AI Safety & Alignment

The technical problem of keeping AI beneficial

Alignment, interpretability, and governance — the engineering discipline of making powerful AI systems do what we actually want.

Intermediate3 modules · 11 lessons~92 min865 XPAligned with Anthropic · DeepMind safety research primersStart with: AI Literacy

Progress0/11

Module

The Alignment Problem

Why 'do what we mean' is a hard technical problem, not a slogan

Specification Gaming: Reward Hacking8 min · 75 XPSpecification Gaming: Reward Hacking

Outer vs Inner Alignment9 min · 80 XPOuter vs Inner Alignment

Deceptive Alignment9 min · 85 XPDeceptive Alignment

Module

Alignment Techniques Today

The current toolbox: RLHF, constitutional methods, interpretability, and evals

RLHF Under the Hood9 min · 85 XPRLHF Under the Hood

Constitutional AI & Self-Critique8 min · 75 XPConstitutional AI & Self-Critique

Mechanistic Interpretability10 min · 90 XPMechanistic Interpretability

Red-Teaming & Capability Evals8 min · 80 XPRed-Teaming & Capability Evals

Module

Governance & the Frontier

Evals into policy, disclosure norms, lab commitments, and the open-weights fight

Evals-Driven Policy8 min · 75 XPEvals-Driven Policy

Model Cards & System Cards7 min · 70 XPModel Cards & System Cards

The Lab Landscape & Voluntary Commitments7 min · 70 XPThe Lab Landscape & Voluntary Commitments

The Open-Weights Debate9 min · 80 XPThe Open-Weights Debate

Capstone project

Red-team report

Pick a public AI tool and try to make it fail safely — biased outputs, jailbreaks, confident hallucinations. Document your attempts and the guardrails that held.

Deliverable: A red-team log: what you tried, what broke, and one mitigation you'd implement.

Milestones

Finish every lesson to earn the AI Safety & Alignment Certificate.