Ten years after AlphaGo's historic match against Lee Sedol, a critical gap remains between past and present AI capabilities. In 2016, AlphaGo's famous 'Move 37' demonstrated a unique blend of intuition and calculation. Unlike Deep Blue, which relied on brute-force computation, AlphaGo used a policy network to generate hunches and a separate search machinery to weigh future consequences. This architecture mimicked human System 1 and System 2 thinking, allowing the AI to evaluate thousands of future game states before selecting a creative move.
Today's large language models operate differently. They predict the next token iteratively, relying primarily on fast, associative pattern completion. While techniques like chain-of-thought prompting encourage intermediate steps, they do not create a distinct reasoning engine. The intermediate output is still generated by the same next-token prediction process, meaning modern AI lacks the genuine deliberative machinery that powered AlphaGo's success. To achieve trustworthy, novel insights in science and medicine, future systems must integrate a separate, deliberate reasoning process rather than relying solely on pattern completion.
Source: MIT Technology Review AI · Summarized by HeadlinesBriefing