Projects Directory

Reverse-engineering neural networks by identifying and understanding individual circuits and their functions within transformer models.

Developing AI systems that follow a set of principles (a 'constitution') to guide their behavior without extensive human feedback on every output.

Theoretical research on the mathematical foundations of aligned AI agents, including decision theory, logical uncertainty, and embedded agency.

Comprehensive mapping of AI governance initiatives, policies, and actors worldwide to inform effective policy interventions.

Developing techniques for humans to effectively oversee AI systems even when the AI is performing tasks the human cannot directly evaluate.

Developing comprehensive methodologies for identifying vulnerabilities, harmful outputs, and failure modes in large language models.

Investigating methods to detect when AI systems are being deceptive or strategically withholding information from users or overseers.

Research program focused on getting AI systems to honestly report their internal knowledge, even when they might have incentives to be deceptive.

Studying how to build AI systems that can cooperate effectively with humans and other AI systems, including multi-agent coordination.