LLM

写给初学者的技术报告精读手册：先用一句话说清每个概念要解决什么问题，再拆开公式与结构，最后用一个可交互的小实验亲手拨动参数。纵向是 DeepSeek 的版本演化，横向是与 GPT / Claude / Gemini 的对比，纵深是 MLA、MoE、GRPO、FP8、DSA、CED 这些名词背后的原理。

How eighteen shipped system prompts are actually written, measured across the seven dimensions they all answer — and the argument the note turns on: every 'never' should be answerable with which layer actually stops the model when it disobeys. Written in Chinese, with twenty-three quoted excerpts, each checked verbatim against the archive it came from.