A structured path from "what is a neuron" to "how a large language model actually works" — written in plain language, backed by the real mathematics, and paired with hands-on labs that run right here in your browser.
Knowing which API to call is a skill with a short shelf life. Understanding why the maths works is a skill that lasts a career.
No lesson assumes knowledge you have not been given yet. Vectors before matrices, matrices before gradients, gradients before backpropagation, backpropagation before transformers.
Every concept ships with an interactive lab. Drag a weight and watch the decision boundary move. Change a learning rate and watch training diverge. Nothing to install.
We do not hide the equations behind analogies — but we always explain the intuition first, then show the formula, then show the code. Three views of the same idea.
A full Python environment with NumPy runs inside your browser tab. Write real training loops, break them, fix them — on any laptop, on any connection.
Two tracks side by side: a gentle path for absolute beginners, and a rigorous path for working professionals. Switch between them at any point.
No account wall, no email capture before you can read, no upsell in the middle of a lesson. Learning material should not have a price tag.
Both tracks cover the same seven modules. They differ in depth, not in scope.
For students, career switchers, and the plain curious. Minimal maths, maximum intuition. Everything is explained with pictures and interactive toys before a single symbol appears.
Prerequisites: school-level algebra. Nothing else.
Pace: roughly 6–8 hours per module.
See the Foundation syllabus →
Track BFor engineers and analysts who already write code. Full derivations, NumPy implementations from scratch, then the PyTorch equivalent, then the production considerations nobody mentions.
Prerequisites: comfortable with Python; some linear algebra helps.
Pace: roughly 10–14 hours per module.
See the Practitioner syllabus →
Each one is complete, with a working lab. No account needed.
Weights, bias, activation — and why a single neuron can separate some problems but never solve XOR. Includes a draggable decision boundary.
18 min read · 1 lab
Module 2 · Lesson 4How a model learns by rolling downhill. Watch what a learning rate that is too small, just right, and far too large each do to training.
22 min read · 1 lab
Module 5 · Lesson 1Tokenization, subwords, and why "strawberry" is three tokens and Indian language text costs more. Type anything and watch it split.
16 min read · 1 lab
Every lesson follows the same four beats, so you always know where you are. You can stop after the picture and still walk away understanding the idea. Or keep going and be able to implement it.
No sign-up form, no credit card, no trial period. Just open it and read.