FIG · About
The AI curriculum we wished existed.
obvix-learn is the free curriculum from Obvix Labs, written by Karan Prasad. Modern AI is scattered across dozens of sites that don't talk to each other; this is the single sequenced path from the math prereqs to mechanistic interpretability, the one we wanted when we were learning the field ourselves. Free to read, no signup, no upsell.
How it was built
Over a focused two-week sprint in May 2026 we (1) ran an exhaustive recon of the existing free-AI-learning web, (2) distilled the sources the field already trusts: Distill, the transformer-circuits papers, Karpathy, Géron, ARENA, Lilian Weng, the CAIS safety textbook, Willison, together with the questions learners get stuck on across open forums, using Argus, Obvix Labs' in-house tooling, (3) wrote 27 chapter drafts totaling more than 190,000 words against a locked editorial template, (4) ran four parallel audit passes for DAG cycles / voice consistency / concept redundancy / safety-lens density, then (5) shipped this site.
Every chapter includes worked library code and from-scratch exercises using the tools appropriate to the topic—NumPy, scikit-learn, or PyTorch. Twenty-three chapters also provide an explicit side-by-side scratch/library mapping. Every chapter ends with a safety lens that re-reads the techniques through the interpretability and red-team angle.
The four signature chapters
These are the chapters obvix-learn was built to make exist:
- Ch. 17 · Efficient Inference: KV cache, paged attention, FlashAttention, speculative decoding, quantization. The systems chapter the rest of the LLM-on-the-web content hand-waves.
- Ch. 19 · RL + RLHF: MDPs → Q-learning → policy gradients → PPO → RLHF pipeline → DPO + GRPO + RLVR. The reward-hacking taxonomy derived from inside the algorithm, not as a hand-wave at the end.
- Ch. 22 · Mechanistic Interpretability: the residual stream view, hookpoints, activation patching, linear probes, sparse autoencoders, induction heads, attribution graphs. Built around finding the induction head in a 2-layer transformer with your own hands.
- Ch. 24 · AI Safety & Red-Team: OWASP LLM Top 10, the lethal trifecta, jailbreak zoo, mech-interp for safety, alignment failure modes, a Level 1 prompt-injection CTF in the lab.
What it is not
obvix-learn is not a cohort program. It is not a certificate mill. There is no upsell to a paid tier. There are no ads. The entire manuscript is free to read on this site, in full, with no signup. The writing is proprietary to Obvix Labs, free to learn from, not open for republication.
We do not believe that one more "Top 10 AI Courses You Must Take" listicle is the missing piece of the AI-education landscape. We do believe that one carefully-written integrated curriculum that takes the reader from print('hello') to building an induction head from scratch, with a red-team lens on every chapter, is a thing the field has not had and ought to.
License
Free to read, in full, with no signup. The writing and code are proprietary to Obvix Labs, free to learn from, not licensed for republication or redistribution. For reuse, write to connect@obvix.io.
Contact
Reach Obvix Labs at connect@obvix.io, or Karan Prasad at hello@karanprasad.com. We respond to corrections fast and to feature requests slowly, by design.