Teach a Small AI to Think

About this course

Understand it first, then use it

Reasoning models like DeepSeek-R1 and OpenAI's o-series changed what AI can do. They don't just answer: they think first, check their work and fix their own mistakes. Most courses stop at using them. This one shows you how they're made, and has you make one. You'll start with a small open model (Qwen2.5, half a billion parameters) that answers grade-school maths badly. Step by step you'll learn how to grade its answers without a human teacher, how a score can nudge a model toward better answers, and how GRPO, the training method DeepSeek made famous, turns that nudge into something stable enough to run. Then you'll run it for real on a free Google Colab GPU, read the training curves like a pro, spot reward hacking before it fools you, teach the model a thinking format, keep its thinking short, and copy a bigger model's reasoning into a smaller one. Every idea comes three ways: a short premium video, an interactive visualization you can play with in the browser (including a real GRPO run you can watch learn), and a written lesson with a line-by-line code walkthrough. Quizzes check your understanding, exercises come with full solutions, and notebooks, cheat sheets and a starter repo are yours to keep. You finish by training and sharing your own reasoning model on a task you choose.

What you will learn

• Explain in plain words what makes a reasoning model different, and how R1-style models were trained
• Write reward functions that grade answers automatically, and spot reward hacking
• Understand policy gradients and GRPO well enough to implement them from scratch in numpy