What Is Machine Learning?

The Old Way: You Write the Rules

For the last sixty years, programming meant one thing: you write the rules.

Say you want a program that checks if someone is old enough to buy alcohol. You already know the rule, so you just type it in:

if age is 18 or older → allowed else → denied

You wrote the rule. The computer only follows it. That’s the whole shape of traditional programming:

Rules (written by you) + Data go in → Answers come out.

This works beautifully, as long as you can actually state the rule.

A Problem That Breaks It

Now try this one:

Write the rules to tell a cat from a dog in a photo.

Go ahead. Every rule you reach for falls apart:

• “Cats have pointy ears” → so do German Shepherds
• “Dogs are bigger” → meet the Chihuahua
• “Cats have whiskers” → so do dogs
• “Dogs have long snouts” → not pugs

For every rule you write, reality has an exception. You could spend your whole life stacking if statements and never finish.

And yet a four-year-old does this instantly. Nobody handed that child a rulebook. They just saw a lot of cats and a lot of dogs, and the pattern formed on its own.

Some knowledge lives in examples, not in rules you can write down.

So We Flip It Around

Machine learning is that exact flip. Instead of writing the rules, you do this:

You show the computer thousands of examples (this photo is a cat, this one is a dog), and the computer figures out the rules itself.

Look carefully at what just swapped places:

That is the entire reframe:

1. In traditional programming, you write the program.
2. In machine learning, you write a program that writes the program by studying examples.

The rules were the thing you used to provide. Now they are the thing you get back.

What “Learning” Actually Means

Here is the part that sounds mystical and absolutely should not.

Inside every model is a big pile of numbers. We call them parameters. When training starts, they’re set randomly, so the model is useless: show it a cat, it confidently blurts “dog.”

So how does it improve? Every time it guesses wrong, we nudge those numbers a little in the direction that would have made the guess less wrong. Then we do it again. And again. Across millions of examples.

• Start: random numbers → mostly wrong
• Each mistake → nudge the numbers slightly
• After enough nudges → the numbers settle into values that just work

Learning is nothing more than slowly adjusting a pile of numbers until the model stops being wrong so often.

There is no magic, no understanding, no spark. Just numbers being tuned toward less wrong.

(How does it know which direction to nudge? That’s a beautiful idea called gradient descent, and it’s the very next note.)

The One Idea That Never Leaves

Hold on to this, because it is the thread through everything ahead.

The cat-or-dog classifier and a frontier model like the ones powering today’s AI assistants are the same loop:

A pile of numbers, getting nudged toward “less wrong,” one example at a time.

We never abandon this idea. From here on, we only do two things to it:

• Make the pile of numbers bigger (from a handful to trillions)
• Make the nudging smarter (better math, better data, better feedback)

That’s the journey. Everything else is detail.