AI systems have advanced dramatically, yet remain narrow in scope, excelling at tasks like image interpretation and language processing but lacking broader human-like intelligence. These achievements rely heavily on vast datasets and computational power rather than fundamental understanding. The authors propose studying human cognitive mechanisms—specifically Daniel Kahneman’s theory of thinking fast and slow—to guide AI development. They suggest a multi-agent architecture where problems are handled by either system 1 (fast) agents, which react based on past experience, or system 2 (slow) agents, activated for deliberate reasoning and optimal solution search. Both agent types are supported by a world model containing domain knowledge and a self-model tracking past actions and solver skills. This approach aims to embed more comprehensive intelligence into AI systems by mimicking human metacognitive processes.

Submitted by Andrea Loreggia on October 5, 2021, the paper explores how architectural design inspired by dual-process theory could bridge the gap between current narrow AI and more general, adaptive intelligence.