Research institute · Research orgs · Wilmington, Delaware · Est. 2022
Orthogonal (orxl.org) is a nonprofit AI alignment research organization focused on “agent foundations” and, specifically, formal-goal alignment. It frames alignment as constructing a coherent, fully formal target (rather than a goal stated in natural language) and designing an AI decision process that reliably pursues that formal goal under plausible assumptions about the world.
Orthogonal’s work has emphasized outer-alignment / formal-goal alignment, with its current flagship candidate described as QACI (Question-Answer Counterfactual Interval) and related theory-of-change posts. In addition to publishing technical writeups, Orthogonal has reported running fellowships and in-person retreats as part of its program service activities.
No curated programs yet. Check the careers page ↗.
Publications are not yet indexed for this organization. Read them on orxl.org ↗.
Orthogonal published a formalization of QACI as a formal-goal alignment approach, describing mathematical constructs for the framework.
Orthogonal's Formal-Goal Alignment theory of changeOrthogonal published its theory-of-change describing the overall plan: construct a formal goal (candidate QACI) and design an AI that takes actions to pursue that formal goal under plausible world assumptions.
Orthogonal: A new agent foundations alignment organizationOrthogonal publicly announced the organization’s formation, stating its focus on agent foundations and formal alignment, and identifying its founder.