蒸馏是模仿，学强模型的输出，把它的「答案形状」复制过来；RL 是探索，模型必须大量自己推理、自己生成、在错误里反复迭代，从试错中提炼能力。
In a recent update made to Cloudflare Workers, I made similar kinds of modifications to an internal data pipeline that reduced the number of JavaScript promises created in certain application scenarios by up to 200x. The result is several orders of magnitude improvement in performance in those applications.
。爱思助手下载最新版本对此有专业解读
Tilly's mother, Jenny said her family felt like they had been left without support
increase productivity