Neural Networks (3Blue1Brown)
by 3Blue1Brown / Grant Sanderson
The visual canon for how neural networks work — and, in the later chapters, how transformers do.
Overview
Grant Sanderson's Deep Learning series is the visual canon for how neural networks work, and it has quietly grown into an LLM course. It opens with 'But what is a Neural Network?', which builds a handwritten-digit classifier neuron by neuron, then 'Gradient descent, how neural networks learn', an interlude analysing what that trained network actually learned, 'What is backpropagation really doing?' and 'Backpropagation calculus'. The later chapters are why it belongs on a working AI engineer's list: Chapter 5, 'Transformers, the tech behind LLMs', walks the full forward pass of a GPT-style model; Chapter 6, 'Attention in transformers, step-by-step', derives queries, keys and values from scratch; and Chapter 7, 'How might LLMs store facts', unpacks the MLP blocks as key-value memories. The series continues past that into newer material on cross-entropy and compression, so it is actively maintained rather than a 2017 artefact. Every lesson on 3blue1brown.com pairs the video with a written text adaptation — Josh Pullen and Justin Sun adapted several — plus interactive figures, quiz questions and footnotes, so you can read at your own pace instead of scrubbing a timeline. The animations are rendered in manim, Sanderson's own open-source Python library. Manning publishes an enhanced 1-hour-2-minute liveVideo edition with 22 exercises, free on their platform. It is explanation, not implementation: there is no code to run and no model you train.
At a Glance
- Topic
- ML
- Level
- Beginner
- Format
- Video
- Cost
- Free
- Duration
- ~3-4 hours across the series, self-paced
- Provider
- 3Blue1Brown / Grant Sanderson
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓What a neural network computes, layer by layer, on a real digit classifier
- ✓How gradient descent minimises a cost function and why the learning rate matters
- ✓What backpropagation is actually doing, both intuitively and in calculus
- ✓The complete forward pass of a GPT-style transformer, from embedding to logits
- ✓How attention derives queries, keys and values to move context between tokens
- ✓How transformer MLP blocks can act as key-value stores for learned facts
Highlights
- •The animations are the reference explanation the rest of the field links to
- •Every lesson has a written text adaptation with interactive figures, not just video
- •Chapters 5 through 7 cover transformers and attention, not only 2017-era MLPs
- •Rendered in manim, Sanderson's own open-source Python animation library
- •Manning publishes a free enhanced liveVideo edition with 22 exercises
Who It’s For
Best For
- ✓Engineers who can use a framework but cannot picture what it computes
- ✓Anyone who needs transformer and attention intuition before reading papers
- ✓Visual learners who bounce off notation-heavy textbook treatments
Prerequisites
- •High-school algebra; the calculus chapters assume derivatives and the chain rule
- •No programming required — nothing in the series asks you to write code
FAQ
What is Neural Networks (3Blue1Brown)?
Grant Sanderson's animated Deep Learning series, free on 3blue1brown.com and YouTube, builds visual intuition from a single neuron up through transformers. It covers gradient descent, backpropagation and its calculus, then the GPT forward pass, attention, and how MLP blocks store learned facts. Each lesson ships a written text adaptation alongside the video. You finish able to picture what your framework is computing, though you write no code.
Is Neural Networks (3Blue1Brown) free?
Neural Networks (3Blue1Brown) is free to access.
What level is Neural Networks (3Blue1Brown) for?
Neural Networks (3Blue1Brown) is aimed at a beginner audience. Recommended background: High-school algebra; the calculus chapters assume derivatives and the chain rule, No programming required — nothing in the series asks you to write code.
How long does Neural Networks (3Blue1Brown) take?
Expect roughly ~3-4 hours across the series, self-paced. Most learners work through it at their own pace.
What will I learn from Neural Networks (3Blue1Brown)?
You'll learn: What a neural network computes, layer by layer, on a real digit classifier; How gradient descent minimises a cost function and why the learning rate matters; What backpropagation is actually doing, both intuitively and in calculus; The complete forward pass of a GPT-style transformer, from embedding to logits; How attention derives queries, keys and values to move context between tokens; How transformer MLP blocks can act as key-value stores for learned facts.
Topics
Sources
This page was written from 3 sources, 1 on domains other than 3blue1brown.com.