<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>楠屿札记 · 手记与创作</title>
    <link>https://7stevejohn.github.io/</link>
    <description>技术手记、小说、游戏创作与随笔，想到什么写什么。</description>
    <language>zh-cn</language>
    <lastBuildDate>Mon, 05 Oct 2026 00:00:00 GMT</lastBuildDate>
    <atom:link href="https://7stevejohn.github.io/rss.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>七、AI 的下一程：算力的账、走出屏幕的智能与治理的规则</title>
      <link>https://7stevejohn.github.io/?post=ai-frontier-mirror</link>
      <guid isPermaLink="false">ai-frontier-mirror</guid>
      <pubDate>Mon, 05 Oct 2026 00:00:00 GMT</pubDate>
      <category>AI 学习</category>
      <description>AI 的下一程看三件事：算力的账、走出屏幕的智能、治理的规则；系列收尾落回终身学习与人机协同。</description>
    </item>
    <item>
      <title>六、AI 会抢走谁的饭碗：为什么不必都挤向开发岗</title>
      <link>https://7stevejohn.github.io/?post=ai-jobs-rethink</link>
      <guid isPermaLink="false">ai-jobs-rethink</guid>
      <pubDate>Sat, 03 Oct 2026 00:00:00 GMT</pubDate>
      <category>AI 学习</category>
      <description>AI 抢的是任务，不是岗位。拆解电网、矿山、政务、医疗落地案例，看清行业复合型人才为何最缺。</description>
    </item>
    <item>
      <title>五、中国 AI 的长板与短板：为什么落地快、原创少</title>
      <link>https://7stevejohn.github.io/?post=ai-china-landscape</link>
      <guid isPermaLink="false">ai-china-landscape</guid>
      <pubDate>Thu, 01 Oct 2026 00:00:00 GMT</pubDate>
      <category>AI 学习</category>
      <description>中国 AI 的真实画像是落地快、原创少——同一结构的一体两面；拆开长板与短板，以及对普通人的含义。</description>
    </item>
    <item>
      <title>刷机为什么会卡开机：验证启动的信任链从哪开始</title>
      <link>https://7stevejohn.github.io/?post=flash-vbmeta-trust-chain</link>
      <guid isPermaLink="false">flash-vbmeta-trust-chain</guid>
      <pubDate>Wed, 30 Sep 2026 00:00:00 GMT</pubDate>
      <category>刷机</category>
      <description>改完分区开不了机，问题通常不在分区写错，在于校验没过。记录 Android 13 的 init_boot 拆分与 AVB 2.0 的校验顺序，以及 9008 这条底层通道到底强在哪。</description>
    </item>
    <item>
      <title>root 藏不住的原因：三层隐藏各管哪一段</title>
      <link>https://7stevejohn.github.io/?post=root-hiding-layers</link>
      <guid isPermaLink="false">root-hiding-layers</guid>
      <pubDate>Wed, 30 Sep 2026 00:00:00 GMT</pubDate>
      <category>刷机</category>
      <description>装了 root 之后支付和银行就不让用，于是有了隐藏。记录注入、命名空间、内核三个层次各自解决什么，以及跨过本机之后那道改不了的云端校验。</description>
    </item>
    <item>
      <title>两年刷机记：从 K60 Pro 到拯救者 Y700</title>
      <link>https://7stevejohn.github.io/?post=two-years-of-flashing</link>
      <guid isPermaLink="false">two-years-of-flashing</guid>
      <pubDate>Wed, 30 Sep 2026 00:00:00 GMT</pubDate>
      <category>随笔</category>
      <description>一台手机、一块平板，从装个 Magisk 开始，到能自己写模块、自己签 vbmeta。记一下这两年是怎么被具体问题推着往前走的。</description>
    </item>
    <item>
      <title>四、Agent 为什么会翻车：智能体的能力边界在哪</title>
      <link>https://7stevejohn.github.io/?post=ai-agent-boundary</link>
      <guid isPermaLink="false">ai-agent-boundary</guid>
      <pubDate>Tue, 29 Sep 2026 00:00:00 GMT</pubDate>
      <category>AI 学习</category>
      <description>智能体翻车是常态而非意外：拆解函数调用与规划幻觉的机制、实验室与现场的落差，以及复合型人才的机会。</description>
    </item>
    <item>
      <title>三、学 AI 要不要先啃高数：普通人现实可走的四条路</title>
      <link>https://7stevejohn.github.io/?post=ai-learning-paths</link>
      <guid isPermaLink="false">ai-learning-paths</guid>
      <pubDate>Sun, 27 Sep 2026 00:00:00 GMT</pubDate>
      <category>AI 学习</category>
      <description>先选路，再补课。拆四条学习路径的门槛与回报，主推深耕行业叠加 AI 的复合路线，附避坑与免费资源清单。</description>
    </item>
    <item>
      <title>8GB 显存跑本地大模型：先算显存账，再挑模型</title>
      <link>https://7stevejohn.github.io/?post=local-llm-8gb-vram</link>
      <guid isPermaLink="false">local-llm-8gb-vram</guid>
      <pubDate>Sat, 26 Sep 2026 00:00:00 GMT</pubDate>
      <category>本地 AI</category>
      <description>第一次拉模型下来加载失败，才知道参数量不等于显存占用。记录 8GB 显卡上按任务分层选型的思路，以及我踩反过的顺序。</description>
    </item>
    <item>
      <title>二、大模型是怎么炼成的：从数据清洗到人类对齐</title>
      <link>https://7stevejohn.github.io/?post=ai-model-training</link>
      <guid isPermaLink="false">ai-model-training</guid>
      <pubDate>Fri, 25 Sep 2026 00:00:00 GMT</pubDate>
      <category>AI 学习</category>
      <description>预训练给智力，后训练给性格。拆解数据清洗到人类对齐这条流水线，以及越狱的本质、对齐税和训练与推理两本算力账。</description>
    </item>
    <item>
      <title>一、Token 是什么：把 AI 圈最容易用错的几个词说清楚</title>
      <link>https://7stevejohn.github.io/?post=ai-concepts-primer</link>
      <guid isPermaLink="false">ai-concepts-primer</guid>
      <pubDate>Wed, 23 Sep 2026 00:00:00 GMT</pubDate>
      <category>AI 学习</category>
      <description>AI 圈一半争论源于词没对齐。拆解 Token、上下文窗口、温度系数，以及三个流传最广的误区。</description>
    </item>
    <item>
      <title>ComfyUI 工作流实践：把一次出图拆成能复用的节点</title>
      <link>https://7stevejohn.github.io/?post=comfyui-workflow-sdxl</link>
      <guid isPermaLink="false">comfyui-workflow-sdxl</guid>
      <pubDate>Fri, 18 Sep 2026 00:00:00 GMT</pubDate>
      <category>AIGC</category>
      <description>一键出图工具满足不了「固定画风 + 保持构图 + 能复用」，换成节点式工作流之后才有解。记录 SDXL 在 8GB 上的节点排法与显存坑。</description>
    </item>
    <item>
      <title>Ollama 显存适配：默认参数是给显存够用的机器配的</title>
      <link>https://7stevejohn.github.io/?post=ollama-vram-fit</link>
      <guid isPermaLink="false">ollama-vram-fit</guid>
      <pubDate>Sat, 12 Sep 2026 00:00:00 GMT</pubDate>
      <category>本地 AI</category>
      <description>模型跑起来了，但接进自己程序之后开始变慢、偶发失败。记录我在 8GB 上收回来的几个参数，以及把本地模型当「可选项」设计这条底线。</description>
    </item>
    <item>
      <title>像素叙事游戏开发手记</title>
      <link>https://7stevejohn.github.io/?post=pixel-narrative-devlog</link>
      <guid isPermaLink="false">pixel-narrative-devlog</guid>
      <pubDate>Wed, 09 Sep 2026 00:00:00 GMT</pubDate>
      <category>游戏创作</category>
      <description>在做的 2D 像素叙事游戏，Godot 4 引擎。记录大纲与卷零的进度，完成后先放卷一 Demo。</description>
    </item>
    <item>
      <title>本地文件管家：秒搜与可撤销怎么同时成立</title>
      <link>https://7stevejohn.github.io/?post=filebutler-local-rag</link>
      <guid isPermaLink="false">filebutler-local-rag</guid>
      <pubDate>Wed, 02 Sep 2026 00:00:00 GMT</pubDate>
      <category>架构设计</category>
      <description>「直接问文件」这件事，难的不是接个模型，是让它既快又能反悔。记录 FileButler 的双索引分工、可撤销设计，以及端侧 AI 必须做的降级底线。</description>
    </item>
    <item>
      <title>帧率防抖：平均 300 帧为什么不跟手</title>
      <link>https://7stevejohn.github.io/?post=gameboost-frame-pacing</link>
      <guid isPermaLink="false">gameboost-frame-pacing</guid>
      <pubDate>Mon, 10 Aug 2026 00:00:00 GMT</pubDate>
      <category>系统调优</category>
      <description>平均 300 帧还是不跟手，问题出在单帧生成时间的尖刺上。记录 GameBoost 针对 1% Low 做的几件事，以及动系统设置时那条不能省的底线。</description>
    </item>
  </channel>
</rss>
