Building Machines that Learn and Think with People (2408.03943v1)

Abstract: What do we want from machine intelligence? We envision machines that are not just tools for thought, but partners in thought: reasonable, insightful, knowledgeable, reliable, and trustworthy systems that think with us. Current AI systems satisfy some of these criteria, some of the time. In this Perspective, we show how the science of collaborative cognition can be put to work to engineer systems that really can be called ``thought partners,'' systems built to meet our expectations and complement our limitations. We lay out several modes of collaborative thought in which humans and AI thought partners can engage and propose desiderata for human-compatible thought partnerships. Drawing on motifs from computational cognitive science, we motivate an alternative scaling path for the design of thought partners and ecosystems around their use through a Bayesian lens, whereby the partners we construct actively build and reason over models of the human and world.

• The paper proposes human-centric AI that acts as a thought partner by leveraging structured probabilistic models to simulate collaborative cognition.
• It outlines modes of collaborative thought such as planning, learning, deliberation, and creation to improve decision-making and problem-solving.
• Case studies in programming, embodied assistance, and storytelling demonstrate the practical impact of integrating cognitive science with AI design.

An Overview of "Building Machines that Learn and Think with People"

The paper "Building Machines that Learn and Think with People" presents a nuanced perspective on engineering machines that engage in collaborative cognition with humans. Unlike traditional AI systems, which often act as tools or extensions of human capability, the authors advocate for developing machines that serve as thought partners, enriching human cognitive processes.

Motivation and Objectives

The authors articulate a clear vision: to design AI systems that not only complement human thought but actively engage in collaborative thinking processes. This entails creating machines that can understand human beliefs, intentions, and resource limitations, allow humans to intuitively understand their actions, and share a common understanding of the world. To achieve this, they argue for leveraging concepts from cognitive science, emphasizing the use of structured probabilistic models to simulate meaningful human-machine interactions.

Modes of Collaborative Thought

The paper delineates several modes of collaborative thought where AI can intersect with human cognition:

• Collaborative Planning: Machine and human jointly engaging in decision-making and task-related assistance.
• Collaborative Learning: Identifying knowledge gaps and facilitating problem-solving.
• Collaborative Deliberation and Sensemaking: Engaging in debate, explanation, visualization, and data analytics.
• Collaborative Creation: Encompassing co-design, idea critiquing, and brainstorming.

The authors stress that these categories are not exhaustive but highlight domains ripe for advancement.

Engineering Thought Partnerships

Desiderata for Thought Partners

The paper underscores three main desiderata for human-centric AI thought partners:

1. You Understand Me: AI should grasp human goals, plans, and limitations.
2. I Understand You: AI systems need to be legible and intuitive.
3. We Understand the World: AI should share a coherent representation of world knowledge compatible with human understanding.

Computational Framework

The utilization of Bayesian models and probabilistic programming surfaces as central to the proposed framework. These models, well-established in computational cognitive science, are suggested for their capacity to represent and infer human cognitive processes, enabling more adaptive and context-aware AI systems. The theoretical underpinnings of probabilistic reasoning provide a principled basis for machines to understand human thought patterns and actions.

Case Studies and Implications

Several case studies demonstrate the application potential of AI thought partners:

• Programming Assistance: Systems like WatChat employ Bayesian models to comprehend and rectify human programming bugs.
• Embodied Assistance: AI systems, such as Cooperative Language-Guided Inverse Plan Search (CLIPS), illustrate sophisticated goal inference and natural language understanding in practical tasks.
• Storytelling and Medicine: Thought partners in these domains exhibit the ability to craft audience-aware narratives and reason over complex medical data, respectively.

These examples elucidate the practical and theoretical implications of integrating probabilistic reasoning with AI development.

Challenges and Future Directions

The paper identifies several challenges and considerations, including:

• Multi-Agent Interactions: The complexity of non-dyadic settings, where multiple humans and machines interact, is acknowledged as a future research frontier.
• Evaluation Metrics: The necessity for interactive evaluation methods that capture the dynamic nature of human-AI collaboration.
• Ethical Concerns: Addressing risks of over-reliance, anthropomorphism, and potential misalignments in AI intentions.

Conclusion

This work advances a compelling paradigm for AI development, proposing a collaborative framework where machines think with, rather than simply for, humans. The integration of cognitive science motifs into AI research opens new avenues for creating systems that align more closely with human thought processes and societal norms. The paper presents a foundational discourse for the evolving landscape of human-centric AI applications, fostering further interdisciplinary exploration.