Area: Implications of a Highly Automated Economy

Project Ideas:

• How can humans create mechanisms to effectively supervise the operation of large teams of autonomous AI agents capable of acting over long (1 day+) time horizons? What does AI agent management look like, and are there unique features of the way AI agents are likely to operate that can reduce the difficulty of this principal-agent problem?

• How can we design infrastructure to simulate wargames between different actors using LLMs? These simulated wargames may be of lower fidelity than an extremely high effort human wargame, but may nonetheless offer valuable information about scenarios.

Herbie Bradley is a PhD Researcher at the University of Cambridge, focused on AI policy that anticipates human-level AI systems. Herbie recently worked at the UK AI Security Institute from its inception, helping to set up the Institute and run the UK AI Safety Summit. Previously, Herbie worked on LLM research in the intersection of synthetic data and open-endedness with EleutherAI and CarperAI.

Ideally, mentees would have either economics background or research/engineering with LLMs background.