Generative AI Glossary of Terms

In your organization, not everyone involved in the adoption of generative AI will have an AI or technical background. And with new terms emerging as quickly as the technology evolves, there can be confusion or miscommunication.

This resource contains the top 40 terms most often used in conversations about generative AI, along with a short description of each. The glossary was created using the AWS PartyRock application with input and oversight from real humans at AWS. It contains terms that business leaders will need to know as they embark on their generative AI journey.

Generative AI Glossary of Terms

Adversarial Training

A technique in machine learning where models are trained on intentionally modified inputs that are designed to fool the model. By training on these adversarial examples, the model learns be more robust.

Accelerators

Hardware components like GPUs that can dramatically speed up the training and inference of generative AI models. By using accelerators, generative AI models are able to be trained on huge datasets and generate high quality outputs in reasonable amounts of time.

Algorithm

A set of step-by-step instructions that computers follow to solve a problem or complete a task. Algorithms allow computers to process data and calculate solutions without human intervention.

Artificial Intelligence

Computer systems that can perform tasks that typically require human intelligence, such as visual perception, speech recognition, and decision-making. Generative AI is a type of artificial intelligence that can create new content like text, images, audio, and video, rather than just analyzing data like other forms of AI.

Backpropagation

A method used in neural network training to calculate the error from model prediction (also known as forward propagation) and update the weights of neural networks to improve the model accuracy. Backpropagation enables networks to learn from mistakes.

Batch Size

The number of samples or instances that are processed together in one iteration during the training of a generative AI model. Using an appropriate batch size can help optimize the training process and model performance.

Bias

When the model exhibits prejudiced behavior or makes unfair assumptions due to limitations in training data or algorithm. This can lead to generative AI producing offensive, stereotypical, or otherwise problematic output (see “Hallucinations” below).

Classification

A common machine learning technique where algorithms are trained to classify data such as images or text into specific categories. Examples are object classification or sentiment analysis.

Computer Vision

The field of AI focused on enabling computers to identify, categorize, and understand visual inputs like images and videos through techniques like image recognition.

Deep Learning

A machine learning technique that uses neural networks with many layers to learn representations of data.

Explainability

The ability to explain how and why an AI model makes certain predictions or decisions. Important for evaluating a model’s fairness and to get transparency in the output.

Fine-Tuning

The process of taking a pre-trained generative AI model and further training it on a new dataset to improve its performance on a specific task or domain. It allows generative AI models to be adapted for new purposes without having to train a model from scratch.

Foundation Models (FMs)

Large artificial intelligence systems which include large language models and other types of models like computer vision and reinforcement learning models. FMs can be adapted and fine-tuned to perform a wide variety of tasks related to natural language, computer vision, and other areas. They serve as the base on which more specialized generative models are built.

Generative Adversarial Network (GAN)

A type of generative model that uses two neural networks competing against each other to generate new, realistic content such as images or new synthetic data sets.

Generative AI

Artificial intelligence systems that can create new content like images, videos, text, and audio. Generative AI models are trained on large datasets so they can generate high-quality, realistic outputs based on patterns they learned from the data.

Generative Model

AI systems that can create new content like images, audio, and text that are similar but not identical to their training data. Examples are systems that generate art, music, and text.

Graphics Processing Unit (GPU)

A hardware component specialized to efficiently handle complex mathematical operations running in parallel. Initially created to handle graphics rendering tasks in gaming and animation, their uses now extended far beyond that.

Hallucinations

When a generative AI model generates content that seems coherent but is actually incorrect or fabricated rather than being grounded in reality. Generative AI models can hallucinate plausible but false information if they are not properly trained on diverse, high-quality data, or not given quality structured inputs and prompts.

Headless

AI systems that can generate content like text, images, or video without needing a visual interface or output. Headless systems focus on the core generative capabilities of the AI model rather than building a complex visual front-end.

Hyperparameter

Settings that are used to control the training process of an AI model. Examples of hyperparameters include the number of layers in the neural network, the learning rate, and the batch size—these are values that are set prior to training and determine how the model learns from the data.

Inference

The ability of generative AI models to make logical deductions and derive new information based on what they have been trained on. Generative AI models use inference to generate new images, text, or other content that plausibly match a given text prompt, even if they have not seen that exact prompt before.

Large Language Model (LLM)

A type of AI system that is trained on massive amounts of text data to generate new text that mimics human writing and conversation. These large neural network models can produce remarkably human-like text and images when prompted.

Machine Learning

A type of artificial intelligence that allows computer systems to learn and improve from experience without being explicitly programmed. In the context of generative AI, machine learning enables systems to produce new, synthetic content like images, text, audio, and video that mimics authentic data but is algorithmically generated rather than copied from existing sources.

Model Migrations

The process of transferring or moving an AI model from one environment to another. For example, a company may need to migrate a large language model from their research and development cloud servers to their production cloud servers when getting ready to launch a new AI product or service.

Natural Language Processing (NLP)

A field of artificial intelligence that focuses on enabling computers to understand, interpret, and generate human language. In the context of generative AI, NLP techniques are used to allow AI systems to produce human-like text and speech outputs.

Neural Network

A type of machine learning model that is designed to mimic the way the human brain works. In generative AI, neural networks can be used to generate new content like images, text, or audio that are similar to the data the network was trained on.

Overfitting

Occurs when a generative AI model is trained too closely on a narrow dataset, causing it to memorize details instead of learning general patterns. This leads the model to perform very well on the data it has seen before but poorly on new data.

Parameters

The settings or configurations that control how a generative AI model behaves. By adjusting parameters, developers can influence the output and capabilities of generative AI systems.

Performant

Something works well and does what it's supposed to do quickly and efficiently. If a system or process is performant, it generates output of the desired quality or accuracy, operates with fast with minimal delays, lags, or errors.

Prompt Engineering

The practice of carefully designing and optimizing the prompts used in AI systems like chatbots and generative models. The goal of prompt engineering is to create prompts that result in more useful, accurate, coherent, and human-like responses from the AI system.

Reinforcement Learning

AI algorithms that learn how to make decisions by receiving rewards or penalties for the actions they take, similar to how humans learn through trial and error. An example of this is to provide feedback from LLM chatbot interactions to the model so that the model can learn the best responses. Used to develop systems that play games and control robots, among other applications.

Responsible AI

The process of developing and using artificial intelligence systems in a way that promotes fairness, transparency, privacy, and the mitigation of bias. The goal of responsible AI is to ensure AI systems are ethical, unbiased, safe, and beneficial to society.

Serverless

AI systems that can generate content like text, images, or video without requiring the user to set up or manage servers or infrastructure. This type of AI system runs on cloud infrastructure that removes the need for servers, allowing users to simply call an API to use the generative capabilities without managing any servers themselves.

Steerability

The ability to guide or control the output of an AI model. Users can influence the content generated to align it with their desired preferences or outcomes. For example, prompting a chatbot to only respond in short, concise sentences.

Supervised Learning

A machine learning approach where algorithms are trained on labeled datasets containing input data and the desired outputs. Used for classification and prediction tasks.

Synthetic Data

Artificial data that is generated by AI systems to mimic real-world data. Generative AI models can create synthetic data sets that have the same statistical properties as real data, allowing them to train on the synthetic data as if it were real.

Training Data

The data that is fed into a generative AI model during the training process. This data provides examples for the model to learn from so it can generate new content that resembles the patterns and characteristics of the training data.

Underfitting

When a machine learning model fails to identify the underlying relationships and patterns in the training data and does not perform well, even during training. Indicative of a model that is too simple, and in the context of a neural network, may need to have more layers or weights.

Unsupervised Learning

Training machine learning models to find hidden patterns and groupings within unlabeled data without human guidance. Used for tasks like customer segmentation, clustering of data sets, and to identify anomalies.

Use Case

The specific problem solved or area where generative AI is applied. These can include generating content, detecting fraud, improving customer experiences, or boosting employee productivity. Often use cases will vary by organization type, size, and industry.