Ai Engineering From Scratch

The weakness of most AI learning material is that it arrives in scattered pieces — a paper here, a fine-tuning post there, a flashy agent demo somewhere else — and the pieces rarely line up. You ship a chatbot but cannot explain its loss curve; you hook a function to an agent but cannot say what attention does inside the model calling it. This curriculum is built as the spine instead: linear algebra at one end, autonomous swarms at the other, with every algorithm implemented by hand before a framework is imported. The stated payoff is real — by the time PyTorch appears you already know what it is doing under the hood. What makes it more than a long reading list is the tooling around the lessons. The GitHub tutor installs into Claude, Cursor, Codex or any SKILL.md-compatible agent from the terminal, runs a placement quiz and then teaches lessons interactively in your own shell — a genuinely different mode from watching a video. The simulations are the strongest idea on the site: an agent is not just an LLM call, so the simulation lets you run one request through a working machine, switch off retrieval, tools or the verifier, and watch which component actually carries the load. The decompiler simulation and the LLM-serving-GPU simulation (NVIDIA L4 or GB10, KV pool, FP8 switch, memory-bandwidth-bound decode, using numbers from the SGLang 101 runs) push the same instinct — break it deliberately and learn the boundary. The certification prep is the second differentiator: 5 tracks, 67 lessons and 505 original practice questions across Claude and Model Context Protocol Associate routes, built around learner-owned artifacts rather than slideware. Be clear on what it is not: the site says the routes are independent and unaffiliated with Anthropic or the Linux Foundation, and it issues no credential of its own. If you need a certificate to show an employer, this is preparation, not the certificate. The honest constraints are structural. It is text-first and terminal-first: no five-minute videos, no copy-paste deploys, no hand-holding, and a 542-lesson catalog that takes months of self-directed commitment. Four languages is a feature until you are debugging Julia when Python would have done. The public REST API and hosted MCP server are read-only — no accounts, no write endpoints, no learner-progress storage, no lesson execution — so anything you want to persist lives on your own machine or in your browser. Progress is saved in the browser only, which means a cleared profile loses it. Where it fits: self-taught engineers who want depth over speed, CS students who need practical from-scratch exposure to real algorithms, and certification candidates who want hands-on labs. Where it does not: absolute beginners who need hand-holding, video learners, and anyone who wants a framework-first approach from day one.