Rules versus learning: what ML actually is

Everything you have programmed so far is RULES: if marks >= 80, grade A. You wrote the logic; the machine executed it. Machine learning inverts this: you supply EXAMPLES — inputs paired with correct answers — and the machine finds the rule itself. Show it a thousand emails labelled spam/not-spam and it learns what spam looks like; show it a thousand student records labelled completed/dropped-out and it learns the shape of dropout risk.

When does learning beat writing rules? When the rule is real but too subtle or too shifting to write down. Nobody can hand-write the rule for 'this photo contains a cat' or 'this transaction is fraudulent' — but examples of each exist in bulk, and patterns hide in them. When a clear rule EXISTS — fee overdue if unpaid past 30 days — writing it remains correct, faster, and explainable; using ML where an if-statement suffices is a beginner disease and, worse, an unexplainable one: the if-statement can be audited, the model must be trusted. The professional question is never 'can we use ML here?' but 'do examples exist, is the pattern real, and is a rule impossible?'.

What ML genuinely cannot do, stated early because the hype will not state it: it cannot find patterns that are not there (a model trained on lottery numbers learns noise, confidently); it cannot exceed its examples (trained on Lahore data, it knows nothing of Quetta — the generalisation problem, lesson 4's whole subject); and it does not know WHY — it finds correlations, not causes, and treating one as the other is how expensive mistakes get automated. This academy's own research history is a standing exhibit: dozens of plausible trading patterns, tested honestly against held-out data, mostly dissolved into noise — and the discipline that caught them (which you will learn in lesson 4) is worth more than any single model.

The field's honest map, so the vocabulary lands: supervised learning — learn from labelled examples (this course); classification predicts a category (spam / dropout / fraud), regression predicts a number (price, demand). Unsupervised learning finds structure without labels (customer groupings). And the LLMs of ai-fundamentals are this same machinery at colossal scale — text in, next-word predictions learned from the internet. One family, one set of honesty rules.