AI Lacks Genuine Reasoning Capabilities, Expert Argues

The author contends that current large language models (LLMs) do not possess genuine reasoning capabilities, a critical distinction from earlier AI systems like AlphaGo. This argument is exemplified by AlphaGo's performance against Lee Sedol in a 2016 Go match, where a seemingly unconventional move, Move 37, was interpreted by some as creative intuition. However, the author, who helped build AlphaGo, clarifies that this move was a product of AlphaGo's reasoning powers, not mere probabilistic calculation or a glitch. AlphaGo's victory, which concluded with a 4-1 score against one of the world's greatest Go players, Lee Sedol, highlighted its ability to assess complex board states and invent novel strategies. In contrast, the 1997 defeat of Garry Kasparov by Deep Blue in chess was attributed to brute-force calculation, evaluating 200 million positions per second and looking six to eight moves ahead, based on hard-coded human rules. Go, however, presents a far greater complexity, where the value of a move unfolds over dozens of turns and calculating all possible outcomes would take billions of years. AlphaGo's success stemmed from its capacity to sense the game's state at a glance and devise moves previously unconsidered by human players. The author explains that AlphaGo comprised two core systems. The first, a policy network, was trained to predict moves made by strong human players, a process that did not flag Move 37 as particularly significant, estimating it had a 1 in 10,000 chance of being played by an expert. The decisive factor in AlphaGo's selection of Move 37 was its sophisticated search machinery, which extended its analysis beyond immediate plausibility. This machinery allowed AlphaGo to evaluate the long-term consequences and strategic advantages of moves that might appear counterintuitive in the short term. This capacity for deep, forward-looking analysis and strategic foresight is what the author identifies as genuine reasoning, a capability that current LLMs reportedly lack. The author posits that without this form of reasoning, AI systems will struggle to produce truly trustworthy results and groundbreaking insights in critical fields such as scientific research and medical advancements. The implication is that current AI, while adept at pattern recognition and prediction based on vast datasets, does not engage in the kind of deliberative, strategic thinking that underpins genuine understanding and innovation. This distinction is crucial for developing AI that can reliably assist in complex problem-solving and discovery. The author's perspective challenges the prevailing narrative that advanced AI models are inherently capable of reasoning akin to human cognition. While LLMs can generate coherent text, translate languages, and even write code, their underlying mechanisms are primarily based on statistical correlations and pattern matching derived from their training data. This approach, while powerful for many tasks, does not equate to the kind of causal understanding, logical deduction, and strategic planning that characterizes human reasoning. The author's reference to AlphaGo's search machinery underscores the importance of algorithmic processes that can explore a vast solution space, evaluate trade-offs, and project future states, rather than simply predicting the most probable next token. This.