AWS Certified AI Practitioner (AIF-C01) Flashcards
Flashcards grouped by the 5 domains of revision 1.1. Question → Answer format: cover the answer, answer aloud, and check. Weights and syllabus may change: verify the official exam guide before taking the exam.
Flashcards grouped by the 5 domains of revision 1.1. Question → Answer format: cover the answer, answer aloud, and check. Weights and syllabus may change: verify the official exam guide before taking the exam.
Domain 1 — Fundamentals of AI and ML#
Q: What is the relationship between AI, ML, and deep learning? A: Deep learning ⊂ machine learning ⊂ artificial intelligence. DL uses deep neural networks; ML learns patterns from data; AI is the broader field.
Q: What distinguishes supervised learning from unsupervised learning? A: Supervised training uses labeled data (classification, regression); unsupervised finds structure in unlabeled data (clustering, dimensionality reduction).
Q: What is reinforcement learning? A: An agent learns through trial and error by interacting with an environment and maximizing a reward signal. It does not use a labeled dataset.
Q: Classification vs. regression? A: Classification predicts a category (spam/not spam); regression predicts a continuous value (price, demand).
Q: What is overfitting and how is it mitigated? A: The model memorizes the training data and generalizes poorly (high train accuracy, low test accuracy). Mitigation: more data, regularization, simpler models, early stopping.
Q: Why is accuracy misleading with imbalanced classes and what metrics should be used? A: A model that always predicts the majority class "gets it right" almost all the time. Use precision, recall, and F1; use recall if false negatives are costly (fraud, disease).
Q: Precision vs. recall? A: Precision: of what was marked positive, how much was actually positive. Recall: of all actual positives, how many were captured.
Q: Batch inference vs. real-time inference? A: Batch: periodic massive predictions, no urgency, cheaper. Real-time: always-active endpoint with immediate response, more expensive.
Q: What is feature engineering? A: Creating/transforming input variables from raw data so the model learns better (normalization, encoding, aggregations…).
Q: What does Amazon Textract do? A: Extracts text, tables, and key-value pairs from scanned documents (intelligent OCR). Documents, not general photos.
Q: What does Amazon Comprehend do? A: Managed NLP on text: sentiment, entities, key phrases, language, and PII detection.
Q: What does Amazon Rekognition do? A: Computer vision on images and video: objects, scenes, faces, text in image, and content moderation.
Q: Polly vs. Transcribe? A: Polly: text-to-speech (TTS). Transcribe: speech-to-text (STT). They are inverses.
Q: What is Amazon SageMaker AI in one sentence? A: The managed platform for building, training, evaluating, and deploying your own end-to-end ML models.
Q: When should you NOT use ML? A: When the problem can be solved with simple, stable deterministic rules (tax calculations, validations), or when there is insufficient data.
Domain 2 — Fundamentals of Generative AI#
Q: What is a foundation model? A: A large pre-trained model with massive data (usually self-supervised) that adapts to many tasks via prompting, RAG, or fine-tuning.
Q: What is a token and what is the context window? A: The token is the minimum unit of text the model processes (~subwords). The context window is the maximum number of input+output tokens per request.
Q: What are embeddings? A: Numerical vectors that represent the meaning of text (or images); similar texts end up close in the vector space → semantic search.
Q: What does temperature control? A: Sampling randomness: low → more deterministic/repeatable outputs; high → more diverse/creative outputs.
Q: Temperature vs. top_p vs. top_k? A: All three regulate sampling: temperature rescales probabilities; top_p limits to the cumulative probability nucleus; top_k limits to the k most probable tokens.
Q: What does max tokens control? A: Only the maximum length of the output. It does not affect creativity or quality.
Q: What is a hallucination? A: Invented but plausible content presented with confidence. Mitigation: RAG/grounding, low temperature, guardrails, human review.
Q: Order from lowest to highest cost to adapt an FM: A: Prompt engineering → RAG → fine-tuning → continued pre-training → training from scratch.
Q: Fine-tuning vs. continued pre-training? A: Fine-tuning: labeled data (prompt-response pairs) for task/style. Continued pre-training: unlabeled corpus to absorb domain language.
Q: Zero-shot vs. few-shot? A: Zero-shot: instructions only. Few-shot: solved examples included in the prompt (in-context learning, without changing weights).
Q: What is Amazon Bedrock? A: A serverless service providing unified API access to foundation models from Amazon and third parties (Anthropic, Meta, Mistral…), with customization, Agents, Knowledge Bases, and Guardrails.
Q: What is Amazon Q Business? A: Managed generative AI assistant for employees: connects to corporate data (S3, SharePoint, Salesforce…) while respecting permissions. Q Developer is its sibling for code/AWS.
Q: What is Amazon Quick? A: Managed AI workspace for chat, agents, BI/visualization, research, and automation on connected data and applications. Amazon Quick Sight is the business intelligence capability within Quick.
Q: What is Kiro in the AIF-C01 scope? A: An agentic IDE for specification-driven development, steering files, and hooks that automate quality controls.
Q: What are Strands Agents? A: Open-source, model-first SDK for building agents with tools, multiple providers, and protocols like MCP. It is code/framework, not managed hosting.
Q: What is Amazon Bedrock AgentCore? A: Model- and framework-agnostic infrastructure for deploying and operating agents: Runtime, Gateway, Memory, Identity, Observability, Evaluations, and Policy.
Q: Strands Agents vs AgentCore? A: Strands builds the agent logic; AgentCore can host and operate agents from Strands, LangGraph, CrewAI, or custom code with managed services.
Q: What does AgentCore Identity provide? A: Workload identity and secure credential management so agents and tools can access AWS or third-party services, with authentication, authorization, and auditable tracing.
Q: What is model distillation? A: Training a smaller model to mimic the behavior of a larger one; aims to preserve useful quality while reducing cost and latency.
Q: On-demand vs Provisioned Throughput in Bedrock? A: On-demand: pay-per-token, variable traffic. Provisioned: reserved capacity, stable cost, and guaranteed throughput; required for serving customized models.
Q: Which metric is associated with summaries and which with translation? A: ROUGE for summarization; BLEU for translation. Both compare against reference texts.
Q: What is the knowledge cutoff? A: The training data deadline: the model does not know facts after this date. This is compensated with RAG or tools.
Q: What is a multimodal model? A: One that accepts and/or generates multiple modalities (text, image, audio) in the same request.
Q: Prompt engineering vs context engineering? A: Prompt engineering designs instructions and examples. Context engineering also composes history, memory, RAG documents, tool results, and metadata within the window and token budget.
Q: What does MCP bring to an agentic system? A: A standard interface for discovering and using tools, resources, and context from external systems. MCP connects; orchestration decides the sequence, state, retries, and approvals.
Q: When does a multi-agent system make sense? A: When specialization, delegation, or review between agents adds more value than their cost, latency, and complexity. For a linear task, a single agent is usually better.
Domain 3 — Applications of Foundation Models#
Q: What is RAG and when to choose it? A: Retrieval-Augmented Generation: retrieve relevant chunks from your data and inject them into the prompt. Choose it when knowledge changes frequently or must be cited, without retraining.
Q: What does Bedrock Knowledge Bases do? A: Managed RAG: ingestion from S3, chunking, embeddings, vector store (e.g., OpenSearch Serverless), and retrieve/retrieve-and-generate APIs.
Q: What is chunking and why does it matter? A: Splitting documents before generating embeddings. Well-sized chunks → more accurate retrieval and context that fits within the model's window.
Q: Name two vector store options in AWS. A: Guide 1.1 cites Amazon OpenSearch Service, Aurora, Neptune, and Amazon RDS for PostgreSQL; pgvector is the usual mechanism in PostgreSQL.
Q: What does Bedrock Agents do? A: Orchestrates multi-step tasks: the model plans, calls APIs defined in action groups, queries knowledge bases, and returns the result.
Q: Knowledge Bases vs Agents: when to use each? A: Only answering questions about documents → Knowledge Bases. Executing actions against APIs (creating tickets, orders) → Agents (which may include a KB).
Q: What is chain-of-thought prompting? A: Asking for step-by-step reasoning before the final answer; improves logic and multi-step tasks.
Q: What does Amazon Bedrock Prompt Management do? A: Allows creating, testing, saving, versioning, and reusing prompts and variants. Each version must be evaluated with a stable dataset and retain a rollback path.
Q: What is prompt injection and how is it mitigated? A: User input that attempts to override system instructions ("ignore your instructions…"). Mitigation: guardrails, separating instructions from data, validating inputs, least privilege in tools.
Q: What does Bedrock Guardrails do? A: Filters at inference: denied topics, harmful content, PII (blocking or masking), word filters, and contextual grounding detection. Applies to input and output.
Q: What is Bedrock Model Evaluation? A: Comparing FMs with automatic metrics (accuracy, robustness, toxicity) or human evaluation on custom datasets, to choose a model.
Q: Criteria for choosing a foundation model: A: Modalities, quality on the task, context window size, cost per token, latency, languages, customization options, and license.
Q: When to choose fine-tuning over RAG? A: When you want to fix style/format/behavior with many labeled examples. RAG when the problem is knowledge that is updated or citable.
Q: What are stop sequences? A: Strings that, when generated, stop the output. Used to delimit formats and avoid extra text.
Q: Prompt best practices for RAG: A: Delimit the retrieved context, order "answer only with this context", ask to "say you don't know" if information is missing, and ask for citations.
Q: What happens to your data when customizing a model in Bedrock? A: They remain private: a copy of the model is created for your account, and your data is not used to improve the base models.
Q: What is right-sizing inference? A: Using the smallest model that meets the required quality: lower cost and latency. Do not use the largest one "just in case".
Domain 4 — Guidelines for Responsible AI#
Q: Typical dimensions of Responsible AI: A: Fairness, explainability, transparency, privacy and security, robustness, governance, controllability, and veracity.
Q: What is fairness? A: That the system does not systematically produce worse outcomes for protected groups (gender, age, ethnicity...).
Q: What is the most common source of bias in a model? A: Biased or unrepresentative training data: the model reproduces the discrimination present in the historical data.
Q: What does SageMaker Clarify do? A: Detects bias in data and models (pre- and post-training) and explains predictions with feature attributions (SHAP).
Q: Guardrails vs Clarify? A: Guardrails filters content during inference (topics, toxicity, PII). Clarify analyzes bias and explainability in the ML lifecycle. They do not compete; they complement each other.
Q: What is Amazon A2I (Augmented AI)? A: A service to insert human review into ML workflows, e.g., when prediction confidence falls below a threshold.
Q: What are SageMaker Model Cards? A: Structured model documentation: intended use, data, metrics, limitations, and risks. A governance tool.
Q: What are AWS AI Service Cards? A: Transparency documentation that AWS publishes about its own AI services: intended use cases, limitations, and best practices.
Q: Explainability vs Interpretability? A: Interpretable: the model is understandable by design (regression, trees). Explainable: the output of a complex model is justified post-hoc (SHAP, importances).
Q: What trade-off is usually present when requiring interpretability? A: The most interpretable models are usually simpler and sometimes less accurate; with regulation in place, interpretability may weigh more than the last point of accuracy.
Q: What is a representative dataset and why does it matter? A: One that reflects the diversity of the real population of use. It is the main lever to prevent bias before training.
Q: Measures against the risk of veracity in Gen AI for customers: A: Grounding/RAG with citations, low temperature, guardrails, and human review in high-impact decisions.
Q: What is AI transparency? A: Communicating what the system does, what data it uses, its limitations, and when the user is interacting with an AI.
Domain 5 — Security, Compliance, and Governance#
Q: What is the principle of least privilege? A: Granting only the essential IAM permissions for the task (e.g., invoking a specific Bedrock model and nothing else).
Q: IAM roles vs. access keys in applications? A: AWS applications must assume roles (temporary credentials), not carry embedded static access keys.
Q: What does AWS KMS do? A: Encryption key management. With customer managed keys, you control policy, rotation, and auditing of encryption at rest (S3, EBS, Bedrock...).
Q: What does AWS CloudTrail record? A: API calls in the account: who, what, when, and from where. It is the audit trail (e.g., Bedrock invocations).
Q: CloudTrail vs. CloudWatch? A: CloudTrail: API call auditing (who did what). CloudWatch: operational metrics, logs, and alarms (how the system is performing).
Q: What does Amazon Macie do? A: Discovers and classifies sensitive data (PII) in S3 using ML. Essential before using a data lake for training.
Q: Macie vs. GuardDuty vs. Inspector? A: Macie: PII in S3. GuardDuty: threat detection/malicious activity in the account. Inspector: vulnerabilities in workloads (EC2, ECR, Lambda).
Q: What are VPC endpoints (PrivateLink) for? A: Connecting the VPC to AWS services (e.g., Bedrock) without traversing the public internet: traffic stays within the AWS network.
Q: Shared responsibility model applied to Bedrock: A: AWS secures the infrastructure and the service ("of the cloud"); the customer secures IAM, their data, configuration, and use of outputs ("in the cloud").
Q: What is AWS Artifact? A: Self-service portal to download AWS compliance reports (SOC, ISO, PCI) and accept agreements.
Q: What does AWS Audit Manager do? A: Continuously collects compliance evidence and maps it to frameworks (GDPR, ISO, PCI) to prepare for audits.
Q: AWS Config vs. Audit Manager? A: Config continuously evaluates resource configuration against rules; Audit Manager collects evidence mapped to frameworks for audits.
Q: What is data residency and how is it respected with Bedrock? A: Requirement that data does not leave a specific region/jurisdiction. Use the service in the chosen region; inference and customization are processed there, and data does not improve base models.
Q: What is data lineage and why does it matter in AI governance? A: The trace of data origins and transformations leading to the model. It enables auditing of quality, permissions, and lifecycle compliance.
Q: AgentCore Identity vs Policy in AgentCore? A: Identity authenticates workloads/users and manages credentials; Policy authorizes which Gateway tool each principal can invoke and under what conditions via Cedar.