For years, making language models smarter meant making them bigger. Then someone asked a different question: what if, instead of training more, you let the model think longer?

In September 2024, OpenAI released o1. In January 2025, DeepSeek released R1. These two models together, invalidated the assumption that was governing the entire field...