EP 325 Joe Edelman on Full-Stack AI Alignment



Jim talks with Joe Edelman about the ideas in the Meaning Alignment Institute’s recent paper “Full Stack Alignment: Co-Aligning AI and Institutions with Thick Models of Value.” They discuss pluralism as a core principle in designing social systems, the informational basis for alignment, how preferential models fail to capture what people truly care about, the limitations of markets and voting as preference-based systems, critiques of text-based approaches in LLMs, thick models of value, values as attentional policies, AI assistants as potential vectors for manipulation, the need for reputation systems and factual grounding, the “super negotiator” project for better contract negotiation, multipolar traps, moral graph elicitation, starting with membranes, Moloch-free zones, unintended consequences and lessons from early Internet optimism, concentration of power as a key danger, co-optation risks, and much more.

Joe Edelman has spent much of his life trying to understand how ML systems and markets could change, retaining their many benefits but avoiding their characteristic problems: of atomization, and of servicing shallow desires over deeper needs. Along the way this led him to formulate theories of human meaning and values (https://arxiv.org/abs/2404.10636) and study models of societal transformation (https://www.full-stack-alignment.ai/paper) as well as inventing the meaning-based metrics used at CouchSurfing, Facebook, and Apple, co-founding the Center for Humane Technology and the Meaning Alignment Institute, and inventing new democratic systems (https://arxiv.org/abs/2404.10636). He’s currently one of the PIs leading the Full-Stack Alignment program at the Meaning Alignment Institute, with a network of more than 50 researchers at universities and corporate labs working on these issues.