AI glossary
Plain-language explanations of AI terms and acronyms.
200 terms & acronyms
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.