Don't be fooled–LLMs don't reason (www.technologyreview.com)

Thore Graepel, a former core member of the AlphaGo team, warns that today's large language models (LLMs) lack genuine reasoning capabilities, which are crucial for high-stakes applications in fields like medicine and science. He contrasts AlphaGo's architecture, which integrates intuitive decision-making with an explicit game tree that tracks possible future positions and their evaluations, with the functioning of contemporary AI models. Current LLMs primarily operate on a system 1 basis—rapid, pattern-based predictions—without maintaining a clear record of their reasoning process or separating knowledge from action. This leads to potentially flawed conclusions where the reasoning path is obscured or inaccurately represented. Graepel argues for a new approach to machine reasoning that mimics AlphaGo’s method of maintaining a detailed epistemic state, allowing AI to systematically evolve its understanding and improve its decision-making process based on evidence. He emphasizes the importance of having AI systems with transparent reasoning mechanisms that can adjust beliefs and generate insights in complex, real-world scenarios. Such advancements are envisioned to enhance AI's potential in pivotal areas like drug discovery and climate modeling, where innovative, well-founded conclusions are essential.