Free AWS AIF-C01 Exam Flashcards

AIF-C01 is the exam code for AWS Certified AI Practitioner, a Foundational AWS Certification for people who understand AI/ML solutions on AWS without necessarily building them.

The exam has 65 total questions and a 90-minute time limit. AWS states that 50 questions are scored and 15 are unscored, but unscored questions are not identified during the exam.

AWS reports scores on a 100-1000 scale, with 700 as the minimum passing score. AWS does not publish an official pass-rate percentage for this exam in the local official-source notes.

Artificial intelligence is the broader capability of making systems perform tasks that seem intelligent. Machine learning is one way to do that by learning patterns from data.

Deep learning uses layered neural networks to learn complex patterns. It commonly supports computer vision, speech recognition, NLP, and many foundation model behaviors.

Training creates or adapts a model from data. Inference uses an already trained model to produce a prediction, classification, score, recommendation, text, or image.

Supervised learning uses labeled examples where the desired answer is known. It is common for classification, such as spam detection, and regression, such as price prediction.

Unsupervised learning looks for patterns in unlabeled data. Typical uses include clustering customers into segments or finding unusual records for further review.

Reinforcement learning trains an agent through rewards and penalties. For AIF-C01, know the concept, but most practitioner decisions focus on more common business AI patterns.

A no-AI approach can be best when rules are deterministic, data is poor, value is weak, risk is unmanaged, or a simple workflow already solves the problem reliably.

Classification predicts a category, such as approved or denied, fraudulent or legitimate, or urgent or non-urgent. The output is a class label.

Regression predicts a continuous number, such as demand, price, risk score, wait time, or quantity. The output is numeric rather than a category.

Forecasting predicts future values from time-series patterns. It is useful for demand, inventory, staffing, or capacity planning when historical time-based data is meaningful.

Recommendation systems rank items or actions for a user, such as products, content, or next-best offers. Good recommendations depend on relevant interaction and item data.

Model drift happens when live data, behavior, labels, or business conditions change enough that model performance degrades after deployment. Monitoring and retraining plans address it.

Overfitting means a model learned the training data too closely and fails to generalize to new data. Strong training results with weak test results are a common clue.

Accuracy can mislead when classes are imbalanced or errors have unequal cost. A fraud model can look accurate by mostly predicting no fraud while missing important cases.

Precision asks: of the items the model marked positive, how many were actually positive? It matters when false positives are costly or disruptive.

Recall asks: of all actual positive cases, how many did the model catch? It matters when missing a positive case is risky, such as fraud or safety detection.

A foundation model is a broad pretrained model that can be adapted to many tasks through prompting, retrieval, or customization instead of training from scratch for every task.

An LLM is a foundation model focused on language tasks such as drafting, summarizing, answering, transforming, classifying, and reasoning over text.

Tokens are the units a generative model reads and writes. Token count affects cost, latency, context size, and how much output the model can produce.

The context window is the space available for instructions, user input, retrieved content, and output. A larger window helps only when the added context is relevant and governed.

Embeddings convert text, images, or other content into numeric vectors that capture meaning. Similar vectors support semantic search, recommendations, and RAG retrieval.

A hallucination is a fluent or confident model response that is unsupported, incorrect, or fabricated. Grounding, evaluation, guardrails, and human review reduce the risk.

Prompt engineering designs the instructions, context, examples, constraints, and output format given to a model. It improves behavior without changing the model weights.

Zero-shot prompting gives instructions without examples. Few-shot prompting includes examples that show the desired pattern, tone, structure, or decision rule.

Temperature controls output randomness. Lower settings usually make responses more predictable; higher settings can increase variety but may reduce consistency.

Top-p controls sampling by limiting generation to a cumulative probability mass of likely tokens. Adjust it deliberately when balancing variety and predictability.

Maximum tokens limits how much output a generative model can produce. It helps control cost, latency, verbosity, and whether responses fit the application experience.

Amazon Bedrock is AWS's managed service for building generative AI applications with foundation models. It supports model access, customization paths, RAG, agents, guardrails, and evaluation patterns.

Retrieval-Augmented Generation retrieves relevant trusted content and gives it to a foundation model as context. It is useful for current, private, or source-cited knowledge.

Knowledge Bases for Amazon Bedrock helps implement RAG by connecting source documents, embeddings, vector storage, retrieval, and model responses in a managed workflow.

Use RAG when facts change often, source grounding is needed, or enterprise documents should stay outside model weights. Use fine-tuning for task behavior, style, or pattern adaptation.

Agents for Amazon Bedrock help a foundation model plan steps, use instructions, retrieve context, and invoke approved tools or APIs to complete a task.

Guardrails for Amazon Bedrock help apply configured controls for safety, denied topics, grounding, word filters, and sensitive information handling in supported GenAI workflows.

Amazon Q Business is a managed generative AI assistant for enterprise knowledge and work. It is different from building a custom Bedrock application from scratch.

Amazon Q Developer supports software development tasks such as code suggestions, code explanation, security-related checks, and developer assistance in supported tools.

Amazon SageMaker AI is the broader managed service family for building, training, deploying, and operating ML models when a team needs custom ML lifecycle capabilities.

SageMaker Canvas gives business users a visual, no-code way to explore and build ML models, while more technical SageMaker tools support deeper ML workflows.

Amazon Textract extracts printed text, handwriting, forms, and tables from documents. It fits document capture and extraction before later analysis or review.

Amazon Comprehend analyzes text for language, entities, sentiment, key phrases, topics, and related NLP insights. It works on text that has already been captured.

Amazon Transcribe converts speech to text, Amazon Polly converts text to speech, and Amazon Translate converts text between languages.

Amazon Rekognition analyzes images and video for visual content. Amazon Textract focuses on extracting structured text and data from documents.

Fairness means the AI system should avoid unjustified harm or discrimination across affected groups. It requires data review, evaluation, process design, and ongoing monitoring.

Transparency tells people how AI is being used and what limits apply. Explainability helps stakeholders understand why a prediction or recommendation occurred.

Human review is important when AI affects customers, regulated decisions, safety, finances, employment, health, legal outcomes, or brand risk. Reviewers need authority and escalation paths.

SageMaker Clarify supports ML bias detection and explainability. It can provide evidence for responsible AI review, but governance teams still set policy thresholds.

AWS secures the cloud infrastructure. Customers remain responsible for identities, data, prompts, application behavior, allowed use, logging choices, and business governance.

AI cost can come from tokens, model choice, endpoints, training jobs, vector storage, logs, data movement, and human review. Use budgets, tags, alerts, quotas, and right-sized models.