<?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>月明的博客</title>
        <link>https://lunalight.top//</link>
        <description>分享AI实践与思考</description>
        <lastBuildDate>Thu, 29 Oct 2026 03:18:40 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, 月明</copyright>
        <item>
            <title><![CDATA[MAI-Image-2.6 系列在 Foundry 正式可用：精度与速度的分工]]></title>
            <link>https://lunalight.top//article/azure-foundry-2026-10-29-4566327</link>
            <guid>https://lunalight.top//article/azure-foundry-2026-10-29-4566327</guid>
            <pubDate>Thu, 29 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[微软宣布 MAI-Image-2.6 与 2.6-Flash 在 Foundry 正式可用，前者偏重精细编辑，后者面向快速迭代。本文梳理其能力分工、适用场景与仍需验证之处。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-40870d5879c8813c87b6e19e2a7b97a8"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40870d5879c8817e919ed5cc9fb0b1fe" data-id="40870d5879c8817e919ed5cc9fb0b1fe"><span><div id="40870d5879c8817e919ed5cc9fb0b1fe" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c8817e919ed5cc9fb0b1fe" title="这篇文章在讲什么"><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">这篇文章在讲什么</span></span></h3><div class="notion-text notion-block-40870d5879c881dcb461d5b7e2b6c72b">微软在 Foundry 博客中宣布，MAI-Image-2.6 与 MAI-Image-2.6-Flash 已进入正式可用阶段。文章的重点不是单张图片画得多好，而是企业团队如何在编辑、适配多种版位和规模化生产中保持可控。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40870d5879c881eaa137f69366b6914e"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/f9083f1e-33e8-42a0-acaf-7f5d6563537d/image-1-image.png?table=block&amp;id=40870d58-79c8-81ea-a137-f69366b6914e&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=0WpmVa8loQEkZXz5YP8ONGwJQ1EasbyGtd0OeZv52Js&amp;t=40870d58-79c8-81ea-a137-f69366b6914e" alt="图片来源：Microsoft Foundry Blog原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：Microsoft Foundry Blog<a class="notion-link" href="https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/mai-image-2-6-and-mai-image-2-6-flash-are-now-generally-available-in-microsoft-f/4566327" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40870d5879c88109a86ad6b4393e0313" data-id="40870d5879c88109a86ad6b4393e0313"><span><div id="40870d5879c88109a86ad6b4393e0313" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c88109a86ad6b4393e0313" title="来源披露的关键能力"><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">来源披露的关键能力</span></span></h3><div class="notion-text notion-block-40870d5879c8818dac5bc808d2dbab8e">按文章说明，MAI-Image-2.6 支持多参考图编辑，最多可使用五张参考图，用来在修改过程中保持产品、角色和品牌素材的一致性。两个模型还提供联网检索（基于 Bing）、动态宽高比，以及最高 1.5K 的输出分辨率。</div><div class="notion-text notion-block-40870d5879c881cd84c9d1f1749c605c">文章对联网检索的表述比较克制：它只是补充上下文，不能代替人工审核。这一点对有事实、品牌或编辑要求的场景尤其重要。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40870d5879c881109bd3c499d33c1477" data-id="40870d5879c881109bd3c499d33c1477"><span><div id="40870d5879c881109bd3c499d33c1477" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c881109bd3c499d33c1477" title="两个模型如何分工"><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">两个模型如何分工</span></span></h3><div class="notion-text notion-block-40870d5879c8816c80a6fc949a87a033">文章的定位是：MAI-Image-2.6 面向复杂编辑、文字渲染和最终的活动主视觉；Flash 面向交互式创作工具、快速迭代、个性化和大批量生成。文中给出的零售示例流程是：先用 Flash 探索和修改，再切换到完整版制作带文字的主视觉。</div><div class="notion-text notion-block-40870d5879c88167b2b9f72c314cf979">定价方面，文章列出了两款模型按百万 token 计费的文本输入、图像输入和图像输出价格，完整版的输出价格高于 Flash。具体数字请以原文和价格计算器为准。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40870d5879c8813bbde3c2036ceeb5d4" data-id="40870d5879c8813bbde3c2036ceeb5d4"><span><div id="40870d5879c8813bbde3c2036ceeb5d4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c8813bbde3c2036ceeb5d4" title="我的分析"><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">我的分析</span></span></h3><div class="notion-text notion-block-40870d5879c88134bfd6c795413c0490">以下为作者观点。双模型分工是一种常见的成本控制思路：草稿阶段用便宜、快的模型，定稿阶段才调用贵的模型。开发者可以在应用中把两者做成可切换的流水线，并把多参考图作为品牌一致性的主要手段。</div><div class="notion-text notion-block-40870d5879c8818d86a7e28c0264a40c">不过，来源文章是厂商发布稿，客户评价与合作方说法（如 Minute、Adobe Firefly 的后续接入）都属于对方陈述，我没有独立核实。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40870d5879c88172b82ff179aa672de0" data-id="40870d5879c88172b82ff179aa672de0"><span><div id="40870d5879c88172b82ff179aa672de0" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c88172b82ff179aa672de0" title="值得继续观察"><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">值得继续观察</span></span></h3><ul class="notion-list notion-list-disc notion-block-40870d5879c8819da1d3d3d1de1f1d7e"><li>多次编辑后的一致性在真实品牌素材上的表现如何。</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c8812ea125fbb818adfd6b"><li>联网检索带来的内容，在合规与事实核查上需要怎样的流程。</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c881fdbbe9ec52c154c3b5"><li>1.5K 输出上限是否满足印刷或大尺寸投放需求（推测：部分场景仍需后期放大）。</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c881f7a8face4f5166d73d"><li>Flash 与完整版在真实负载下的实际成本差异。</li></ul><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40870d5879c881ad8c2de3eaafd24f04" data-id="40870d5879c881ad8c2de3eaafd24f04"><span><div id="40870d5879c881ad8c2de3eaafd24f04" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c881ad8c2de3eaafd24f04" title="来源"><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">来源</span></span></h3><ul class="notion-list notion-list-disc notion-block-40870d5879c88107aa25fa71955b3fbf"><li>Microsoft Foundry Blog 原文：https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/mai-image-2-6-and-mai-image-2-6-flash-are-now-generally-available-in-microsoft-f/4566327</li></ul></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[一天四条消息：模型竞速、云业务增长与安全能力的外溢]]></title>
            <link>https://lunalight.top//article/ai-daily-2026-10-29-001928</link>
            <guid>https://lunalight.top//article/ai-daily-2026-10-29-001928</guid>
            <pubDate>Thu, 29 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Google 的新模型与财报、Anthropic 的漏洞研究、微软开放 Agent 运行框架，以及对 Anthropic 招股书的批评，都围绕 AI 能力与商业化展开。本文区分来源事实与个人分析，并标出尚未核实的部分。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-40870d5879c8816a9d1ae44be04cea6c"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-40870d5879c88156a45ef3b9fb867fd3" data-id="40870d5879c88156a45ef3b9fb867fd3"><span><div id="40870d5879c88156a45ef3b9fb867fd3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c88156a45ef3b9fb867fd3" title="能力、资本与安全同时加速"><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">能力、资本与安全同时加速</span></span></h2><div class="notion-text notion-block-40870d5879c88150b3b3c23f0a0ba1b7">2026-10-29 综述</div><div class="notion-text notion-block-40870d5879c881a29877cbb2c6e3b871">今天的几条消息看似分散，指向的却是同一个问题：模型能力在变强，围绕它的商业叙事、安全风险和工程基础设施也在同步变化。下文先写来源中可核实的事实，再写我的判断。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40870d5879c8811facb6e205f7582094" data-id="40870d5879c8811facb6e205f7582094"><span><div id="40870d5879c8811facb6e205f7582094" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c8811facb6e205f7582094" title="Google：新模型与财报互相印证"><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">Google：新模型与财报互相印证</span></span></h3><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40870d5879c881168c8de61d05a8d589"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/bf9d1bdd-d791-4ecf-a0ff-e12668744a5c/image-3-image.webp?table=block&amp;id=40870d58-79c8-8116-8c8d-e61d05a8d589&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=eVRKT5jq7HwntaAfT5pMuXrpcSEDg8WO44qMRfahHb0&amp;t=40870d58-79c8-8116-8c8d-e61d05a8d589" alt="图片来源：Sundar Pichai on X: &quot;I did say we&#x27;d be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model. Following Argon&#x27;s release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Int… / X原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：Sundar Pichai on X: &quot;I did say we&#x27;d be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model. Following Argon&#x27;s release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Int… / X<a class="notion-link" href="https://x.com/sundarpichai/status/2115489219804352921" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><div class="notion-text notion-block-40870d5879c88195949edb8d4b5a5014">来源事实：Sundar Pichai 在 X 上介绍了 Gemini 4 Carbon，称其在更新后的 Artificial Analysis Intelligence Index v5.0 上排名第一，得分 59。帖子还提到它在 SEC-bench Pro 上得分 88.2%，用于衡量发现漏洞的能力。这些数字都是 Google 方面的自述，目前没有第三方复核。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40870d5879c88154aeeecb7ed8bce6e6"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/11fa9b6e-a4cd-4207-8890-71550565e57b/image-1-image.webp?table=block&amp;id=40870d58-79c8-8154-aeee-cb7ed8bce6e6&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=gul-KiThCHaPvSEtKoA9Y609kQ4KdNqMx3y8Wb2U2hg&amp;t=40870d58-79c8-8154-aeee-cb7ed8bce6e6" alt="图片来源：Sundar Pichai on X: &quot;Q3 is in the books! We’re now on a half-trillion annual revenue run rate, with Alphabet revenues growing at 25%. - Our AI investments are driving overall Search query growth, including commercial queries - our YoY query growth rate is the highest it’s been in the last five yea… / X原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：Sundar Pichai on X: &quot;Q3 is in the books! We’re now on a half-trillion annual revenue run rate, with Alphabet revenues growing at 25%. - Our AI investments are driving overall Search query growth, including commercial queries - our YoY query growth rate is the highest it’s been in the last five yea… / X<a class="notion-link" href="https://x.com/sundarpichai/status/2115535455156642066" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><div class="notion-text notion-block-40870d5879c88173bcd7f81ffc09bb42">同日，Pichai 公布 Alphabet 第三季度业绩：年化收入达到约 5000 亿美元，总收入同比增长 25%，Google Cloud 收入增长 103%，并称搜索查询同比增速是五年来最高。</div><div class="notion-text notion-block-40870d5879c881848f71d70246fa0f76">我的观察：模型发布节奏与云收入增长被放在同一叙事中，说明模型竞争已经直接和基础设施需求绑定。不过 103% 的增速只是公司披露的单项数字，我没有看到分业务的完整财务细节，因此不做更深的推断。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40870d5879c881b49199f45b71fb7b9e" data-id="40870d5879c881b49199f45b71fb7b9e"><span><div id="40870d5879c881b49199f45b71fb7b9e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c881b49199f45b71fb7b9e" title="Anthropic：安全研究显示模型的双面性"><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">Anthropic：安全研究显示模型的双面性</span></span></h3><div class="notion-text notion-block-40870d5879c8816da39adef4399e687c">来源事实：据 IT之家报道，Anthropic 于 10 月 27 日发布研究报告，称一个未公开发布的内部 Claude 研究模型在 Chrome、Safari、Android、Windows 和 Linux 中发现漏洞，并把多个漏洞串联成 5 条可实际运行的攻击链。报道称相关漏洞已按协调披露流程报告厂商并修复，其中提交给谷歌的报告涉及 62 个严重或高危 CVE。</div><div class="notion-text notion-block-40870d5879c88164a115f2dca42bcd4f">我的观察：Google 的新模型在漏洞发现基准上领先，Anthropic 的研究又展示了模型在真实软件中找漏洞的能力。两者合在一起，意味着防守方和攻击方都会更快地用上这类能力。对开发者来说，补丁响应速度和依赖更新流程的重要性会上升。</div><blockquote class="notion-quote notion-block-40870d5879c881f69ca3d4979d3ae40c"><div>以上漏洞数量与场景来自二手报道对原报告的转述，详细技术内容应以 Anthropic 原文为准。</div></blockquote><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40870d5879c881d99874c257731218cf" data-id="40870d5879c881d99874c257731218cf"><span><div id="40870d5879c881d99874c257731218cf" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c881d99874c257731218cf" title="微软：把 Copilot harness 开放给开发者"><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">微软：把 Copilot harness 开放给开发者</span></span></h3><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40870d5879c88169bc8efa0d39084db7"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/588f6847-865b-42d8-97cd-a806bd9ac35e/image-4-image.png?table=block&amp;id=40870d58-79c8-8169-bc8e-fa0d39084db7&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=kMRMde52qUZkdheGXS0zlvZtI1u0M9fz-2WCcoJ6kHA&amp;t=40870d58-79c8-8169-bc8e-fa0d39084db7" alt="图片来源：Introducing the Microsoft Copilot harness: Unlocking the potential of truly multi-model agentic work - The Official Microsoft Blog Your Privacy Choices Opt-Out Icon Your Privacy Choices Opt-Out Icon原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：Introducing the Microsoft Copilot harness: Unlocking the potential of truly multi-model agentic work - The Official Microsoft Blog Your Privacy Choices Opt-Out Icon Your Privacy Choices Opt-Out Icon<a class="notion-link" href="https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness-unlocking-the-potential-of-truly-multi-model-agentic-work/" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><div class="notion-text notion-block-40870d5879c88117a602f8650c08952c">来源事实：微软官方博客发布了介绍 Microsoft Copilot harness 的文章，标题强调多模型 Agent 工作。据任务摘要，该 harness 将通过 Microsoft Foundry 向开发者和企业开放。我能核实的只有文章标题与发布位置，产品范围与可用时间未能从抓取到的页面摘录中确认。</div><div class="notion-text notion-block-40870d5879c881558838ed4683a985f1">我的推测：如果开放属实，Agent 的竞争点可能从单个模型转向运行框架、工具调用和多模型调度。这只是推测，需要等产品文档确认。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40870d5879c8817a9614eb995d80a319" data-id="40870d5879c8817a9614eb995d80a319"><span><div id="40870d5879c8817a9614eb995d80a319" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c8817a9614eb995d80a319" title="对 Anthropic S-1 的批评"><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 Anthropic S-1 的批评</span></span></h3><div class="notion-text notion-block-40870d5879c881b8b374d64f02b30cd7">来源事实：Ed Zitron 在 10 月 28 日发表长文 Exit Liquidity，主题是 Anthropic 的 S-1 招股书。他的具体论点属于个人观点，且我取得的页面摘录没有覆盖正文论证，因此无法核实其对招股书内容的具体描述。</div><div class="notion-text notion-block-40870d5879c8819cb6f0c30a1ace8100">我的观察：无论该批评是否成立，AI 公司进入公开市场后，财务披露的完整性都会成为关注焦点。读者应对照招股书原文，而不是只看评论。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40870d5879c881e6806ecb576b810630" data-id="40870d5879c881e6806ecb576b810630"><span><div id="40870d5879c881e6806ecb576b810630" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c881e6806ecb576b810630" title="值得继续观察"><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">值得继续观察</span></span></h3><ul class="notion-list notion-list-disc notion-block-40870d5879c8812c912ad6815a954b60"><li>Gemini 4 Carbon 的成绩能否被独立评测复现。</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c8815393aaca66d697318b"><li>漏洞发现能力增强后，厂商的修复与披露流程能否跟上。</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c881edb319eab1170d66e1"><li>Copilot harness 的具体开放范围与支持的模型。</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c88181b2bacc538cf0e356"><li>Anthropic 招股书的披露内容与外界评论之间的差异。</li></ul><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40870d5879c88149b412e586faec086b" data-id="40870d5879c88149b412e586faec086b"><span><div id="40870d5879c88149b412e586faec086b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40870d5879c88149b412e586faec086b" title="来源清单"><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">来源清单</span></span></h3><ul class="notion-list notion-list-disc notion-block-40870d5879c8818b8218e99bb7093d6c"><li>https://www.ithome.com/1/019/268.htm</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c88142be89db401d1a43a8"><li>https://x.com/sundarpichai/status/2115535455156642066</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c881c3a3d7de542bce4ffa"><li>https://x.com/sundarpichai/status/2115489219804352921</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c88183b369fe07f91bb36f"><li>https://www.wheresyoured.at/exit-liquidity/</li></ul><ul class="notion-list notion-list-disc notion-block-40870d5879c88168ab11cf5169ee8f61"><li>https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness-unlocking-the-potential-of-truly-multi-model-agentic-work/</li></ul></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[多智能体系统跨区域上线：先证明“上线的就是测过的”]]></title>
            <link>https://lunalight.top//article/azure-foundry-2026-10-28-4564916</link>
            <guid>https://lunalight.top//article/azure-foundry-2026-10-28-4564916</guid>
            <pubDate>Wed, 28 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Microsoft Foundry 博客的第二部分围绕托管智能体的发布与运维，提出七项实践。这篇分析梳理其中的主线，并讨论它们对开发团队的影响。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-40770d5879c88192ad80fb9e15c739ce"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-40770d5879c8817d91a5f83b75a0071f" data-id="40770d5879c8817d91a5f83b75a0071f"><span><div id="40770d5879c8817d91a5f83b75a0071f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c8817d91a5f83b75a0071f" title="发布证据比部署成功更重要"><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">发布证据比部署成功更重要</span></span></h2><div class="notion-text notion-block-40770d5879c881ba91d0fd6196a05b3a">托管智能体跨区域上线时的七项实践</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40770d5879c8819c80dcca691cc225a2" data-id="40770d5879c8819c80dcca691cc225a2"><span><div id="40770d5879c8819c80dcca691cc225a2" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c8819c80dcca691cc225a2" title="这篇文章在讨论什么"><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">这篇文章在讨论什么</span></span></h3><div class="notion-text notion-block-40770d5879c88128b46bcc2b82b422df">这是 Microsoft Foundry 博客“全球企业多智能体系统”系列的第二篇。上一篇偏重起步、设计与安全，这一篇转向发布与运维。作者的核心观点是：智能体能够启动，不等于它就是被测试过的那个版本，也不等于配置正确、业务结果可靠。</div><div class="notion-text notion-block-40770d5879c881d6877bd030624766cb">文章给出第 8 到第 14 条实践，依次涉及：区分部署与晋升、明确配置归属、分层评估、保护诊断数据、衡量有效进展、衡量单次有效产出的成本、演练恢复与交接。最后还有一个发布后的持续改进闭环。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40770d5879c881558a5cc388b90cb65b" data-id="40770d5879c881558a5cc388b90cb65b"><span><div id="40770d5879c881558a5cc388b90cb65b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c881558a5cc388b90cb65b" title="来源事实：托管运行时已经提供了什么"><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">来源事实：托管运行时已经提供了什么</span></span></h3><div class="notion-text notion-block-40770d5879c881b09a8dd98a12632f27">据文章介绍，托管智能体具备独立身份与端点、不可变版本，并默认把 OpenTelemetry 追踪发送到 Application Insights。这让“测试并晋升同一个候选版本”成为可能，也为发布留下证据。编排逻辑、应用权限、依赖和业务结果仍由团队自己负责。</div><div class="notion-text notion-block-40770d5879c881c697e2ea6faef89926">文章还强调，每个区域都应视为一次独立发布。服务版本、模型与工具可用性、生效配置、访问控制和诊断数据边界，都要逐区域核对。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40770d5879c88157bc4cc1b030cbc340"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/3a205e0f-338b-47f0-a092-975eaa01b558/image-1-image.png?table=block&amp;id=40770d58-79c8-8157-bc4c-c1b030cbc340&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=fi9w686hQJLnqQDWlR17npl-Q3rqg3EPEz2F6DDzpm8&amp;t=40770d58-79c8-8157-bc4c-c1b030cbc340" alt="图片来源：Microsoft Foundry Blog原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：Microsoft Foundry Blog<a class="notion-link" href="https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/global-enterprise-multi-agent-systems-in-production-release-and-operations-part-/4564916" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40770d5879c8816e95fad91556795bde" data-id="40770d5879c8816e95fad91556795bde"><span><div id="40770d5879c8816e95fad91556795bde" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c8816e95fad91556795bde" title="部署与晋升要分开"><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">部署与晋升要分开</span></span></h3><div class="notion-text notion-block-40770d5879c881618a6bc2d0bde790d7">这是文章最具操作性的一条。文章指出，默认路由会跟随最新版本，所以要先固定当前服务版本，再创建候选版本，并且只记录一次候选版本号，测试之后不再重新解析“最新”。</div><div class="notion-text notion-block-40770d5879c8817c8b04debba0271484">文中的一个团队用配置服务保存当前生效的版本号：流水线先创建并验证新版本，最后才改这个值，失败时流量留在旧版本。文章还提醒，端点选择器同一时刻只服务一个版本，渐进放量需要网关或独立端点；已有会话也不会随选择器迁移。</div><div class="notion-text notion-block-40770d5879c881e49546cb6ca5f1d565">我的看法是，这套顺序的价值在于把“已合并”“已部署”“已生效”“已复测”拆成不同里程碑，团队可以据此设置明确的晋升门槛。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40770d5879c88128906fff5f7fac480b" data-id="40770d5879c88128906fff5f7fac480b"><span><div id="40770d5879c88128906fff5f7fac480b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c88128906fff5f7fac480b" title="配置归属与评估"><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">配置归属与评估</span></span></h3><div class="notion-text notion-block-40770d5879c881819155f069bd458478">文章认为，门户用户、流水线和安全自动化都可能写入设置，所以每个设置都需要负责人，并要验证从后端行为到最终生效状态的整条链路。基础设施即代码有帮助，但不会自动覆盖所有字段。</div><div class="notion-text notion-block-40770d5879c881929056d2e5795f3c85">评估部分建议分别看智能体、工具和系统结果三个层面，并按用户群体或任务类型拆分结果，避免平均分掩盖小群体的严重问题。文章特别提到，被跳过或不适用的检查不能算通过；智能体说“任务已完成”也不能代替业务系统的回执。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40770d5879c88128ad8eee7625918c8f" data-id="40770d5879c88128ad8eee7625918c8f"><span><div id="40770d5879c88128ad8eee7625918c8f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c88128ad8eee7625918c8f" title="诊断数据：追踪会复制敏感内容"><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">诊断数据：追踪会复制敏感内容</span></span></h3><div class="notion-text notion-block-40770d5879c881e38602f49e562f76e8">追踪会复制提示词、检索内容、工具参数和输出。文章建议盘点所有落点，包括遥测、治理与安全副本、运维副本，并为每处明确用途、归属、访问、保留与删除规则。</div><div class="notion-text notion-block-40770d5879c88108870cf92917f604de">文章还介绍了把敏感内容路由到专用表、再用受保护表控制访问的做法，同时强调路由不等于保护。上线前应先用唯一标记做测试，确认数据能写入，再验证各类访问路径；无法证明保护到位时，应抑制内容采集。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40770d5879c8812cb3b3f73c65d82c84"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/daa46caa-68e4-43a4-8ac1-ed3c97123303/image-2-image.png?table=block&amp;id=40770d58-79c8-812c-b3b3-f73c65d82c84&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=u1MS166wYpnyMphAjditbYj9-OJ92ojouq9epgGhJyM&amp;t=40770d58-79c8-812c-b3b3-f73c65d82c84" alt="图片来源：Microsoft Foundry Blog原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：Microsoft Foundry Blog<a class="notion-link" href="https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/global-enterprise-multi-agent-systems-in-production-release-and-operations-part-/4564916" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40770d5879c8818c9a01f9d7c24d4ee4" data-id="40770d5879c8818c9a01f9d7c24d4ee4"><span><div id="40770d5879c8818c9a01f9d7c24d4ee4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c8818c9a01f9d7c24d4ee4" title="对开发者的影响"><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">对开发者的影响</span></span></h3><ul class="notion-list notion-list-disc notion-block-40770d5879c881cfbbbaee9dd69af2f3"><li>发布清单需要增加版本固定、候选版本验证、晋升审批和回滚材料，不能只看部署状态。</li></ul><ul class="notion-list notion-list-disc notion-block-40770d5879c8817ea93bc3fa0a89b3ab"><li>多区域意味着每个区域都有独立的验证与成本，需要提前规划容量和归属。</li></ul><ul class="notion-list notion-list-disc notion-block-40770d5879c8819eaf7be1ccc292ad86"><li>评估集、阈值和跳过项应进入发布清单，作为版本化证据保存。</li></ul><ul class="notion-list notion-list-disc notion-block-40770d5879c881fa8e74f9dd8e3cce54"><li>可观测性设计要把隐私当作前置条件，而不是事后补丁。</li></ul><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40770d5879c8815a8b93d3e5ebe9421c" data-id="40770d5879c8815a8b93d3e5ebe9421c"><span><div id="40770d5879c8815a8b93d3e5ebe9421c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c8815a8b93d3e5ebe9421c" title="适用边界与待观察"><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">适用边界与待观察</span></span></h3><div class="notion-text notion-block-40770d5879c8817b90f6c24195e157d9">文章明确说明，文中多项能力在撰写时处于预览阶段，并且预览功能不提供服务等级协议，不建议用于生产。因此，团队落地时需要逐项确认当前状态。</div><div class="notion-text notion-block-40770d5879c88150a56eff4c76eb382a">待观察的问题：预览功能何时正式可用；原生的渐进放量能否取代自建网关（这只是推测）；多区域成本与恢复演练在实际团队中会占多大工作量。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40770d5879c881be92dbf2752ad965e5" data-id="40770d5879c881be92dbf2752ad965e5"><span><div id="40770d5879c881be92dbf2752ad965e5" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c881be92dbf2752ad965e5" title="来源"><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">来源</span></span></h3><ul class="notion-list notion-list-disc notion-block-40770d5879c881e781e1d99cb5f0149a"><li>Microsoft Foundry Blog：Global enterprise multi-agent systems in production: Release and operations (Part 2)，https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/global-enterprise-multi-agent-systems-in-production-release-and-operations-part-/4564916</li></ul></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[一天五条消息：AI 的安全责任、可信标识与落地成本]]></title>
            <link>https://lunalight.top//article/ai-daily-2026-10-28-001925</link>
            <guid>https://lunalight.top//article/ai-daily-2026-10-28-001925</guid>
            <pubDate>Wed, 28 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[今天的五条动态分别涉及医疗场景下的 AI 责任争议、生成内容的来源检测、Anthropic 的财务披露、智能体的工程成本，以及世界模型的新进展。它们都指向同一个问题：AI 如何从演示走向可被信任的日常使用。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-40770d5879c8817e94fbcf4dbb39d965"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40770d5879c881e99017eaf650647ae4" data-id="40770d5879c881e99017eaf650647ae4"><span><div id="40770d5879c881e99017eaf650647ae4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c881e99017eaf650647ae4" title="开篇"><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">开篇</span></span></h3><div class="notion-text notion-block-40770d5879c881e5a30ae85cd9804fde">今天的几条消息看似分散，但有一条共同的线索：AI 越深入真实场景，围绕它的问责、验证和成本问题就越具体。下文先写来源里能核实的事实，再写我的判断。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40770d5879c881699926fae13c26b6bb" data-id="40770d5879c881699926fae13c26b6bb"><span><div id="40770d5879c881699926fae13c26b6bb" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c881699926fae13c26b6bb" title="医疗场景里的责任争议"><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">医疗场景里的责任争议</span></span></h3><div class="notion-text notion-block-40770d5879c881b0b816c2e5e8603fb2">据 TechCrunch 报道，一名癫痫患者的家长对 OpenAI 提起诉讼，指控 GPT-4o 的回答影响了她对抗癫痫药物的使用。这目前只是原告的指控，案情细节和因果关系有待法庭认定，我不会把它写成定论。</div><div class="notion-text notion-block-40770d5879c88182a80ef925e69f5402">我的观察是：当用户把模型当作健康顾问，产品的边界提示、对高风险建议的拒答策略和对话记录的可审计性，都会变成开发者必须回答的问题。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40770d5879c88194b775e71829b9f241" data-id="40770d5879c88194b775e71829b9f241"><span><div id="40770d5879c88194b775e71829b9f241" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c88194b775e71829b9f241" title="Gemini 加入来源检测"><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">Gemini 加入来源检测</span></span></h3><div class="notion-text notion-block-40770d5879c881b2865eef92ebd42ef9">Google Gemini 官方账号宣布，用户可以直接在 Gemini 中询问内容是否为 AI 生成，借助 SynthID 检测来检查图片、视频或音频是否出自 Google 的 AI。</div><div class="notion-text notion-block-40770d5879c88147960fc5dc282a678b">需要注意的是，这类检测依赖水印，其结论天然限于 Google 自家模型的产出。没有检出，并不等于内容不是 AI 生成。这是我的分析，不是来源的表述。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40770d5879c881b59db6c70d3a9337cb" data-id="40770d5879c881b59db6c70d3a9337cb"><span><div id="40770d5879c881b59db6c70d3a9337cb" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c881b59db6c70d3a9337cb" title="Anthropic 的 S-1 与商业化信号"><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">Anthropic 的 S-1 与商业化信号</span></span></h3><div class="notion-text notion-block-40770d5879c8813eb561c7187f8be9bf">Azeem Azhar 在文章中解读了 Anthropic 的 S-1 文件，核心判断是公司已具备走向可靠盈利的迹象。文章同时提到增长放缓，以及营收高度依赖新近发布的模型。全文为付费内容，以上细节仅来自页面可见的片段与摘要，详细数字我没有逐项核对。</div><div class="notion-text notion-block-40770d5879c8816a87d0c430a9a912cd">推测：如果营收确实集中在新模型上，模型迭代速度就会直接影响财务表现。这一点值得在后续披露中继续观察。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40770d5879c881d59febd6de6afe1e81" data-id="40770d5879c881d59febd6de6afe1e81"><span><div id="40770d5879c881d59febd6de6afe1e81" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c881d59febd6de6afe1e81" title="LangChain 的设计 agent"><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">LangChain 的设计 agent</span></span></h3><div class="notion-text notion-block-40770d5879c8810da8befc480d522209">LangChain 介绍了内部使用的设计 agent Content Studio，并开放了相关 harness。按摘要所述，它能在 Slack 中根据一条提示生成视频、幻灯片等代码化资产，单次请求的模型成本也有明显下降。</div><div class="notion-text notion-block-40770d5879c8818b96b0f6d2951a5231">对开发者来说，真正可借鉴的是工程化路径：把设计产出表示为代码，再用成本和复用度衡量 agent 是否值得投入。具体成本数字我未能从证据摘录中核实。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40770d5879c881c785fec4249fa39a67" data-id="40770d5879c881c785fec4249fa39a67"><span><div id="40770d5879c881c785fec4249fa39a67" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c881c785fec4249fa39a67" title="Genie 4 与世界模型"><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">Genie 4 与世界模型</span></span></h3><div class="notion-text notion-block-40770d5879c88115b803d1ec0cb7e249">Google DeepMind 发布了世界模型 Genie 4，页面标题称其为新一代世界模拟。摘要还提到实时交互和原生音频生成，但这些细节在证据摘录中没有出现，我暂时把它们当作未核实信息。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40770d5879c881bdb301de94de79a616" data-id="40770d5879c881bdb301de94de79a616"><span><div id="40770d5879c881bdb301de94de79a616" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c881bdb301de94de79a616" title="值得继续观察"><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">值得继续观察</span></span></h3><ul class="notion-list notion-list-disc notion-block-40770d5879c881f9bff9cee242d9d7cf"><li>医疗类诉讼的进展，会如何影响通用聊天产品对健康问题的处理方式。</li></ul><ul class="notion-list notion-list-disc notion-block-40770d5879c881de98deef3f12c253c0"><li>来源检测能否走向跨厂商的统一标准。</li></ul><ul class="notion-list notion-list-disc notion-block-40770d5879c88121a91ac7503309b3b3"><li>Anthropic 正式披露后，财务数据与文章解读是否一致。</li></ul><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-40770d5879c8815ebe2fd8897dd44c84" data-id="40770d5879c8815ebe2fd8897dd44c84"><span><div id="40770d5879c8815ebe2fd8897dd44c84" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40770d5879c8815ebe2fd8897dd44c84" title="来源清单"><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">来源清单</span></span></h3><ul class="notion-list notion-list-disc notion-block-40770d5879c88111a0aed94ef77f10ce"><li>https://techcrunch.com/2026/10/27/parents-sue-openai-alleging-gpt-4o-coached-their-daughter-off-seizure-medication/</li></ul><ul class="notion-list notion-list-disc notion-block-40770d5879c881d29e9ccb0dd6cc0afc"><li>https://x.com/GeminiApp/status/2115171788812669225</li></ul><ul class="notion-list notion-list-disc notion-block-40770d5879c881478404fad1cfbfbb0d"><li>https://www.exponentialview.co/p/anthropics-s-1-in-charts</li></ul><ul class="notion-list notion-list-disc notion-block-40770d5879c881ecbe76e202e5b772e3"><li>https://www.langchain.com/blog/design-agent</li></ul><ul class="notion-list notion-list-disc notion-block-40770d5879c881aabbdbd2d7b6dc6165"><li>https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/</li></ul></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[给生产级智能体做记忆：写入、读取、复核与遗忘要分开治理]]></title>
            <link>https://lunalight.top//article/azure-foundry-2026-10-27-4564392</link>
            <guid>https://lunalight.top//article/azure-foundry-2026-10-27-4564392</guid>
            <pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Microsoft Foundry 博客的这篇文章主张，智能体记忆应拆成四条由应用策略层把关的独立路径，而不是一个更大的向量库。本文梳理其要点，并从开发者角度做了分析。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-40670d5879c881d6b400f0a24ed89128"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-40670d5879c881a9ac1ff726524be714" data-id="40670d5879c881a9ac1ff726524be714"><span><div id="40670d5879c881a9ac1ff726524be714" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c881a9ac1ff726524be714" title="智能体记忆是一套治理架构"><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">智能体记忆是一套治理架构</span></span></h2><div class="notion-text notion-block-40670d5879c881b6a5a0f8b30cef0075">写入、读取、复核、遗忘四条路径，各有责任方</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c881d5a357c116e4ab1d79" data-id="40670d5879c881d5a357c116e4ab1d79"><span><div id="40670d5879c881d5a357c116e4ab1d79" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c881d5a357c116e4ab1d79" title="文章在讨论什么问题"><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">文章在讨论什么问题</span></span></h3><div class="notion-text notion-block-40670d5879c88106b0afe35fafaa0228">Microsoft Foundry 博客在 2026 年 10 月 26 日发表了一篇文章，讨论生产环境里的智能体该如何管理记忆。作者的核心判断是：记忆能力不能只靠更好的提示词或更大的向量库。它应当拆成四项相互配合的能力，即 Remember（记住）、Retrieve（检索）、Revalidate（复核）和 Forget（遗忘）。</div><div class="notion-text notion-block-40670d5879c88158bc56c8a57e462963">这四项能力由智能体、应用策略层、平台记忆服务以及各自拥有数据的业务系统分担。每一项都有自己的身份校验、策略关卡、失败处理和审计记录。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c88156bc10e74e18605544" data-id="40670d5879c88156bc10e74e18605544"><span><div id="40670d5879c88156bc10e74e18605544" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c88156bc10e74e18605544" title="文章用什么例子说明风险"><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">文章用什么例子说明风险</span></span></h3><div class="notion-text notion-block-40670d5879c8813fbeb0dbf98884f973">文中设想了一个发布协调场景。面向用户的编排智能体记住了负责人的汇报格式偏好，这类内容适合作为用户级长期记忆。但项目日期应由另一个有权读取项目系统的数据智能体提供。</div><div class="notion-text notion-block-40670d5879c881f7971dfb1fc6981a08">如果编排智能体把用户曾说过的某个日期存成了长期事实，而权威系统之后已经改期，那么回答的格式可能很规整，内容却是错的。作者的结论是：偏好可以记住，日期应当检索并回到权威系统复核，过期的副本应当被清除。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40670d5879c88145a81cc4231c03cbb6"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/e7d3a32e-a801-4915-bfc1-fa63087040ae/image-1-image.png?table=block&amp;id=40670d58-79c8-8145-a81c-c4231c03cbb6&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=X8B9qSODphoBOMBHCKYPIkLdKDtijvJRbsx1yTb5BV8&amp;t=40670d58-79c8-8145-a81c-c4231c03cbb6" alt="图片来源：Microsoft Foundry Blog原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：Microsoft Foundry Blog<a class="notion-link" href="https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/remember-retrieve-revalidate-or-forget-architecting-memory-for-production-ai-age/4564392" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c88186b8dde6d52cd951dc" data-id="40670d5879c88186b8dde6d52cd951dc"><span><div id="40670d5879c88186b8dde6d52cd951dc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c88186b8dde6d52cd951dc" title="四条路径的分工"><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">四条路径的分工</span></span></h3><div class="notion-text notion-block-40670d5879c881819e00d8641010041b">文章用同一套契约描述四项操作，包括触发条件、决策方、各层职责、身份授权、来源记录、失败行为和审计。我的理解是，这套做法的重点是“谁有权决定”，而不是“模型能不能做到”。</div><ul class="notion-list notion-list-disc notion-block-40670d5879c8811c9fcde025c3bf86b8"><li>记住：智能体只能提名候选事实，不能把模型输出直接写进持久记忆。是否写入由应用层的写入关卡决定。</li></ul><ul class="notion-list notion-list-disc notion-block-40670d5879c881a68868e623daafe1b8"><li>检索：智能体只声明信息需求。文章强调先授权再排序，因为语义相似只代表相关，不代表有权访问。</li></ul><ul class="notion-list notion-list-disc notion-block-40670d5879c8817caf95dc21d0ace599"><li>复核：当任务需要当前事实时，由策略决定新鲜度要求，再向权威系统核对版本或标记。</li></ul><ul class="notion-list notion-list-disc notion-block-40670d5879c8811daf0dc296fb1ddcfa"><li>遗忘：按保留、数据主体请求或事故处置策略执行，需要覆盖索引、衍生摘要和缓存，并验证结果。</li></ul><div class="notion-text notion-block-40670d5879c88123927cd259553f34ea">文章还提到，Microsoft 的智能体追踪规范已经把创建、搜索、更新、写入或更新、删除记忆定义为可单独观测的操作。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c88179acc0c1ba8a5a2aa0" data-id="40670d5879c88179acc0c1ba8a5a2aa0"><span><div id="40670d5879c88179acc0c1ba8a5a2aa0" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c88179acc0c1ba8a5a2aa0" title="写入与读取的实现要点"><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">写入与读取的实现要点</span></span></h3><div class="notion-text notion-block-40670d5879c8812dbe38d3e3018aa361">在写入侧，文章给出了一个与具体 API 无关的伪代码关卡。它按来源与敏感度分类，丢弃凭据、失败工具调用的输出和低置信推断，把不可信内容隔离，并在涉及敏感信息时要求用户同意。通过之后才做范围解析、抽取、合并去重，并带着有效期和删除谱系写入存储。</div><div class="notion-text notion-block-40670d5879c881e08f67fb24f5571e03">文章还列出了一份受治理记忆记录应有的字段，例如租户、所有者、来源、置信度、有效期、被取代关系和保留类别。在读取侧，伪代码先按身份缩小候选集，再混合检索、排序，并剔除缺少来源或已被取代的条目。必要时再向权威系统复核，最后只组装最小上下文并附上引用。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40670d5879c881a584bac8345728b745"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/7a41ff7c-06f6-46ef-a2b2-e10cab8661fc/image-2-image.png?table=block&amp;id=40670d58-79c8-81a5-84ba-c8345728b745&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=6Z8Yfw1NMDcrhBCW9yGVmD4JSjTu39T-1Z4-KDJWbkQ&amp;t=40670d58-79c8-81a5-84ba-c8345728b745" alt="图片来源：Microsoft Foundry Blog原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：Microsoft Foundry Blog<a class="notion-link" href="https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/remember-retrieve-revalidate-or-forget-architecting-memory-for-production-ai-age/4564392" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c88169b980f68d77eb8861" data-id="40670d5879c88169b980f68d77eb8861"><span><div id="40670d5879c88169b980f68d77eb8861" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c88169b980f68d77eb8861" title="对开发者的影响"><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">对开发者的影响</span></span></h3><div class="notion-text notion-block-40670d5879c8812ea7fde38df76c15f4">这是我的分析，不是文章原话。</div><div class="notion-text notion-block-40670d5879c881b38b1ffac53622bd20">1. 记忆应当作为一个有独立权限与生命周期的服务来设计，不要把它当成提示词的附属品。 2. 多租户场景下，授权过滤要放在检索之前。共享向量库如果缺少过滤，即使结果很相关，也构成边界故障。 3. 日期、状态、价格这类会变化的事实，更适合每次查询权威系统，而不是缓存为长期记忆。 4. 删除需求要在设计之初就纳入，因为只删主记录可能遗留嵌入、摘要和缓存。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c88116b1d1f6b409acdb38" data-id="40670d5879c88116b1d1f6b409acdb38"><span><div id="40670d5879c88116b1d1f6b409acdb38" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c88116b1d1f6b409acdb38" title="仍待观察之处"><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">仍待观察之处</span></span></h3><ul class="notion-list notion-list-disc notion-block-40670d5879c881fdb55ec771b3ddb185"><li>我读到的文章内容在检索章节处中断，复核与遗忘两节的具体实现、跨范围与多智能体的应用，以及治理部分没有核对，上文对这些部分只做了概括。</li></ul><ul class="notion-list notion-list-disc notion-block-40670d5879c881509fd6e6888d6c6d5b"><li>文章给出的是厂商中立的架构模式。这些路径在 Foundry 产品中的具体对应与成熟度，需要对照官方文档确认。</li></ul><ul class="notion-list notion-list-disc notion-block-40670d5879c8814ba607fb2631904fb3"><li>增加复核与审计环节会带来延迟和成本。推测：高频、低风险的场景可能需要分级策略，这一点文章是否展开，我没有核实。</li></ul><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c88180913aea976a2b0732" data-id="40670d5879c88180913aea976a2b0732"><span><div id="40670d5879c88180913aea976a2b0732" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c88180913aea976a2b0732" title="来源"><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">来源</span></span></h3><ul class="notion-list notion-list-disc notion-block-40670d5879c8812d81f4e87599dd5ad0"><li>Microsoft Foundry Blog：Remember, Retrieve, Revalidate-or Forget? Architecting Memory for Production AI Agents（链接见来源清单）。</li></ul></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[本周三条 AI 动态：科学证明、电脑操作与演示文稿智能体]]></title>
            <link>https://lunalight.top//article/ai-daily-2026-10-27-001929</link>
            <guid>https://lunalight.top//article/ai-daily-2026-10-27-001929</guid>
            <pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[OpenAI 称模型在数学、计算机科学与物理领域取得三项突破，GPT-6.1 Terra 和 Gamma 5 则把智能体推向日常操作与内容制作。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-40670d5879c881f3a744ea1a44ce2d7a"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-40670d5879c881c091acca4517850e87" data-id="40670d5879c881c091acca4517850e87"><span><div id="40670d5879c881c091acca4517850e87" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c881c091acca4517850e87" title="从证明定理到操作电脑"><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">从证明定理到操作电脑</span></span></h2><div class="notion-text notion-block-40670d5879c881858b8df53352cd65d4">三条消息里的智能体与推理能力</div><div class="notion-text notion-block-40670d5879c88103b673f96ae7c14b46">今天值得关注的三条消息有一个共同点：模型正在从回答问题走向独立完成任务。下面先写来源中的事实，再写我的判断。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c88134a675f45ea6cd136b" data-id="40670d5879c88134a675f45ea6cd136b"><span><div id="40670d5879c88134a675f45ea6cd136b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c88134a675f45ea6cd136b" title="OpenAI：三项科学突破的宣称"><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">OpenAI：三项科学突破的宣称</span></span></h3><div class="notion-text notion-block-40670d5879c8810e8ed4ea0a4234fe5f">据 IT之家报道，OpenAI 宣布其 AI 模型在数学、计算机科学和物理学领域取得三项突破。数学方面，模型证明了适用于所有维度的局部光滑猜想，并据此解决了局部光滑猜想与挂谷猜想。计算机科学方面，模型推翻了指数时间假设，这意味着未来可能出现更快的算法。物理方面，涉及量子色动力学与杨–米尔斯理论的严格数学存在性。</div><div class="notion-text notion-block-40670d5879c881918788d30582416692">我的看法：这些都是 OpenAI 的宣布，我没有看到独立的数学界验证。形式化证明与同行评议通常需要时间，因此在验证完成前，应当把它们当作待检验的结论，而不是定论。</div><blockquote class="notion-quote notion-block-40670d5879c881c7bdb0d723cd538286"><div>以上结论目前仅来自 OpenAI 的宣布及媒体转述，尚未见到第三方验证。</div></blockquote><div class="notion-text notion-block-40670d5879c88173b66cc3944771837f">原文链接：https://www.ithome.com/1/018/203.htm</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c88181bc7eefb336a23cd2" data-id="40670d5879c88181bc7eefb336a23cd2"><span><div id="40670d5879c88181bc7eefb336a23cd2" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c88181bc7eefb336a23cd2" title="GPT-6.1 Terra：电脑操作的速度门槛"><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">GPT-6.1 Terra：电脑操作的速度门槛</span></span></h3><div class="notion-text notion-block-40670d5879c881d4a2f0cae2a8255336">Greg Brockman 在 X 上表示，Terra 的电脑操作能力已跨过一个质变门槛，他感觉它比自己不借助 AI 时用电脑更快。他引用的 OpenAI 介绍称，GPT-6.1 Terra 以更低价格提供快速且强大的 computer use，能够理解屏幕内容并跨应用执行操作。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40670d5879c8817eb8eecfd7f3377655"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/4b319500-b973-4843-9ef9-93896265a7bd/image-1-image.webp?table=block&amp;id=40670d58-79c8-817e-b8ee-cfd7f3377655&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=4J10WFio-fjZt5QY9XMxD_Pbme00W8n04-0YWiB5LQM&amp;t=40670d58-79c8-817e-b8ee-cfd7f3377655" alt="图片来源：Greg Brockman on X: &quot;computer use with terra has crossed a qualitative threshold where it feels faster at using a computer than what i can do unassisted&quot; / X原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：Greg Brockman on X: &quot;computer use with terra has crossed a qualitative threshold where it feels faster at using a computer than what i can do unassisted&quot; / X<a class="notion-link" href="https://x.com/gdb/status/2114814747254432243" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><div class="notion-text notion-block-40670d5879c88110a706d67d34730ee4">我的看法：这是个人的主观体验，不是基准测试结果。对开发者而言，真正的问题是在真实业务流程中的稳定性、权限边界与出错后的可恢复性。速度只是其中一项指标。</div><div class="notion-text notion-block-40670d5879c881f6af39dceb49f8022c">原文链接：https://x.com/gdb/status/2114814747254432243</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c881edb1d0edcb100a582d" data-id="40670d5879c881edb1d0edcb100a582d"><span><div id="40670d5879c881edb1d0edcb100a582d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c881edb1d0edcb100a582d" title="Gamma 5：把智能体放进演示文稿流程"><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">Gamma 5：把智能体放进演示文稿流程</span></span></h3><div class="notion-text notion-block-40670d5879c8816fbca2edf99ccb9cc5">Gamma 在更新日志中介绍了 Gamma 5，称其重建了思考、设计和编辑的方式。用户可以提供一句话、文档、PDF、图片或旧演示文稿，由智能体在几分钟内完成构思、调研和设计。日志还提到可以引入 PDF 和 PPT 设计以便持续使用。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-40670d5879c8810b8f43dbedbadada45"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img src="https://file.notion.com/f/f/9394689f-e281-41a6-a46a-d163dff3e5d8/70d23f1f-fef9-416e-9088-1f74d747b02b/image-2-image.jpg?table=block&amp;id=40670d58-79c8-810b-8f43-dbedbadada45&amp;spaceId=9394689f-e281-41a6-a46a-d163dff3e5d8&amp;expirationTimestamp=1793268000000&amp;signature=F90sz-MyDZUmp4MxbUI8tjFU1iQJx0iqCkUd7-yXcWU&amp;t=40670d58-79c8-810b-8f43-dbedbadada45" alt="图片来源：What’s new in Gamma | September–October 2026 | Gamma Changelog Facebook X (formerly Twitter)原文链接" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">图片来源：What’s new in Gamma | September–October 2026 | Gamma Changelog Facebook X (formerly Twitter)<a class="notion-link" href="https://meetgamma.canny.io/changelog/whats-new-in-gamma-septemberoctober-2026" target="_blank" rel="noopener noreferrer">原文链接</a></figcaption></div></figure><div class="notion-text notion-block-40670d5879c88149b67afcc0ade3b2ce">我的看法：输入形式越宽松，质量就越取决于调研环节是否可靠。生成的内容仍需要人工核对事实和引用，这一点与上面两条消息一样。</div><div class="notion-text notion-block-40670d5879c8812e9d99da0cc540887d">原文链接：https://meetgamma.canny.io/changelog/whats-new-in-gamma-septemberoctober-2026</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c881e49d01ccfcb253844c" data-id="40670d5879c881e49d01ccfcb253844c"><span><div id="40670d5879c881e49d01ccfcb253844c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c881e49d01ccfcb253844c" title="值得继续观察"><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">值得继续观察</span></span></h3><div class="notion-text notion-block-40670d5879c881f09efff3c8935e4d0e">第一，OpenAI 的三项证明能否得到独立验证。第二，Terra 的电脑操作在多步骤任务中是否同样稳定。第三，推测而言，智能体类产品的竞争重点可能从演示效果转向可审计性与可控性。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-40670d5879c88180b164e3f58711d025" data-id="40670d5879c88180b164e3f58711d025"><span><div id="40670d5879c88180b164e3f58711d025" class="notion-header-anchor"></div><a class="notion-hash-link" href="#40670d5879c88180b164e3f58711d025" title="来源清单"><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">来源清单</span></span></h3><ul class="notion-list notion-list-disc notion-block-40670d5879c8813dad5bdf416bd8924a"><li>https://www.ithome.com/1/018/203.htm</li></ul><ul class="notion-list notion-list-disc notion-block-40670d5879c8813691cec235ea27b417"><li>https://x.com/gdb/status/2114814747254432243</li></ul><ul class="notion-list notion-list-disc notion-block-40670d5879c8814aa4ace49e45ba284e"><li>https://meetgamma.canny.io/changelog/whats-new-in-gamma-septemberoctober-2026</li></ul></main></div>]]></content:encoded>
        </item>
    </channel>
</rss>