Alignment: models that do what we intend.

Capability and controllability diverge as models scale. We work on the training methods and oversight mechanisms that maintain a closeable gap between intent and output, and on what genuine corrigibility requires at scale.

• 1.1Scalable oversight for tasks that resist direct human evaluation.
• 1.2Honesty, calibration, and principled abstention: the conditions under which a model should acknowledge the limits of its knowledge.
• 1.3Corrigibility under distributional shift and resistance to specification gaming.

Interpretability: understanding how a model works.

We reverse-engineer the features, circuits, and reasoning trajectories behind model behavior. A central focus is reasoning geometry: how inference traces paths through latent space, where those paths are stable, and where they bifurcate.

• 2.1Reasoning manifolds: the geometric structure of how models traverse latent space during inference.
• 2.2Probing and editing internal representations.
• 2.3Translating circuit-level findings into falsifiable safety cases.

Biology: reading and designing biological systems.

We treat crystal forms not as static endpoints but as terminal states of distinct nucleation pathways. An MCMC framework samples trajectory distributions over bonding transition space; an RL component scores those trajectories by kinetic accessibility, identifying manufacturing risk where pathways diverge.

• 3.1Nucleation pathway modeling: MCMC over bonding transition space.
• 3.2Kinetic accessibility scoring for polymorph stability and manufacturing risk.
• 3.3Generalization across polymorphic systems toward biologic-class compounds.

Societal: how AI affects work and society.

How interpretability findings translate into deployment policy, and the structure of accountability for systems whose internals are not directly accessible.

• 4.1Diffusion, labor, and the economics of capable AI.
• 4.2Governance and incentive design for deployment.
• 4.3Measuring real-world impact, not benchmarks.