<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/">
    <channel>
        <title>智汇AI</title>
        <link>http://easyai.fyi/</link>
        <description>每天看懂AI新闻、知识与工具，给产品经理、运营和AI初学者的人工智能科普站。</description>
        <lastBuildDate>Thu, 02 Jul 2026 15:50:22 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>https://github.com/jpmonette/feed</generator>
        <language>zh-CN</language>
        <copyright>All rights reserved 2026, 智汇AI</copyright>
        <item>
            <title><![CDATA[Fable 5 回归与 Agentic 工作流 - 2026-07-02]]></title>
            <link>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-02</link>
            <guid>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-02</guid>
            <pubDate>Thu, 02 Jul 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[今日全量沉淀包括 17 位 X builders 的 35 条推文、0 篇博客和 1 期 AI & I 播客。主线是 Fable 5 回归、agentic mapreduce 和 AI 工作流进入更真实的产品与企业场景。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-3915eac3813c814388e9f31f014c3fe2"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-a0aa78fd6cff4e3db2d1f0b6dda15e86" data-id="a0aa78fd6cff4e3db2d1f0b6dda15e86"><span><div id="a0aa78fd6cff4e3db2d1f0b6dda15e86" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a0aa78fd6cff4e3db2d1f0b6dda15e86" title="今日主线"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">今日主线</span></span></h3><div class="notion-text notion-block-2483361381644c67b79d429f3be873af">今天这批 follow-builders 内容的核心是 Fable 5 正式回到开发者视野，同时 Agent 正在从单个 coding assistant 走向更复杂的工作流：sandboxing、world models、agentic mapreduce、dry-run deployment、技能沉淀、企业 AI 工作流都在同一天出现。全量数据包括 17 位 X builders 的 35 条推文、0 篇博客和 1 期 AI &amp; I 播客。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-ba0f290da16447869fa66ae9a2b16630" data-id="ba0f290da16447869fa66ae9a2b16630"><span><div id="ba0f290da16447869fa66ae9a2b16630" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ba0f290da16447869fa66ae9a2b16630" title="重点解读"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">重点解读</span></span></h3><div class="notion-text notion-block-81c5af53184444d2b0233b294cb4bcb0">Fable 5 是今天最明显的热点。Claude 官方说付费计划可以在 7 月 7 日前使用 Fable 5，最多可占 weekly usage limit 的 50%，误触发分类器时可以用反馈入口上报。Peter Yang 做了 Fable 5 vibe check，认为它仍然比其他模型高一个台阶；Alex Albert、Nikunj Kothari、Dan Shipper 也都在庆祝它回归。这说明前沿 coding model 的访问状态已经会直接影响 builder 圈的工作节奏。</div><div class="notion-text notion-block-3bf278d67d1e40749f7f47b85029e9c6">Agent 的系统形态也在继续变复杂。Swyx 提到 AIE 里 sandboxing 和 world models 的双倍时长 keynote 受到欢迎；Aaron Levie 转述 Devin 的 agentic mapreduce，把大量代码信号拆成 shards，让多个 focused agents 并行处理，再汇总验证。Guillermo Rauch 则说 Vercel 正在做 agentic deployments 的 dry-run step，用来降低成本和风险。这些内容放在一起看，Agent 正在从“一个模型做任务”变成“多 agent、可验证、可部署、可控风险”的工程系统。</div><div class="notion-text notion-block-dc5f7a1686ec44e7a627473b86811492">今天的播客也接上了这个方向。AI &amp; I 采访 Every 咨询团队，重点不是展示一个酷工具，而是讲企业里的 AI 工作流怎么真正落地：Codex 帮人搭出类似操作系统的流程，邮件和资料能比人更快掌握上下文，知识工作越来越像园艺，不是亲手造出每片叶子，而是创造条件，让系统自己持续生长。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-1d6ae4cc65ef4746b902717fa1cc651b" data-id="1d6ae4cc65ef4746b902717fa1cc651b"><span><div id="1d6ae4cc65ef4746b902717fa1cc651b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#1d6ae4cc65ef4746b902717fa1cc651b" title="X Builders 全量记录"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">X Builders 全量记录</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-c10eecc94ac247acaabc8878ebe24e7c" data-id="c10eecc94ac247acaabc8878ebe24e7c"><span><div id="c10eecc94ac247acaabc8878ebe24e7c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c10eecc94ac247acaabc8878ebe24e7c" title="1. Swyx"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">1. Swyx</span></span></h4><div class="notion-text notion-block-68ce78e819424f9a8b1ec3158bd0c8f9">Swyx 今天提到 AIE 的双倍时长 track keynote。他说只有在 speaker 和内容都值得时，才会安排 double-length keynote；这次 Chris Manning 和 Abhishek Bhardwaj 分别围绕 sandboxing 和 world models 做了更深讲解，现场房间很满，线上观众未来会多很多倍。重点不是活动热闹，而是 sandboxing 和 world models 已经从研究词变成 AI engineering 社区真正想深入学的主题。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-c75b1ccc8eff435b81566d183ad97f17"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072562702703046855" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-ff8b0102be294b03a853d94e6a0bcdba" data-id="ff8b0102be294b03a853d94e6a0bcdba"><span><div id="ff8b0102be294b03a853d94e6a0bcdba" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ff8b0102be294b03a853d94e6a0bcdba" title="2. Thibault Sottiaux"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">2. Thibault Sottiaux</span></span></h4><div class="notion-text notion-block-6982e223c0044762bd4b7561d1fdde03">Thibault Sottiaux 只发了一条短帖：“It&#x27;s happening”。上下文来自原帖里的图片和链接，信息本身很短，但按全量沉淀要求保留。结合今天 Fable 5 回归和 Codex 相关讨论，这条更像是在指向某个正在发生的产品或模型访问变化。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-29062648b88647249bb34266376475ec"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072410623380468190" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-a05783563af24f77bf3fa40e55f2c491" data-id="a05783563af24f77bf3fa40e55f2c491"><span><div id="a05783563af24f77bf3fa40e55f2c491" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a05783563af24f77bf3fa40e55f2c491" title="3. Peter Yang"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">3. Peter Yang</span></span></h4><div class="notion-text notion-block-05bbe7bceacd434ea79fd5af9d3df65e">Peter Yang 今天三条都和 AI 编程有关。他说 explain-diff skill 对还在学习读代码的人很有帮助，准备马上安装；随后给 Fable 5 做 vibe check，认为它仍然非常强，是高出其他模型的 step function，并期待 GPT 5.6 能追上。第三条是他的新教程，强调 Fable 5 已经回来，但 Claude 订阅用户只能用到 7 月 7 日，他列了 5 个值得尝试的场景：找 Fable-worthy work、生活和商业建议、让项目 ready to ship、规划大事、重构项目或代码库。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-3f3a0d1f5fe4468bb323f4f7e3dc0e95"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072525669704384612" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-78b82d4214c8457e90b78428ea9bc5c5"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072470191511113732" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-46cf6fc9d98344808bd7c670cbaf259d"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072458983886205333" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-0ad9bfb905d747a5a4588e3fe9c67900" data-id="0ad9bfb905d747a5a4588e3fe9c67900"><span><div id="0ad9bfb905d747a5a4588e3fe9c67900" class="notion-header-anchor"></div><a class="notion-hash-link" href="#0ad9bfb905d747a5a4588e3fe9c67900" title="4. Thariq"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">4. Thariq</span></span></h4><div class="notion-text notion-block-667cc23057fc4a0cafa816e54d293b63">Thariq 今天两条都来自 AI Engineer 现场。一条提到 HTML mentioned，另一条是 hello from AI engineer。信息量不大，但和 Swyx 的现场内容一起看，AIE 这波讨论明显围绕 Agent 工程、HTML、sandboxing、world models 等具体构建主题，而不是单纯模型发布。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-c5137bd331b74440b7ae00b68f232577"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072366310416425053" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-238b84aea8b54f319a9c69c46063f28f"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072360902964511171" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-0c888bfa329a4196965c97b40ba67481" data-id="0c888bfa329a4196965c97b40ba67481"><span><div id="0c888bfa329a4196965c97b40ba67481" class="notion-header-anchor"></div><a class="notion-hash-link" href="#0c888bfa329a4196965c97b40ba67481" title="5. Google Labs"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">5. Google Labs</span></span></h4><div class="notion-text notion-block-76ffbffe043744859443865ab85bd843">Google Labs 宣布为了专注建设 Google Flow Music，会在 2026 年 7 月 31 日告别 MusicFX 和 MusicFX DJ。这两个早期实验推动过实时 AI 音乐创作边界，现在 Google 会把学到的东西带到更长期的音乐项目入口里。它说明 Google 正在把实验性 AI 音乐工具收束进更明确的产品线，而不是一直并行维护多个实验入口。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-6a7c400178fc40a3a41bb7170ef425af"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072417166952136789" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-2bfe8f527a264fd8ac363bc40b32f67d" data-id="2bfe8f527a264fd8ac363bc40b32f67d"><span><div id="2bfe8f527a264fd8ac363bc40b32f67d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2bfe8f527a264fd8ac363bc40b32f67d" title="6. Amjad Masad"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">6. Amjad Masad</span></span></h4><div class="notion-text notion-block-c450f395561241a6a4f409a3aed0ef4d">Amjad Masad 说，现在 building 已经变容易，Replit 越来越关注怎么帮创业者走向市场，拿到第一个客户和第一美元。Whop 是他认为适合变现创作的地方，所以现在可以把 Replit apps 放到 Whop 上卖。这个动作很实际：AI coding 工具如果只停在“做出来”，价值还不完整，下一步是分发、收费和商业闭环。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-9d5c9ca5eb43432c807dd0e4f884f0fa"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072385092824260748" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-9ef1462c227f40b2b2aa78a8de55b803" data-id="9ef1462c227f40b2b2aa78a8de55b803"><span><div id="9ef1462c227f40b2b2aa78a8de55b803" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9ef1462c227f40b2b2aa78a8de55b803" title="7. Guillermo Rauch"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">7. Guillermo Rauch</span></span></h4><div class="notion-text notion-block-cc5a17168c4d4f389d5e3ab1f5db8bd4">Guillermo Rauch 今天三条都和 Vercel 工程体验有关。他提到 portless 对本地 WordPress 开发很有帮助；又展示了 WordPress on Vercel Fluid，用单个 Dockerfile、PlanetScale MySQL、云端 docker build 和 30 秒部署跑起来。最重要的是第三条：Agents 很喜欢在 push 前检查自己的工作，比如 node --check、tsc --noEmit、next build；Vercel 正在发布 agentic deployments 的 dry-run step，用来减少成本和风险。这是把 Agent 的“自检习惯”产品化。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-2682cf5659cf407b986a29a71c74c355"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072463961597878762" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-fc2b8858e0ac43598d2d75c7c907b721"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072463293654942090" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-0114ecdc9929487cb2d78c572e54c05f"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072398926175404250" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-ae4b732ad4db4f5394f2b862ef91b5e7" data-id="ae4b732ad4db4f5394f2b862ef91b5e7"><span><div id="ae4b732ad4db4f5394f2b862ef91b5e7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ae4b732ad4db4f5394f2b862ef91b5e7" title="8. Alex Albert"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">8. Alex Albert</span></span></h4><div class="notion-text notion-block-94d6331d43e54280938fcb309d8f69b7">Anthropic researcher Alex Albert 发了一句“Welcome back to the world Fable!!”。这条很短，但很明确：Fable 5 回来了，而且连 Anthropic 内部研究员也在公开强调它的回归。和 Claude 官方的说明、Peter Yang 的 vibe check 放在一起看，今天 Fable 5 是 builder 圈最强信号之一。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-0d3444e35275488e84b225526cdcdb12"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072404717490360727" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-35f91f7a33f94883ae31eb6227b30a08" data-id="35f91f7a33f94883ae31eb6227b30a08"><span><div id="35f91f7a33f94883ae31eb6227b30a08" class="notion-header-anchor"></div><a class="notion-hash-link" href="#35f91f7a33f94883ae31eb6227b30a08" title="9. Aaron Levie"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">9. Aaron Levie</span></span></h4><div class="notion-text notion-block-028c97bcf83740bd96d32b32cd617af9">Aaron Levie 把 Devin 的 agentic mapreduce 当成未来需要 100 倍 AI inference 的例子。这个模式是：Devin 先在代码库里映射相关信号，再把 bounded shards 分给多个 focused agents，汇总成报告，最后在隔离 sandbox 里验证严重漏洞。Aaron 的判断是，这不只适用于代码安全，也会出现在企业知识工作里，比如 Box 客户想处理数百万文档，理解风险、洞察和关系。这里的重点是 inference 需求会被“多 agent 并行处理大量数据”放大。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-ddddd0c9bee54fbdb9d69771d15100f7"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072519377371459836" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-8fd45db8c37242b49fcde7fe89c709b4" data-id="8fd45db8c37242b49fcde7fe89c709b4"><span><div id="8fd45db8c37242b49fcde7fe89c709b4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#8fd45db8c37242b49fcde7fe89c709b4" title="10. Garry Tan"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">10. Garry Tan</span></span></h4><div class="notion-text notion-block-55a98e99f48049d48f0317253b024485">Garry Tan 今天三条里，一条是 link-only，信息有限；一条说“Most underreported story in all of tech”，但原帖需要上下文；第三条提到 UC Berkeley EECS 负责人加入 Anthropic，称 Anthropic 最近吸引人才非常强。结合最近 labs 人才流动，这条说明 Anthropic 和 OpenAI 这类前沿实验室仍然是顶级人才的强磁场。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-6ea61ddad6fb438c828ac3aca086b195"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072461457195950446" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-7025456d11c94d0fb46741ef3c31440e"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072402517397573717" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-e6570a2bad88415f932b393a7d916db0"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072331451270606933" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-9874123719af4302a650ad55a9eef4c4" data-id="9874123719af4302a650ad55a9eef4c4"><span><div id="9874123719af4302a650ad55a9eef4c4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9874123719af4302a650ad55a9eef4c4" title="11. Matt Turck"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">11. Matt Turck</span></span></h4><div class="notion-text notion-block-d592b0326b8241cfb433e5d9b88dd3f7">Matt Turck 今天讨论的是 IPO 和 Lime。他调侃 2026 年热门 IPO 品牌包括 AOL、Evernote、Vimeo 和 Lime scooters；另一条认真分析 Lime 上市：虽然有 10 亿美元债务，还曾表达持续经营疑虑，但 IPO 还清了高风险贷款、把剩余债务转成股权，Uber 持有 22% 并导流，而且 Lime 连续 3 年 FCF 为正、收入接近 30% 年增长。这条不是 AI 主线，但属于 tech/business 观察，完整保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-2ad70ef361784db2b516de17483629eb"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072462125474181623" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-bb5bbdf57aec4af39f9cad5accce08f3"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072419592354529712" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-71d4cf1872c84dc889369498333c0e34" data-id="71d4cf1872c84dc889369498333c0e34"><span><div id="71d4cf1872c84dc889369498333c0e34" class="notion-header-anchor"></div><a class="notion-hash-link" href="#71d4cf1872c84dc889369498333c0e34" title="12. Zara Zhang"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">12. Zara Zhang</span></span></h4><div class="notion-text notion-block-fba66094f2a34d729dd1eed3530ecd67">Zara Zhang 今天三条都围绕 builders 和技能。一条提醒可以把 Codex 的模型改成 GLM，这和中国 open model 在开发者工具里的渗透有关；一条放了 YouTube 完整视频；最有方法论价值的是“HOW TO BUILD A SKILL”：她说不要从 skill 开始，而是以 skill 结束，先完成真实工作流，再把它沉淀成 skill。这句话很实用，因为 skill 的价值来自被验证过的流程，而不是一开始就写抽象规范。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-34c4fdde4cb64a5f9a526fb5062114d3"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072391971721884073" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-4c46c4441c2e467db7089b88f23684ee"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072384777785888875" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-f349c3cce3364a71af1fbc9aebca147a"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072381929366987087" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-e775196650b84ad695c10a04116aeff5" data-id="e775196650b84ad695c10a04116aeff5"><span><div id="e775196650b84ad695c10a04116aeff5" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e775196650b84ad695c10a04116aeff5" title="13. Nikunj Kothari"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">13. Nikunj Kothari</span></span></h4><div class="notion-text notion-block-96787f48a6ab40d3b6e73c015a502564">Nikunj Kothari 今天三条里，一条是美国足球，非 AI 主线；一条很直接地说 Fable is back；第三条讨论 OpenAI 和 Anthropic 的人才吸引力。他说过去两个月有四位朋友从非常成熟的岗位离开，加入 labs，原因是公司重要性、pre-IPO 和流动性都太有吸引力。他也补充，如果选择自己创业，就需要强 conviction 和巨大 ambition。这里的信号是：前沿实验室仍在持续吸走顶级人才。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-22fc3b2ca3bc414d87f63a5e491250d7"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072522778327371819" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-20f84cce0b014ea582a1c3a7ef7cad52"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072406317617262753" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-fcae179a466a46c3a3be1df68a98653a"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072344802570756121" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-aff8b9aec52148e0a02b44d60ab82ba9" data-id="aff8b9aec52148e0a02b44d60ab82ba9"><span><div id="aff8b9aec52148e0a02b44d60ab82ba9" class="notion-header-anchor"></div><a class="notion-hash-link" href="#aff8b9aec52148e0a02b44d60ab82ba9" title="14. Peter Steinberger"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">14. Peter Steinberger</span></span></h4><div class="notion-text notion-block-056a243f56fb470aa6e9ab351bb60304">Peter Steinberger 今天三条比较生活化但都和 builder 氛围有关。他提到 Steve Yegge 很早就看到了方向，现在大家都在 build factories；又说自己在旧金山找半私密 hack space，准备和 OpenClaw maintainers 一起高强度开发；最后感叹没有 AI 以前自己是怎么工作的。这几条合在一起看，重点是 AI 工具正在改变个人和团队的工作强度，也让“工厂式构建”变成常态。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-11dae8a76f6a4e6ba17f53d36333c82c"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072532278476148881" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-67cdb9a4b58d4c77b7110f7d05f0198f"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072475858435276840" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-f75c58e8448042bbb5bf4032ac806e35"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072447453622882338" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-97c150eb8bde43c0969551095ac3fa2d" data-id="97c150eb8bde43c0969551095ac3fa2d"><span><div id="97c150eb8bde43c0969551095ac3fa2d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#97c150eb8bde43c0969551095ac3fa2d" title="15. Dan Shipper"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">15. Dan Shipper</span></span></h4><div class="notion-text notion-block-e5221f6b530047df9318be9115736863">Dan Shipper 今天继续围绕 Fable。他先说某件事太酷，自己一直想做；然后在 jacuzzi 里兴奋地说要 rip Fable tokens；最后直接喊 Fable is back。信息密度不高，但情绪非常强，说明 Fable 5 回归已经成为 AI builder 圈的即时事件。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-89e71298eca34504a5c615b0b772450f"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072436587665797518" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-5d8f2dd220ec4834802cdd84039205d3"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072402843819212906" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-0a7d61ae478b45a19e0662147bd19607"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072402230041272669" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-74cd9c94f43644deb928e4ccd07526c3" data-id="74cd9c94f43644deb928e4ccd07526c3"><span><div id="74cd9c94f43644deb928e4ccd07526c3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#74cd9c94f43644deb928e4ccd07526c3" title="16. Aditya Agarwal"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">16. Aditya Agarwal</span></span></h4><div class="notion-text notion-block-11d9ddb0aae34596b3b63c05f72cc086">Aditya Agarwal 说，旧金山偶尔会遇到非常聪明但默认悲观的人，很多来自增长或华尔街背景；他的第一反应是，为什么这样的人要来一个靠乐观驱动的城市。这条不是 AI 产品动态，但它解释了湾区 builder 文化的一部分：强乐观和行动偏好仍然是这里的底层气质。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-848e69e78eb14bb8a0d962e2abc10a88"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072449611550380526" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-7e224e77935e4260ab19265c06c30403" data-id="7e224e77935e4260ab19265c06c30403"><span><div id="7e224e77935e4260ab19265c06c30403" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7e224e77935e4260ab19265c06c30403" title="17. Claude"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">17. Claude</span></span></h4><div class="notion-text notion-block-ec1f88bd6f50467686b635a2d5a2f073">Claude 官方账号给出了 Fable 5 访问规则：所有包含 usage 的付费计划都可以在 7 月 7 日前使用 Fable 5；最多可占 weekly usage limit 的 50%，之后可以切换到其他模型，也可以继续用 usage credits。它还说，如果 Claude Code 里的请求被误判，可以用 /feedback 上报；Claude 网页和 Cowork 则可以用 thumbs 按钮反馈，用来继续调分类器、减少 false positives。这说明 Fable 5 回归不是无限开放，而是带着限制、分类器和反馈回路一起回来。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-70f1a871ccbc4e75837211f513e5a943"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072402642836615273" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-4b65f589850e44e18ecd2ebb6ace4485"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072402640907162072" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-e146e75002b34909ad2a3dc6c33eb499"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072402639644766602" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-850620ebd93c485ca3086f5764fc6bfd" data-id="850620ebd93c485ca3086f5764fc6bfd"><span><div id="850620ebd93c485ca3086f5764fc6bfd" class="notion-header-anchor"></div><a class="notion-hash-link" href="#850620ebd93c485ca3086f5764fc6bfd" title="Blog 全量记录"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Blog 全量记录</span></span></h3><div class="notion-text notion-block-830eecff52474cb09e247580f9c3d46c">今天 feed 里没有新的 blog 条目。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-7e742bd3422e49d897a381eaa74091c3" data-id="7e742bd3422e49d897a381eaa74091c3"><span><div id="7e742bd3422e49d897a381eaa74091c3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7e742bd3422e49d897a381eaa74091c3" title="Podcast 全量记录"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Podcast 全量记录</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-b22e011d6b4a4cdb962c23036fde3d29" data-id="b22e011d6b4a4cdb962c23036fde3d29"><span><div id="b22e011d6b4a4cdb962c23036fde3d29" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b22e011d6b4a4cdb962c23036fde3d29" title="1. AI &amp; I by Every：The AI Workflows Behind Every&#x27;s Consulting Team"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">1. AI &amp; I by Every：The AI Workflows Behind Every&#x27;s Consulting Team</span></span></h4><div class="notion-text notion-block-e12388fe94c34be4abb67a50415e4a81">AI &amp; I 这期聊 Every 咨询团队内部如何使用 AI 工作流。核心不是“用了哪个工具”，而是把 AI 变成一种操作系统：Codex 帮人搭流程，邮件和上下文能主动支持决策，行政任务会逐渐被系统接管。节目里有个比喻很好：知识工作越来越像 gardening，不是亲手制造植物，而是创造条件，让增长自己发生。这个判断对企业 AI 落地很有价值，因为很多团队真正缺的不是模型，而是能让模型持续进入工作现场的 workflow。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-62b812433f7b4d07b9901924a4a7f435"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://www.youtube.com/embed/videoseries?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-8d820d2cfb6a46888fa2aa03b9f4a46f" data-id="8d820d2cfb6a46888fa2aa03b9f4a46f"><span><div id="8d820d2cfb6a46888fa2aa03b9f4a46f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#8d820d2cfb6a46888fa2aa03b9f4a46f" title="今日沉淀结论"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">今日沉淀结论</span></span></h3><div class="notion-text notion-block-41d2db0c98e7489db40d46ab9546df38">今天所有内容放在一起看，Fable 5 回归是表层热点，深层变化是 AI builder 生态正在从“单模型能力”进入“多 agent、可验证、可部署、可变现、可沉淀”的阶段。谁能把模型接进真实 workflow，并把成本、风险和反馈机制管住，谁就更接近下一阶段的 AI 产品形态。</div><div class="notion-text notion-block-fd483be0d1134fc197cf4f4b9a3c4f80">Generated through the Follow Builders skill: <a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/zarazhangrui/follow-builders">https://github.com/zarazhangrui/follow-builders</a></div></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sonnet 5、Fable 回归与 Agentic Web - 2026-07-01]]></title>
            <link>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-01</link>
            <guid>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-01</guid>
            <pubDate>Wed, 01 Jul 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[今日全量沉淀包括 17 位 X builders 的 38 条推文、0 篇博客和 1 期 Training Data 播客。主线是 Sonnet 5 推向默认模型、Fable 访问恢复引发关注，Vercel 和 Replit 继续把 AI 应用往真实工程与硬件协同推进。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-3905eac3813c8102a3dff59fe3cb275d"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-301236c03b144722b00c5c8710fd41ca" data-id="301236c03b144722b00c5c8710fd41ca"><span><div id="301236c03b144722b00c5c8710fd41ca" class="notion-header-anchor"></div><a class="notion-hash-link" href="#301236c03b144722b00c5c8710fd41ca" title="今日主线"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">今日主线</span></span></h3><div class="notion-text notion-block-cc10b1eee4ca40d18133aff6fd8f2046">今天这批 follow-builders 内容的主线很清楚：Claude Sonnet 5 正在从“新模型”变成默认体验，Fable 访问恢复成为 builder 圈的共同关注点；同时，Vercel、Replit、Box、Linear 等人的讨论都在指向一个方向，AI 产品不只是更聪明的聊天框，而是要进入 agentic web、真实企业工作流、专用推理硬件和更深的基础设施。全量数据包括 17 位 X builders 的 38 条推文、0 篇博客和 1 期 Training Data 播客。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-866dba33a9944180989a2fad0fc7de17" data-id="866dba33a9944180989a2fad0fc7de17"><span><div id="866dba33a9944180989a2fad0fc7de17" class="notion-header-anchor"></div><a class="notion-hash-link" href="#866dba33a9944180989a2fad0fc7de17" title="重点解读"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">重点解读</span></span></h3><div class="notion-text notion-block-c8893398faeb4472b766075b55e9572d">Claude 官方宣布 Sonnet 5 已成为 Free 和 Pro 的默认模型，也向 Max、Team、Enterprise 用户开放，并进入 Claude apps 和 Claude Platform。它的定位很明确：接近 Opus 4.8 的表现，但价格更低，并且在 reasoning、tool use、coding、knowledge work 上比 Sonnet 4.6 有明显提升。Aaron Levie 也分享了 Box AI Complex Work Eval 的结果，说 Sonnet 5 在复杂企业文档任务上保持 frontier-class 质量，并在能源、零售、专业服务等领域超过 Sonnet 4.6。这说明 Sonnet 5 的重点不是单点 benchmark，而是更长链路的企业任务。</div><div class="notion-text notion-block-222c2838664e4110bf53abb12fe7af8a">Fable 也是今天的高频词。Peter Yang、Nikunj Kothari、Dan Shipper 都在等 Fable 回归，Aaron Levie 则把它放到更大的模型发布框架里看：如果前沿模型拥有强 coding 和 cyber 能力，之后行业可能需要一套共同标准来判断 jailbreak 的严重程度，并决定如何修复和放行。这个点值得关注，因为模型发布节奏正在从“直接上线”变成“能力、风险、访问权限一起管理”。</div><div class="notion-text notion-block-37da9ab3eed547c989ec0e341cb654f3">硬件和基础设施线也很重。Amjad Masad 转发 Etched，强调 AI 推理贵的一部分原因是今天很多 workload 还跑在 pre-LLM 时代的通用硬件上；Training Data 这期和 SemiAnalysis 的 Dylan Patel 聊 hardware-software co-design，也在讲同一件事：真正的 100x 不只来自模型，而来自模型、软件、芯片、网络、内存、封装和数据中心一起设计。Vercel Services 则把部署层往多服务、多运行时、可观测和 internal networking 推进，和“agentic web”这条线接上了。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-2b829a1f411446edaf27e111e91c0f14" data-id="2b829a1f411446edaf27e111e91c0f14"><span><div id="2b829a1f411446edaf27e111e91c0f14" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2b829a1f411446edaf27e111e91c0f14" title="X Builders 全量记录"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">X Builders 全量记录</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-b7c92f0ce987454bbeecd94f1abcba8e" data-id="b7c92f0ce987454bbeecd94f1abcba8e"><span><div id="b7c92f0ce987454bbeecd94f1abcba8e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b7c92f0ce987454bbeecd94f1abcba8e" title="1. Swyx"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">1. Swyx</span></span></h4><div class="notion-text notion-block-f865b4e6ce79431daa22ca941f69ea4a">Swyx 今天继续在 AIEWF 现场。他提到第一天 poster sessions 很成功，还感谢 Heather Miller 和 ACM President 的支持；明天会做 Poaster sessions，让大家把热门推文提交给 Vibhu 打印出来。这条更偏社区现场，但能看出 AI Engineer 这类 builder 活动已经不只是演讲，而是在把社区表达、demo 和现场交流变成内容本身。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-3f82e72bc01244e08dd1dd27009a3c71"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072048730581405800" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-3e3bb6a1c60d463ea0299c019a7bfea7" data-id="3e3bb6a1c60d463ea0299c019a7bfea7"><span><div id="3e3bb6a1c60d463ea0299c019a7bfea7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3e3bb6a1c60d463ea0299c019a7bfea7" title="2. Boris Cherny"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">2. Boris Cherny</span></span></h4><div class="notion-text notion-block-bedbf6cc47aa456eabcf5dac4c44d8fd">Boris Cherny 有两条。一条只是表示同意，信息量有限；另一条更实在：Claude Desktop on Linux 终于来了，并附了下载链接。对 Claude Code 和 Linux 开发者来说，这是一个补齐平台覆盖的动作，意味着 Claude 不再主要围绕 macOS 和网页体验。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-154b3f1731d149be91ec2bdc88193462"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072105851457019944" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-a211eeb6f81d4a4caf17623a93bbc7b2"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072000214634742243" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-75654343c94c4538954c786f1d03386a" data-id="75654343c94c4538954c786f1d03386a"><span><div id="75654343c94c4538954c786f1d03386a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#75654343c94c4538954c786f1d03386a" title="3. Peter Yang"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">3. Peter Yang</span></span></h4><div class="notion-text notion-block-a494dfb814324a5899338ad7ed15c0c1">Peter Yang 今天主要围绕 Fable。他在问达到 50% weekly usage 后 Fable 是否会不可用，也说自己不知道它是不是午夜发布，还开玩笑说要半夜定闹钟让 Fable 修所有项目。内容不算正式公告，但很能反映 builder 圈对高能力 coding model 恢复访问的期待。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-1bab0468762a4933b4e5608369393e62"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072165346476511583" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-031627dfe2cd4d8e9f3f33053cea53fd"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072159132908695772" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-00ff0eb76c324841bd77a5e429cdab03"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072111311396438410" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-e75284cd251e47e6a9a6f76ef48d8545" data-id="e75284cd251e47e6a9a6f76ef48d8545"><span><div id="e75284cd251e47e6a9a6f76ef48d8545" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e75284cd251e47e6a9a6f76ef48d8545" title="4. Nan Yu"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">4. Nan Yu</span></span></h4><div class="notion-text notion-block-f7a472be574845caa2f462148bd7f2d4">Linear 的 Nan Yu 讨论了 distillation 的定义。他顺着别人的逻辑指出，如果这么定义，Cursor 早期训练数据几乎都可以说是从 Claude distill 来的。这条不是产品发布，但抓住了一个长期问题：AI 工具之间的数据来源、模型输出、训练边界和“借鉴”定义会越来越难分清。另一条只是轻量认同，也保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-bd5e4a4f6c5d4fb892a868b9ddafc97a"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072003481779478848" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-3707925dfecf4affacd40749a0cfdc18"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071973229070033322" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-25151f05603842f6a2d5a228b7f220a0" data-id="25151f05603842f6a2d5a228b7f220a0"><span><div id="25151f05603842f6a2d5a228b7f220a0" class="notion-header-anchor"></div><a class="notion-hash-link" href="#25151f05603842f6a2d5a228b7f220a0" title="5. Madhu Guru"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">5. Madhu Guru</span></span></h4><div class="notion-text notion-block-91ec52439d7b4b01bffd2f1458381526">Madhu Guru 说传统 PM 适应 AI native building 的最大障碍，是缺少 magical thinking。过去十年的框架、agile 和指标迷信，让很多团队先想约束，再想增量改进；他以前会让团队想象自己拥有 100 年后的技术，然后倒推体验。现在的问题是，这种“未来技术”已经来了，应该直接去做更有想象力的东西。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-d02e5e02c8604e16b340132f62ce6a29"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071970221477470694" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-d6867d3e19974a31a052bd5bd918c724" data-id="d6867d3e19974a31a052bd5bd918c724"><span><div id="d6867d3e19974a31a052bd5bd918c724" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d6867d3e19974a31a052bd5bd918c724" title="6. Thariq"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">6. Thariq</span></span></h4><div class="notion-text notion-block-75a093b868af4a8692d53d34b076383d">Thariq 今天一边准备 AIE 演讲 deck，一边解释 updated classifiers。他说新的 classifier 仍会让一小部分常规 coding/debugging 任务被 flag，然后 fallback 到 Opus；团队还在继续改进 safeguards，目标是更好地区分真实滥用和正常请求，减少 false positives。这和 Fable 访问恢复放在一起看很重要：模型能力越强，产品越要在误杀和防滥用之间找平衡。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-60fba8a5df7c4b1da0fe0873f15ded5a"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072202267563082180" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-4ca55f34c08d4ebb94252c68f5136a9b"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072185566695977161" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-31681fe380f040d6960521b48938cc57"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072185565076988326" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-ad72d649e8094e5b83cf6ac9d3008a92" data-id="ad72d649e8094e5b83cf6ac9d3008a92"><span><div id="ad72d649e8094e5b83cf6ac9d3008a92" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ad72d649e8094e5b83cf6ac9d3008a92" title="7. Amjad Masad"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">7. Amjad Masad</span></span></h4><div class="notion-text notion-block-021e46f623a449df9ed47994c6d5c648">Replit CEO Amjad Masad 今天一条是世界首个 subway hackathon，偏活动；另一条更关键，他转发 Etched 并说 AI 运行昂贵的一部分原因，是大多数 workload 还跑在 LLM 出现前设计的通用硬件上。Etched 的定位是从头为现代 inference 设计的系统。这个判断和今天播客里的 hardware-software co-design 主线高度一致。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-41afbeeec09d437aacb89be32070f2db"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071992670394765407" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-6a3f686bd3704dc3b2cbc10b2bea56ea"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071992110132117740" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-d147f732d3c3428cb3fd32db665a6159" data-id="d147f732d3c3428cb3fd32db665a6159"><span><div id="d147f732d3c3428cb3fd32db665a6159" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d147f732d3c3428cb3fd32db665a6159" title="8. Guillermo Rauch"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">8. Guillermo Rauch</span></span></h4><div class="notion-text notion-block-872dc98674fb4b6b8661969cf86cc255">Guillermo Rauch 今天的主线是 Vercel Services 和 agentic web。他说 Vercel 现在可以把 Python backend API、ExpressJS server、React SPA 放在同一个项目里，本地一起跑、一起部署回滚、一起观察监控和调试，并支持 internal networking。他也表达了和 Shopify 团队一起推进 agentic web 的兴奋。另一条是晚餐时同时收到公司高管和对方 12 岁儿子对 Vercel 的反馈，说明他在强调“Vercel is for everyone”。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-38455e808e174ae79bba622e8c71281a"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072123264843886709" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-6aeced7f45e84ff194b48e67f71d4b6d"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072044844965400589" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-28e2e985d8984df7a0f593f9e8a37841"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071966055308607765" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-e83c86d95a3749adb54516085509a4ea" data-id="e83c86d95a3749adb54516085509a4ea"><span><div id="e83c86d95a3749adb54516085509a4ea" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e83c86d95a3749adb54516085509a4ea" title="9. Aaron Levie"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">9. Aaron Levie</span></span></h4><div class="notion-text notion-block-10de92bc631143b68a3baf59b43fe7d1">Aaron Levie 今天三条都很有信息量。第一条是对 Fable 和未来 GPT-5.6 这类模型发布节奏的判断：强 coding 和 cyber 能力模型可能需要共享行业框架来评估 jailbreak 严重程度和修复流程。第二条是 Box AI Complex Work Eval，Sonnet 5 在复杂企业文档任务上比 Sonnet 4.6 更强，尤其在能源、零售、专业服务等领域领先。第三条是 AI adoption 和就业，他引用 Ramp 与 Box 调研，认为更成熟采用 AI 的公司反而更可能增加 headcount，但这不等于“用了 AI 就会自动扩张”，也可能是增长公司更有能力采用 AI。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-087a10a1bbd24d09a7e87d505bc85e9b"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072172275017879829" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-05cf462a38eb45aab22fa66623a438b8"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072046374045249671" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-1088c2b021ef4c6580aa71950a74fe4d"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071992799109824562" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-60870985f96c4bdb93d0ab86ce540dee" data-id="60870985f96c4bdb93d0ab86ce540dee"><span><div id="60870985f96c4bdb93d0ab86ce540dee" class="notion-header-anchor"></div><a class="notion-hash-link" href="#60870985f96c4bdb93d0ab86ce540dee" title="10. Garry Tan"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">10. Garry Tan</span></span></h4><div class="notion-text notion-block-55101770be7d4c58a21edac31def03b1">Garry Tan 一条是加州治理相关，偏公共议题；另一条提到 Gbrain 在个人或公司 brain 达到 10000+ markdown 文件时最有用。这条对知识库和 AI memory 有参考意义：很多个人知识工具在小规模时看不出差异，真正的价值可能在大规模、长周期、可检索的个人或组织记忆里出现。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-4f0acba0da5b43c4837a6a62f141b90e"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072100337478160856" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-b001f83457614ba18cd79eef0ef4d3ff"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071910876496757145" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-b94cc0d72e0d498c8d0910d5aa53127b" data-id="b94cc0d72e0d498c8d0910d5aa53127b"><span><div id="b94cc0d72e0d498c8d0910d5aa53127b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b94cc0d72e0d498c8d0910d5aa53127b" title="11. Matt Turck"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">11. Matt Turck</span></span></h4><div class="notion-text notion-block-103cf8d319bb4570972af9c262a31f7e">Matt Turck 今天发的是世界杯相关内容，写了“LE DICTATEUR #fraswe”。这条和 AI 主线没有直接关系，但按全量沉淀要求保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-79da3b70e7cf4bbcaf55788d92a5a3dd"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072088791393669144" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-19957a86580645f18fe80943337d6334" data-id="19957a86580645f18fe80943337d6334"><span><div id="19957a86580645f18fe80943337d6334" class="notion-header-anchor"></div><a class="notion-hash-link" href="#19957a86580645f18fe80943337d6334" title="12. Zara Zhang"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">12. Zara Zhang</span></span></h4><div class="notion-text notion-block-313cb70226294d2faa5a8ddac309d530">Zara Zhang 今天几条更偏 taste 和社交观察。她转发“taste 的价值不在于不能被复制，而在于它定义了别人选择复制什么”；还感叹有人非常早，以及发了一条 link-only 内容。这里最值得记的是 taste 这条，因为 AI 工具越强，判断什么值得做、什么值得模仿，反而更重要。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-9d438573f9d24ff69a3a435d085d7a71"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072197929138602079" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-229621b18d444c5dab3ddf08786645fe"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072145334428680365" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-6c206c4a4f954499977c556eb49cdc64"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072120868042641686" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-ef1b57802128437ea8ffa8ceeda9d344" data-id="ef1b57802128437ea8ffa8ceeda9d344"><span><div id="ef1b57802128437ea8ffa8ceeda9d344" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ef1b57802128437ea8ffa8ceeda9d344" title="13. Nikunj Kothari"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">13. Nikunj Kothari</span></span></h4><div class="notion-text notion-block-8f133c75000a45279dfb0627aa23e53b">Nikunj Kothari 今天两条都围绕 Fable 5。他说“早了一天，明天会来”，又调侃旧金山开发者听到 Fable 5 要回来后都跑回桌前。这不是正式产品说明，但能反映 Fable 在开发者圈里的期待值。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-da476f5e6dc44ea7a132a006158f79c6"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072111272859246943" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-c29f685a0c8a4ed9a5fd402f8b56351b"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072106393256218883" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-9d79d1db8c6144fea65065bc6f2df6df" data-id="9d79d1db8c6144fea65065bc6f2df6df"><span><div id="9d79d1db8c6144fea65065bc6f2df6df" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9d79d1db8c6144fea65065bc6f2df6df" title="14. Peter Steinberger"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">14. Peter Steinberger</span></span></h4><div class="notion-text notion-block-0c491e0c603641e08266bd28fdcba60d">Peter Steinberger 今天几条短内容里，最值得看的是“price per token != cost per task”。这句话抓住了 AI coding 和 agent 计费里的关键：便宜 token 不一定等于便宜任务，如果模型更容易跑偏、需要更多轮、或者不能完成任务，总成本可能更高。他还提到 loops，说明 agent 工作流讨论还在继续发酵。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-1267c4c6523447dab3c2b9f71f42b373"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072183656739983397" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-a69fe5e4b58142c18f7b4da60f568b82"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072144627474579925" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-da6cd4106c9e49fb82c4ac02a074aa8a"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072143124496097302" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-319e5d1f8c0045e98c9ae50d211d0639" data-id="319e5d1f8c0045e98c9ae50d211d0639"><span><div id="319e5d1f8c0045e98c9ae50d211d0639" class="notion-header-anchor"></div><a class="notion-hash-link" href="#319e5d1f8c0045e98c9ae50d211d0639" title="15. Dan Shipper"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">15. Dan Shipper</span></span></h4><div class="notion-text notion-block-97e4bd9b5d4c4803873bc3c5dec050d6">Dan Shipper 今天也是围绕 Fable 可能回归。他在度假中表示如果 Fable 今晚回来，自己会直播并“ripping tokens”。语气很兴奋，信息本身不多，但它和 Peter Yang、Nikunj 的内容一起说明，builder 圈对 Fable 的访问恢复非常敏感。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-18375857bfe94b1f8038ed3c477469c3"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072106715286225313" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-e6dd5d18ea1143b6bf8292bef7603222"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072099953254584462" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-2e1b07f3f8b8489fae1358a74c54debf"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072094163378622823" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-3677ca3eba40453f96b116477b864242" data-id="3677ca3eba40453f96b116477b864242"><span><div id="3677ca3eba40453f96b116477b864242" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3677ca3eba40453f96b116477b864242" title="16. Aditya Agarwal"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">16. Aditya Agarwal</span></span></h4><div class="notion-text notion-block-baccadf5c3a74c239e98e46050859aba">Aditya Agarwal 的判断很尖锐：现在很奇怪，美国创新所依赖的一些模型，竟然是中国开源模型。这条和 open-weight、GLM、模型开放生态的讨论接在一起看很有意思：如果开源模型持续接近前沿能力，地缘和产业竞争会变得更复杂。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-5cec2118ea27449c8d9d414d010327ca"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071983952894837062" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-3901606956c14d9c91c72d7ba367b2e3" data-id="3901606956c14d9c91c72d7ba367b2e3"><span><div id="3901606956c14d9c91c72d7ba367b2e3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3901606956c14d9c91c72d7ba367b2e3" title="17. Claude"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">17. Claude</span></span></h4><div class="notion-text notion-block-325c733462fc4d76bbf7f96a14cafe14">Claude 官方账号宣布 Sonnet 5 成为 Free 和 Pro 默认模型，也向 Max、Team、Enterprise 用户开放，并已进入 Claude apps 和 Claude Platform。它强调 Sonnet 5 在 reasoning、tool use、coding、knowledge work 上比 Sonnet 4.6 有明显提升，表现接近 Opus 4.8，但价格更低。早期伙伴反馈是，Sonnet 5 能完成以前 Sonnet 容易半途停下来的复杂任务，也会主动检查自己的输出。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-81049378521045af95d7474b56858ca5"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072017457057853480" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-73c3fa638f6a4039a01c0526551ffa82"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072017455833100494" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-a298050d70e64578a8d83033c179d863"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2072017452335087996" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-369885212a9c4203a99cfa28a1520fb4" data-id="369885212a9c4203a99cfa28a1520fb4"><span><div id="369885212a9c4203a99cfa28a1520fb4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#369885212a9c4203a99cfa28a1520fb4" title="Blog 全量记录"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Blog 全量记录</span></span></h3><div class="notion-text notion-block-d9d4f0880aa14fbd94c77c0c45b93007">今天 feed 里没有新的 blog 条目。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-d17dba04efdb4b6494f9b485ae9362bb" data-id="d17dba04efdb4b6494f9b485ae9362bb"><span><div id="d17dba04efdb4b6494f9b485ae9362bb" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d17dba04efdb4b6494f9b485ae9362bb" title="Podcast 全量记录"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Podcast 全量记录</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-885dcc5747694125a6bbb17ed5d3b1e6" data-id="885dcc5747694125a6bbb17ed5d3b1e6"><span><div id="885dcc5747694125a6bbb17ed5d3b1e6" class="notion-header-anchor"></div><a class="notion-hash-link" href="#885dcc5747694125a6bbb17ed5d3b1e6" title="1. Training Data：Why Hardware-Software Co-Design Is AI&#x27;s Real 100x: Dylan Patel of SemiAnalysis"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">1. Training Data：Why Hardware-Software Co-Design Is AI&#x27;s Real 100x: Dylan Patel of SemiAnalysis</span></span></h4><div class="notion-text notion-block-fd05c09a01e94900ba025740e7a93ae5">Training Data 这期和 SemiAnalysis 的 Dylan Patel 聊 AI 硬件和软件如何一起设计。最核心的判断是，AI 的真正 100x 不只是模型算法，而是从模型结构、kernel、compiler、networking、memory、芯片、封装到数据中心的整体协同。Dylan 提到 SemiAnalysis 内部同时有工程师和前 hedge fund 人，所以技术可行性和成本会不断互相拉扯。这个视角很实用：AI 基础设施的胜负不是单点参数，而是谁能把性能、成本、供给、软件栈和部署现实一起优化。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-be8827826fad4b5ca71c250c392cce58"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://www.youtube.com/embed/f6D_aiy8qyU" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-34662007a2cc4ea3be89663ebcc950e8" data-id="34662007a2cc4ea3be89663ebcc950e8"><span><div id="34662007a2cc4ea3be89663ebcc950e8" class="notion-header-anchor"></div><a class="notion-hash-link" href="#34662007a2cc4ea3be89663ebcc950e8" title="今日沉淀结论"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">今日沉淀结论</span></span></h3><div class="notion-text notion-block-befa8d2ef70843f28c452a0af359f482">今天的内容放在一起看，AI 正在进入两个同时发生的阶段：上层是 Sonnet 5、Fable、Vercel Services、Claude Desktop Linux 这类产品把模型能力带进更广的真实工作流；底层是 Etched、SemiAnalysis、硬件软件协同、专用推理系统在回答一个更硬的问题：这些能力怎样以可接受成本稳定跑起来。</div><div class="notion-text notion-block-e3eaca3383f943aa84bdc845462d403d">Generated through the Follow Builders skill: <a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/zarazhangrui/follow-builders">https://github.com/zarazhangrui/follow-builders</a></div></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Code as Agent Harness 深度分析 - 2026-07-02]]></title>
            <link>http://easyai.fyi/article/ai-agent-code-as-agent-harness-2026-07-02</link>
            <guid>http://easyai.fyi/article/ai-agent-code-as-agent-harness-2026-07-02</guid>
            <pubDate>Thu, 02 Jul 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[这篇报告分析 Code as Agent Harness：为什么代码正在从 agent 的输出物，变成承载状态、执行反馈、工具治理和多 agent 协作的运行层。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-3905eac3813c816a8e9ed245c3063c2b"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-8748edef08154efbbe89357f789768d9" data-id="8748edef08154efbbe89357f789768d9"><span><div id="8748edef08154efbbe89357f789768d9" class="notion-header-anchor"></div><a class="notion-hash-link" href="#8748edef08154efbbe89357f789768d9" title="原始链接"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">原始链接</span></span></h3><ul class="notion-list notion-list-disc notion-block-9e103df81d0945a6ae6c46e91e177748"><li>论文页面：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/abs/2605.18747">Code as Agent Harness</a></li></ul><ul class="notion-list notion-list-disc notion-block-11555e9398fc4f75955a4be667453742"><li>论文 HTML：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/html/2605.18747v1">arXiv HTML</a></li></ul><ul class="notion-list notion-list-disc notion-block-c96e5880a76d412899a54b7688ae45c5"><li>论文 PDF：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/pdf/2605.18747">arXiv PDF</a></li></ul><ul class="notion-list notion-list-disc notion-block-3d2dc3f7d8af479085dcc2ba92773bb0"><li>项目页：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://code-as-harness.github.io/code-as-harness-webpage/">https://code-as-harness.github.io/code-as-harness-webpage/</a></li></ul><ul class="notion-list notion-list-disc notion-block-0bcfeb2fe5cd4b9e860c27eb09511f34"><li>配套资料库：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/YennNing/Awesome-Code-as-Agent-Harness-Papers">Awesome Code as Agent Harness Papers</a></li></ul><ul class="notion-list notion-list-disc notion-block-a7e7e37071884897acc8da7edd7b983a"><li>Hugging Face 论文页：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://huggingface.co/papers/2605.18747">Code as Agent Harness</a></li></ul><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-512c3dc709824a0a85dcf73398046131" data-id="512c3dc709824a0a85dcf73398046131"><span><div id="512c3dc709824a0a85dcf73398046131" class="notion-header-anchor"></div><a class="notion-hash-link" href="#512c3dc709824a0a85dcf73398046131" title="为什么今天选它"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">为什么今天选它</span></span></h3><div class="notion-text notion-block-01044e37cb534dbab03080a6d28c32ed">今天我比较了几个候选：<code class="notion-inline-code">earendil-works/pi</code>、<code class="notion-inline-code">Natural-Language Agent Harnesses</code>、<code class="notion-inline-code">Agentic Harness Engineering</code> 和 <code class="notion-inline-code">Code as Agent Harness</code>。</div><div class="notion-text notion-block-5447237f65584c4eb2b12681f2441fac">Pi 更像一个正在变热的开源 coding agent 产品，GitHub 页面显示约 66.9k stars，并把自己定义成可自定义的 agent harness；它很适合作为工程案例。<code class="notion-inline-code">Natural-Language Agent Harnesses</code> 把 harness 写成可执行的自然语言文档，想法新，但范围更窄。<code class="notion-inline-code">Agentic Harness Engineering</code> 更偏实验系统，重点是让 agent 根据轨迹自动改自己的 harness，实验结果具体，但讨论对象主要是 coding agent。</div><div class="notion-text notion-block-d0ce1a35b1fd4f328ceec6148d5ea374">我最后选 <code class="notion-inline-code">Code as Agent Harness</code>，主要因为它把这几条线放到了一张图里。论文 2026 年 5 月 18 日提交，Hugging Face 标为 <code class="notion-inline-code">#1 Paper of the day</code>，有 223 个 upvote；配套 GitHub 资料库有 531 左右 star，并且整理了三层框架和大量相关论文。它不是在介绍一个新工具，而是在追问一个更底层的问题：为什么现在的 agent 不再只是“模型加工具”，而是越来越像一个围绕代码、状态、执行反馈和协作流程搭起来的运行系统。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-cb6275a5225f427f9adb2b34ad8dfd2e" data-id="cb6275a5225f427f9adb2b34ad8dfd2e"><span><div id="cb6275a5225f427f9adb2b34ad8dfd2e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#cb6275a5225f427f9adb2b34ad8dfd2e" title="它要解决的问题"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">它要解决的问题</span></span></h3><div class="notion-text notion-block-645a1850d1ea47bb885bcd8767af0ce0">过去讲 AI Agent，经常会拆成模型、工具调用、记忆、计划、反思、评测几个模块。这个拆法有用，但它漏掉了运行时最麻烦的一层：这些模块真正跑起来时，总要落到某种可执行、可检查、可回滚的东西上。对 coding agent、GUI agent、科学发现 agent、机器人 agent 来说，这个东西经常就是代码。</div><div class="notion-text notion-block-623a21d1f1164877a67c7edec6874131">论文的核心判断是：代码不再只是 agent 最后生成的结果。代码正在变成 agent 的操作介质。agent 用代码表达中间推理，用代码调用工具，用代码建模环境，用测试和执行结果验证自己，用仓库、日志和工作流保存状态，多 agent 还会围绕同一份代码做分工、审查和修复。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-3ac6ea2a5e094583a0342007759ccdb3"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://arxiv.org/html/2605.18747v1/x1.png?t=3ac6ea2a-5e09-4583-a034-2007759ccdb3" alt="图 1：论文的总览图，把 code as agent harness 拆成接口、机制和扩展三层。来源：arXiv HTML Figure 1" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图 1：论文的总览图，把 code as agent harness 拆成接口、机制和扩展三层。来源：arXiv HTML Figure 1</figcaption></div></figure><div class="notion-text notion-block-b82ca274ff9e4fc597af2962e5548fbe">这张图最值得看的不是分类有多全，而是它把 agent 的“想”和“做”放到同一个闭环里：左边是 agent 生成代码，右边是外部环境执行并返回验证信号，中间的 harness infrastructure 负责把推理、动作、环境状态、记忆、调试和多人协作串起来。harness 不是提示词外壳，它更像 agent 能不能长期工作的运行层。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-e3b25925c21c42358a32c89d49a6f855" data-id="e3b25925c21c42358a32c89d49a6f855"><span><div id="e3b25925c21c42358a32c89d49a6f855" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e3b25925c21c42358a32c89d49a6f855" title="核心思想：代码是可执行的上下文"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">核心思想：代码是可执行的上下文</span></span></h3><div class="notion-text notion-block-da4edde2e35d4b4b9b0ee5cd1596b31a">论文把 code as agent harness 分成三层。</div><div class="notion-text notion-block-84e8b32c0f684db392a23557c882ffd1">第一层是 harness interface。代码把模型输出变成可执行、可检查的结构。数学推理里，代码可以把中间计算交给解释器或求解器；GUI/OS agent 里，代码可以变成 DOM 操作、系统命令或 API 调用；在软件工程里，代码库、测试、日志和执行轨迹本身就是环境状态。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-a74ffc4b98544230b068a546c0b7e6e9"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://raw.githubusercontent.com/YennNing/Awesome-Code-as-Agent-Harness-Papers/main/figs/harness_interface.png?t=a74ffc4b-9854-4230-b068-a546c0b7e6e9" alt="图 2：Harness Interface，把代码分成 reasoning、acting、environment modeling 三个入口。来源：配套 GitHub README" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图 2：Harness Interface，把代码分成 reasoning、acting、environment modeling 三个入口。来源：配套 GitHub README</figcaption></div></figure><div class="notion-text notion-block-324f51fe5d6f4646bbdd68df3931b467">这张图说明了一个容易被忽略的点：工具调用并不是“模型想做什么就调哪个函数”。可靠的 agent 要先把意图翻译成动作，再通过环境反馈确认动作有没有生效。代码在这里的价值，是把意图变成可执行对象，也把反馈变成可检查证据。</div><div class="notion-text notion-block-afefce74ed1b4ed9a1ea6eab8dddd9e5">第二层是 harness mechanisms。这里包括 planning、memory、tool use、feedback control 和 harness optimization。论文对 tool use 的表述很关键：工具不是附加能力，而是模型意图和外部系统之间的治理接口。一个可靠的 harness 要决定哪些工具可用、schema 怎么暴露、权限怎么给、执行在哪里发生、结果怎么清洗或压缩、哪些危险动作要人类确认。</div><div class="notion-text notion-block-f39e68e1377340aba2d4942df8956559">memory 也不是简单的“向量库记忆”。在 code-as-harness 视角下，memory 是状态管理：哪些信息留在当前上下文，哪些被压缩，哪些写入外部存储，哪些变成可复用技能、测试、脚本或仓库状态。对长任务来说，这比“记得更多”更重要。agent 真正需要的是在正确时刻拿到正确状态。</div><div class="notion-text notion-block-a4cf4d4d0dd247d98fef74bb9824606a">第三层是 scaling the harness。单个 agent 可以靠对话推进任务，多 agent 就必须有共享状态，否则很快变成角色扮演。论文指出，代码的优势在于它可以执行，可以产生非语言的客观信号。多 agent 系统里的收敛，不应该只靠“大家同意了”，更应该靠测试通过、行为一致、审查完成、状态冲突解决。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-eb99b145357549219c0c7e4be36ed1bf" data-id="eb99b145357549219c0c7e4be36ed1bf"><span><div id="eb99b145357549219c0c7e4be36ed1bf" class="notion-header-anchor"></div><a class="notion-hash-link" href="#eb99b145357549219c0c7e4be36ed1bf" title="它和常见 agent 设计有什么不同"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">它和常见 agent 设计有什么不同</span></span></h3><div class="notion-text notion-block-5a1d5f07b4064f5cb6e405c9b3448d5a">常见 agent 设计往往从 loop 开始：plan -&gt; act -&gt; observe -&gt; reflect。这个模型直观，但有一个缺口：每一步留下来的东西是什么？如果只是聊天记录，系统很难复用、审计和纠错。</div><div class="notion-text notion-block-74055f8af4814557b87d321a7865649b"><code class="notion-inline-code">Code as Agent Harness</code> 的新意在于，它把“留下来的东西”摆到中心位置。测试、脚本、工具封装、执行轨迹、环境模型、技能库、仓库状态、工作流配置，都不是附属产物，而是 agent 系统的主干。这样一来，agent 的能力就不只取决于模型单轮回答，而取决于它能不能不断生产、执行、观察、修改这些可检查的中间对象。</div><div class="notion-text notion-block-c8748ea1b78a4f2b9a6b36446b27b47a">这里也有一些只是工程包装的地方。比如把 planning、memory、tool use 重新归类，本身并不新；很多 coding agent 和自动化平台早就在做。论文的价值不在于发明这些模块，而在于给它们一个更统一的解释：凡是能让 agent 长期运行、保存状态、接受反馈、支持协作和验证的代码对象，都应该被看成 harness 的一部分。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-a330796f81e642be9d58ac22795fae17" data-id="a330796f81e642be9d58ac22795fae17"><span><div id="a330796f81e642be9d58ac22795fae17" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a330796f81e642be9d58ac22795fae17" title="对产品和研发的启发"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">对产品和研发的启发</span></span></h3><div class="notion-text notion-block-4a24f16ffad1419597715a49478f409f">第一，别只优化 prompt。很多 agent 失败不是因为模型不会想，而是环境接口、权限边界、状态保存、错误反馈和验证方式太粗。把工具、执行日志、测试、失败样本和人工审批设计好，往往比多加几句系统提示更有效。</div><div class="notion-text notion-block-ebac832c1d734836838c4b9cdbc0812d">第二，把中间产物产品化。一个 agent 做完任务后，如果只给最终答案，价值会很薄。更好的形态是留下可复用脚本、测试用例、操作记录、决策说明、可回滚变更和失败原因。用户下次不是从零开始，而是在一套已经积累过的 harness 上继续。</div><div class="notion-text notion-block-f7b08a231c1242b898d7c8740d86744d">第三，评测要看过程。只看最后任务成功率，会掩盖很多问题：agent 是否用了危险工具，是否绕过权限，是否靠偶然通过测试，是否留下不可维护状态，是否在多 agent 协作中产生冲突。论文提到的 open problems 里，evaluation beyond final success 是我认为最实际的一条。</div><div class="notion-text notion-block-cf3434f7564f422b980ccb4a3a7a092b">第四，人类介入要变成状态，而不是临时打断。危险操作、权限申请、审查意见、回滚选择，都应该进入系统状态，供后续 agent 读取和遵守。否则 human-in-the-loop 只是一个弹窗，不是治理机制。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-cf5db16ca0b2452da40ef4a5294cdc59" data-id="cf5db16ca0b2452da40ef4a5294cdc59"><span><div id="cf5db16ca0b2452da40ef4a5294cdc59" class="notion-header-anchor"></div><a class="notion-hash-link" href="#cf5db16ca0b2452da40ef4a5294cdc59" title="风险和还没验证的问题"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">风险和还没验证的问题</span></span></h3><div class="notion-text notion-block-164f36790797484f8d5cfa5c64509f56">最大风险是把“代码可执行”误解成“代码就可靠”。执行反馈很有用，但反馈经常是不完整的：测试可能覆盖不足，环境可能有隐藏状态，GUI 操作可能成功但语义错误，机器人仿真可能过不了现实世界。论文也把 incomplete feedback 下的 verification 列为开放问题。</div><div class="notion-text notion-block-eb1a70ae480741ef81dd35c22964ce0c">第二个风险是共享状态会变成新的混乱源。多 agent 都能改仓库、记忆、工具和工作流时，冲突不只是文本冲突，还包括语义冲突、权限冲突和版本回退。没有事务、审计和回滚机制，多 agent 很容易把问题扩大。</div><div class="notion-text notion-block-487384c74bd349f7bb64938cc92b3056">第三个风险是自我优化可能带来回归。像 <code class="notion-inline-code">Agentic Harness Engineering</code> 这类方向已经在尝试让 agent 自动改 harness，但如果没有足够好的对照、预测和回滚，系统可能在某个 benchmark 上变好，同时破坏真实任务里的稳定表现。</div><div class="notion-text notion-block-798ad946a7ed4244a8535c641d91fde6">第四个问题是成本。代码作为 harness 往往意味着更多执行、更长日志、更复杂的状态管理和更多验证。对产品来说，最后要算的是 cost per completed task，而不是 token 单价。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-f8e81182a4914e6aabacf41bf33446bd"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://arxiv.org/html/2605.18747v1/x12.png?t=f8e81182-a491-4e6a-abac-f41bf33446bd" alt="图 3：论文把 code-as-harness 扩展到 coding assistant、GUI/OS agent、scientific discovery、personalization 和 embodied agent。来源：arXiv HTML Figure 12" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图 3：论文把 code-as-harness 扩展到 coding assistant、GUI/OS agent、scientific discovery、personalization 和 embodied agent。来源：arXiv HTML Figure 12</figcaption></div></figure><div class="notion-text notion-block-c85c0b50fee347e6b689452b5399e513">这张应用图提醒我们，code-as-harness 不只是 coding agent 的话题。GUI agent 需要把界面变成可观察状态，科学发现 agent 需要把假设、实验、数据分析串成程序世界，机器人 agent 需要把技能写成可复用、可验证的控制代码。场景不同，但共同点很清楚：真正有用的 agent 都需要一个能承载状态和反馈的执行层。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-bb8f2fd553f3472594a0ad642e49cab8" data-id="bb8f2fd553f3472594a0ad642e49cab8"><span><div id="bb8f2fd553f3472594a0ad642e49cab8" class="notion-header-anchor"></div><a class="notion-hash-link" href="#bb8f2fd553f3472594a0ad642e49cab8" title="今日沉淀"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">今日沉淀</span></span></h3><ol start="1" class="notion-list notion-list-numbered notion-block-49748bee40394d249997ab9896adfc56"><li>Agent 的核心不只是 loop，而是 loop 里留下了什么可执行、可检查、可复用的东西。</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-8b650fb1aad04b4aab8e0e356051d4c1"><li>Memory 更像状态管理，不只是把历史塞进向量库。</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-d4a290bad92c41ea9eda15c7d082eb7a"><li>Tool use 是治理接口，权限、执行位置、结果清洗和人工确认都要设计。</li></ol><ol start="4" class="notion-list notion-list-numbered notion-block-a12738eebc7041dd8e7f32e0d2b49487"><li>多 agent 协作不能只靠角色分工，要有共享状态、客观验证和冲突处理。</li></ol><ol start="5" class="notion-list notion-list-numbered notion-block-ebebdb9098f74c059a6b1eda4535730f"><li>评测不能只看最终成功率，还要看过程是否安全、可审计、可回滚。</li></ol></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agent 工程化与算力供应链 - 2026-06-30]]></title>
            <link>http://easyai.fyi/article/follow-builders-ai-summary-2026-06-30</link>
            <guid>http://easyai.fyi/article/follow-builders-ai-summary-2026-06-30</guid>
            <pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[今日全量沉淀包括 15 位 X builders 的 32 条推文、3 篇 Claude/Anthropic 文章和 1 期 No Priors 播客。主线是 Agent 正在进入企业执行环境，同时 AI 算力、CPU、foundry 和数据中心继续成为底层瓶颈。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-38e5eac3813c81ae8ba0cf9f4f4515e9"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-eb4a45d407c54c6588aeeb01425ce1f9" data-id="eb4a45d407c54c6588aeeb01425ce1f9"><span><div id="eb4a45d407c54c6588aeeb01425ce1f9" class="notion-header-anchor"></div><a class="notion-hash-link" href="#eb4a45d407c54c6588aeeb01425ce1f9" title="今日主线"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">今日主线</span></span></h3><div class="notion-text notion-block-4c81ce5e662b4b2c860f71ac89fd49a6">今天的 follow-builders 内容集中在两条线：一条是 Agent 正在变成真正的工程系统，开始处理后台子任务、权限边界、企业沙箱、MCP tunnel 和质量事故；另一条是 AI 底层供应链继续吃紧，CPU、foundry、先进封装、数据中心和电力都重新变成关键变量。全量数据包括 15 位 X builders 的 32 条推文、3 篇 Claude/Anthropic 文章和 1 期 No Priors 播客。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-23aa8437b5634f5b8f46153db5144255" data-id="23aa8437b5634f5b8f46153db5144255"><span><div id="23aa8437b5634f5b8f46153db5144255" class="notion-header-anchor"></div><a class="notion-hash-link" href="#23aa8437b5634f5b8f46153db5144255" title="重点解读"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">重点解读</span></span></h3><div class="notion-text notion-block-37c0812ac5f44208a9cfba3680d392fb">Claude 这边今天的信息密度最高。Claude Code 要让 subagents 默认在后台运行，Codex 也在推进更细的权限配置，这说明 coding agent 的重点正在从“能不能写代码”转向“能不能长期、安全、可控地跑在真实工作流里”。Anthropic 的几篇文章也在补同一条主线：Managed Agents 要支持自托管 sandbox、MCP tunnels、可恢复 session、凭证隔离和更稳定的 harness。换句话说，Agent 产品正在从 demo 工具进入企业执行环境。</div><div class="notion-text notion-block-ad659df09ecd465aa3afd875e30d1302">另一个主线是算力和半导体。No Priors 采访 Intel CEO Lip-Bu Tan，他强调 AI 不只是拉动 GPU，也会重新提高 CPU、foundry、先进封装和供应链韧性的价值。他提到 agent 编排、reinforcement learning 和数据中心需求会让 CPU 重新重要，同时电力、helium、memory、fab 建设周期都会限制 AI 增长。这和今天 builder 圈里关于数据中心、电力、云基础设施的讨论放在一起看，很清楚：AI 的瓶颈越来越像工程和供应链问题。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-eb3db84a53c94b1fa6eef454724ef2cb" data-id="eb3db84a53c94b1fa6eef454724ef2cb"><span><div id="eb3db84a53c94b1fa6eef454724ef2cb" class="notion-header-anchor"></div><a class="notion-hash-link" href="#eb3db84a53c94b1fa6eef454724ef2cb" title="X Builders 全量记录"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">X Builders 全量记录</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-41220c2938e4466da8d24e7268aebb02" data-id="41220c2938e4466da8d24e7268aebb02"><span><div id="41220c2938e4466da8d24e7268aebb02" class="notion-header-anchor"></div><a class="notion-hash-link" href="#41220c2938e4466da8d24e7268aebb02" title="1. Swyx"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">1. Swyx</span></span></h4><div class="notion-text notion-block-2cbddb65ba434e3892b7e1111d7ea052">Swyx 今天主要在 AI Engineer Expo 现场。他发了 workshop day 的现场内容，也提到周一早上 9 点、旁边还有 OpenAI workshop 的情况下，Snyk、Atlassian、Neo4j、Arize、Akamai、Together 等房间仍然很满。他的判断很直接：builder 社区对 AI 工程实践的需求非常强，不只是追模型发布。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-38e5eac3813c800c8c3ce88965efcd1e"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071692683182252317" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-02d406d0aade425ab0763e401aab85e8"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071634789669777716" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-2f81d7a5445e47eb967ca33e847f4519"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071613383380770823" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-64793aa27cf9444b98d9c7f8f3fa28a4" data-id="64793aa27cf9444b98d9c7f8f3fa28a4"><span><div id="64793aa27cf9444b98d9c7f8f3fa28a4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#64793aa27cf9444b98d9c7f8f3fa28a4" title="2. Boris Cherny"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">2. Boris Cherny</span></span></h4><div class="notion-text notion-block-e872fb89f0df43b38f5e4b8913acce87">Boris Cherny 是 Claude Code 团队成员。他透露下一版 Claude Code 会让 subagents 默认在后台运行，这样用户可以继续和 Claude 对话，同时让子 agent 继续处理任务。如果需要前台执行，也可以直接告诉 Claude。这个变化小但关键，因为它把 coding agent 从单线程对话推向更接近“并行工作台”的形态。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-bf761122d8f248cc863e8e825172844c"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071647677591466098" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-926e24df03ac497cacba90e8ce9a9077" data-id="926e24df03ac497cacba90e8ce9a9077"><span><div id="926e24df03ac497cacba90e8ce9a9077" class="notion-header-anchor"></div><a class="notion-hash-link" href="#926e24df03ac497cacba90e8ce9a9077" title="3. Thibault Sottiaux"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">3. Thibault Sottiaux</span></span></h4><div class="notion-text notion-block-050ae4e25e524944b35f2de8aadf88ff">Thibault Sottiaux 今天有两条 Codex 相关的重要更新。第一条是 Codex 使用额度会再次重置，并额外补一个 24 小时 reset；他们调查后发现不是单点事故，而是 auto-review 更主动、更多 subagent work、background suggestions 等几个变化叠加，导致部分用户觉得消耗异常。第二条是 Codex 推出替代粗粒度 sandbox mode 的 permission profiles，可以把文件读写规则、网络域名、Unix socket、管理员 allowlist 绑定到可复用权限配置里。中间那条 reset button 是轻量玩笑，也一并保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-c8a287565ec442ca88d273d2a1634d10"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071740419030053227" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-5c5a5ef2facf48639e84c92ce5a6894f"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071710834527523030" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-5dcb88bfed45433aa5d47f5c93e4cfaa"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071636285807059315" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-07481f3cb9f54f9aad9a3524d1213785" data-id="07481f3cb9f54f9aad9a3524d1213785"><span><div id="07481f3cb9f54f9aad9a3524d1213785" class="notion-header-anchor"></div><a class="notion-hash-link" href="#07481f3cb9f54f9aad9a3524d1213785" title="4. Peter Yang"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">4. Peter Yang</span></span></h4><div class="notion-text notion-block-0773e4a14bb34562a6e71de58096614a">Peter Yang 的几条内容都围绕 AI 写作和模型开放争议。他认为纯写作和编辑场景里，普通 Claude web 仍然比 Codex、Claude Code 这类 coding agent 更好，猜测可能是 coding agent 的系统提示影响了写作质量。他还说自己对 restricted access 和 open source 有很多强观点，会整理成一篇文章，并补充已有 written post。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-9607bd49b704442b975f556200a31988"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071731343390851519" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-2213bb6ee0814ceca676a6539f9001e2"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071730786886435261" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-02462dc55a474cf6826a314b1309265f"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071690806440898916" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-2d64b2dc230b436c8d48b8af7aa47189" data-id="2d64b2dc230b436c8d48b8af7aa47189"><span><div id="2d64b2dc230b436c8d48b8af7aa47189" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2d64b2dc230b436c8d48b8af7aa47189" title="5. Madhu Guru"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">5. Madhu Guru</span></span></h4><div class="notion-text notion-block-4c8820d469894c8b866c25e83368dcaa">Madhu Guru 的判断是，强 open-weight 模型比如 GLM 反而会加强 Google 的位置。理由是更多企业会尝试 fine-tuning open-weight 模型，但最后价值会流向基础设施，因为企业需要托管平台、可靠性、安全、支持和完整 compute stack。他特别点出 Google Cloud 在这类需求里位置不错，因为 Google 掌握很多底层 compute stack。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-4c03ae9005e54c2492810e140f15dae1"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071637885154148785" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-1f2835376633489d824bbca927a29382" data-id="1f2835376633489d824bbca927a29382"><span><div id="1f2835376633489d824bbca927a29382" class="notion-header-anchor"></div><a class="notion-hash-link" href="#1f2835376633489d824bbca927a29382" title="6. Thariq"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">6. Thariq</span></span></h4><div class="notion-text notion-block-686a64a27c384d78ac34d27832eaf60e">Thariq 分享了他现在的写作流程：先做工程工作，再和很多人聊，用 Claude 做 brainstorm 和 research，写文章，讲一两次，再反复改。这个细节值得保留，因为它说明 AI builder 的写作不是简单让模型代写，而是把工程、交流、模型辅助和公开表达串成一个循环。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-196d5605a69b407f95bed1351343a525"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071787401475960892" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-255bf8d27fa24e60be3f27825ec39817" data-id="255bf8d27fa24e60be3f27825ec39817"><span><div id="255bf8d27fa24e60be3f27825ec39817" class="notion-header-anchor"></div><a class="notion-hash-link" href="#255bf8d27fa24e60be3f27825ec39817" title="7. Guillermo Rauch"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">7. Guillermo Rauch</span></span></h4><div class="notion-text notion-block-186315df2ef7468c994a71004edac1e7">Guillermo Rauch 今天主要在为 Vercel 的新能力预热。他提到 Vercel functions 变成 20 倍更大，又说“anything and everything” 都可以部署到 Vercel，更多内容明天公布。另一条是一个通过 curl 触发的 demo。整体看，Vercel 仍在把自己往更通用的 AI/应用运行平台上推。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-42230009c32a48f48bb35b92d68ffafc"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071718135799927224" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-75217278b6e94ce0990cada4abe64354"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071716510389662153" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-486831fe6ce24374b8dae4c58a99eccb"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071710688150528443" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-7b50fd7f810141b49a6fef765cfe4471" data-id="7b50fd7f810141b49a6fef765cfe4471"><span><div id="7b50fd7f810141b49a6fef765cfe4471" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7b50fd7f810141b49a6fef765cfe4471" title="8. Aaron Levie"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">8. Aaron Levie</span></span></h4><div class="notion-text notion-block-ea0c2b4fdefe4f57a3ecb37b2911a3b1">Aaron Levie 继续讨论 AI 开放和监管。他的核心判断是：如果闭源垂直栈永远大幅领先，那么 gatekeeping 和严格控制可能成立；但如果 open weights 能长期接近 frontier intelligence，那么过度监管会让局面反过来，限制方可能失去扩散优势。这条和最近模型开放、访问限制、国家竞争放在一起看，很有价值。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-861b2b5797fd437d823b82fc4edfd47e"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071775583072375214" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-abf454ac353740be8e86f0bbd52f26aa" data-id="abf454ac353740be8e86f0bbd52f26aa"><span><div id="abf454ac353740be8e86f0bbd52f26aa" class="notion-header-anchor"></div><a class="notion-hash-link" href="#abf454ac353740be8e86f0bbd52f26aa" title="9. Ryo Lu"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">9. Ryo Lu</span></span></h4><div class="notion-text notion-block-6c720fa009044e1c931fcbf8cedd3cc2">Ryo Lu 今天的两条更像 Cursor 相关产品表达：一条是“get the app”，另一条强调 ideas 可以随时出现，不需要桌子和笔记本电脑。信息不长，但它反映了一个方向：AI coding 工具正在从桌面 IDE 场景，往更轻、更移动、更随时记录想法的入口延伸。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-bc7f53f1a4bd495eb6b5fc24af5f11cb"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071655130152493297" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-8dbc7221ec834d6b90150ee2a09e2594"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071652629890088964" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-a977a9f281924aa0ba046224824f4b71" data-id="a977a9f281924aa0ba046224824f4b71"><span><div id="a977a9f281924aa0ba046224824f4b71" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a977a9f281924aa0ba046224824f4b71" title="10. Garry Tan"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">10. Garry Tan</span></span></h4><div class="notion-text notion-block-361b5b68f673403badbbf588488858ef">Garry Tan 今天一条是 link-only 内容，信息量有限；另一条很直接：build power and datacenters。后者和今天的 AI 算力主线很一致。模型和 agent 走到真实规模后，电力和数据中心建设不再是背景问题，而是产品增长和国家竞争的底座。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-d2c1e59349484deeb888b15e419f4f44"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071817410303393840" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-063a947742374ba792b9571b747cc615"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071600933210100074" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-7e6c9a71b2df46a8a5566469adbee5fc" data-id="7e6c9a71b2df46a8a5566469adbee5fc"><span><div id="7e6c9a71b2df46a8a5566469adbee5fc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7e6c9a71b2df46a8a5566469adbee5fc" title="11. Matt Turck"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">11. Matt Turck</span></span></h4><div class="notion-text notion-block-ffc80544e8d649a8aa43c0263f048ef1">Matt Turck 一条说“反复做 underdog 是最大优势”，偏创业心态；另一条是 World Cup 门票价格的玩笑，和 AI 主线关系不大。按全量沉淀要求，这两条也保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-0124c23dccbc4bfc981f15c6781b708d"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071806129001164934" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-e43629575ce64684afdd4a00fc58c47f"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071772069742756333" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-5acd450c58e44441b9984e9bfe463c59" data-id="5acd450c58e44441b9984e9bfe463c59"><span><div id="5acd450c58e44441b9984e9bfe463c59" class="notion-header-anchor"></div><a class="notion-hash-link" href="#5acd450c58e44441b9984e9bfe463c59" title="12. Zara Zhang"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">12. Zara Zhang</span></span></h4><div class="notion-text notion-block-c73d1343908c4d39b777e167665e85a6">Zara Zhang 今天发了一个小工具：Chrome extension，把 read later 列表自动变成 Google Calendar 上的阅读时间块。比如保存 5 篇文章，就自动安排 30 分钟 reading block，链接也带上；没有账号、没有服务器、本地运行、开源。另一条是她转述 Anthropic PM 的观点：写作的市场价值上升了，因为好写作既能帮助 steer models，也能帮助建立受众。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-f3520ec4bbb9434490969adf8eb59cec"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071766865245012255" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-8f8d5ee5ccfd41d6964cdd68550fd320"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071766827345285411" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-f3d6df196ad749abb92cf40847024e96"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071670108033073365" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-8b3a9f6747aa45258e280256569df6a5" data-id="8b3a9f6747aa45258e280256569df6a5"><span><div id="8b3a9f6747aa45258e280256569df6a5" class="notion-header-anchor"></div><a class="notion-hash-link" href="#8b3a9f6747aa45258e280256569df6a5" title="13. Nikunj Kothari"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">13. Nikunj Kothari</span></span></h4><div class="notion-text notion-block-69b70edb78584c9e8a0cde148a4a9a10">Nikunj 今天三条都是 World Cup 相关：比赛、点球大战、球队混淆修正。它们不是 AI 主线，但属于本次 follow-builders 抓取内容，完整保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-73e05829b387447d90496ade04b357c7"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071807436307222968" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-222b46724767435a857c087c82d89d15"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071803912785666483" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-b39ce4566c0440648c889c735fa6ce22"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071798024536572123" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-ee37d741d4a34bd084f3208e7d35c15a" data-id="ee37d741d4a34bd084f3208e7d35c15a"><span><div id="ee37d741d4a34bd084f3208e7d35c15a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ee37d741d4a34bd084f3208e7d35c15a" title="14. Peter Steinberger"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">14. Peter Steinberger</span></span></h4><div class="notion-text notion-block-650266ffe2584ffb839b7b226f137eff">Peter Steinberger 今天两条都比较短。一条问“was this vibed”，另一条说某个东西几年前会很惊艳，但在 AI 时代意义变小。上下文不完整，但核心情绪是：AI 正在快速改变人们对产品、demo 或自动化能力的惊艳阈值。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-6e1733cf0b8e404184ed6257949a06f0"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071770560875671831" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-ccf6ddb26125469ca834d3522424b3bd"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071769993151398074" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-d6a41e0a826c4149b4319f6301e0eab8" data-id="d6a41e0a826c4149b4319f6301e0eab8"><span><div id="d6a41e0a826c4149b4319f6301e0eab8" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d6a41e0a826c4149b4319f6301e0eab8" title="15. Claude"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">15. Claude</span></span></h4><div class="notion-text notion-block-4c82a9bcc87847239607461b327951e5">Claude 官方账号宣布 Claude in Microsoft Foundry 已经 GA，托管在 Azure 上。Azure 客户可以使用 Claude Opus 4.8 和 Claude Haiku 4.5，并接入 Azure authentication、billing 和 commitment retirement。另一条补充说 inference 运行在 Anthropic 操作的 Azure infrastructure 上，prompt caching 和 extended thinking 已支持，更多能力会继续加入。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-1d920c526ca446dd981dd28b0f7b89c6"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071653962013446586" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-51bc4b07016943e8b8bf551d3814d298"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://platform.twitter.com/embed/Tweet.html?id=2071653958905467027" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-3c6381e164334d7b8e391e1ab7ac168b" data-id="3c6381e164334d7b8e391e1ab7ac168b"><span><div id="3c6381e164334d7b8e391e1ab7ac168b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3c6381e164334d7b8e391e1ab7ac168b" title="Blog 全量记录"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Blog 全量记录</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-50942acb71904bca9de77f0e0d972cd1" data-id="50942acb71904bca9de77f0e0d972cd1"><span><div id="50942acb71904bca9de77f0e0d972cd1" class="notion-header-anchor"></div><a class="notion-hash-link" href="#50942acb71904bca9de77f0e0d972cd1" title="1. Anthropic Engineering：An update on recent Claude Code quality reports"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">1. Anthropic Engineering：An update on recent Claude Code quality reports</span></span></h4><div class="notion-text notion-block-380c00840ae0493e9083d32ca432bcc0">Anthropic Engineering 这篇是一次质量问题复盘。核心结论是，过去一个月一些用户感觉 Claude 变差，并不是 API 整体退步，而是三个不同变化分别影响了 Claude Code、Claude Agent SDK 和 Claude Cowork；这些问题已经在 4 月 20 日的 v2.1.116 修复。值得关注的是 Anthropic 把问题拆得很具体：前沿模型产品的“质量”不只来自模型本身，还来自工具层、agent loop、评测、发布流程和回滚机制。原文：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.anthropic.com/engineering/april-23-postmortem">An update on recent Claude Code quality reports</a></div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-c03584fd61ec4a47a329118c9278a10d" data-id="c03584fd61ec4a47a329118c9278a10d"><span><div id="c03584fd61ec4a47a329118c9278a10d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c03584fd61ec4a47a329118c9278a10d" title="2. Anthropic Engineering：Scaling Managed Agents: Decoupling the brain from the hands"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">2. Anthropic Engineering：Scaling Managed Agents: Decoupling the brain from the hands</span></span></h4><div class="notion-text notion-block-d6fce2b385184decb40843fc32f56bc0">这篇是今天最重要的技术长文之一。Anthropic 讲 Managed Agents 为什么要把“brain”和“hands”拆开：brain 负责推理和编排，hands 是 sandbox、工具和外部执行环境；中间用 durable session log 记录事件，让 harness 崩了也能恢复。文章还提到，把执行环境和凭证隔离后，可以降低 prompt injection 直接摸到 token 的风险；并且因为不用每次先启动完整容器，p50 time-to-first-token 降了约 60%，p95 降了超过 90%。原文：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.anthropic.com/engineering/managed-agents">Scaling Managed Agents</a></div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-725d50c685be43dabc565bce4fa08b6a" data-id="725d50c685be43dabc565bce4fa08b6a"><span><div id="725d50c685be43dabc565bce4fa08b6a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#725d50c685be43dabc565bce4fa08b6a" title="3. Claude Blog：New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">3. Claude Blog：New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels</span></span></h4><div class="notion-text notion-block-47b061d77dcf40d3a967b4e5156326c0">Claude Blog 宣布 Managed Agents 支持 self-hosted sandboxes 和 MCP tunnels。意思是 agent 的工具执行可以跑在企业自己的 sandbox 或 Cloudflare、Daytona、Modal、Vercel 等托管环境里，敏感文件、包和内部服务不用离开企业边界；MCP tunnels 则让 agent 能访问私有网络里的 MCP server，不需要把内部服务暴露到公网。这个功能很企业向，重点是把 agent 从“云端黑盒”推进到可审计、可控、可接入私有系统的部署方式。原文：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://claude.com/blog/claude-managed-agents-updates">New in Claude Managed Agents</a></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-d598b47594ae40189ebe3197644c4f7e" data-id="d598b47594ae40189ebe3197644c4f7e"><span><div id="d598b47594ae40189ebe3197644c4f7e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d598b47594ae40189ebe3197644c4f7e" title="Podcast 全量记录"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Podcast 全量记录</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-88401d10fd5d40e69450463797266be9" data-id="88401d10fd5d40e69450463797266be9"><span><div id="88401d10fd5d40e69450463797266be9" class="notion-header-anchor"></div><a class="notion-hash-link" href="#88401d10fd5d40e69450463797266be9" title="1. No Priors：Re-engineering the Semiconductor Supply Chain with Intel CEO Lip-Bu Tan"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">1. No Priors：Re-engineering the Semiconductor Supply Chain with Intel CEO Lip-Bu Tan</span></span></h4><div class="notion-text notion-block-8fc1ac4b8b204c78aac38ee94bdc82cb">No Priors 这期和 Intel CEO Lip-Bu Tan 聊半导体供应链。最值得记的是他对 AI 硬件瓶颈的判断：AI 不只需要 GPU，agentic AI、reinforcement learning 和 agent 编排也会重新拉高 CPU 需求。他还反复强调 foundry 是服务和信任生意，先进制程、良率、cycle time、先进封装、材料、memory、电力、helium 都会成为限制项。对 Intel 来说，政府入股、NVIDIA 投资、SoftBank 支持、和 Elon Musk 的 TeraFab 协作，都是为了让美国半导体供应链更有韧性。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-b302858fbbf54b2287150cffd640a22c"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:320px"><iframe class="notion-asset-object-fit" src="https://www.youtube.com/embed/asCgCv2XB4s" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-c722065579c2437bb9f55ff7b0ff2671" data-id="c722065579c2437bb9f55ff7b0ff2671"><span><div id="c722065579c2437bb9f55ff7b0ff2671" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c722065579c2437bb9f55ff7b0ff2671" title="今日沉淀结论"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">今日沉淀结论</span></span></h3><div class="notion-text notion-block-59772ea6de4143c98cd8dfdf635b9169">今天所有内容放在一起看，AI 的主线已经很清楚：Agent 正在进入真实企业系统，所以权限、sandbox、MCP、session、质量复盘和后台并行执行会变得越来越重要；同时，模型能力越往前走，底层供应链越不能被忽略，CPU、数据中心、电力、memory、foundry 和封装都会继续决定 AI 产品能跑多远。</div><div class="notion-text notion-block-67b7b5b12a8642eca546054d107e7480">Generated through the Follow Builders skill: <a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/zarazhangrui/follow-builders">https://github.com/zarazhangrui/follow-builders</a></div></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Anthropic AI Agents科普]]></title>
            <link>http://easyai.fyi/article/ai-agents-intro</link>
            <guid>http://easyai.fyi/article/ai-agents-intro</guid>
            <pubDate>Tue, 22 Jul 2025 00:00:00 GMT</pubDate>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-1df5eac3813c80ec9cb5c2aabb4a65b4"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><div class="notion-row"><a target="_blank" rel="noopener noreferrer" class="notion-bookmark notion-block-2385eac3813c80958d63d2806ebd10c8" href="https://www.anthropic.com/engineering/building-effective-agents"><div><div class="notion-bookmark-title">Building Effective AI Agents</div><div class="notion-bookmark-description">Discover how Anthropic approaches the development of reliable AI agents. Learn about our research on agent capabilities, safety considerations, and technical framework for building trustworthy AI.</div><div class="notion-bookmark-link"><div class="notion-bookmark-link-icon"><img src="https://www.anthropic.com/images/icons/favicon-32x32.png?t=2385eac3-813c-8095-8d63-d2806ebd10c8" alt="Building Effective AI Agents" loading="lazy" decoding="async"/></div><div class="notion-bookmark-link-text">https://www.anthropic.com/engineering/building-effective-agents</div></div></div><div class="notion-bookmark-image"><img style="object-fit:cover" src="https://cdn.sanity.io/images/4zrzovbb/website/76b5733c669f0dfb9c7aa7fc512a495867cf12e6-2400x1260.png?t=2385eac3-813c-8095-8d63-d2806ebd10c8" alt="Building Effective AI Agents" loading="lazy" decoding="async"/></div></a></div></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[NotebookLLM]]></title>
            <link>http://easyai.fyi/article/1b85eac3-813c-808d-931f-d347a3ec39a3</link>
            <guid>http://easyai.fyi/article/1b85eac3-813c-808d-931f-d347a3ec39a3</guid>
            <pubDate>Sun, 16 Mar 2025 00:00:00 GMT</pubDate>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-1b85eac3813c808d931fd347a3ec39a3"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><div class="notion-text notion-block-2005eac3813c8077ac06cee312b4a60b">关于NotebookLLM</div><figure class="notion-asset-wrapper notion-asset-wrapper-tweet notion-block-2005eac3813c80318f24e23c2cd31a96"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><div style="max-width:420px;width:100%;margin-left:auto;margin-right:auto"><div class="notion-tweet-embed"><a class="notion-tweet-link" href="https://x.com/i/web/status/1926555449899233515" target="_blank" rel="noopener noreferrer">查看 x.com 原文</a></div></div></div></figure><div class="notion-row"><a target="_blank" rel="noopener noreferrer" class="notion-bookmark notion-block-2005eac3813c80ceabaaf7b3ad1eed82" href="https://v2ex.com/t/1098014"><div><div class="notion-bookmark-title">大模型时代，应该用什么知识管理软件？ - V2EX</div><div class="notion-bookmark-description">问与答 - @afterzero - 在笔记时代，我每天在用 flomo 、微信收藏、OneNote 、Markdown 文件、浏览器收藏夹等工具。但是 [我不想把收藏变成一个**把喜欢的东西丢进垃圾堆**的过程。] 大模型可以帮我捡</div><div class="notion-bookmark-link"><div class="notion-bookmark-link-icon"><img src="https://v2ex.com/static/icon-192.png?t=2005eac3-813c-80ce-abaa-f7b3ad1eed82" alt="大模型时代，应该用什么知识管理软件？ - V2EX" loading="lazy" decoding="async"/></div><div class="notion-bookmark-link-text">https://v2ex.com/t/1098014</div></div></div><div class="notion-bookmark-image"><img style="object-fit:cover" src="https://cdn.v2ex.com/gravatar/f5cb03c1357f8c1899ffff051e75603d?s=73&amp;d=retro&amp;t=2005eac3-813c-80ce-abaa-f7b3ad1eed82" alt="大模型时代，应该用什么知识管理软件？ - V2EX" loading="lazy" decoding="async"/></div></a></div></main></div>]]></content:encoded>
        </item>
    </channel>
</rss>