Build It, Then Use It: How I wrote 435 AI engineering lessons from scratch

Описание и применение

История о том, как автор создал 435-уроков курса по инженерии ИИ, следуя правилу «Сначала напиши, потом используй». Курс охватывает 20 фаз, от линейной алгебры до многоагентных систем. Каждый алгоритм сначала реализуется с нуля на NumPy, а затем сравнивается с производственной библиотекой. Это помогает разработчикам понять внутреннее устройство фреймворков и стать уверенными в работе с ними.

Задача

There is a gap between using AI tools daily and feeling ready to ship production code. Frameworks like PyTorch and libraries like tiktoken act as black boxes, hiding the underlying mechanics. Most courses either teach math without code or code without math, leaving developers unable to debug or scale their implementations when issues arise.

Результат

The author successfully created a 435-lesson, MIT-licensed curriculum called 'ai-engineering-from-scratch'. The curriculum covers 20 phases of AI engineering, from linear algebra to multi-agent systems. By following the 'Build It, Then Use It' rule, the author produced a resource that helps developers understand the internals of frameworks and build production-ready skills.

Рабочий процесс

• Identify a core AI algorithm (e.g., attention, gradient descent).
• Implement the algorithm from scratch using only NumPy and standard libraries (Build It).
• Run the same input and data through the production framework (e.g., PyTorch, tiktoken).
• Compare the outputs of the custom implementation and the framework to verify numerical precision.
• Iterate through 20 phases, building increasingly complex systems (e.g., small LLM, agent loop, multi-agent system).
• Generate reusable artifacts like prompt templates, skill specs, and MCP servers for each lesson.

Доказательства

The first time I wrote a tokenizer, I did it with a for loop. I counted byte pairs by hand, merged the most common ones, and waited about forty seconds for it to chew through a small corpus. The output was slow. The output was ugly. The output was correct.

Источник

Then I ran the same input through tiktoken and watched it finish in forty milliseconds. That was the moment tiktoken stopped being magic. It was the same thing I had written the night before, in Rust, with the loop unrolled and the cache warm.

Источник

I wrote 435 lessons under that rule. They live in a free, MIT-licensed curriculum called **ai-engineering-from-scratch**.

Источник