The work

Four research areas, one practice that pays for them, and a funding model with no one to impress.

Alignment

The gap between what a model was trained to do and what it actually does. We approach it empirically: less grand theory, more careful observation of where behaviour and intention part ways.

Evaluations and benchmarks

Measuring what models actually do, independent of what anyone hoped they would do. Evaluation is most useful when it is boring, reproducible, and slightly unwelcome.

Safeguard robustness

Adversarial testing of the safety measures wrapped around models, finding where the fences fail before someone less friendly does.

Agentic safety

What autonomous systems get up to when nobody is watching: tool use, delegation, and the failure modes that only appear when a model is given hands.

Consulting, in support of the research

A small consulting practice keeps the lab self-funded: model evaluations, adversarial testing, software engineering, and cybersecurity services. In practice, anything that supports the work of the lab. Engagements are scoped, written up properly, and taken on when they are interesting.

Write to us to start one. Inquiries are answered by the people who would do the work.

Self-funded, grant-supported

The lab runs on its own revenue and on research grants. It does not manage outside capital, offer securities, or take investment. Grantmakers with an interest in alignment and safety are warmly encouraged to write.