Supervised, unsupervised, and reinforcement learning

Before picking any technique, one question matters more than any other: what shape is the data you actually have? Do you have labeled correct answers for every example? No labels at all, just raw data? Or no fixed dataset at all, just a way to tell whether an outcome was good or bad? Each answer points to a genuinely different category of machine learning.

Supervised learning is what the previous lesson described: every training example comes with the correct answer attached, and the model learns to map inputs to those answers.

This is by far the most common category in real production systems, precisely because labeled data — historical examples where you already know the right answer — is exactly what most businesses already have sitting in their own databases (past transactions, past support tickets, past purchases).

Unsupervised learning works on data with no correct answers provided at all — the goal is finding structure in the data itself, not predicting a known label.

Reinforcement learning doesn't use a fixed dataset of examples at all — an agent takes actions in an environment, receives a reward signal (a number indicating how good or bad the outcome was), and gradually learns a policy (a strategy for choosing actions) that maximizes reward over time, through repeated trial and error rather than being shown correct answers directly.

This is the paradigm behind game-playing AI (chess, Go, video games — there's no dataset of "correct moves," just win/lose outcomes to learn from) and, significantly, a key technique behind how modern large language models are fine-tuned to be more helpful and less harmful after their initial training — RLHF (Reinforcement Learning from Human Feedback) uses human preference ratings between different model outputs as the reward signal, directly connecting this lesson to the LLM-specific lessons ahead in this domain.

If someone tells you a feature "uses AI," the honest next question is which kind, trained on what — because it changes everything about how to evaluate and trust it. A supervised classifier is only as good as the labeled examples it was trained on, and inherits any bias present in that labeling. An unsupervised clustering result has no "correct" grouping to check against — evaluating it is inherently more subjective. A reinforcement-learned policy can behave unpredictably in situations meaningfully different from what it was trained against, since it optimized for reward in its training environment specifically, not for correctness in the abstract.