Chain-of-thought extraction is a technique used to distill proprietary large language models