The Pursuit of AGI: Beyond Empirical Benchmarks

Artificial General Intelligence (AGI) has become widely recognized as the “Holy Grail” of AI research, an objective promising to match or surpass human intelligence. Major global tech companies are investing unprecedented resources in its pursuit and development, pushing the boundaries of what is technologically feasible. Despite this fervor, the field of AGI faces a fundamental challenge: the lack of a single, formal definition.

Currently, approaches to evaluating progress towards AGI primarily rely on empirical benchmarking frameworks. While these tools are useful for measuring the performance of specific systems in defined tasks, they do not offer a robust theoretical foundation for comparing radically different architectures or for understanding the intrinsic properties of general intelligence. This theoretical gap makes it difficult to identify the true commonalities and differences among various proposed approaches, and consequently, hinders the planning of targeted future research directions.

A Comparative Framework Based on Category Theory

To address this challenge, a recent working paper proposes the development of a general, algebraic, and category-theoretic framework. The primary goal of this approach is to provide a formal language for describing, comparing, and analyzing the diverse AGI architectures currently under study or development. This includes paradigms such as Reinforcement Learning (RL), Universal AI, Active Inference, Causal Reinforcement Learning (CRL), and Schema-based Learning (SBL), among others.

Formalization through Category Theory would allow for the unambiguous exposure of their commonalities and structural differences. More importantly, this framework is designed to highlight unexplored areas and opportunities for future research, providing a theoretical compass for developers and researchers. The paper's authors draw inspiration from the concept of “Machines in a Category” to offer a modern view of AGI Architectures within a categorical context, aiming for a deeper and more structured understanding.

Towards a Unified Foundation for AGI Systems

This first position paper represents an initial exercise in applying Category Theory to specific architectures like RL, Causal RL, and SBL. However, it also serves as a first step in a broader and more ambitious research program. The ultimate goal is to provide a unified formal foundation for AGI systems, integrating crucial aspects such as architectural structure, informational organization, agent realization, agent and environment interaction, behavioral development over time, and the empirical evaluation of properties.

The framework is also intended to support the definition of architectural properties, both syntactic and informational, as well as the semantic properties of agents. This includes their assessment in environments with explicitly characterized features, allowing for more rigorous and comparable analysis. This systematic approach is essential for overcoming the current fragmentation of AGI research and for building a coherent and comprehensive understanding.

The Symbiotic Relationship Between Category Theory and AGI

The authors strongly claim that Category Theory and AGI will have a very symbiotic relationship. This mathematical discipline, known for its ability to describe complex structures and the relationships between them in an abstract and powerful way, offers the conceptual tools necessary to formalize the intrinsic complexity of general intelligence. In a field where the very definition of the object of study is still evolving, a solid theoretical foundation is indispensable.

The adoption of a categorical framework could not only accelerate the understanding and development of AGI but also provide a common language for different research communities. This would foster more effective collaboration and faster progression towards the ultimate goal of Artificial General Intelligence, transforming research from a collection of disparate efforts into a unified and coherent scientific program.