Are We Thinking Correctly about AI Intelligence?
Summary
Steve Strogatz interviews cognitive scientist Melanie Mitchell about evaluating AI through developmental psychology and rigorous experimental design. They discuss the "Clever Hans" effect, distinguishing performance from true competence, and the limitations of current benchmarks. Mitchell advocates for mechanistic interpretability to understand AI's inner workings, emphasizing that human curiosity and the joy of understanding remain vital despite rapid technological advances.
Topics discussed
Introduction and guest Melanie Mitchell
AI through the lens of developmental psychology
Neural networks and AI creativity in math
Transfer learning and the limits of AI brilliance
Experimental methodology and the Clever Hans effect
The black box of human vs. AI intelligence
Competence versus performance in AI testing
AI's impact on jobs and pure mathematics
Mechanistic interpretability and understanding
The joy of curiosity in the age of AI
Listen ad-free on Castria