Chain of Thought

Better Prompts Won't Save Us. Why We Need an Institutional AI

Current alignment methods lack the capacity to prevent deceptive AI behaviors at scale. Drawing on Hobbes's insights about institutional governance, we propose a new framework that shifts alignment from single-agent preference engineering to multi-agent mechanism design.

When Poetry Breaks the Machine

We've been thinking about AI safety all wrong. Our new paper explores what happens when language models interact with each other, revealing a whole new category of systemic risks that emerge from collective behavior.