2026.06.10 2026.06.10

AIF-C01 Domain 2 Complete Guide: Fundamentals of Generative AI (24%)

swiftwand

Domain 2, Fundamentals of Generative AI, carries 24% of the AIF-C01 exam — the second-largest weight. Together with Domain 3 it forms the generative-AI core that distinguishes this certification from earlier AWS foundational exams. This guide covers the vocabulary, the model types, the foundation-model lifecycle, the honest limits of generative AI, and the AWS stack you are expected to recognize.

The shape of Domain 2 — three task statements
Core concepts — tokens, embeddings, chunking
Types of generative models — LLM, foundation, multimodal, diffusion
The foundation-model lifecycle — the official seven stages
Strengths and limits — hallucination, non-determinism, interpretability
Model-selection factors and business-value metrics
The AWS generative-AI stack — Bedrock, Amazon Q, PartyRock, JumpStart
Why build on AWS — pricing model and security
Domain 2 on the workbench — applying it to making
Summary — the scoring strategy for Domain 2
References

忍者AdMax

The shape of Domain 2 — three task statements

The three task statements are: explain basic concepts of generative AI; understand the capabilities and limitations of generative AI for solving business problems; and describe AWS infrastructure and technologies for building generative-AI applications. The thread running through all three is judgment about fit and trade-offs, not implementation detail.

Core concepts — tokens, embeddings, chunking

A token is the unit a model reads and generates — roughly a word fragment; pricing and context limits are measured in tokens. An embedding is a numeric vector that captures meaning, so that similar concepts sit close together in vector space. Chunking splits a long document into passages before embedding, which is the first step of retrieval. Learn the chain document to chunking to embedding to vector — ordering questions test exactly this flow.

Types of generative models — LLM, foundation, multimodal, diffusion

Large language models (LLMs) generate and understand text.
Foundation models are large models pre-trained on broad data and adaptable to many tasks; LLMs are a subset.
Multimodal models handle more than one modality — text plus images or audio.
Diffusion models generate images by iteratively denoising, the basis of most image generators.

Match the model type to the use case — a matching-question favorite. Image generation points to diffusion; a document chatbot points to an LLM; a system that reads a photo and answers questions points to multimodal.

The foundation-model lifecycle — the official seven stages

The exam guide frames the foundation-model lifecycle in seven stages: data selection, pre-training, optimization (such as fine-tuning), evaluation, deployment, feedback, and continuous improvement. Do not confuse this seven-stage FM lifecycle with the nine-stage ML lifecycle from Domain 1 — questions deliberately mix them.

Strengths and limits — hallucination, non-determinism, interpretability

Generative AI excels at drafting, summarizing, translating, coding, and ideation. Its limits are testable and worth memorizing as four points: hallucination (confident but false output), inaccuracy on facts, non-determinism (the same prompt can yield different answers), and limited interpretability (you cannot fully explain why a given output was produced). These limits drive design choices such as RAG and human review in later domains.

Model-selection factors and business-value metrics

Choosing a model balances capability, latency, cost, context-window size, modality, and customizability. On the business side, the exam expects value framing: time saved, conversion lift, cost avoided, and customer satisfaction. A larger, pricier model is not automatically better — the right answer is usually the smallest model that meets the requirement.

The AWS generative-AI stack — Bedrock, Amazon Q, PartyRock, JumpStart

Service	Role
Amazon Bedrock	Managed access to foundation models from multiple providers via one API
Amazon Q Developer	Coding assistant — free tier plus Pro at $19/user/month
Amazon Q Business	Enterprise assistant over your documents — Lite $3, Pro $20 per user/month
PartyRock	No-code playground for building generative-AI apps
SageMaker JumpStart	Hub of pre-trained models and solution templates

Why build on AWS — pricing model and security

The exam frames AWS advantages as consumption-based pricing (pay per token or per hour, no upfront model training), a wide choice of models through Bedrock, and built-in security: your data is not used to train the base models, and it stays within your AWS environment with IAM and encryption. These points answer “why AWS” questions directly.

Domain 2 on the workbench — applying it to making

For a solo maker, Domain 2 is the most immediately useful section. Choosing a smaller model for a product-description generator, using embeddings to search your own build logs, or recognizing when non-determinism will frustrate users — these are everyday decisions. The exam vocabulary simply gives you a precise way to reason about them.

Summary — the scoring strategy for Domain 2

Domain 2 (24%) rewards clean vocabulary: tokens, embeddings, and chunking; the four model types; the seven-stage FM lifecycle; the four limits; and the AWS stack. Lock these in and you carry momentum into the heaviest domain, Domain 3.