363 | Chandra Sripada on How LLMs and Humans are Cognitive Cousins
Summary
Sean Carroll interviews philosopher and cognitive scientist Chandra Sripada, who argues that LLMs likely use cognitive mechanisms similar to humans rather than alien intelligence. Sripada presents evidence that LLMs exhibit human-like cognitive effects, such as dual-process thinking and in-context learning, suggesting they have rediscovered core human reasoning strategies through prediction training.
Topics discussed
Intro and Progressive ad
Defining thinking and introducing guest Sean
Overview of how LLMs work
Autoregressive generation and next-word prediction
Babbel and DeleteMe ads
Do LLMs think like humans?
Prediction as a core cognitive principle
Psycholinguistic effects in LLMs
Instagram ad and common sense reasoning
Dual process theory: In-weight vs In-context
System 2 reasoning and Chain of Thought
Anthropomorphism and dual process debate
Production systems and Transformer architecture
LLMs as production systems and prediction machines
Data efficiency and mechanistic interpretability
Architectural canalization and representational alignment
Consciousness and sentience in AI
Impact of LLMs on human learning and society
Conclusion: LLMs as cognitive cousins
Listen ad-free on Castria