AWS Certified AI Practitioner (AIF-C01) Vocabulary Flashcards

Comprehensive vocabulary flashcards covering the core topics, AWS services, and key concepts for the AWS Certified AI Practitioner (AIF-C01) exam based on the lecture transcript.

Agentic AI

AI that plans and carries out multi-step tasks by using tools.

Supervised Learning

A machine learning approach where every example comes with the correct answer or label, such as predicting house prices from past sales that list their prices.

Unsupervised Learning

A machine learning approach that works on data with no answers or labels, requiring the model to find natural patterns or groupings itself.

Clustering

A kind of unsupervised learning technique used to find natural groups in unlabeled data, with K-means as a common example.

Classification

A task where the model predicts a category, such as deciding whether an email is spam or not spam.

Regression

A task where the model predicts a numerical value, such as a house price.

Reinforcement Learning

A learning method where an agent takes actions in an environment and learns to maximize its total reward.

RLHF (Reinforcement Learning from Human Feedback)

A process where human evaluators rank a model's answers, and those rankings are used to train and improve the model.

Transformer

A model architecture that reads an entire sequence at once using attention mechanisms instead of processing one word at a time; LLMs are built on transformers.

Algorithm vs. Model

The algorithm is the procedure used to train, whereas the model is the trained result that makes predictions.

Real-time Inference

An inference method using an endpoint that is always running to deliver instant predictions for individual user requests.

Batch Inference

An inference method used for scoring large datasets offline without an always-running endpoint.

Asynchronous Inference

An inference method where large requests wait in a queue; it handles payloads up to 1 GB and processing durations up to one hour.

Serverless Inference

An inference option that scales to zero during idle periods so you pay nothing, with the tradeoff that initial requests may experience a cold start.

Overfitting

A condition where a model learns the training data too closely, performing very well on training data but poorly on new data due to high variance.

Underfitting

A condition where a model suffers error from wrong assumptions (bias), missing real patterns and performing badly on both training and new data.

Confusion Matrix

A table that displays a classifier's true positives, false positives, true negatives, and false negatives.

Recall

A metric measuring how many real positives were caught, used when missed cases are critical to prevent (such as catching fraud).

Precision

A metric measuring the proportion of positive identifications that were correct, used when false alarms are costly (such as spam filters blocking real email).

F1 Score

The harmonic mean of precision and recall, ideal when class distributions are imbalanced and a single balanced number is needed.

MLOps

DevOps principles applied to machine learning, covering version control, automated project stages, CI/CD, continuous retraining, and continuous monitoring.

Amazon Textract

An AWS service that extracts text, forms, and tables directly from scanned documents like invoices.

Amazon Comprehend

An AWS NLP service that analyzes text to find sentiment, entities, key phrases, and topics, or learn custom categories.

Amazon Transcribe

An AWS service that converts speech into text and can redact personal information and detect toxic speech.

Amazon Polly

An AWS service that converts text into lifelike speech.

Amazon Rekognition

An AWS computer vision service used to identify objects, faces, text, or unsafe content in images and video.

Amazon Lex

An AWS service for building conversational chatbots that detects user intent, gathers details, and calls Lambda functions.

Trainium vs. Inferentia

Trainium chips (Trn1, Trn2, Trn3 instances) are specialized for model training, whereas Inferentia chips (Inf1, Inf2 instances) are built for running inference.

SageMaker JumpStart

A feature in Amazon SageMaker providing open-source pre-trained models that can be directly deployed.

Data Wrangler vs. Feature Store

Data Wrangler prepares and transforms data visually, while Feature Store serves as a central location to store and share the resulting features.

SageMaker Ground Truth

A SageMaker feature used to get training data labeled by human workers or collect human preference data.

SageMaker Clarify

A SageMaker feature designed to detect bias in datasets or models and explain predictions.

SageMaker Model Monitor

A SageMaker tool that alerts you when a deployed model's data quality or performance drifts.

SageMaker Model Cards

A SageMaker documentation feature used to record a model's intended use, risk rating, training details, evaluation results, and ethical considerations.

SageMaker Canvas

A SageMaker feature that allows users to build machine learning models without writing code.

Foundation Model vs. LLM

A foundation model is a large model pre-trained on broad unlabeled data capable of many tasks; an LLM is a foundation model specific to text.

Diffusion Model

A type of generative model that creates images by starting with random noise and progressively removing noise step by step.

Token

The basic unit of text that a model reads and bills for, which can represent a whole word or part of a word.

Context Window

The maximum number of tokens a model can consider at one time during processing.

Embedding

A numerical vector representing semantic meaning, allowing items with similar meanings to be located near each other for semantic search.

Chunking

The practice of splitting documents into smaller fragments prior to generating embeddings for RAG.

Vector Database

A specialized store designed to locate embeddings closest in distance to a given search query.

Prompt Engineering vs. Context Engineering

Prompt engineering determines what to ask the model, while context engineering designs the overall system and information (instructions, tools, memory, retrieved documents) provided to the model.

Model Context Protocol (MCP)

An open standard designed to connect AI applications to external data, tools, and workflows, functioning like a USB-C port for AI apps.

Amazon Bedrock

A managed AWS service serving multiple foundation models through a single API under pay-per-use pricing.

Amazon Nova Models

Amazon's foundation model family comprising text models (Premier, Pro, Lite, Micro) and specialized media models (Nova Canvas for images, Nova Reel for video, Nova Sonic for speech).

Amazon Bedrock AgentCore

A managed AWS platform used to run, secure, and monitor agents created with any framework and model.

Strands Agents

An open-source SDK for writing agents in Python or TypeScript with built-in MCP and multi-agent patterns.

Kiro

An AI coding tool for developers that turns prompts into specifications, tasks, and code (formerly known as Amazon Q Developer).

AWS Transform

An AWS service that uses AI agents to automate the migration or modernization of mainframe, VMware, or .NET workloads.

Provisioned Throughput

A Bedrock billing option that reserves model capacity billed by the hour for steady, heavy traffic workloads.

Prompt Caching

A Bedrock capability that saves the repeated opening portion of a prompt to bypass reprocessing on future requests, lowering input costs and latency.

RAG (Retrieval-Augmented Generation)

A method that retrieves relevant snippets from private documents and appends them to a prompt, enabling the model to answer from current data without retraining.

Bedrock Knowledge Bases

A managed Bedrock feature that chunks, embeds, and stores documents in a vector database to perform automated RAG with source citations.

System Prompt

A foundational prompt that defines how an AI model must behave across every turn of a conversation.

Chain-of-Thought Prompting

A technique that asks the model to reason step by step to solve complex, multi-step problems.

Data Poisoning vs. Prompt Injection

Poisoning injects malicious data into the training set, whereas prompt injection (hijacking) hides malicious instructions inside the input prompt.

Exposure vs. Prompt Leaking

Exposure occurs when a model reveals sensitive user/training data, while prompt leaking occurs when a model exposes its own system instructions.

Jailbreaking

The act of crafting inputs to bypass a model's safety and policy rules.

Bedrock Prompt Management

A Bedrock feature to create, version, test, and reuse prompts containing variables across different applications.

Continued Pre-training

Training an existing model further on large amounts of unlabeled domain-specific text to teach it domain knowledge.

Instruction Tuning

Training a model on labeled prompt-and-answer pairs to teach it how to follow specific tasks.

Model Distillation

A technique where a large teacher model answers prompts, and those outputs are used to train a smaller, cheaper student model.

ROUGE vs. BLEU

ROUGE measures word overlap as recall to evaluate text summaries; BLEU measures word overlap as precision and penalizes short outputs to evaluate translations.

BERTScore

An evaluation metric that compares text embeddings using cosine similarity to check if two texts share the same meaning.

Amazon Bedrock Guardrails

A security feature offering content filters, denied topics, word filters, PII masking, contextual grounding checks, and automated reasoning checks.

Contextual Grounding Check

A Guardrails feature that verifies whether a response is factually supported by the source text (grounding) and directly answers the user query (relevance).

Shared Responsibility Model

An AWS security framework where AWS secures the cloud infrastructure, and the customer secures data, encryption, and access controls within the cloud.

Amazon Macie

An AWS security service that uses pattern matching and machine learning to discover PII in Amazon S3 buckets.

Amazon Inspector

An AWS security scanner that checks EC2 instances, container images, and Lambda functions for software vulnerabilities and network exposure.

Interface VPC Endpoint

An endpoint powered by AWS PrivateLink that allows private access to services like Bedrock and SageMaker without crossing the public internet.

AWS CloudTrail

An AWS governance service that records API calls, detailing who requested an action, when, and whether it was authorized.

Model Invocation Logging

A Bedrock setting that saves full prompt inputs and generated model outputs to Amazon CloudWatch or Amazon S3.

AWS Config

An AWS service that tracks changes to resource settings over time, evaluates configurations against compliance rules, and sends alerts.

AWS Trusted Advisor

An AWS tool providing best-practice checks across cost, performance, security, fault tolerance, service limits, and operational excellence.

AWS Artifact

A central portal to download AWS compliance reports (SOC, PCI, ISO) and accept legal agreements like the HIPAA Business Associate Addendum.