You already know: how to fit weights by minimizing a loss with gradient descent (Optimization), cross-entropy for classification (Information theory §3), and a learned model beating a hand rule (Bridge D2 → D3).
Why this matters for Atlas. "Will this student quit?", "How long will this answer take to generate?", "Is this message spam?" — each is a question with known answers in historical data. Supervised learning turns those examples into a function that answers new cases. Even in the LLM era, these models run fraud checks, routing, ranking and monitoring everywhere — and they're the baseline every deep model must beat.
1. The setup: features, labels, and a function to learn
The question. What exactly does "learning from examples" mean?
Start from nothing. We have n examples. Each has: