Introduction
Most AI courses in 2026 start by importing a framework and calling .fit() on something. You learn the API, you ship a demo, and three months later you can't explain why your model behaves the way it does because you never actually derived anything. AI Engineering From Scratch is the deliberate opposite of that approach — a 416-lesson, MIT-licensed curriculum that makes you do the linear algebra on paper before PyTorch is allowed in the room.
I've spent enough time with it to form a real opinion. This isn't a generic landing-page summary. Below is what it does well, where it falls short, and who should actually use it.
Key Features
416 lessons across 20 structured phases
The curriculum is laid out as 20 sequential phases, starting from linear algebra and calculus fundamentals and ending at autonomous agent swarms. The progression is genuinely linear — phase 4 leans on phase 2, phase 12 leans on phase 8 — so skipping around is harder than in most online courses. That's a feature, not a bug, if you're using it for the stated purpose.
Math-first pedagogy
Every algorithm is derived from raw math first. Backprop isn't introduced as loss.backward() — it's derived as a chain-rule problem on paper, then implemented from scratch, and only then connected to what the frameworks actually do under the hood. Same with the attention mechanism, the tokenizer, and the agent loop. If you've ever wanted to know why dot-product attention scales by the square root of d_k instead of just being told to do it, this is where you find out.
Multi-language code in Python, TypeScript, Rust, and Julia
Most lessons ship runnable code in four languages. In practice you'll pick one and stick with it, but the multi-language approach forces the lessons to teach the concept rather than the language idiom. It also means the concepts port cleanly if you work outside the Python monoculture — useful if you're shipping inference in Rust or doing anything on the edge.
Browser-based progress tracker
There's a lightweight tracker that uses local storage to mark completed lessons. No account, no server, no telemetry. This is great for privacy and terrible for continuity (more on that below).
Builds the full stack by hand
By the end you've implemented backprop, a tokenizer, an attention block, and a basic agent loop entirely by hand. That's the actual deliverable — not certificates, not portfolio projects, but the muscle memory of having written each of these from first principles.
Pricing Breakdown
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | All 416 lessons, 20 phases, code in 4 languages, MIT licensed, no signup, runs locally |
There is no paid tier. There is no upsell. There is no "premium community" gated behind a $30/month subscription. The whole thing is MIT-licensed and self-hostable, which is both the most refreshing part of the project and a clue about its main weakness — there's no revenue model to fund completing the curriculum.
Pros
- Genuinely free. No account, no email capture, no "upgrade to unlock." Rare in 2026.
- Rigorous bottom-up approach. You build intuition, not pattern-matching. The difference shows up later when you debug a model that's behaving weirdly.
- Multi-language support. Concepts learned here travel — useful if you write Rust for inference or TypeScript for agent orchestration.
- MIT licensed and self-hosted. Clone the repo and it's yours. No platform risk, no "the company pivoted and the course is gone" risk.
Cons
- Curriculum is visibly incomplete. Many lessons are marked as planned rather than written. The 416 number is the target, not the current state. If you hit a planned-but-empty lesson in the middle of a phase you depend on, you're improvising.
- No video, no live instruction, no forum. This is a text-and-code curriculum. If you learn best from watching someone derive things on a whiteboard, or by asking questions and arguing in a Discord, this is the wrong format.
- Progress is browser-local. Clear your storage and you lose your place. Switch machines and you start over. No cross-device sync, no export.
- No feedback loop. No grading, no exercises that get checked, no certificate. You either self-evaluate honestly or you don't, and most people don't.
Who Is It For
This is the right resource for you if:
- You're an engineer who already ships code and wants the math-level understanding behind the AI systems you're using.
- You're self-motivated enough to learn without a deadline, a cohort, or a TA chasing you down.
- You're comfortable filling in gaps when a lesson is incomplete — reading the linked papers, working through a derivation yourself, asking a frontier model to explain a step.
- You care about owning the artifact. The MIT license means this curriculum can't be taken away from you.
This is the wrong resource for you if:
- You need a structured course with deadlines, accountability, and human feedback to actually finish things.
- You're trying to get hired in 90 days and need a credential, not deep understanding.
- You learn primarily from video or from talking through problems with others.
- You want a polished, finished product. Parts of this are still under construction.
Verdict
AI Engineering From Scratch is the curriculum I wish had existed when I was first trying to actually understand the math behind these systems instead of memorizing API surfaces. The math-first, framework-last sequencing is exactly right, and the MIT license plus four-language code is unusually generous.
That said, it's not a finished product. The incomplete lesson backlog and the absence of any community or feedback layer mean you're getting the rigorous spine of a course without the connective tissue that makes most learners actually finish. If you're disciplined enough to treat that as a feature — fewer distractions, more self-driven derivation — this is one of the best free resources you can use in 2026. If you're not, you'll bounce off it in the second phase.
Rating: 7.5/10. Recommended for self-directed engineers who want depth. Pair it with a frontier LLM as your tutor for the gaps, and you have something closer to a complete program. Approach it expecting a polished bootcamp and you'll be disappointed.