See exactly how LLMs work

21chapters, about 14 hours

10hands-on labs

3real-model teardowns

118papers, each linked

Every number is inspectable

A small transformer runs in your browser, and every value it computes is on the screen. Nothing is animated for effect: if a number appears, you can see where it came from.

TraceFollow a token through 26 operationsStep from prompt to next-token probabilities. Pick the token, layer and head, and share the exact state as a link. MathRecompute it by handThe real q, k and v values are substituted into each equation, so one attention score can be checked on paper. CodeRead the reference implementationThe operation you are inspecting is highlighted in plain JavaScript, with downloadable test fixtures. CompareTrained against random weightsRun the identical arithmetic with and without learning, and see exactly what training changes.

The browser engine is checked against an independent NumPy implementation to within 1e-9, and the hand-written backpropagation is checked against finite differences. The model is deliberately tiny and says so wherever it differs from production systems.

Start where you are

Curious beginnerStart with functions and gradients. No machine-learning background assumed. Software engineerConnect tensor shapes and operations to code in the Explorer and the labs. ML practitionerVerify the mathematics, then compare against real open-weight models. Instructor or studentShare reproducible examples by link, then check understanding with the challenges and labs.

Curriculum

Each chapter opens with a box saying who can skip it, works through numbers before stating general rules, includes widgets you can poke, and ends with the research behind it.

01

Foundations

Functions, gradients, networks, probability and tokens. Everything later is built from these.

1. 1.Functions, parameters and gradient descent30 min
2. 2.Neural networks and the multilayer perceptron35 min
3. 3.Backpropagation and optimizers40 min
4. 4.Probability, softmax and next-token prediction30 min
5. 5.Tokens and embeddings30 min

02

The transformer

Attention, the block, position, and the complete model.

1. 6.Attention45 min
2. 7.The transformer block40 min
3. 8.Position and rotary embeddings40 min
4. 9.The full model and its training loop35 min

03

Fast attention and inference

Why generation is memory-bound, and the designs that cut the cost.

1. 10.Generating text: the KV cache and inference cost35 min
2. 11.Grouped-query and latent attention35 min
3. 12.FlashAttention30 min
4. 13.Local, sparse and linear attention; state-space models45 min

04

Scaling the network

Experts and low-precision numbers.

1. 14.Mixture of experts40 min
2. 15.Floating point and quantization40 min

05

Training

Pretraining at scale, post-training and adaptation.

1. 16.Pretraining at scale40 min
2. 17.Post-training: instruction tuning, preferences and reinforcement learning50 min
3. 18.Adapting models: LoRA, QLoRA and distillation25 min

06

Systems

Prompting, retrieval, tools and agents.

1. 19.Prompting, retrieval, tools, agents and MCP40 min

07

Security and evaluation

The attacks and defenses, tied back to the papers, and how to measure any of it honestly.

1. 20.Attacks on language models and their defenses60 min
2. 21.Evaluating models and systems45 min

Also here

• Model teardowns: follow one token through gpt-oss-120b and DeepSeek-V3, layer by layer, with their real configurations.
• Hands-on labs: 10 exercises that make the chapters concrete.
• Public security guidance for AI systems: what NSA, CISA and allied agencies have published.
• All papers: every reference, with the chapters that cite it.

Every paper links to arXiv, and every id and title was checked against the arXiv API. Statements about specific models come from their technical reports or model cards, and anything that could not be confirmed is marked. Reading progress stays private to this browser.