1.1 Basic AI concepts and terminology

Key terms

• AI: Any technique that lets machines perform tasks that normally need human intelligence.
• ML: A subset of AI where systems learn patterns from data instead of following hand-written rules.
• Deep learning: ML using neural networks with many layers. Best for unstructured data (images, audio, text).
• Neural network: Layers of connected nodes whose weights are adjusted during training.
• Computer vision: AI that interprets images and video (Amazon Rekognition).
• NLP: AI that understands or generates human language (Amazon Comprehend).
• Algorithm vs model: The algorithm is the method (e.g., XGBoost). The model is the trained artifact you get after running it on data.
• Training vs inference: Training learns from data and is expensive. Inference uses the trained model to predict on new data.
• Bias: Systematic error — either unfair skew against a group, or a model too simple to capture the pattern.
• Fairness: The model treats demographic groups equitably.
• Fit: Overfitting = great on training data, poor on new data (high variance). Underfitting = poor on both (high bias). Good fit = performs well on both.
• LLM: A large transformer model trained on huge text corpora that predicts the next token.
• GenAI: AI that creates new content: text, images, audio, video, code.
• Agentic AI: A system where a model plans steps, calls tools or APIs, and takes actions toward a goal with some autonomy.

How they relate: AI ⊃ ML ⊃ deep learning ⊃ GenAI. Agentic AI sits on top of GenAI: the model is the brain, and tools, memory, and orchestration let it act. A rule-based system ("flag transactions over $10,000") is automation, not ML, because nothing is learned.

Types of inference (SageMaker AI deployment options)

Type	Choose it when	Traits
Real-time	Instant responses, steady traffic	Always-on endpoint, low latency, pay while running
Serverless	Intermittent or unpredictable traffic	Scales to zero, possible cold starts
Asynchronous	Large payloads or long processing	Requests queued, results delivered later, can scale to zero
Batch	Score a whole dataset offline	No persistent endpoint, cheapest for bulk

Clue words: "milliseconds" → real-time; "sporadic traffic, minimize cost" → serverless; "large files, minutes to process" → asynchronous; "nightly, entire dataset" → batch.

Types of data

• Labeled vs unlabeled: Labeled data includes the answer ("this email = spam"); unlabeled is raw.
• Structured / tabular: Rows and columns. Unstructured: text, images, audio, video.
• Time-series: Values ordered over time (sales, sensor readings); used for forecasting.

Types of learning

• Supervised: Learns from labeled data. Classification predicts a category; regression predicts a number.
• Unsupervised: Finds patterns in unlabeled data — clustering, anomaly detection, dimensionality reduction.
• Reinforcement learning: An agent learns by trial and error from rewards and penalties (games, robotics, AWS DeepRacer).
• Also know: semi-supervised (few labels + much unlabeled data), self-supervised (how LLMs pre-train: labels come from the data itself), RLHF (aligns LLMs using human preference rankings).