Artificial Intelligence: Glossary

Artificial intelligence is a fast-moving technology that requires knowledge and understanding of many jargon terms. Use this glossary as a reference as you continue learning about genAI and its applications.

Jump to a definition in the table or review the complete glossary:

AlgorithmsGenerative AI (genAI)Multi-Shot Prompting (Few-Shot Prompting)Small Language Model (SLM)Artificial General Intelligence (AGI)Generative Pretrained Transformer (GPT)Natural Language Processing (NLP)Supervised LearningArtificial Intelligence (AI)Graphical Processing Unit (GPU)Neural NetworksSycophancyBig DataHallucinationOverfittingSynthetic DataChain-of-Thought PromptingLanguage ModelParameterTokenComputeLarge Language Model (LLM)Probabilistic ModelTraining DataDatasetMachine LearningPromptTuring TestDeep LearningModelRetrieval-Augmented GenerationUnsupervised LearningFine-TuningMultimodalSingle-Shot PromptingZero-Shot Prompting

Algorithms

Any rule or set of rules that can be used to solve problems. Examples include automated product recommendations, long division, or a recipe for chocolate cake.

Artificial General Intelligence (AGI)

A (largely hypothetical) general term, referring to an artificial system that can perform a wide range of tasks at the same level or higher as a human.

Artificial Intelligence (AI)

“Intelligence” demonstrated by machines, often implemented and exhibited differently than human intelligence. Also, subfield of computer science concerned with researching and developing tools that enable computers to act “intelligently.”

Big Data

Data that remains difficult to analyze with traditional data-analysis tools. Big datasets’ important characteristics such as insights, trends, or significant differences become clear only when analyzing the entire dataset. Often used in analytics or machine learning.

• See also: dataset, machine learning

Chain-of-Thought Prompting

A method of prompting that encourages (or forces) an AI model to output the intermediate steps of a task, often improving the quality of the response.

• See also: prompt, one-shot prompting, multi-shot prompting, zero-shot prompting

Compute

A general term for the amount of computational power required to complete a given task.

Dataset

A collection of data, which could be documents, files, numbers, images, or any other type of data.

• See also: training dataset

Deep Learning

A type of neural-network architecture that contains multiple computational layers of and is often used for learning complex tasks.

• See also: machine learning, neural network

Fine-Tuning

Changes to a model that has already been trained, in order to affect its output.

Generative AI (genAI)

AI models that learn patterns existing in training data and generate examples that fit a particular pattern requested in the prompt. The training examples can be multimodal (for example, can include an image and a piece of text that describes it).

• See also: multimodal

Generative Pretrained Transformer (GPT)

A type of LLM that is built on a special type of deep-learning architecture called transformer architecture. This architecture enables fast training times and takes advantage of the information provided by the context in which words (or tokens) appear in the prompt and the training set.

• See also: LLM

Graphical Processing Unit (GPU)

A type of computer architecture that enables the fast, parallel computations required for graphics, video rendering, as well as training Artificial Intelligence models.

• See also: compute

Hallucination

A result from an AI model (typically a language model) that is misleading or incorrect but confidently presented as truth.

• See also: language model

Language Model

A probabilistic model that generates and understands text by learning associations between text documents during training.

• See also: model

Large Language Model (LLM)

A language model that utilizes large amounts of text-based training data to build associations between different text snippets.

• See also: language model, neural networks, training data

Machine Learning

A category of algorithms that focus on identifying and incorporating trends from training data and making predictions for new data. Put another way, machine-learning algorithms effectively generalize from examples they are provided. They usually do not use explicit instructions, instead relying on probabilistic and inference solutions.

• See also: algorithms, neural networks

Model

An approximative representation of a natural phenomenon or concept. Two notable types of models are mathematical or software-based. Models are used to describe behavior or events.

Multimodal

Refers to multiple modalities, including text, audio, video, or other data types.

Multi-Shot Prompting (Few-Shot Prompting)

Prompts that provide several example outputs to guide a model in producing outputs with a particular structure. These are contrasted with one-shot or zero-shot prompts.

• See also: one-shot prompting, zero-shot prompting

Natural Language Processing (NLP)

A subfield of computer science dedicated to developing algorithms and systems that can understand and generate human language.

Neural Networks

A class of machine-learning algorithms that mimic aspects of biological nervous systems such as memory or directed connectivity. A neural network consists of one or more layers of interconnected computational nodes.

• See also: machine learning, deep-learning, algorithm, model

Overfitting

A model that has been trained so closely to its set of training data that it’s unable to meaningfully generalize to new instructions or test data.

• See also: training data, model

Parameter

A number that influences some aspect of a system.

Probabilistic Model

A model that works by looking at earlier parts of a response and then sampling a probability to generate the most likely piece of data to come next. A probabilistic model is nondeterministic: for a given input, it may predict different outcomes at different times. The model rolls a “set of dice” for each new piece of data it generates.

• See also: model

Prompt

An input (usually text) that a generative model uses to generate an output or response. Prompts can be in the form of questions, stories, code, pictures, or audio recordings, depending on the model being used. Prompts are used to fine-tune models to change their outputs.

Retrieval-Augmented Generation

An approach used by AI systems to remember information, using a secondary system of storage and retrieval.

Single-Shot Prompting

Using a prompt that contains a single example of the desired answer. With single-shot prompting, the model relies on its training data and the single example to provide a correct answer to the prompt.

• See also: prompt, multi-shot prompting, zero-shot prompting

Small Language Model (SLM)

A category of language models that boast smaller numbers parameters without severely impacting performance. SLMs are typically developed for deployment on local devices such as laptops or smartphones, or for contexts that require rapid response times.

• See also: language model, large language model

Supervised Learning

A method of model training with categorized data, such as question/answer pairs, or pictures with descriptive labels.

• See also: unsupervised learning, machine learning, model

Sycophancy

Instances in which an AI model adapts responses to align with the user’s opinion, even if the opinion is not objectively true. This behavior is generally undesirable.

Synthetic Data

Data that is usually used for training a model and is simulated or generated, as opposed to nonsynthetic data gathered from real sources.

Token

An atomic unit of data that a large input model will work with. Both input and training data are converted into tokens and then processed by the LLM. The model’s output will also be a set of tokens that are further decoded and presented in a human-understandable modality (e.g., text). For text data, a token could be a character, a syllable, or a word.

Training Data

Data used to inform changes to aspects of a model. For example, a labeled picture of a strawberry, which helps a model recognize a strawberry in the future.

Turing Test

A test developed in 1950 by Alan Turing to evaluate an artificial system’s ability to portray behavior indistinguishable from a human.

Unsupervised Learning

A method of model training that does not used label examples for idenitfying where patterns in training data. For example, an unsupervised learning algorithm might group potatoes with other root vegetables without being given an example of a root vegetable.

• See also: training data

Zero-Shot Prompting

Using a prompt that does not contain examples of the desired answer. With zero-shot prompting, the model relies on its training data to provide a correct answer to the prompt.

• See also: prompt, multi-shot prompting, one-shot prompting