README.md

The RustingBrain Course

Eighteen chapters, from what a neural network is to training a mixture-of-experts language model on a GPU and generating images with a published diffusion model. No prior machine-learning knowledge assumed. Rust assumed at the level of "I can read a for loop".

Every chapter has runnable code in code/ and ends with exercises. Chapters 1–10 need nothing but Rust. Chapters 11 and 14–18 are much faster with an NVIDIA GPU, and all of them still run without one.

---

Part I — Foundations

#	Chapter	You need	Time
1	Introduction — How a Neural Network Actually Works	nothing	30 min
2	Setup — From Nothing to a Running Model	a computer	10 min
3	The XOR Problem — Your First Trained Network	1–2	30 min
4	Working With Real Data	1–3	45 min

Read 1 even if you are in a hurry. Everything after it assumes it.

Part II — The Dense Toolkit

#	Chapter	You need	Time
5	Regression — Predicting Numbers	1–4	45 min
6	Classification — Choosing Between Categories	1–5	45 min
7	The Training Loop — Taking Control	1–6	1 hr
8	Evaluation — Measuring Honestly	1–7	45 min
9	Saving, Loading, and Using a Model	1–8	40 min
10	A Complete Project	1–9	90 min

Chapter 8 is the one people skip and regret. A model that looks good and is not is worse than one that looks bad.

Part III — Going Further

#	Chapter	You need	Time
11	Training on the GPU with CUDA	1–10, an NVIDIA GPU	1 hr
12	Importing Models From TensorFlow and PyTorch	1–10	45 min
13	When Things Go Wrong	1–10	read once, return often

Part IV — Language Models

#	Chapter	You need	Time
14	Tokens and a Transformer Language Model	1–10	2 hr
15	Mixture of Experts	14	90 min
16	Training a Language Model End to End	14–15	an afternoon
17	Fine-Tuning With LoRA	16	45 min

Part V — Images

#	Chapter	You need	Time
18	Generating Images From Text	1–10, a downloaded model	45 min

---

If you are here for one thing

You want to	Start at
Understand what any of this means	1
Predict a number from a spreadsheet	5
Sort things into categories	6
Know whether your model is actually any good	8
Ship a trained model	9
Make it run on your GPU	11
Run a model someone else trained	12
Export a model for another runtime	12
Fix a loss curve that looks wrong	13
Train a language model	14
Fit a bigger model in the same memory	15
Plan a run that takes days	16
Fine-tune a model to answer questions	16
Adapt a trained model on a small GPU	17
Turn a prompt into a picture	18

Running the code

Each chapter links its full source in code/. Set up a project once, as chapter 2 describes, and paste a chapter's file into src/main.rs:

cargo run --release

--release is not optional for anything past chapter 10. A debug build is ten to fifty times slower, and chapter 13.8 is a list of people who forgot.

What is not here

Training a convolutional network, encoder-decoder models, reinforcement learning, multi-GPU and distributed training. The library runs convolutions (chapter 18 is built on them) but no chapter trains one yet; the rest are not implemented, so they are not taught here.

The reference documentation is the source. Every public type carries the reasoning behind it, which is the part that does not fit in a tutorial.