Princeton Scientist: We Don't Understand AI - Tom Griffiths - #553

Cognitive science and artificial intelligence share a fundamental goal: defining the mathematical structures that govern thought. While modern AI systems excel at processing vast datasets, they lack the innate inductive biases that allow humans to generalize effectively from limited information. Historical efforts to codify reasoning—from Aristotle’s syllogisms to Leibniz’s vector embeddings and Boole’s algebraic logic—provide the necessary framework for understanding current computational challenges. Despite the rapid advancement of large language models, these systems remain prone to sycophancy and lack the robust, cross-domain creativity inherent in human cognition. Rather than viewing AI as a linear path toward human-like intelligence, it is more accurate to treat these systems as distinct computational agents optimized under different constraints, highlighting the critical need for better alignment between human cognitive models and machine learning architectures.