Current State (2024)

What We've Solved: Almost Nothing

• RLHF: Superficial, easily circumventable
• Constitutional AI: Better than nothing, insufficient
• Interpretability: Progress but doesn't scale
• Formal verification: Theoretical only

What We HAVEN'T Solved (Critical)

• Inner alignment (mesa-optimization)
• Deceptive alignment (detection)
• Corrigibility (possibly impossible)
• Scalable oversight (supervising superintelligence)
• Value specification (defining our true values)

P(doom) Estimates (Researchers)

• Eliezer Yudkowsky: ~99%
• Paul Christiano: ~50-70%
• Nate Soares (MIRI): ~90%+
• Community median: ~60-80%

Resources

• 2023 AI Alignment Research Overview - Alignment Forum
• AI Safety State of the Field Report