Inference-time scaling
Spend more compute at inference (longer chain-of-thought, beam search, MCTS) instead of training a larger model. Made famous by OpenAI o1 (September 2024).
The "Bitter Lesson 2.0" — for some tasks (math, coding) it's cheaper to think longer than to train bigger. Used by o1, o3, DeepSeek-R1, Claude (extended thinking mode), Gemini 2.0 Flash Thinking. Drawback: each query takes 10-100× longer, can cost minutes of compute.