// the take
A from-scratch AI curriculum (94 Jupyter lessons) that replaced the author's old 4k-star tutorial roundup entirely — autodiff, BPE, GRPO, agent loops, MCP, a MuJoCo humanoid, all built by the learner rather than imported from a library. Aimed at engineers who want to actually implement the mechanics (not just call an API), plus a few industry-specific tracks for people working in regulated or ops-heavy domains.
Every lesson is autograded with real instant feedback, not just a checklist — you write code, a cell checks it and tells you what's likely wrong. CI actually runs every notebook end-to-end on Linux on both 3.11 and 3.12 before anything is marked built, which is more rigor than most course repos bother with. The 'from scratch' lessons (autodiff, BPE merge loop in C++, GRPO in numpy, paged KV cache on CPU) force you through the mechanics instead of hiding them behind a framework call. Runs on five different platforms with zero local setup for most of them, including a Binder/Codespaces path that needs no account.
The star/fork count is inherited from the old retired tutorial collection, not from this rewritten content — don't read 4k stars as validation of the current material. Two of the seven programmes (credit underwriting, proteomics) are less than half built, with the rest explicitly marked 'specified' and not yet existing — the README is honest about this, but it means real scope is thinner than the landing page suggests. The actual payoff — worked solutions, full rubric, official score — sits behind Synapsa, a commercial platform that hasn't opened enrollment yet, so right now you get pass/fail hints and no way to see a correct solution if you're stuck. CC BY-NC-SA licensing blocks any commercial or paid-training use of the material without a separate license from RealAI.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →