AlphaGo won against Lee Sedol in 2016 using reasoning not just intuition. Its policy network guessed human-like moves. Its search machinery weighed future consequences explicitly. Today’s large language models only use fast pattern completion. Chain-of-thought methods still rely on next-token prediction. They lack an explicit inspectable epistemic state. They do not separate knowledge from reasoning. They often fabricate reasoning steps after the fact. The author argues future AI needs AlphaGo-like reasoning for trustworthy results. Such systems should maintain and update an epistemic state. They must separate knowledge from reasoning. They should audit conclusions via evidence and inference. The author left Google DeepMind to pursue this approach.