AWS AI Practitioner Study Guide

Core Concepts

• • Artificial Intelligence (AI): machines that can perform tasks that typically require human intelligence
• • Machine Learning (ML): subset of AI where systems learn from data without explicit programming
• • Deep Learning: subset of ML using neural networks with many layers
• • Training: feeding data to a model so it learns patterns
• • Inference: using a trained model to make predictions on new data

Types of Machine Learning

• • Supervised Learning: learns from labeled data (classification, regression)
• • Unsupervised Learning: finds patterns in unlabeled data (clustering, anomaly detection)
• • Semi-supervised Learning: mix of labeled and unlabeled data
• • Reinforcement Learning: learns by trial and error with rewards/penalties
• • Self-supervised Learning: generates labels from the data itself

Common ML Tasks

• • Classification: categorize into discrete classes (spam detection, image recognition)
• • Regression: predict continuous values (price prediction, demand forecasting)
• • Clustering: group similar data points (customer segmentation)
• • Anomaly Detection: find unusual patterns (fraud detection)
• • Recommendation: suggest items based on patterns (product recommendations)

ML Pipeline

1. Data Collection and preparation
2. Feature Engineering (select and transform input variables)
3. Model Training (choose algorithm, train on data)
4. Model Evaluation (accuracy, precision, recall, F1 score)
5. Model Deployment (serve predictions in production)
6. Monitoring and Retraining

AWS AI/ML Services

• • Amazon SageMaker: build, train, and deploy ML models
• • Amazon Rekognition: image and video analysis
• • Amazon Comprehend: natural language processing (NLP)
• • Amazon Transcribe: speech to text
• • Amazon Polly: text to speech
• • Amazon Translate: language translation
• • Amazon Textract: extract text from documents
• • Amazon Forecast: time-series forecasting
• • Amazon Personalize: real-time recommendations
• • Amazon Kendra: intelligent enterprise search

What is Generative AI?

• • AI that can create new content: text, images, code, audio, video
• • Built on foundation models trained on massive datasets
• • Uses patterns learned during training to generate new outputs
• • Key difference from traditional ML: creates vs classifies/predicts

Foundation Models and LLMs

• • Foundation Model: large-scale model trained on broad data, adaptable to many tasks
• • Large Language Model (LLM): foundation model specialized for language tasks
• • Transformer Architecture: attention mechanism, parallel processing
• • Tokens: basic units of text (words, subwords, characters)
• • Context Window: how much text the model can process at once
• • Parameters: learned weights that define model behavior (billions)

Prompt Engineering

• • Zero-shot: prompt without examples
• • Few-shot: prompt with a few examples
• • Chain-of-thought: ask model to show reasoning steps
• • System prompts: set behavior and persona for the model
• • Temperature: controls randomness (0 = deterministic, 1 = creative)
• • Top-p (nucleus sampling): controls diversity of responses
• • Max tokens: limit the length of generated output

Model Customization Techniques

• • Fine-tuning: train model further on domain-specific data
• • RAG (Retrieval Augmented Generation): enhance responses with external knowledge
• • In-context learning: guide model through examples in the prompt
• • RLHF: Reinforcement Learning from Human Feedback
• • RAG vs Fine-tuning: RAG for real-time data, fine-tuning for behavior change

Generative AI Applications

• • Text generation: chatbots, summarization, translation
• • Code generation: code completion, debugging, documentation
• • Image generation: art, design, product visualization
• • Content creation: marketing copy, reports, emails
• • Knowledge assistants: Q&A over enterprise documents

Amazon Bedrock ⭐

• • Fully managed service to access foundation models via API
• • Models from: Anthropic (Claude), Meta (Llama), AI21 Labs, Cohere, Stability AI, Amazon (Titan)
• • No infrastructure management — serverless
• • Customize models with your own data (fine-tuning, continued pre-training)
• • Knowledge Bases: connect models to your data sources for RAG
• • Agents: enable models to take actions and call APIs
• • Guardrails: filter harmful content, enforce policies
• • Model evaluation: compare models for your use case

Amazon Titan Models

• • Titan Text: text generation, summarization, Q&A
• • Titan Embeddings: convert text to numerical vectors for search
• • Titan Image Generator: create and edit images from text
• • Titan Multimodal Embeddings: combine text and image understanding
• • Built-in safety protections and watermarking

Amazon Q ⭐

• • Amazon Q Business: AI assistant for enterprise (connects to data sources)
• • Amazon Q Developer: AI coding assistant (code suggestions, transformation)
• • Understands enterprise context through connectors
• • Respects existing access controls and permissions

Amazon SageMaker for Foundation Models

• • SageMaker JumpStart: deploy pre-trained foundation models
• • SageMaker Canvas: no-code ML for business users
• • SageMaker Studio: full IDE for ML development
• • Fine-tune open-source models (Llama, Falcon, etc.)
• • Deploy custom models with managed endpoints

Model Selection Considerations

• • Task type: text, code, image, multimodal
• • Performance: quality vs latency tradeoffs
• • Cost: on-demand pricing, provisioned throughput
• • Context window: how much input the model can handle
• • Customization: fine-tuning support, RAG capability
• • Compliance: data residency, regulatory requirements

Responsible AI Principles

• • Fairness: AI systems should not discriminate against groups
• • Explainability: understand why a model made a decision
• • Transparency: be open about AI usage and limitations
• • Privacy: protect personal and sensitive data
• • Safety: prevent harmful outputs and misuse
• • Robustness: model performs reliably across conditions
• • Governance: policies, oversight, and accountability

Bias in AI

• • Training data bias: unrepresentative or skewed datasets
• • Selection bias: non-random data selection
• • Measurement bias: inconsistent data collection
• • Confirmation bias: model reinforces existing patterns
• • Mitigation: diverse data, bias metrics, human review, Amazon SageMaker Clarify

AI Challenges

• • Hallucination: model generates plausible but incorrect information
• • Mitigation: RAG, grounding, guardrails, temperature control
• • Toxicity: harmful, offensive, or inappropriate content
• • Intellectual property: copyright concerns with training data
• • Model drift: performance degrades over time as data changes

AWS Tools for Responsible AI

• • Amazon SageMaker Clarify: detect bias, explain predictions
• • Amazon Bedrock Guardrails: content filtering, topic avoidance
• • Amazon Augmented AI (A2I): human review workflows
• • Model Cards: document model performance, limitations, intended use
• • SageMaker Model Monitor: detect data/model drift in production

Shared Responsibility Model for AI

• • AWS responsibility: security OF the cloud (infrastructure, managed services)
• • Customer responsibility: security IN the cloud (data, access, model outputs)
• • For Bedrock: AWS manages infrastructure; you manage data, prompts, and outputs
• • For SageMaker: more customer responsibility for model training and deployment

Data Security for AI

• • Encryption at rest: AWS KMS, S3 encryption for training data
• • Encryption in transit: TLS/SSL for API calls and data transfer
• • Data isolation: Bedrock does not use your data to train base models
• • VPC endpoints: private connectivity without internet exposure
• • Amazon Macie: discover and protect sensitive data (PII)

Identity and Access Management

• • IAM policies: control who can access AI services and models
• • IAM roles: grant temporary access for SageMaker training jobs
• • Service-linked roles: automatic permissions for AWS services
• • Least privilege: grant minimum permissions needed
• • Resource-based policies: control access to specific models

Compliance and Governance

• • AWS CloudTrail: log all API calls for AI services (audit trail)
• • AWS Config: track configuration changes to AI resources
• • AWS Artifact: access AWS compliance reports (SOC, ISO, HIPAA)
• • GDPR, HIPAA, SOC 2: know which apply to your AI workloads
• • Model governance: track model versions, approvals, lineage
• • SageMaker Model Registry: version control for models

Cost Management

• • Bedrock pricing: per input/output token for on-demand
• • Provisioned throughput: reserved capacity for consistent workloads
• • SageMaker: pay for compute instances, storage, data transfer
• • AWS Cost Explorer: analyze AI service spending
• • AWS Budgets: set alerts for AI spending thresholds