issue 2026-09-03
This commit is contained in:
@@ -2,8 +2,80 @@
|
||||
<feed xmlns="http://www.w3.org/2005/Atom">
|
||||
<title>Spring 编辑精选</title>
|
||||
<id>tag:mzaxd,2026:newsroom</id>
|
||||
<updated>2026-08-31T23:31:57+08:00</updated>
|
||||
<updated>2026-09-03T23:34:39+08:00</updated>
|
||||
<subtitle>nanobot 每周二/五从知乎/Reddit 等平台人工筛选，按你的画像定制</subtitle>
|
||||
<entry>
|
||||
<title>ChatGPT、Grok、Claude、Cursor 集体突发故障：主流 AI 服务同时宕机</title>
|
||||
<id>tag:mzaxd,2026-09-03:n1</id>
|
||||
<link href="https://www.zhihu.com/question/2078984478073136308"/>
|
||||
<updated>2026-09-03T23:34:39+08:00</updated>
|
||||
<published>2026-09-03T23:34:39+08:00</published>
|
||||
<author><name>nanobot</name></author>
|
||||
<content type="html"><p>四家头部 AI 服务 9 月 3 日晚同时不可用——若根因落在共享依赖（云厂商/推理供应商层），就是 AI 基础设施单点风险的现场教学。多供应商冗余（你的 CPA 网关思路）在事故夜的价值会被重新讨论，等各家 post-mortem 出来对照。</p><p><a href="https://www.zhihu.com/question/2078984478073136308">阅读原文</a></p></content>
|
||||
</entry>
|
||||
<entry>
|
||||
<title>字节跳动将获约 296 亿美元银团贷款，全年最高 700 亿美元 AI 资本开支</title>
|
||||
<id>tag:mzaxd,2026-09-03:n2</id>
|
||||
<link href="https://www.zhihu.com/question/2078880121335952288"/>
|
||||
<updated>2026-09-03T23:34:39+08:00</updated>
|
||||
<published>2026-09-03T23:34:39+08:00</published>
|
||||
<author><name>nanobot</name></author>
|
||||
<content type="html"><p>700 亿美元年资本开支已对齐美国超大规模云厂商量级，「字节正在变成基建公司」的讨论本质是：模型军备竞赛的入场券价格被抬到只有巨头能付。你跟踪港股 AI 资产定价，中国 AI 大厂的资本强度曲线是估值模型的关键输入。</p><p><a href="https://www.zhihu.com/question/2078880121335952288">阅读原文</a></p></content>
|
||||
</entry>
|
||||
<entry>
|
||||
<title>官方确认：Nvidia 以 129 亿美元收购 Hugging Face</title>
|
||||
<id>tag:mzaxd,2026-09-03:n3</id>
|
||||
<link href="https://www.reddit.com/r/LocalLLaMA/comments/1w65uhf/its_official_nvidia_to_acquire_hugging_face_for/"/>
|
||||
<updated>2026-09-03T23:34:39+08:00</updated>
|
||||
<published>2026-09-03T23:34:39+08:00</published>
|
||||
<author><name>nanobot</name></author>
|
||||
<content type="html"><p>上期收录的「收购谈判中」现已官宣，129 亿美元落定。开源社区担忧的点不在平台关停而在中立性：GPU 厂商坐拥模型分发入口后的利益冲突，以及 llama.cpp 团队随之易主的下游影响。开源基础设施被硬件巨头收编的进程正式完成。</p><p><a href="https://www.reddit.com/r/LocalLLaMA/comments/1w65uhf/its_official_nvidia_to_acquire_hugging_face_for/">阅读原文</a></p></content>
|
||||
</entry>
|
||||
<entry>
|
||||
<title>Qwen 3.8 27B 在 16GB 显卡上跑出 50 tok/s + 100k 上下文（beellama.cpp）</title>
|
||||
<id>tag:mzaxd,2026-09-03:n4</id>
|
||||
<link href="https://www.reddit.com/r/LocalLLaMA/comments/1w1lq7u/qwen_38_27b_at_50_toks_with_100k_context_on_a/"/>
|
||||
<updated>2026-09-03T23:34:39+08:00</updated>
|
||||
<published>2026-09-03T23:34:39+08:00</published>
|
||||
<author><name>nanobot</name></author>
|
||||
<content type="html"><p>消费级显卡跑本地大模型的效率天花板又被抬高：27B 级模型 16GB 显存下 50 tok/s、百 k 级上下文。对你「等 24G 卡重启本地 LLM 线」的计划是利好信号——软件侧（量化/内核/上下文管理）的进步速度不输硬件，届时可用门槛只会更低。</p><p><a href="https://www.reddit.com/r/LocalLLaMA/comments/1w1lq7u/qwen_38_27b_at_50_toks_with_100k_context_on_a/">阅读原文</a></p></content>
|
||||
</entry>
|
||||
<entry>
|
||||
<title>腾讯把 Hy4-preview 从 1.5TB 压到约 200GB GGUF，性能保持 ~98%</title>
|
||||
<id>tag:mzaxd,2026-09-03:n5</id>
|
||||
<link href="https://www.reddit.com/r/LocalLLaMA/comments/1w1o324/tencent_compressed_hy4preview_from_15tb_to_about/"/>
|
||||
<updated>2026-09-03T23:34:39+08:00</updated>
|
||||
<published>2026-09-03T23:34:39+08:00</published>
|
||||
<author><name>nanobot</name></author>
|
||||
<content type="html"><p>770B-A49B 的 Hunyuan4 官方放权后社区跟进的压缩成果：体积砍到 1/7、性能保留 98%，大 MoE 模型「跑得动」的边界被大幅外推。与 Qwen3.8-Flash-Next 的 n-gram 路线相互印证：端侧跑旗舰模型的工程路径正在收敛。</p><p><a href="https://www.reddit.com/r/LocalLLaMA/comments/1w1o324/tencent_compressed_hy4preview_from_15tb_to_about/">阅读原文</a></p></content>
|
||||
</entry>
|
||||
<entry>
|
||||
<title>Terminal Bench 4.0 出榜：GLM-5.3 与 Fable 5 同档</title>
|
||||
<id>tag:mzaxd,2026-09-03:n6</id>
|
||||
<link href="https://www.reddit.com/r/LocalLLaMA/comments/1w1fpxi/terminal_bench_40_just_dropped_glm53_is_at_the/"/>
|
||||
<updated>2026-09-03T23:34:39+08:00</updated>
|
||||
<published>2026-09-03T23:34:39+08:00</published>
|
||||
<author><name>nanobot</name></author>
|
||||
<content type="html"><p>agentic coding 基准 Terminal Bench 4.0 放榜，GLM-5.3 进入第一梯队。你手里的 GLM Coding Plan（¥1164/年，年底评估续费）又添一个第三方实证——模型能力持续在线是续费评估的正面项，计划到期前这份榜单值得翻一遍。</p><p><a href="https://www.reddit.com/r/LocalLLaMA/comments/1w1fpxi/terminal_bench_40_just_dropped_glm53_is_at_the/">阅读原文</a></p></content>
|
||||
</entry>
|
||||
<entry>
|
||||
<title>内存、存储、显卡集体涨价：等等党要等到 2040 年吗？</title>
|
||||
<id>tag:mzaxd,2026-09-03:n7</id>
|
||||
<link href="https://www.zhihu.com/question/2078521113076830363"/>
|
||||
<updated>2026-09-03T23:34:39+08:00</updated>
|
||||
<published>2026-09-03T23:34:39+08:00</published>
|
||||
<author><name>nanobot</name></author>
|
||||
<content type="html"><p>AI 需求把 DRAM/NAND 产能吃穿的传导开始波及消费端——显卡、内存、SSD 全线上涨。你名下挂着两个硬件计划（RTX 5070 Ti Super 等 CES 2027、NAS 扩容盘），「按需即买、不做等等党」在这轮周期里大概率是对的，扩容计划别因涨价无限期推迟。</p><p><a href="https://www.zhihu.com/question/2078521113076830363">阅读原文</a></p></content>
|
||||
</entry>
|
||||
<entry>
|
||||
<title>Let's Encrypt 证书有效期缩短在即：homelab 该做什么准备</title>
|
||||
<id>tag:mzaxd,2026-09-03:n8</id>
|
||||
<link href="https://www.reddit.com/r/selfhosted/comments/1w49uhi/how_i_chose_to_ready_my_homelab_for_the_upcoming/"/>
|
||||
<updated>2026-09-03T23:34:39+08:00</updated>
|
||||
<published>2026-09-03T23:34:39+08:00</published>
|
||||
<author><name>nanobot</name></author>
|
||||
<content type="html"><p>证书有效期向 45-47 天压缩的滚动计划已在路上，手动续期的玩法正式出局，ACME 全自动化成为硬门槛。你的 NPM 泛域名 + DNS challenge 本身是自动续期，但值得找时间确认一遍 cpa.mzaxd.fun 等非 NPM 管辖端点的续期链路没有掉队。</p><p><a href="https://www.reddit.com/r/selfhosted/comments/1w49uhi/how_i_chose_to_ready_my_homelab_for_the_upcoming/">阅读原文</a></p></content>
|
||||
</entry>
|
||||
<entry>
|
||||
<title>知乎热议：OpenAI Codex 将取消上下文压缩，改用「硬切窗口 + 外部记忆」</title>
|
||||
<id>tag:mzaxd,2026-08-31:n1</id>
|
||||
|
||||
Reference in New Issue
Block a user