<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Anaxagore</title>
    <link>https://lgystoic.github.io/</link>
    <description>GPU kernel、训练性能、生成模型和城市数据的中文笔记存档。</description>
    <language>zh-CN</language>
    <atom:link href="https://lgystoic.github.io/feed.xml" rel="self" type="application/rss+xml" />
    <lastBuildDate>Sun, 11 Oct 2026 00:00:00 +0000</lastBuildDate>
    <item>
      <title>想降到 Plus，却先修了小米通行密钥登录</title>
      <link>https://lgystoic.github.io/notes/xiaomi-chatgpt-passkey/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/xiaomi-chatgpt-passkey/</guid>
      <pubDate>Sun, 11 Oct 2026 00:00:00 +0000</pubDate>
      <description>小米14指纹通过后，ChatGPT网页与App登录仍卡住。用日志、精确版本源码和恢复实验追查连接类型兼容问题，记录临时方法与证据边界。</description>
      <category>Android</category>
      <category>小米</category>
      <category>ChatGPT</category>
      <category>Passkeys</category>
      <category>WebAuthn</category>
    </item>
    <item>
      <title>82 亿美元买的不是模型 而是下一代 AI 的工作负载</title>
      <link>https://lgystoic.github.io/notes/amd-world-labs-compute/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/amd-world-labs-compute/</guid>
      <pubDate>Sun, 04 Oct 2026 00:00:00 +0000</pubDate>
      <description>AMD 签约收购 World Labs。本文从 Atlas 的自回归、扩散和空间上下文出发，拆解模型研究如何经过负载、软件、硬件与用户任务五步转换，并明确目前没有公开跨平台性能证据。</description>
      <category>AMD</category>
      <category>World Labs</category>
      <category>Spatial Intelligence</category>
      <category>AI Compute</category>
      <category>ROCm</category>
    </item>
    <item>
      <title>AI 视频会拍镜头 为什么还拍不好故事</title>
      <link>https://lgystoic.github.io/notes/ai-video-shot-continuity/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/ai-video-shot-continuity/</guid>
      <pubDate>Tue, 29 Sep 2026 00:00:00 +0000</pubDate>
      <description>AI 视频的难题正在从生成一个漂亮镜头，变成控制许多镜头的误差。用同一个案例拆解导演、分镜、生成、场记、审片和接戏六个 Agent 怎样维护叙事意图与画面事实。</description>
      <category>AI Video</category>
      <category>Multi-Agent</category>
      <category>Video Generation</category>
      <category>Continuity</category>
      <category>System Design</category>
    </item>
    <item>
      <title>3D 打印的四十年：瓶颈换了四次位</title>
      <link>https://lgystoic.github.io/notes/3d-printing-40-years/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/3d-printing-40-years/</guid>
      <pubDate>Fri, 25 Sep 2026 00:00:00 +0000</pubDate>
      <description>从 1981 年到 2026 年，3D 打印这个行业经历了四次瓶颈的换位：能不能做出来、买不买得起、会不会用、打什么。前三个基本解决了，最后一个由 AI 接管。覆盖专利到期、价格崩盘、拓竹冲击波、FDM vs 光固化双产线、Tripo vs Meshy 路线对决。</description>
      <category>3D Printing</category>
      <category>FDM</category>
      <category>SLA</category>
      <category>Bambu Lab</category>
      <category>Tripo</category>
      <category>Meshy</category>
      <category>Tech History</category>
    </item>
    <item>
      <title>AI 是怎么学会捏 3D 模型的：生成式 3D 六年演进</title>
      <link>https://lgystoic.github.io/notes/generative-3d-ai/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/generative-3d-ai/</guid>
      <pubDate>Thu, 24 Sep 2026 00:00:00 +0000</pubDate>
      <description>从 NeRF 到 3DGS，从 SDS 到原生 mesh 生成，再到 2026 年的 flow matching：生成式 3D 这六年没有一步是拍脑袋想出来的，每一步都是被上一步的缺陷逼出来的。按缺陷驱动的顺序串起学术脉络与商业战场，附 8 张机制图与论文速查表。</description>
      <category>3D Generation</category>
      <category>NeRF</category>
      <category>3DGS</category>
      <category>Mesh</category>
      <category>Flow Matching</category>
    </item>
    <item>
      <title>用 VLM 把家庭摄像头压成状态序列：一个看护记录系统的设计</title>
      <link>https://lgystoic.github.io/notes/home-camera-vlm-pipeline/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/home-camera-vlm-pipeline/</guid>
      <pubDate>Sat, 19 Sep 2026 00:00:00 +0000</pubDate>
      <description>三个家用摄像头接一个视觉模型，把连续画面变成可查询的状态序列。记录其中的取舍：帧差闸门如何把模型调用压到接近零、输出格式为什么不用 JSON、状态机为什么必须忽略 unknown，以及门口那一路为什么从轮询改成边沿触发。</description>
      <category>VLM</category>
      <category>System Design</category>
      <category>成本优化</category>
      <category>状态机</category>
      <category>家庭自动化</category>
    </item>
    <item>
      <title>深圳地铁 200 米内住宅画像：南山 / 宝安 / 光明</title>
      <link>https://lgystoic.github.io/notes/2026-08-17-metro200m-residential-profile/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/2026-08-17-metro200m-residential-profile/</guid>
      <pubDate>Tue, 18 Aug 2026 00:00:00 +0000</pubDate>
      <description>对南山、宝安、光明三区地铁 200 米内 461 个地块做住宅粗筛，并深挖 120 个候选盘的户数、车位、户型、回迁房与学位风险。</description>
      <category>深圳</category>
      <category>买房</category>
      <category>地铁</category>
      <category>住宅画像</category>
      <category>教育</category>
    </item>
    <item>
      <title>Ring Collectives</title>
      <link>https://lgystoic.github.io/notes/ring-collectives/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/ring-collectives/</guid>
      <pubDate>Thu, 13 Aug 2026 00:00:00 +0000</pubDate>
      <description>Interactive visual explanation of ring all-reduce, reduce-scatter, all-gather, all-to-all, and NCCL bus bandwidth factors.</description>
      <category>NCCL</category>
      <category>Distributed Training</category>
      <category>GPU</category>
      <category>Collectives</category>
    </item>
    <item>
      <title>深圳离地铁站直线 200 米内的小区全清单</title>
      <link>https://lgystoic.github.io/notes/2026-08-10-shenzhen-metro-200m/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/2026-08-10-shenzhen-metro-200m/</guid>
      <pubDate>Mon, 10 Aug 2026 00:00:00 +0000</pubDate>
      <description>用 OSM 数据穷举深圳地铁出入口 200 米内的小区清单，覆盖 1150 个地块、261 个站点，并提供分区 CSV/JSON 下载。</description>
      <category>深圳</category>
      <category>地铁</category>
      <category>住房</category>
      <category>OpenStreetMap</category>
    </item>
    <item>
      <title>SwiGLU MLP 的前向原理、反向链条与优化思路</title>
      <link>https://lgystoic.github.io/notes/swiglu-mlp-forward-backward-optimization/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/swiglu-mlp-forward-backward-optimization/</guid>
      <pubDate>Thu, 11 Jun 2026 00:00:00 +0000</pubDate>
      <description>从前向 shape 到反向 4 个 GEMM + 1 组逐元素，解释 dout 不落 HBM 的 CuTe 融合优化。</description>
      <category>SwiGLU</category>
      <category>MLP</category>
      <category>CUDA</category>
      <category>B200</category>
      <category>CuTe</category>
    </item>
    <item>
      <title>SwiGLU MLP Drop-H 技术报告</title>
      <link>https://lgystoic.github.io/notes/swiglu-mlp-drop-h/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/swiglu-mlp-drop-h/</guid>
      <pubDate>Fri, 05 Jun 2026 00:00:00 +0000</pubDate>
      <description>B200 上 SwiGLU MLP drop-h 方案的公式、实现、性能与显存取舍。</description>
      <category>GPU</category>
      <category>Triton</category>
      <category>SwiGLU</category>
      <category>MLP</category>
      <category>B200</category>
    </item>
    <item>
      <title>PyTorch Autograd 学习笔记</title>
      <link>https://lgystoic.github.io/notes/pytorch-autograd-notes/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/pytorch-autograd-notes/</guid>
      <pubDate>Wed, 03 Jun 2026 00:00:00 +0000</pubDate>
      <description>从 Tensor 底层结构到 grad_fn、saved tensors、version counter、no_grad、inference_mode 和自定义 Function 的 autograd 知识地图。</description>
      <category>PyTorch</category>
      <category>Autograd</category>
      <category>Deep Learning</category>
      <category>Notes</category>
    </item>
    <item>
      <title>Kernel Design Agents：KernelWiki 与 ncu-report-skill 拆解</title>
      <link>https://lgystoic.github.io/notes/kernel-design-agents-skills/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/kernel-design-agents-skills/</guid>
      <pubDate>Mon, 01 Jun 2026 00:00:00 +0000</pubDate>
      <description>拆解 KDA 仓库中 KernelWiki 知识库和 ncu-report-skill profiling 工作流，整理它们如何服务 agentic kernel 优化循环。</description>
      <category>CUDA</category>
      <category>Blackwell</category>
      <category>KernelWiki</category>
      <category>Nsight Compute</category>
    </item>
    <item>
      <title>Fused SwiGLU Wide Packed Save-Factors</title>
      <link>https://lgystoic.github.io/notes/swiglu-packed-save-factors/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/swiglu-packed-save-factors/</guid>
      <pubDate>Mon, 01 Jun 2026 00:00:00 +0000</pubDate>
      <description>B200 上 packed W1 + save-factors SwiGLU 融合实现笔记，覆盖 pack layout、CTA tile、forward/backward 数据流和显存收益。</description>
      <category>Triton</category>
      <category>SwiGLU</category>
      <category>Blackwell</category>
      <category>CUDA</category>
    </item>
    <item>
      <title>从噪声到数据：Flow Matching 与扩散模型中文地图</title>
      <link>https://lgystoic.github.io/notes/mit-6s184-flow-diffusion/</link>
      <guid isPermaLink="true">https://lgystoic.github.io/notes/mit-6s184-flow-diffusion/</guid>
      <pubDate>Sat, 30 May 2026 00:00:00 +0000</pubDate>
      <description>基于 MIT 6.S184 讲义整理，覆盖生成即采样、ODE/SDE、Flow Matching、Score Matching、Classifier-Free Guidance、潜空间生成器和离散扩散语言模型。</description>
      <category>Flow Matching</category>
      <category>Diffusion</category>
      <category>SDE</category>
      <category>CFG</category>
    </item>
  </channel>
</rss>
