FIG 0.0 · A free curriculum from Obvix Labs
A first-principles guide to modern AI.
Twenty-seven chapters, from math basics to mechanistic interpretability. Write algorithms from scratch, test them here, and see what broke. Free, no account.Twenty-seven chapters from math prerequisites to mechanistic interpretability. Write core algorithms from scratch, run real tests in your browser, and inspect the failed assertion and observed values. Free, no account. Classical labs run here; deep-learning labs open in Colab, always clearly labelled.
Chapters
Words of original writing
Signature chapters
Hands-on notebooks
FIG 0.1 · The full table of contents
All 27 chapters.
Scroll · 27 chapters
FIG 1.0 · The signature chapters
Four chapters that earn their place on the curriculum.
The rest of the manuscript is solid prerequisite work. These four are what obvix-learn was built to make exist.
Ch. 17 · Signature
Efficient Inference
KV cache, paged attention, quantization, FlashAttention, speculative decoding, vLLM. The signature systems chapter.
Ch. 19 · Signature
RL + RLHF
MDPs, PPO, RLHF pipeline, DPO derived, GRPO + RLVR. Reward hacking taxonomy from the inside.
Ch. 22 · Signature
Mechanistic Interpretability
The residual stream view, induction heads, sparse autoencoders, attribution graphs. The obvix flagship.
Ch. 24 · Signature
AI Safety & Red-Team
OWASP LLM Top 10, the lethal trifecta, mech-interp for safety, a Level-1 prompt-injection CTF. Signature.
FIG 2.0 · The dependency map
Twenty-seven chapters, and what each one needs.
Not a roadmap you abandon at chapter four. Choose any chapter and the map lights up everything you need to read first. Click or tap to keep a chapter selected.
Hover or focus to preview a chapter. Click or tap to keep it selected; choose it again to clear. The dashed rings mark the four signature chapters.
FIG 3.0 · What makes it different
Three things the curriculum is designed to do.
Library and scratch, side by side.
The notebooks make you implement the load-bearing operations before using the production API. Twenty-three chapters also include an explicit scratch-to-library mapping under ‘Under the hood’; the remaining chapters use worked notebook exercises without pretending there is a one-call equivalence. The implementation language follows the topic—NumPy, scikit-learn, or PyTorch—so the scratch path teaches the mechanism while the library path teaches the API you would actually ship.
Every chapter carries a safety lens.
The final section asks how the technique fails in practice: leakage, distribution shift, reward hacking, prompt injection, or fragile evaluation. Some exercises run real local checks; exercises that would require a multi-billion-parameter model are clearly labeled toy models or precomputed replays. Chapter 24 uses a toy classifier to teach refusal-direction geometry and a Level-1 prompt-injection CTF to teach capability boundaries—it does not claim to run Llama-2-7B in your browser.
Distilled from what the field already trusts.
The chapters synthesize primary papers, established textbooks, framework documentation, Distill, transformer-circuits work, Karpathy, Géron, ARENA, Lilian Weng, the CAIS safety textbook, and Simon Willison's prompt-injection writing. Sources are listed in the chapters so claims can be checked rather than accepted because the platform says so. Karan Prasad assembled and edited the curriculum with Argus, Obvix Labs' in-house tooling.
FIG 4.0 · The problem this curriculum is built to fix
Learning stalls when the explanation, implementation, exercises, and failure modes live in different tabs. This curriculum keeps them in one sequence so each detour returns to the thing you are building.
Obvix editorial premise—not a quotation or testimonial.