What Does It Mean For a Machine to "Learn"?
A few weeks back, I was passing through a small town and stopped near a tea stall. I don't drink tea myself, never have, but I ended up standing there for a few minutes just watching the stall owner work. He was making cup after cup, for one customer after another, and every single time, without measuring anything, without checking a recipe, the tea came out exactly the way that customer wanted it.
One person asked for it less sweet. Another wanted it strong. He just nodded, and somehow got it right every time.
I got curious and asked him how he manages this without ever measuring anything. He laughed and said something like: in the beginning, he used to get it wrong all the time. Now he can just look at a customer's face and tell.
That one line stuck with me. Because what he described, without knowing it, is almost exactly what happens inside something called machine learning.
So What Actually Is "Learning" For a Machine?
When we say a person is learning something, we usually mean they are getting better at a task over time, through experience. They try something, see how it turns out, and adjust for next time.
A machine "learning" something means almost the same thing. A computer program is given some data, which acts as its experience, and it slowly adjusts itself to get better at a task based on that data. No robot brain, no mystery. A program starts out bad at something, and keeps adjusting until it gets good at it.
The real question is: adjusting what, exactly?
Every Model Has Something Like a Slider
Think about the brightness slider on your phone. You drag it up, the screen gets brighter. You drag it down, it gets dimmer. Somewhere in between, it feels just right for the room you're sitting in.
A machine learning model works in a similar way. Inside it, there are numbers that control how it behaves. We call these numbers parameters, and you can think of them as sliders.
When a model is created, its parameters are typically set to random or near-random starting values, so it behaves badly at first, kind of like a screen brightness set to the wrong level for the room you're in. Training is the process of automatically adjusting those parameters based on how wrong the model's predictions are, until the results stop being bad and start being useful. Nobody is reaching in and nudging one slider at a time by hand — the model adjusts itself, guided by its own errors.
The tea stall owner had his own sliders too, even if he never called them that. Sugar amount. Tea leaf quantity. Milk quantity. Boiling time. Over the years, without writing any of it down, he adjusted every one of these based on what customers told him. A machine does the same thing, except its sliders are just numbers, and its "customers" are the data we feed it.
Where Does This "Data" Come From?
Let's make this even more concrete with something you use almost every day — food delivery apps.
Say a food delivery app wants to predict how long your order will take to arrive. To teach a model to do this, they don't sit down and write rules like "if distance is 3 km, then time is 20 minutes." Instead, they show the model thousands of past orders. For each one, they already know:
How far the restaurant was from your house
What time of day it was
How many items were in the order
And most importantly — how long that delivery actually took
This past record — the inputs, and the correct answer that already happened — is called training data.
Here's roughly what that adjustment looks like in practice. Early on, the model might see a 3 km order placed during lunch rush and predict 18 minutes. The order actually took 31. That's a big miss, so the model's parameters shift. Next time it sees a similar order, it predicts 24. Still short, but closer. Then 28. Then 30. It never "understands" traffic or lunch rush the way you do — it's just narrowing the gap between its guess and reality, one adjustment at a time, across thousands of examples.
Once it has adjusted enough, you can hand it a brand new order it has never seen before, and it will give you a reasonable guess for the delivery time. That guess is the entire point of the exercise.
Two Different Ways a Model Can Learn
Not every machine learning problem looks the same. There are several categories in practice, but two of the most important — and the ones you'll run into constantly as a beginner — are supervised and unsupervised learning.
Sometimes, the model is given the correct answer while it learns. Think of a student practicing from a book that has an answer key printed at the back. The student solves a question, checks the answer key, sees how far off they were, and adjusts. Do this with enough questions, and the student genuinely improves. This is called supervised learning, and the delivery time example above falls right into it, since for every past order, we already know exactly how long it took.
Other times, there is no answer key at all. Imagine someone hands you five hundred random photos from your gallery and says, "sort these however you like, I won't tell you the categories." You would probably still manage to group them somehow — maybe by the people in them, maybe by screenshots versus actual photos, maybe by color. Nobody told you what the "correct" groups were. You just found some structure on your own. This is called unsupervised learning, and you'll see it again later in this series.
The thing worth remembering here is the distinction itself: does the data come with correct answers attached, or does the model have to find patterns entirely on its own?
The Part That Surprised Me
A machine learning model does not actually understand anything, not the way a person does. It cannot tell you why a longer distance means a longer wait. It has no idea what a restaurant even is. All it is really doing, underneath everything, is adjusting sliders so its output lands closer to being correct.
That might sound underwhelming. Just sliders being nudged around? But this one idea, repeated enough times, on enough data, is powerful enough to recognize a face in a photo, translate a sentence, recommend a movie you end up actually liking, or hold a conversation with you. Most of what feels like "intelligence" in these systems comes down to that same loop, running at a scale most of us never sit down and think about: guess, compare with the correct answer, adjust a little, try again.
Where We Go From Here
Right now, this probably still feels a bit abstract. Adjusting sliders, getting closer to the correct answer — what does that actually look like with real numbers, on a real problem?
In the next post, we'll take the simplest version of this idea and build it from scratch. We'll take something you can picture easily, a straight line, and teach a machine to draw that line itself, purely by adjusting its sliders based on data.
It might sound too simple to even count as machine learning. But once you see how this one works underneath, every model you learn after this — no matter how advanced or intimidating it sounds on paper — will feel like a variation of the exact same trick you're about to see.
See you in the next one.
