AIF-C01 (exam guide version 1.4) covers five domains: AI and ML fundamentals (20%), generative AI fundamentals (24%), applications of foundation models (28%), responsible AI (14%) and security, compliance and governance for AI (14%). It is a foundational exam for people who use AI on AWS rather than build models, so coding, hyperparameter tuning and building pipelines are out of scope. Tick topics off as you cover them. Your ticks are saved in this browser.
This outline follows the official AIF-C01 exam guide . Check it for the exact wording.
Fundamentals of AI and ML 20% The vocabulary of AI and ML, learning types and data types, when ML is and isn't the right tool, the AWS managed AI services, the ML lifecycle and how models are judged.
Basic AI concepts and terms AI, ML, deep learning, neural networks, computer vision, NLP, model, algorithm and LLM Training versus inferencing; bias, fairness and fit (underfitting and overfitting) How AI, ML and deep learning relate and differ Batch and real-time inferencing (plus asynchronous and serverless options) Data types: labeled and unlabeled, tabular, time-series, image, text, structured and unstructured Supervised, unsupervised and reinforcement learning Practical use cases for AI Where AI/ML adds value: decision support, scale and automation When AI/ML is the wrong choice: poor cost-benefit, or when an exact outcome is needed instead of a prediction Choosing a technique: regression, classification, clustering Real-world applications: computer vision, NLP, speech, recommendations, fraud detection, forecasting What SageMaker, Transcribe, Translate, Comprehend, Lex and Polly (and Rekognition, Textract, Kendra, Personalize) do The ML development lifecycle Pipeline stages: data collection, EDA, pre-processing, feature engineering, training, tuning, evaluation, deployment, monitoring Model sources: open-source pre-trained models versus custom-trained models Using a model in production: managed API versus self-hosted API AWS services per stage: SageMaker, Data Wrangler, Feature Store, Model Monitor MLOps: experimentation, repeatability, scale, technical debt, production readiness, monitoring, retraining Model metrics (accuracy, AUC, F1) and business metrics (cost per user, ROI, customer feedback)
Fundamentals of generative AI 24% What generative AI is made of, what it's used for, where it fails, and which AWS services and pricing models apply.
Basic generative AI concepts Tokens, chunking, embeddings and vectors Prompt engineering, transformer-based LLMs, foundation models, multimodal and diffusion models Use cases: image, video and audio generation, summarisation, chatbots, translation, code, customer service agents, search, recommendations The foundation model lifecycle: data selection, model selection, pre-training, fine-tuning, evaluation, deployment, feedback Capabilities and limitations for business problems Advantages: adaptability, responsiveness, simplicity Disadvantages: hallucinations, interpretability, inaccuracy, nondeterminism Factors for choosing a model: type, performance, capabilities, constraints, compliance Business value and metrics: cross-domain performance, efficiency, conversion rate, ARPU, accuracy, customer lifetime value AWS infrastructure and technologies for generative AI SageMaker JumpStart, Amazon Bedrock, PartyRock and Amazon Q Advantages of AWS generative AI services: accessibility, low barrier to entry, efficiency, cost, speed to market, meeting business goals Benefits of AWS infrastructure: security, compliance, responsibility, safety Cost tradeoffs: responsiveness, availability, redundancy, performance, Regional coverage, token-based pricing, provisioned throughput, custom models
Applications of foundation models 28% Choosing and customising a foundation model, RAG and agents, prompt engineering, training and fine-tuning, and evaluating results.
Design considerations for foundation model applications Model selection criteria: cost, modality, latency, multilingual support, size, complexity, customisation, input and output length Inference parameters such as temperature and response length What RAG is and its business uses, including Bedrock Knowledge Bases Vector stores for embeddings: OpenSearch Service, Aurora, Neptune, DocumentDB, RDS for PostgreSQL Cost tradeoffs of customisation: pre-training, fine-tuning, in-context learning, RAG Agents for multi-step tasks (Agents for Amazon Bedrock) Prompt engineering Constructs: context, instruction, negative prompts, model latent space Techniques: chain-of-thought, zero-shot, single-shot, few-shot and prompt templates Benefits and best practices: quality, experimentation, guardrails, discovery, specificity and concision, multiple comments Risks and limits: exposure, poisoning, hijacking and jailbreaking Training and fine-tuning foundation models Key elements: pre-training, fine-tuning, continued pre-training Fine-tuning methods: instruction tuning, domain adaptation, transfer learning, continued pre-training Preparing fine-tuning data: curation, governance, size, labeling, representativeness, RLHF Evaluating foundation model performance
Guidelines for responsible AI 14% Building AI that is fair, safe, transparent and explainable, and the AWS tools that help.
Developing responsible AI systems Features of responsible AI: bias, fairness, inclusivity, robustness, safety, veracity Using tools such as Guardrails for Amazon Bedrock Responsible model selection: environmental impact and sustainability Legal risks of generative AI: IP claims, biased output, lost trust, end-user risk, hallucinations Dataset characteristics: inclusivity, diversity, curated sources, balance Effects of bias and variance: demographic impact, inaccuracy, overfitting, underfitting Detecting and monitoring bias and truthfulness: label quality, human audits, subgroup analysis, Clarify, Model Monitor, A2I Transparent and explainable models
Security, compliance and governance for AI 14% Securing AI workloads on AWS, documenting data origins, and meeting governance and compliance requirements.
Securing AI systems IAM roles, policies and permissions, encryption, Macie, PrivateLink and the shared responsibility model Source citation and documenting data origins: lineage, cataloguing, Model Cards Secure data engineering: data quality, privacy-enhancing technologies, access control, integrity Security and privacy for AI: application security, threat detection, vulnerability management, infrastructure protection, prompt injection, encryption at rest and in transit Governance and compliance for AI Compliance standards for AI: ISO, SOC, algorithm accountability laws AWS services for governance and compliance: Config, Inspector, Audit Manager, Artifact, CloudTrail, Trusted Advisor Data governance: lifecycles, logging, residency, monitoring, observation, retention Following governance protocols: policies, review cadence and strategies, the Generative AI Security Scoping Matrix, transparency standards, team training