AWS AI Practitioner: ML Fundamentals Cheat Sheet

Posted Sep 29, 2026 Updated Sep 30, 2026

By Abhishek K

26 min read

Disclaimer: I put this review cheat sheet together using Claude AI and tweaked it a bit to make it easier for my own study sessions in Q2 2026. Hope it helps you on your journey!

This cheat sheet only contains AI and ML fundamentals and concepts (Domain 1). Please review the official AWS AI Practitioner Study Guide.

Other Resources:

Tutorials Dojo Practice Sets (highly recommended)
CloudNinja AI Practitioner Cheat Sheet

Quick Exam Domain Map

Domain	Weight	What it Tests
Domain 1 — ML Fundamentals	20%	ML & AI concepts and definitions
Domain 2 — Fundamentals of Generative AI	24%	GenAI concepts, LLMs, prompt engineering, FM types
Domain 3 — Applications of Foundation Models	28%	Bedrock, RAG, fine-tuning, agentic AI, Amazon Q
Domain 4 — Guidelines for Responsible AI	14%	Bias, fairness, explainability, safety, human review
Domain 5 — Security, Compliance & Governance	14%	IAM, encryption, Macie, GuardDuty, CloudTrail, shared responsibility

Domain 1: ML Fundamentals Cheat Sheet

Module 1: The AI/ML/Deep Learning Hierarchy

Understanding the relationship between these three terms is a foundational exam topic.

    Artificial Intelligence (AI)
    ┌───────────────────────────────────────────────────┐
    │  Any technique that enables machines to mimic     │
    │  human intelligence (rules, search, reasoning).   │
    │                                                   │
    │   Machine Learning (ML)                           │
    │   ┌────────────────────────────────────────────┐  │
    │   │  Subset of AI. Systems learn from data     │  │
    │   │  (find patterns) without being explicitly  │  │
    │   │  programmed.                               │  │
    │   │                                            │  │
    │   │   Deep Learning (DL)                       │  │
    │   │   ┌──────────────────────────────────────┐ │  │
    │   │   │  Subset of ML using multi-layered    │ │  │
    │   │   │  Neural Networks to learn from       │ │  │
    │   │   │  very large datasets automatically.  │ │  │
    │   │   └──────────────────────────────────────┘ │  │
    │   └────────────────────────────────────────────┘  │
    └───────────────────────────────────────────────────┘

Term	Plain Definition	Example
AI	Broad field — any system exhibiting intelligent behavior	Chess rules engine, chatbots
ML	AI that learns patterns from data without explicit rules	Spam filter trained on emails
Deep Learning	ML using deep neural networks; excels at unstructured data	Image recognition, LLMs
Neural Network	Computing architecture modeled on the brain; layers of connected nodes (neurons)	Any DL model
LLM	Very large transformer-based language model trained on massive text corpora	GPT, Claude, Amazon Titan
NLP	AI field focused on understanding and generating human language	Comprehend, Lex, Translate
Computer Vision	AI field focused on understanding images and video	Rekognition
Algorithm	The mathematical procedure/method used to learn from data	Linear Regression, XGBoost
Model	The output artifact produced after an algorithm trains on data	A trained fraud-detection model
Training	The process of feeding data to an algorithm so it learns parameters (weights)	Fitting a model to labeled data
Inferencing	Using a trained model to make predictions on new, unseen data	Running the model in production
Bias	Systematic error in a model’s outputs, often from skewed training data	Model underperforms for a group
Fairness	Ensuring model outputs are equitable across all groups	No discrimination by demographics
Fit	How well a model captures patterns — can be under-fit or over-fit	See Module 6

NLP Evaluation Metrics

ROUGE = RECALL - Did you COVER everything? Used for Summarization. Think: ROUGE = Remembering.

BLEU = PRECISION - Is what you wrote CORRECT? Used for Translation. Think: BLEU = Blue tick = Exact.

BERTScore = MEANING - Not exact words, but same SENSE? Think: BERT = Brain / Semantics.

ROUGE has R -> Recall -> Summary. BLEU has B -> Best precision -> Translation. BERT has Brain -> Meaning.

Evaluation metrics should be used to assess the model’s performance:

Root Mean Square Error (RMSE) squares the errors before averaging. This gives a high penalty to large misses. It helps when big forecasting errors cost the business a lot.
Mean Absolute Percentage Error (MAPE) shows the error as a percentage of real revenue. This makes it easy to explain financial accuracy to non-technical teams.

Module 2: Data Types in AI/ML

Exam Tip: The exam tests whether you can identify the right data type for a given scenario and the right ML technique to apply to it.

A. By Structure

Type	Definition	Examples
Structured	Organized into rows and columns with a defined schema	SQL databases, CSV files, spreadsheets
Unstructured	No predefined format or schema	Images, video, audio, raw text, PDFs
Semi-structured	Has some organizational properties but not a rigid schema	JSON, XML, email logs

B. By Label Status

Type	Definition	Used In
Labeled Data	Each data point has a known target value (y)	Supervised Learning
Unlabeled Data	Raw data with no target value; only features (X)	Unsupervised Learning, Self-supervised

C. By Domain/Format

Data Format	Description	Common ML Task	AWS Service
Tabular	Rows and columns (spreadsheet-like)	Classification, Regression	SageMaker, Autopilot
Time-Series	Data points indexed chronologically	Forecasting, Anomaly detection	Amazon Forecast
Image	Pixel grids	Object detection, Classification	Amazon Rekognition
Text	Natural language sequences	Sentiment analysis, Translation	Amazon Comprehend, Translate
Audio/Speech	Sound waveforms	Speech-to-text, Speaker ID	Amazon Transcribe

Module 3: ML Learning Paradigms & Problem Types

                            Machine Learning
                                   |
      ┌────────────────────────────┼────────────────────────────┐
      ▼                            ▼                            ▼
Supervised Learning      Unsupervised Learning        Reinforcement Learning
(Labeled Data: X → y)    (Unlabeled Data: X only)     (Trial, Error & Rewards)
  ├── Classification       ├── Clustering               └── Agent + Environment
  └── Regression           └── Dimensionality Red.

1. Supervised Learning: Classification vs. Regression

A. Classification

What it is: Mapping input features (X) to a discrete, categorical label (y).
Why/When to use: When the target is a category, flag, or label (Binary: Yes/No; Multi-class: Red/Blue/Green; Multi-label: Tagging multiple topics).
Use Cases: Email spam filtering, image recognition, fraud detection, customer churn prediction, sentiment analysis.
AWS Services: Amazon Rekognition (objects/moderation), Amazon Comprehend (sentiment/topic classification), Amazon SageMaker.
Exam Trigger Words: “Categorize”, “Is this A or B”, “Identify sentiment”, “Classify”, “Binary”, “Labels”.
Don’t Confuse With:
- Regression: Regression outputs a continuous number; classification outputs a discrete class.
- Logistic Regression: Despite having “Regression” in its name, it is a classification algorithm.

B. Regression

What it is: Mapping input features (X) to a continuous numerical value (y).
Why/When to use: When the target value can be any number on a continuous scale.
Use Cases: House price prediction, sales revenue forecasting, temperature estimation, insurance claim amount prediction.
AWS Services: Amazon Forecast, Amazon SageMaker (Linear Learner in regression mode, XGBoost).
Exam Trigger Words: “Predict numerical value”, “Forecast future amount”, “How much / How many”, “Continuous variable”.
Don’t Confuse With:
- Time-Series Classification: Predicting whether a stock goes Up/Down is classification; predicting the exact stock price ($152.45) is regression.

2. Unsupervised Learning: Clustering & Dimensionality Reduction

A. Clustering

What it is: Grouping unlabeled data points based on inherent mathematical similarities or distances (e.g., K-Means).
Why/When to use: When you have raw data without target labels and want to discover natural groupings.
Use Cases: Customer market segmentation, document grouping, genetic sequence grouping.
AWS Services: Amazon SageMaker (K-Means algorithm).
Exam Trigger Words: “Unlabeled”, “Find natural groups”, “Segment customers without prior labels”, “Clusters”.

B. Dimensionality Reduction

What it is: Reducing the number of input features while retaining as much meaningful variance/information as possible (e.g., PCA — Principal Component Analysis).
Why/When to use: When dealing with the “curse of dimensionality” (hundreds/thousands of columns), high compute costs, multicollinearity, or data visualization.
Use Cases: Compressing image data, reducing feature set before training, removing correlated redundant noise.
AWS Services: Amazon SageMaker (PCA algorithm).
Exam Trigger Words: “Reduce feature count”, “High-dimensional data”, “Compress inputs without losing variance”, “PCA”.
Don’t Confuse With:
- Feature Selection: Selecting a subset of existing columns. Dimensionality reduction creates new, combined lower-dimensional components.

3. Reinforcement Learning (RL)

What it is: An Agent interacts with a dynamic Environment, observes the State, takes Actions, and receives Rewards or penalties to learn an optimal policy.
Why/When to use: For sequential decision-making problems where immediate feedback isn’t available, but an overall goal must be optimized.
Use Cases: Autonomous driving, game playing (Chess, Go), robotic arm manipulation, dynamic pricing strategies.
AWS Services: AWS DeepRacer (1/18th scale autonomous car learning RL), AWS DeepComposer.
Exam Trigger Words: “Reward function”, “Agent and environment”, “Policy optimization”, “Sequential decisions”, “Trial and error”, “DeepRacer”.

Linear regression refers to supervised learning models that use one or more inputs to predict a value on a continuous scale. It is used to predict housing prices. After training a model using a set of historical sales training data that includes those characteristics, you could forecast the price of a property based on its location, age, and number of rooms.

Logistic regression is supervised learning algorithm, typically used for binary classification, and is not applicable for the unsupervised tasks.

Probability density is generally used to estimate the likelihood or possibility of a random variable falling within a particular range of values, not for predicting future values based on historical data.

Dimensionality reduction is primarily used to reduce the number of features in a dataset while retaining as much information as possible. It is not used for prediction tasks directly.

Anomaly detection is only used to identify outliers or unusual patterns in data, which is not the primary goal of predicting continuous portfolio values.

Semantic segmentation tags every pixel in an image with a class label from a predefined set of classes.

Module 4: Practical AI/ML Use Cases

When AI/ML IS Appropriate

Scenario	Why AI/ML Fits
Human decision support	Augments experts (e.g., medical imaging AI flags anomalies for a doctor)
Scalability	A single trained model can serve millions of predictions simultaneously
Automation of repetitive tasks	Document classification, invoice processing, spam filtering
Pattern recognition in large data	Fraud detection across billions of transactions
Personalization	Recommendation engines (content, products)
Prediction / Forecasting	Demand forecasting, predictive maintenance

When AI/ML is NOT Appropriate

Exam Tip: This is a commonly tested nuance. Not every problem needs AI.

Situation	Why AI/ML is NOT the Right Tool
A perfectly deterministic rule exists	If `age > 65 → senior discount`, a simple rule is more reliable and cheaper than a model
Cost outweighs benefit	Training and running models is expensive; small-scale problems may not justify the cost
A specific, exact outcome is required	ML models produce probabilities/predictions, not guaranteed exact answers
Insufficient data	ML needs sufficient data to learn; sparse data leads to unreliable models
High explainability required	Some regulated domains (legal, medical) may require fully auditable deterministic logic

Real-World AI Application Categories

Application	Description	AWS Service
Computer Vision	Analyzing images/video for objects, faces, text	Amazon Rekognition
NLP / Text Analytics	Sentiment, entity extraction, language detection	Amazon Comprehend
Speech Recognition	Convert audio to text (speech-to-text)	Amazon Transcribe
Text-to-Speech	Convert text to natural-sounding audio	Amazon Polly
Machine Translation	Translate between languages	Amazon Translate
Conversational AI / Chatbots	Build voice and text chatbots	Amazon Lex
Recommendation Systems	Personalized product/content recommendations	Amazon Personalize
Fraud Detection	Identify anomalous transactions in real time	Amazon Fraud Detector
Forecasting	Predict future values from time-series data	Amazon Forecast
Document Processing	Extract text and structured data from documents	Amazon Textract

Module 5: The ML Development Lifecycle

[Problem Definition] ➔ [Data Collection] ➔ [EDA] ➔ [Data Preprocessing] ➔ [Feature Engineering] ➔ [Model Training] ➔ [Hyperparameter Tuning] ➔ [Evaluation] ➔ [Deployment] ➔ [Monitoring & Retraining]

Stage-by-Stage Breakdown

Stage	What Happens	Key AWS Service
Problem Definition	Define the business objective and success metrics	—
Data Collection & Ingestion	Gather raw data from sources (databases, APIs, S3)	Amazon S3, AWS Glue
Exploratory Data Analysis (EDA)	Understand distributions, correlations, outliers, missing values; visualize data to form hypotheses	SageMaker Studio, Data Wrangler
Data Preprocessing	Handle missing values, remove duplicates, fix errors, encode categoricals, scale features	SageMaker Data Wrangler
Data Labeling	Assign ground truth labels (y) to raw data for supervised learning	Amazon SageMaker Ground Truth
Feature Engineering	Create, transform, or select input variables (X) to improve model learning	SageMaker Data Wrangler, Feature Store
Model Training	Feed prepared data to an algorithm so it learns parameters (weights/biases)	Amazon SageMaker Training
Hyperparameter Tuning	Optimize training configurations (learning rate, epochs, batch size) — set before training	SageMaker Automatic Model Tuning
Evaluation	Measure model performance on the held-out test set using metrics	SageMaker Experiments, Clarify
Deployment	Serve the model for inference in production	SageMaker Endpoints, Bedrock
Monitoring & Retraining	Track for drift and model degradation; retrain when performance drops	SageMaker Model Monitor

1. Data Labeling vs. Feature Engineering

Concept	What It Is	Purpose	Key AWS Service
Data Labeling	Assigning ground truth target labels (y) to raw unlabeled data.	Creates the training dataset required for Supervised Learning.	Amazon SageMaker Ground Truth (uses active learning + human workforces).
Feature Engineering	Transforming, creating, scaling, or selecting input variables (X).	Makes mathematical patterns easier for the ML algorithm to learn.	Amazon SageMaker Data Wrangler, Amazon SageMaker Feature Store.

2. Core Data Preprocessing & Feature Engineering Techniques

A. Encoding Categorical Variables

One-Hot Encoding: Converts categorical strings (Red, Green, Blue) into binary columns ([1,0,0], [0,1,0], [0,0,1]). Used when there is no inherent order.
Label / Ordinal Encoding: Converts ordered categories into numbers (Small = 1, Medium = 2, Large = 3).

B. Feature Scaling

Normalization (Min-Max Scaling): Rescales values to a range of [0, 1].
Standardization (Z-score): Centers data around a mean of 0 with standard deviation of 1.
Why it matters: Prevents features with large raw numbers (e.g., Annual Salary: $120,000) from mathematically overpowering features with small numbers (e.g., Age: 30).

C. Data Splitting & Leakage

Train / Validation / Test Split:
- Training Set (~70–80%): Model learns the weights/patterns.
- Validation Set (~10–15%): Used during training for hyperparameter tuning and model selection.
- Test Set (~10–15%): Unseen data used for the final, unbiased performance evaluation.
Target / Data Leakage: When information from outside the training dataset (or from the future target label) inadvertently leaks into the training data, producing unrealistically high training accuracy that fails in production.

3. Model Sources

Source	Description	Example
Open-source pre-trained models	Publicly available models already trained on large datasets; can be used directly or fine-tuned	Hugging Face models, open-source LLMs via Amazon Bedrock
Custom-trained models	Models trained from scratch on your own proprietary data for a specific task	Training a fraud model on your transaction data in SageMaker

4. Deployment Methods

Method	Description	When to Use
Managed API Service	AWS hosts and manages the model endpoint (scaling, patching, infra)	Most use cases; low operational overhead
Self-hosted API	You deploy the model on your own EC2/container infrastructure	Full control, custom hardware, or compliance requirements
Serverless Inference	Model runs only when invoked; no idle cost	Infrequent or unpredictable traffic
Batch Inference	Runs predictions on a large dataset all at once on a schedule	Non-real-time, bulk processing needs

5. Inferencing Types

Type	Description	Use Case
Real-Time Inference	Low-latency, synchronous prediction for individual requests	Fraud check at point of sale
Batch Inference	Asynchronous bulk predictions on large datasets	Overnight report scoring
Asynchronous Inference	Requests are queued; response returned when ready (for large payloads)	Long-running predictions on large files
Serverless Inference	Auto-scales to zero; cost based purely on invocations	Intermittent or spiky traffic

Module 6: Model Evaluation & Performance Tuning

1. Classification Metrics & The Confusion Matrix

                      Predicted Positive       Predicted Negative
Actual Positive       True Positive (TP)       False Negative (FN)  <- Type II Error
Actual Negative       False Positive (FP)      True Negative (TN)
                      (Type I Error)

Metrics Formula & Purpose Cheat Sheet

Accuracy: (TP + TN) / (TP + TN + FP + FN)
- When to use: Only when classes are balanced. Misleading for imbalanced datasets (e.g., 99% legitimate, 1% fraud).
Precision: TP / (TP + FP) — Out of all we predicted positive, how many were right?
- When to prioritize: When the cost of a False Positive (FP) is high (e.g., Spam filter — don’t send crucial client emails to spam).
Recall (Sensitivity): TP / (TP + FN) — Out of all actual positives, how many did we catch?
- When to prioritize: When the cost of a False Negative (FN) is high (e.g., Cancer detection, fraud detection — never miss a sick patient or fraudster).
F1-Score: Harmonic mean of Precision and Recall: 2 × (Precision × Recall) / (Precision + Recall)
- When to prioritize: Imbalanced datasets where you need a balance between Precision and Recall.
ROC-AUC (Receiver Operating Characteristic - Area Under Curve): Measures classifier performance across all classification thresholds. (AUC of 1.0 is perfect; 0.5 is random guessing).

2. Regression Metrics

MAE (Mean Absolute Error): Average of absolute differences. Easy to interpret in original units; treats all errors linearly.
MSE (Mean Squared Error): Averages squared differences. Heavily penalizes large outlier errors.
RMSE (Root Mean Squared Error): Square root of MSE. Penalizes large errors, but brought back into original target units.
R-Squared (Coefficient of Determination): Explains the proportion of variance captured by the model (1.0 = perfect fit, 0 = performs no better than the mean).

3. Business Metrics (vs. Model Metrics)

Exam Tip: The exam distinguishes between technical model metrics and business impact metrics.

Metric Type	Examples	Purpose
Model Metrics	Accuracy, F1, AUC, RMSE	Measure technical ML performance
Business Metrics	Cost per user, Development cost, Customer satisfaction (CSAT), Revenue uplift, ROI	Measure real-world business impact

A model with high accuracy may still fail if the cost to build/run it exceeds the business value it generates (negative ROI). Always evaluate both.

4. The Bias-Variance Tradeoff (Overfitting vs. Underfitting)

        High Bias (Underfitting)             Balanced (Good Fit)             High Variance (Overfitting)
        ─────────────────────────             ───────────────────             ───────────────────────────
        • Model is too simple                • Generalizes well              • Model memorized training data
        • High Train & Test Error            • Low Train & Test Error        • Very Low Train Error, High Test Error

Problem	Symptoms	Root Cause	How to Fix It
Overfitting (High Variance)	High training accuracy, poor validation/test accuracy.	Model is too complex; memorized training noise.	1. Add more training data 2. Regularization (L1/L2, Dropout) 3. Feature reduction / Pruning 4. Early stopping during training
Underfitting (High Bias)	Poor accuracy on both training and test data.	Model is too simple; cannot capture patterns.	1. Use a more complex model/architecture 2. Add more relevant features 3. Train for more epochs / reduce regularization

Module 7: Neural Networks & Deep Learning Architectures

Foundational deep learning models often serve as the building blocks for Generative AI and Foundation Models.

                                  Neural Networks
                                         |
        ┌────────────────────────────────┼────────────────────────────────┐
        ▼                                ▼                                ▼
  CNN (Convolutional)            RNN / LSTM (Recurrent)          Transformer (Attention)
  • Grid/Spatial Data            • Sequential/Time Data          • Parallelized Sequences
  • Images, Video, OCR           • Speech, Historical Series     • LLMs, Modern GenAI

CNN (Convolutional Neural Network): Uses convolutional filter matrices to process grid/spatial data.
- Exam Association: Computer Vision, Image Classification, Object Detection (Amazon Rekognition).
RNN / LSTM (Recurrent Neural Network / Long Short-Term Memory): Processes sequential data by passing hidden states step-by-step.
- Exam Association: Legacy NLP, speech-to-text, time-series forecasting.
Transformers (Self-Attention Mechanism): Processes entire sequences in parallel using attention weights to capture context.
- Exam Association: Modern Foundation Models (FMs), Large Language Models (LLMs), Amazon Bedrock, Amazon Titan, Claude, GPT.

Module 8: MLOps — Operationalizing ML

MLOps applies DevOps principles to the ML lifecycle to make model development repeatable, scalable, and production-ready.

MLOps Concept	Description
Experimentation	Tracking model versions, hyperparameters, and metrics across runs to compare results
Repeatable Processes	Automated, version-controlled pipelines that produce consistent results
Scalable Systems	Infrastructure that scales training and inference with demand
Managing Technical Debt	Avoiding shortcuts (e.g., hard-coded thresholds) that cause future maintenance problems
Production Readiness	Ensuring models pass validation gates before serving real traffic
Model Monitoring	Continuously tracking a deployed model’s performance for degradation
Model Retraining	Triggering a new training run when performance drops below a threshold

Data Drift vs. Concept Drift

Drift Type	Definition	Example
Data Drift	The distribution of input features (X) changes over time	Customer age range shifts after targeting a new demographic
Concept Drift	The relationship between features (X) and target (y) changes over time	Fraud patterns change after fraudsters adapt to detection

Detection: Amazon SageMaker Model Monitor detects both data drift and model quality degradation automatically.

Module 9: AWS AI/ML Ecosystem Mapping

Match each ML concept directly to its managed AWS service:

ML Lifecycle Stage / Need	AWS Service	Key Role
Data Labeling	Amazon SageMaker Ground Truth	Managed human workforce + active learning automated labeling.
Data Prep & Feature Engineering	Amazon SageMaker Data Wrangler	Visual interface to clean, normalize, and engineer features with 300+ built-in transforms.
Feature Storage & Reuse	Amazon SageMaker Feature Store	Central repository to store, share, and serve features for training and real-time inference.
Automated ML (No-Code/Low-Code)	Amazon SageMaker Autopilot	Automatically cleans data, selects algorithms, trains, tunes hyperparameters, and outputs full code notebooks.
No-Code ML for Business Analysts	Amazon SageMaker Canvas	Drag-and-drop visual ML for non-technical business users; no code required.
Bias Detection & Explainability	Amazon SageMaker Clarify	Detects pre/post-training bias and provides feature attribution (SHAP values) for explainability.
Model Monitoring & Drift	Amazon SageMaker Model Monitor	Monitors deployed endpoints for data drift, concept drift, and model quality degradation.
Computer Vision	Amazon Rekognition	Image and video analysis — object detection, face recognition, content moderation.
NLP / Text Analytics	Amazon Comprehend	Sentiment analysis, entity recognition, language detection, topic modeling.
Speech-to-Text	Amazon Transcribe	Converts audio to text with speaker identification and medical vocabulary support.
Text-to-Speech	Amazon Polly	Converts text to lifelike speech in multiple languages.
Machine Translation	Amazon Translate	Neural machine translation between 75+ languages.
Conversational AI	Amazon Lex	Build voice and text chatbots (same engine as Alexa).
Recommendations	Amazon Personalize	Real-time personalized recommendations using your own data.
Fraud Detection	Amazon Fraud Detector	Managed service for detecting online fraud using ML.
Forecasting	Amazon Forecast	Time-series forecasting using ML (no ML expertise required).
Document Analysis	Amazon Textract	Extracts text, forms, and tables from scanned documents.
Reinforcement Learning Education	AWS DeepRacer	Cloud-based 3D racing simulator and 1/18th scale physical race car for practicing RL.

Module 10: High-Yield Exam “Versus” Trap Matrix

Review this table right before sitting exam questions:

Concept A	vs.	Concept B	How to Distinguish on the Exam
AI	vs.	ML	AI is the broad field of intelligent systems; ML is a subset that learns from data.
ML	vs.	Deep Learning	Deep Learning is a subset of ML using multi-layer neural networks; requires much more data and compute.
Supervised	vs.	Unsupervised	Supervised requires labels/ground truth (y); Unsupervised works on raw, unlabelled data (X).
Classification	vs.	Regression	Classification outputs a category/class; Regression outputs a continuous number.
Precision	vs.	Recall	Prioritize Precision to avoid False Alarms (FP); prioritize Recall to avoid Missing Positives (FN).
Hyperparameters	vs.	Model Parameters	Hyperparameters are set before training (learning rate, batch size, epochs); Parameters are learned during training (weights, biases).
Data Drift	vs.	Concept Drift	Data Drift: Input feature distribution (X) changes. Concept Drift: The X→y relationship changes.
Model Metrics	vs.	Business Metrics	Model metrics (F1, AUC) measure ML performance; Business metrics (ROI, cost/user) measure real-world value.
SageMaker Autopilot	vs.	SageMaker Canvas	Autopilot: AutoML with full code access (data scientist-focused). Canvas: No-code visual drag-and-drop (business analyst-focused).
Real-Time Inference	vs.	Batch Inference	Real-time: single requests, low latency. Batch: large bulk datasets, scheduled, non-urgent.
Pre-trained Model	vs.	Custom-trained Model	Pre-trained: reuse existing model (fast, less data needed). Custom: trained from scratch on proprietary data (more control, more effort).
Labeled Data	vs.	Unlabeled Data	Labeled: has target y, required for Supervised Learning. Unlabeled: no target, used in Unsupervised Learning.

Capabilities of Generative AI

Capability	Short explanation
Adaptability	Can adapt responses to different tasks, contexts, instructions, and domains without being specifically programmed for each one.
Responsiveness	Can quickly generate relevant responses to user inputs, often enabling interactive, conversational experiences.
Simplicity	Makes complex tasks easier by allowing users to interact with AI using natural language rather than specialized interfaces or code.
Creativity and exploration	Can generate new text, images, code, ideas, designs, and alternatives, supporting brainstorming and experimentation.
Data efficiency	Can perform useful tasks with relatively little task-specific training data, especially through prompting or few-shot learning.
Personalization	Can tailor outputs to individual users based on their context, preferences, history, or provided information.
Scalability	Can serve many users and handle large volumes of content/tasks through cloud-based infrastructure and automated inference.

Challenges of Generative AI

Intellectual Property
Hallucinations
Toxicity
Plagiarism and Cheating
Disruption of Nature of Work

Challenge	Short explanation
Regulatory violations	AI outputs or use of data may violate laws, regulations, intellectual-property requirements, or industry rules.
Social risks	AI can contribute to bias, discrimination, misinformation, job impacts, or loss of trust if used irresponsibly.
Data security and privacy concerns	Sensitive or proprietary information can be exposed, misused, or improperly processed if data and access are not adequately protected.
Toxicity	Models may generate harmful, offensive, abusive, or inappropriate content.
Hallucinations	The model may confidently generate information that is false, fabricated, or unsupported by facts. (AWS Documentation)
Interpretability	It can be difficult to understand why a model produced a particular output or decision, making transparency challenging. (AWS Documentation)
Nondeterminism	The same or similar input can produce different outputs because generative AI is probabilistic rather than guaranteed to produce identical results.

Responsible AI

AWS Docs/Blogs worth reading on Responsible AI and AI Safety:

Responsible AI - AWS Docs

Responsible AI - AWS Detailed Blog - Considerations for addressing the core dimensions of responsible AI for Amazon Bedrock applications

Fairness: Considering impacts on different groups of stakeholders.
Explainability: Understanding and evaluating system outputs.
Privacy and Security: Appropriately obtaining, using, and protecting data and models.
Safety: Reducing harmful system output and misuse.
Controllability: Having mechanisms to monitor and steer AI system behavior.
Veracity and Robustness: Achieving correct system outputs, even with unexpected or adversarial inputs.
Governance: Incorporating best practices into the AI supply chain, including providers and deployers.
Transparency: Enabling stakeholders to make informed choices about their engagement with an AI system.

Fairness

When your AI system interacts with customers, those interactions reflect your brand values. Consider a mortgage application system: it must evaluate applications based on relevant financial criteria while verifying that decisions aren’t influenced by unwanted discriminatory factors. This means implementing robust testing frameworks to detect potential bias, regularly auditing outcomes across different customer segments, and maintaining clear documentation of decision criteria. Leading organizations are integrating these considerations into their development processes, which means that their AI systems enhance rather than compromise their commitment to equitable service.

Explainability

The ability to understand and communicate how AI makes decisions is both a good practice and is essential for business operations. A wealth management AI advisor must be able to articulate the reasoning behind its investment recommendations. This requires implementing interpretability techniques and developing frameworks to translate complex model decisions into understandable explanations for both customers and regulators. Organizations that excel here find they build deeper customer trust and navigate regulatory requirements more effectively.

Privacy and Security

Customers want to trust generative AI applications with their most sensitive information. This trust is earned by strong application security and data privacy controls. When implementing AI systems, this trust must be preserved through robust data protection and infrastructure security mechanisms. Leading organizations are implementing sophisticated data governance frameworks that include encryption, access controls, and data minimization practices. They also develop clear policies about data usage, verify that AI systems access only the information necessary for their specific functions, and seek to maintain regulatory compliance.

Safety

AI systems must operate within a clearly defined use case scope that aligns with your organization’s risk tolerance and values. Consider a trading recommendation system: it needs guardrails to remove suggestions that could violate regulatory requirements or exceed risk thresholds. Forward-thinking organizations are implementing comprehensive safety frameworks that include content filtering, output validation, and clear escalation paths for edge cases.

Controllability

The ability to monitor and adjust AI system behavior aligns with business objectives and risk parameters. Leading organizations implement robust monitoring systems that track performance metrics, user feedback, and system outputs. They maintain clear procedures for adjusting or disabling AI systems when necessary, keeping human oversight effective even as systems scale.

Veracity and Robustness

AI systems must deliver reliable, veratious results consistently, even when facing unexpected situations. Organizations at the forefront of AI adoption are implementing comprehensive testing frameworks that challenge their systems with diverse inputs, monitoring accuracy across different scenarios, and maintaining clear protocols for handling edge cases. They’re building systems that not only perform well in ideal conditions but remain reliable under stress. At the forefront of this field are concepts like automated reasoning, which use mathematically provable statements to capture and correct hallucinations in real-time.

Governance

Clear governance frameworks align AI systems with organizational policies and regulatory requirements. Leading organizations are establishing AI governance committees that include technical, business, and risk management perspectives. They’re developing comprehensive documentation practices, clear escalation paths, and regular review processes so that AI systems continue to serve business objectives while managing risk effectively.

Transparency

Building trust requires openness about AI system capabilities and limitations. Successful organizations clearly communicate when and how AI is being used, what data informs decisions, and what controls are in place. This commitment to transparency enhances user trust in the AI system, encouraging adoption.

Cheat Sheets, AWS

AWS Cheat Sheet

This post is licensed under CC BY 4.0 by the author.