Running some models locally and seeing these thinking tokens was quite the experience. I never saw an LLM so "unsure" about virtually everything.