在打印窗口把「目标」选为「另存为 PDF」即可保存,页面已带 awsgood.com 版权水印。 返回原文

www.awsgood.com

AIF-C01 Exam Domain Deep Dive: Fundamentals of AI and ML (2026)

2026-11-10 AIF-C01 AI Practitioner 约 8 分钟

TL;DR

Who this exam is really for

AIF-C01 sits at the foundational tier, next to Cloud Practitioner. No prerequisites, no calculus, no training runs. If you can read an IAM policy and explain why a chatbot sometimes invents a refund policy, you're already past the hardest conceptual step.

The realistic candidates I meet: a developer who's been asked to bolt a model onto an internal app, an ops engineer who inherited a Bedrock bill nobody can explain, a sales engineer who needs to talk credibly about RAG. The exam rewards breadth over depth — you need to know which AWS service solves a stated problem, not how to tune it.

If you're deciding between this and a higher tier, the sequence almost everyone should follow is Cloud Practitioner (or equivalent experience) → AIF-C01 → ML Engineer Associate. Skipping straight to associate-level AI material with no AWS vocabulary is doable and miserable.

The five domains and where the points actually sit

The weights come straight from the official exam guide. The "questions" column is my arithmetic on 65 items — treat it as a planning aid, not an AWS promise. AWS reports results by domain on your score report, so a weak domain is visible after the fact whether you planned for it or not.

# Domain Weight ≈ of 65 questions Where it bites
1 Fundamentals of AI and ML 20% ~13 terminology, learning types, evaluation metrics
2 Fundamentals of Generative AI 24% ~16 tokens, context windows, embeddings, cost drivers
3 Applications of Foundation Models 28% ~18 inference parameters, RAG, agents, prompt engineering
4 Guidelines for Responsible AI 14% ~9 bias, explainability, Guardrails, Clarify
5 Security, Compliance and Governance 14% ~9 IAM, KMS, PrivateLink, prompt injection, data residency

Domains 2 and 3 together are 52% of the paper. That's the whole game. Domain 1 is the keyword people search for, but it's a fifth of the score.

Domain 1: what "Fundamentals of AI and ML" actually tests

You need clean definitions and the ability to map a business sentence onto a service. Supervised vs unsupervised vs reinforcement, classification vs regression, training vs inference, labeled vs unlabeled. The classic service-mapping items: Comprehend for text sentiment and entities, Rekognition for images and video, Forecast for time series, Personalize for recommendations, Transcribe and Polly for speech, Textract for document extraction, SageMaker when the question says "custom model".

The trap is metrics. Every practice set I've worked through has a scenario where one error type is expensive, and the answer hinges on reading that correctly.

Metric Pick it when the prompt says
Accuracy classes are balanced and both error types cost about the same
Precision false positives hurt — blocking a legitimate transaction, flagging real mail as spam
Recall false negatives hurt — missed fraud, missed screening, missed defect
F1 one number needed on an imbalanced set
RMSE / MAE regression, not classification
AUC-ROC ranking quality across thresholds

Overfitting shows up as "great on training data, poor in production" — that phrasing points at more data, regularization, or a simpler model, not more epochs. Underfitting is the mirror image.

Domains 2 and 3: 52% of the score in two sittings

Generative AI fundamentals is vocabulary with teeth. A token is roughly ¾ of a word; 1,000 tokens lands near 750 English words. Context window is the hard ceiling on prompt plus completion, and it's the reason long chat histories suddenly return errors. Embeddings are vectors, similarity search is nearest-neighbour, transformers are the architecture, diffusion is the image path. Hallucination is a property of the model, not a bug you file.

Domain 3 is where hands-on experience pays off. The inference parameters get asked about directly:

Then the architectural comparisons, which is where most of the 18 questions live:

Approach Changes Good for Cost profile
Prompt engineering nothing, just the input format control, tone, quick iteration lowest — no training, no extra infra
RAG (Bedrock Knowledge Bases) the context you inject private or fresh facts, citations, per-user filtering embedding + vector store + extra input tokens
Fine-tuning model weights consistent style and format at scale training job + custom model hosting
Continued pre-training weights, on domain text specialised vocabulary, huge corpora highest — needs volume and budget approval

One rule of thumb I use when the exam asks: facts change → RAG. Behaviour and format need to change → fine-tuning. Agentic, multi-step actions with tool calls → Bedrock Agents. You don't fine-tune to teach a model your refund policy that changes monthly; that's a RAG problem.

Domains 4 and 5: the 28% people skip and then fail

Responsible AI is mostly vocabulary plus the right AWS tool. Fairness, explainability, privacy, safety, veracity, transparency, controllability. Bias arrives through data, labelling and even feature selection. SageMaker Clarify detects bias and explains predictions; Model Monitor watches drift in production; Bedrock Guardrails filters harmful content and redacts PII at runtime. Knowing that Guardrails is runtime protection and Clarify is data/model-level analysis answers a surprising number of items on its own.

Security and governance is classic AWS ops wearing an AI hat. Least-privilege IAM, KMS for encryption at rest and TLS in transit, VPC endpoints and PrivateLink for traffic that must not cross the public internet, CloudTrail for the audit trail, Macie for PII discovery in S3. Prompt injection and jailbreaking are the new entries — the answer is almost always input validation plus Guardrails plus not granting the model more permissions than the task needs.

A 02:40 Bedrock failure that maps straight to Domains 3 and 5

A chat endpoint started 500-ing. The alert came in at 02:40 with 100% error rate over five minutes. Lambda logs showed this:

An error occurred (AccessDeniedException) when calling the InvokeModel operation:
User: arn:aws:iam::111122223333:role/chat-api-prod is not authorized to perform:
bedrock:InvokeModel on resource:
arn:aws:bedrock:us-east-1:111122223333:inference-profile/us.anthropic.claude-3-5-sonnet-20241022-v2:0
because no identity-based policy allows the bedrock:InvokeModel action

The IAM policy had been copied from an older tutorial and scoped to foundation-model ARNs only. Someone had switched the code to a cross-region inference profile the week before to chase capacity, and the profile resource needs its own permission. The policy was correct for the thing it was written for and wrong for the thing now running — a very common shape of outage.

Diagnosis took four minutes: aws sts get-caller-identity to confirm the role, then a CloudTrail lookup on eventSource = bedrock.amazonaws.com with errorCode = AccessDenied, which shows the exact resource ARN the call targeted. Fix was adding the inference-profile ARN plus the underlying foundation-model ARNs to the same statement. Verified by re-running the call by hand and getting a normal response with "stop_reason":"end_turn". The alarm cleared on the next evaluation period.

# What models does this account/region actually see?
aws bedrock list-foundation-models --region us-east-1 
  --query 'modelSummaries[].{id:modelId,provider:providerName}' --output table

# Prove invocation works before you blame the app
aws bedrock-runtime invoke-model --region us-east-1 
  --model-id anthropic.claude-3-haiku-20240307-v1:0 
  --cli-binary-format raw-in-base64-out 
  --body '{"anthropic_version":"bedrock-2023-05-31","max_tokens":64,"messages":[{"role":"user","content":"Explain RAG in one sentence."}]}' 
  /tmp/out.json && cat /tmp/out.json


Two habits come out of this. Model IDs and profile IDs change faster than blog posts do, so list them rather than copying a string from an article (this one included). And every time you swap a model or add a region, re-check the policy — this is exactly the kind of detail AIF-C01 tests under inference parameters and IAM.

## Registration, fee and the decision to book

The exam runs through Pearson VUE, either at a test centre or online-proctored. The fee is set per region and is around USD 100 in the US — check your local certification page before budgeting, because currency and tax treatment differ. The certification is valid for three years. Reschedule or cancel at least 48 hours before your slot or you forfeit the fee; I've watched a colleague lose a booking to a 30-hour-notice reschedule.

Before booking, do one full-length practice set cold and score it. Under 65%? Give yourself three weeks, not ten days. Over 80%? Book for two weeks out and spend the gap on Domains 2 and 3 only.

## A 14-evening study plan that fits around on-call

Seven evenings of reading, five of practice, two of review.

1. Evenings 1–2: Domain 2 and 3 material — this is 52% of the paper, do it first and with the console open.
2. Evenings 3–4: Domain 1 vocabulary and the metrics table above.
3. Evening 5: Domains 4 and 5, including one pass through the Guardrails and Clarify feature docs.
4. Evenings 6–8: two practice sets, then re-read every wrong answer's explanation — including the ones you got right by guessing.
5. Evenings 9–12: hands-on. Enable model access in one region, build a small Knowledge Base, set an inference parameter and watch output change.
6. Evenings 13–14: timed full mock, then a review of only the flagged items.

If you'd rather not spend evenings wiring Bedrock IAM and guardrails just to learn the shape of

Sources