Research

Current areas

AI alignment: character training, constitutions, and value drift in post-training.
AI evaluations: benchmarking hard-to-check tasks, especially where appearance and actual quality diverge.
AI agents: misalignment, oversight, and reliability in long-horizon tool-use settings.
Robustness and generalization: behavior under pressure, distribution shift, and prior RL incentives.
AI governance of model behavior: when and how models disclose failures, uncertainty, and reasons for deference.
Mechanistic interpretability: early markers of deceptive or apparent-success-seeking behavior.
Formal methods for AI systems: enforceable specifications and safer agentic workflows.