Learning From Examples
Modern LLMs are built with machine learning. You do not need the full theory to build with them — but a little ML literacy makes every earlier decision sharper. This module is a bridge, not a full ML course.
Supervised Learning
The most common shape of machine learning — and the one most classic models use.
- Learn from labeled examples — each example pairs an input with its correct output.
- Map inputs to answers — it is how a model learns to turn an input into the right output.
- Most classic ML is this — spam-or-not, price prediction, image labels.
Reinforcement Learning
A different idea: learn by trial and reward. An agent tries actions and is rewarded for good outcomes, so over time it favours what works.
Where LLMs Fit
Large language models draw on both shapes.
- Trained on huge text — they first learn from vast amounts of text (self-supervised).
- Refined with human feedback — then tuned so their answers are more helpful (an RL-style step).
- You do not need the math — just this mental model of how they came to be.
Why It Matters
Knowing these two shapes helps you reason about what a model can actually learn — when it is worth fine-tuning (Module 5), and why data quality dominates almost everything else.
Build It
How to implement: for one problem in your tool, ask — is this a supervised-learning task (map an input to a label), or something an LLM prompt already solves? Often prompting wins; the skill is knowing when it does not.
- Weekly AI Tasks tracker — turning a message into a task is a mapping problem; an LLM prompt handles it, but with thousands of labeled examples a small supervised classifier could be cheaper.
- Personal brand site — pure generation; no training needed.
What you learned
Supervised learning maps labeled inputs to outputs; reinforcement learning learns from reward. LLMs use both. You do not need the math — just the mental model, so you can judge when to prompt, when to fine-tune, and why data quality wins.