See It Before You Code It: 3Blue1Brown on Neural Networks
Grant Sanderson's 3Blue1Brown series on deep learning is the best visual introduction I know to how neural networks, transformers, and LLMs actually work. Watch these in order before you fight with the code; the intuition will save you hours later.
The Deep Learning series, in order
Each video stands on its own, but they build on each other from the ground up.
-
Intro
Large Language Models explained briefly
A lightweight, high-level intro to what LLMs are and what they do. Good starting point that sets the stage for everything below.
-
Chapter 1
But what is a Neural Network?
An overview of what a neural network is, introduced in the context of recognizing hand-written digits. The one to send someone who has never seen a network before.
-
Chapter 2
Gradient descent, how neural networks learn
An overview of gradient descent in the context of neural networks; a method used widely throughout machine learning to optimize how a computer performs on a task.
-
Chapter 3
What is backpropagation really doing?
An overview of backpropagation, the algorithm behind how neural networks learn. Builds real intuition for what the gradient signal is telling each weight to do.
-
Chapter 4
Backpropagation calculus
The math behind backpropagation: chain rule, partial derivatives, and how the error propagates back through the network. Watch if you want the formalism behind the picture from chapter 3.
-
Follow-up
Analyzing our neural network
A closer look at what the trained network from the earlier chapters actually learned; what the neurons represent and where the pretty story starts to break.
-
Chapter 5
Transformers, the tech behind LLMs
A visual introduction to transformers. Focuses on the overall structure and word embeddings, the pieces you need before attention makes sense.
-
Chapter 6
Attention in transformers, step-by-step
Demystifies attention, the key mechanism inside transformers and LLMs. This is the video to rewatch until Q, K, V click.
-
Chapter 7
How might LLMs store facts?
Unpacks the multilayer perceptrons inside a transformer and how they may store facts. Where "the model knows things" starts to become a mechanical picture.
How to use these: watch the video that matches whatever you are struggling with in the course, then read the equivalent chapter in the course notes or a textbook. Video builds the intuition, text nails down the details.