Why spend compute at inference: CoT, o1/o3, DeepSeek-R1 (GRPO, the "aha moment"), best-of-N, PRMs, search, scaling laws, and the limits.
