Learn AI from First Principles

Modern models write code, solve hard problems, and use tools. LLM Primer teaches the ideas underneath them, one visible step at a time.

Method

How each topic is taught

Each topic moves through the same simple loop: understand the idea, compute it, then test it.

When scale allows it, we also inspect what the model learned: weights, activations, attention patterns, errors, and failure modes.

  1. 1

    Intuition

    Why the idea exists.

  2. 2

    Theory

    The math and derivation.

  3. 3

    Worked example

    A small case by hand.

  4. 4

    Code

    An implementation you can inspect.

  5. 5

    Experiment

    See where it works and fails.

Approach

Framework-neutral by default.

The early path uses Python, NumPy, and Matplotlib. That keeps the math, arrays, gradients, and experiments visible.

Frameworks enter only when they help the idea, not when they replace it.

Who it is for

For readers who want to understand AI from the inside.

You may be new to programming. You may know Python but not the math. You may have used models and APIs without understanding what happens underneath.

LLM Primer is for the reader who wants the pieces to connect.

Curriculum

From Foundation to Frontier

Python prepares NumPy. Math prepares gradients. Deep learning prepares language modeling. Language modeling prepares transformers.

Frontier

Advanced material will develop carefully because the field changes quickly. The goal is to teach lasting ideas, then connect them to new papers and systems.

What you build

Each stage ends in something inspectable.

A plot. A small numerical library. A gradient check. A training loop. A tokenizer. An attention block. An experiment that fails for a reason you can explain.

The project is not decoration after the lesson. It is how the lesson becomes real.

Updates

Follow the releases.

Chapters will be released in phases. Join the update list when it opens to hear about new lessons, curriculum changes, and important site milestones.

No account is needed for the early site. The update form will stay simple until sign-in and personal features are worth adding.