About
SOAR Initiative — Science of Agency Research Initiative — is a research group working on the foundations of agency and their consequences for advanced AI.
Why this, why now
Safety arguments about advanced AI lean heavily on claims about agency: whether a system pursues goals, how coherently, over what horizon, and how robustly under distribution shift. These claims do real work in determining what gets built and what gets regulated.
They are also, at present, largely informal. We think that gap is tractable, under-attended, and worth a dedicated research effort.
How we work
Small, theory-first, and biased toward results that other groups can build on. We publish openly, including negative results, and we would rather be usefully wrong in public than vaguely right in private.