Skip to content

AI and language models

Start with one question: how does a model turn some text into a prediction for the next token? From there, the rest of the terminology has something concrete to refer to.

Transformers follows a prediction from token IDs to vectors, attention, and output probabilities. Small numerical examples explain the operations; optional sections give the formulas and code. See the reading order.

LLMs in the Wild covers the choices around that computation: architecture, prompts, training data, evaluation, and failure. Each article can also be read on its own.

Research papers, from first principles connects those operations to the original papers. Start with the DeepSeek-V4.1-Flash architecture walkthrough, or work through attention and recurrence with small numerical examples.

What do you want to understand?Read
What do the matrix shapes and dot products mean?Linear algebra
How does an error become a weight update?Partial derivatives and backpropagation
Why use logarithms, softmax, and cross-entropy?The math beneath
How can a model use words on both sides of a gap?Masked language models
How do demonstrations and preferences change behavior?Post-training
How does audio become a transcript?Speech processing
How do models use tools or process images?Agents and images

NumPy covers arrays and numerical operations. PyTorch introduces tensors and automatic differentiation. Use these when you want to reproduce a calculation; they are not prerequisites for the main reading.

Definition

Read the full glossary entry →