AIF-C01 Exam Domain Deep Dive: Fundamentals of Generative AI

TL;DR
- Two domains carry 52% of the score: Applications of Foundation Models (28%) and Fundamentals of Generative AI (24%). Everything else is supporting context.
- The format is fixed — 65 questions, 90 minutes, USD 100, pass mark 720/1000. That's 83 seconds per question, and unanswered questions score zero.
- Three weeks at ~1.5 hours a day is enough, but only if you spend one real hour calling Bedrock and reading the error messages instead of memorizing transformer diagrams.
The exam most people fail by 30 points, not by 300
The people who miss AIF-C01 usually don't miss it on attention-mechanism trivia. They miss it because they studied "AI" as a broad topic and then met questions about inference parameters, retrieval data flows, and IAM conditions on bedrock:InvokeModel.
I've watched ops engineers with three years of console time bounce off this exam. The AWS knowledge transfers fine; the vocabulary doesn't. The paper assumes you have sent at least one prompt through an API and looked at usage.input_tokens in the response body. Without that, Domain 3 questions read like noise.
This is a breadth exam. Going deep on one domain buys you nothing — the weights are published and the spread is deliberate.
The five published domains and what the weights imply
| # | Domain | Weight | Approx. questions of 65* |
|---|---|---|---|
| 1 | Fundamentals of AI and ML | 20% | ~13 |
| 2 | Fundamentals of Generative AI | 24% | ~16 |
| 3 | Applications of Foundation Models | 28% | ~18 |
| 4 | Guidelines for Responsible AI | 14% | ~9 |
| 5 | Security, Compliance, and Governance for AI Solutions | 14% | ~9 |
*My arithmetic from the published percentages; AWS doesn't publish per-domain question counts.
Domain 3 is the largest block and it's the practical one: RAG versus fine-tuning, agents, prompt engineering patterns, Bedrock features, cost and latency trade-offs. Domain 2 is the vocabulary layer underneath it. Together they're 52% of the paper, so an hour on those two is worth roughly two hours on governance.
Domain 1 covers supervised versus unsupervised, regression versus classification, precision/recall/F1, and — a favorite trap — when ML is the wrong answer and a simple rules engine is cheaper. Domain 4 is bias, fairness, explainability, human-in-the-loop review, and the AWS responsible-AI tooling. Domain 5 is where your day job pays off: IAM, KMS encryption, VPC endpoints, PrivateLink, data residency, and how the shared responsibility model shifts for foundation models.
Inside the 24%: what "Fundamentals of Generative AI" actually asks
The domain reads small on paper and then shows up in questions like "which parameter would you lower to make output more predictable" or "why does the same prompt return different summaries 20 minutes apart". Concrete sub-topics:
- Tokens and context windows. The model sees tokens, not characters. English runs roughly 4 characters per token; a 200K-token context is not a memory of your previous sessions.
- Embeddings. Vectors, cosine similarity, and why chunk size in a RAG pipeline changes answer quality more than model choice does.
- Model families. LLMs for text, diffusion for images, multimodal input/output — and the fact that generative models produce new content while classifiers pick from a fixed set.
- Prompt mechanics. Zero-shot versus few-shot, system versus user prompt, chain-of-thought as a latency-for-accuracy trade.
- Non-determinism. Same prompt, different output. Hallucination and grounding are consequences, not separate bugs.
Then the parameters, which is where easy points live:
| Parameter | What it does | Trap |
|---|---|---|
temperature |
Scales randomness in token selection | 0 reduces variance, it does not guarantee identical output |
top_p |
Nucleus sampling cutoff | Tune one of temperature/top_p, not both |
max_tokens |
Caps the response length | The answer gets truncated, not errored |
stop_sequences |
Halts generation on a string | Truncated output still bills as output tokens |
That temperature row is the one people argue about. In practice, setting temperature to 0 on Bedrock gets you close to repeatable output but not byte-identical — batching and floating-point non-determinism on the inference side still apply. If an exam option says "temperature 0 guarantees deterministic output", it's wrong.
The 2am Bedrock failure that covered Domains 3 and 5 at once
2025-11-18, 02:14 UTC. A client's ticket-summarizer Lambda started throwing every invocation, and the pager went to me because their only ops person was asleep. The CloudWatch line was blunt:
An error occurred (AccessDeniedException) when calling the InvokeModel operation: User: arn:aws:iam::111122223333:user/ticket-summarizer-role is not authorized to perform: bedrock:InvokeModel on resource: arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-3-5-sonnet-20241022-v2:0
First reaction was to widen the IAM policy. Wrong instinct. The real cause was that model access had been requested in `us-west-2` months earlier, while
Leave a Reply