00 · What is machine learning

Machine learning video

Nobody told this line where to go. It looks at 40 houses, guesses prices, measures how wrong it was, and nudges two numbers. After 2,000 tries it has learned a rule for pricing houses, without anyone writing that rule.

Watch: YouTube · TikTok · Instagram (35 seconds)

The idea

Normal programming: you write the rules. Machine learning: you show examples and let the program find the rule.

Here the rule is a straight line, price = w × size + b. The model is just two numbers, w (how steep the line is) and b (where it starts). Training means adjusting those two numbers until the line sits close to the examples.

Run it

python3 src/learn.py

price = 3.33 * size + 2.44

Files

00-what-is-machine-learning/
├── data/
│   └── houses.csv        40 houses: size and price
├── scripts/
│   └── make_houses.py    made houses.csv (you don't need to run it)
└── src/
    ├── learn.py          the 8 lines from the video
    └── houses.py         loads houses.csv into x and y

Data and code each get their own folder, so you can swap in different data without touching the model.

make_houses.py made the 40 examples. The prices follow 3 × size + 4 plus some random noise, the way real data is never perfectly on a line. It uses a fixed random seed, so running it again gives the same 40 houses.

The code

src/learn.py, line by line:

Line	Code	What it does
1	`import numpy as np`	NumPy lets us do math on all 40 houses at once.
2	`from houses import x, y`	`x` is the size of each house, `y` its real price.
3	`w, b = 0.0, 0.0`	The model starts clueless: a flat line at zero.
4	`for epoch in range(2000):`	Repeat the lesson 2,000 times. One pass over the data is an epoch.
5	`err = w * x + b - y`	Guess every price, subtract the real price. Positive means we guessed too high.
6	`w -= 0.01 * (err * x).mean()`	Nudge the slope against the error. Big houses with big errors push hardest.
7	`b -= 0.01 * err.mean()`	Nudge the starting point against the average error.
8	`print(...)`	Show the rule it learned.

The 0.01 is the learning rate: how big each nudge is. Lines 5 to 7 are gradient descent, covered properly in lesson 01.

Why the answer isn’t exactly 3 and 4

The data was made from 3 × size + 4, but the model learns 3.33 × size + 2.44. Two reasons: the noise moves the best-fitting line a little, and 2,000 small steps is not quite enough for b to settle. Try more epochs and watch it move.

Try this

Change range(2000) to range(200). How far off is the line?
Use the learned rule to price a house of size 7.6 (the “house it has never seen” in the video).
Change the learning rate to 0.05, then 0.1. One of them breaks. Why?
Add print(epoch, w, b) inside the loop for the first 10 epochs and watch the numbers move.

Next: 01 · Gradient descent · All lessons