AI Engineering from Scratch: 260 Lessons, No Fluff — Is It Worth Your Time?
Nearly 2,000 stars in under three weeks. That's the kind of momentum that either means something genuinely useful landed, or someone got lucky on a Hacker News thread. I went through rohitg00/ai-engineering-from-scratch to figure out which one it is.
Short answer: it's more real than I expected, less complete than it claims, and worth bookmarking if you're serious about building a structured AI knowledge base from the ground up.
What It Actually Is
This is a structured, self-paced AI curriculum hosted as a GitHub repo with a companion website. It's not a library, not a framework, not a tool you install and run. It's educational content — markdown lessons, runnable Python notebooks, and some scaffolding around a lesson-viewer web app.
The scope is genuinely ambitious: 20 phases covering math foundations, classical ML, deep learning, NLP, computer vision, transformers, LLMs, agents, MCP servers, swarm intelligence, and more. The stated language coverage is Python, TypeScript, Rust, and Julia. The target audience is developers who want to go from "I've used ChatGPT" to "I can build and ship AI systems."
The curriculum is structured so each lesson produces something reusable — a prompt template, a Claude Code skill, a deployable agent. That's the core design philosophy and it's a good one. You're not just reading theory; you're supposed to end each lesson with an artifact you can actually use.
Why This Exists and Why Now
The AI tooling space has a real problem: there's an enormous gap between "I watched a 3Blue1Brown video" and "I can build a production RAG pipeline." Most courses fill one slice — fast.ai does deep learning, Andrej Karpathy does LLMs from scratch, various bootcamps do prompt engineering. Nobody has tried to stitch the whole stack together in one open, free, structured curriculum.
The timing also makes sense. The MCP (Model Context Protocol) ecosystem is exploding right now. Agentic workflows are moving from demos to production. Developers who understood the full stack — from the math to the deployment — are suddenly very employable. A curriculum that treats these things as connected rather than separate domains is filling a real gap.
The "AI-native learning" angle is also interesting: the repo ships Claude Code skills (/find-your-level, /check-understanding) so you can use an AI assistant to quiz yourself as you work through the material. Meta, but actually useful.
What I Actually Liked
1. The lesson structure is solid where it exists.
The completed lessons I reviewed (Phase 0, Phase 1, Phase 2, Phase 3) follow a consistent format: learning objectives up front, concept explanation, runnable code, a quiz, and an output artifact. The math foundations phase in particular is better than I expected — it's not just "here's numpy," it actually builds intuition through code before introducing abstractions. The loss functions lesson producing both a prompt-loss-function-selector.md and a prompt-loss-debugger.md as outputs is exactly the kind of thing that makes this more than a reading list.
2. The dependency stack is honest.
The requirements.txt is pinned with reasonable minimum versions: PyTorch 2.0+, Transformers 4.30+, the Anthropic and OpenAI SDKs. No mystery dependencies, no pip install everything. It tells you what you're getting into.
3. The glossary is genuinely useful.
277 searchable terms. I've seen repos that claim to teach AI and can't define "attention" clearly. The glossary here is a useful standalone reference, and the recent commits show them actively expanding it as new phases land.
4. The pedagogical improvements are real, not cosmetic.
The most recent commits added interactive quizzes and learning objectives to all 96 completed lessons in a single PR. That's a meaningful quality-of-life improvement, not just a star-farming commit. They also fixed Mermaid diagram rendering, which tells me someone is actually using this and filing issues.
5. The roadmap is public and tracked.
There's a ROADMAP.md. The README shows exactly which phases are complete (96/260+ lessons done). This is honest about what exists versus what's planned, which I respect.
Who Should Use This
Use this if: - You're a software developer who wants to understand AI engineering end-to-end, not just API calls - You learn better from structured curricula than from picking random tutorials - You're already using Claude Code or Cursor and want to integrate AI-assisted learning into your workflow - You want a free alternative to paid courses that actually goes deeper than surface level - You're building a team curriculum and need something you can fork and customize
Skip this if: - You need production-ready code to copy-paste into a real project today - You're already past the fundamentals and looking for advanced research-level content - You need video instruction — this is entirely text and code - You want something finished. Right now, 164 of 260+ lessons don't exist yet.
My Concerns
Let me be direct about what bothers me.
The completion gap is significant. 96 of 260+ lessons are done. That's 37%. The phases covering agents, swarm intelligence, Rust, and TypeScript are largely stubs. If you're here specifically for agentic systems or MCP server development — the topics that are most relevant right now — you may hit a wall fast. The README markets 20 phases but only a handful are actually usable today.
One contributor is doing almost all the work. 183 of ~189 commits are from the repo owner. That's fine for a personal project, but the README frames this as a community curriculum. If rohitg00 loses interest or gets busy, this stops getting updated. The 6 commits from thereisnotime and the 4 open issues suggest the community hasn't really formed around this yet.
The "290 hours" claim is unverifiable. There are per-lesson time estimates, but I can't tell how they were calculated. Are these time-to-read estimates? Time-to-complete-including-exercises? This matters a lot when someone is planning a learning schedule.
No releases, no versioning. There's no tagged release. If you're building a team curriculum on top of this, there's no stable version to pin to. Lessons can change under you. The recent commit history shows active rewrites ("technical accuracy audit across 15 lesson files"), which is good for quality but means the content isn't stable.
The multi-language promise is mostly aspirational right now. The completed lessons are overwhelmingly Python. The TypeScript, Rust, and Julia content that's advertised is either minimal or not yet written. Don't come here expecting a Rust ML tutorial today.
The website exists but I can't assess it. The README links to aiengineeringfromscratch.com. I can see the lesson-viewer code in the repo, but without actually running through the site experience, I can't tell if it's a polished learning interface or a thin wrapper around the markdown files.
The Honest Verdict
This is a good foundation for something that could become genuinely great. The completed content is better quality than most free AI curricula I've seen. The design philosophy — every lesson ships a reusable artifact — is the right philosophy. The AI-native learning angle is clever and practical.
But right now, you're buying into a curriculum that's 37% complete, maintained by one person, with the most exciting phases (agents, MCP, swarms, Rust) still largely unwritten.
My recommendation: star it, watch it, but don't commit your team's learning plan to it yet. Come back in 3-6 months. If the owner keeps the current pace and the community starts contributing, this could be the definitive open-source AI engineering curriculum. Right now it's a very promising draft.
If you're an individual developer who wants to start at Phase 0 and work forward, the existing content is worth your time. Just know you'll hit incomplete phases before you finish.
If you want to contribute, this is a good place to do it. The structure is clear, the lesson format is documented, and there are 164+ lessons that need to be written.