FIG · Our safety posture
Safety-curious by default.
obvix-learn is the educational arm of Obvix Labs. The curriculum was sequenced so that by the time you can build a transformer from scratch, you can also reason about what it could fail at, and why.
FIG · Three habits
How the safety lens shows up in every chapter.
We do not bolt a single safety chapter on at the end and call it done. Instead, three editorial habits sit inside every chapter, regardless of topic.
FIG · 01
A safety lens on every technique
Each chapter ends with a section that re-reads the chapter's algorithm through interpretability and red-team eyes: what could go wrong, what's measurable, what's still open.
FIG · 02
Mech-interp as a first-class topic
Mechanistic interpretability is treated as a core engineering discipline, not a niche. The flagship chapter walks through induction heads, residual streams, and sparse autoencoders with your own hands.
FIG · 03
Red-team labs with real stakes
The safety signature chapter ships a Level-1 prompt-injection CTF in the lab, OWASP LLM Top 10 with worked attacks, and the lethal trifecta diagrammed against real agent stacks.
FIG · Signature safety chapters
Three chapters where safety is the whole point.
Ch. 22Signature
Mechanistic Interpretability
The residual stream view, induction heads, sparse autoencoders, attribution graphs. The obvix flagship.
Read chapterCh. 23
Eval Science
Why evals are hard. LLM-as-judge failure modes, Elo arenas, custom-eval recipe, contamination detection.
Read chapterCh. 24Signature
AI Safety & Red-Team
OWASP LLM Top 10, the lethal trifecta, mech-interp for safety, a Level-1 prompt-injection CTF. Signature.
Read chapterFIG · The lab behind the curriculum
obvix is a research lab first. Teaching is how we audit our own understanding.
Writing chapters that hold up to a first-principles reader is the same discipline as writing internal research notes that hold up to peer review. The curriculum exists because the process of building it makes the lab better.