AI: Broadest concept — systems that perform tasks requiring human intelligence (language, decisions, predictions).

ML: Subset of AI — models learn patterns from data instead of hard-coded rules (fraud detection, recommendations).

Deep Learning: Subset of ML using multi-layer neural networks (computer vision, NLP, speech).

Generative AI: Focuses on generating new content — text, images, audio, code — often via foundation models.

Supervised: Trained on labeled data — classification (spam/not spam) and regression (price prediction).

Unsupervised: Finds hidden patterns in unlabeled data — clustering, anomaly detection, dimensionality reduction.

Reinforcement: Agent learns via reward/penalty through trial and error — robotics, game playing, autonomous systems.

Tokens: Small chunks of text (words, subwords) that foundation models process as input/output units.

Embeddings: Numerical representations of text/images that capture semantic meaning for similarity search.

Vectors: The mathematical arrays produced by embeddings — used in vector databases for RAG.

Bedrock: Fully managed, serverless API access to pre-trained foundation models (Claude, Titan, Llama) — no infrastructure management.

SageMaker: End-to-end platform to build, train, tune, and deploy custom ML models — includes Studio, Canvas, JumpStart, Clarify.

Prompt Engineering: Crafting inputs (zero-shot, few-shot, chain-of-thought) — no model changes, lowest effort.

RAG: Retrieves external/company data at query time — ideal for knowledge gaps, current facts, private documents.

Fine-Tuning: Retrains model weights on labeled data — for behavior/style adaptation, not knowledge updates.

Zero-Shot: No examples provided — model relies on its training.

Few-Shot: A small number of examples guide the model's response format.

Chain-of-Thought: Prompts model to reason step-by-step before answering.

Role-Playing / Persona: Assigns the model a specific role (e.g., "You are a security analyst").

Accuracy: Overall correctness — can be misleading on imbalanced datasets.

Precision: Of predicted positives, how many were actually positive (minimizes false positives).

Recall: Of actual positives, how many were correctly identified (minimizes false negatives).

F1 Score: Harmonic mean of precision and recall — balanced metric.

ROUGE / BLEU: Text generation quality — summary overlap / translation quality.

Fairness: AI treats all users equitably — mitigates bias in training data and outcomes.

Transparency: Users can understand how and why the AI reached a decision.

Explainability: The model's reasoning can be articulated in human-understandable terms.

Privacy: Personal data is protected — aligns with GDPR, CCPA.

Safety: AI systems operate reliably without causing harm.

Accountability: Clear ownership for AI system outcomes.

Prompt Injection: Malicious input manipulates the model to bypass instructions — focuses on the application layer.

Jailbreaking: Bypassing safety filters to make the model violate guidelines — focuses on the model itself.

Data Poisoning: Corrupting training data to degrade model performance or introduce bias.

Model Inversion: Extracting sensitive training data from a model's outputs.

IAM: Controls access to Bedrock, SageMaker, and AI resources via roles and policies.

GuardDuty: Threat detection — monitors for malicious activity in AWS accounts.

Macie: Discovers and protects sensitive data (PII) in S3 — critical for AI training data.

AWS Config: Tracks resource configuration and compliance for AI workloads.

CloudTrail: Audits API activity — who accessed what AI service and when.

Amazon Q: Generative AI assistant for business — answers questions, generates content, and automates tasks.

Bedrock: Managed API for foundation models — the "engine" behind generative AI applications.

JumpStart: SageMaker hub with pre-trained models and solutions for quick deployment.

Bias: Model makes systematic errors — underfitting, too simple, misses patterns.

Variance: Model is too sensitive to training data — overfitting, poor generalization.

Overfitting: Model memorizes training data instead of learning patterns — performs poorly on new data.

1. Data Collection: Gather and prepare training data.

2. Data Preparation: Clean, label, and split data (train/validation/test).

3. Model Training: Train the algorithm on training data.

4. Evaluation: Assess model performance against validation metrics.

5. Deployment: Deploy the model to production.

6. Monitoring: Track performance and retrain as needed.

Text Generation: Produces coherent text — summaries, articles, dialogue.

Image Generation: Creates images from text prompts (e.g., Amazon Titan Image Generator).

Code Generation: Writes and explains code — Amazon Q Developer.

Multimodal: Processes and generates across text, image, and audio modalities.

Bedrock Guardrails: Runtime controls — content filters, PII redaction, denied topics, grounding checks.

Bedrock Model Evaluation: Pre-deployment assessment — automatic and human evaluation of model quality.

Classification: Predicts a category — spam/not spam, fraud/not fraud (supervised).

Regression: Predicts a continuous value — house price, temperature (supervised).

Clustering: Groups similar data points without labels — customer segmentation (unsupervised).