About
I'm a researcher at METR, where I lead embedded control assessment. This means I go into frontier AI companies like OpenAI and Anthropic, and I figure out how misaligned agents could evade their monitoring and control systems (hopefully before the AIs do).
Unfortunately, it seems like we're on track to create AI systems that are substantially more powerful than all of humanity combined, and that we don't know how to control, supervise, or align. We must do everything in our power to not cede control of the future to AIs.
Previously, I created GPQA, a widely used AI capability benchmark; I developed scalable oversight methods (i.e., for supervising systems smarter than their overseers); and I was one of the first few employees at Cohere, where I trained embedding models for semantic search.
Writing & talks
- Red-Teaming OpenAI's Internal Agent Monitoring Systems
- How an assessor could investigate control and monitoring of internal AI agents
- Sub-agent delegation chaining
- What I Found Stress-Testing Anthropic's Internal Agent Monitoring
- MirrorCode: Evidence that AI can already do some weeks-long coding tasks
- Research Update: Algorithmic vs. Holistic Evaluation
- NYU Code Debates Update/Postmortem