Technical and Value Alignment

Covers the problem of making systems behave in line with human goals and values: scalable oversight, reward and control, goal drift and value alignment. Excludes fine-tuning and preference-learning techniques as training methods, which belong to 5.5, normative questions about values, which belong to 12.1, and interpretability, which belongs to 11.4.

• 5.5 Fine-tuning, RLHF, DPO and ORPO
• 11.4 Interpretability and Model Reverse-Engineering
• 12.1 Dignity and Human Agency

• 11.1 Assurance, TEVV, Audit and Certification
• 11.2 Cybersecurity and Adversarial Machine Learning
• 11.3 Physical Safety of Robots, Vehicles and Medical Devices
• 11.4 Interpretability and Model Reverse-Engineering
• 11.5 Deepfakes, Authenticity and Watermarking
• 11.7 Incidents and Systemic Risks
• 11.8 Dual Use and Offensive Uses