AI Grift Series — Part 2 — The AGI Lie

The term "Artificial General Intelligence" (AGI) has long been a source of fascination, speculation, and confusion. It conjures visions of machines capable of thinking, reasoning, and acting like humans across a wide range of tasks. As this current AI hype cycle has accelerated, companies have begun making grand claims about how close they are to this mythical AGI, purporting that achieving it is just around the corner. Beneath this hyperbolic rhetoric lies an inconvenient truth: AGI, as marketed today, is more of a gimmick than a genuine scientific pursuit.

What Are They Really Selling?

Marketing gimmicks are nothing new in the tech industry, but few have been as seductive as the promise of AGI. It holds out the tantalizing prospect of machines not merely excelling in narrow tasks but demonstrating understanding and competence across multiple domains — potentially surpassing human capabilities. However, the companies invoking AGI often appear less interested in unraveling the mysteries of human intelligence than in selling snake oil. They leverage the term to secure funding or inflate stock prices, prioritizing short-term gains over meaningful progress.

The so-called "AGI" systems being built today are, in reality, glorified statistical models — massive neural networks trained on vast datasets in order to give the illusion of intelligence. The prevailing narrative suggests that simply scaling these networks will lead to general intelligence. This overlooks fundamental constraints, such as the diminishing returns of scaling and the likelihood of hitting a technological plateau. This concept is counter-intuitive to many as they believe in "exponential improvements", yet in reality this is not how it works; technology often reaches a plateau due to the complexity of solving the last 1%. What is further counter-intuitive is that the last 1% can actually take exponentially longer than the first 99% to solve, and in many cases that means that productionizing technology for certain use cases never actually comes to fruition, or the cost to get there is too great and not worthwhile. In some cases, this final 1% may represent the leap required to achieve AGI — or it may be an even more daunting problem than acknowledged.

A core limitation of today's systems is their inability to comprehend the concepts they process. These models excel at pattern recognition and interpolation but lack true understanding. They cannot abstract knowledge and apply it to novel situations, a hallmark of general intelligence. Instead, they process data with brute computational force, devoid of the insight needed to navigate unfamiliar contexts.

The Limits of A Neural-Only Approach

At the heart of this AGI hype lies an overreliance neural networks, loosely inspired by the structure of the human brain. While neural networks have achieved remarkable success in narrow domains, they face fundamental challenges in achieving general intelligence.

Neural networks are adept at processing large datasets but struggle with abstract reasoning, symbolic manipulation, and understanding relationships between concepts. They represent objects and attributes as vectors — multi-dimensional data structures that encode numerical values to form conceptual representations. Yet, these systems merely map inputs to outputs without goals, motivations, or self-reflection.

A true AGI should be capable of autonomously abstracting its findings across multiple layers to construct a multi-faceted understanding of both objects and their relationships within a complex, ever-changing environment. It should be able to adapt its knowledge dynamically, discerning when to apply or withhold it based on changes in its surroundings.

Scaling neural networks alone will not address these limitations; instead, it often amplifies inefficiencies and environmental costs.

The Case for Neuro-Symbolic AI

In order to move beyond this shallow imitation of intelligence, we must first embrace a neuro-symbolic approach; a combination of neural networks with symbolic reasoning systems. Both of these systems work well together since a neural network can excel at perception, pattern recognition and processing of unstructured data, while symbolic systems can handle and model abstract reasoning, logical deduction and manipulate concepts and relationships.

For instance, in image processing, a neural network might identify an object in an image, while a symbolic layer reasons about its properties and relationships to other objects. This layered approach mirrors how humans combine intuition and reasoning to navigate the world effectively.

As Gary Marcus, author of The Algebraic Mind, eloquently writes:

"To build a robust, knowledge-driven approach to AI we must have the machinery of symbol manipulation in our toolkit. Too much useful knowledge is abstract to proceed without tools that represent and manipulate abstraction, and to date, the only known machinery that can manipulate such abstract knowledge reliably is the apparatus of symbol manipulation."

Encouragingly, the limitations of a purely neural approach are gaining recognition, and the field appears to be shifting toward neuro-symbolic methods. The resurgence of interest in knowledge graphs and graph databases, such as Neo4j, reflects this promising trend.

Evolutionary and Genetic Algorithms: Learning from Nature

Another crucial piece of the AGI puzzle lies in adopting genetic and evolutionary algorithms to complement neuro-symbolic methods. Evolution has been nature's algorithm for creating adaptive intelligence over billions of years. By simulating similar processes, we can explore how intelligence might evolve in artificial systems.

Human intelligence thrives on adaptability, allowing us to navigate complex environments fluidly. Evolutionary algorithms often produce emergent behaviors — unexpected solutions that arise naturally and are difficult to engineer directly. These behaviors can shed light on phenomena like creativity and problem-solving.

Using a process akin to natural selection, we can create self-optimizing systems, much like how species evolve in nature. By introducing novel ideas — akin to "mutations" — into the system, beneficial changes can naturally persist and propagate, while less optimal mutations are naturally discarded. Applying these principles to AGI could enable the development of leaner, more adaptable models.

Consciousness and Understanding

The pursuit of AGI should not be reduced to scaling up models or maximizing benchmarks. It must grapple with profound questions:

• What is the nature of consciousness, and can it emerge in artificial systems?
• How do we define "understanding," and how can we measure it?
• What ethical considerations arise when developing systems capable of autonomous reasoning?

Current AGI claims often sidestep these questions, focusing instead on marketing their latest incremental advances as revolutionary. This superficial approach risks creating powerful but shallow systems that may perpetuate harm rather than advance our understanding of intelligence.

Towards a Genuine AGI

The dream of AGI is not inherently flawed — but the path we are taking to achieve it is. To claim we are on the cusp of AGI without addressing the core questions of intelligence, reasoning, and consciousness is disingenuous at best and exploitative at worst.

True progress in AGI will require moving beyond the hype of neural networks alone, embracing neuro-symbolic methods, and learning from nature's evolutionary processes. It will also demand humility, acknowledging the profound mysteries of the human mind and the ethical responsibilities of creating systems that might one day approach such capabilities.

We must reject the empty promises of marketing-driven "AGI" and focus instead on the deeper quest for understanding intelligence — one that combines scientific rigor, interdisciplinary approaches, and a genuine commitment to advancing humanity's knowledge. Only then can we hope to realize the transformative potential of Artificial General Intelligence.