AWS Certified AI Practitioner (AIF-C01) Cheat Sheet
High-yield architectural matrices, inference tuning reference tables, model catalog comparisons, Bedrock Guardrail layers, and responsible AI principles for last-minute exam day review.
📌 Quick Summary / Core Test Principles:AIF-C01 Quick Rules: DL ⊂ ML ⊂ AI; GenAI is DL with Foundation Models; Supervised = labeled (Regression = continuous / Classification = discrete); Unsupervised = unlabeled (K-Means clustering); Overfitting = high train / low val (fix with regularization/dropout); Recall = TP / (TP+FN) (prioritize for medical/security); Precision = TP / (TP+FP) (prioritize for spam); F1 = harmonic mean; Temperature: low (0.0) for compliance, high (0.8+) for creative; Top-P = cumulative probability sampling; 1,000 tokens ≈ 750 words; RAG = dynamic knowledge + citations; Fine-tuning = style/format adaptation; Bedrock = serverless API; SageMaker = dedicated instance control; Bedrock Guardrails = Denied topics + Content filters + PII masking + Contextual Grounding; Customer data in Bedrock is NEVER used to train base models!
Core Taxonomy
1. AI, ML, Deep Learning & Generative AI Taxonomy
| Discipline | Definition & Core Mechanism | Key Characteristics | Primary AWS Services |
|---|---|---|---|
| Artificial Intelligence (AI) | Broad discipline of creating computer systems capable of performing tasks requiring human intelligence. | Encompasses rule-based expert systems, heuristic search, automation, and statistical algorithms. | Amazon Bedrock, SageMaker, Amazon Q |
| Machine Learning (ML) | Subset of AI where algorithms learn statistical patterns directly from data without explicit procedural programming. | Categorized into Supervised, Unsupervised, and Reinforcement Learning; requires feature engineering. | Amazon SageMaker, SageMaker Canvas, Rekognition |
| Deep Learning (DL) | Subset of ML based on multi-layered artificial neural networks inspired by biological neural connections. | Automatically extracts hierarchical representations from raw unstructured data (images, audio, text). | SageMaker Training, EC2 Trn1/Inf2, PyTorch/TensorFlow |
| Generative AI (GenAI) | Subset of Deep Learning powered by foundation models that generate novel text, images, audio, or synthetic code. | Built predominantly on Transformer self-attention; trained self-supervised on internet-scale corpora. | Amazon Bedrock, SageMaker JumpStart, Amazon Q Developer |
💡 Pro Tip:Exam rule of thumb: Deep Learning is a subset of Machine Learning; Generative AI is a subset of Deep Learning.
Architectural Decision
2. Generative AI Customization Spectrum: From Prompting to Pre-Training
| Customization Method | Effort & Cost | When to Use | Key Architectural Consideration |
|---|---|---|---|
| Prompt Engineering | Lowest (Zero GPU cost) | Quick formatting, summarization, zero-shot/few-shot tasks using off-the-shelf foundation models. | Operates entirely within runtime context window; does not modify model weights. |
| Retrieval-Augmented Generation (RAG) | Low-Medium (Vector storage) | Frequently changing private enterprise data, internal policy Q&A, and verifiable source citations. | Separates knowledge base retrieval from generative reasoning; prevents hallucinations without retraining. |
| Fine-Tuning (PEFT / LoRA) | Medium-High (GPU training compute) | Specialized jargon, highly unique brand voice, domain-specific syntax, or structured output formats. | Updates model weights (or lightweight adapter matrices); requires high-quality labeled JSONL datasets. |
| Continued Pre-Training | High (Days of GPU compute) | Injecting massive volumes of unlabeled domain text (e.g., medical journals, legal statutory codes). | Adapts vocabulary embeddings and base knowledge before task-specific fine-tuning. |
| Pre-Training from Scratch | Extremely High ($ Millions in compute) | Radically new modalities or languages with zero existing foundation model representation. | Requires trillions of tokens, thousands of GPUs, and deep AI research teams. Almost never needed for enterprise. |
💡 Pro Tip:Default to RAG (Bedrock Knowledge Bases) for dynamic enterprise data. Only fine-tune when you must teach a unique style, tone, or niche format.
Inference Tuning
3. Inference Hyperparameters & Sampling Controls
| Parameter | Range / Settings | Functional Impact | Recommended Exam Use Case |
|---|---|---|---|
| Temperature | 0.0 to 1.0 (or 2.0) | Scales probability distribution of next-token logits. Lower = deterministic & focused; Higher = creative & varied. | Use 0.0–0.2 for financial compliance, code generation, and factual extraction; use 0.7–1.0 for creative copy. |
| Top-P (Nucleus Sampling) | 0.0 to 1.0 | Dynamically samples from the smallest pool of tokens whose cumulative probability reaches P. | Set Top-P to 0.9 to eliminate the long tail of bizarre tokens while preserving conversational fluency. |
| Top-K | Integer (e.g., 1 to 500) | Restricts the candidate pool to a static count (K) of the most probable next tokens before sampling. | Set Top-K to 50 to strictly bound vocabulary selection regardless of probability distribution shape. |
| Max Generation Length | Token count (e.g., 2048) | Enforces a hard upper ceiling on the number of tokens the model can produce in a single completion. | Prevents runaway generation, controls API token costs, and bounds latency. |
| Stop Sequences | List of strings (e.g., ['\n\n', 'User:']) | Instructs the model to immediately halt token generation upon encountering any specified character sequence. | Essential for multi-turn conversational agents to prevent the model from hallucinating simulated user replies. |
💡 Pro Tip:For deterministic, repeatable, and factual answers: Lower Temperature (0.0) and lower Top-P (0.1–0.3). Never set high temperature for compliance tasks!
Model Selection
4. Amazon Bedrock Foundation Model Catalog Matrix
| Model Family | Provider | Core Strengths | Target Workloads |
|---|---|---|---|
| Anthropic Claude (3.5 Sonnet / Opus / Haiku) | Anthropic | Industry-leading analytical reasoning, advanced software engineering/coding, vision analysis, nuanced writing. | Complex code generation, deep multi-step reasoning, contract review, complex RAG synthesis. |
| Amazon Titan (Text G1 / Embeddings / Multimodal) | Amazon | Cost-effective text generation, high-speed vector embeddings, and image generation with built-in watermarking. | Enterprise semantic search embeddings, general document summarization, cost-optimized text processing. |
| Meta Llama (Llama 3 / 3.1) | Meta | High-performance open weights, strong general knowledge, instruction following, and multilingual support. | General-purpose enterprise assistants, fine-tuning for domain specialization, open-architecture alignment. |
| Mistral AI (Mistral Large / Mixtral 8x7B) | Mistral AI | High efficiency, strong reasoning, open mixture-of-experts (MoE) architecture, low latency. | Low-latency classification, multilingual European language tasks, structured JSON extraction. |
| AI21 Labs Jurassic | AI21 Labs | Exceptional natural language generation, complex text rewriting, and grammatical precision. | Long-form content drafting, marketing copy, and multi-paragraph synthesis. |
| Cohere (Command & Embed) | Cohere | Highly optimized for enterprise business tasks and multilingual retrieval vector embeddings. | Enterprise search ranking, reranking, and business document classification. |
💡 Pro Tip:When the exam mentions 'advanced coding, mathematical reasoning, and complex logic on Bedrock', the answer is Anthropic Claude.
Platform Comparison
5. Amazon Bedrock vs. Amazon SageMaker Architectural Matrix
| Dimension | Amazon Bedrock | Amazon SageMaker | SageMaker JumpStart |
|---|---|---|---|
| Management Model | 100% Serverless (Zero infrastructure) | Managed instances (User chooses EC2 instance types) | Pre-trained model hub deploying to dedicated SageMaker instances |
| API & Invocations | Unified REST API (InvokeModel) | Dedicated SageMaker hosting endpoints | Dedicated hosting endpoints with direct PyTorch container control |
| Pricing Paradigm | Pay per 1,000 tokens (or Provisioned Model Units) | Pay per instance-hour for compute running 24/7 | Pay per instance-hour for hosting instances + storage |
| Model Artifact Access | Managed weights; no direct file access | Full root access to S3 model weights and code | Direct access to open-source model weights in private S3 |
| Best Used For | Fast generative AI deployment, managed RAG, and guardrails without infrastructure overhead. | End-to-end custom machine learning, complex ETL pipelines, and specialized neural network training. | Hosting open-weight models with custom VPC networking, custom containers, and dedicated hardware control. |
💡 Pro Tip:Bedrock is serverless API consumption. SageMaker is full lifecycle control on dedicated infrastructure.
Safety & Security
6. Guardrails for Amazon Bedrock: 5 Safety Filters
| Protection Layer | Purpose | Configuration Options | Exam Use Case |
|---|---|---|---|
| Denied Topics | Blocks conversations on defined off-limit subjects. | Natural language topic descriptions with optional example prompts. | Preventing a financial chatbot from offering investment advice. |
| Content Filters | Moderates harmful or offensive text across both prompts and outputs. | Threshold levels (None, Low, Medium, High) across Hate, Insults, Sexual, Violence, Misconduct, and Prompt Attack. | Blocking toxic language and mitigating direct jailbreak prompt injection attempts. |
| Sensitive Information (PII) | Protects personally identifiable information. | Pre-built entity filters (SSN, credit card, phone, email) or custom regex. Actions: Mask or Block. | Redacting patient SSNs and financial account numbers in customer support chats. |
| Word Filters | Filters exact prohibited terms or custom profanity lists. | Custom word lists or predefined profanity blocklists. | Blocking competitor brand names or offensive corporate slang. |
| Contextual Grounding | Detects hallucinations in RAG systems. | Grounding score (fidelity to source) and Relevance score (alignment with user query). | Ensuring customer policy answers are strictly derived from retrieved knowledge base documents. |
💡 Pro Tip:Contextual Grounding is Bedrock's specialized filter for catching hallucinations in RAG by cross-referencing retrieved source text.