It's interesting to note that the newer LLMs like deepseek v4 or kimi k3 basically use caveman mode natively for their thinking traces. Lot word dropping when thinking.