3Blue1Brown Neural Networks Series

What the series covers

The deep learning series runs to seven chapters, released in two waves:

• Chapters 1 to 4 (the classic set): what a neural network is structurally, using handwritten digit recognition as the running example; how gradient descent lets it learn; an intuitive walk through backpropagation; and a follow-up on the calculus behind backpropagation.
• Chapters 5 to 7 (added in 2024): how transformers and GPT-style models work, what attention is doing, and how the multilayer perceptron blocks inside a transformer might store facts, including the idea of superposition.

Each chapter is a video of roughly 15 to 30 minutes, and all of them are free on YouTube and on the 3Blue1Brown site.

Why the visual approach works

Most of the ideas in machine learning have a shape. Gradient descent is rolling downhill on a surface of errors. Backpropagation is responsibility for an error flowing backwards through the network. Attention is words looking at other words and deciding what matters. Seeing these move makes the equations readable rather than opaque, and it gives you a mental model you can come back to when a course or paper gets dense.

How to watch it

• Before a course, not instead of one. Watch chapters 1 to 3 before any machine learning course that assumes mathematical comfort. They make the following material substantially easier.
• Pause and predict. Before a result is shown, guess what it will be. Getting it wrong is where the learning happens.
• Chapter 4 is optional at first. If calculus is rusty, skip the backpropagation calculus on a first pass and return to it later.
• Take the transformer chapters when you need them. Chapters 5 to 7 are the right preparation before building with language models or reading about their internals.
• Rewatch after coding. Once you have trained a small network yourself, a second viewing lands very differently.

Pair it with

• A hands-on course, such as Google's Machine Learning Crash Course or Kaggle Learn, to turn the pictures into working code.
• The Illustrated Transformer by Jay Alammar alongside chapters 5 and 6, since diagrams and animation reinforce each other.
• Andrej Karpathy's lectures where a network and a small GPT are built in code, for the implementation side.

The series does not teach you to train or deploy models. It builds understanding, which is what makes the practical material stick.

Common questions

How much maths do I need? School algebra is enough for chapters 1 to 3. Chapter 4 uses derivatives and the chain rule; it is optional on a first pass.

How long does it take? Each chapter runs roughly 15 to 30 minutes. Allow an evening or two for the full series, more if you pause to work through the ideas.

Will I be able to build a neural network afterwards? You will understand what one does. To build one, follow it with a hands-on course or a tutorial that trains a small network in Python.

Is it outdated? The first four chapters date from 2017 but explain ideas that have not changed. The 2024 chapters cover transformers, the architecture behind current language models.

Best for: beginners who want to understand what a neural network and a transformer actually do before writing code.