<?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>Tue, 14 Jul 2026 04:00:24 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[模型额度、Agent 工作流与能源底座 - 2026-07-13]]></title>
            <link>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-13-model-agent-energy</link>
            <guid>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-13-model-agent-energy</guid>
            <pubDate>Mon, 13 Jul 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[本轮 follow-builders 新增 15 个 X builders、30 条 tweet 和 1 期 podcast，已按输出全量沉淀。今天主线是 GPT-5.6 Sol 与 Claude Fable 5 的付费访问和额度调整、Agentic coding 进入更真实的工作流，以及 AI 算力背后的能源底座。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-39c5eac3813c81f9af4aecdbd16d1bd2"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-b5cbc5887d444cbb85533da50c763a2c" data-id="b5cbc5887d444cbb85533da50c763a2c"><span><div id="b5cbc5887d444cbb85533da50c763a2c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b5cbc5887d444cbb85533da50c763a2c" 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-1db09bfe775a449fb2becce73a095d9b">今天这批 follow-builders 内容有三条主线。第一，GPT-5.6 Sol 和 Claude Fable 5 都在围绕付费访问、额度、成本和社区沟通继续调整，模型能力之外，产品可用性开始变成用户最敏感的问题。第二，Agentic coding 不是只停在“能写代码”，它开始进入会议转 PRD、并行实验、企业模型路由、eval 和数据控制这些真实工作流。第三，No Priors 这期把 AI compute 的电力问题拉到核能层面：如果算力需求继续上升，能源会成为 AI 产业绕不开的底座。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-94ae48a590804e02ba58667b716ac806" data-id="94ae48a590804e02ba58667b716ac806"><span><div id="94ae48a590804e02ba58667b716ac806" class="notion-header-anchor"></div><a class="notion-hash-link" href="#94ae48a590804e02ba58667b716ac806" 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-63229fdf18d64c61986d67baa4bdd5fd">Thibault Sottiaux 对 Codex 和 ChatGPT Work 的说明，是今天最具体的产品信号。他说 GPT-5.6 Sol 做了推理优化，会把节省下来的成本反映到订阅里，带来大约 10% 的额外可用量；同时，团队发现把 Sol 上下文限制从 GPT-5.5 的 272k 提到 372k 后，实际扣量比预期更高，所以先回退到 272k，并继续调整 reasoning effort、多 agent 和 auto-review 的消耗。这个信息比单纯说“模型变强”更重要：用户真正关心的是模型能不能稳定用、扣量是否可预期、长任务是否值得交给它。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-cfc74369e4a14539ba93d9218c3ae9d8"><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=2076495156757577895" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><div class="notion-text notion-block-4e3866a6c5ad457fb9a23ad71a954097">Claude 官方也在做相似的访问管理：Fable 5 对所有付费计划继续开放到 7 月 19 日，Claude Code 的周额度也继续保持 50% 上调。另一条说明补充说，用户最多可以把每周额度的一半用在 Fable 5 上，之后可以继续用 usage credits，或者切到其他模型。这说明前沿模型的竞争正在从“谁最强”转向“谁能让用户持续、可控地用”。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-698663489691461e9ebeb92bcd1377ea"><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=2076351399999557669" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-eb43faa12a0542dab0d62bfcd7575857"><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=2076351401006154204" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><div class="notion-text notion-block-56cdcd36b0774666a5d0447ec5133dcf">企业 AI 层的讨论也更清楚了。Guillermo Rauch 说，企业应该把模型当成自己机器里的一个齿轮，而不是把大脑外包出去：数据、eval、模型选择、软件层都要掌握在自己手里。Aaron Levie 的说法和这个方向一致：当所有公司都能访问前沿智能，真正有价值的是怎么把企业自己的决策、洞察、流程模式和最佳实践沉淀进 AI 工作流里。换句话说，模型本身会越来越通用，企业自己的 workflow、trace、eval 和信息复利会变得更值钱。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-16af200dc1334e1eb4b371b233c22205"><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=2076364176252191222" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-c63e8310f17341629f2627635a6e00d0"><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=2076338364635287637" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><div class="notion-text notion-block-0c91c443e4e64f3eb05fd6ff51b205da">No Priors 这期和 Valar Atomics 创始人 Isaiah Taylor 聊核能。节目一开始就把问题讲得很直接：AI compute 正在推高美国电力需求，而核能如果要真正成为答案，就不能只停留在建模、仿真和审批里，要走向硬件迭代、制造化和规模化。Taylor 的核心判断是，能源是商品，价格下降会创造需求；如果能把核裂变做成更像制造业的产品，AI 产业的能源瓶颈才可能被系统性缓解。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-30d7f4a398ff43128cd7d401b31af105"><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/5Xvbq_zvOQ4" 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-4f6ffee4ac304249bdf6e00de035bdbc" data-id="4f6ffee4ac304249bdf6e00de035bdbc"><span><div id="4f6ffee4ac304249bdf6e00de035bdbc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4f6ffee4ac304249bdf6e00de035bdbc" 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-62b644968bc24947ba8069a755277c03" data-id="62b644968bc24947ba8069a755277c03"><span><div id="62b644968bc24947ba8069a755277c03" class="notion-header-anchor"></div><a class="notion-hash-link" href="#62b644968bc24947ba8069a755277c03" title="1. Swyx @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 @swyx</span></span></h4><div class="notion-text notion-block-6feae41ff39f4f08acbe2338d6627289">Swyx 今天两条都偏研究和内容分发。一条用很口语化的方式谈 introspection / backpropagation 和多次 rollout 的差别，大意是光做多次尝试但不根据优势信号改进，不能算真正学习；另一条是补充 Latent Space 的 writeup。前者和 agent / RL 讨论有关，信息量在于提醒大家：多跑几次不等于会变聪明，关键是反馈怎么进入下一轮。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-11af205aca9c40828065059b795ee120"><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=2076345087634620528" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-bc3144875cdb4958b94c852e7b579420"><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=2076216180529156097" 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-45538ab320534713b1984221099e6860" data-id="45538ab320534713b1984221099e6860"><span><div id="45538ab320534713b1984221099e6860" class="notion-header-anchor"></div><a class="notion-hash-link" href="#45538ab320534713b1984221099e6860" title="2. Thibault Sottiaux @thsottiaux"><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 @thsottiaux</span></span></h4><div class="notion-text notion-block-4cec886734294178861629d012bf69fc">Thibault 今天三条都围绕 GPT-5.6 Sol 的订阅和额度。核心是：Sol 会继续留在 Go、Plus、Pro、Team、Edu 等付费订阅里，直到更好的模型上线；同时 Codex 和 ChatGPT Work 的推理优化、上下文限制、reasoning effort、多 agent 和 auto-review 消耗都在被重新校准。对开发者来说，这不是小修小补，而是前沿模型进入高频工作流后必须面对的运营问题：强模型不仅要强，还要让用户知道自己能用多久、会怎么扣量。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-150e188882504f6fa6c8502be43e0eac"><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=2076495156757577895" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-98c4e16c647c47c29db502b8238cfaef"><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=2076460408437887268" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-e30e44eb96f7482392f9f1b4017d8fcb"><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=2076459871021736245" 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-e8d766ff47724dbe80f46a37b4469974" data-id="e8d766ff47724dbe80f46a37b4469974"><span><div id="e8d766ff47724dbe80f46a37b4469974" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e8d766ff47724dbe80f46a37b4469974" title="3. Peter Yang @petergyang"><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 @petergyang</span></span></h4><div class="notion-text notion-block-d9f927e6de8f42fd8f8fdc2a1d0f80ea">Peter 今天的三条重点在社区沟通。他观察到很多人可能都在用 GPT-5.6 Sol，而 Terra / Luna 的使用占比可能很低；另外两条讲的是公司在社区情绪转差时，反而应该更透明、更像真人一样沟通，而不是更少说话、更企业腔。他还拿 OpenAI 的沟通方式和 Anthropic 做对比。这里值得记的是：模型产品已经像开发者基础设施一样，需要公开、及时、能解释取舍的沟通。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-50b9ea521e0f4054b9aa2e85e5042243"><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=2076519927843000448" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-8290e7ad001b4e78af3df708b445883c"><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=2076512796481880270" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-8ebe0ce5fdf745179284adc159618d38"><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=2076510899490480228" 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-9cb2eadc8bd34d25b2babc0fca75b6d6" data-id="9cb2eadc8bd34d25b2babc0fca75b6d6"><span><div id="9cb2eadc8bd34d25b2babc0fca75b6d6" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9cb2eadc8bd34d25b2babc0fca75b6d6" title="4. Cat Wu @_catwu"><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. Cat Wu @_catwu</span></span></h4><div class="notion-text notion-block-df971394806a4ddaa24d2edf2d7896b5">Cat Wu 发了一条很短的“Enjoy!”并引用了外部内容。单看这条信息量不大，但她本身在 Claude Code / Anthropic cowork 相关方向上，属于值得持续跟踪的 builder。这里按全量沉淀保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-acccfc9ad4e24142a64262ef35995873"><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=2076358263688569314" 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-72efc99a6c1e432fb9fa671e35be8204" data-id="72efc99a6c1e432fb9fa671e35be8204"><span><div id="72efc99a6c1e432fb9fa671e35be8204" class="notion-header-anchor"></div><a class="notion-hash-link" href="#72efc99a6c1e432fb9fa671e35be8204" title="5. Amjad Masad @amasad"><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. Amjad Masad @amasad</span></span></h4><div class="notion-text notion-block-3b0f411334324abab0720e1295dfc04c">Amjad 的两条很有 builder 味道。他一边让 Replit 的 computer use 模型和自己新写的棋类引擎对弈，一边说自己在 Replit 上 fine-tune 一个 Qwen-8B 模型下棋，同时开了三个并行分支做不同实验。他的判断是，模型做 ML 工作的能力已经比以前强很多，有直觉的人即使不是传统 ML 背景，也能带着模型做出有意思的实验。这是 AI 编程从写应用代码往研究/实验自动化扩展的一个小信号。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-868157d9c0cf4c69a9b6cd1d76dd5a28"><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=2076356893736673507" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-a18de9ec74594bb08769509d1e05019e"><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=2076227936202662357" 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-d6013fd8d6c24800a091d4f2e9124b91" data-id="d6013fd8d6c24800a091d4f2e9124b91"><span><div id="d6013fd8d6c24800a091d4f2e9124b91" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d6013fd8d6c24800a091d4f2e9124b91" title="6. Guillermo Rauch @rauchg"><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. Guillermo Rauch @rauchg</span></span></h4><div class="notion-text notion-block-24b98e6dee764f4c8937e7350346ad84">Guillermo 这条是今天企业 AI 层最清楚的一句话：把模型变成你自己机器里的一个齿轮。AI SDK、open Agent API、AI Gateway、ZDR inference 这些词背后是同一个判断：企业和创业公司不能把数据、eval、模型选择和软件层都交出去。越到 agent 时代，谁掌握工作流和评估闭环，谁才有长期控制权。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-5db5496d9b4b402abcdb8c233eb5f425"><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=2076364176252191222" 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-062dcf78a8e04c58a8186a5bf80f73b1" data-id="062dcf78a8e04c58a8186a5bf80f73b1"><span><div id="062dcf78a8e04c58a8186a5bf80f73b1" class="notion-header-anchor"></div><a class="notion-hash-link" href="#062dcf78a8e04c58a8186a5bf80f73b1" title="7. Aaron Levie @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">7. Aaron Levie @levie</span></span></h4><div class="notion-text notion-block-89dee26f15a441f5abc03d33cc4db978">Aaron Levie 讨论的是企业 IP 在 AI 时代怎么继续产生价值。他认为，当每家公司都能用上前沿智能时，真正的问题反而更突出：如何把企业自己的决策、洞察、工作流模式和最佳实践嵌入 AI；如何做 workflow eval；如何在不同智能层级之间路由模型；如何捕捉 trace，让信息价值随着 AI 能力一起复利。这是 applied AI layer 的机会所在。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-ad685faf1bb1436bbd99e5c69bd330f0"><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=2076338364635287637" 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-489c04d910f14fc2b82904807c23a2e5" data-id="489c04d910f14fc2b82904807c23a2e5"><span><div id="489c04d910f14fc2b82904807c23a2e5" class="notion-header-anchor"></div><a class="notion-hash-link" href="#489c04d910f14fc2b82904807c23a2e5" title="8. Garry Tan @garrytan"><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. Garry Tan @garrytan</span></span></h4><div class="notion-text notion-block-1eba00a2521f47cfbaf1172c0e6bacaf">Garry Tan 今天这条主要是公共安全技术和政治讨论，不是 AI 主线。按不过滤要求保留：他的观点是，出于文化战和表态目的关闭公共安全技术，会带来真实安全代价。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-154794e72f684e75aaa2b4b0d5eb8950"><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=2076534860064416115" 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-690eace85369454f88a8b6dc6e08101a" data-id="690eace85369454f88a8b6dc6e08101a"><span><div id="690eace85369454f88a8b6dc6e08101a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#690eace85369454f88a8b6dc6e08101a" title="9. Matt Turck @mattturck"><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. Matt Turck @mattturck</span></span></h4><div class="notion-text notion-block-aa794cd87c2145d29004099c454328ac">Matt 今天一条是偏玩笑的体育内容，另一条和 agentic coding 有关：当大家说“任何人现在都能用 agentic coding tool 做应用”时，他用了一个反应图表达怀疑或复杂心情。它不是系统性论证，但反映了一个社区情绪：agentic coding 正在降低门槛，同时也让人担心“能做 demo”和“能做产品”之间的差距被低估。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-5bc4ff312baf4d6bb3c57dd5d39f4ecd"><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=2076343266291626064" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-6b9b0efc8e4c45df8ac758a52c28b1af"><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=2076311766049374598" 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-7ece9255d5944710b8bb8979965bd05c" data-id="7ece9255d5944710b8bb8979965bd05c"><span><div id="7ece9255d5944710b8bb8979965bd05c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7ece9255d5944710b8bb8979965bd05c" title="10. Zara Zhang @zarazhangrui"><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. Zara Zhang @zarazhangrui</span></span></h4><div class="notion-text notion-block-0b05d94b1e974be49e569e7fe7f70372">Zara 的“meeting transcript as PRD”是今天最实用的一条工作流信号：她和同事讨论一个功能实现，把会议记录发给 Codex，然后让它按讨论内容做原型。她总结得很直接：会议就是 prompt。另一条“Passion is the biggest moat”更像个人判断。放在一起看，重点是 Agentic coding 的输入正在从精心写 prompt，变成会议、讨论、上下文这些自然工作材料。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-073894f3638c4b0ea2d51e5920cf6ca8"><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=2076300222884626754" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-43b221bcb8d64f39ad53a03d84c93eaf"><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=2076284012339843546" 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-33a57c0e5d2e4a058c16bc9134e04d40" data-id="33a57c0e5d2e4a058c16bc9134e04d40"><span><div id="33a57c0e5d2e4a058c16bc9134e04d40" class="notion-header-anchor"></div><a class="notion-hash-link" href="#33a57c0e5d2e4a058c16bc9134e04d40" title="11. Nikunj Kothari @nikunj"><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. Nikunj Kothari @nikunj</span></span></h4><div class="notion-text notion-block-d1f8a0fc25614991bab2b9cb0b4df36d">Nikunj 今天三条里，第一条最相关。他说很多人在旧金山会讲自己有很多 subagents 在循环、token 跑得很猛，但当被问到到底在为谁、做什么时，回答反而不清楚。他的提醒很朴素：AI 时代也不能替代方向感，token 多不等于事情重要。后两条分别是提醒出去走走，以及说 outbound sales 这种能力会越来越重要，按全量沉淀保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-f84cc9e22c4f4713aea150ab13b761b0"><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=2076458876816540144" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-6a465eb41f744507bea360778dd91640"><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=2076416145255731677" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-43b936be0e6848c393f3ee319c5e5fd8"><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=2076370608833827124" 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-ae52726451e74ad6a0693d3baa00b3d2" data-id="ae52726451e74ad6a0693d3baa00b3d2"><span><div id="ae52726451e74ad6a0693d3baa00b3d2" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ae52726451e74ad6a0693d3baa00b3d2" title="12. Peter Steinberger @steipete"><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. Peter Steinberger @steipete</span></span></h4><div class="notion-text notion-block-d73563a0a847438eb266bbdc377580f9">Peter 今天三条都和自己的多机器、多会话工作流有关。他说自己大概把工作分散在 5 台机器上，通过 Jump Desktop 操作；还提到 Mac Studio 能承受的 session 数量，以及周末做了一点界面 facelift。它们不是模型发布，但很贴近 AI power user 的真实状态：多 agent、多 session、多设备已经变成一部分人的日常工作方式。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-2675ccbb24cf451abf7fb75d325c8eb6"><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=2076553742883930455" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-48cb34f18ecd48e6b022f9ac7a55b9a4"><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=2076552605262872904" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-8a08503d876f43dca800b5c552430485"><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=2076551622227095828" 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-8c871eda021d4dd2866a39cb3a64d418" data-id="8c871eda021d4dd2866a39cb3a64d418"><span><div id="8c871eda021d4dd2866a39cb3a64d418" class="notion-header-anchor"></div><a class="notion-hash-link" href="#8c871eda021d4dd2866a39cb3a64d418" title="13. Dan Shipper @danshipper"><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. Dan Shipper @danshipper</span></span></h4><div class="notion-text notion-block-cc61b1dc2d644377990610af39467673">Dan 今天三条都比较短，偏转发和轻量评论。一条说某个结果相当可疑，一条说 capitalism stays winning，另一条说 extremely relatable。信息量不高，但作为 follow-builders 的原始抓取内容完整保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-69dcab5c3a03476e865d50b6bf82bee1"><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=2076455432546066826" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-4731a723ff904394b30a8d9b58830db4"><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=2076351869782286707" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-fc1c761d9f924e2cbc78557e573ad3da"><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=2076340879787237562" 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-ce396d5d7626432f850820c8e8f7e2e2" data-id="ce396d5d7626432f850820c8e8f7e2e2"><span><div id="ce396d5d7626432f850820c8e8f7e2e2" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ce396d5d7626432f850820c8e8f7e2e2" title="14. Sam Altman @sama"><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. Sam Altman @sama</span></span></h4><div class="notion-text notion-block-39cb9b5e06e84ffa96f6289bd870a839">Sam Altman 这条是一个很直接的 Sol builder 激励：他想看大家用 GPT-5.6 Sol 做了什么有意思的东西，并会给最酷的作品送一个 OpenAI archives 里的特别礼物。这是产品增长和开发者生态动作，不是技术细节；但它说明 OpenAI 想把 Sol 的讨论从 benchmark 和额度，推向“真实作品展示”。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-85e83ac604184978b6fc58b59a9bbe9d"><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=2076398253332140410" 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-c1098bf5c1cc41e0b3c978ef5e142b06" data-id="c1098bf5c1cc41e0b3c978ef5e142b06"><span><div id="c1098bf5c1cc41e0b3c978ef5e142b06" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c1098bf5c1cc41e0b3c978ef5e142b06" title="15. Claude @claudeai"><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 @claudeai</span></span></h4><div class="notion-text notion-block-567b088ec90448b3878dcd2840d974fc">Claude 官方两条说明 Fable 5 访问继续延期：所有付费计划继续开放到 7 月 19 日，Claude Code 周额度继续保持 50% 上调；同时用户最多可把每周额度的一半用于 Fable 5，之后可以用 credits 或切换到其他模型。这和 OpenAI Sol 的额度调整放在一起看，说明前沿模型可用性已经进入精细化运营阶段。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-2fb6d6774cc14ed18773d43b4468b8a9"><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=2076351401006154204" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-179fd222e8a04159baf75701a439027b"><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=2076351399999557669" 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-7c83ec18d5174c2a943eb3a1fe3ceaca" data-id="7c83ec18d5174c2a943eb3a1fe3ceaca"><span><div id="7c83ec18d5174c2a943eb3a1fe3ceaca" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7c83ec18d5174c2a943eb3a1fe3ceaca" 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-f3847433195c42ae81748f055c5802b8">今天 follow-builders 输出里没有新增 blog。这里不额外补外部网页。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-d0673563e4784833a3b92488a152b0f0" data-id="d0673563e4784833a3b92488a152b0f0"><span><div id="d0673563e4784833a3b92488a152b0f0" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d0673563e4784833a3b92488a152b0f0" 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-ffc6864d91e44dd49ad1fd17c864227f" data-id="ffc6864d91e44dd49ad1fd17c864227f"><span><div id="ffc6864d91e44dd49ad1fd17c864227f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ffc6864d91e44dd49ad1fd17c864227f" title="1. No Priors：How Nuclear Will Unlock Energy Abundance with Valar Atomics Founder Isaiah Taylor"><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：How Nuclear Will Unlock Energy Abundance with Valar Atomics Founder Isaiah Taylor</span></span></h4><div class="notion-text notion-block-3b1e1c8aae5e4ab7bb81f6333c0d27ab">这期播客把 AI compute 的能源问题讲得很硬。Valar Atomics 创始人 Isaiah Taylor 的核心观点是：核能如果要支撑未来 AI 和工业需求，不能继续主要停留在建模、仿真和大型工程审批里，而要走向硬件迭代、制造化、规模化和更便宜的能量供给。他还强调，能源价格下降会创造新需求，这和 AI 算力需求增长是同一条产业链上的问题。节目里提到他们做出了由核反应堆供电的 AI 芯片演示，以及 TRISO 反应堆相关进展。对 AI 从业者来说，这期不是能源科普，而是在提醒：模型和 agent 的上限最终会碰到电力、制造、监管和基础设施。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-4ba49822a0944b66b9422b4c1c7fa327"><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/5Xvbq_zvOQ4" 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-746d5ac1b76c4aa8bed677c842365214" data-id="746d5ac1b76c4aa8bed677c842365214"><span><div id="746d5ac1b76c4aa8bed677c842365214" class="notion-header-anchor"></div><a class="notion-hash-link" href="#746d5ac1b76c4aa8bed677c842365214" 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-0c47f9e73bd54bf0b93bc556111bde1d">今天所有内容放在一起看，AI 的关注点正在从单一模型能力扩展到三层：第一层是模型服务本身，Sol 和 Fable 5 的额度、价格、访问边界会直接影响用户体验；第二层是真实工作流，会议记录、subagents、多会话、企业 eval 和模型路由正在把 agent 拉进日常生产；第三层是底座，算力增长最终要回到能源、硬件和基础设施。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[Sol 模型热度升温，Agent 需求扩散 - 2026-07-12]]></title>
            <link>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-12-sol-agent-demand</link>
            <guid>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-12-sol-agent-demand</guid>
            <pubDate>Sun, 12 Jul 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[本轮 follow-builders 新增 10 个 X builders、15 条 tweet 和 1 期 podcast，已按输出全量沉淀。主线是 GPT-5.6 Sol 在 coding / frontend 场景继续升温，同时 builder 们开始把 AI 放进更大的需求扩张、软件岗位和旅行 agent 场景里看。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-39b5eac3813c8169af3ecd0826b5706d"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-b067ab2d4e9f4e29bbc7c2a959fab647" data-id="b067ab2d4e9f4e29bbc7c2a959fab647"><span><div id="b067ab2d4e9f4e29bbc7c2a959fab647" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b067ab2d4e9f4e29bbc7c2a959fab647" 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-9fe3db98611e4c8c9cdf159641c5fdf7">今天这批内容有三条真正值得抓住的线。第一，GPT-5.6 Sol 继续被推到 coding 和前端场景里，OpenAI 相关成员、Sam Altman 和 builder 都在围绕它的速度、成本、前端质量和接入方式发声。第二，Swyx 和 Aaron Levie 都在讲同一个反直觉判断：AI 让软件生产更便宜后，需求可能不是下降，而是被放大。第三，No Priors 访谈 <a target="_blank" rel="noopener noreferrer" class="notion-link" href="http://Booking.com">Booking.com</a> CEO Glenn Fogel，核心不是“AI 会不会做旅行”，而是旅行 agent 真要落地，必须处理库存、合作伙伴、监管、出错后的连锁反应这些脏活。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-c21bb4eef71c40a685fc0990efd68557" data-id="c21bb4eef71c40a685fc0990efd68557"><span><div id="c21bb4eef71c40a685fc0990efd68557" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c21bb4eef71c40a685fc0990efd68557" 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-32cd84569f464efe85d78dbb291070f3">GPT-5.6 Sol 今天的信号比较集中。Thibault Sottiaux 直接把 Sol 的卖点压到几个很具体的产品问题上：更快、更省 token、后端更强、前端更好，并且少写乱用 <code class="notion-inline-code">useEffect</code> 的代码。他还给出一套把 GPT-5.6 Sol 接进 Claude Code 工作流的方式。Zara Zhang 的反馈更短，但方向一致：她认为 Sol 的前端表现很好。Sam Altman 则从模型地位和使用成本侧面补了一刀，称很多 benchmark 显示 Sol 可能是当前最强模型，也提到 Fable 在某种使用水平下吃掉了 30% 成本。放在一起看，Sol 不是单纯模型宣传，它正在被推成更适合长任务、coding agent 和真实工程工作流的模型。</div><div class="notion-text notion-block-6f1be3f4e38447bf93b0befb119cb837">Swyx 和 Aaron Levie 的两条内容可以放在一起看。Swyx 用 Jevons paradox 解释 agentic engineering：当 coding agent 让知识工作的单位成本下降，总工作需求可能会上升。Aaron Levie 讲得更直白：软件岗位没有按“AI 替代”的剧本消失，反而因为软件生产变便宜，企业想做更多软件项目，仍需要懂系统的人长期维护、决策和运营。这个判断很重要，因为它把“AI 是否替代人”换成了另一个问题：当生产成本下降，哪些工作会被重新打开。</div><div class="notion-text notion-block-f86367c1ccb64154bbaeaf9ad1a67d29">No Priors 这期和 <a target="_blank" rel="noopener noreferrer" class="notion-link" href="http://Booking.com">Booking.com</a> CEO Glenn Fogel 聊 AI 旅行。Fogel 的基本结论是：旅行 agent 很有价值，因为复杂行程本来就痛苦，AI 能记住偏好、比较方案、在航班酒店和行程之间来回推演。但他也反复强调，旅行不是把库存放进数据库就完事。真正难的是合作伙伴关系、全球监管、酒店和物业系统、异常天气或航班变更后的连锁处理。也就是说，AI travel 的壁垒不只在聊天界面，而在能不能处理真实世界的复杂执行。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-7c6fd797f6aa42149e47ca06b7aaec8e" data-id="7c6fd797f6aa42149e47ca06b7aaec8e"><span><div id="7c6fd797f6aa42149e47ca06b7aaec8e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7c6fd797f6aa42149e47ca06b7aaec8e" 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-9e05079b323a42068fed7012779a57d7" data-id="9e05079b323a42068fed7012779a57d7"><span><div id="9e05079b323a42068fed7012779a57d7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9e05079b323a42068fed7012779a57d7" title="1. Swyx @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 @swyx</span></span></h4><div class="notion-text notion-block-ec7a0597805545c996a1278c34b05aa8">Swyx 这条是今天最清楚的“AI 需求扩张”观点。他说如果你只从软件需求理解 Jevons paradox，可能还没完全意识到它对更广义知识工作的影响：coding agent 把单位劳动成本打下来后，真正上升的可能是总工作量和对更好知识工作的需求。他把 coding 发生的事称为一个前兆，而不是例外。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-fd3adc4f791144c7b84797587f825aba"><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=2076155833428431012" 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-00262b521f8c4c6eb9ffb6b0f3c09dd7" data-id="00262b521f8c4c6eb9ffb6b0f3c09dd7"><span><div id="00262b521f8c4c6eb9ffb6b0f3c09dd7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#00262b521f8c4c6eb9ffb6b0f3c09dd7" title="2. Thibault Sottiaux @thsottiaux"><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 @thsottiaux</span></span></h4><div class="notion-text notion-block-7edffcf56866435fa43e56cb483523b0">Thibault 今天两条都围绕 GPT-5.6 Sol。第一条直接回应用户对速度、代码质量、前端能力的抱怨，强调 Sol 更快、更省 token、后端强、前端好，也少出现一些常见 React 代码坏味道。第二条给出一个把 GPT-5.6 Sol 接进 Claude Code 风格工作流的 CLIProxyAPI / alias 玩法，重点是让不装 Codex app 的人也能试到 Sol。这两条合起来说明，Sol 正在被当作 coding agent 模型来推，而不是只放在聊天窗口里展示。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-b85dbc8a159b4b3dbaf8a7929f80f1ae"><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=2076145711922696371" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-b63ac5cbf5e94a7b8daa745656a5311a"><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=2076119366647894371" 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-a1788997325a4f4b94f66bf91da44316" data-id="a1788997325a4f4b94f66bf91da44316"><span><div id="a1788997325a4f4b94f66bf91da44316" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a1788997325a4f4b94f66bf91da44316" title="3. Peter Yang @petergyang"><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 @petergyang</span></span></h4><div class="notion-text notion-block-f09e9c4699174e33af5ad59c588c7388">Peter 今天三条都是阿根廷比赛相关，不是 AI 主线。按全量沉淀要求保留在这里，主要价值是说明这批 follow-builders 输出里混有明显生活化/体育内容，不适合作为短视频主题。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-7d95332f11484a83a2a6ffc0ee3d4dad"><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=2076151105080447404" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-7149f8f2c1984a0a9c0c83fe5154986f"><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=2076143769192484900" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-380d64315c6844afba9111b702b38366"><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=2076132077054095751" 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-a56d4bf8e8bf4041af499c67a2d763c9" data-id="a56d4bf8e8bf4041af499c67a2d763c9"><span><div id="a56d4bf8e8bf4041af499c67a2d763c9" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a56d4bf8e8bf4041af499c67a2d763c9" title="4. Guillermo Rauch @rauchg"><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. Guillermo Rauch @rauchg</span></span></h4><div class="notion-text notion-block-5e26332e6bbf478db2e449c8e64c459a">Guillermo 这条是很短的“Easy money (Swiss francs)”式评论，信息量不高，也不是 AI 主线。按原始抓取结果保留。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-b1b55432e5fb4c3d9ab89068310d0e9b"><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=2076151956440261008" 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-b4a972da3ca64ea5b476a4271c1f50ba" data-id="b4a972da3ca64ea5b476a4271c1f50ba"><span><div id="b4a972da3ca64ea5b476a4271c1f50ba" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b4a972da3ca64ea5b476a4271c1f50ba" title="5. Aaron Levie @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">5. Aaron Levie @levie</span></span></h4><div class="notion-text notion-block-65050faa4bea46589b747509636ecab2">Aaron Levie 讲的是 AI 与软件岗位的反直觉关系。他认为软件岗位没有按“被 AI 替代”的方向走，是因为软件生产成本下降后，人们反而想生产更多软件。企业会启动更多项目，但只要工作没有完全自动化，就仍需要人来决定做什么、维护系统、长期运营和更新。他把这个逻辑推广到其他知识工作：agent 带来的可能是更多供给和需求，而不是简单替代。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-1865cfe8287a4061a4b256c4e87bc349"><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=2076116544980214164" 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-31d6a52cf84a49a5957e8ae711db9a84" data-id="31d6a52cf84a49a5957e8ae711db9a84"><span><div id="31d6a52cf84a49a5957e8ae711db9a84" class="notion-header-anchor"></div><a class="notion-hash-link" href="#31d6a52cf84a49a5957e8ae711db9a84" title="6. Garry Tan @garrytan"><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. Garry Tan @garrytan</span></span></h4><div class="notion-text notion-block-d03b64827b204ab1b0252ecdff6d6a2a">Garry Tan 今天两条主要是加州建设和创业心态，不是 AI 主线。第一条谈 CEQA 改革和住房建设，第二条是偏 founder 心态的表达：很多事可能出错，但更有意思的问题是如果事情做成会怎样。按全量要求保留。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-434c28538e6d4740bc190223eea1d812"><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=2075944103867830352" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-0c7e8391d19a45ce853cf8c2f3a12993"><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=2075933358660730901" 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-a9806df6b64f4ef99ad30b548fee67ae" data-id="a9806df6b64f4ef99ad30b548fee67ae"><span><div id="a9806df6b64f4ef99ad30b548fee67ae" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a9806df6b64f4ef99ad30b548fee67ae" title="7. Matt Turck @mattturck"><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. Matt Turck @mattturck</span></span></h4><div class="notion-text notion-block-a1cbe2565c344601a475578bd9519795">Matt Turck 这条也是阿根廷比赛评论，不是 AI 主线。由于他同时也是 MAD Podcast 主理人，账号本身常和 AI / data 内容相关，但这条具体内容没有可沉淀的 AI 信息。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-d6ce91052afe402eb43a61eb3cabf7d7"><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=2076141940484010189" 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-847b182265a6436cbb6b27ccd2c1062c" data-id="847b182265a6436cbb6b27ccd2c1062c"><span><div id="847b182265a6436cbb6b27ccd2c1062c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#847b182265a6436cbb6b27ccd2c1062c" title="8. Zara Zhang @zarazhangrui"><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. Zara Zhang @zarazhangrui</span></span></h4><div class="notion-text notion-block-23aff5bc6bce4e80b52951ac0a44e1af">Zara 的反馈很短，但和今天 Sol 主线对上了：她认为 5.6 Sol 的前端能力很好。这条价值在于它是 builder 使用侧的即时反馈，不是官方 benchmark。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-f2d69c2b4cdf46c09d85afaabc2c88c0"><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=2076130810143367453" 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-9e3787f90a1741748d63c4ea138d254c" data-id="9e3787f90a1741748d63c4ea138d254c"><span><div id="9e3787f90a1741748d63c4ea138d254c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9e3787f90a1741748d63c4ea138d254c" title="9. Nikunj Kothari @nikunj"><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. Nikunj Kothari @nikunj</span></span></h4><div class="notion-text notion-block-5373a54ef4c54e4c8596a4461bae82d4">Nikunj 这条也是足球相关，不是 AI 主线。按全量抓取内容保留。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-89108a1ae81f49389c6f52b5ed7f8a70"><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=2076150076087611433" 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-a087aad0b0d24269b9bc8d5f1a55c7ee" data-id="a087aad0b0d24269b9bc8d5f1a55c7ee"><span><div id="a087aad0b0d24269b9bc8d5f1a55c7ee" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a087aad0b0d24269b9bc8d5f1a55c7ee" title="10. Sam Altman @sama"><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. Sam Altman @sama</span></span></h4><div class="notion-text notion-block-37fa42712a074544b00da1358aaeef71">Sam 今天两条都和模型竞争有关。第一条说很多 benchmark 显示 5.6 Sol 可能是现在世界上最好的模型，但他用了比较调侃的方式表达。第二条提到在某种使用水平下，Fable 吃掉了 30% 成本。两条合起来看，一个在讲模型位置，一个在暗示成本结构仍是前沿模型部署绕不开的问题。原帖如下：</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-b3481d46a36a4e409834fafb7974e852"><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=2075983427019612242" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-7cd2b74053a14fa1a43d44e1430d428d"><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=2075982820322025788" 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-d15e2cbd209f4cd68cb18ee5626bc54c" data-id="d15e2cbd209f4cd68cb18ee5626bc54c"><span><div id="d15e2cbd209f4cd68cb18ee5626bc54c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d15e2cbd209f4cd68cb18ee5626bc54c" 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-294b715346b24036807b19e52b430625">今天 follow-builders 输出里没有 blog 新内容。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-94de460357364056b0783269c2e73783" data-id="94de460357364056b0783269c2e73783"><span><div id="94de460357364056b0783269c2e73783" class="notion-header-anchor"></div><a class="notion-hash-link" href="#94de460357364056b0783269c2e73783" 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-cb296fd71f1346c3bc71aea721b18998" data-id="cb296fd71f1346c3bc71aea721b18998"><span><div id="cb296fd71f1346c3bc71aea721b18998" class="notion-header-anchor"></div><a class="notion-hash-link" href="#cb296fd71f1346c3bc71aea721b18998" title="1. No Priors：Travel Through the Lens of AI with Booking.com CEO Glenn Fogel"><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：Travel Through the Lens of AI with <a target="_blank" rel="noopener noreferrer" class="notion-link" href="http://Booking.com">Booking.com</a> CEO Glenn Fogel</span></span></h4><div class="notion-text notion-block-8b19e054668f4309bfade11854a7d739">这期访谈最值得记住的是 Glenn Fogel 对 AI travel 的边界判断。他承认 personalized travel agent 是非常自然的方向：复杂家庭行程、里程和现金选择、多目的地、酒店、餐厅、突发变更，这些都适合交给 AI 先做大量推演。但他也强调，旅行行业不是一个“模型接个库存库”就能打穿的领域。真正难的是全球监管、酒店和物业方协作、供应关系、异常情况处理，以及用户最终确认权。换句话说，AI 会改造旅行入口，但真正的护城河会在执行层。节目来源：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.youtube.com/watch?v=8nj_0wZkbtA">No Priors / YouTube</a></div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-ae62ab2af01344248f6272ff88a02800"><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/8nj_0wZkbtA" 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-4ea0f355cf8f417b84842ab3b4baf930" data-id="4ea0f355cf8f417b84842ab3b4baf930"><span><div id="4ea0f355cf8f417b84842ab3b4baf930" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4ea0f355cf8f417b84842ab3b4baf930" 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-bd44afb7ee9e409ebb84247d48f1b1b4">今天所有内容放在一起看，主线不是一个单点发布，而是 AI 正在进入三个更真实的层面：模型层，GPT-5.6 Sol 被推向 coding 和 frontend 工作流；工作层，agentic engineering 可能让软件需求继续变大；应用层，旅行 agent 这种看起来简单的场景，真正难点在执行、关系和责任边界。非 AI 的体育、政策和生活内容也已按要求完整保留。</div><div class="notion-text notion-block-9b19e225ec1245ad83b5cae6367c06ad">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[MagenticLite 深度分析 - 2026-07-12]]></title>
            <link>http://easyai.fyi/article/magenticlite-deep-analysis-2026-07-12</link>
            <guid>http://easyai.fyi/article/magenticlite-deep-analysis-2026-07-12</guid>
            <pubDate>Sun, 12 Jul 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Microsoft 把浏览器 agent 拆成小型编排模型、专用 computer-use 模型和受控 harness，重点不在更大模型，而在执行环境、上下文管理和人在环。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-39b5eac3813c81459546f166b0f18b1a"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-4e0f08c9379f49809774b31ee0d7e17b" data-id="4e0f08c9379f49809774b31ee0d7e17b"><span><div id="4e0f08c9379f49809774b31ee0d7e17b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4e0f08c9379f49809774b31ee0d7e17b" 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-42fa75b8dfe744d3aa800dc6e30f4050"><li>GitHub 仓库：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/microsoft/magentic-ui">microsoft/magentic-ui</a></li></ul><ul class="notion-list notion-list-disc notion-block-5493628dec5b427ab59da6b79d166887"><li>Microsoft Research 发布文章：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.microsoft.com/en-us/research/blog/magenticlite-magenticbrain-fara1-5-an-agentic-experience-optimized-for-small-models/">MagenticLite, MagenticBrain, Fara1.5: An agentic experience optimized for small models</a></li></ul><ul class="notion-list notion-list-disc notion-block-a2c8733c030c4866945540063a61a7dc"><li>Fara1.5 技术文章：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/">Fara1.5 - A family of frontier computer use agent models</a></li></ul><ul class="notion-list notion-list-disc notion-block-f076c71165914d7780a02d80b7991f59"><li>透明说明：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/microsoft/magentic-ui/blob/main/docs/TRANSPARENCY_NOTE.md">MagenticLite Transparency Note</a></li></ul><ul class="notion-list notion-list-disc notion-block-4b067206e9384542876b7257ce42ef6e"><li>相关论文：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/abs/2507.22358">Magentic-UI: Towards Human-in-the-loop Agentic Systems</a></li></ul><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-74713abf632048cb84970fa87c9c65be" data-id="74713abf632048cb84970fa87c9c65be"><span><div id="74713abf632048cb84970fa87c9c65be" class="notion-header-anchor"></div><a class="notion-hash-link" href="#74713abf632048cb84970fa87c9c65be" 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-0b0bc9f3ff444ceeb4138bfdb4da8810">今天的候选里，有几类东西都值得看：AI agent harness 的资料库、agent memory 综述、近期 human-in-the-loop 论文，还有 MagenticLite。最后选 MagenticLite，原因很简单：它不是再讲“agent 应该会规划、会用工具”这种老话，而是把一个能跑的 browser/file-system agent 拆到模型、harness、沙箱、人机交互、评测几层一起做。</div><div class="notion-text notion-block-0074a478e538461ab388a6c81ad8d745">这也避开了《智汇AI》最近几篇的重复。前几天已经写过 Browser Harness、Loop Engineering、Agents&#x27; Last Exam、Hermes Agent。MagenticLite 仍然是 harness 主题，但它的核心问题换了：如果不用最大模型，能不能靠更合适的执行系统，让小模型完成真实浏览器任务？</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-4c654d11286b46ca92b17d46179f1de0"><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://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/magentic_releases.png?t=4c654d11-286b-46ca-92b1-7d46179f1de0" alt="图 1：MagenticLite、MagenticBrain、Fara1.5 三个组件来自同一次 Microsoft Research 发布。来源：Microsoft Research" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图 1：MagenticLite、MagenticBrain、Fara1.5 三个组件来自同一次 Microsoft Research 发布。来源：Microsoft Research</figcaption></div></figure><div class="notion-text notion-block-eea8ebdb1e924831964d0e40fffcaa69">图 1 很直接：MagenticLite 是应用和执行体验；MagenticBrain 是负责推理、写代码、调度的 14B 编排模型；Fara1.5 是专门做浏览器操作的 computer-use model。这个拆法比“一个大模型加一堆工具”更像真实产品架构。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-98a993577ca84ac295742fbfab15a624" data-id="98a993577ca84ac295742fbfab15a624"><span><div id="98a993577ca84ac295742fbfab15a624" class="notion-header-anchor"></div><a class="notion-hash-link" href="#98a993577ca84ac295742fbfab15a624" title="它解决的不是“会不会点网页”，而是“能不能受控地做事”"><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">它解决的不是“会不会点网页”，而是“能不能受控地做事”</span></span></h3><div class="notion-text notion-block-f1b3e46ac915479785483fe9b4b751bc">MagenticLite 面向的是浏览器和本地文件系统之间的连续任务，比如查资料、填表、预约、整理本地文件。GitHub README 里写得很明确：它可以跨 browser 和 file system 工作，也会在关键动作前停下来让人确认，并把浏览器会话放在 Quicksand 沙箱里，避免 agent 直接碰到主机环境。</div><div class="notion-text notion-block-aae07a6609084e638eafb24856a6a3ad">这点值得重视。很多 computer-use agent 的演示看起来像“模型能操作鼠标键盘”，但产品真正卡住的地方是三件事：</div><ul class="notion-list notion-list-disc notion-block-b731e47f8941456db153ff616a0c211f"><li>任务会跑很久，上下文会变脏。</li></ul><ul class="notion-list notion-list-disc notion-block-f6d0944724e642918240e3d1e0c92313"><li>网页里有登录、付款、提交表单这类不可逆动作。</li></ul><ul class="notion-list notion-list-disc notion-block-b546f197626546e18ce1f393dde31fba"><li>模型错一步，用户很难知道它为什么错、错在哪里。</li></ul><div class="notion-text notion-block-b597186a78f7472fafc4f9ec88b90023">MagenticLite 的答案不是让模型更“聪明”，而是给模型一个更窄、更稳的工作台：编排模型只做规划、工具选择、代码和委派；浏览器模型只做视觉观察和单步动作；harness 负责压缩上下文、暂停恢复、关键点审批和沙箱隔离。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-5c76b1d5f3f3480dae9b9e38c452aaf5" data-id="5c76b1d5f3f3480dae9b9e38c452aaf5"><span><div id="5c76b1d5f3f3480dae9b9e38c452aaf5" class="notion-header-anchor"></div><a class="notion-hash-link" href="#5c76b1d5f3f3480dae9b9e38c452aaf5" title="核心循环：Fara1.5 只做一步，循环很多次"><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">核心循环：Fara1.5 只做一步，循环很多次</span></span></h3><div class="notion-text notion-block-f7edece262514c528c7c0a2f758a3543">Fara1.5 的 loop 是 observe-think-act。每一步，它看最近三张浏览器截图和对话历史，然后输出一次思考和一个原子动作。动作包括鼠标键盘、网页动作，也包括记住事实、询问用户、结束任务这类 context management 动作。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-4b8f6876dc3e431ba252e6f160b5c0bb"><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://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/fig2_fara_agentic_loop.png?t=4b8f6876-dc3e-431b-a252-e6f160b5c0bb" alt="图 2：Fara1.5 的 observe-think-act loop。来源：Microsoft Research Fara1.5 技术文章" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图 2：Fara1.5 的 observe-think-act loop。来源：Microsoft Research Fara1.5 技术文章</figcaption></div></figure><div class="notion-text notion-block-16ccee0476354dd58137fe939cefbcc6">这张图说明了一个常被忽略的问题：browser agent 不应该一次性“规划完整网页操作”。网页状态一直在变，弹窗、表单校验、登录页、加载延迟都会打断计划。Fara1.5 把动作压到单步，harness 再反复喂回新的截图和历史。它牺牲了一点速度，换来更强的可纠错性。</div><div class="notion-text notion-block-31ef7eb247c7468f8b68eaa0ffdd22cf">这里的新东西不是 ReAct 本身。observe-think-act 很像 ReAct 的 GUI 版本。真正有价值的是三点：</div><ul class="notion-list notion-list-disc notion-block-3790a9a850504f6fa6beb1de375667db"><li>输入不是 DOM 树，而是截图和对话历史，更接近人类看网页的方式。</li></ul><ul class="notion-list notion-list-disc notion-block-234c42c04aa54946bed7abd2fd6f32f6"><li>动作空间里放进了 memorise、ask_user、finish 这类元动作，说明“上下文管理”和“请求人介入”已经变成 agent loop 的一部分。</li></ul><ul class="notion-list notion-list-disc notion-block-f6a71c905b2f48229016c33f99a71fbb"><li>每步只执行一个动作，减少了长链操作里一次性犯大错的机会。</li></ul><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-3bf2ab4ce63847929955556d200bbd7a" data-id="3bf2ab4ce63847929955556d200bbd7a"><span><div id="3bf2ab4ce63847929955556d200bbd7a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3bf2ab4ce63847929955556d200bbd7a" title="MagenticBrain：小模型能编排，前提是训练时就待在同一个 harness 里"><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">MagenticBrain：小模型能编排，前提是训练时就待在同一个 harness 里</span></span></h3><div class="notion-text notion-block-43af8ad47e01448a861f3b70923b5ad1">Microsoft Research 说 MagenticBrain 是 14B 参数的 orchestration model，负责把自然语言请求拆成计划、选择工具、写代码、在需要时把浏览器任务委派给 Fara1.5。关键点不是“14B 也能推理”，而是它在 MagenticLite harness 里端到端训练，训练时见到的工具 schema 和执行环境就是推理时用的那套。</div><div class="notion-text notion-block-12758e71f478402e812313f6c1324631">这和常见 agent 设计差别很大。很多团队先拿通用模型，再在外面套工具调用、prompt、路由和重试。这样能快，但模型其实没有学过你的真实运行环境，只是在临场猜格式。MagenticBrain 的思路更像把“harness 行为”也纳入训练分布：模型不是只学知识，而是学怎么在这个系统里做事。</div><div class="notion-text notion-block-09d1122cd25842cca8065b3b1803918b">我认为这是 MagenticLite 最值得沉淀的地方。小模型路线如果只比参数和榜单，很容易变成缩水版大模型。MagenticLite 的路径是：缩小模型负责的范围，把环境、工具、委派关系和上下文整理做得更明确。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-4c41dcfcff6548cab66913c56b6ce9a1" data-id="4c41dcfcff6548cab66913c56b6ce9a1"><span><div id="4c41dcfcff6548cab66913c56b6ce9a1" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4c41dcfcff6548cab66913c56b6ce9a1" title="训练数据：用合成环境补上真实网页不能训练的部分"><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">训练数据：用合成环境补上真实网页不能训练的部分</span></span></h3><div class="notion-text notion-block-41538acf8b7e42148afce85456cbadbf">Fara1.5 没有只靠公开网页轨迹。公开网页能训练搜索和比较，但很难训练登录、发邮件、提交表单、订票这类动作，因为这些任务要账号、可能不可逆，也可能触碰真实用户数据。Microsoft 用 FaraGen1.5 做合成环境，把邮件、日历、媒体平台、ML 实验、市场等场景放进沙箱，再用 solver 和 verifier 过滤轨迹。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-c67f3140fbe34fd4a824be129168913f"><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://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/fig6_faragen15_v2.png?t=c67f3140-fbe3-4fd4-a824-be129168913f" alt="图 3：FaraGen1.5 用环境、solver、verifier 生成并筛选 computer-use 轨迹。来源：Microsoft Research Fara1.5 技术文章" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图 3：FaraGen1.5 用环境、solver、verifier 生成并筛选 computer-use 轨迹。来源：Microsoft Research Fara1.5 技术文章</figcaption></div></figure><div class="notion-text notion-block-329efd85930546aa8507663d23164708">这张图里最重要的是 verifier。它不是只看最后有没有成功，还检查三件事：轨迹是否符合任务意图、是否有多余或绕路动作、遇到缺信息/任务不明确/不可逆动作时是否正确请求用户。换句话说，它把“好 agent”从结果正确扩展到过程可控。</div><div class="notion-text notion-block-d5a36f9f3dac4e08b68a8c5436c497fc">这对产品落地很有启发。很多 agent 评测只看 success rate，最后会鼓励模型赌一把。真实产品不能这样。一个会停下来问用户的 agent，可能比一个盲目提交表单但偶尔成功的 agent 更可用。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-e025abd6acc142348dfccf14eca52b68" data-id="e025abd6acc142348dfccf14eca52b68"><span><div id="e025abd6acc142348dfccf14eca52b68" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e025abd6acc142348dfccf14eca52b68" title="效果数据：小模型进步明显，但别只看榜单"><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">效果数据：小模型进步明显，但别只看榜单</span></span></h3><div class="notion-text notion-block-30f4b2df879e4d91967bedd4180e9543">Fara1.5 的数据不错。Microsoft Research 报告称，Fara1.5-9B 在 Online-Mind2Web 上达到 63.4%，相比 Fara-7B 的 34.1% 提升很大；在 WebVoyager 上从 73.5% 提到 86.6%。Fara1.5-27B 在 Online-Mind2Web 上达到 72.0%，WebVoyager 为 88.6%。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-078960e293b0492483a09c7e86cf4590"><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://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/fig7_fara7b_vs_fara15-9b.png?t=078960e2-93b0-4924-83a0-9c7e86cf4590" alt="图 4：Fara1.5-9B 相比 Fara-7B 在多个 benchmark 上提升。来源：Microsoft Research Fara1.5 技术文章" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图 4：Fara1.5-9B 相比 Fara-7B 在多个 benchmark 上提升。来源：Microsoft Research Fara1.5 技术文章</figcaption></div></figure><div class="notion-text notion-block-3928c2f427f84c9a9ea34b41cc31e03c">但这些数字要谨慎读。WebVoyager、Online-Mind2Web 这类 benchmark 能说明模型在网页任务上变强了，却不能完整代表真实用户任务。真实任务里会有登录状态、支付风险、公司权限、隐私数据、网页改版、慢加载、弹窗、反爬和多标签页状态。Microsoft 自己的透明说明也提醒：MagenticLite 是研究原型，不建议在没有进一步测试的情况下用于商业或真实高风险场景。</div><div class="notion-text notion-block-822ccc93ad8b4f1290916195a59abb70">所以这次发布的价值不在于“9B 打赢谁”，而在于它给了一个比较完整的工程假设：小模型 agent 的上限，不只由模型参数决定，还由 action space、训练轨迹、harness、上下文压缩、用户审批和沙箱共同决定。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-d66e72123cdd4f1eb44807ba13f5051b" data-id="d66e72123cdd4f1eb44807ba13f5051b"><span><div id="d66e72123cdd4f1eb44807ba13f5051b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d66e72123cdd4f1eb44807ba13f5051b" title="和常见 agent 设计相比，新在哪里"><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">和常见 agent 设计相比，新在哪里</span></span></h3><div class="notion-text notion-block-18499445b5014349add993dbc39ca588">常见 agent 设计大致是：一个大模型，接一组工具，外面套 planning prompt、memory、reflection、retry，再加一点日志。MagenticLite 不是完全推翻这个模式，但把几个薄弱点做实了。</div><div class="notion-text notion-block-8c31cd0e0e5946b2b9830eccbc0f6e72">第一，它把编排和浏览器操作拆开。MagenticBrain 负责计划、代码、工具和委派；Fara1.5 负责视觉网页动作。这样做的好处是每个模型训练目标更窄，harness 也更容易判断什么时候该委派。</div><div class="notion-text notion-block-2775497ccb0b46ddb68ec971eae7fb76">第二，它把人在环放进系统协议里，而不是失败后的补丁。Magentic-UI 论文里提到 co-planning、co-tasking、action approval、answer verification、memory、multi-tasking。这里的人不是旁观者，而是 multi-agent team 里的一个特殊角色。agent 可以把某些步骤委派给用户，用户也可以暂停、接管浏览器、修改计划、恢复执行。</div><div class="notion-text notion-block-2802c3d6072849bdb289a5cc7a8f5eab">第三，它把 memory 收窄为“可复用计划”。这不是最通用的 agent memory，但很实用。用户完成过一次任务后，可以把执行轨迹总结成计划，下次相似任务直接复用。它没有试图一口吃掉个人偏好、事实记忆、长期关系和世界知识，而是先解决重复任务的 workflow memory。</div><div class="notion-text notion-block-61c1aca31ded477fb38f12efb5b4e442">第四，安全不是一句“有 guardrail”。GitHub README 和透明说明都强调关键动作前暂停、人类监督、Quicksand 沙箱、研究用途限制。Fara1.5 的训练和评测里也把 critical points 单独处理，包括缺少用户信息、任务不明确、不可逆动作未获批准。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-43c92822385d43189d680a38d762db1f" data-id="43c92822385d43189d680a38d762db1f"><span><div id="43c92822385d43189d680a38d762db1f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#43c92822385d43189d680a38d762db1f" title="哪些地方只是工程包装"><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">哪些地方只是工程包装</span></span></h3><div class="notion-text notion-block-531c000b0c53478092f8e41f07d775d9">MagenticLite 也有不少熟悉的东西。observe-think-act 不是新范式，multi-agent orchestration 也不是新概念，plan editor 和审批流程在企业自动化产品里早就有类似形态。它的“新”不在单个模块，而在组合方式。</div><div class="notion-text notion-block-8d0ed1987f684e9b979f5a59e707fab0">如果把它拆开看，每个模块都能找到前身：ReAct、AutoGen/Magentic-One、browser-use agent、workflow memory、human approval、sandbox。真正值得学的是它把这些模块放在同一个任务闭环里，并且用训练数据和执行环境对齐它们。也就是说，它不是发明了所有零件，而是把零件之间的接口做得更像产品。</div><div class="notion-text notion-block-1e108791990147ad9a173571231b403e">这也意味着复刻 MagenticLite 不能只抄 UI。只做一个漂亮的浏览器控制台，没有专门动作空间、上下文压缩、关键点识别、轨迹验证和沙箱，最后还是会退回“模型在网页上裸奔”。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-4063195ba8594a92bf85ef1b3a4d5a39" data-id="4063195ba8594a92bf85ef1b3a4d5a39"><span><div id="4063195ba8594a92bf85ef1b3a4d5a39" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4063195ba8594a92bf85ef1b3a4d5a39" title="可复用的方法论"><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">可复用的方法论</span></span></h3><div class="notion-text notion-block-8510e6d46e3a46a5af5c9f8494874346">第一，先定义 agent 的受控动作空间，再谈模型能力。好的 agent 不是想做什么都能做，而是每一步能被观察、回放、暂停、验证。</div><div class="notion-text notion-block-667a72e59802459d9393dfd0776b0675">第二，把用户介入设计成正常路径。不要等失败才问用户。计划阶段让用户修正方向，执行阶段允许接管，关键动作前必须确认，结束后提供可验证轨迹。</div><div class="notion-text notion-block-edc55328edef4294a264560c023233e5">第三，小模型要配窄职责。一个 9B browser model 可以很强，但前提是它只需要看截图、思考一步、执行一步。编排、代码、文件、浏览器不要全压给一个模型。</div><div class="notion-text notion-block-cdd100c924da49538b0112b68b9d6f48">第四，memory 不必一开始就做成“人格记忆”。从可复用计划开始更稳。重复任务、固定流程、可检查结果，才是 agent memory 最容易落地的地方。</div><div class="notion-text notion-block-68e0fe5d2b8f44fe8275ab067ccc936c">第五，评测要同时看结果和过程。success rate 很重要，但真实 agent 还要看有没有绕路、有没有该问不问、有没有误触不可逆动作、有没有把用户带到无法理解的状态。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-5303d582c7d74a608571e0fc87ce9f5f" data-id="5303d582c7d74a608571e0fc87ce9f5f"><span><div id="5303d582c7d74a608571e0fc87ce9f5f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#5303d582c7d74a608571e0fc87ce9f5f" title="风险和没验证清楚的问题"><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">风险和没验证清楚的问题</span></span></h3><div class="notion-text notion-block-68774524df9b4b71a62f9032c8bd779e">MagenticLite 仍然是研究原型。透明说明里明确写到，它不建议在没有进一步测试的情况下用于商业或真实场景，也不适合高风险决策。它主要用英语设计和测试，其他语言表现需要单独评估。</div><div class="notion-text notion-block-0323571cc51c4da296351c102b6ef84e">它还有几个问题没完全回答：</div><ul class="notion-list notion-list-disc notion-block-f0cad9ad49d44405b47d76e1b485fc7d"><li>小模型在长任务里靠上下文压缩能撑多久？压缩错了，后面的行动会被污染。</li></ul><ul class="notion-list notion-list-disc notion-block-7ae6b97e97454c738e41eda619c51649"><li>合成环境训练出来的能力，迁移到真实企业系统时会掉多少？企业网页通常权限复杂、状态隐蔽、流程定制重。</li></ul><ul class="notion-list notion-list-disc notion-block-4197c867a89946009970cefb4e3d520b"><li>人在环会不会变成频繁打断？审批太少不安全，审批太多又没人愿意用。</li></ul><ul class="notion-list notion-list-disc notion-block-50c38b06687b4f73946c65d34f98e148"><li>关键点识别能不能覆盖 prompt injection、伪造按钮、恶意网页文案这类对抗场景？</li></ul><ul class="notion-list notion-list-disc notion-block-3492fd9d755d4b5a84e803ea4a941505"><li>计划记忆会不会复用过头？网页变了、政策变了、用户偏好变了，旧计划可能变成错误捷径。</li></ul><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-ef498691c276434f9591232b75f97adb" data-id="ef498691c276434f9591232b75f97adb"><span><div id="ef498691c276434f9591232b75f97adb" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ef498691c276434f9591232b75f97adb" 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><ul class="notion-list notion-list-disc notion-block-88564d62250f4572a8784cd0f02fd2bd"><li>小模型 agent 的关键不是“缩水”，而是缩小职责边界。</li></ul><ul class="notion-list notion-list-disc notion-block-718deb8938984557ba8a2b4dceff1ca1"><li>Harness 是 agent 产品的运行时，不是 prompt 外壳。</li></ul><ul class="notion-list notion-list-disc notion-block-e7a022efb821487f871c4de4cdf9f96c"><li>人在环要变成协议，不能只当兜底客服。</li></ul><ul class="notion-list notion-list-disc notion-block-325b1f7d5ba947249dd7627298a3c3e5"><li>Browser agent 的 memory，先从可复用计划做起。</li></ul><ul class="notion-list notion-list-disc notion-block-a6c1429384204b78a891bd35e63ad7f9"><li>评测 agent，要同时评结果、过程和该停不停。</li></ul></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agent 支付基础设施成形 - 2026-07-11]]></title>
            <link>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-11-agent-payments</link>
            <guid>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-11-agent-payments</guid>
            <pubDate>Sat, 11 Jul 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[本轮 follow-builders 新增 1 期 podcast，X builders 和 blog 没有可用新内容，已按输出全量沉淀。主线是 Stripe 对 agentic commerce 的判断：Agent 正在从帮人找商品，走向能授权、下单、支付和参与交易。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-39a5eac3813c810cb6cbde368828fcda"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-15f9327bd6ef4808ba8357319e62ce6b" data-id="15f9327bd6ef4808ba8357319e62ce6b"><span><div id="15f9327bd6ef4808ba8357319e62ce6b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#15f9327bd6ef4808ba8357319e62ce6b" 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-9d9aa00fe89048dbb065c8fa1dc336ba">这轮 follow-builders 的新内容集中在一件事：AI agent 正在从“帮你找答案”走向“帮你完成交易”。The MAD Podcast 这期和 Stripe 的 Head of Data and AI Emily Sense 聊 agentic commerce，她的判断很直接：一年前 agent 买东西还偏假设，现在已经有基础设施、商家和平台在真实接入。重点不只是让 agent 自动下单，而是让商品目录、价格、授权、支付凭证、风控和商家结算都能被 AI 表面调用。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-4a6aec298b904450b5eec137032b4693" data-id="4a6aec298b904450b5eec137032b4693"><span><div id="4a6aec298b904450b5eec137032b4693" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4a6aec298b904450b5eec137032b4693" 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-5d7ebdadd8ed412bb9292d4f984cc561">Stripe 的判断是，agentic commerce 不是单一路径，而是一条光谱。一端是完全自主的 agent：自己发现服务、判断是否购买、完成交易；另一端是人在 AI 应用里找商品，AI 给出答案，并把购买按钮放进流程里。Google 的 AI Mode 和 Gemini、Copilot、ChatGPT、Meta 广告里的 checkout，都是这种变化的早期形态。</div><div class="notion-text notion-block-3d19e649f39544a0a45fe05f97bf127b">真正难的地方不是“让模型推荐商品”，而是交易基础设施。商家要能暴露产品目录、库存和价格；消费者要能授权 agent 代为支付；agent 要能安全地执行交易。Emily Sense 提到 Stripe 的 Agentic Commerce Suite、catalog 暴露、shared payment token、Radar 风控和 Link wallet guardrails，核心都是为了解决一个问题：让 agent 能花钱，但不能乱花钱，也不能拿到不该拿的底层凭证。</div><div class="notion-text notion-block-1008e70a18744d578df6ac3a93880251">这期里最值得警惕的信号是 token theft。Emily Sense 说，在 AI 领域，攻击者不一定要偷钱或账号，偷 token 就能获得价值；她提到 AI 公司里超过六分之一的新注册存在这类滥用。这个判断很重要，因为 AI 应用的支付、试用、额度和 API 调用，本质上都会变成新的风控对象。以后企业做 AI 产品，不只是防盗刷，还要防“薅 token”。</div><div class="notion-text notion-block-7a858cf04c95481180d706306bc75f80">另一个长期信号是商业模式会变。传统 SaaS 的 per-seat pricing 在 agent 场景下会变得别扭，因为 agent 不一定按人头工作。它可能替一个团队跑任务、替一个商家处理采购，甚至被赋予买东西、卖东西、赚利润的目标。也就是说，Agent 不只是用户界面里的助手，正在变成新的经济行为主体。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-8f5a8450a521466ea882e1511a61b0e3" data-id="8f5a8450a521466ea882e1511a61b0e3"><span><div id="8f5a8450a521466ea882e1511a61b0e3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#8f5a8450a521466ea882e1511a61b0e3" 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><div class="notion-text notion-block-24beceb8a2ac41e68361f70263eb206f">本轮 follow-builders 输出中没有可用 X builder/tweet 内容，因此没有 X embed 可写入。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-b34b682f73094372872c26b73069191f" data-id="b34b682f73094372872c26b73069191f"><span><div id="b34b682f73094372872c26b73069191f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b34b682f73094372872c26b73069191f" 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-a4ce80277a6c45b29443db09a104ce8f">本轮 follow-builders 输出中没有 blog 更新。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-48a54cb132a34964b3f47e1274ff3997" data-id="48a54cb132a34964b3f47e1274ff3997"><span><div id="48a54cb132a34964b3f47e1274ff3997" class="notion-header-anchor"></div><a class="notion-hash-link" href="#48a54cb132a34964b3f47e1274ff3997" 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-653faa3ead3a4c749877b77772723b5b" data-id="653faa3ead3a4c749877b77772723b5b"><span><div id="653faa3ead3a4c749877b77772723b5b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#653faa3ead3a4c749877b77772723b5b" title="1. The MAD Podcast with Matt Turck：Stripe&#x27;s AI Chief: How AI Agents Will Buy, Sell, and Pay"><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. The MAD Podcast with Matt Turck：Stripe&#x27;s AI Chief: How AI Agents Will Buy, Sell, and Pay</span></span></h4><div class="notion-text notion-block-a6d4c48de10c470493a8e1c91ed5c5ce">这期内容的基本结论是：agentic commerce 已经从概念进入基础设施建设期。Stripe 看到的方向不是“AI 帮你点一下购买按钮”这么简单，而是 agent 能在授权边界内发现商品、比较选择、完成支付，并和商家系统对接。最先落地的地方可能不是完全自主交易，而是 AI 搜索、AI 应用、广告和 checkout 混在一起的半自动购买流程。真正影响后续速度的，是消费者信任、商家目录标准、支付凭证安全、风控评分和新的定价方式。来源：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.youtube.com/@DataDrivenNYC/videos">The MAD Podcast with Matt Turck</a></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-fe65332ba00d4b1d8f247555f529aa7d" data-id="fe65332ba00d4b1d8f247555f529aa7d"><span><div id="fe65332ba00d4b1d8f247555f529aa7d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#fe65332ba00d4b1d8f247555f529aa7d" 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-5a43e745fe5847249f4913e00118499f">这条内容值得单独沉淀，是因为它把 AI agent 从“工作流工具”推到了“交易参与者”。当 agent 可以被授权花钱，产品形态、支付风控、商家获客和 SaaS 定价都会被重新改写。下一阶段的关键问题不是 agent 会不会买东西，而是谁能提供让它安全买、可控买、可结算买的基础设施。</div><div class="notion-text notion-block-9b7c7789389e4e4e99624ac3d3aba45a">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[SelfMem 深度分析 - 2026-07-11]]></title>
            <link>http://easyai.fyi/article/selfmem-ai-agent-memory-2026-07-11</link>
            <guid>http://easyai.fyi/article/selfmem-ai-agent-memory-2026-07-11</guid>
            <pubDate>Sat, 11 Jul 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[SelfMem 把 Agent 记忆从固定检索/摘要管线，改成可被模型自己检查、修订和优化的记忆策略。它在 BEAM 长对话记忆评测中优于 RAG、MemGPT、A-Mem、Mem0 等基线，但仍需要更多真实任务验证。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-39a5eac3813c81568570d8dc8eb648cd"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-847bda8f696d48528f9e31d906bbfb2a" data-id="847bda8f696d48528f9e31d906bbfb2a"><span><div id="847bda8f696d48528f9e31d906bbfb2a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#847bda8f696d48528f9e31d906bbfb2a" 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-c48e5996a494464aa87a109d29b45a13"><li>论文页面：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/abs/2607.03726">SelfMem: Self-Optimizing Memory for AI Agents</a></li></ul><ul class="notion-list notion-list-disc notion-block-6d8529865ef14260bca7c577ecb5ff67"><li>PDF：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/pdf/2607.03726">arXiv PDF</a></li></ul><ul class="notion-list notion-list-disc notion-block-bee67a8c7720468e918666ebe4079be9"><li>DOI：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://doi.org/10.48550/arXiv.2607.03726">10.48550/arXiv.2607.03726</a></li></ul><ul class="notion-list notion-list-disc notion-block-127e9a7b81ff45c998dee91bf23907f9"><li>论文 HTML 原图来源：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/html/2607.03726v1">arXiv HTML</a></li></ul><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-558a27c30434405a8af20c7438faf37d" data-id="558a27c30434405a8af20c7438faf37d"><span><div id="558a27c30434405a8af20c7438faf37d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#558a27c30434405a8af20c7438faf37d" title="今天为什么选 SelfMem"><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">今天为什么选 SelfMem</span></span></h3><div class="notion-text notion-block-9455618f5f52495d94557204883f409d">今天看了几个近期活跃的 Agent 方向：Orca 把多个 coding agent 放进并行 worktree，Rowboat 做本地长期工作记忆，agent-device 给移动端和真实设备测试补上“眼睛和手”，OpenConnector 解决 agent 调工具时的授权和动作目录问题。这些都很实用，但更像产品形态或基础设施。</div><div class="notion-text notion-block-2500dae0005e408d8b0c88bdeb489af0">SelfMem 值得单独写，是因为它碰的是更底层的问题：Agent 长期运行时，记忆到底该由人写规则来管，还是让 Agent 在受控环境里自己学会“该记什么、该删什么、该回源查什么”。这和最近几天写过的 harness、loop、评测不同，焦点落在 memory policy 本身。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-f550e9b25ea145a08acffec618de0497" data-id="f550e9b25ea145a08acffec618de0497"><span><div id="f550e9b25ea145a08acffec618de0497" class="notion-header-anchor"></div><a class="notion-hash-link" href="#f550e9b25ea145a08acffec618de0497" title="它解决的问题：不是记住更多，而是记得更会取舍"><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">它解决的问题：不是记住更多，而是记得更会取舍</span></span></h3><div class="notion-text notion-block-b2eae1ad9c7b48aa9a0893656682785e">很多 Agent 记忆方案会在三条路里选一条：全量上下文、RAG 检索、摘要/画像/知识图谱。问题是长任务里信息有新旧、有稳定和临时、有事实和偏好。把所有东西塞进上下文很贵，压缩会丢细节，RAG 又常常只在提问时捞几个片段，没形成可维护的工作记忆。</div><div class="notion-text notion-block-70696f2323234c77871c586491262a83">SelfMem 的核心判断很朴素：记忆不是一个静态仓库，而是一套可被检查和修订的策略。论文把这句话画成“教 Agent 钓鱼，而不是直接给鱼”。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-87fcc69af74d475f8c97470e6a9dad8e"><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/2607.03726v1/x1.png?t=87fcc69a-f74d-475f-8c97-470e6a9dad8e" alt="SelfMem Figure 1" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">SelfMem Figure 1</figcaption></div></figure><div class="notion-text notion-block-090eded4a90748b7a08dffbfaf1466ed">图 1 来自论文 HTML 原图。它表达的是 SelfMem 的出发点：不要预先规定一种固定记忆格式，而是给 Agent 原则、工具和反馈，让它自己学习记忆取舍。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-624a46228db44e51aa057753781f5793" data-id="624a46228db44e51aa057753781f5793"><span><div id="624a46228db44e51aa057753781f5793" class="notion-header-anchor"></div><a class="notion-hash-link" href="#624a46228db44e51aa057753781f5793" title="系统怎么运作"><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">系统怎么运作</span></span></h3><div class="notion-text notion-block-539283ccb2074c87825457faed3e4aa4">SelfMem 把系统拆成两层。第一层是不可变的原始对话记录，论文实现里用 SQLite 表保存 turn_index、role、time_anchor、content 等字段。它是事实来源，Agent 只能读，不能改。第二层是 Agent 管理的 memory workspace，里面可以是用户画像、偏好、项目状态、时间线，也可以是 Agent 自己整理出来的策略笔记。</div><div class="notion-text notion-block-0951c3f286c0424a940ff5a510a1f999">Agent 不直接“生成一份总结”就结束，而是在工具循环里做几类动作：读原始 transcript，读当前 memory，写入或修订 memory，检查 memory 质量，必要时做 self-test。检查项包括来源是否可靠、事实是否过期、是否有矛盾、摘要是否太粗、以后是否容易检索。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-de7c3617686b41928c86fb373ce4a114"><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/2607.03726v1/x2.png?t=de7c3617-686b-4192-8c86-fb373ce4a114" alt="SelfMem Figure 2" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">SelfMem Figure 2</figcaption></div></figure><div class="notion-text notion-block-4fe3516ae3c649288a700d85229f2f77">图 2 是 SelfMem 的主流程。左边保留原始 transcript，中间是 memory action space，右边是 Agent 生成的可用记忆。关键不在 SQLite，也不在某个具体工具名，而在这个闭环：inspect -&gt; write -&gt; review -&gt; revise。记忆第一次写错并不可怕，可怕的是系统没有让它发现错误和修正错误的地方。</div><div class="notion-text notion-block-4a94412d590a496da579f7101f05f3e8">回答问题时，SelfMem 会同时使用最终 memory workspace 和从原始 transcript 找回来的相关片段。这样可以避免两种极端：只靠记忆导致事实失真，只靠检索又缺少长期结构。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-3ef4dd8b55244e9790c465c4c8e63ccc" data-id="3ef4dd8b55244e9790c465c4c8e63ccc"><span><div id="3ef4dd8b55244e9790c465c4c8e63ccc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3ef4dd8b55244e9790c465c4c8e63ccc" title="它和常见 Agent 记忆方案的区别"><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 记忆方案的区别</span></span></h3><div class="notion-text notion-block-ecb21f97690f49ee9f20cdb14b9a2900">RAG 的强项是便宜、直接、可扩展，但它通常是“提问时检索”，不是“长期维护记忆”。SelfMem 不是否定 RAG，而是把 RAG 放回一个更大的记忆管理过程里：哪些内容该长期保存，哪些内容该临时回源查，由 Agent 根据反馈决定。</div><div class="notion-text notion-block-fbfb2a16bb5d4fd9bcb4cde5e1faf685">MemGPT、MemoryBank、Mem0、A-Mem 都比普通 RAG 更接近 Agent 记忆。MemGPT 让模型在不同记忆层之间搬运信息，MemoryBank 做长期用户记忆，Mem0 更偏生产化抽取和检索，A-Mem 允许 Agent 写、索引、链接记忆。SelfMem 往前推了一步：不只让 Agent 操作记忆，还让它调整“记忆策略”本身。</div><div class="notion-text notion-block-e6fe41b2209b42bc8112dbcf67e437cc">但这里也别夸过头。SelfMem 仍然依赖人设计好的 harness：原始记录怎么存、工具能做什么、review 看哪些风险、反馈怎么给，都是系统边界。真正新的地方，是把 memory policy 暴露给模型，让模型在工具反馈下修订策略。工程包装的部分，是 SQLite、工具接口、review checklist 这些实现形式。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-b4517488cd4f4366b43360b39aa7ccf7" data-id="b4517488cd4f4366b43360b39aa7ccf7"><span><div id="b4517488cd4f4366b43360b39aa7ccf7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b4517488cd4f4366b43360b39aa7ccf7" title="实验结果怎么看"><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">实验结果怎么看</span></span></h3><div class="notion-text notion-block-42b21704011641e79606f8d6c0d1dfb6">论文在 BEAM 长对话记忆评测上比较了 RAG、Full context、Compression、LoCoMo、ReadAgent、MemoryBank、MemGPT、A-Mem、Mem0 和 SelfMem。测试覆盖 100K、500K、1M token 三个规模，问题类型包括信息抽取、多轮推理、时间推理、偏好、更新、摘要等。</div><div class="notion-text notion-block-a578f3f5361749f9916870c847f51877">主表里，SelfMem 在三个规模上都拿到最高 Score 和 Pass0.5。100K 时 Score 是 0.504，RAG 是 0.339；500K 时 SelfMem 是 0.487，RAG 是 0.346；1M 时 SelfMem 是 0.454，RAG 是 0.320。论文还给了成本：1M token 规模下，SelfMem 成本为 2.004 美元，Mem0 是 18.830 美元，RAG 是 1.843 美元。也就是说 SelfMem 不是最便宜，但比很多重型记忆系统更准、更省。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-23e3d024de6b476ab1a6ed70d4ed7d53"><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/2607.03726v1/x3.png?t=23e3d024-de6b-476a-b1a6-ed70d4ed7d53" alt="SelfMem Figure 3" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">SelfMem Figure 3</figcaption></div></figure><div class="notion-text notion-block-033fc62d45464a53a7b4b461a6553bbc">图 3 来自论文的 BEAM 分问题类型结果。它比较重要，因为它说明 SelfMem 的优势不只来自某一类题。论文报告称，SelfMem 在 100K 规模下 10 类问题中拿下 9 类最佳，500K 下 8 类最佳，1M 下 7 类最佳。对 Agent 产品来说，这比单一总分更有参考价值，因为真实用户问的问题不会只落在一种记忆模式里。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-bb89b1e6ac1f4062b709006cc5ff7e7a" data-id="bb89b1e6ac1f4062b709006cc5ff7e7a"><span><div id="bb89b1e6ac1f4062b709006cc5ff7e7a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#bb89b1e6ac1f4062b709006cc5ff7e7a" title="更有意思的是策略优化"><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">更有意思的是策略优化</span></span></h3><div class="notion-text notion-block-07fc667853954d37aa44f1bb870631eb">SelfMem 不只构造记忆，还测试了“记忆策略能不能继续被优化”。论文把 conversation 0-8 用作训练反馈，把 conversation 9-19 作为 held-out 测试。Agent 看不到测试答案，只能根据训练分数和记忆工具诊断，生成新的策略说明。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-d339e13df22a4d3fad0a069d1d4c15f1"><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/2607.03726v1/x4.png?t=d339e13d-f22a-4d3f-ad0a-069d1d4c15f1" alt="SelfMem Figure 4" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">SelfMem Figure 4</figcaption></div></figure><div class="notion-text notion-block-18927d1c07704b0aa88d2fbd55e73432">图 4 显示默认策略的 held-out 分数是 0.472，最终 Agent 合成策略到 0.497，搜索过程中最好的策略到 0.510。这里最值得注意的不是涨了多少，而是最强策略并不出现在最多迭代或最多训练样本时。换句话说，记忆优化不是“跑越久越好”，而是需要找到合适的程序性规则：哪些事实长期保存，哪些事实保持可回源，哪些摘要必须带来源，哪些细节不能提前压扁。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-a8d4b60b71e34439b505f98f8e8785f9" data-id="a8d4b60b71e34439b505f98f8e8785f9"><span><div id="a8d4b60b71e34439b505f98f8e8785f9" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a8d4b60b71e34439b505f98f8e8785f9" 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-1b93febb86b242f1ac8b214e0c1d7b82">第一，长期记忆不要只做成资料库。真正难的是“记忆治理”：过期事实怎么降级，稳定偏好怎么保留，冲突信息怎么提示，重要细节怎么回源。</div><div class="notion-text notion-block-b9e2ca8f766c4daa87e2ebc781449550">第二，原始记录要当作事实底座保留。SelfMem 的 transcript store 很关键，因为 memory workspace 可以错、可以改、可以压缩，但原始证据不能跟着一起变形。</div><div class="notion-text notion-block-fa59f7e9f7a94151836f21e9ba001d9e">第三，把诊断反馈做成工具，而不是只靠提示词提醒。比如“这个记忆有没有来源”“是否和旧记忆冲突”“未来能不能检索到”，这些都可以变成 Agent loop 里的显式检查。</div><div class="notion-text notion-block-71cc29e468ee41c5b1ae8c46264b1c93">第四，memory 和 eval 要一起设计。没有长周期、多类型、带成本的评测，团队很容易把“记得多”误判成“记得好”。SelfMem 的结果提醒我们，记忆系统至少要同时看正确率、成本、请求次数、缓存利用和可审计性。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-a50e6fc289d549abae8a69236c848888" data-id="a50e6fc289d549abae8a69236c848888"><span><div id="a50e6fc289d549abae8a69236c848888" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a50e6fc289d549abae8a69236c848888" title="风险和没验证的地方"><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">风险和没验证的地方</span></span></h3><div class="notion-text notion-block-01979127f6074221a6d08bbc111ed3e0">SelfMem 目前主要在 BEAM 上验证，论文自己也说还需要更多 long-horizon benchmark、模型家族和真实部署场景验证。策略优化实验只在 100K 规模做了，500K 和 1M 下能否稳定优化，还没有结果。</div><div class="notion-text notion-block-034d2847e7ba4db792e7bbc7ca3a2bd0">它还依赖模型自己判断记忆质量。review 工具能降低风险，但不能保证 Agent 不会把错误事实写得很自信。真实产品里还要考虑隐私、用户可见性、删除权、团队共享记忆、多 Agent 同时写 memory 时的冲突处理。</div><div class="notion-text notion-block-385820c283a54d2389ba29604a51f19c">还有一点现实问题：我在 arXiv 页面没有看到作者提供代码链接。论文给了相当多流程和指标，但如果没有可复现实装，外部团队要判断它在自己业务里的成本和收益，还需要重做一遍 harness。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-34f00e00792449e29c36e88c277a893e" data-id="34f00e00792449e29c36e88c277a893e"><span><div id="34f00e00792449e29c36e88c277a893e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#34f00e00792449e29c36e88c277a893e" 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-d30d3c1cae7848f58737491dad96b759"><li>长期记忆的难点不是容量，是取舍。</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-6c2d7e7a1a4e4c5e90f145f78ce29f42"><li>好的 Agent memory 应该能回源、能审计、能修订。</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-f237c73d6a3f4ddd8b9ac04115c324cf"><li>记忆策略本身可以成为 Agent 优化对象。</li></ol><ol start="4" class="notion-list notion-list-numbered notion-block-aa5ca378db894a38805326a6798f1720"><li>RAG 适合找证据，不等于长期工作记忆。</li></ol><ol start="5" class="notion-list notion-list-numbered notion-block-81af865695cd4114aaafdd46f5be2622"><li>做 Agent 产品时，要把 memory、loop、eval 放在同一个设计里看。</li></ol></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[模型发布周，Agent 入口开打 - 2026-07-11]]></title>
            <link>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-11</link>
            <guid>http://easyai.fyi/article/follow-builders-ai-summary-2026-07-11</guid>
            <pubDate>Sat, 11 Jul 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[今日 follow-builders 全量沉淀：19 位 X builders、44 条 tweet、1 期 podcast。主线是模型发布周继续推进，Codex、Gemini、开放模型、企业 eval 和 Agent 运行底座都在往真实工作流里挤。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-3995eac3813c8100bd57f891e58bed05"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-9cb9a69d426d4baca8b0e39e913024df" data-id="9cb9a69d426d4baca8b0e39e913024df"><span><div id="9cb9a69d426d4baca8b0e39e913024df" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9cb9a69d426d4baca8b0e39e913024df" 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-192ade2175a0437d8c838c578590ae0c">今天这批内容的主线很集中：模型发布周不是单纯比谁的模型更聪明，而是在抢 Agent 的工作入口。OpenAI 这边把 GPT-5.6 Sol、ChatGPT Work 和 Codex 放在一起推，Gemini 团队公开整理用户对 Workspace、tool calling、MCP 的反馈，Box 用自己的复杂文档 eval 测 Sol，Replit 和 Anthropic 相关 builder 也在谈 coding agent、runtime 和未知问题管理。另一条线是开放模型和价格压力，大家开始意识到“够聪明、够快、够便宜”会比单点 benchmark 更影响真实采用。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-43907a65b7824f9d9c97084491b561f3" data-id="43907a65b7824f9d9c97084491b561f3"><span><div id="43907a65b7824f9d9c97084491b561f3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#43907a65b7824f9d9c97084491b561f3" 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-7dded1c3f953477e8f767ca3c111d91d">Thibault Sottiaux 的两条 Codex 更新最能代表今天的气氛：GPT-5.6 Sol 发布后，OpenAI 给 ChatGPT Work 和 Codex 重置额度，让用户有空间试更大的任务。这不是普通促销，而是把模型能力直接推向工作场景。Sam Altman 也说 Codex 是新 work product 的核心，这说明 coding agent 正从一个工具按钮，变成 OpenAI 想要占住的工作入口。</div><div class="notion-text notion-block-d6de1642fc354537b28584b00659e985">Josh Woodward 公开整理 Gemini 用户反馈，重点集中在 Workspace、tool calling、Projects、MCP 等真实工作流需求上。这个信号很明确：用户已经不满足于单次问答，他们要的是模型接入工具、上下文和日常生产系统。</div><div class="notion-text notion-block-4100e61c7ab249df9dde001b4bcfde06">Podcast 里 Jürgen Schmidhuber 的观点给这轮模型热降了温。他认为真正的 AGI 不能只在屏幕里完成，机器人硬件和真实世界能力仍然是短板；同时，开放模型追赶和算力降价会压缩闭源模型公司的定价空间。这和今天 X 上关于模型发布、开放模型速度、企业 eval 的讨论放在一起看，很有参考价值。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-79908cb56e5f40d286b63fe5fe34fb86" data-id="79908cb56e5f40d286b63fe5fe34fb86"><span><div id="79908cb56e5f40d286b63fe5fe34fb86" class="notion-header-anchor"></div><a class="notion-hash-link" href="#79908cb56e5f40d286b63fe5fe34fb86" 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-616d722d54a440d1843ca03351f6104d" data-id="616d722d54a440d1843ca03351f6104d"><span><div id="616d722d54a440d1843ca03351f6104d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#616d722d54a440d1843ca03351f6104d" title="1. Swyx @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 @swyx</span></span></h4><div class="notion-text notion-block-87f8c66fa14b4ee5bbf0b1201de1154f">Swyx 今天三条都偏 builder 生态观察。一条说前沿模型在需要发邮件时经常倾向使用 Resend，像是 AEO 做得很好；一条把 Greptile 和 OpenAI 的新动作联系起来，调侃早期争议后来可能成了更大产品方向的影子；最后一条是轻量社区梗。核心可记的是：AI 工具会改变“默认被模型选择”的产品分发路径。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-fbd6e90f695a4766a55e0e4b83f1f84e"><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=2075376938676621752" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-418298545f324fe99eb708e934774cb1"><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=2075336806661509513" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-7ff1a36d2ea846ce97073f0696af6756"><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=2075269966438494303" 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-e090dc7430bd4bb2997c1ab764c0e8b5" data-id="e090dc7430bd4bb2997c1ab764c0e8b5"><span><div id="e090dc7430bd4bb2997c1ab764c0e8b5" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e090dc7430bd4bb2997c1ab764c0e8b5" title="2. Josh Woodward @joshwoodward"><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. Josh Woodward @joshwoodward</span></span></h4><div class="notion-text notion-block-3c5aebbe54cc46f9bc9f77c284aa6e50">Josh Woodward 说 Gemini 在 12 小时内收到 1400+ 条反馈，他正在整理 Top 10 需求和团队进展。能看到的关键词是 Workspace integrations、tool calling、Projects、MCP 这些真实工作流能力。这里的重点不是反馈数量，而是用户把模型产品往“能接工具、能进日常工作系统”的方向推。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-e996e555bfe447e6ba2f203bc2ce81fe"><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=2075241749048401936" 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-29a6eeb4f406495ba04c2fc6233594a4" data-id="29a6eeb4f406495ba04c2fc6233594a4"><span><div id="29a6eeb4f406495ba04c2fc6233594a4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#29a6eeb4f406495ba04c2fc6233594a4" title="3. Thibault Sottiaux @thsottiaux"><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 @thsottiaux</span></span></h4><div class="notion-text notion-block-8f66b28a73624afbbff8a34249c80c35">Thibault 连发三条 OpenAI / Codex 相关更新：为庆祝 GPT-5.6 Sol 发布，ChatGPT Work 和 Codex 的使用额度会在 24 小时内重置两次；随后又说用户会先拿到一次完整重置；中间还有一条 3D 内容展示。这里最值得记的是额度策略：OpenAI 想让用户真的去试更复杂、更长的工作任务，而不是只看模型发布说明。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-c15c0522d997499d8b9dd3fdd2dfc378"><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=2075452680760443190" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-8e92bc70dbf74248b82e34c0636475d7"><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=2075389936564588714" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-784480fe1ec645dc9923601dee1acd7e"><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=2075330198887940337" 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-5fd48fd59615475ba95f5f2ad307497d" data-id="5fd48fd59615475ba95f5f2ad307497d"><span><div id="5fd48fd59615475ba95f5f2ad307497d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#5fd48fd59615475ba95f5f2ad307497d" title="4. Peter Yang @petergyang"><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 @petergyang</span></span></h4><div class="notion-text notion-block-8101499216eb4d04a775256bd170979b">Peter 今天一条是世界杯相关，属于非 AI 主线；一条是对 OpenAI 发布的称赞和反馈，核心是 OpenAI 比其他实验室更有机会把 agent 工作方式推向主流，因为 ChatGPT 已经有图片、实时语音、浏览器、computer use 和插件入口；另一条也是足球内容。真正值得沉淀的是第二条：Agent 的胜负不只在模型，而在入口、交互和已有用户习惯。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-38c388d208714afea1f9641230e1ac6f"><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=2075448068783407571" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-2ecd0b30ee5649eba2fc8ced4f5ca554"><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=2075345016437039600" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-4e348f5e935a4d2db1c46cfa1da5f794"><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=2075331828161138943" 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-9ed2b255c7024686a625990651b1e842" data-id="9ed2b255c7024686a625990651b1e842"><span><div id="9ed2b255c7024686a625990651b1e842" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9ed2b255c7024686a625990651b1e842" title="5. Nan Yu @thenanyu"><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. Nan Yu @thenanyu</span></span></h4><div class="notion-text notion-block-f7315e9e6f5e4753b0b72c1326bca4af">Nan Yu 今天三条都比较短：一条把 ChatGPT Codex 和 Adeptus Astartes 放在一起玩梗；两条吐槽融资视频里大数字和炫技表达。它们不是 AI 产品主线，但反映 builder 圈对“夸张传播”的疲劳感。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-0b27c3c4f7734bbe9fd892ed3490fbdb"><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=2075372186731270618" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-80a2c7dbd80b43e39081ac358736eaef"><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=2075369895408095511" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-8414555fe93e4a91b3a1cce2f3803460"><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=2075369080006087127" 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-0dd2e6ddd7d2466a90d44249b17ac52c" data-id="0dd2e6ddd7d2466a90d44249b17ac52c"><span><div id="0dd2e6ddd7d2466a90d44249b17ac52c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#0dd2e6ddd7d2466a90d44249b17ac52c" title="6. Madhu Guru @realmadhuguru"><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. Madhu Guru @realmadhuguru</span></span></h4><div class="notion-text notion-block-7696d0675d564910ba945de40bb03bca">Madhu Guru 宣布加入 Meta 做 AI 产品。他的判断很直接：软件工程 agent 已经明显改变了开发，但大多数复杂系统里的 agent 还处在早期，普通人还没有真正感受到 AI agent 的力量。这条和今天的 Codex、Gemini、runtime 讨论连在一起看，说明 agent 竞争正在从开发者场景向更多复杂系统扩散。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-a835cd3be908469ca5132cd0d1a5ac18"><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=2075243087325217038" 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-9e3dc26077d741b4a85b186964950a75" data-id="9e3dc26077d741b4a85b186964950a75"><span><div id="9e3dc26077d741b4a85b186964950a75" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9e3dc26077d741b4a85b186964950a75" title="7. Amanda Askell @AmandaAskell"><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. Amanda Askell @AmandaAskell</span></span></h4><div class="notion-text notion-block-60e246a9508e4a6ebdc31ddfb4c765bb">Amanda 两条都在讲纽约建筑倒塌的个人记忆和事实校正，属于生活化内容，不是 AI 主线。按全量沉淀要求保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-77b830cbffef48129a8f130528cbc540"><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=2075247953309311043" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-f717bce83dce4b0795b059250941c349"><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=2075245939548455009" 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-bbb3a3641ea24789bab23fe0d170811a" data-id="bbb3a3641ea24789bab23fe0d170811a"><span><div id="bbb3a3641ea24789bab23fe0d170811a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#bbb3a3641ea24789bab23fe0d170811a" title="8. Thariq @trq212"><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. Thariq @trq212</span></span></h4><div class="notion-text notion-block-c14df4a77d9f4a528dcfbbd888ae895d">Thariq 的第一条很值得记：agentic coding 的核心技能之一，是减少未知数。这比“让 agent 多写代码”更接近真实工程经验。第二条是转发 Fable 相关可用性。放在一起看，Anthropic / Claude Code 生态里的重点仍然是让 agent 在复杂代码库里更可控、更能推进任务。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-820fd044ed044c0180d8e4de68e5f6b7"><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=2075283841758183674" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-acd3d6fda15c471caa46ec9422566d33"><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=2075280416995705312" 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-1988e7cdab804fc9b2f729c661b536e7" data-id="1988e7cdab804fc9b2f729c661b536e7"><span><div id="1988e7cdab804fc9b2f729c661b536e7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#1988e7cdab804fc9b2f729c661b536e7" title="9. Amjad Masad @amasad"><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. Amjad Masad @amasad</span></span></h4><div class="notion-text notion-block-2ba813f926414188b0180cfda48e8130">Amjad Masad 今天三条都和 AI 产品落地有关。他说 AI 让 coding 更灵活，但 Replit 反而在让 runtime 更刚性：更多 formal specs、更确定的系统、更有韧性的基础设施。另一条判断 LLM 市场已经变得更动态，反对单一模型公司会垄断的看法；第三条鼓励用户试 Fable in Replit。核心信号是：agent 越能快速生成和修改代码，底层 runtime 越要稳定。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-b1cc11f0be2647b18ec8b14d52b9b611"><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=2075423115052790054" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-47301996a1684a80912b524fe5a1d2c3"><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=2075413916491075755" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-b21fa41c6fd54938855d535121d9b4f9"><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=2075358353686208741" 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-41eb16ff8ae342a292aeb2543d8cb501" data-id="41eb16ff8ae342a292aeb2543d8cb501"><span><div id="41eb16ff8ae342a292aeb2543d8cb501" class="notion-header-anchor"></div><a class="notion-hash-link" href="#41eb16ff8ae342a292aeb2543d8cb501" title="10. Guillermo Rauch @rauchg"><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. Guillermo Rauch @rauchg</span></span></h4><div class="notion-text notion-block-de6492bb92bd45389b2e7b80ad07f2a0">Guillermo Rauch 今天谈的是模型市场。他说开放模型会变得非常快，并判断 Meta Spark 1.1、Grok 4.5、GLM 5.2 可能显著挤占 token 市场份额，因为多数 agentic tasks 需要的是“足够高的智能 + 快速度”。这条很关键：模型竞争正在从“最强”走向“高性价比、低延迟、可大规模跑任务”。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-ecad40201fac48ceafbe36a93eb36877"><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=2075294327354577256" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-77fab796795d4e558dc660983f82402b"><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=2075294130327196152" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-a7d2275cb8f4429997603789c0ac68af"><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=2075255565627080813" 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-8b7d261c06cf49ba92e67f6775d99a28" data-id="8b7d261c06cf49ba92e67f6775d99a28"><span><div id="8b7d261c06cf49ba92e67f6775d99a28" class="notion-header-anchor"></div><a class="notion-hash-link" href="#8b7d261c06cf49ba92e67f6775d99a28" title="11. Alex Albert @alexalbert__"><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. Alex Albert @alexalbert__</span></span></h4><div class="notion-text notion-block-4957297309cb421a8953380461e3d153">Alex Albert 转发 More Fable，信息量不大，但和 Thariq、Amjad 的 Fable/Replit 线索呼应：Anthropic 相关模型和 coding agent 工具正在继续放量。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-85c9b8d0c7d940549f65245f557e0f9e"><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=2075285305096343583" 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-abf80e0ce2b4451d99f327fce857ef64" data-id="abf80e0ce2b4451d99f327fce857ef64"><span><div id="abf80e0ce2b4451d99f327fce857ef64" class="notion-header-anchor"></div><a class="notion-hash-link" href="#abf80e0ce2b4451d99f327fce857ef64" title="12. Aaron Levie @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">12. Aaron Levie @levie</span></span></h4><div class="notion-text notion-block-c244ae72694943c382baaf6f0941d384">Aaron Levie 一条讨论模型共享后的竞争优势：当行业智能越来越多来自相同模型，企业真正的差异会转向流程、数据、产品集成和分发。另一条是 Box 对 GPT-5.6 系列的评估，重点是 Sol 在 Box AI Complex Work eval 里，相比 GPT-5.5 在复杂文档和数据任务上有明显提升。企业不会只看榜单，而会看模型能不能在自己的文档集里给出可校验结果。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-8be459874ade48e189cb71790053b519"><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=2075416313481290077" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-37380a9b29594e3a87c65f89752c9d3e"><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=2075287443411222628" 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-d6920985977444908aa96bf4bc825641" data-id="d6920985977444908aa96bf4bc825641"><span><div id="d6920985977444908aa96bf4bc825641" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d6920985977444908aa96bf4bc825641" title="13. Garry Tan @garrytan"><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. Garry Tan @garrytan</span></span></h4><div class="notion-text notion-block-d2947f4a41fb44ff902eed958a42bb98">Garry Tan 提到 Meta Muse Spark 1.1 在 OpenClaw 上表现很好，并点名早期版本 Hornbill。这条和今天的模型发布周主线相关：除了 OpenAI，新模型和开放/半开放模型也在快速进入 builder 的实际工具链。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-de5fce0efa6342e7bd3ed0c1e6f170da"><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=2075445455438385255" 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-e179f61d4265474eaa91c3fbebbdd009" data-id="e179f61d4265474eaa91c3fbebbdd009"><span><div id="e179f61d4265474eaa91c3fbebbdd009" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e179f61d4265474eaa91c3fbebbdd009" title="14. Matt Turck @mattturck"><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. Matt Turck @mattturck</span></span></h4><div class="notion-text notion-block-1f7ce5869b2d49ac9b5b7b344f542f82">Matt 今天三条都是世界杯 / 巴黎相关，属于非 AI 主线。按“不过滤”要求完整保留。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-78728d486487464ebb1d756bf238d82b"><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=2075386931924414686" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-d6825b18b80240a2a4f0a22b57e4effd"><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=2075361793904439489" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-805acfc07c834a9aa8f117adc137229f"><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=2075342627436712362" 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-27e3acabbb1541c2a0ceb6feaf1e105e" data-id="27e3acabbb1541c2a0ceb6feaf1e105e"><span><div id="27e3acabbb1541c2a0ceb6feaf1e105e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#27e3acabbb1541c2a0ceb6feaf1e105e" title="15. Nikunj Kothari @nikunj"><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. Nikunj Kothari @nikunj</span></span></h4><div class="notion-text notion-block-2f62a03b04194f8580d3b725c3ce9e11">Nikunj 一条用生活化方式解释这周模型发布局势：GPT-5.6 有 Sol、Terra、Luna 三个版本，Sol 最强，还要通过政府安全评估；另一条是电气工程相关轻松内容。第一条说明模型发布节奏已经复杂到需要“面向普通人解释版本、能力、限制和放量方式”。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-38f5e5bc433f4852a6d62bfa02b8e4e3"><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=2075411514773967261" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-093e43e6954d47ff92d4d85c7fa1ad2c"><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=2075248134457041341" 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-878a0e9cbdd745e49d3fa37bc2f382f7" data-id="878a0e9cbdd745e49d3fa37bc2f382f7"><span><div id="878a0e9cbdd745e49d3fa37bc2f382f7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#878a0e9cbdd745e49d3fa37bc2f382f7" title="16. Peter Steinberger @steipete"><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. Peter Steinberger @steipete</span></span></h4><div class="notion-text notion-block-e2813917949943fea68bce45e7155253">Peter 今天三条都是短转发和活动相关：称赞某个内容很强、说会做 livestream、以及祝贺别人做出好东西。原帖信息量有限，但保留为全量记录。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-bbf18ecfe88d45bf8a2043ee1374bf48"><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=2075350572560191630" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-819f54a67f154bc4b6ea8f207a473cfc"><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=2075328495677521934" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-1bf85ffd4b06434b9e17252b2158ae82"><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=2075313523237019686" 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-aeb0e76ab9a940a09b82d4b3b1321fcd" data-id="aeb0e76ab9a940a09b82d4b3b1321fcd"><span><div id="aeb0e76ab9a940a09b82d4b3b1321fcd" class="notion-header-anchor"></div><a class="notion-hash-link" href="#aeb0e76ab9a940a09b82d4b3b1321fcd" title="17. Dan Shipper @danshipper"><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. Dan Shipper @danshipper</span></span></h4><div class="notion-text notion-block-62d9f7af5dde4f6cadabef0c3524a0c1">Dan Shipper 今天三条：一条 Every 内部/内容梗，一条吐槽“developers don&#x27;t do work”的说法，另一条转发 GPT-5.6 Sol 作为知识工作标准的内容。它们都围绕 AI 工作方式的叙事，尤其第三条和今天的 Sol / knowledge work 主线直接相关。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-2980596e8be4407499d0b17cadb9a3b6"><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=2075389275969826879" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-cbcf2218e707436aafaf4d1552230714"><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=2075330044289802584" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-4a50147bd41f40d38f3adcf8eb390200"><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=2075264022988116280" 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-90d1d885657f4de9955337f78bbb1a6b" data-id="90d1d885657f4de9955337f78bbb1a6b"><span><div id="90d1d885657f4de9955337f78bbb1a6b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#90d1d885657f4de9955337f78bbb1a6b" title="18. Aditya Agarwal @adityaag"><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">18. Aditya Agarwal @adityaag</span></span></h4><div class="notion-text notion-block-ccad752cef754991a77dbdc271d30afd">Aditya 两条都来自 South Park Commons 访谈内容，讲 Gagan Shux 从申请印度空军到登上 ISS 的经历。不是 AI 主线，但属于 builder 内容，保留记录。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-44c349f62ee44f7dad70f9c3ac422a30"><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=2075414473695703054" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-3db61a341db348fab856649413a09ea5"><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=2075414469497270557" 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-3ca2400919a64ae39dfbb03641c698b6" data-id="3ca2400919a64ae39dfbb03641c698b6"><span><div id="3ca2400919a64ae39dfbb03641c698b6" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3ca2400919a64ae39dfbb03641c698b6" title="19. Sam Altman @sama"><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">19. Sam Altman @sama</span></span></h4><div class="notion-text notion-block-b9d2f962d9204814af98c281cd56493b">Sam Altman 今天三条里，一条是对 Fidji 个人情况的祝福，不展开；一条明确说 Codex 是新 work product 的核心，Codex 不会消失；另一条说企业对 AI 成本的担忧已经被听到，5.6 Sol、Terra、Luna 在 dollars-per-task 上是重要进步。这三条连起来看，OpenAI 今天的关键词是工作产品、Codex 和单位任务成本。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-dc2956bf92c44b09b211732e073c3d0f"><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=2075354679031067058" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-000f8e2239ee42809830dbfd35755a8d"><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=2075293792048136572" title="iframe embed" frameBorder="0" allowfullscreen="" loading="lazy" scrolling="auto"></iframe></div></figure><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-f5082155b5864d00a31a64e3024c78c9"><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=2075267201058426944" 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-9a3a4fff857e4a5c81951be76138a856" data-id="9a3a4fff857e4a5c81951be76138a856"><span><div id="9a3a4fff857e4a5c81951be76138a856" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9a3a4fff857e4a5c81951be76138a856" 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-a0270d7df855485f828e887553956b2d">今天 follow-builders 输出里没有 blog 更新。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-262151e89f8849eba31e0bd5080cc6e6" data-id="262151e89f8849eba31e0bd5080cc6e6"><span><div id="262151e89f8849eba31e0bd5080cc6e6" class="notion-header-anchor"></div><a class="notion-hash-link" href="#262151e89f8849eba31e0bd5080cc6e6" 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-bda9b1ad5fb14b46b9c6e149f750b5d6" data-id="bda9b1ad5fb14b46b9c6e149f750b5d6"><span><div id="bda9b1ad5fb14b46b9c6e149f750b5d6" class="notion-header-anchor"></div><a class="notion-hash-link" href="#bda9b1ad5fb14b46b9c6e149f750b5d6" title="1. Unsupervised Learning：Ep 90: AI Pioneer Jürgen Schmidhuber on the State of AI Today"><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. Unsupervised Learning：Ep 90: AI Pioneer Jürgen Schmidhuber on the State of AI Today</span></span></h4><div class="notion-text notion-block-957af2b899b34e5d95d8e2aaffd2b0e9">这期 podcast 的核心观点很清楚：Schmidhuber 对 AI 技术长期乐观，但对当前模型公司的商业结构更谨慎。他认为真正的 AGI 不能只在屏幕里完成，机器人硬件和真实世界能力还远弱于人类身体；他也认为当前数据中心和 GPU 投入可能被高估，因为开放模型追赶和未来算力降价会持续压缩闭源模型的定价空间。对今天的模型发布周来说，这是一盆有用的冷水：模型能力会继续进步，但护城河不一定在模型本身。</div><figure class="notion-asset-wrapper notion-asset-wrapper-embed notion-block-e12c4443a0404564862d4c87bac98e7d"><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/RKjR8DQ40po" 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-a990be6f065b45e68e1cc9405631a798" data-id="a990be6f065b45e68e1cc9405631a798"><span><div id="a990be6f065b45e68e1cc9405631a798" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a990be6f065b45e68e1cc9405631a798" 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-90f1b12e859343fbbbf5340bef3a0d5e">今天所有内容放在一起看，主线不是“又有几个模型发布了”，而是模型正在被拉进真实工作：Codex 要成为 work product 核心，Gemini 用户在要工具和 Workspace，企业用自己的 eval 评估 Sol，Replit 强调 runtime 稳定，开放模型在速度和价格上继续施压。下一轮 AI 产品竞争，很可能不是单点能力，而是谁能把模型、工具、运行环境和成本控制接成一个稳定工作入口。</div><div class="notion-text notion-block-197cd55324d447f5a94a51a92d0ba816">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>
    </channel>
</rss>