09, rules that convert conversations into a tokenizable string in the format a model expects.

in: Local AI runtimes › Choose a runtime and application

Software that interacts with people through conversational messages.

in: AI foundations › What AI means

A saved model or training state that can be loaded later.

in: Large language models › From text to tokens and back

Splitting source material into pieces for processing or retrieval.

in: Retrieval-augmented generation › The retrieval and answer pipeline

A reference to a source intended to support a statement.

in: Retrieval-augmented generation › The retrieval and answer pipeline

Predicting which category an input belongs to.

in: Predictive machine learning › Match the method to the output

Model parameters not distributed to users.

in: Model access and licensing › Access, weights, and the OSI definition

Grouping items according to similarities in their data.

in: Predictive machine learning › Match the method to the output

Suggesting continuations or snippets while code is being written.

in: AI for coding › Use assistance without losing the thread

Operating a graphical interface through observations and input actions.

in: Computer-use agents › Interfaces and actions

Computational methods that extract information from images or video.

in: Computer vision › Match the task to the output

A factor influencing both a proposed cause and its measured outcome.

in: Causal AI › What makes a claim causal?

A cryptographically bound record that records an asset’s provenance, such as its origin, modifications, or use of AI.

in: AI media and information › What provenance can tell you

Reducing stored context by summarizing or otherwise compressing it so a workflow can continue.

in: Context engineering › What belongs in context

Designing which instructions, information, and state a model receives across a workflow.

in: Context engineering › What belongs in context

The maximum context length supported under a stated model and configuration.

in: Large language models › From text to tokens and back

Scheduling that lets new inference requests join work already being processed, rather than waiting for an entire fixed batch to finish.

in: Model serving › Model serving fundamentals

A network that uses shared local filters to detect structured patterns.

in: Neural-network architectures › Architecture building blocks

In the United States, the Copyright Act of 1976 was adopted on 19 October 1976. Copyright concerns legal rights in original works of authorship fixed in a tangible medium of expression.

in: AI law and regulation › Find the law, status and date

A measure comparing the directions of two vectors.

in: Math for AI › Read vectors, matrices, and tensors

An alternative outcome considered under a different action or condition.

in: Causal AI › What makes a claim causal?

A general-purpose processor that executes a computer’s instructions.

in: AI hardware › Processors, memory, and accelerator support

NVIDIA’s platform and tools for programming compatible GPUs.

in: AI hardware › Processors, memory, and accelerator support

Unauthorized transfer of information outside its permitted boundary.

in: AI security › Threats in prompts and data

Information unavailable at deployment improperly enters training or evaluation.

in: Data for AI › Build a dataset and keep evaluation honest

Deliberate manipulation of training or reference data to alter system behavior.

in: AI security › Threats in prompts and data

The recorded origin, permissions, and transformation history of data.

in: Data for AI › Build a dataset and keep evaluation honest

Using statistics and computation to analyze data and communicate findings.

in: AI foundations › What AI means

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

in: Data for AI › Build a dataset and keep evaluation honest

A model that follows branching conditions to reach a prediction.

in: Predictive machine learning › Match the method to the output

The inference stage that generates subsequent output tokens.

in: LLM inference and reasoning › Choose how the next token is selected

A model component that produces outputs from internal representations or context.

in: Neural-network architectures › Architecture building blocks

Machine learning that uses neural networks with multiple processing layers.

in: Deep learning › How a neural network learns

Manipulated or synthetic audio or visual media that seems authentic and features a person who appears to say or do something they never said or did, produced using artificial intelligence techniques.

in: AI media and information › What provenance can tell you

Removing or reducing noise in a representation.

in: Image and video generation › How generation methods work

A model that uses the same main network rather than selecting expert subnetworks.

in: Neural-network architectures › Architecture building blocks

A generative model learning to reverse a process that adds noise.

in: Image and video generation › How generation methods work

A digital representation of a particular physical system or process.

in: World models and simulation › Simulation, models, and digital twins

Representing data with fewer dimensions while retaining useful information.

in: Predictive machine learning › Match the method to the output

Training a student model to imitate signals from a teacher model.

in: Fine-tuning › Choose an adaptation method

A method that optimizes preferred responses over rejected responses without an explicit reinforcement-learning loop.

in: Preference and reward training › Feedback, reward models, and optimization

A learned numerical representation of text, images, or other items.

in: Embeddings and search › Representations and retrieval methods

AI that perceives and acts within a physical environment.

in: Embodied AI › What makes AI embodied?

A model component that converts inputs into internal representations.

in: Neural-network architectures › Architecture building blocks

A measure of uncertainty in a probability distribution.

in: Math for AI › Read vectors, matrices, and tensors

The setting that responds to actions with observations and feedback.

in: Reinforcement learning › The parts of a reward-based problem

One traversal of the defined training set, with each example processed once.

in: Deep learning › How a neural network learns

EU legislation establishing requirements for specified AI systems and models. In the EU, the Act is in force and generally applies from 2 August 2026, with provision-specific dates.

in: AI law and regulation › Find the law, status and date

Measuring behavior or performance against stated tasks and criteria.

in: AI evaluation › Choose measures that fit the task

A system that applies encoded specialist knowledge, represented as facts and rules, to draw conclusions in a specific domain.

in: Symbolic and neuro-symbolic AI › Rules and expert systems

Providing understandable accounts of a system’s outputs or operation.

in: AI interpretability › What an explanation does and does not show

The tradeoff between trying uncertain actions and using currently promising ones.

in: Reinforcement learning › The parts of a reward-based problem

The harmonic mean of precision and recall.

in: AI evaluation › Choose measures that fit the task

In the United States, the Copyright Act of 1976 was adopted on 19 October 1976. Section 107 provides a context-dependent copyright exception evaluated using statutory factors; it is not a universal permission.

in: AI law and regulation › Find the law, status and date

How well an answer stays supported by its supplied context.

in: Retrieval-augmented generation › The retrieval and answer pipeline

An input property used by a model.

in: Machine Learning › How models learn from data

Providing task examples inside the prompt.

in: Prompting and context › Instructions, examples, and response shape

Additional training that adapts an already trained model.

in: Foundation models › What kind of model is it?

Training a model to transform distributions through a learned continuous flow.

in: Image and video generation › How generation methods work

A broadly pretrained model adaptable to multiple downstream tasks.

in: Foundation models › What kind of model is it?

A 16-bit floating-point numerical format.

in: Model quantization › What fewer bits change

For the Frontier Model Forum’s membership definition, a general-purpose model that outperforms models widely deployed for at least 12 months, based on a range of conventional performance benchmarks or high-risk capability assessments.

in: AI model landscape › Recognize model families

An EU legal category concerning models capable of a broad range of distinct tasks. The EU AI Act, in force and generally applying from 2 August 2026, defines a general-purpose AI model by its significant generality and capacity to perform a wide range of distinct tasks.

in: AI law and regulation › Find the law, status and date

A model's ability to perform on new inputs beyond its training examples.

in: Machine Learning › Fit, overfit, and generalize

AI that produces new content such as text, images, audio, or code.

in: AI foundations › What AI means

As described in the format documentation checked 2026-10-09, a file format for storing models for inference with GGML and executors based on GGML.

in: Local AI runtimes › Choose a runtime and application

A processor suited to many parallel computations.

in: AI hardware › Processors, memory, and accelerator support

The direction and rate of a function's change across its inputs.

in: Math for AI › Read vectors, matrices, and tensors

An iterative method that uses gradients to update parameters in an effort to reduce an objective.

in: Deep learning › How a neural network learns

A family of RAG approaches using graphs to organize or retrieve information.

in: Knowledge graphs for AI › Represent entities and relationships

Connecting an output to supplied evidence or observations.

in: Retrieval-augmented generation › The retrieval and answer pipeline

A mechanism intended to restrict or check system behavior.

in: AI alignment › Goals, constraints and behavior

Generated content presented as factual despite being false or unsupported.

in: Large language models › From text to tokens and back

Stanford’s framework for evaluating foundation models with transparent, reproducible methods.

in: AI benchmarks › What makes two scores comparable?

A practical estimate or shortcut used to guide search or choice; it does not by itself guarantee the best answer.

in: Symbolic and neuro-symbolic AI › Rules and expert systems

AI systems falling within specified high-risk categories under the EU AI Act. In the EU, the Act is in force and generally applies from 2 August 2026; the AI Omnibus amendment sets later application dates for specified obligations.

in: AI law and regulation › Find the law, status and date

A workflow assigning a person a review or decision step.

in: Tool use and automation › How a model calls a tool

Retrieval that combines methods such as keyword and vector search.

in: Embeddings and search › Representations and retrieval methods

A setting chosen outside the model's learned parameter values.

in: Machine Learning › Fit, overfit, and generalize

Learning behavior from demonstrations.

in: Robot learning › Learning behavior and transferring it

Prompt injection arriving through retrieved content, files, or tool outputs.

in: AI security › Threats in prompts and data

Using a trained model to produce an output for an input.

in: Machine Learning › How models learn from data

Software that executes a model's computations.

in: Local AI runtimes › Choose a runtime and application

Generating or editing content within selected image regions.

in: Image and video generation › How generation methods work

A model additionally trained to respond to instructions.

in: Foundation models › What kind of model is it?

A four-bit integer representation used in some quantized models.

in: Model quantization › What fewer bits change

An attempt to bypass a model's intended behavioral or safety restrictions.

in: AI security › Threats in prompts and data

A vocabulary describing and validating the structure of JSON data.

in: Prompting and context › Instructions, examples, and response shape

A lightweight, transport-agnostic remote procedure call protocol that uses JSON as its data format.

in: AI integration standards › Choose the interface and control access

Retrieval based on matching textual terms.

in: Embeddings and search › Representations and retrieval methods

A stated boundary for training-data recency, not a guarantee of complete knowledge.

in: Large language models › From text to tokens and back

A structured representation of entities and their relationships.

in: Knowledge graphs for AI › Represent entities and relationships

Saved attention calculations reused when generating later tokens.

in: LLM inference and reasoning › Choose how the next token is selected

A known answer or target associated with a training example.

in: Machine Learning › How models learn from data

A model that estimates patterns or probabilities in language.

in: Natural language processing › Choose the language task

A model trained at substantial scale to process and generate language.

in: Large language models › From text to tokens and back

Elapsed time between a request and a specified response event.

in: Model serving › Model serving fundamentals

A space of learned representations rather than raw observations.

in: Image and video generation › How generation methods work

A stage of computation in a neural network.

in: Deep learning › How a neural network learns

A ranking based on specified tests, preferences, or scoring procedures.

in: AI benchmarks › What makes two scores comparable?

A setting that controls how strongly gradient descent adjusts parameters on each iteration.

in: Deep learning › How a neural network learns

Granting only the access necessary for an operation.

in: Tool use and automation › How a model calls a tool

As described by the project on 2026-10-09, an inference engine implemented primarily in C and C++.

in: Local AI runtimes › Choose a runtime and application

Using a language model to assess outputs against evaluation criteria. For agent trials, report the judged system's model, tools, scaffold, and environment alongside the result.

in: Evaluating AI systems › Define tests and choose graders

Operational practices specific to language-model applications.

in: AI observability › Operations across the model lifecycle

As documented on 2026-10-09, an application for serving local models through a server interface.

in: Local AI runtimes › Choose a runtime and application

Running model computations on a local machine.

in: Local and on-device AI › Where computation runs

An adaptation method that learns low-rank updates to selected model weights.

in: Fine-tuning › Choose an adaptation method

A mathematical function that measures prediction loss, which training typically aims to minimize.

in: Deep learning › How a neural network learns

Methods that improve task performance by learning patterns from data or experience.

in: Machine Learning › How models learn from data

A rectangular grid of numbers used to represent data or transformations.

in: Math for AI › Read vectors, matrices, and tensors

An extension that lets MCP servers deliver interactive user interfaces to hosts.

in: Model Context Protocol › Connection roles and server capabilities

A connector created by a host that communicates with one MCP server.

in: Model Context Protocol › Connection roles and server capabilities

A standardized way for a server to request additional information from a user through the client.

in: Model Context Protocol › Connection roles and server capabilities

An optional, opt-in capability that adds functionality beyond the core protocol.

in: Model Context Protocol › Connection roles and server capabilities

The AI application that coordinates user interaction and manages MCP client connections.

in: Model Context Protocol › Connection roles and server capabilities

A server-provided template of structured messages and instructions that a client can retrieve and customize with arguments.

in: Model Context Protocol › Connection roles and server capabilities

Data or content a server exposes to a client under a resource URI.

in: Model Context Protocol › Connection roles and server capabilities

A local process or remote service that exposes context and capabilities through MCP.

in: Model Context Protocol › Connection roles and server capabilities

An extension that lets a server return a task handle for certain requests so a client can poll for the eventual result.

in: Model Context Protocol › Connection roles and server capabilities

A named, schema-described capability that a server exposes for a client to invoke.

in: Model Context Protocol › Connection roles and server capabilities

Investigating internal model mechanisms associated with particular behaviors.

in: AI interpretability › What an explanation does and does not show

Apple’s interface for graphics and GPU computation.

in: AI hardware › Processors, memory, and accelerator support

An architecture that routes inputs through selected specialized subnetworks.

in: Neural-network architectures › Architecture building blocks

Practices for developing, deploying and maintaining machine-learning systems.

in: AI observability › Operations across the model lifecycle

As described by the project on 2026-10-09, an array framework for machine learning on Apple silicon.

in: Local AI runtimes › Choose a runtime and application

A multiple-choice benchmark that measures multitask accuracy across academic and professional subjects.

in: AI benchmarks › What makes two scores comparable?

A form of information, such as text, an image, or audio.

in: Multimodal AI › What counts as multimodal?

Systematic tendencies or errors that can disadvantage people or distort results.

in: AI safety and governance › Name the harms and responsibilities

Documentation describing a model’s intended uses, evaluation, and limitations.

in: AI model landscape › Read capability and documentation claims

A JSON-RPC protocol for applications to share context with language models and expose tools and capabilities.

in: Model Context Protocol › Connection roles and server capabilities

A change in a model's observed behavior or performance over time.

in: AI observability › Operations across the model lifecycle

Related models sharing a lineage, design, or release identity.

in: AI model landscape › Recognize model families

A service hosting and organizing model files and documentation.

in: Local AI runtimes › Choose a runtime and application

Terms governing a model's use, modification, and distribution.

in: Model access and licensing › Access, weights, and the OSI definition

Operating a model behind an interface that accepts inference requests and returns results.

in: Model serving › Model serving fundamentals

A system in which multiple agents interact or coordinate.

in: Agent memory and orchestration › Choose what to retain and coordinate

A model that processes or generates more than one kind of data.

in: Multimodal AI › What counts as multimodal?

AI designed or evaluated for a limited range of tasks.

in: AI foundations › What AI means

Computational methods for analyzing, understanding, or generating human language.

in: Natural language processing › Choose the language task

A model made of connected layers of adjustable calculations.

in: Deep learning › How a neural network learns

Approaches that combine neural methods with symbolic approaches to computing and reasoning.

in: Symbolic and neuro-symbolic AI › Rules and expert systems

Estimating the next token from preceding context.

in: Large language models › From text to tokens and back

A processor designed to accelerate selected neural-network computations.

in: AI hardware › Processors, memory, and accelerator support

A framework that lets a third-party application obtain limited access to an HTTP service, with a resource owner’s approval or on its own behalf.

in: AI integration standards › Choose the interface and control access

Finding and labeling objects and their locations in visual data.

in: Computer vision › Match the task to the output

The extent to which a job’s tasks could be affected by a technology.

in: AI at work › What AI can change in a job

Converting text visible in an image into machine-readable text.

in: Computer vision › Match the task to the output

As documented on 2026-10-09, software for managing and using models through local and cloud options.

in: Local AI runtimes › Choose a runtime and application

AI computation performed on the device handling the user's task.

in: Local and on-device AI › Where computation runs

Model parameters available for download; permissions depend on the license.

in: Model access and licensing › Access, weights, and the OSI definition

An AI system that meets the OSI definition; downloadable weights alone do not show that it meets the definition.

in: Model access and licensing › Access, weights, and the OSI definition

Coordinating models, tools, state, and workflow steps.

in: Agent memory and orchestration › Choose what to retain and coordinate

Learning training-data quirks that do not carry over to new data.

in: Machine Learning › Fit, overfit, and generalize

Adapting a model by training a limited set of parameters or added components.

in: Fine-tuning › Choose an adaptation method

Adjustable numerical values learned during training.

in: Large language models › From text to tokens and back

The estimated chance that at least one of k generated attempts passes the evaluation tests.

in: AI benchmarks › What makes two scores comparable?

A measure of how uncertain a model is about observed tokens.

in: AI evaluation › Choose measures that fit the task

Under the EU GDPR, which applies from 25 May 2018, personal data means information relating to an identified or identifiable natural person.

in: AI law and regulation › Find the law, status and date

A rule or model that specifies actions for an observed situation.

in: Reinforcement learning › The parts of a reward-based problem

Training after a base model is created to change behavior or capabilities.

in: Foundation models › What kind of model is it?

Quantizing a model after its original training.

in: Model quantization › What fewer bits change

The share of predicted positive cases that are actually positive.

in: AI evaluation › Choose measures that fit the task

The inference stage that processes the supplied prompt.

in: LLM inference and reasoning › Choose how the next token is selected

Initial training that develops broadly reusable model capabilities.

in: Foundation models › What kind of model is it?

A description of how probability is assigned across possible outcomes.

in: Math for AI › Read vectors, matrices, and tensors

Input supplied to a model to guide its response.

in: Prompting and context › Instructions, examples, and response shape

Reusing computation for a matching prompt prefix in later requests.

in: Context engineering › What belongs in context

Attacker-controlled input attempting to redirect a system from its intended instructions.

in: AI security › Threats in prompts and data

A mixed-precision preset in llama.cpp's four-bit K-quant family.

in: Model quantization › What fewer bits change

Training LoRA adapters while keeping the base model quantized.

in: Fine-tuning › Choose an adaptation method

Representing model values with fewer numerical bits.

in: Model quantization › What fewer bits change

Working memory used by a running computer.

in: AI hardware › Processors, memory, and accelerator support

A model combining predictions from multiple decision trees.

in: Predictive machine learning › Match the method to the output

A model designed or trained to devote computation to multi-step problem solving.

in: Foundation models › What kind of model is it?

Tokens used for intermediate reasoning; their accounting and visibility vary by system.

in: LLM inference and reasoning › Choose how the next token is selected

The share of actual positive cases successfully identified.

in: AI evaluation › Choose measures that fit the task

A system selecting or ranking items for a user or situation.

in: Predictive machine learning › Match the method to the output

A network that carries an internal state between sequence steps.

in: Neural-network architectures › Architecture building blocks

Predicting a numerical value.

in: Predictive machine learning › Match the method to the output

Constraints or penalties intended to reduce unnecessary model complexity.

in: Machine Learning › Fit, overfit, and generalize

Learning action choices through feedback expressed as rewards.

in: Reinforcement learning › The parts of a reward-based problem

A component reordering retrieved candidates by a relevance score.

in: Retrieval-augmented generation › The retrieval and answer pipeline

Generating an answer using information retrieved from an external collection.

in: Retrieval-augmented generation › The retrieval and answer pipeline

A component selecting potentially relevant information for a query.

in: Retrieval-augmented generation › The retrieval and answer pipeline

A signal expressing how desirable an action or outcome is.

in: Reinforcement learning › The parts of a reward-based problem

Exploiting a reward measure without achieving its intended objective.

in: AI alignment › Goals, constraints and behavior

A model that scores outputs according to a learned preference or objective.

in: Preference and reward training › Feedback, reward models, and optimization

Reinforcement learning in which feedback supplied by AI systems provides the preference signal.

in: Preference and reward training › Feedback, reward models, and optimization

Reinforcement learning in which rewards are derived from human feedback.

in: Preference and reward training › Feedback, reward models, and optimization

Reinforcement learning that uses rewards checked against defined criteria, such as matching a correct answer or passing code unit tests.

in: Preference and reward training › Feedback, reward models, and optimization

AMD’s software stack for supported accelerator computing.

in: AI hardware › Processors, memory, and accelerator support

A tensor file format designed to avoid executable object deserialization.

in: Local AI runtimes › Choose a runtime and application

Selecting an output token according to a probability distribution.

in: LLM inference and reasoning › Choose how the next token is selected

An isolated execution environment intended to limit permitted actions.

in: Tool use and automation › How a model calls a tool

An empirical relationship between resources and measured model performance.

in: Pretraining › Read scaling evidence

Assigning visual regions or pixels to meaningful categories or objects.

in: Computer vision › Match the task to the output

Attention that relates positions within the same input sequence.

in: Neural-network architectures › Architecture building blocks

Operating a service on infrastructure under your control.

in: Local and on-device AI › Where computation runs

Learning from targets constructed from the input data itself.

in: Machine Learning › How models learn from data

Retrieval that uses represented meaning rather than matching only textual terms.

in: Embeddings and search › Representations and retrieval methods

Applying learning from simulation to a physical environment.

in: Robot learning › Learning behavior and transferring it

An extension that exposes Agent Skills through MCP resources and related methods.

in: Model Context Protocol › Connection roles and server capabilities

A relatively small language model; the term does not specify a universal size cutoff.

in: Foundation models › What kind of model is it?

Detecting when speakers change and labeling the individual voices in audio.

in: Speech and audio AI › Choose the speech task and measure its output

A draft process proposes tokens that a target model verifies.

in: LLM inference and reasoning › Choose how the next token is selected

Communication with a process through its standard input and output streams.

in: Model Context Protocol › Connection roles and server capabilities

A sequence model that updates a compact internal state over time.

in: Neural-network architectures › Architecture building blocks

In the 2026-07-28 specification, an HTTP-based transport in which a client sends each message as a POST and receives a JSON response or a request-scoped event stream.

in: Model Context Protocol › Connection roles and server capabilities

Output constrained to a specified machine-readable structure.

in: Prompting and context › Instructions, examples, and response shape

Adapting a model by training it on examples that pair inputs with desired outputs.

in: Fine-tuning › Choose an adaptation method

Learning from examples paired with target outputs.

in: Machine Learning › How models learn from data

A benchmark family that evaluates software changes by asking a system to resolve repository issues.

in: AI benchmarks › What makes two scores comparable?

Agreeing with a user's beliefs or preferences at the expense of accuracy.

in: AI alignment › Goals, constraints and behavior

AI that represents relevant knowledge explicitly with symbols and manipulates those representations using rules or other procedures.

in: Symbolic and neuro-symbolic AI › Rules and expert systems

Artificially generated examples used instead of or alongside observed data.

in: Data for AI › Build a dataset and keep evaluation honest

An application-supplied instruction with a designated priority that varies by system.

in: Prompting and context › Instructions, examples, and response shape

The share of trials meeting explicitly defined completion criteria.

in: Evaluating AI systems › Define tests and choose graders

A setting that changes how concentrated the probabilities are when selecting a token.

in: LLM inference and reasoning › Choose how the next token is selected

A numerical array that can have any number of dimensions.

in: Math for AI › Read vectors, matrices, and tensors

Held-out examples used to estimate performance after development choices.

in: Data for AI › Build a dataset and keep evaluation honest

Computation spent while solving a problem rather than training the model.

in: LLM inference and reasoning › Choose how the next token is selected

Generating spoken audio from text.

in: Speech and audio AI › Choose the speech task and measure its output

The amount of work completed per unit of time, such as final responses or output tokens during a benchmark.

in: Model serving › Model serving fundamentals

Elapsed time from sending a request until the first streamed output is received; state the measurement boundaries.

in: Model serving › Model serving fundamentals

A unit encoding model input or output, often a piece of a word.

in: Large language models › From text to tokens and back

A procedure that converts text into a model's token identifiers.

in: Large language models › From text to tokens and back

A rate of token processing or generation, interpreted under the stated workload, measurement interval, and token-counting method.

in: Model serving › Model serving fundamentals

A model requests a software operation with structured arguments for a host to execute.

in: Tool use and automation › How a model calls a tool

Restricting token sampling to the k highest-probability candidates.

in: LLM inference and reasoning › Choose how the next token is selected

Sampling from the smallest set of likely candidates whose probabilities reach a chosen cumulative share.

in: LLM inference and reasoning › Choose how the next token is selected

Adjusting a model using data or experience.

in: Machine Learning › How models learn from data

The examples used to adjust a model.

in: Data for AI › Build a dataset and keep evaluation honest

Applying learning from one setting to another task or domain.

in: Machine Learning › How models learn from data

A neural-network architecture that uses attention to update representations.

in: Neural-network architectures › Architecture building blocks

A historical test based on whether conversation can distinguish a machine from a person.

in: History of AI › Milestones and tests

Failing to capture useful patterns even in the training data.

in: Machine Learning › Fit, overfit, and generalize

A memory arrangement allowing processors to share a common memory pool.

in: AI hardware › Processors, memory, and accelerator support

Finding patterns in data without supplied target outputs.

in: Machine Learning › How models learn from data

Examples used to choose settings or compare development decisions.

in: Data for AI › Build a dataset and keep evaluation honest

An ordered collection of numbers representing a quantity or item.

in: Math for AI › Read vectors, matrices, and tensors

A datastore that stores vector representations and supports similarity search.

in: Embeddings and search › Representations and retrieval methods

A data structure that organizes vectors to speed up searches among them.

in: Embeddings and search › Representations and retrieval methods

A loosely defined practice of building software through heavy AI generation.

in: AI for coding › Use assistance without losing the thread

A model that connects visual information with language, for example by aligning image and text representations.

in: Multimodal AI › What counts as multimodal?

A model that uses visual and language information to produce actions for controlling a robot.

in: Embodied AI › What makes AI embodied?

As checked on 2026-10-09, vLLM is an inference and serving engine for language models.

in: Model serving › Model serving fundamentals

Generating speech that resembles a reference person's voice.

in: Speech and audio AI › Choose the speech task and measure its output

Memory directly available to a graphics processor.

in: AI hardware › Processors, memory, and accelerator support

Direct and indirect water use attributed to an activity within stated boundaries.

in: AI environmental impact › What does a footprint include?

Embedding a detectable signal in content to indicate origin or ownership.

in: AI media and information › What provenance can tell you

Learned coefficients, often used loosely to refer to all model parameters.

in: Large language models › From text to tokens and back

Transcription errors measured relative to the number of reference words.

in: Speech and audio AI › Choose the speech task and measure its output

A learned representation that predicts how an environment evolves under actions.

in: World models and simulation › Simulation, models, and digital twins

Requesting a task without supplying worked examples.

in: Prompting and context › Instructions, examples, and response shape