Practical AI Safety: Alignment, Red-Teaming, and Responsible Deployment

Descrizione del corso

A practical course on AI safety for engineers. Covers alignment fundamentals (reward hacking, specification gaming, goal misgeneralization), red-teaming techniques for LLMs (jailbreaks, prompt injections, data poisoning), input/output filtering strategies, content moderation APIs (OpenAI moderation, Anthropic guidelines), safe system design patterns, EU AI Act compliance overview, NIST AI RMF framework, bias detection and mitigation, and building internal AI safety review processes.

Cosa imparerai

• Curated from public resources. Understand AI safety fundamentals and implement practical safety measures in production AI systems.

Programma del corso

1. Reward hacking, specification gaming, and goal misgeneralization (30 min)
2. RLHF, Constitutional AI, and preference learning (35 min)
3. LLM red-teaming: jailbreaks, injections, and extraction attacks (40 min)
4. Input/output filtering: guardrails and content policies (32 min)
5. Bias detection and fairness metrics for LLMs (25 min)
6. EU AI Act: high-risk system classification and requirements (35 min)
7. NIST AI RMF: risk management framework for AI systems (30 min)
8. Building an internal AI safety review process (28 min)

Struttura del corso

3 sezioni, 8 lezioni, 9h 0m di durata complessiva. Il corso si segue interamente online, al proprio ritmo.

Domande frequenti