Syllabus Lesson 1 of 246 · Module 0: Start line
Module 0: Start line

How this course works, and your first prediction

Why it matters Most people learn neural networks by watching someone else's code run. Here you build every piece, and you predict what will happen before it does. Predicting first is what turns watching into understanding.

Before you start

  • What you need. School maths (the course teaches every slope, log and probability it uses) and some Python: loops, functions, lists and dictionaries. Nothing to install: everything runs in this browser.
  • How long. Part 1 (Build and Train, Modules 0 to 14) is about 90 hours of work, about 80 if you skip the Python module: at 6 hours a week, about 15 weeks. Part 2 (Generate, Scale and Understand, Modules 15 to 24: generative models, diffusion, self-supervised learning, scaling laws, adapting and looking inside models, judging claims, a capstone and the final exam) adds about 50 hours, 9 more weeks. Each part has its own certificate. Each module shows its own time on the course page.
  • Already fluent in Python? Module 1 ends with a gate. Take it first: passing it skips the module, and you go straight to the maths.
  • Experienced programmer? The times shown are for someone new to the subject. People who already code for a living often need a third of them, most of all in the maths and first network modules (2 to 4): read each lesson's exercise first, and slow down only where it is new to you.
  • Already know a module's topic? Modules 1 (Python), 2 (maths), 3 (neuron to network), 4 (autograd) and 7 (new data) each end with a gate you may take first, with the AI tutor off. Passing it marks the module done and opens the next one; the certificate counts the gate for that module.

Four habits you will practise

  • Call the Curve. Before a training run, you guess its loss at a few points. Then it runs and you see how close you were. Over the course your guesses get sharper; that sharpening is real intuition.
  • Earn the Layer. The course has a small library of ready-made layers. Each one stays locked until you have built it yourself in plain NumPy and it has passed a gradient check. After that, you use the fast version.
  • The FLOP Ledger. A FLOP is one multiply or add. Every run in the course reports how many it used, so you learn what is cheap and what is expensive.
  • Four dials. Your progress is measured on four skills: Build (write it), Diagnose (find what is broken), Predict (say what will happen) and Trust (judge when an AI answer is wrong).

Your first Call the Curve

Below, a single neuron learns the line y = 3x + 1 by gradient descent. The loss starts high and falls. Guess the loss at step 0, 5, 20 and 100, then run the code: it trains and prints your guess beside the truth, with a score. Do not worry about being right. Make a real guess.

Loss: one number that says how wrong the model is right now. Training is the job of making it smaller.

Words and symbols in this lesson
loss
one number saying how wrong the model is; training makes it smaller
Call the Curve
guessing the loss at a few points before a run, then comparing
FLOP
one multiply or add; how the course measures the cost of a run
four dials
Build, Diagnose, Predict and Trust: the four skills the exams measure
Your turn

Replace each None in my_guess with your guess for the loss at steps 0, 5, 20 and 100. Then run. Any honest guess passes; the score is for you.

Spotted a problem in this lesson? Report it · Ask about this lesson

Code · runs in your browser
Output

        

Your progress is counted anonymously. Change this