AI Engineering from Scratch: The University-Level Curriculum You Don't Pay For
The internet is flooded with AI tutorials. You can find a five-minute video on fine-tuning a model, a blog post on building a RAG pipeline, and a GitHub Gist for a basic agent. The problem? They're all disconnected fragments. You can follow the steps, but you're left with a shallow understanding, unable to debug when things go wrong or innovate beyond the template. ai-engineering-from-scratch is the antidote to this fragmented, copy-paste culture. It's not a tutorial; it's a comprehensive, deeply principled curriculum designed to teach you how AI actually works, from the linear algebra up to autonomous agent swarms.
A Spine for Your Knowledge
The project's central premise is that to truly master a subject, you have to build it yourself before you use the polished tools. Most courses jump straight to PyTorch or TensorFlow, hiding the complex machinery under layers of abstraction. This curriculum forces you to confront the machinery head-on. It's structured into 20 distinct phases, starting with the absolute fundamentals (Phase 1 — Math Foundations) and progressively building up to the cutting edge (Phase 16 — Multi-Agent & Swarms and Phase 19 — Capstone Projects).
This structure is the project's greatest strength. It provides a clear, linear path through an intimidatingly complex field. You don't just learn about transformers in isolation; you arrive at them after mastering deep learning fundamentals, NLP concepts, and the attention mechanism they're built on. By the time you get to Phase 11 — LLM Engineering, you're not just using a large language model; you've already built a smaller version from scratch in Phase 10. This foundational approach demystifies the "magic" of AI and replaces it with solid engineering principles.
The Rhythm of a Lesson
This philosophy is baked into the structure of every single one of the 503 lessons. Each lesson follows a consistent six-step pattern: Motto, Problem, Concept, Build It, Use It, Ship It. The core of this loop is the "Build It / Use It" split. First, you implement an algorithm using raw math and standard libraries—no deep learning frameworks allowed. You write backpropagation by hand. You build a tokenizer from scratch. Then, in the "Use It" phase, you solve the same problem using a production library like PyTorch or scikit-learn.
This is a brilliant pedagogical device. It ensures you understand exactly what the high-level framework is doing for you. You appreciate the convenience of model.backward() because you've already calculated the gradients yourself. This approach builds intuition that you simply can't get from reading API documentation. At the end of each lesson, you "Ship It" by creating a tangible artifact: a reusable prompt, a skill for an AI agent, or even a server. This focus on producing practical outputs keeps the material grounded and helps you build a portfolio of work as you learn.
Strengths and Modern Touches
Beyond its core philosophy, the project has several standout features. The sheer scope is staggering—at an estimated 320 hours, it's a serious commitment but also an exhaustive resource. The inclusion of code in multiple languages (Python, TypeScript, Rust, Julia) is a thoughtful touch, acknowledging that AI engineering isn't a Python monolith.
Perhaps the most modern and impressive feature is its integration with AI agents. The curriculum includes built-in skills for assistants like Claude. You can run /find-your-level to get a personalized starting point based on a ten-question quiz. After completing a phase, you can use /check-understanding to quiz yourself. This is a clever, interactive way to navigate a massive amount of content, turning a static repository into a dynamic learning partner. It shows the creator isn't just teaching AI; they're using AI to improve the teaching itself.
The Price of Depth
The project's greatest strength—its depth and rigor—is also its biggest barrier. This is not a quick path to building a weekend demo. A ~320-hour time commitment is equivalent to two full university courses. Anyone looking for a fast-track into AI will be better served elsewhere. This curriculum demands patience and dedication.
While the project is actively maintained, with a release just last week, the fast-paced nature of AI means some material can risk becoming dated. What is state-of-the-art today might be legacy in 18 months. A structured curriculum like this has to fight a constant battle against obsolescence. Finally, while the provision of machine-translated READMEs is a nice gesture for accessibility, non-English speakers may find the quality of these translations to be a hurdle for understanding nuanced technical concepts. The 106 open issues also suggest that while the project is popular and active, users will likely encounter small bugs, typos, or areas needing clarification along the way.
Is It For You?
With over 45,000 stars and significant web traffic, ai-engineering-from-scratch has clearly found a massive audience. It's an active, healthy project that fills a genuine need. It's not a library to install or a framework to use in your next production app. It's an educational resource, and a uniquely powerful one at that.
If you're a developer who feels like you're just skimming the surface of AI by calling APIs, this curriculum is your deep dive. If you're a student looking for a free, open-source alternative to a formal master's program, this is the closest you'll get. It's a long, challenging road, but it's one that leads to genuine, foundational understanding. It won't just teach you how to use AI tools; it will teach you how to build them.