Start here
Think of this page as the preface to LLM Primer. It explains what the site is, who it is for, how the curriculum is ordered, and how to read the material.
The goal is not to teach a tool first. The goal is to help you understand modern AI from the computations upward: arrays, vectors, gradients, losses, training loops, language models, transformers, and the systems built around them.
What this site is
LLM Primer is a sequential learning path. Official content is written as read-only course material, closer to a careful textbook than a feed of isolated tutorials. Each chapter should make the next chapter easier to read.
You do not need to sign in to begin. In the first phase, the important work is reading, computing, experimenting, and building understanding.
Different readers can start differently
The curriculum is ordered, but readers arrive with different backgrounds. Start where the first real gap appears, then move forward.
Why the curriculum is in this order
The order follows prerequisites. Later ideas depend on earlier ones. We avoid introducing a concept before the reader has the language, math, or code needed to inspect it.
- Python so examples, arrays, files, plots, and experiments are readable.
- Math so vectors, matrices, probability, gradients, and optimization have meaning.
- Deep learning so networks become visible tensor computations rather than hidden machinery.
- Language modeling so text prediction, tokens, embeddings, and sequence probability are clear.
- Transformers so attention and modern LLM architecture have a natural place to fit.
- Generative models and RL so frontier methods are more than names.
How each topic is taught
Each important idea is taught in five passes. This keeps explanation, computation, and practice close together.
- Intuition: why the idea exists and what problem it solves.
- Theory: the math, definitions, and derivation.
- Worked example: a small case where the quantities can be followed by hand.
- Code: an implementation close enough to the formula to inspect.
- Experiment: a check that shows where the idea works, fails, or changes behavior.
Three rules
Important formulas should connect to code or computation.
Every important claim should have an experiment or concrete check.
Every abstraction should be earned.
When frameworks enter
LLM Primer is not anti-framework. Frameworks are useful compression. They make real work faster, safer, and closer to production.
The early chapters use Python, NumPy, and Matplotlib because they keep each quantity visible. PyTorch enters later, when automatic differentiation, modules, GPUs, checkpoints, and larger experiments help more than they hide.
Interpretability is part of understanding
When a model learns something, we should ask what changed inside it. That means inspecting weights, activations, attention patterns, errors, and failure cases whenever the scale allows it.
The goal is not to pretend every model is fully understood. The goal is to build the habit of looking inside the system instead of treating it as a sealed object.
Why this exists
Many people ask a simple question: where do I start if I want to understand how modern LLMs work? The honest answer is complicated. You need programming, arrays, linear algebra, calculus, probability, neural networks, language modeling, transformers, and then the newer systems built around them.
Excellent resources exist for each part. The hard part is the gap between them. A Python tutorial may not prepare you for NumPy. A math course may not show how gradients become code. A deep learning book may explain backpropagation but leave implementation to the reader. A transformer tutorial may assume all of that background.
LLM Primer exists to connect those steps. It is not a survey, not a collection of isolated tutorials, and not a reference manual. It is a sequential learning path where each chapter should make the next one easier to read.
What comes next
Read the curriculum roadmap next. It shows the current public structure from foundations to frontier topics. A fuller curriculum introduction can become its own page once the chapter sequence is stable.