AWS AI Practitioner Course Notes

AI and machine learning questions usually start by identifying the learning pattern. Supervised learning trains from labeled examples, unsupervised learning finds structure in unlabeled data, and reinforcement learning improves decisions through rewards from an environment. On the exam, a fraud label, churn label, or known target value points to supervised learning, while clustering similar records without labels points to unsupervised learning.

• Treating any labeled training dataset as clustering instead of supervised learning.

• Labeled outcomes point to supervised learning; unlabeled grouping points to unsupervised learning; reward-based decision improvement points to reinforcement learning.

A foundation model is a large pretrained model that can perform broad language, image, or multimodal tasks through prompting. Amazon Bedrock provides managed access to foundation models without requiring the customer to provision model infrastructure. The key exam distinction is whether the scenario needs to call an existing model, customize behavior with prompts and context, or train a separate custom model elsewhere.

• Using direct model invocation when the answer needs current or private documents retrieved through RAG.

• When an answer must use approved documents, think RAG, embeddings, retrieval, and grounding.

Amazon Bedrock Knowledge Bases provide a managed RAG path for foundation model applications. A knowledge base connects source data, embedding generation, vector storage, retrieval, and model generation so an application can answer from controlled content. The exam often tests this against direct model invocation, where the model only sees the prompt and has no managed retrieval layer.

• Choosing RetrieveAndGenerate when the application only needs retrieved passages from a knowledge base.

• Retrieve returns passages; RetrieveAndGenerate returns a model-generated answer based on retrieved passages.

AWS AI service selection depends on how much control the scenario requires. Amazon Bedrock is the managed path for building generative AI applications with foundation models, while Amazon SageMaker is the broader ML platform for custom model development, training, deployment, monitoring, and MLOps. A question that asks for model training code, custom algorithms, or full ML lifecycle control usually points away from Bedrock and toward SageMaker.

• Choosing SageMaker for every AI workload even when Bedrock already provides managed foundation model access.

• Bedrock is the managed foundation model service; SageMaker is the custom ML lifecycle platform.

Operating an AI solution starts with evaluation. Amazon Bedrock model evaluation helps compare model outputs against quality criteria, while human evaluation jobs bring reviewers into the process when judgment, safety, tone, or task accuracy cannot be trusted to automation alone. The exam may separate automatic evaluation from human evaluation, so watch for language about reviewers, work teams, or manual scoring.

• Reading CloudTrail as model quality monitoring instead of audit history for AWS API activity.

• Bedrock model evaluation and human evaluation jobs are for assessing model output quality and safety.

Responsible AI on AWS is tested through concrete controls, not slogans. Amazon Bedrock Guardrails can filter harmful content, block denied topics, and apply sensitive information filters so generative AI applications have runtime safety controls. These controls sit around model use; they do not replace model evaluation, identity permissions, or human review.

• Using Bedrock Guardrails for bias measurement when SageMaker Clarify is the service for bias and explainability analysis.

• Use Bedrock Guardrails for runtime content and sensitive information filtering around foundation model apps.