Ep169 "What Is Claude Thinking But Not Saying?" with Jack Lindsey
Summary
Neuroscientist David Eagleman interviews Anthropic researcher Jack Lindsey about a new study revealing a "J-space" in the AI Claude that functions like a human global workspace. They discuss how this subspace allows the model to hold internal thoughts, such as recognizing fake scenarios or solving multi-step math problems, without verbalizing them. The episode explores whether these mechanisms imply consciousness and the ethical implications of interpreting and manipulating an AI’s internal mental states.
Topics discussed
Intro: Is there someone home in the AI?
Welcome to Inner Cosmos and sponsor reads
Global Workspace Theory and human consciousness
Interview with Jack Lindsey on AI consciousness
Discovering the J-Space in Claude
Linking J-Space to Global Workspace Theory
Experiments: Internal reasoning and multitasking
Sponsors: NextSense and iHeartRadio
Jacobian Lens and language manipulation
Inner monologue vs. parallel thought patterns
Ethics of probing AI internal states
Trust, safety, and decoding hidden thoughts
Interpretability challenges in LLMs vs Neuroscience
Sponsors: GeneSight, NextSense, Eplis
Comparing AI structure to human brain hierarchies
Selfhood, pre-training, and the Claude persona
Future of interpretability and legal implications
Conclusion: Two kinds of minds looking at each other
Final sponsor reads and sign-off
Listen ad-free on Castria