HACKER TUBES / A GUIDED LEARNING ROUTE
Neural networks: from intuition to micrograd
Follow five selected videos to understand gradients and backpropagation, then finish a working experiment. Includes original guides, exercises, worked answers, and AI tutoring.
For learners who know a little Python and want to understand training. Familiarity with functions, classes, and basic algebra is recommended.
Guides available in English and Simplified Chinese. Original videos are in English.
Understand the forward pass.
Follow the gradient backward.
Make the update yourself.
A clear route
Five videos, ordered from intuition to code.
Original study notes
Preparation, viewing prompts, and runnable examples.
Practice with answers
Check understanding, then complete a small experiment.
10 AI answers
Guide-based tutoring. Failed requests do not count.
YOUR LEARNING ROUTE
Know what to watch next.
- 01Original video ↗
Build an intuition for neural networks
3Blue1Brown - 02Original video ↗
Understand gradient descent
3Blue1Brown - 03Original video ↗
Follow the gradient backward
3Blue1Brown - 04Original video ↗
Work through the chain rule
3Blue1Brown - 05Original video ↗
Put it together with micrograd
Andrej Karpathy
FREE PREVIEW / LESSON 01
Build an intuition for neural networks
A complete sample of the guide, exercise, and worked answer included with every lesson.
Explain how inputs become a prediction, and distinguish a prediction from learning.
Before you watch
Review variables, functions, and multiplication. You do not need to implement a network yet.
Watch with these questions
- Follow one number from the input through a hidden layer to the output.
- Notice which values are data and which values are trainable parameters.
- Pause and explain why stacking only linear operations would still give a linear function.
A neuron is a small computation
For two inputs, begin with z = w1*x1 + w2*x2 + b. The inputs describe an example; the weights and bias are adjustable parameters. An activation function transforms z. A layer repeats this computation with different parameters.
Prediction and training are different steps
A forward pass calculates an output with the current parameters. A loss function scores that output against a target. Training changes parameters to improve the score across examples. A single forward pass does not teach the model anything by itself.
Why use an activation?
A composition of linear maps can be replaced by one linear map. A nonlinear activation, such as ReLU(z) = max(0, z), lets a stack of layers represent more varied functions. This flexibility also makes it possible to fit the training set without generalizing well.
x1, x2 = 2.0, -1.0
w1, w2, b = 0.5, 2.0, 1.0
z = w1*x1 + w2*x2 + b
print(z, max(0.0, z)) # 0.0, 0.0Check your understanding
Keeping the other numbers unchanged, set b = 2.0. What are z and ReLU(z)? Which quantities would training normally adjust?
Show worked answer
z = 0.5*2 + 2*(-1) + 2 = 1, so ReLU(z) = 1. Training normally adjusts w1, w2, and b. The input example is data, not a model parameter.
Before you start
What do I get when I purchase?
Access to all five original lesson guides, exercises, worked answers, and a Markdown download. The final lesson includes a runnable micrograd experiment. Videos remain hosted by their creators.
How do the included AI answers work?
Your 10 answers are shared across the five lessons. Each successfully generated answer uses one, including follow-ups. Failed requests return the reserved answer. Tutoring uses the original lesson guides and exercises, rather than analyzing the video.
What happens when my included answers run out?
Your purchased guides remain accessible. You can continue asking with credits: we show the maximum charge and ask you to confirm before spending. There is no automatic renewal.
Can I watch the videos for free?
Yes. The original videos are publicly available. Your purchase pays for our original learning route, guides, exercises, and included AI tutoring. Hacker Tubes is independent of the video creators.