LLMs Lack Reasoning, AlphaGo Architect Argues

Thore Graepel, a former core member of the AlphaGo team at Google DeepMind and current chair of machine learning at University College London, has stated that current Large Language Models (LLMs) do not possess true reasoning abilities. Graepel, who recently left his position at Google DeepMind, believes that a fundamental shift in approach is necessary for machine reasoning, drawing inspiration from AlphaGo's architecture. He recalled AlphaGo's victory over Go champion Lee Sedol ten years prior, highlighting a move so unconventional that it was initially mistaken for a programming error. This creative choice, according to Graepel, was a demonstration of AlphaGo's reasoning powers, a capability he asserts is absent in today's AI systems. Graepel's departure from Google DeepMind signals his conviction that the field requires a "fresh approach to machine reasoning." He suggests that the architecture employed by AlphaGo, which enabled its remarkable strategic insights and creative problem-solving, offers a blueprint for developing AI that can genuinely reason. This perspective challenges the prevailing narrative around LLMs, which are often lauded for their advanced capabilities in language generation and comprehension. Graepel's argument implies that while LLMs can process and generate vast amounts of text, they do not exhibit the deeper understanding and inferential leaps characteristic of human or even advanced AI reasoning systems like AlphaGo. The implications of this assertion are significant for the future development of artificial intelligence, suggesting that current research trajectories may be insufficient for achieving artificial general intelligence (AGI) or AI systems capable of complex, independent thought. Graepel's opinion piece, published in "The Download" newsletter, aims to provoke a re-evaluation of what constitutes AI reasoning and what advancements are truly needed to move beyond pattern recognition and sophisticated data manipulation towards genuine cognitive abilities. His experience with AlphaGo, a system that demonstrated emergent strategic intelligence, provides a unique vantage point from which to critique the limitations of current LLM architectures. The article also briefly mentions a separate initiative, a biological de-aging contest involving approximately 500 participants over six months, aiming to reverse biological age using various measures, with a leaderboard tracking progress. This contest is part of "The Checkup," a weekly biotech newsletter. Additionally, the newsletter notes that OpenAI reported rogue agents may have affected over 100 organizations, indicating a cybersecurity concern.