Cheat Sheet & Exam Strategy
Every high-yield AIF-C01 association on one page, plus how to attack scenario questions on exam day.
10 min read
Concept associations
| When you see… | Think… |
|---|---|
| Labeled data | Supervised learning |
| No labels, find groups | Unsupervised learning / clustering |
| Reward, trial and error | Reinforcement learning |
| Predict a category / a number | Classification / regression |
| Great on training data, bad on new data | Overfitting |
| Can't miss a positive case | Optimize recall |
| False alarms are costly | Optimize precision |
| Imbalanced dataset | Don't trust accuracy; use F1/AUC |
| Confidently wrong answers | Hallucination → fix with RAG + human review |
| Same prompt, different outputs | Nondeterminism → lower temperature |
| Creativity dial | Temperature |
| Examples inside the prompt | Few-shot (in-context learning) |
| "Think step by step" | Chain-of-thought |
| Malicious instructions in input | Prompt injection |
| Company documents + citations, data changes often | RAG / Knowledge Bases |
| Change model behavior/style with labeled pairs | Fine-tuning |
| Unlabeled domain corpus | Continued pre-training |
| Multi-step tasks calling APIs | Bedrock Agents |
| Filter harmful content / mask PII in GenAI | Bedrock Guardrails |
| Summarization metric / translation metric | ROUGE / BLEU |
| Bias detection & explainability | SageMaker Clarify |
| Drift in production | SageMaker Model Monitor |
| Human review of predictions | Amazon A2I |
| Document your model | SageMaker Model Cards |
| Keep GenAI traffic off the internet | PrivateLink / VPC endpoints |
| Who invoked the model (audit) | CloudTrail |
| Configuration compliance | AWS Config |
Service one-liners
| Service | One line |
|---|---|
| Amazon Bedrock | FMs from many providers via one serverless API |
| Bedrock Knowledge Bases | Managed RAG over your documents |
| Bedrock Agents | LLM that plans steps and calls your APIs |
| Bedrock Guardrails | Safety filters + PII masking for GenAI |
| Amazon Q Business | Ready-made employee assistant on company data |
| Amazon Q Developer | AI coding assistant |
| AWS PartyRock | Free no-code GenAI playground |
| Amazon SageMaker AI | Build/train/deploy custom ML |
| SageMaker JumpStart | Pre-built models to start from |
| SageMaker Clarify | Bias detection + explainability |
| SageMaker Model Monitor | Production drift monitoring |
| SageMaker Ground Truth | Human data labeling |
| Amazon Rekognition | Images & video analysis |
| Amazon Comprehend | Text NLP: sentiment, entities, PII |
| Amazon Transcribe / Polly | Speech→text / text→speech |
| Amazon Translate | Translation |
| Amazon Textract | Extract text/forms from documents |
| Amazon Lex | Chatbots |
| Amazon Kendra | Intelligent enterprise search |
| Amazon Personalize / Forecast | Recommendations / time-series forecasts |
| Amazon OpenSearch | Vector database for semantic search |
| Trainium / Inferentia | AWS chips for training / inference |
Exam-day tactics
- Find the constraint: "no ML expertise" → AI service; "lowest cost, no training" → prompting/RAG; "must cite sources" → RAG; "consistent behavior" → fine-tuning.
- Eliminate the extremes: "train from scratch" and answers removing human oversight are nearly always wrong.
- Watch superlatives: LEAST effort → most managed option; MOST cost-effective → smallest sufficient model / prompting before tuning.
- Budget ~80 seconds per question; flag and return rather than stall.
- Answer everything — no penalty for guessing.
Tip
Take Mock Exam 1 now. Review misses, reread those lessons, drill the practice bank, then confirm with Mock Exam 2. Two 750+ scores = book the real exam with confidence.
Knowledge check
Question 1 of 3Match the need: "Our chatbot must answer from current company documents AND show its sources, with no model training."