1. Linear Algebra

Dot Product

Dot products compress two directions of two vectors into one signal: how much they point the same way (their alignment). Read More.

Vectors

Vectors are the simplest building block of LLM geometry. They are how LLMs package meaning: embeddings, activations, and even weights are just numbers in a direction. Read more.

2. Artificial Neuron, Neural Networks

Layers

Layers are “combinations of combinations”: stacking simple units so each stage remixes the last, letting the network build richer and more abstract features step by step. Read More.

Perceptrons

A perceptron is the simplest learnable decision-maker: it combines several numbers into one score and uses that to separate one kind of input from another with a straight line. Read more about what a perceptron is, and it’s comparison to biological neurons.

3. Training

Loss

Gradient descent

Backpropagation

Model training techniques

4. Architectures

RNN

CNN

Transformer

5. Transformer

Embedding

Encoder, Decoder

Attention

6. Evolving Higher Concepts

Polysemanticity

Thinking mode

Mechanistic Interpretability

Scaling laws

Reward hacking