Choose a route before setting up an environment
AI Engineering from Scratch gives developers several ways into the same curriculum: read on the website, use a coding-agent tutor, or clone the repository and run code. The useful decision is which route matches the work you want to learn.
The supplied current README advertises 542 lessons across 20 phases. The README pinned to our measured commit, 1bafaa8, advertised 523 lessons and approximately 342 hours; Edition v2026.10 also describes 523 lessons. Those figures belong to different source snapshots. Neither lesson totals nor study-time estimates are our measurements of coverage or completion.
The current README lists entry points for model foundations, LLM applications, and agents, plus focused MCP and Agent Skills routes. Its estimates include about 60 hours starting at Phase 14, 23 hours 15 minutes for MCP, and 9.5 hours for Agent Skills. Treat these as publisher estimates. The learning-paths page further warns that its displayed guided lesson minutes exclude foundation work, independent projects, and professional experience.
The prerequisite is straightforward: you should already be able to write code. The course is a possible next step for a developer learning AI engineering, rather than a substitute for an introductory programming course.
Reading, local lessons, and tutor installation are separate choices
The README says you can read completed lessons on the website without setup or cloning. For local work, it documents a Phase 0 preflight followed by a dependency-free vector lesson. That is a smaller starting commitment than reproducing the dependency environment recorded by our harness.
The tutor uses npx skills add rohitg00/ai-engineering-from-scratch. The README requires Node.js, npx, and a skill-capable coding agent, and asks users to choose a host and installation scope. It also tells users to verify that the host discovers the selected destination. Installing files and confirming host discovery are separate steps.
The placement flow is documented as saving a study plan to LEARNING.md. Reading tutor material does not require a clone, but copied repository code commands and executable MCP or Agent Skills labs do. Our run did not verify tutor installation, invocation, or those labs.
What our recorded run establishes
Our recorded run took place on September 30, 2026, at commit 1bafaa88bb4668356791150bec3a6d7df38387eb, dated September 29. Its configuration recorded 3 CPUs, 8 GB of memory, and image ghcr.io/astral-sh/uv:python3.12-bookworm. The retained install output identifies Python 3.12.12 in /work/home/venv.
The harness marked installation successful in 86 seconds and the build successful in 8 seconds. It recorded 207 dependencies and 6,346 MB for their environment. Total harness time was 104 seconds; that total is distinct from either individual step. The checkout summary recorded 5,423 files, approximately 322,874 lines of source, and 60.9 MB.
The exact install and build commands are absent from the supplied record, and the build tail is empty. These successful step statuses do not establish that every lesson runs, that an installed binary works, or that a particular site or book artifact was built. No learning workload or model benchmark was measured.
The test step was skipped with reason "no script". That is neither a failed test suite nor proof that the repository has no tests. The detector recorded four CI workflows, no root Dockerfile or Compose file, and no tests directory. These are detector observations at that commit, not an exhaustive inventory of checks throughout the curriculum.
The recorded pip-audit advisory total was 0. The record does not establish the scanned package set, so coverage remains unknown. The dependency count must not be read as the audit's coverage count.
Upstream checks and issue reports have their own scope
Edition v2026.10 says CI runs lessons' own tests on changes while skipping lessons requiring scientific libraries such as PyTorch. That describes upstream checking; it is not a passing-test result from our sandbox. The release also reports 58 revised lesson texts and 46 repaired dead links. Those are release-note claims, not our independent audit.
Issue 510 reported incorrect model-parameter and KV-cache calculations in Phase 10 lesson 14. The supplied issue extract shows it closed as completed through PR 511 on October 13, 2026. It should not be presented as an unresolved current defect, and we did not rerun the corrected calculator.
Issue 523, opened October 1, reports that the Phase 0 Python verification section shows Python code without a shell command to run it. The supplied extract still marks the issue open and references a proposed correction. That report concerns the lesson's verification section; the supplied current README separately includes a runnable preflight command. Neither report establishes that the whole curriculum is broken.
For a class or team, pin a commit, select the lessons you will assign, and retain their commands and outputs. A successful generic build step cannot replace checking the particular examples learners will use.
How to choose between curricula
Microsoft's AI for Beginners advertises a 12-week, 24-lesson curriculum with practical lessons, quizzes, and labs. It is a relevant option when you want that defined beginner structure. We have not measured whether learners finish it faster.
Sebastian Raschka's LLMs from Scratch repository accompanies Build a Large Language Model (From Scratch). Its stated scope is developing, pretraining, and finetuning a GPT-like LLM, using PyTorch without external LLM libraries. Choose it for that concentrated implementation path.
Choose AI Engineering from Scratch when moving between foundations, applications, protocols, and agents is useful to you. Start with one documented route and keep evidence of the examples you run before expanding your environment.
How the ratings work
These are editorial ratings of the available evidence, not benchmark results. On the 1-to-5 scale, 1 means substantial documented barriers, 2 means limited support with material gaps, 3 means a useful documented basis with important verification limits, 4 means strong support with narrower gaps, and 5 requires unusually complete corroboration for the category. Setup considers entry requirements and recorded execution; docs considers navigation and runnable guidance; community considers visible participation rather than popularity; maturity considers releases and verification scope. Unknown adoption, compatibility, or test coverage does not earn a high score.