Don't Worry About the Vase Podcast Claude's Constitutional Structure
4 snips
Jan 26, 2026 A deep dive into how a constitutional framework shapes an AI's values and behavior. Discussion of decision theory, negotiation via observable commitments, and why AI alignment is a philosophical test. Exploration of virtue-ethics training, core values and priority order for instructions. Concerns about harmful dependence and the balance between helpfulness and long-term safety.
AI Snips
Chapters
Transcript
Episode notes
Constitution As Training Anchor
- Claude's Constitution aims to guide powerful AI through a human-readable, revisable document.
- Anthropic intends it to shape Claude's training, values, and behavior as a long-term authority.
Missing Functional Decision Theory
- Zvi highlights the absence of explicit Functional Decision Theory despite its relevance.
- He argues FDT should be made explicit because it aligns agent and developer decision-making.
Constitution As Negotiation
- The Constitution functions as a negotiation with Claude, aiming to exhibit policies that elicit cooperative responses.
- This embodies an implicit FDT-style strategy without requiring Claude to be conscious or moral.