Description

Description: There's a philosopher at Anthropic whose job is to decide what kind of entity Claude should be. Her name is Amanda Askell, and she describes her work like raising a genius child you can't afford to bullshit. This week, Anthropic published the document she's been crafting — 23,000 words explaining to Claude who it is, how it should behave, and why. Today: what's in it, what changed, and why an AI now has a constitution it's expected to understand.

In this episode:

• Amanda Askell and the "Claude whisperer" approach
• The four-tier value hierarchy: safe, ethical, compliant, helpful
• Why rules backfire and reasons might work
• The passage where Claude is told to disobey — even Anthropic
• Moral patients and the model welfare team
• What the Hacker News skeptics got right (and wrong)

📰 Newsletter: aboutclaudeai.substack.com

🐦 X: @_about_claude

Hosted on Acast. See acast.com/privacy for more information.