What Is Machine Learning?
On her first day at Plateful — a food-delivery app, and the company we will follow through this whole course — Iris, the new product manager, hears an engineer say the app predicts when your order will arrive. She asks the obvious question: who wrote the prediction rules? Rush hour, rain, a slow kitchen — someone must have sat down and written all of that logic. The answer she gets is the reason this course exists: nobody did.
That is machine learning: instead of a person writing the instructions, the computer works out its own rules from many examples of the past. Every other page in this book stands on that one idea, so this page's only job is to make it solid.
Software Before Learning
Almost all the software you use is the first kind. A person thought through the logic — if the cart is over thirty dollars, delivery is free; if the password is wrong three times, lock the account — and wrote it down as instructions. The computer follows those instructions exactly, every time, and does nothing it was not told to do.
It is worth saying clearly: this is not the old-fashioned way that machine learning replaced. It is still how most software is built, and for most jobs it is the right way. When a rule can be written down, written rules are exact, cheap, and easy to check.
The Problem Rules Can't Solve
But some tasks refuse to become rules. Try to write the instructions that decide whether a photo shows a cat. Pointed ears? So has a fox. Fur? Not in a cartoon. Whiskers? Sometimes hidden. Every rule you write has a thousand exceptions, and the exceptions have exceptions. Yet a four-year-old does the task instantly — not because she knows the rules, but because she has seen a lot of cats.
Plateful has the same problem in less cute form. Which reviews are fake? A human moderator develops a feel for it after reading a few thousand — something about the wording, the timing, the pattern. Ask her to write her feel down as exact rules and she can't. Nobody can. That is the signature of the tasks machine learning was invented for: humans do them by feel, built from experience, and there is no checklist underneath to copy out.
Learning from Examples
So instead of writing rules, we do what experience does. Think of learning to pick a ripe melon at the market. Nobody can hand you a reliable written rule for it — but let an experienced neighbor show you fifty ripe melons and fifty unripe ones, and you start getting it right. You could not say exactly what tipped you off; the knack formed on its own, out of examples.
Machine learning gives computers that same path. Show the computer ten thousand delivery records — the distance, the hour, the restaurant, and how long each one actually took — and let it find the patterns connecting the details to the outcome. The patterns it finds become the program. From here on we will call that finished thing a model, and the next pages make it concrete; for now, what matters is where it came from: examples, not instructions.
What "Learning" Really Means Here
The word "learning" carries a lot of baggage, so let's unpack exactly what happens and no more. During learning, the computer adjusts internal numbers — thousands of little dials — until its predictions match the examples it was shown. That is the whole event. Nothing understands anything; no insight arrives; there is no little mind in there developing opinions about traffic.
The honest picture is the melon knack, not a student having a breakthrough. The result is real and useful — the delivery predictions genuinely work — but it is a learned pattern-matcher, nothing more mysterious. Keeping that picture will protect you later, when the machines start writing poetry and it becomes very tempting to imagine someone is home.
- "Machine learning means the computer is intelligent." It means the computer found patterns in examples. The result is a knack, like your melon-picking — genuinely useful, with no understanding behind it.
- "ML is replacing normal programming." Most software is still hand-written rules, and should be. ML is the tool for the specific tasks where rules cannot be written down.
- "The computer learns by itself, from nothing." It learns only from examples that people collected, chose, and often labeled by hand. No examples, no learning — a fact so important it gets its own chapter (Chapter 2).
- "Learning continues while the software runs." Almost never. The patterns are found once, in a separate step, and then used. Topic 04 makes this concrete.
- Written rules versus learned rules is the single distinction that unlocks every AI headline, product pitch, and data-team conversation you will meet — including all the ones in this book.
- Knowing which kind of task calls for learning — the ones humans do by feel, with no writable checklist — is the first real judgment call of the field, and you can already make it.
Knowledge Check
What is the core difference between machine learning and traditional software?
- In ML the computer finds its own rules from examples, not ones a person wrote
- Machine learning programs run much faster than programs written by hand
- ML requires special computers, while traditional software runs on any machine
- In ML the computer actually understands the task instead of blindly following orders
Which of these tasks is the strongest candidate for machine learning?
- Giving free delivery on orders over thirty dollars
- Deciding whether a restaurant review is fake
- Locking an account after three wrong password attempts
- Calculating the sales tax on an order
During "learning", what actually happens inside the computer?
- It memorizes all the examples so it can look them up later
- It gradually comes to understand the task the way an experienced person would
- It adjusts internal numbers until predictions match the examples shown
- It writes out a list of rules that a programmer then checks
Why couldn't Plateful's engineers just write down the rules for predicting delivery time by hand?
- Regulations forbid hand-written rules in delivery apps
- Writing the rules would simply take too much effort, so the team took a shortcut
- The pattern has too many interacting factors to state as explicit rules
- They had no records of past deliveries to base the rules on
You got correct