1. Capabilities

We need to understand what AI models are currently capable of, and what the trends for their capabilities indicate.

• 1.1. Introduction
• 1.2. State-of-the-Art AI
• 1.3. Foundation Models
• 1.4. Defining and Measuring AGI
• 1.5. Scaling
• 1.6. Forecasting
• 1.7. Takeoff
• 1.8. Appendix: Takeoff
• 1.9. Appendix: Expert Opinions
• 1.10. Appendix: Discussion on LLMs

3. Strategies

How can we mitigate the risks of advanced AI? This chapter surveys the vast strategic landscape to mitigate misuse, alignment and systemic risks.

• 3.1. Introduction
• 3.2. Definitions
• 3.3. Misuse Prevention Strategies
• 3.4. AGI Safety Strategies
• 3.5. ASI Safety Strategies
• 3.6. Socio-Technical Strategies
• 3.7. Combining Strategies
• 3.8. Challenges
• 3.9. Conclusion
• 3.10. Appendix: Long-term questions

5. Evaluations

We need a systematic way of measuring progress on safety.

• 5.1. Introduction
• 5.2. Benchmarks
• 5.3. Evaluated Properties
• 5.4. Evaluation Techniques
• 5.5. Evaluation Frameworks
• 5.6. Dangerous Capability Evaluations
• 5.7. Dangerous Propensity Evaluations
• 5.8. Control Evaluations
• 5.9. Evaluation Design
• 5.10. Limitations
• 5.11. Conclusion

7. Goal Misgeneralization

• 7.1. Introduction
• 7.2. Multi Objective Generalization
• 7.3. Learning Dynamics
• 7.4. Goal-Directedness
• 7.5. Scheming
• 7.6. Detection
• 7.7. Mitigations [WIP]
• 7.8. Appendix: Singular Learning Theory [WIP]