Artificial general intelligence (AGI) is a term that generates much discussion yet remains ill-defined. The latest volume from the Neuraxon Intelligence Academy delves into the various interpretations of AGI, emphasizing why definitions vary so broadly and how current narrow AI, despite outperforming humans in specific tasks, fails to encapsulate the essence of general intelligence. Central to Neuraxon's philosophy is the idea that intelligence consists of a functional system rather than a numerical score.
The disagreements surrounding what constitutes AGI stem largely from individual labs framing definitions based on their specific strengths or aspirations. Some view AGI as the ability to perform any human task, while others see it as surpassing human capabilities in economically productive tasks or achieving proficiency in most cognitive endeavors. Each definition holds merit but fails to withstand critical examination and leads to inevitable conflicts when discussing AGI's attainment.
Academics have sought to provide more objective definitions free from human biases. Researchers like Legg and Hutter suggest conceptualizing intelligence as the capacity to meet goals across various environments. Others, such as Wang, focus on adaptability, while Chollet emphasizes the importance of how efficiently a system can acquire new skills. Despite their rigor, these frameworks still lack a universally accepted benchmark for determining when a machine may achieve general intelligence.
We often find clarity in what AGI is not. A calculator may excel in arithmetic, yet few would deem it intelligent. Similarly, chess engines may outmatch the best players without evoking thoughts of general intelligence. For decades, narrow AI systems have demonstrated superhuman abilities without fulfilling the broader criteria of generality, lending itself to the notion that AGI cannot be simply derived from a collection of narrow skills.
Industry voices have proposed practical metrics, such as an economic variant of the Turing test, emphasizing performance in real-world jobs as a criterion for AGI. A machine capable of doing a job indistinguishably from a human might meet this criterion, yet this approach conflates economic capability with intelligence, potentially misrepresenting the true nature of the system.
Science fiction shapes much of our perception regarding AI, often embedding incompatible definitions of intelligence in our thinking. Films such as "Minority Report" and "I, Robot" promote notions of predictive capacity and rule obedience that don't necessarily correlate to the intelligence we're attempting to replicate in machines.
In nature, intelligence is manifested as a gradient rather than a strict binary between intelligent and non-intelligent beings. Various species exhibit unique problem-solving abilities that illustrate intelligence in diverse forms, reminding us that generality does not equate to a singular hierarchy.
Human intelligence itself is complex, organized hierarchically rather than reduced to a single quantified measure. The insights from the Cattell, Horn, and Carroll model illustrate that intelligence consists of interrelated abilities rather than being confined to a solitary score.
Traditional assessments of intelligence, particularly those that depend on closed environments or established formats, obscure true general intelligence. These tests often lead to superficial success without reflecting the genuine adaptability needed in fluid, unpredictable contexts. Emerging interactive examinations encourage systems to learn dynamically, offering more authentic insight into their capabilities.
Given the complexity of intelligence in living systems, it might be more fruitful to question whether a machine is viable rather than intelligent. Stafford Beer introduced the concept of a viable system years ago, emphasizing sustainability and adaptability in ever-changing environments. This perspective shifts the focus from intelligence as a trait to intelligence as a function of a system.
General intelligence may better represent itself not as an individual achievement but as a synergy of different agents within a network. This multifaceted approach is pivotal as we ponder how to construct systems that reflect the way natural organisms adapt and respond.
While imagining a singular, supreme AI can be tempting, it fails to recognize the inherently decentralized nature of intelligence in society. Knowledge exists in manifold places rather than being concentrated in a singular entity, and complex systems thrive on a diversity of perspectives and abilities. To centralize decision-making in a single intelligence undermines the adaptive, pluralistic nature that defines functionality in society.
Ultimately, artificial intelligence can assist in revealing patterns and refining decisions but cannot substitute for the collective, iterative learning process integral to societal advancement. By recognizing intelligence as a viable, systemic construct, both artificial and human systems can coexist and enrich each other, rather than establishing a hierarchical supremacy where one displaces the other.
This overview of AGI principles and ideals is part of a broader exploration found in the Neuraxon Intelligence Academy series, which continually seeks to expand our understanding of decentralized, brain-inspired AI. For more insights, one can explore the ongoing conversation in Qubic's platforms and resources.
