What is Meant by AGI? On the Definition of Artificial General Intelligence

Abstract: This paper aims to establish a consensus on AGI's definition. General intelligence refers to the adaptation to open environments according to certain principles using limited resources. It emphasizes that adaptation or learning is an indispensable property of intelligence, and places the controversial part within the principles of intelligence, which can be described from different perspectives.

• The paper presents a unified AGI definition by integrating axioms of learning and resource limitations into a coherent framework.
• It distinguishes general intelligence from specific problem-solving, emphasizing the need for autonomous adaptation in dynamic environments.
• The study underlines consensus-building in AGI research, laying a foundation for evaluating and advancing future adaptive systems.

Defining Artificial General Intelligence: A Review of "What is Meant by AGI? On the Definition of Artificial General Intelligence" (2404.10731)

The paper, "What is Meant by AGI? On the Definition of Artificial General Intelligence" (2404.10731), presents an examination of the concept of AGI, with a primary focus on establishing a clear and unified definition of AGI within the research community. Previous attempts at defining AGI lacked clarity and consensus, leading to differing views and debates. This paper therefore endeavors to distill the essential characteristics of AGI into a coherent framework that aligns with existing artificial intelligence research while paving the way for future development in AGI.

Establishing a Common Foundation for AGI

The paper underscores the importance of gaining a consensus on the definition of AGI as a prerequisite for advanced research and development in the field. The author highlights that understanding intelligence and AGI is not just an intellectual exercise but a foundational endeavor that influences the broader conversation and inquiries in AGI research. Historically, AGI has been described in multiple, sometimes conflicting, ways: as the ultimate aim of AI research, as systems that adapt to non-specified problems, and as autonomous entities that operate without designer intervention post-deployment. However, for a consensus, the reportedly minimal common understanding is essential to prevent ambiguous discourse in the field.

Basic Concepts of Intelligence as the Foundation

The paper identifies intelligence as a complex concept with multiple interpretations derived from different domains, such as human cognitive abilities, problem-solving capacities, and adaptability to environments. The foundation of the discussion is rooted in a set of axioms.

First, intelligence is linked intrinsically with learning and adaptation:

• Axiom 1 posits that any intelligent information system must have the ability to learn and adapt to its environment.
• Axiom 2 asserts that intelligent systems operate with limited computational resources, both in terms of memory and processing capabilities.

The acceptance of the first axiom is predicated on an intuitive understanding that intelligence cannot be attributed to systems incapable of learning. For the second axiom, it underscores that ideal intelligence must address resource constraints akin to human limitations, diverging from Turing’s theoretical model of infinite computational resources.

Definitions: Intelligence and General Intelligence

The paper examines previously proposed definitions, specifically referencing Pei Wang's view of intelligence as the adaptive capacity of an information-processing system acting under constraints of limited knowledge and resources [wang2019defining]. Building on this, the paper suggests a bifurcated definition of intelligence, which sees it both as the capability to adapt to one's environment (with resource limitations acknowledged) and as a collection of cognitive principles, which might vary depending on observer perspective.

There are further discussions on the distinction between intelligence and problem-solving skills. An intelligent system may solve novel problems autonomously, while problem-solving capability achievable only through specific predefined algorithms does not imply intelligence.

Distinguishing Artificial and General Intelligence

The paper clarifies conceptual difficulties surrounding "artificial intelligence" by restricting the interpretation of artificial intelligence to non-biological systems, excluding artificial life forms created using biotechnology. On the notion of “general” in AGI, the paper underlines that general intelligence must handle non-specific problems and adapt to dynamic environments. This highlights how AGI’s ability to learn widely contrasts with specific problem-solving systems.

Defining AGI

The paper proposes a definition specifically for AGI, incorporating a stipulation for AGI systems to possess a defined set of principles, noted as $\mathcal{P}_G$, in addition to possessing a capacity for adaptation to open environments with limited resources. This is contrasted with the broader definition of intelligence and restricted to emphasize the vision of AGI as a system that not only can perform across various domains but does so independently, without direct intervention, and possesses general adaptability.

Conclusion

The paper posits an attempt to bring clarity to the often-debated definitions of intelligence, articulating it as a meta-capability tied to learning and environmental adaptation with limited resources. In advancing towards AGI, the understanding put forward seeks congruity with traditional AI research while acknowledging the necessity for a broader abstraction. This framework invites further critical analysis, encouraging researchers to clarify and hone the foundational principles that will guide future advancements in the field of AGI. The potential implications of achieving AGI are vast, with possibilities ranging from mundane automation to significant impacts on human existence.