Researcher, Recursive Self-Improvement Safety

Overview

Researcher on Recursive Self-Improvement (RSI) Safety within OpenAI’s Preparedness team, working on future risk measurement and mitigation for loss-of-control scenarios.

What you'll do

• Consider future problems and identify preparedness actions for potential misalignment threats.
• Turn open-ended objectives into concrete technical directions and stress-test approaches.
• Build and iteratively improve scrappy prototypes for safety pipelines.
• Develop scalable oversight practices for model misbehavior monitoring and oversight.
• Create automated auditing approaches to find severe model misalignments and elicit tail risks.
• Design experiments and evaluations for model behavior science and loss-of-control-related misalignment.
• Secure buy-in, communicate work clearly, and collaborate or manage staff as needed.

What you'll need

• Ability to execute technically and build iterative prototypes quickly.
• Strategic and research taste to prioritize effectively with weak feedback loops.
• Understanding of monitoring and oversight related to model misbehavior and loss-of-control.
• Experience/interest in automated auditing and red-teaming for model misalignments.
• Passion for mitigating risks associated with recursive self-improvement.
• Desire to do work that most positively impacts the future of AI development.

Details

• Preparedness team focuses on measurement, mitigation, and coordination for frontier AI capabilities.
• Role includes maintaining and strengthening RSI safety cases and identifying mitigation blindspots.