Phase 13 · 3–4 weeks
Safety, Alignment & Interpretability
Red-team your own model. Reproduce a basic SAE. Understand current limits of mech interp.
What this phase teaches
Jailbreaks (DAN, prompt-injection, many-shot, GCG/AutoDAN). Defenses (system prompts, classifier guardrails, constitutional classifiers, NeMo Guardrails, Llama Guard, Prompt Guard). Mechanistic interpretability — circuits, induction heads, IOI, superposition, sparse autoencoders (SAEs), transcoders, feature steering, activation patching.
Anchor resources
- Neel Nanda’s ARENA Mechanistic Interpretability curriculum
- Anthropic Mapping the Mind materials
- Papers: transformer-circuits.pub series, Templeton 2024 Scaling Monosemanticity, 2307.15043 (GCG), 2401.05566 (Sleeper Agents)