AIF-C01 cheat sheet

Fundamentals of AI and ML

AI vs ML vs Deep Learning: Core AIF-C01 Concepts and Terminology

• AI contains ML, ML contains deep learning, and generative AI is built on deep learning; agentic AI is GenAI that plans and acts using tools.
• An algorithm is the learning method; a model is the trained result. Training builds the model; inferencing uses it on new data.
• Supervised learning needs labeled data; unsupervised learning finds structure in unlabeled data; reinforcement learning learns by rewards through trial and error.
• Structured versus unstructured is about format (tabular and time-series versus image and text); labeled versus unlabeled is about whether answers are provided.
• Batch = many records offline; real-time = instant single answer; asynchronous = large or slow jobs queued; serverless = auto-scaling with no servers to manage.
• Overfitting means great on training data but poor on new data; underfitting means poor on both.
• Bias is a systematic skew in outputs; fairness is the goal of equitable treatment across groups.
• Computer vision handles images, NLP handles language, and an LLM is a large neural network that generates text.

Practical AI Use Cases and the AWS AI Service Catalog (AIF-C01)

• AI/ML adds value when a task needs prediction or pattern-finding at scale, assists human decisions, or automates repetitive work.
• AI/ML is the wrong tool when a guaranteed exact answer is required rather than a probabilistic prediction, or when a simple rule already works.
• Regression predicts continuous numbers, classification assigns known categories, and clustering groups unlabeled items.
• Match the verb to the service: speech to text is Transcribe, text to speech is Polly, translation is Translate, NLP and sentiment is Comprehend, chatbots are Lex.
• Use pre-trained managed services for common tasks; use SageMaker AI to build, train, and deploy a custom model on your own data.
• Choose traditional ML over a foundation model when explainability, regulation, or tight cost and latency constraints dominate.
• Choose a foundation model when the task is generating content or broad, flexible language understanding.
• Real-world mapping: images to computer vision, audio to speech recognition, suggestions to recommendation systems, anomalies to fraud detection.

The AI/ML Development Lifecycle & MLOps: Pipeline, Models & Metrics

• The AI/ML pipeline runs in order and loops: collect data, prepare it, train, evaluate and tune, deploy, monitor, then retrain.
• Three ways to source a foundation model: use a provider's managed FM (Amazon Bedrock), adapt an open-source pre-trained model, or train a custom model — moving from least to most control, cost, and effort.
• A managed API service means low operational overhead but less control; a self-hosted API gives full control but you run the infrastructure.
• Amazon SageMaker AI builds, trains, and deploys ML models; Amazon Bedrock serves foundation models via API; Amazon Q is an AI assistant; Kiro helps developers build software; Amazon S3 stores the data.
• Precision matters when false positives are costly; recall matters when you must catch every true case; F1 score balances the two; accuracy alone misleads on imbalanced data.
• Models decay over time as real-world data drifts, so monitoring detects the drop and retraining restores performance.
• Business metrics — cost per user, development cost, customer feedback, and ROI — decide whether a model is worth running, not just whether it is accurate.
• MLOps makes ML repeatable, automated, scalable, and production-ready while managing technical debt.

Fundamentals of GenAI

Generative AI Concepts: Tokens, Embeddings, and Foundation Models (AIF-C01)

• Generative AI creates new content; a foundation model (FM) is the broadly pre-trained base, and an LLM is a foundation model specialized for text.
• A token is the small unit of text a model reads and writes; chunking splits a large document into smaller pieces an application can handle.
• An embedding is a numerical representation of meaning expressed as a vector; similar meanings produce nearby vectors, which powers semantic search.
• Transformer-based LLMs handle text, multi-modal models handle several data types at once, and diffusion models generate images from noise.
• The FM lifecycle runs data selection, model selection, pre-training, fine-tuning, evaluation, deployment, then feedback; pre-training is the costly step that builds broad knowledge.
• Token-based pricing charges for input plus output tokens, so longer prompts and longer answers cost more.
• Context engineering means giving the model the right information at the right time, balancing answer quality against token cost.
• An AI agent adds tool usage, memory, and orchestration to a foundation model; the Model Context Protocol (MCP) connects agents to external systems.

Generative AI Capabilities and Limitations for Business (AIF-C01)

• GenAI's core advantages are adaptability, responsiveness, conversational capability, and the ability to generate original content.
• A hallucination is when a model states false or made-up information confidently; it is the most heavily tested GenAI limitation.
• Interpretability is limited because large models are black boxes, so you often cannot explain why they produced a given output.
• Nondeterminism means the same prompt can produce different answers on different runs, making behavior hard to test.
• Inaccuracy, plus data-privacy and cost concerns, round out the practical limitations to weigh.
• Selecting a GenAI model balances capabilities, performance, cost, latency, model size, constraints, and compliance; no model wins on every dimension.
• Business value is measured with metrics like ROI, efficiency, conversion rate, average revenue per user, accuracy, and customer lifetime value.
• Good GenAI judgment weighs benefits against risks rather than adopting or rejecting the technology outright.

AWS Infrastructure and Services for Building GenAI Applications

• Amazon Bedrock is a fully managed, serverless service for accessing foundation models from multiple providers through one API; it is the fast path to GenAI.
• Amazon SageMaker AI is for building, training, fine-tuning, and hosting your own models with full control; SageMaker JumpStart deploys pre-built foundation models inside it.
• Amazon Q is an AI assistant (Q Business for the business, Q Developer for developers), and Kiro is AI-assisted development.
• Strands Agents and Amazon Bedrock AgentCore are technologies for building and running AI agents; know them at definition level.
• Managed AWS GenAI services offer accessibility, a lower barrier to entry, efficiency, cost-effectiveness, and faster speed to market.
• AWS infrastructure adds security, compliance, responsibility, and safety benefits for GenAI workloads.
• Bedrock is priced per token; use on-demand for variable workloads and Provisioned Throughput for predictable, high-volume traffic at a lower cost per request.

Applications of Foundation Models

Foundation Model Design: Selection, RAG, and Customization (AIF-C01)

• Select a foundation model by matching the scenario's constraint to a criterion: cost, modality, latency, multi-lingual support, model size and complexity, customization, input/output length, or prompt caching.
• Inference parameters steer each response; lower temperature makes output more deterministic and factual, higher temperature makes it more creative, and max tokens caps answer length.
• Retrieval Augmented Generation (RAG) retrieves relevant information and adds it to the prompt so the model answers from current or private data without retraining, reducing hallucination.
• Amazon Bedrock Knowledge Bases is the managed RAG service on AWS; embeddings can be stored in Amazon OpenSearch Service, Amazon Aurora, Amazon Neptune, or Amazon RDS for PostgreSQL.
• Customization runs cheapest to most expensive: prompt engineering / in-context learning, then RAG, then fine-tuning, then continuous pre-training, then pre-training from scratch.
• Model distillation creates a smaller, cheaper, faster model from a larger one to reduce inference cost.
• An AI agent is a foundation model plus tools plus orchestration that takes actions to complete multi-step tasks, unlike a model that only returns text.

Prompt Engineering Techniques: Zero-Shot to Chain-of-Thought (AIF-C01)

• Prompt engineering shapes a model's output by improving the prompt, with no training, making it the cheapest way to change model behavior.
• The core prompt constructs are instruction (the task), context (background), input data, examples, and negative prompts (what to avoid).
• Techniques differ by example count: zero-shot uses no examples, single-shot uses one, and few-shot uses several to teach a pattern.
• Chain-of-thought prompting asks the model to reason step by step, improving accuracy on complex, multi-step problems.
• Prompt templates are reusable structures with placeholders that keep prompting consistent and repeatable across many requests.
• Best practices are specificity, clear instructions, concision, providing context and examples, and iterating; they most improve response quality.
• Prompting risks include prompt injection (hidden malicious instructions), jailbreaking (bypassing safety rules), hijacking, poisoning, and exposure of sensitive data.
• Amazon Bedrock Prompt Management stores, versions, and reuses prompts so teams can test changes and roll back to known-good versions.

Training and Fine-Tuning Foundation Models

• Pre-training builds broad knowledge from scratch and is the most expensive step; you usually consume a provider's pre-trained model rather than create one.
• Fine-tuning adapts a pre-trained model's behavior to your task by updating its weights using labeled examples.
• Continuous pre-training extends an existing model with new unlabeled domain data; distillation shrinks a large model into a smaller, cheaper one.
• Fine-tuning methods include instruction tuning (follow instructions), domain adaptation (specialize for a field), and transfer learning (reuse learned features).
• Good fine-tuning data must be curated, governed, correctly labeled, adequately sized, and representative, because unrepresentative data creates bias.
• RLHF aligns a model by having humans rank outputs, training a reward model that steers the FM toward preferred responses.
• Only fine-tuning changes the model; RAG adds external knowledge at request time without retraining, and prompt engineering changes nothing in the model.

Evaluating Foundation Model Performance

• The three evaluation approaches are human evaluation for subjective quality, benchmark datasets for standardized comparison, and Amazon Bedrock Model Evaluation as a managed service offering automatic and human evaluation.
• ROUGE measures summarization quality by overlap with a reference summary.
• BLEU measures machine-translation quality against reference translations.
• BERTScore measures semantic similarity of meaning rather than exact word overlap, and LLM-as-a-judge uses a model to score another model's outputs at scale.
• Use human evaluation when quality is subjective, such as a chatbot's helpfulness, where no single correct answer exists.
• Beyond metrics, judge whether the FM meets business objectives like productivity, user engagement, and task engineering.
• Evaluate whole applications: RAG by retrieval relevance plus answer quality, agents by task completion, and workflows end to end, using alignment metrics such as task completion rate, user satisfaction, and cost per interaction.

Guidelines for Responsible AI

Responsible AI: Bias, Fairness, and Amazon Bedrock Guardrails

• The features of responsible AI are bias, fairness, inclusivity, robustness, safety, and veracity — bias is the flaw, fairness the goal, and veracity (truthfulness) the feature generative AI most threatens.
• Amazon Bedrock Guardrails is the managed safety layer for generative AI: content filters, denied topics, word filters, PII redaction, and contextual grounding checks block harmful content and personal data.
• Responsible model selection includes sustainability — bigger models use more energy, so right-size the model and reuse pretrained ones.
• Legal risks of generative AI include intellectual-property infringement claims, biased outputs, hallucinations, end-user risk, and loss of customer trust.
• Bias enters through data: unbalanced, non-diverse datasets create biased models, while curated, inclusive, diverse, and balanced data is the fix.
• Bias underfits (model too simple) and variance overfits (model too tuned); their errors concentrate on underrepresented demographic groups, making them a fairness problem.
• To detect and monitor: SageMaker Clarify detects bias, SageMaker Model Monitor watches production drift, and Amazon A2I routes low-confidence predictions to humans, alongside label-quality analysis, human audits, and subgroup analysis.

Transparent and Explainable AI Models: Model Cards and Tradeoffs

• Transparency is understanding how a model works and was built; explainability is understanding why it produced a specific output.
• Interpretable models (decision trees, linear and logistic regression) show their reasoning, while black-box models (deep neural networks, large language models) are accurate but opaque — the black-box problem.
• High-stakes, regulated decisions such as loans, hiring, and medical use need explainability; low-stakes ones such as movie recommendations can tolerate a black box.
• Amazon SageMaker Model Cards document a model's purpose, training data, performance, and limitations — the model-documentation answer.
• SageMaker Clarify provides explainability through feature attribution as well as bias detection; Amazon Bedrock Model Evaluation compares foundation models; open-source models, data, and licensing give transparency through openness.
• The most accurate model is often the least explainable — the interpretability-versus-performance tradeoff — and responsible AI picks the right balance for the stakes.
• Human-centered design keeps people in control: tell users AI is used, collect feedback, and let people understand and contest decisions.

Security, Compliance, and Governance for AI Solutions

Securing AI Systems on AWS: IAM, Encryption, Guardrails, and Grounding

• Under the AWS shared responsibility model, AWS secures the cloud infrastructure while you secure your data, access, and configuration — that split does not change for AI workloads.
• Use IAM roles with least privilege so each AI component gets only the permissions it needs and receives temporary credentials instead of long-lived keys.
• Encrypt AI data both at rest (with AWS KMS) and in transit (with TLS); a complete answer always covers both states.
• Amazon Macie finds sensitive data such as PII in training data, and AWS PrivateLink keeps traffic to AI services off the public internet.
• Prompt injection, data leakage, and toxicity are AI-specific risks — defend with output filtering and validation, data-leakage prevention, and audit logging of every interaction.
• Document data origins with data lineage, data cataloging, and Amazon SageMaker Model Cards, and follow secure data-engineering practices like data quality, access control, and integrity.
• Reduce hallucinations by grounding answers in verified sources with RAG, validating output, and using confidence scoring; Amazon Bedrock Guardrails filters unsafe responses.

AI Governance and Compliance on AWS: Config, CloudTrail, and Audit Manager

• Governance and compliance prove an AI system is run responsibly and meets the rules — think of the chain regulation to control to evidence, run continuously.
• AWS Config tracks resource configuration and compliance over time, while AWS CloudTrail records the audit trail of who called which API and when.
• AWS Audit Manager collects evidence that you meet a compliance framework; AWS Artifact lets you download AWS's own compliance reports such as SOC 2 and ISO.
• Amazon Inspector scans workloads for vulnerabilities, and AWS Trusted Advisor checks your account against AWS best practices.
• Data-governance strategies cover data lifecycles, retention, residency (where data lives), logging, and monitoring — policy decisions you set and then enforce.
• Governance processes include written policies, a regular review cadence, transparency standards, and team training requirements.
• The AWS Generative AI Security Scoping Matrix is a framework for scoping generative AI security by how you use the model; responsibility grows from using a public app up to training your own model.