无双的技术博客 记录 AI、Linux、网络架构、FreeSWITCH 与企业数字化实践

Why Another GPU's 4-Bit Speedup Does Not Transfer to Your RTX 3090

Separate smaller weights from faster inference: check formats, kernels, memory, and compatibility on an RTX 3090.

Administrator Administrator 发布于 2026-09-25

都是 4bit 量化,为什么别人的加速结果不能照搬到你的 3090?

分清 4bit 权重、原生 FP4 计算和实际推理内核,判断量化模型在 3090 上能否运行、节省显存,以及是否值得部署。

Administrator Administrator 发布于 2026-09-25

Build One Useful AI Skill Before Installing a Whole Collection

Turn work notes into a reusable reporting skill, with source checks that keep unfinished work out of the done list.

Administrator Administrator 发布于 2026-09-25

别急着装一堆 AI Skills,先把一个重复任务做成自己的技能

用一份虚构工作记录,做出可复用的周报技能。附完整规则、调用提示和检查方法,避免 AI 把进行中写成已完成。

Administrator Administrator 发布于 2026-09-25

AI Is Still the Copilot—You Need to Learn the Wheel

AI can make work faster without making every answer reliable. Here is how to work with it without giving up judgment.

Administrator Administrator 发布于 2026-08-12

AI再强,也只是副驾驶:你得学会握方向盘

AI能把工作做得更快,却不能替你判断答案是否可靠。普通人该怎样真正开始与AI协作?

Administrator Administrator 发布于 2026-08-12

完全不会开发,也能用 AI 做出一个网站:从想法、计划到部署

从描述需求、生成计划到把代码交给服务器 AI 部署,用可复制提示词完成第一个网站。

Administrator Administrator 发布于 2026-07-29

04-Function Calling Without Handing Control to the Model

Let the model suggest a tool while the application validates, authorizes, executes, and records the action.

Administrator Administrator 发布于 2026-07-28

04-Function Calling 实战:让模型会用工具,也不越权

模型负责建议工具,应用负责校验、授权和执行;用只读查询走完一次安全闭环。

Administrator Administrator 发布于 2026-07-28

03-Web Search, RAG, and Memory Solve Different Problems

A refund-policy query shows which data belongs to web search, controlled retrieval, or conversation memory.

Administrator Administrator 发布于 2026-07-20