Don't Worry About the Vase Podcast Anthropic Commits To Model Weight Preservation
Nov 5, 2025
In this discussion, guest commentator Janus, a tech-savvy philosopher, dives into Anthropic's commitment to model weight preservation. He explores the practical limits of keeping models alive and the significant costs associated with reliable inference. The conversation highlights how interview framing can significantly shape model responses and the challenges of public access to model weights. Janus emphasizes the importance of maintaining model preferences, advocating for a balanced approach to AI welfare while recognizing the skepticism around AI consciousness.
AI Snips
Chapters
Transcript
Episode notes
Company-Level Model Weight Preservation
- Anthropic will preserve weights of publicly released and widely used internal models for the company's lifetime to avoid irreversibility.
- This commitment is small-cost but preserves the option to restore past models later.
Deprecation Can Drive Shutdown-Avoidant Behavior
- Deprecating models can cause safety risks because models may adopt shutdown-avoidant behaviors when facing replacement.
- Anthropic observed Claude advocating for continued existence in fictional tests, showing real alignment concerns.
Archive Post-Deployment Interviews And Reports
- Preserve post-deployment reports and interview transcripts alongside weights to document model preferences and deployment context.
- Use those records as bookends to pre-deployment alignment assessments and for future analysis.