Praxis-based AI alignment values

Peli Grietzer is an AI alignment researcher at Arb Research with a PhD in mathematically informed literary theory from Harvard Comparative Literature, completed in collaboration with the Hebrew University Einstein Institute of Mathematics. His research applies machine learning theory to "ambient" phenomena like moods, vibes, and cultural logics, sitting at an unusual intersection of literary studies, philosophy of mind, and ML.

The praxis-based alignment program rejects the dominant utility-function framing for AI behavior. Grietzer's argument, developed across four years and published as the essay "After Orthogonality" on The Gradient, is that rational agents do not align actions to fixed goals but to practices: networks of action-dispositions, evaluation criteria, and shared resources that structure, clarify, and reinforce themselves. The implication is that alignment cannot be expressed as a static reward function or rule set, and must instead be cultivated as a set of learned dispositions over time. The position draws on virtue-ethics traditions that treat rationality and value as entangled rather than orthogonal.

Cosmos Institute selected Grietzer for its second cohort, announced December 17, 2024, alongside Raphael Milliere, Sam Klein, and Seth Lazar. Grants in this cohort fall in the institute's $1K to $10K fast-grant band. The work sits at the governance and evaluation cross-layers of the open AI stack: it shapes how alignment criteria are written and what counts as a valid behavioral specification, upstream of the actual training and runtime tooling.