This list is not meant to be exhaustive: it includes example interventions to help orient you, as opposed to describing the entirety of the work needed to make AI go well.

Prevent the training of dangerous AI systems

• Expediting verification methods and increasing political will for a global AI treaty. Links: one, two, three
• Compute governance: Control advanced chip manufacturing and distribution, require licensing for large training runs (e.g. above 10^27 FLOPs) akin to nuclear materials licensing
• International monitoring: Create IAEA-style inspections for frontier AI labs
• Algorithmic secrecy: Classify key training breakthroughs as state secrets
• Scientist AI: Developing non-agentic superintelligent AI

Constrain dangerous AI capabilities

• **Input data filtering**: Implement automated safety classifiers to detect and remove harmful, biased, or poisoned content from training datasets before model training (like content policy filters for violence, hate speech, and PII)
• AI Alignment: Deploy constitutional AI with explicit value specifications and RLHF to ensure models follow defined principles for helpfulness, harmlessness, and honesty
• Interpretability: Understand the internal reasoning process of AI systems, and use this to detect misbehaviour of undesirable objectives.
• AI Control: Maintain human oversight over potentially dangerous AI by using trusted models to monitor untrusted models, requiring human approval for consequential decisions, and more.
• Robustness: Conduct adversarial red teaming exercises to test model resilience against jailbreaks, prompt injections, and evasion attacks before deployment
• Model weight security: Secure model parameters through encryption, confidential computing, and access-controlled storage systems to prevent theft by adversaries (following RAND's five security levels framework)

Withstand dangerous AI actions

Biosecurity

• Prevent patient 0: Establish norms against dangerous virology research. Improve biosafety at BSL4 labs. Reduce human-animal contact and zoonosis by reducing animal farming.
• Detect and contain early infections: Build environmental pathogen surveillance in all countries. Establish rapid containment teams (e.g. GERM teams) to prevent further spread.
• Minimise population-level harms: Develop vaccine candidates for all human-infecting viral families. Build vaccine manufacturing capacity. Set up testing procedures for all known drugs against new infections. Build large-scale PPE stockpiles for critical workers.

Cybersecurity