AI Safety

David Matolcsi's OSAC framework tackles infinite ethics for AI alignment

A new decision theory for allocating care across infinite worlds and moments.

Deep Dive

David Matolcsi introduces OSAC (Optimization with Subjective Allocation of Care), a tentative framework for moral decision-making under infinities. OSAC does not accept Scott Garrabrant's proposal of allocating care to worlds in proportion to mathematical simplicity, instead using weighted, bounded-utility distributions. Developed alongside ideas about the Long Reflection, OSAC aims to increase optionality and set up good processes for reflection.

Key Points
  • OSAC proposes allocating subjective caring across worlds via weighted, bounded-utility distributions, avoiding UDASSA's simplicity paradoxes.
  • The framework is designed for a 'Long Reflection'—a safe period for moral reasoning with AI advisors and intelligence enhancement.
  • Matolcsi emphasizes increasing optionality and trade between value systems rather than making irreversible utility calculations.

Why It Matters

Provides a practical decision-theoretic approach for AI alignment in infinite universes, preserving subjective values.