AIF study plan

Distinguish supervised, unsupervised, and reinforcement learning and match each to a concrete business use case

Explain the roles of training, validation, and test splits and identify what each detects (overfitting vs generalization)

Differentiate classification, regression, clustering, and ranking tasks and pick the correct task type for a stated problem

Compute and interpret precision, recall, F1, and accuracy on a small confusion matrix you construct yourself

Describe the AWS shared responsibility model as it applies to AI/ML workloads and separate customer duties from provider duties

Explain tokenization, embeddings, and positional encoding and describe how each affects model behavior and cost

Contrast foundation model architectures (transformer-based text, diffusion image, multimodal) and state which output type each produces

Distinguish pre-training, fine-tuning, instruction tuning, and RLHF by data needed and outcome achieved

Define temperature, top-p, and max tokens and predict how changing each alters determinism and length of output

Describe retrieval-augmented generation end to end, naming the role of the embedding model, vector store, retriever, and generator

Select an appropriate foundation model for a given latency, cost, modality, and context-window constraint and justify the choice

Design a prompt that constrains output format, supplies context, and includes guardrails against hallucination for a support-triage scenario

Choose between zero-shot, few-shot, and fine-tuning approaches based on data volume and task specialization

Map a stated workload to the correct AWS service (e.g., model hosting versus managed generative API versus vector search) with rationale

Identify when to use agents or tool-calling versus a single-turn prompt and state the added failure modes

Estimate inference cost and latency for a workload given token counts, request volume, and model tier

Design an evaluation approach combining automated metrics, human review, and task-specific rubrics

Diagnose common failure modes (hallucination, prompt injection via retrieved content, context overflow) and propose one mitigation per failure

Explain how prompt caching, batching, and smaller-model routing reduce cost without unacceptable quality loss

Match chunking and embedding strategies to document types (long policy PDFs, short FAQs, tabular records)

Integrate retrieval, guardrails, and evaluation into a single architecture diagram described in text only

Compare model customization options by effort, data requirement, and when each is not worth the cost

Identify observability signals (input/output logging, latency, token usage, quality sampling) needed to operate a generative workload

Re-derive ML fundamentals that underlie model selection, including bias-variance tradeoff and data quality effects

Classify a described AI system's risks across bias, toxicity, hallucination, privacy, and intellectual property

Select bias detection and explainability approaches appropriate to a stated model type and audience

Design human-in-the-loop review points and define escalation criteria for a high-stakes workflow

Apply transparency and disclosure practices, including when to label AI-generated content

Write an acceptable-use and model-card style summary for one scenario

Map data privacy obligations (classification, minimization, retention, residency) to concrete controls in an AI pipeline

Identify threats specific to generative systems, including prompt injection, data poisoning, model extraction, and insecure output handling, and pair each with a mitigation

Describe identity, access, and network controls applied to model endpoints and training data stores

Draft audit and monitoring requirements that produce evidence for governance review

Distinguish compliance-relevant documentation artifacts and state what each must contain

Rehearse timed full-length practice sets and hold pacing to the per-question target

Re-teach the 10 weakest sub-areas identified from prior quizzes using self-generated explanations and examples

Resolve every remaining flagged uncertainty by verifying against primary documentation

Build a one-page recall sheet of decision rules, thresholds, and service mappings