OpenAI co-founder Andrei Kapaci has a habit of putting all teaching code on GitHub for free. Recently, a developer ranked his seven warehouses in order of learning, and this list was collected by nearly 1,500 people.

The first stop is nn-zero-to-hero, a set of video lessons on building a neural network from scratch. This is the best place for people with no basic knowledge to enter.

Second stop micrograd. The core of letting AI "learn things" is automatic derivation and backpropagation. This warehouse uses very little code to write it clearly. After reading it, you will have a feel for the word "gradient".

The third stop is makemore. Feed it a string of names and it will learn to make up new names. It’s a small toy, but the complete process of character-level language model is in it.

The fourth stop, minbpe, talks about how AI cuts text into tokens. Many people have used large models for several years and have never thought about why Chinese costs so much tokens. After reading this, you will understand.

The fifth stop is nanoGPT, train your own GPT. At this point, all the things learned previously are connected.

The sixth stop was nanochat, which Kapasi himself described as "the best ChatGPT you can buy for $100", from word segmentation, pre-training, fine-tuning to chat interface.

The seventh stop, autoresearch, allows an AI agent to run training experiments overnight on a GPU and retain effective changes. This has gone from "learning how to build a model" to "let the model do research for you".

What the seven warehouses have in common is that they are short. Kapasi's code has always been known for being able to be read to the end. It does not have the layers of encapsulation like industrial-level projects, making it suitable for holding a computer on weekends and following it line by line.