AIF key terms

The broad field of building computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognizing images, or making decisions. AI is the umbrella discipline that contains machine learning, deep learning, and generative AI.

A branch of AI where systems learn patterns from data instead of following hand-written rules. The model is trained on examples, then uses what it learned to make predictions or decisions on new, unseen inputs.

A machine learning approach that stacks many layers of simple computational units into a neural network. Each layer learns progressively more abstract features, which makes deep learning effective for images, audio, and natural language tasks.

A training method where each example in the dataset includes the correct answer, called a label. The model learns to map inputs to those labels, so it can later predict labels for new data such as spam or not spam.

A training method that uses data without correct-answer labels. The model looks for structure on its own, for example grouping similar customers together or detecting unusual patterns that differ from the norm.

A training approach where an agent learns by taking actions in an environment and receiving rewards or penalties. Over many attempts it discovers the sequence of actions that maximizes cumulative reward.

The collection of examples used to teach a model its patterns and relationships. The quality, volume, and representativeness of training data strongly determine how accurate and fair the resulting model will be.

The stage where a trained model is used to produce an output, such as a prediction, a classification, or a generated sentence. Inference happens after training and is what end users actually experience in an application.

A failure mode where a model memorizes the specific details of its training data, including noise, and performs poorly on new data. It typically appears as very high training accuracy but weak real-world results.

A managed AWS service for building, training, tuning, and deploying machine learning models at scale. It provides notebooks, built-in algorithms, automatic model tuning, hosting endpoints, and monitoring tools in one environment.

A very large model pretrained on broad, general data that can be adapted to many downstream tasks without being built from scratch. Examples include large language models for text and diffusion models for images.

A foundation model trained on massive amounts of text to predict the next token. It can summarize, translate, answer questions, write code, and hold a conversation when given a suitable prompt.

The neural network architecture behind most modern language models. Its attention mechanism lets the model weigh how relevant every part of the input is to every other part, enabling strong handling of long-range context.

The small unit of text a model reads and generates, roughly a word fragment. Prompts, outputs, and context limits are all measured in tokens, and token counts drive usage costs and length constraints.

A numeric vector that represents the meaning of text, an image, or another item so that computers can compare them mathematically. Items with similar meaning end up with vectors that are close together in vector space.

The input text or instruction a user gives to a generative model to guide its output. A prompt can include a task description, context, examples, and formatting requirements that shape the response.

The practice of designing and refining prompts so a foundation model produces useful, accurate, and well-formatted output. Techniques include clear instructions, role framing, examples, step-by-step reasoning requests, and output format specifications.

A prompting technique that includes a small number of worked examples inside the prompt so the model can imitate the desired pattern. Zero-shot prompting gives instructions only, with no examples at all.

The maximum amount of text, measured in tokens, that a model can consider at one time when generating a response. Content beyond that limit must be shortened, summarized, or retrieved in stages.

A confident but factually wrong or invented output from a generative model, such as a fake citation or nonexistent policy. It happens because models predict plausible text rather than verify facts against a trusted source.

A generative model that learns to create images or other media by starting from random noise and progressively removing noise until a coherent result appears. It is the basis of many text-to-image systems.

Further training of an existing foundation model on your own labeled examples so it better fits a specific task, tone, or domain. It adjusts the model's weights and requires curated data plus compute resources.

A pattern that retrieves relevant documents from a knowledge source and inserts them into the prompt so the model answers using grounded facts. It reduces hallucination and keeps responses current without retraining the model.

A fully managed AWS service that provides access to foundation models from multiple providers through a single API. It supports customization, knowledge retrieval, agents, and safety guardrails without managing underlying infrastructure.

A managed capability that connects a foundation model to your organization's documents for retrieval-augmented generation. It handles ingesting, chunking, embedding, storing, and retrieving relevant passages automatically.

A feature that lets a foundation model plan and carry out multi-step tasks by calling approved APIs and company systems. The agent breaks a request into actions, executes them, and returns a final answer.

A family of AWS generative AI assistants, including one for business data and one for software developers. They answer questions, summarize content, generate code, and help troubleshoot using connected enterprise systems.

A data store optimized for saving embeddings and finding the vectors most similar to a query vector. It powers semantic search and retrieval steps in retrieval-augmented generation applications.

Splitting long documents into smaller overlapping pieces before generating embeddings for retrieval. Good chunk size and overlap balance context completeness against retrieval precision, directly affecting answer quality.

A foundation model that accepts and produces more than one type of data, such as text plus images, audio, or video. Multimodal models can describe a photo, answer questions about a chart, or generate an image from a sentence.

The practice of designing, building, and operating AI systems so they are fair, safe, transparent, privacy-preserving, and accountable. It means considering human impact and societal risk across the entire lifecycle of a model.

A systematic skew in model outputs caused by unbalanced or unrepresentative data, flawed labeling, or design choices. It can lead to unfair treatment of groups and must be measured and mitigated deliberately.

The degree to which humans can understand why an AI system produced a particular output. Higher explainability supports trust, debugging, regulatory review, and appeals when a model's decision affects a person.

A design pattern that places a person in the workflow to review, approve, or correct AI output before it takes effect. It is essential for high-stakes decisions where errors carry significant consequences.

A tool that detects statistical bias in data and models and explains how individual features influenced predictions. It produces reports that help teams document and reduce unfair or opaque model behavior.

Configurable safety controls that filter or block harmful content, denied topics, sensitive information, and prompt attacks in generative AI applications. They apply consistently across supported foundation models without changing the model itself.

The division of security duties between AWS and the customer. AWS secures the cloud, including facilities, hardware, and managed service infrastructure, while the customer secures what they put in the cloud, such as data, identities, and configurations.

Protecting data both while it is stored on disk and while it moves across networks. At-rest encryption protects saved data; in-transit encryption such as TLS protects data during transmission between clients and services.

A managed service for creating and controlling the encryption keys used to protect data across AWS. It centralizes key policies, rotation, and audit logging so teams can enforce who may decrypt what.

The AWS system of users, groups, roles, and policies that determines who can call which services and resources. Least-privilege policies are the primary control for limiting what an AI application or its users can access.

The requirement that data, including prompts and model outputs, stays within specified geographic boundaries. Teams meet it by choosing AWS Regions and configuring services so data is not replicated or processed elsewhere.

A self-service portal that provides AWS compliance reports and agreements on demand. It helps auditors and customers verify that AWS controls meet standards such as ISO, SOC, and PCI requirements.