Scalable Oversight by Learning to Decompose Tasks
How do we control superintelligent AI?
We are exploring one route to supervising increasingly capable AI systems: an AI learns to break a hard task into “safe” subtasks that a human-level model, or a human, can reliably judge. This site documents the project as it develops: results, methods, setbacks and explanations.
Latest updates
Progress update
Decomposing difficult tasks is ... difficult
The system in which reasoning LLM decomposes complex tasks into simpler ones for a weaker LLM to implement loses with single zero-shot thinking LLM but wins with zero-shot non-thinking model. This leaves a lot of room for improvement.
Vision & Goals
Scalable oversight by learning to decompose tasks
Our vision and goals for scalable oversight over AI systems.