This AI Spent 28 Minutes Thinking Before Answering — And Got It Right

DeepSeek R1 generates 100K thinking tokens before responding. Inference scaling trades speed for accuracy — and it's working.