AI glossary

A

Accuracy: The fraction of evaluated examples a model gets right. Accuracy alone can be misleading when one category is much more common than the others.
Agent: An AI system that can choose actions, use tools, and work through steps toward a goal. How independently it operates depends on its design and permissions.
Agentic workflow: A process in which an AI system helps decide which steps or tools to use, rather than following only a fixed sequence.
AGIArtificial general intelligence: A proposed kind of AI able to perform a broad range of intellectual tasks at a human level or beyond. There is no single agreed test for when a system qualifies.
AIArtificial intelligence: Software that performs tasks associated with human intelligence, such as recognizing speech, making predictions, or generating text.
Alignment: The effort to make an AI system behave consistently with intended goals and human expectations. It includes how a system handles instructions, uncertainty, and harmful requests.
APIApplication programming interface: A defined way for software to request data or actions from another service. An AI API lets an application send inputs to a model and receive outputs.
API key: A credential used by software to identify itself to a service. For an AI service, it may also determine which account pays for requests.
Artificial neural network: A machine-learning model built from connected layers of adjustable calculations. Training changes those calculations so the network becomes better at its task.
ASRAutomatic speech recognition: Technology that converts spoken audio into written text. It is used for transcription, captions, and voice interfaces.
Attention: A mechanism that helps a model weigh relationships between parts of its input. In a language model, it helps connect a word or token with relevant surrounding information.
Autoregressive model: A model that predicts the next part of a sequence using the parts already available. Many language models generate text one token at a time this way.

B

Backpropagation: A training method that works backward through a neural network to calculate how its adjustable values contributed to an error.
Base model: A model after its initial training, before additional adaptation for a particular product or task. A base language model may predict text without reliably following chat instructions.
Batch: A group of examples processed together during training, or a group of requests processed together during inference. Batching can improve hardware efficiency.
Batch size: The number of examples or requests processed together. A larger batch often needs more memory and can change training behavior or response timing.
Benchmark: A defined test used to compare AI systems. A benchmark score describes performance on that test, not every task a user might attempt.
BF16Bfloat16: A 16-bit number format often used in AI computation. Its range helps large models train and run with less memory than 32-bit numbers.
Bias: A systematic tendency in data, predictions, or behavior. Depending on context, it can mean a statistical pattern or an unfair difference in how a system treats people.
BPEByte pair encoding: A method often used to build a tokenizer's vocabulary by combining frequently occurring pieces of text. Tokens may be words, word parts, or smaller pieces.

C

Chain of thought: Intermediate reasoning steps used while working toward an answer. Some systems use internal reasoning that is not shown; a visible explanation is not proof of the model's exact internal process.
Chat model: A language model adapted for conversation, usually with messages labeled by role. It may also support tool calls, images, audio, or other inputs.
Checkpoint: A saved state of a model during or after training. It allows training to resume or a particular model version to be used later.
Chunking: Splitting a document into smaller pieces for storage, search, or model input. Chunk size and overlap affect what information can be retrieved.
Classification: Assigning an input to one or more categories. A model might label a message as a billing question, technical issue, or cancellation request.
CLIPContrastive Language–Image Pre-training: A model approach that learns relationships between images and text by comparing matching and nonmatching pairs. It is useful for tasks such as image search.
Clustering: Grouping similar examples without requiring a category label for every example. It can help reveal patterns in a dataset.
CNNConvolutional neural network: A neural network that looks for local patterns using filters that move across an input. CNNs are widely used for image-related tasks.
Coding agent: An AI agent that works with software projects using tools such as file editing, code search, terminals, and tests. Its access and approval settings determine what it can do.
Completion: Output generated by a language model in response to an input. The term comes from systems that continue a supplied piece of text.
Compute: The processing work and hardware capacity needed to train or run a model. More compute can mean more time, more processors, or both.
Computer vision: AI methods for interpreting images and video, such as identifying objects, reading a scene, or measuring movement.
Context: The information available to a model for a particular request. It can include instructions, conversation history, documents, images, and tool results.
Context compression: Shortening or summarizing information to fit more useful material into a model's available context. Compression can lose details that later turn out to matter.
Context engineering: Designing how an AI system receives the information it needs. This includes selecting, organizing, retrieving, and trimming instructions, history, and supporting material.
Context window: The amount of information a model can handle in one request, usually measured in tokens. How inputs and generated output share that limit depends on the model and service.
Contrastive learning: Training a model to bring related examples closer together in its representation and push unrelated examples farther apart.
Cosine similarity: A measure of how closely two vectors point in the same direction. It is commonly used to compare embeddings in semantic search.
Cross-attention: An attention mechanism that connects information from different sequences or representations, such as a text description and an image representation.
CUDACompute Unified Device Architecture: NVIDIA's platform for running general-purpose computation on its GPUs. Many AI libraries use it to accelerate training and inference.

D

Data augmentation: Creating additional training examples by changing existing ones, such as cropping an image. The changes should preserve the information relevant to the task.
Data contamination: Evaluation material appearing in training data or other information available to the system. It can make a reported score look better than performance on genuinely new material.
Data drift: A change in the data a model sees over time. A model trained on older patterns may become less useful when real-world inputs change.
Data labeling: Adding the answers or categories a model is meant to learn from. Labels might identify an object in an image or mark whether a response is helpful.
Dataset: A collection of examples used for training, evaluation, or analysis. A dataset can contain text, images, audio, structured records, or a mixture of formats.
Decoder: A model component that produces an output from a representation or earlier output. Decoder-based language models often generate text token by token.
Decoding: The process of turning a model's predictions into an output sequence. Decoding settings influence which next token is selected.
Deep learning: Machine learning that uses neural networks with many processing layers. It powers many modern language, speech, and image systems.
Dense model: A model that uses most of its parameters for each input, in contrast with architectures that activate selected groups of experts.
Diffusion model: A generative model that learns to create data through a process that reverses added noise. Diffusion-based approaches are widely used for image generation.
Distillation: Training one model to learn from the outputs or behavior of another model. It is often used to make a smaller or faster model approximate a larger one.
DPODirect preference optimization: Adapting a model using examples of preferred and less-preferred responses. It learns from those comparisons without a separate reinforcement-learning loop of the usual RLHF kind.
Dropout: A training technique that temporarily disables selected parts of a neural network to reduce overreliance on particular connections.

E

Embedding: A numerical representation of information, such as a passage or image, designed so useful relationships can be compared mathematically. It is commonly used for semantic search.
Encoder: A model component that converts an input into a representation useful for another task, such as search, classification, or generation.
Encoder–decoder model: An architecture with one component that represents the input and another that generates the output. It is used for tasks such as translation and summarization.
Endpoint: A network address an application calls to use a service. An inference endpoint accepts model inputs and returns the model's results.
Ensemble: A system that combines multiple models or predictions. Combining different approaches can improve reliability, although it adds complexity or cost.
Epoch: One complete pass through a training dataset. Training may involve several epochs, depending on the data and method.
Evaluation: Testing how well an AI system performs for a defined purpose. Useful evaluations check actual tasks and failure cases, rather than relying only on a general benchmark score.
Expert: A specialized component within a mixture-of-experts model. A routing mechanism selects which experts process a particular input.
Explainability: Methods for helping people understand why a model produced a result. An explanation can be an approximation rather than a complete account of internal computation.

F

F1 score: A score combining precision and recall using their harmonic mean. A strong F1 score generally requires both few false alarms and few missed positive cases.
Feature: An input property used by a model, such as a measurement, word pattern, or representation learned from data.
Few-shot prompting: Including a small number of examples in a prompt to demonstrate the desired task or response format. The examples guide the request without retraining the model.
Fine-tuning: Further training an existing model on additional examples to adapt its behavior or task performance. This changes model parameters, unlike simply supplying a longer prompt.
Foundation model: A broadly trained model that can be adapted to many tasks. Language models and some image or multimodal models fall into this category.
FP1616-bit floating point: A number format used to reduce the memory and computation needed for AI workloads. Its precision and numerical range differ from FP32 and BF16.
FP3232-bit floating point: A floating-point number format with more precision and memory use than 16-bit formats. It is also called single precision.
Function calling: A way for a model to request a structured call to an application's function or tool. The application, rather than the model's text alone, decides whether to execute the call.

G

GANGenerative adversarial network: A generative approach that trains one network to produce examples while another learns to distinguish generated examples from real ones.
Generative AI: AI that creates outputs such as text, images, audio, video, or code from patterns learned during training.
GPUGraphics processing unit: A processor that can perform many calculations in parallel. GPUs are widely used to train AI models and run them efficiently.
Gradient: A calculation describing how a change in a model's adjustable values would affect its training error. Training uses gradients to choose updates.
Gradient descent: An optimization method that repeatedly adjusts model parameters in a direction intended to reduce training error.
Grounding: Connecting an AI response to relevant evidence, such as documents, search results, or database records. Grounding can improve factual support but does not guarantee a correct answer.
Ground truth: The reference answers or labels used to train or evaluate a model. These references can still contain mistakes or subjective judgments.
Guardrail: A rule, check, or control used to constrain an AI system's inputs, outputs, or actions. Guardrails can operate before, during, or after a model call.

H

Hallucination: An AI output that presents invented or unsupported information as if it were established fact. A fluent or confident answer can still be a hallucination.
Hidden state: An internal numerical representation maintained or produced by a model while processing information. Its structure depends on the model architecture.
Human in the loop: A process that includes a person in reviewing, approving, correcting, or guiding an AI system's work.
Hybrid search: Search that combines methods such as keyword matching and embedding-based retrieval. It can find both exact names and conceptually related material.
Hyperparameter: A setting chosen to control a training or model process, rather than a value learned directly from examples. Learning rate and batch size are common training hyperparameters.

I

Image-to-image: Generating or editing an image using another image as an input. A text instruction or other controls may also guide the result.
Inference: Using a trained AI model to produce an output from new input. Chat replies, image generation, speech transcription, and classification can all involve inference.
Inference server: Software that loads a model and handles requests to run it. It may manage batching, memory use, streaming, and access to accelerators.
Inpainting: Using an AI model to replace or create content inside a selected part of an image while using the surrounding image as context.
Input token: A token representing information supplied to a model, such as instructions or conversation context. Some services charge different rates for input and output tokens.
Instruction hierarchy: Rules describing which instructions take priority when they conflict. Applications may distinguish system rules, developer instructions, user requests, and untrusted retrieved content.
Instruction tuning: Training a model on examples of instructions and desired responses so it becomes better at following requests.
INT88-bit integer: An integer number format often used when quantizing model values. It can reduce memory use, but quality depends on how quantization is performed.

J

JSONJavaScript Object Notation: A text format for structured data using objects, lists, and values. AI applications often use JSON to exchange tool arguments and structured responses.

K

Knowledge cutoff: The point up to which a model's training data is generally described as current. It does not guarantee knowledge of every earlier fact, and tools can supply newer information.
Knowledge graph: A representation of entities and the relationships between them. It can help a system connect facts across people, places, products, or concepts.
KV cacheKey–value cache: Stored attention information from tokens a model has already processed. Reusing it can speed up text generation, but it uses memory.

L

Label: An expected answer or category attached to a training example. An image might be labeled as containing a bicycle.
Language model: A model trained to represent patterns in language, often by predicting missing or following text. Language models can support tasks beyond producing chat replies.
Latency: The delay between a request and a response. For streaming output, the wait for the first token and the wait for the complete answer are different measurements.
Latent space: An internal representation in which a model expresses features of data as numbers. Generative models can work in this representation before producing a visible or audible result.
Learning rate: A setting that controls the size of parameter updates during training. Updates that are too large or too small can make training ineffective.
LLMLarge language model: A language model with a large number of learned parameters, trained on substantial text data. Many AI assistants use LLMs to understand requests and generate responses.
LLMOpsLarge language model operations: The practices used to run and maintain language-model applications, including prompt versions, model selection, evaluation, monitoring, and cost controls.
Logit: A model's raw score for a possible output, before scores are converted into probabilities. In text generation, logits help determine the next token.
LoRALow-rank adaptation: A fine-tuning technique that trains small added parameter sets while keeping most of the original model fixed. It can make adaptation less expensive than updating every parameter.
Loss: A numerical measure of how far a training result is from the desired result. Training aims to improve this measure, but a lower loss does not automatically mean better real-world behavior.

M

Machine learning: Methods that let software learn patterns from data rather than requiring a person to write a separate rule for every case.
MCPModel Context Protocol: An open protocol through which AI applications can connect to tools, data, and other capabilities supplied by servers. The application controls permissions and how those capabilities are used.
MCP client: The part of an AI application that connects to an MCP server and exchanges protocol messages with it.
MCP server: A program that exposes capabilities through the Model Context Protocol. These can include tools, data resources, or prompt templates.
Memory: Information an AI application preserves for use in later interactions. It may be stored in a database or file; it does not necessarily mean the model's trained parameters changed.
Metadata: Information describing other data, such as an author, date, file type, or topic. Search systems use metadata to organize and filter results.
MLMachine learning: An abbreviation for machine learning: methods that learn patterns from data to make predictions or perform tasks.
MLOpsMachine learning operations: The practices used to deploy, monitor, and maintain machine-learning systems, including model versions, data pipelines, and performance checks.
Model: The learned system that maps inputs to outputs. A deployed AI product may combine a model with tools, interfaces, retrieval, permissions, and other software.
Model card: Documentation describing a model's intended uses, training information, evaluation results, and known limitations. The amount and quality of detail vary by publisher.
Model merging: Combining parameters or adaptations from multiple models to create another model. Success depends on the models and method; it does not simply combine all their abilities.
MoEMixture of experts: An architecture that routes inputs through selected expert components. It can provide a large total parameter count while using only part of the model for each token.
Multi-agent system: A system in which multiple AI agents work together or divide tasks. Coordination adds overhead and does not by itself guarantee a better result.
Multimodal model: A model that works with more than one kind of information, such as text and images or audio and text. Supported input and output types vary by model.

N

NLPNatural language processing: Methods for working with human language in text or speech, including translation, classification, search, and language generation.

O

OCROptical character recognition: Technology that reads text from images or scanned documents and converts it into machine-readable text.
On-device AI: AI inference performed on a user's device rather than entirely on a remote server. Available models and performance depend on the device's hardware and software.
Open source: Software distributed under a license that grants specified rights to inspect, use, modify, and share its source code. A downloadable AI model is not automatically open-source software.
Open weights: Model parameters made available for download. The license determines how they may be used or redistributed; open weights do not necessarily include training code or data.
Optimizer: An algorithm that decides how to update model parameters during training using information such as gradients. Adam and stochastic gradient descent are common examples.
Outpainting: Generating new image content beyond the boundaries of an existing image, using the original image as context.
Output token: A token generated by a model. Depending on the service, reported or billable output can include internal reasoning tokens as well as the visible answer.
Overfitting: When a model learns training examples so closely that it performs poorly on new examples. Strong training results alone do not prove that a model generalizes well.

P

Parameter: A value learned during model training, such as a neural-network weight. Parameter count describes model size, but it is not a complete measure of capability.
PEFTParameter-efficient fine-tuning: Methods that adapt a model by training a relatively small set of parameters. LoRA is one example.
Perplexity: A measure of how well a language model predicts a token sequence. Comparisons need the same data and tokenization; lower perplexity does not guarantee a better assistant.
Pipeline: A series of processing steps that pass information onward. An AI pipeline might extract text, retrieve relevant passages, and then generate an answer.
Precision: For a chosen positive category, the fraction of predicted positive results that are actually correct. It describes how trustworthy those positive detections are.
Prefill: The stage of language-model inference that processes the input context before generating new output tokens.
Pretraining: The initial broad training used to build a model before more specific adaptation. For many language models, it involves learning to predict text from large collections of examples.
Prompt: The input and instructions supplied to guide a model's response. In a chat application, the full prompt can include more than the message typed by the user.
Prompt engineering: Writing and testing instructions, examples, and context to improve a model's results for a particular task.
Prompt injection: An attempt to make an AI system treat untrusted content as instructions that override its intended task or rules. It can arrive through a webpage, document, tool result, or user input.
Prompt template: A reusable instruction structure with places for changing information, such as a customer name, document, or desired output format.
Pruning: Removing or disabling selected model parameters or structures to reduce size or computation. Quality depends on what is removed and how the model is adapted afterward.

Q

QLoRAQuantized low-rank adaptation: A fine-tuning approach that combines a quantized base model with trainable LoRA adapters to reduce the memory needed for adaptation.
Quantization: Representing model values with fewer bits to reduce memory use or computation. Effects on speed and output quality depend on the method, hardware, and model.

R

RAGRetrieval-augmented generation: Retrieving relevant information and supplying it to a model when generating an answer. The material supplements the request without requiring retraining on that information.
Rate limit: A service's limit on how often or how much an application can request within a period. Limits may apply to requests, tokens, concurrent work, or another measure.
Reasoning model: A model designed or trained to spend additional computation working through a problem before answering. The approach and visibility of intermediate reasoning vary by system.
Recall: For a chosen positive category, the fraction of actual positive examples the system finds. It describes how much relevant material or how many true cases were missed.
Red teaming: Deliberately testing an AI system with challenging or adversarial cases to discover weaknesses and harmful failure modes.
Regression: Predicting a numerical value, such as a price or temperature, rather than assigning only a category. In software testing, regression instead means a previously working behavior breaks.
Reinforcement learning: Training an agent or model using feedback about the outcomes of its actions. The goal is to learn behavior that achieves more reward over time.
Reranking: Taking an initial set of search results and using another scoring step to put the most relevant results first.
Retrieval: Finding information relevant to a request. Retrieval can use keywords, embeddings, metadata filters, or combinations of these methods.
RLHFReinforcement learning from human feedback: A family of methods using human preference information to guide behavior through reinforcement learning. Human feedback is often used to train a reward model first.
RNNRecurrent neural network: A neural network that carries a state from one step of a sequence to the next. It has been used for text, time series, and speech.

S

Sampling: Selecting an output from a model's predicted possibilities rather than always choosing the highest-scoring option. Sampling settings influence output variation.
Sandbox: An isolated or restricted environment for running software or tools. Its protection depends on the permissions and boundaries the system enforces.
SDKSoftware development kit: Libraries and tools that help developers use a platform. An AI SDK can simplify authentication, model requests, streaming, and tool integration.
Seed: A value used to initialize a random process. Reusing a seed may improve reproducibility, but identical results also depend on the model, software, settings, and hardware.
Self-attention: An attention mechanism that compares positions within the same sequence, helping a model connect relevant parts of an input with each other.
Self-hosted model: A model run on infrastructure controlled by the operator, rather than accessed only through another provider's hosted model service.
Self-supervised learning: Learning from training targets derived from the data itself, rather than a separate human label for every example. Predicting the next token in text is one example.
Semantic search: Search based on meaning or related concepts, rather than only exact word matches. It often uses embeddings to compare a query with candidate results.
SFTSupervised fine-tuning: Adapting a model with examples of inputs and desired outputs. It is commonly used to teach response styles, formats, or instruction-following behavior.
Skill: A reusable package of instructions, and sometimes supporting files, that an AI application or agent can use for a task. The format and installation process depend on the application.
Small language model: A relatively compact language model that uses fewer parameters or less compute than larger alternatives. There is no universal size boundary for the label.
Speculative decoding: An inference technique that proposes output tokens with a faster draft process and checks them with the target model. It can reduce generation time when the setup is suitable.
Speech-to-text: Converting spoken audio into written text. The output might be a transcript, captions, or text used by another part of an application.
Streaming: Delivering output in pieces as it becomes available instead of waiting for the complete result. The request may still be running while earlier pieces are displayed.
Structured output: A model response constrained or validated against a specified format, such as a JSON schema. A valid format does not by itself guarantee factual correctness.
Supervised learning: Training with examples that include expected answers or labels. The model learns to map inputs to those supplied targets.
Synthetic data: Data produced artificially rather than collected directly from the original real-world process. It can help with training or testing, but may introduce its own errors or biases.
System prompt: An instruction message used by an application to guide a model's behavior. Its priority and treatment depend on the model service and application.

T

Temperature: A setting that changes how sharply a model favors higher-scoring next tokens during sampling. Higher settings usually allow more varied choices; they do not make a model more knowledgeable.
Tensor: A numerical array with one or more dimensions. AI systems use tensors to represent inputs, model parameters, and intermediate calculations.
Test set: Examples held back to evaluate a model after training and model selection. Repeatedly tuning to a test set weakens its value as an independent check.
Text-to-image: Generating an image from a text description. The result depends on the model, prompt, settings, and any additional guidance.
Text-to-speech: Generating spoken audio from written text. Systems may offer different voices, languages, speaking styles, or controls.
Throughput: The work a system completes in a period, such as requests per second or tokens per second. High throughput is different from a short wait for one request.
Token: A unit of input or output used by a model's tokenizer. A token may be a word, part of a word, punctuation, or another encoded piece of information.
Tokenization: Splitting and encoding information into the units a model processes. Different models can tokenize the same text differently.
Tool calling: A model requesting an external capability through structured arguments, such as a search or calculation. The surrounding application decides which tools exist and whether a requested action runs.
Top-k: A sampling setting that limits the next-token candidates to a specified number of the highest-scoring choices.
Top-p: A sampling setting that keeps next-token candidates whose combined probability reaches a chosen threshold. It is also called nucleus sampling.
TPUTensor processing unit: An accelerator developed by Google for machine-learning computation. Its software and deployment options differ from those of a general-purpose GPU.
Training: Adjusting a model using data and a learning objective. Training changes the model, while ordinary inference applies an existing model to an input.
Training data: The examples used to adjust a model during training. A model can learn useful patterns and also inherit gaps, errors, or biases from those examples.
Training split: The portion of a dataset used to fit a model. Separate validation and test splits help assess decisions and performance on examples not used for that fitting.
Transfer learning: Using a model or representation learned for one setting as a starting point for another task or dataset.
Transformer: A neural-network architecture built around attention mechanisms. It is used in many language models and also in systems for images, audio, and other data.
TTFTTime to first token: The time from sending a request until the first output token becomes available. It measures responsiveness for streaming language models.
TTSText-to-speech: An abbreviation for technology that turns written text into spoken audio. Voice and language options depend on the system.

U

Underfitting: When a model has not learned enough of the relevant patterns to perform well, including on its training examples.
Unsupervised learning: Learning patterns from data without an expected answer label for every example. Clustering is one common example.

V

Validation set: Examples used to guide training choices or model selection without directly fitting the model to those examples in the usual training loop. A separate test set supports final evaluation.
Vector: An ordered collection of numbers. In AI search, a vector often represents an embedding that can be compared with other embeddings.
Vector database: A database designed to store vectors and find nearby or similar vectors efficiently. It is often used to retrieve passages or other items by their embeddings.
VLMVision-language model: A model that works with both visual information and language. It might describe images, answer image questions, or support other visual tasks.
VRAMVideo random-access memory: Memory associated with a graphics processor. Model weights, working data, and attention caches can all compete for this memory during inference.

W

Weights: Learned numerical values that influence a neural network's calculations. Downloading weights does not necessarily provide training data or the right to redistribute the model.
Workflow: A sequence of tasks and decisions organized to achieve an outcome. AI may handle individual steps, while others can be ordinary software or human review.

Z

Zero-shot prompting: Asking a model to perform a task without supplying examples of that task in the prompt. The model relies on its training and the current instructions.

Terms & definitions

No matching terms

A

B

C

D

E

F

G

H

I

J

K

L

M

N

O

P

Q

R

S

T

U

V

W

Z