Free. No account. No fee. Ever.
Nobody explains AI. They just describe it.
This is a course you watch working. Every diagram here is a live simulation — a real network, really training, on your own device. Start from two numbers on a taxi receipt and go as deep as you like.
You need no maths beyond multiplication, no degree, and no laptop better than the thing you are reading this on.
Every line is one weight. Thickness is its size, colour is its sign. Watch them move.
Above: a network learning the problem that stopped this entire field dead in 1969. Each line is one weight — thickness is its size, colour is its sign. This is not a video.
What makes this different
You can take the machinery apart
Most courses show you a picture of a neural network. This one hands you a running one and lets you pause it mid-thought, drag it backwards, and watch a single weight change sign.
The explanations use real-world things — a taxi fare, a receipt, a dial you turn — because that is how anyone actually understands a mechanism. It is not simplified for children. It is explained properly, in plain words.
The whole territory
152 concepts, mapped end to end
From what a parameter is, through convolution and attention, to fine-tuning, reinforcement learning, robotics, and where all of this can hurt people. Each concept states its claim in one plain sentence.
- 12
What learning is
You can explain, without hand-waving, what it means for a machine to learn.
- 11
The little maths you need
You can read an equation in a paper and know what it is asking for.
- 12
Data and honesty
You can tell a real result from a fooled one.
- 10
Classical machine learning
You can solve most business problems without a neural network.
- 11
Neural networks
You can build a network from nothing but multiply, add, and bend.
- 14
How training really goes
You can diagnose a training run that is going wrong.
- 08
Machines that see
You understand why a filter sliding over pixels changed everything.
- 06
Order and memory
You know why sequences broke every model that came before attention.
- 08
Attention and transformers
You can draw the architecture behind every modern model from memory.
- 08
Large language models
You know what is actually happening when you type into a chat box.
- 09
Making a model yours
You can adapt a pretrained model to your own problem, cheaply.
- 07
Learning from consequences
You understand how a model is taught what people prefer.
- 07
Getting it in front of people
You can serve a model and know what it costs you.
- 16
AI in the physical world
You understand how a machine turns sensors into decisions, and why that is so much harder than a benchmark.
- 13
Limits, harms, and honesty
You can say clearly what a model cannot and should not do — and who gets hurt when it is wrong.
Being straight with you: 16 of these have a finished, interactive lesson today — the ones lesson 1 covers. The rest are mapped, written, and queued. This page will always say which is which.
Open the full map →Where to start
How machines learned to learn
Eighty years, four collapses, and one idea that refused to die. You need no background to follow this — only curiosity.
Watch it →Lesson 1 · 12 minutesThe smallest brain that learns
Eight taxi receipts. Two unknown numbers. By the end of this page you will have trained a model by hand, and you will know what that sentence means.
Train it yourself →Lesson 2 · 15 minutesTeach a machine its first words
ChatGPT was trained by one game: guess the next piece of text. You are about to run that exact game on two hundred characters, watch a machine learn to talk, and save its brain as a file.
Train it yourself →