Current AI systems excel at narrow tasks but lack broader human-like intelligence, according to a paper submitted to arXiv in October 2021. While recent advances in image recognition, natural language processing, and prediction have been dramatic, researchers argue these successes rely heavily on large datasets and computational power rather than fundamental improvements in how AI systems reason.
To address these limitations, researchers propose a new multi-agent AI architecture inspired by psychologist Daniel Kahneman’s theory of thinking fast and slow. The approach divides problem-solving into two distinct systems: System 1 agents operate quickly by drawing on past experience and stored patterns, while System 2 agents engage in slower, more deliberate reasoning when optimal solutions require searching beyond what System 1 can provide.
According to the paper, both types of agents are supported by two models: a “model of the world” containing domain knowledge about the environment, and a “model of self” tracking the system’s past actions and the skills of different solvers. This dual-model approach mirrors how humans distinguish between what they know from experience versus what they must actively think through.
The researchers argue that studying human cognitive mechanisms more closely could help AI systems develop capabilities that are currently missing. State-of-the-art AI systems still lack many competencies that humans take for granted—abilities that extend beyond the specific tasks they were trained to perform.
The work represents an attempt to move beyond purely algorithmic improvements toward architectures that incorporate metacognitive principles—essentially, the ability of AI systems to monitor and regulate their own thinking processes. By implementing this two-speed approach, the system could theoretically know when to trust immediate, experience-based responses and when to engage in more computationally intensive reasoning.
Key facts
- Current AI systems remain mostly narrow, excelling at specific tasks but lacking broader human-like intelligence
- The proposed architecture uses two types of agents inspired by Kahneman’s fast-and-slow thinking theory
- System 1 agents make quick decisions based on past experience; System 2 agents engage in deliberate reasoning
- The system includes both a world model and a self model to support decision-making
- The research was submitted to arXiv in October 2021