Read this the night before or on exam morning. It assumes you’ve studied the notes. It’s here to refresh, not to teach.
Exam day
- 65 questions in 90 minutes: 50 scored and 15 unscored questions you can’t spot. Scaled score 100 to 1,000; you need 700.
- Question types: multiple choice, multiple response, ordering, matching and case study. No penalty for guessing, so never leave one blank.
- Scoring is compensatory: you pass on the overall score, not per domain.
- Domain weights: foundation model applications 28%, generative AI fundamentals 24%, AI and ML fundamentals 20%, responsible AI 14%, security, compliance and governance 14%.
- About 1.4 minutes a question. Flag long ones and move on.
How to read an AIF question
Read the last sentence first, then look for the constraint words.
| The question says |
Pick |
| Least effort, no ML expertise, quickest |
A managed AI service (Rekognition, Comprehend, Textract) or Bedrock; not a custom SageMaker model |
| Most cost-effective way to adapt a model |
Prompt engineering, then RAG, before fine-tuning |
| Current or private company data |
RAG (Knowledge Bases) |
| Block topics, mask PII, filter harmful content |
Guardrails for Amazon Bedrock |
| Detect bias or explain predictions |
SageMaker Clarify |
| Human review of predictions |
Amazon A2I |
| Prove compliance to auditors |
AWS Artifact or AWS Audit Manager |
| Exact, guaranteed outcome |
Not ML; use rules |
Elimination: options that call for training a model from scratch are almost always wrong for a foundational exam. Options that expose data publicly or skip encryption are wrong.
Service to task
| Task |
Service |
| Image and video analysis, faces, moderation |
Rekognition |
| Text and forms from documents |
Textract |
| Sentiment, entities, PII in text |
Comprehend |
| Speech to text |
Transcribe |
| Text to speech |
Polly |
| Translation |
Translate |
| Chatbots and voice bots |
Lex |
| Enterprise search |
Kendra |
| Recommendations |
Personalize |
| Foundation models by API, RAG, agents, guardrails |
Bedrock |
| Open or custom models on your endpoints |
SageMaker JumpStart / SageMaker |
| No-code app experiments |
PartyRock |
| Company-data assistant |
Amazon Q Business |
| Coding assistant |
Amazon Q Developer |
| No-code ML for analysts |
SageMaker Canvas |
SageMaker features
| Feature |
Job |
| Data Wrangler |
Prepare data and features visually |
| Feature Store |
Share and reuse features |
| Ground Truth |
Label data with humans |
| JumpStart |
Pre-trained models and templates |
| Clarify |
Bias and explainability |
| Model Monitor |
Drift in production |
| Model Cards |
Document a model |
| Pipelines |
Orchestrate ML workflows |
Customisation ladder (cost low to high)
Prompt engineering, RAG, fine-tuning, continued pre-training, pre-training from scratch. Fine-tuned models in Bedrock need provisioned throughput.
Prompt and inference cheat sheet
| Idea |
Remember |
| Zero-shot, one-shot, few-shot |
No example, one example, several examples |
| Chain-of-thought |
“Think step by step” |
| Temperature low / high |
Focused and repeatable / creative and varied |
| Top-p, top-k |
Limit the pool of candidate tokens |
| Prompt injection |
Untrusted text overrides instructions |
| Jailbreak |
Bypass safety rules |
| Poisoning |
Bad data planted in training or retrieval |
Metrics
| Metric |
Use |
| Accuracy |
Balanced classes only |
| Precision |
Costly false positives |
| Recall |
Costly false negatives |
| F1 |
Balance of both, imbalanced data |
| AUC |
Compare classifiers across thresholds |
| ROUGE |
Summaries |
| BLEU |
Translation |
| BERTScore |
Semantic similarity |
Learning types
| Type |
Cue |
| Supervised |
Labeled data; classification or regression |
| Unsupervised |
No labels; clustering |
| Reinforcement |
Agent, reward, environment |
| Overfitting |
Great on training data, poor on new data (high variance) |
| Underfitting |
Poor everywhere (high bias) |
Vector stores
OpenSearch, Aurora PostgreSQL, RDS for PostgreSQL (pgvector), Neptune, DocumentDB, MemoryDB. Embeddings are vectors; chunking splits documents before embedding.
Security and governance mappings
| Need |
Service |
| Who can call the model |
IAM |
| Encrypt data at rest |
KMS |
| Keep traffic private |
PrivateLink (VPC endpoint) |
| Find PII in S3 |
Macie |
| Store secrets |
Secrets Manager |
| Record API activity |
CloudTrail |
| Check resource configuration |
Config |
| Vulnerability scanning |
Inspector |
| AWS compliance reports |
Artifact |
| Collect audit evidence |
Audit Manager |
| Best-practice checks |
Trusted Advisor |
| Framework for scoping generative AI risk |
Generative AI Security Scoping Matrix (five scopes: consumer, enterprise, pre-trained, fine-tuned, self-trained) |
Responsible AI words
Fairness, explainability, privacy and security, safety, controllability, veracity, robustness, governance, transparency, inclusivity. Smaller models mean a lower environmental footprint.
Pairs people mix up
- Bedrock versus JumpStart: serverless API of many FMs versus open models on your own endpoints.
- Fine-tuning versus continued pre-training: labeled task data versus unlabeled domain data.
- Clarify versus Model Monitor: bias and explanations versus production drift.
- Artifact versus Audit Manager: AWS’s reports versus your evidence.
- Guardrails versus Clarify: runtime content control versus bias analysis.
- Batch versus real-time inference: scheduled bulk scoring versus interactive requests.