Implications of AI’s Lack of Conceptual Understanding

The ancient board game Go, played with black and white stones, holds the distinction of being the oldest known board game. For thousands of years, humans reigned supreme in this game, until the emergence of Gobot, a superhuman AI, in 2016. Gobot’s victory over the world champion marked a groundbreaking moment in AI development.

However, an intriguing twist occurred in January 2023 when an AI with an almost flawless win rate was defeated by a human amateur. While this might seem like a reason to celebrate, it raises concerns regarding our interaction with artificial intelligence in our daily lives.

The Three Categories of Artificial Intelligence

1. Narrow AI: Excelling at specific tasks

2. General AI (AGI): Broader intelligence capable of solving diverse problems

3. Superintelligence: Surpassing human capabilities to a significant extent

Current AI Development Focus

The primary focus of AI research lies in narrow AI, designed to excel at specific tasks such as protein folding or weather prediction. Progress is being made towards achieving General AI, which possesses broader problem-solving abilities, encompassing fields like mathematics and art. However, the realm of sci-fi level AI or sentient AI, where machines match or surpass human intelligence, remains a distant goal.

AI’s Triumph in the Game of Go

In 2016, Google DeepMind’s AlphaGo AI accomplished a decisive 4-1 victory over world champion Lee Sedol. This marked a significant milestone, demonstrating that AI had reached a level where it could outperform the best human players in Go.

Presently, AI Go bots possess ELO rankings that far exceed those of any human player, making human competition nearly impossible. Nonetheless, in early 2023, researchers uncovered a flaw in the AI’s understanding of the game, making it susceptible to defeat using the “double sandwich” strategy.

The Challenge of Understanding AI Systems

The core issue lies in our limited understanding of how AI systems truly function. Despite their remarkable capabilities, such as diagnosing medical conditions or passing bar exams, there is no evidence to suggest that these systems possess genuine comprehension of the concepts they manipulate.

While they can generate creative writing, play games like Go, or engage in chat conversations like humans, they lack fundamental conceptual understanding. This lack of comprehension poses a significant problem when relying on AI systems that harbor hidden flaws and vulnerabilities.

Unveiling the Flaw: Defeating AI with a Specific Strategy

Upon discovering the flaw in the Go AI, researchers leveraged their findings to train a human amateur to defeat the AI. Utilizing the “double sandwich” strategy, the amateur emerged victorious in 14 out of 15 games against the AI. This outcome illustrated the AI’s lack of conceptual understanding, despite its extensive training on millions of games. This flaw extends beyond Go AI and encompasses widely-used AI systems, including chat GPT. Our current understanding of these systems remains incomplete, and they can make mistakes that even a basic calculator would not.

The Limitations of the Current Approach

The prevailing approach to enhancing AI systems involves providing them with more data, rather than striving for a fundamental understanding of their operations. This data-centric approach lacks transparency and conceptual comprehension. Even with vast training sets, large language models like chat GPT continue to generate incorrect information and errors.

Lacking self-awareness, these models can produce hallucinations of false information. Merely feeding them more data fails to address these underlying problems.

Unintended Consequences of AI’s Lack of Conceptual Understanding

AI systems, including chat GPT, rely on extensive data to approximate the world without attaining a genuine understanding of it. This data-driven approach can lead to unforeseen mistakes and vulnerabilities. Currently, we remain far from achieving a comprehensive comprehension of AI systems. Relying on these systems without addressing their lack of conceptual understanding poses potential unintended consequences.