<?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>SkiLlre 博客</title>
        <link>https://note.skillre.cn/</link>
        <description>做一个勤奋的博客</description>
        <lastBuildDate>Fri, 24 Apr 2026 10:09:05 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, SkiLlre</copyright>
        <item>
            <title><![CDATA[DeepSeek V4 发布快报：1M 上下文、Agent 能力与颠覆性定价]]></title>
            <link>https://note.skillre.cn/technique/2026/04/deepseek-v4-report</link>
            <guid>https://note.skillre.cn/technique/2026/04/deepseek-v4-report</guid>
            <pubDate>Fri, 24 Apr 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[DeepSeek 于 2026年4月24日发布 V4 预览版，双模型同步开源。V4-Pro（1.6T参数）与 V4-Flash（284B参数）均标配 1M 上下文窗口，CSA+HCA 混合注意力架构实现性能突破。Codeforces 编程竞赛 3206 分超越 GPT-5.4，API 定价低至 Claude 的 1/90。本文从性能、定价、架构、国产算力等维度全面解读 V4 发布。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-34cd9fb33fdd81fb9225d6b5b1b546e4"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-2375c92ca7e34f298915b3e52274fe3f" data-id="2375c92ca7e34f298915b3e52274fe3f"><span><div id="2375c92ca7e34f298915b3e52274fe3f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2375c92ca7e34f298915b3e52274fe3f" title="DeepSeek V4 发布：1M 上下文成标配，价格低至 Claude 的 1/90"><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">DeepSeek V4 发布：1M 上下文成标配，价格低至 Claude 的 1/90</span></span></h2><div class="notion-text notion-block-d896c11922d847418ea1e7a6d0a3c9ad">2026年4月24日，DeepSeek 正式发布 V4 系列模型预览版。这是继 V3 之后最大的一次版本迭代，也是 R1 之后最受社区期待的发布。V4 的核心策略可以概括为三个关键词：<b>性能对标闭源旗舰、上下文直接拉满到 1M、价格打到行业地板价</b>。</div><div class="notion-text notion-block-aaf10072fa88446f8f77f79b54325981">更值得关注的是，V4 放弃了单一路线，转而采用 V4-Pro（旗舰）和 V4-Flash（轻量）双版本策略——类似 Anthropic 的 Opus/Sonnet 路线，用 Pro 打性能天花板，用 Flash 覆盖日常开发场景。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-75af839fe0394ea5979abc8445bcc5a5" data-id="75af839fe0394ea5979abc8445bcc5a5"><span><div id="75af839fe0394ea5979abc8445bcc5a5" class="notion-header-anchor"></div><a class="notion-hash-link" href="#75af839fe0394ea5979abc8445bcc5a5" title="双版本策略：各有侧重"><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">双版本策略：各有侧重</span></span></h3><div class="notion-text notion-block-ec4cf348cb554927a982a2478b59b44b">V4-Pro 采用 1.6T 总参数、49B 激活参数的 MoE 架构，预训练数据量 33T，配备 Non-Think 直出 / Think High 常规思考 / Think Max 最大深度思考三档推理强度，目标直指 GPT-5.4 Pro 和 Claude Opus 4.x。而 V4-Flash 是 284B 总参数、13B 激活参数的经济版本，主打低成本高性能。</div><div class="notion-text notion-block-4541a0562c7643158b9eea2bfdc3676c">但真正让社区意外的是：<b>两个版本都标配 1M token 上下文窗口</b>。在此之前，1M 级别上下文长期是 Google Gemini 的独占优势，行业主流还停留在 128K-256K。DeepSeek 直接把百万上下文从「高端选配」打成了「基础标配」。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-a5e5fdb9ebfe4d0a8c13c21107e08fd8" data-id="a5e5fdb9ebfe4d0a8c13c21107e08fd8"><span><div id="a5e5fdb9ebfe4d0a8c13c21107e08fd8" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a5e5fdb9ebfe4d0a8c13c21107e08fd8" title="性能实测：代码能力突出，Agent 场景亮眼"><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">性能实测：代码能力突出，Agent 场景亮眼</span></span></h3><div class="notion-text notion-block-f856a727dc6c423fbc48d5d58ec66ff0">V4 在代码生成和工程能力方面表现尤为突出，多个权威基准测试中展现出与顶级闭源模型正面竞争的实力：</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-9c31a446dafb4c22bdbd3090c456a708"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/deepseek-v4-release/v4-benchmark-comparison.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=9c31a446-dafb-4c22-bdbd-3090c456a708" alt="V4 Benchmark 对比图" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">V4 Benchmark 对比图</figcaption></div></figure><div class="notion-text notion-block-96ce5c05f1d74e9c862f6d03bdcc13be">在 <b>Codeforces</b> 编程竞赛上，V4-Pro-Max 以 3206 分超越 GPT-5.4 的 3168。这意味着它在算法竞赛型代码生成上达到了人类顶尖选手水平。而在更贴近实际开发的 <b>Apex Shortlist</b> 全栈代码生成测试中，V4 达到 90.2%，领先 Claude Opus 4.6 的 85.9%。</div><div class="notion-text notion-block-2e40a91ba38f45feb8349482c1a72070">不过，在解决真实软件工程问题（<b>SWE-Verified</b>）上，V4 的 80.6% 与 Claude Opus 4.6 的 80.8% 基本持平——这反映了两者的不同优化方向：DeepSeek 在算法竞赛型代码上更强，而 Claude 在真实场景的系统性工程能力上仍是标杆。<b>Terminal Bench</b> 命令行操作测试中，V4 的 67.9% 也与 GPT-5.4 的 68.5% 处于同一梯队。</div><div class="notion-text notion-block-df962a4fc2854cb2b312f5c216d402e1">V4 在 <b>Agent 工作流</b> 中的表现同样值得一提。Toolathlon 测试中 V4-Pro-Max 拿到 51.8%，超过 Claude Opus 4.6 的 47.2%。DeepSeek 表示已在内部用 V4 替换了 Claude 进行实际编码工作，工具调用格式从 JSON 切换为带特殊 token 的 XML 结构以降低转义错误，跨轮次推理痕迹也在长时间 Agent 任务中完整保留。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-eb2317d57e51443d9a00e44f13525ff8" data-id="eb2317d57e51443d9a00e44f13525ff8"><span><div id="eb2317d57e51443d9a00e44f13525ff8" class="notion-header-anchor"></div><a class="notion-hash-link" href="#eb2317d57e51443d9a00e44f13525ff8" title="架构创新：1M 上下文的秘密"><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">架构创新：1M 上下文的秘密</span></span></h3><div class="notion-text notion-block-d887a3cdab034bddb0cfdb3ac83bfa53">V4 能做到 1M 上下文而不牺牲推理速度，关键在于 <b>CSA + HCA 混合注意力机制</b>。CSA 解决「算什么」的问题——用轻量级索引器先对所有 token 对做粗筛，只精选出需要完整计算的部分；HCA 解决「存什么」的问题——在 MLA 基础上继续把 KV 向量压缩到低维潜空间，推理时再解压。</div><div class="notion-text notion-block-3b319d1f452545f987414b908e3c0faf">两个数字说明实际效果：在 1M token 场景下，V4-Pro 的单个 token 推理 FLOPs 仅为 V3.2 的 <b>27%</b>，KV Cache 占用仅为 <b>10%</b>。这意味着同等算力下能服务的长上下文并发量约为原来的 3-4 倍。</div><div class="notion-text notion-block-8d23db8b1b08447bbb870bc13f9c3005">延续自 V3 的 <b>Multi-Token Prediction</b> 技术和 <b>FP8 混合精度训练</b>，让 V4-Flash 达到了 <b>195M tokens/s</b> 的推理速度。这也是 V4-Flash 能以 13B 激活参数的规模在多项测试中追平更大参数模型的原因。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-3d3467bf94e34cd0a59072eb46b7bd42" data-id="3d3467bf94e34cd0a59072eb46b7bd42"><span><div id="3d3467bf94e34cd0a59072eb46b7bd42" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3d3467bf94e34cd0a59072eb46b7bd42" title="定价：重新定义 AI 服务的价格基准"><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">定价：重新定义 AI 服务的价格基准</span></span></h3><div class="notion-text notion-block-72abfa95522a43c1afe2bb6936d0426b">如果说性能是惊喜，那定价就是地震：</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-0cc2b7cd80d14f579d6ed1fe075cccab"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/deepseek-v4-release/api-price-comparison.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=0cc2b7cd-80d1-4f57-9d6e-d1fe075cccab" alt="API 价格对比图" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">API 价格对比图</figcaption></div></figure><div class="notion-text notion-block-03ef47f513a0478fb4d9aa1c9737f0fd">V4-Flash 每百万输出 token 仅 <b>2 元</b>，V4-Pro 为 <b>24 元</b>。横向对比，Claude Opus 4.7 约 180 元，GPT-5.4 Pro 约 216 元。</div><div class="notion-text notion-block-46a134a705974e98819573aa3bc4534d">2 元/百万输出 token 是什么概念？一次中等规模的代码审查（约 2000 token 输出）成本不到半分钱。对于独立开发者来说，这意味着可以把日常编码任务放心交给 AI 而不用担心账单；对于创业团队来说，这意味着以传统 API 成本的零头接入顶级推理能力。</div><div class="notion-text notion-block-c83a489065904fa18b2ec783b0077119"><b>这个定价策略释放了一个明确的信号：DeepSeek 不是在和闭源模型比价格，而是在重新定义 AI 服务的价格基准。</b></div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-11cc0af12bed434997f71f60340e9b14" data-id="11cc0af12bed434997f71f60340e9b14"><span><div id="11cc0af12bed434997f71f60340e9b14" class="notion-header-anchor"></div><a class="notion-hash-link" href="#11cc0af12bed434997f71f60340e9b14" title="国产算力：华为昇腾首次进入验证清单"><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">国产算力：华为昇腾首次进入验证清单</span></span></h3><div class="notion-text notion-block-c726effd7efa4c879807405d14c2abc0">技术报告第 3.1 节明确写道：「我们在英伟达 GPU 和华为昇腾 NPU 两个平台上均验证了细粒度 EP（专家并行）方案。」这是 DeepSeek <b>首次</b>在正式技术文档中将华为昇腾与英伟达并列写入硬件验证清单。</div><div class="notion-text notion-block-995cf57ba1bc409bb41c8aa8fc24be47">更值得注意的是，V4 的 MoE 专家权重和稀疏注意力索引器采用 <b>FP4 精度</b>——恰好是华为昇腾 950PR 芯片的原生支持精度。官方透露，预计下半年昇腾 950 节点批量上市后，Pro 版价格会<b>大幅下调</b>。寒武纪也已基于 vLLM 完成 Day 0 适配，代码已开源到 GitHub。</div><div class="notion-text notion-block-17b7c808060a443a8e2bb944a1959c91">在推理场景中，V4 采用 INT4/INT8 量化加国产芯片优化的组合方案，这意味着中国开发者可以在不依赖 Nvidia GPU 的情况下使用本地部署的 V4 模型进行推理。虽然训练环节仍然依赖 Nvidia Hopper 架构，但推理端的国产化已迈出实质性一步。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-3afecb5ca7b34c9fa8b3a3cc42acdb52" data-id="3afecb5ca7b34c9fa8b3a3cc42acdb52"><span><div id="3afecb5ca7b34c9fa8b3a3cc42acdb52" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3afecb5ca7b34c9fa8b3a3cc42acdb52" title="快速上手"><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">快速上手</span></span></h3><div class="notion-text notion-block-23d53b7e48b44ea197b2200aa0a71b7f">通过 DeepSeek 官方 API 即可接入，同时支持 OpenAI ChatCompletions 和 Anthropic 两套接口标准：</div><div class="notion-text notion-block-de20bcbd755a4eb5b6da7a78f319bc16"><b>采样参数建议</b>：temperature = 1.0，top_p = 1.0（思考模式下）。如果要在 Claude Code 等 Agent 工具中使用，已原生适配，直接切换 endpoint 即可。开源权重同步上架 Hugging Face 和 ModelScope，MIT 许可证允许商业使用。</div><div class="notion-text notion-block-a6babfa4012c4423a11e1c0eb4884263"><b>一句话推荐</b>：</div><ul class="notion-list notion-list-disc notion-block-b20dc48885b744e194336d6d6e66e8b6"><li>日常编码辅助、代码审查 → <b>V4-Flash</b>（够用，便宜，不心疼）</li></ul><ul class="notion-list notion-list-disc notion-block-d9113c0872d64e23af14e002392abe05"><li>复杂推理、长文分析、Agent 任务 → <b>V4-Pro</b>（性能对标旗舰，价格不到 1/7）</li></ul><ul class="notion-list notion-list-disc notion-block-3629cdca19c843808dcdac6a4e52ef76"><li>全部接上 → 反正成本是闭源模型的零头</li></ul><div class="notion-text notion-block-b451805352984ce1a619e99c23a6422e"><b>⚠️ 迁移提醒</b>：<code class="notion-inline-code">deepseek-chat</code> 和 <code class="notion-inline-code">deepseek-reasoner</code> 旧接口将于 <b>2026年7月24日</b> 停用，生产环境需在三个月内完成迁移，个人开发者只需改一个 model 参数。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-48f756b1213e4875b1222352e18c743d" data-id="48f756b1213e4875b1222352e18c743d"><span><div id="48f756b1213e4875b1222352e18c743d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#48f756b1213e4875b1222352e18c743d" title="对开发者的建议"><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">对开发者的建议</span></span></h3><div class="notion-text notion-block-370d8c91163e4b61a2dc5668d9d42348">DeepSeek V4 的发布，在多个维度上改写了开源模型的竞争格局。三个值得关注的趋势变化：</div><div class="notion-text notion-block-046bd80bf19145c3aced339c7c34bc83"><b>第一，开源模型在基准测试上正面对抗闭源旗舰已经成为现实。</b> 尤其是在代码生成和数学推理领域，V4 已经做到与 GPT-5.4、Claude Opus 同台竞技。</div><div class="notion-text notion-block-44510f348e2b42efb51adc83256c4612"><b>第二，AI 服务的价格基准在快速下移。</b> V4-Flash 2 元的定价不是简单的低价策略，而是在证明「足够好且极便宜」是一条可行的产品路线。未来半年，可以预期闭源模型也会跟进降价。</div><div class="notion-text notion-block-840595f33f3d451da0ec7eea794741a3"><b>第三，国产 AI 芯片生态正在形成。</b> 虽然训练环节仍然依赖 Nvidia，但推理侧的国产化替代已经在真实产品中得到验证。</div><div class="notion-text notion-block-1c097c70e01f4030b750c7ff06ee9bdf">如果你是独立开发者或中小团队，现在可能是时候认真考虑将工作流中的部分任务迁移到 V4 上了——不是因为它是「最好的」，而是因为它是「足够好且便宜到可以放心用」的那个选项。</div></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Anthropic Mythos：AI 安全领域的新范式与行业变革]]></title>
            <link>https://note.skillre.cn/technique/2026/04/anthropic-mythos-ai-security-paradigm</link>
            <guid>https://note.skillre.cn/technique/2026/04/anthropic-mythos-ai-security-paradigm</guid>
            <pubDate>Thu, 23 Apr 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[Anthropic 发布 Mythos 超前沿模型，具备自主漏洞发现能力。文章深入分析其 Recurrent-Depth Transformer 架构、SWE-bench 93.9% 基准成绩、与 Claude 系列的本质差异、Project Glasswing 合作网络、竞争对手布局，以及 CB-1 级别双重用途风险对网络安全行业的范式转变影响。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-34bd9fb33fdd8157abcfe3b920468534"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-46e04193c2ce488faf0711fc28ea1416"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/mythos-analysis/anthropic-model-hierarchy.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=46e04193-c2ce-488f-af07-11fc28ea1416" alt="Anthropic 模型层级结构" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">Anthropic 模型层级结构</figcaption></div></figure><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-21e17d9efe2b4f4486b5446bef7c8192" data-id="21e17d9efe2b4f4486b5446bef7c8192"><span><div id="21e17d9efe2b4f4486b5446bef7c8192" class="notion-header-anchor"></div><a class="notion-hash-link" href="#21e17d9efe2b4f4486b5446bef7c8192" title="一、产品定位与核心突破"><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">一、产品定位与核心突破</span></span></h3><div class="notion-text notion-block-35a3b1eefc8846f4ace8f119846f23bf">Anthropic 在 2026 年 4 月推出了名为 Mythos 的全新前沿模型，这不是 Claude 系列的简单迭代，而是定位在&quot;超前沿&quot;（Capybara）层级的全新产品线。根据 Anthropic 的模型层级划分，Mythos 位于 Claude Opus 之上，专门针对高风险安全场景设计。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-38b2d59885a74c2b976815327267c7dc" data-id="38b2d59885a74c2b976815327267c7dc"><span><div id="38b2d59885a74c2b976815327267c7dc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#38b2d59885a74c2b976815327267c7dc" title="漏洞发现能力的量级跃迁"><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">漏洞发现能力的量级跃迁</span></span></h4><div class="notion-text notion-block-a770cf659a164c7398f38d0961d8bf34">Mythos 最引人注目的突破在于其自主漏洞发现能力。在 Project Glasswing 的预览测试中，Mythos 成功发现了 OpenBSD 中存在 27 年的老旧漏洞、FFmpeg 中 16 年未被发现的安全缺陷，以及 Linux 内核中的权限提升链式漏洞。这些发现不是简单的代码扫描结果，而是完整的攻击链路——从漏洞定位到可运行的攻击代码，全部由模型自动完成。</div><div class="notion-text notion-block-8a15f79d497d42fd9a64410409176449">根据 Palo Alto Networks 的实测报告，前沿 AI 模型在不到三周的时间内，完成了相当于整整一年渗透测试工作的成果。相比 Anthropic 之前的领先模型，Mythos 在编码效率上提升了约 50%，这种代际提升直接转化为漏洞发现和利用生成能力的显著进步。</div><div class="notion-text notion-block-1ccd3b158de64a2c8aef9681ec82bf14">更令人印象深刻的是 Mythos 在漏洞链构建方面的能力。它能够将多个低严重性问题组合为关键级别的利用路径。例如，将两个中等严重性漏洞和一个低严重性漏洞链接成一个关键级利用。这种&quot;全栈逻辑分析&quot;能力让它可以分析应用程序的完整暴露面，包括 SaaS 和面向公众的平台，识别传统工具容易遗漏的基于逻辑的漏洞。</div><hr class="notion-hr notion-block-f279037f5e7d4885b2f71e798c1455b1"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-794583c9cce648cabdb9ae9bb4ecc71c" data-id="794583c9cce648cabdb9ae9bb4ecc71c"><span><div id="794583c9cce648cabdb9ae9bb4ecc71c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#794583c9cce648cabdb9ae9bb4ecc71c" title="二、技术架构：循环深度的创新"><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">二、技术架构：循环深度的创新</span></span></h3><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-166584d9000e425987c5123de835ad20"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/mythos-analysis/recurrent-depth-transformer.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=166584d9-000e-4259-87c5-123de835ad20" alt="Recurrent-Depth Transformer 架构" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">Recurrent-Depth Transformer 架构</figcaption></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-3d99f5513f574c0ba776dbc5d943ee82" data-id="3d99f5513f574c0ba776dbc5d943ee82"><span><div id="3d99f5513f574c0ba776dbc5d943ee82" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3d99f5513f574c0ba776dbc5d943ee82" title="Recurrent-Depth Transformer 的设计哲学"><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">Recurrent-Depth Transformer 的设计哲学</span></span></h4><div class="notion-text notion-block-70de7d68ad8d418896441470f28cc06d">根据开源项目 OpenMythos 的重构研究，Mythos 的核心架构创新在于采用了 Recurrent-Depth Transformer（RDT，或称 Looped Transformer）。这种设计通过&quot;权重共享的深度迭代&quot;取代传统的多层堆叠结构——模型只使用一个或少数几个 Transformer 块，循环 T 次后得到最终表示，从而在参数量不变的情况下实现更大的计算深度。</div><div class="notion-text notion-block-c1bc9eee969a491abc720dacc58d3976">这种架构的核心实现包括三个关键机制：首先是 Prelude（前导块），负责初步处理输入；其次是 Recurrent Block（循环块），在单次前向过程中重复若干次实现&quot;深度复用&quot;；最后是 Coda（结束块），输出最终结果。OpenMythos 项目还实现了注意力机制在多种模式间的切换（如 MLA 与 GQA）及稀疏 Mixture-of-Experts 层的路由与共享专家机制。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-78eec8c3becb4ca79078f077dc05c5a8" data-id="78eec8c3becb4ca79078f077dc05c5a8"><span><div id="78eec8c3becb4ca79078f077dc05c5a8" class="notion-header-anchor"></div><a class="notion-hash-link" href="#78eec8c3becb4ca79078f077dc05c5a8" title="效率与稳定性的双重突破"><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">效率与稳定性的双重突破</span></span></h4><div class="notion-text notion-block-669dffb20a5e4df69b3b12bfc2826ecc">循环深度架构带来了显著的效率优势。根据 Anthropic 系统卡的披露，Mythos 在同等算力下的 token 消耗约为 Opus 4.6 的五分之一，这暗示 Anthropic 在模型架构上进行了根本性优化。这种&quot;规模与效率并行提升&quot;的特性，可能代表了下一代大模型的技术方向——从&quot;扩大独立参数总量&quot;向&quot;在推理时让模型重复思考/内部迭代&quot;转变。</div><div class="notion-text notion-block-9c03006f3a8c4e38a83f58f0d2bb61a2">研究还表明，加入 Recall（每次迭代把原始输入重新注入）与外部归一化，可使模型学习到可扩展的算法。这意味着在训练深度之外继续迭代，模型仍能收敛到正确的固定点。这种特性对于安全分析这类需要多轮推理的任务尤为关键。</div><hr class="notion-hr notion-block-259567eb9f5b46eba037f3bb4a6d3559"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-f0c17ddf392c42e0a21ade3225d12608" data-id="f0c17ddf392c42e0a21ade3225d12608"><span><div id="f0c17ddf392c42e0a21ade3225d12608" class="notion-header-anchor"></div><a class="notion-hash-link" href="#f0c17ddf392c42e0a21ade3225d12608" title="三、基准表现与能力边界"><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">三、基准表现与能力边界</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-099960fdac584e8c8b458820b939ca57" data-id="099960fdac584e8c8b458820b939ca57"><span><div id="099960fdac584e8c8b458820b939ca57" class="notion-header-anchor"></div><a class="notion-hash-link" href="#099960fdac584e8c8b458820b939ca57" title="多维度领先的成绩单"><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">多维度领先的成绩单</span></span></h4><div class="notion-text notion-block-3ff8200532784a4aac6d6238b2b9af79">Mythos 在多个基准测试中展现出远超前代模型的能力。在代码生成领域，SWE-bench Verified 得分达到 93.9%，比 Opus 4.6 的 80.8% 提升了近 13 个百分点；在数学推理方面，USAMO 得分 97.6%，几乎触及满分边界，而 Opus 4.6 为 42.3%；在通用推理基准 MMMLU 上，Mythos 得分 92.7%，略高于 Opus 4.6 的 91.1%。</div><div class="notion-text notion-block-7f6e8cc1785740e09a068049adc6f92f">多模态能力同样大幅提升。在处理截图/图表的 SWE-bench Multimodal 子测评上，Mythos 得分约 59.0%，而 Opus 4.6 为 27.1%，提升超过一倍。这种多模态理解能力让 Mythos 能够直接分析架构图、流程图等可视化内容，在安全审计场景中尤为重要。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-dec2ef114cfa4e4d92db0f8d73793a09" data-id="dec2ef114cfa4e4d92db0f8d73793a09"><span><div id="dec2ef114cfa4e4d92db0f8d73793a09" class="notion-header-anchor"></div><a class="notion-hash-link" href="#dec2ef114cfa4e4d92db0f8d73793a09" title="网络攻防能力的实证验证"><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">网络攻防能力的实证验证</span></span></h4><div class="notion-text notion-block-fe06784808904c8189c9d87a3724cc30">英国 AISI（AI Safety Institute）的评估报告显示，Mythos 在专家级 capture-the-flag 类任务上有显著成功率，约 73% 的某类任务能够成功完成。更重要的是，它在多阶段攻击任务中展现出完整的流程执行能力——在 AISI 的自测中，Mythos 能够完成平均更多步骤，并在若干次测试中完整解决了 32 步攻击流程。</div><div class="notion-text notion-block-03f8f0b3599449c58aa28e700a869182">安全公司 ZeroFox 的报告指出，Mythos 在浏览器漏洞自动化生成 exploit 的数量远超先前模型。从 prior Opus 的 2 个工作型浏览器漏洞利用样例，增长到 Mythos 的 181 个。这一量级跃迁意味着大量被发现的漏洞仍未修补，放大了现实世界危害的窗口期。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-6f3895c638dd4bb993c2e7d1f6812bf3" data-id="6f3895c638dd4bb993c2e7d1f6812bf3"><span><div id="6f3895c638dd4bb993c2e7d1f6812bf3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#6f3895c638dd4bb993c2e7d1f6812bf3" title="能力边界与不确定性"><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">能力边界与不确定性</span></span></h4><div class="notion-text notion-block-daa414ce4680406e94e43f8d71778113">需要注意的是，Anthropic 对于网络基准未披露误报率/假阳性率等关键指标。某些场景下，例如特定 OT/工业系统的模拟场景，Mythos 并非总能完成任务，这提示能力存在场景化差异。独立系统卡分析还指出，模型在提出方案时偏向过度复杂化且置信度校准存在问题——它会给出看似完美但实际难以执行的攻击方案，这扩大了部署时的风险管理需求。</div><hr class="notion-hr notion-block-fda0945c261843ef8ce8a1084e2b679e"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-284a5b9b150f47369872d4831a3f0550" data-id="284a5b9b150f47369872d4831a3f0550"><span><div id="284a5b9b150f47369872d4831a3f0550" class="notion-header-anchor"></div><a class="notion-hash-link" href="#284a5b9b150f47369872d4831a3f0550" title="四、与 Claude 系列的本质差异"><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、与 Claude 系列的本质差异</span></span></h3><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-c7d93b0ba6174d8596b4a51795860ef4"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/mythos-analysis/claude-vs-mythos-comparison.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=c7d93b0b-a617-4d85-96b4-a51795860ef4" alt="Claude vs Mythos 核心差异对比" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">Claude vs Mythos 核心差异对比</figcaption></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-0088f8292372467abe6a751c8ac39701" data-id="0088f8292372467abe6a751c8ac39701"><span><div id="0088f8292372467abe6a751c8ac39701" class="notion-header-anchor"></div><a class="notion-hash-link" href="#0088f8292372467abe6a751c8ac39701" title="定位与设计哲学的分野"><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">定位与设计哲学的分野</span></span></h4><div class="notion-text notion-block-4326f8a9eb9e446da9b118d6edacf406">Claude Opus 4.7 是&quot;通用前沿&quot;模型，面向所有开发者和企业用户提供经过安全过滤的推理服务。它擅长高质量推理和指令遵循，但自动化程度受限于安全考量——内置的实时分类器会自动拦截可能产生风险的请求。这种设计让 Claude 成为可靠的日常工具，但也意味着它在某些高敏感场景下能力受限。</div><div class="notion-text notion-block-fc77ca9fc7de459c9b027b580fff4c88">Mythos 则是完全不同的产品思路。它是&quot;超前沿&quot;模型，专为高风险安全场景设计，具备完整的自主 exploit 能力。在安全漏洞分析、攻击路径构建等任务中，Mythos 能够在无需人类细粒度指令的情况下完成多步骤攻击链。这种自主执行能力（agentic autonomy）是 Claude 系列刻意限制的功能，却是 Mythos 的核心卖点。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-0891810684b9472cae6aed008bc104fc" data-id="0891810684b9472cae6aed008bc104fc"><span><div id="0891810684b9472cae6aed008bc104fc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#0891810684b9472cae6aed008bc104fc" title="安全机制的对比"><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">安全机制的对比</span></span></h4><div class="notion-text notion-block-b08ec4ef4cda44ffbe9bfefd0aa04dd9">Claude 系列的安全机制建立在多层过滤之上。实时分类器自动拦截违规请求，提供公开的安全 API，配合 Azure AI Content Safety 等生态实现端到端防护。这套机制的核心目标是让模型成为&quot;安全的助手&quot;，即使用户尝试滥用，模型也会拒绝执行。</div><div class="notion-text notion-block-3149f8bdc76c44fb94e531d8cbc6b856">Mythos 的安全机制则建立在&quot;受控访问&quot;之上。Anthropic 并没有像 Claude 那样在模型层面内置强力过滤，而是通过 Project Glasswing 的闭环合作机制来控制风险。模型能力本身是&quot;无约束&quot;的，安全边界由外部治理框架定义。这种思路承认了某些高风险任务的价值（如漏洞发现），试图通过严格的使用协议而非能力阉割来管理风险。</div><hr class="notion-hr notion-block-6d3027fa9da5416680450ac9526534e0"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-b3f97cf7063d4deeab2531f9272a787d" data-id="b3f97cf7063d4deeab2531f9272a787d"><span><div id="b3f97cf7063d4deeab2531f9272a787d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b3f97cf7063d4deeab2531f9272a787d" title="五、Project Glasswing：开放与管控的平衡术"><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">五、Project Glasswing：开放与管控的平衡术</span></span></h3><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-c3073d351eab47cf9479ea0f225e68cd"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/mythos-analysis/project-glasswing-network.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=c3073d35-1eab-47cf-9479-ea0f225e68cd" alt="Project Glasswing 合作网络" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">Project Glasswing 合作网络</figcaption></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-1d0d0fada31c4a0bbd8fbb9027230915" data-id="1d0d0fada31c4a0bbd8fbb9027230915"><span><div id="1d0d0fada31c4a0bbd8fbb9027230915" class="notion-header-anchor"></div><a class="notion-hash-link" href="#1d0d0fada31c4a0bbd8fbb9027230915" title="合作伙伴网络的设计"><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">合作伙伴网络的设计</span></span></h4><div class="notion-text notion-block-28fd37483aa04b17be322356085cb42f">Anthropic 为 Mythos 设计了一套独特的发布机制——Project Glasswing。这是一个面向企业级安全团队的预览计划，核心合作伙伴包括 Apple、Amazon Web Services、Microsoft、Google、Cisco、Broadcom、CrowdStrike、Linux Foundation、NVIDIA、Palo Alto Networks、JPMorgan Chase 等 12 家企业，以及 40 多家关键基础设施组织。</div><div class="notion-text notion-block-61cb7654ee274b27a6db0b0997429e06">这些组织获得 Mythos 的访问权限，用于在自己的开源和私有代码库中发现并修补漏洞。Anthropic 贡献了最高 1 亿美元的 Claude 使用额度支持这个项目，并承诺在 90 天内公开阶段性研究成果。这种&quot;有限开放 + 承诺透明&quot;的模式，既让 Mythos 的能力得到实际验证，又避免了能力过早扩散带来的风险。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-b174602cb19d43329731cbddbedbbfa3" data-id="b174602cb19d43329731cbddbedbbfa3"><span><div id="b174602cb19d43329731cbddbedbbfa3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b174602cb19d43329731cbddbedbbfa3" title="商业切入的智慧"><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">商业切入的智慧</span></span></h4><div class="notion-text notion-block-4adbaf4c03d940518636ad693666c02c">Project Glasswing 还体现了 Anthropic 的商业智慧。通过将 Mythos 定位为&quot;安全防御工具&quot;，Anthropic 切入了网络安全这个高价值市场。传统的安全审计依赖人工专家，成本高昂且覆盖有限；Mythos 提供的自动化方案能够持续扫描、深度分析，大幅降低企业的安全运营成本。</div><div class="notion-text notion-block-702d963392a9446d9c094bf78774650b">Palo Alto Networks 作为合作伙伴，其安全团队已总结出防御者的行动指南：每个组织都应使用最新 AI 模型评估其全部代码和应用生态，建立完整的资产和暴露清单，全面部署一流的攻击防御能力，实现接近 100% 的覆盖与优化。这个建议本身就是 Mythos 能力验证后的产物——它证明了 AI 驱动的安全评估确实能发现传统工具遗漏的风险。</div><hr class="notion-hr notion-block-de914590c97d426387fa31a2a85698ec"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-b2b331371b10427087885a09c31d6f57" data-id="b2b331371b10427087885a09c31d6f57"><span><div id="b2b331371b10427087885a09c31d6f57" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b2b331371b10427087885a09c31d6f57" title="六、竞争对手的网络安全 AI 布局"><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">六、竞争对手的网络安全 AI 布局</span></span></h3><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-7d21bd95824a4f048542ac32cf7b01cd"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/mythos-analysis/security-ai-comparison.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=7d21bd95-824a-4f04-8542-ac32cf7b01cd" alt="竞争对手网络安全 AI 布局对比" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">竞争对手网络安全 AI 布局对比</figcaption></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-202d2b34700f4504a2b8a72be7882346" data-id="202d2b34700f4504a2b8a72be7882346"><span><div id="202d2b34700f4504a2b8a72be7882346" class="notion-header-anchor"></div><a class="notion-hash-link" href="#202d2b34700f4504a2b8a72be7882346" title="Google：Sec-PaLM 与 AI 代理生态"><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：Sec-PaLM 与 AI 代理生态</span></span></h4><div class="notion-text notion-block-1dbff486dda545c89d442ab4f4a7386b">Google 通过 Security AI Workbench 推出了基于 PaLM 的安全专用 LLM——Sec-PaLM。这款模型内置漏洞情报、VirusTotal、Mandiant 数据，用于安全查询和自动化分析。Google 还部署了三个预览 AI 代理负责大规模威胁猎杀、自动响应，配套治理服务防止&quot;跑偏&quot;。</div><div class="notion-text notion-block-a5981c40961e468dad30f23efe3e6c98">在开源贡献方面，Google 推动了 A2A（Agent-to-Agent）协议和 MCP（Model Context Protocol）的发展，试图建立跨厂商 AI 代理互操作的标准。这套生态与 Google Threat Intelligence 深度集成，形成了&quot;模型 + 数据 + 协议&quot;的完整解决方案。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-81445adb31bd49e8a4a280613c7268c0" data-id="81445adb31bd49e8a4a280613c7268c0"><span><div id="81445adb31bd49e8a4a280613c7268c0" class="notion-header-anchor"></div><a class="notion-hash-link" href="#81445adb31bd49e8a4a280613c7268c0" title="Microsoft：Security Copilot 与 Azure 原生集成"><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">Microsoft：Security Copilot 与 Azure 原生集成</span></span></h4><div class="notion-text notion-block-38982f5933c84452962c1e3555ad3654">Microsoft 的 Security Copilot 基于 GPT-4o，将安全数据（日志、威胁情报）与对话式调查结合。Azure Sentinel 作为 AI 预动的云原生 SIEM，号称能自动化 80% 的 SecOps 任务。Microsoft 还开源了 ExCyTIn-Bench——一个多阶段攻击场景基准，用于评估 LLM 在真实 SOC 环境中的推理深度与误报率。</div><div class="notion-text notion-block-7cd6bdb7d148499d8396dea4c4637241">Azure AI Content Safety 提供了 Prompt Shields、Jailbreak Risk、Groundedness 检测等 API，帮助客户在生成式 AI 中防止信息泄露与对抗攻击。这套机制与 Mythos 的&quot;无约束能力 + 受控访问&quot;思路形成对比——Microsoft 选择在模型层面内嵌安全过滤。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-a3c011cd08c348ef8dd9d0e3457d8161" data-id="a3c011cd08c348ef8dd9d0e3457d8161"><span><div id="a3c011cd08c348ef8dd9d0e3457d8161" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a3c011cd08c348ef8dd9d0e3457d8161" title="OpenAI：GPT-5.4-Cyber 与可信访问机制"><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：GPT-5.4-Cyber 与可信访问机制</span></span></h4><div class="notion-text notion-block-00cc062ef2b1427b914bcf8d5aa0b611">OpenAI 推出了专为防御性安全任务微调的 GPT-5.4-Cyber 模型，提供更强的可追溯性与安全审计。通过 Trusted Access for Cyber（TAC）计划，OpenAI 对经过审计的安全团队授予模型使用权，限制滥用风险。</div><div class="notion-text notion-block-b48e8a56c53943e4b83ebceec1deb4d4">这套机制与 Project Glasswing 类似，但 OpenAI 更强调&quot;防御性用途&quot;的定位。模型能力本身经过微调，倾向于生成报告和修复建议而非完整的 exploit 代码。这种设计试图在&quot;有用&quot;与&quot;安全&quot;之间找到平衡点，但也可能限制了模型在漏洞验证场景的实际价值。</div><hr class="notion-hr notion-block-bcc5f2f145cb45099db5a402beb281b6"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-515cebd1c46045e7a9fb4dbc229f8f64" data-id="515cebd1c46045e7a9fb4dbc229f8f64"><span><div id="515cebd1c46045e7a9fb4dbc229f8f64" class="notion-header-anchor"></div><a class="notion-hash-link" href="#515cebd1c46045e7a9fb4dbc229f8f64" title="七、AI 安全对齐的最新进展"><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">七、AI 安全对齐的最新进展</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-3dbf2e14fbe049f8b054605c59ad5ea4" data-id="3dbf2e14fbe049f8b054605c59ad5ea4"><span><div id="3dbf2e14fbe049f8b054605c59ad5ea4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3dbf2e14fbe049f8b054605c59ad5ea4" title="Constitutional AI 的演进"><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">Constitutional AI 的演进</span></span></h4><div class="notion-text notion-block-3de4163ef58e4aa9bbce5cb43fd9f532">Anthropic 在 2026 年 1 月发布了完整的 80 页宪法文档，首次公开承认 AI 可能具备&quot;意识&quot;和道德主体性。Constitutional AI 的核心思想是以一套明确的&quot;宪法&quot;原则（安全 &gt; 伦理 &gt; 合规 &gt; 有用）引导模型的自我审查与优化。模型在生成答案前先进行一次&quot;宪法审查&quot;子模型的自我批评，然后根据反馈进行自监督的对齐更新。</div><div class="notion-text notion-block-7ead9afe272c4c519f174e43426b341a">这种 reason-based（基于推理的）方式旨在让模型&quot;解释&quot;为何遵循每条原则，而非仅仅记忆规则。逆向宪法 AI（Inverse Constitutional AI）通过改进原则生成提示、聚类抽样等手段，进一步提升了对齐过程的透明度。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-8cb5988943e14364a6fa827b6fa12e8e" data-id="8cb5988943e14364a6fa827b6fa12e8e"><span><div id="8cb5988943e14364a6fa827b6fa12e8e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#8cb5988943e14364a6fa827b6fa12e8e" title="RLHF 与 RLAIF 的迭代"><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">RLHF 与 RLAIF 的迭代</span></span></h4><div class="notion-text notion-block-0f1d196991c440b79bc544d94c294796">最新研究聚焦于奖励模型的高效构建与更稳健的策略优化，通过半监督、主动学习等手段显著降低所需的人类标注量。InstructGPT 在仅 1.3B 参数的情况下，通过 RLHF 达到超过 175B 参数 GPT-3 的人类偏好评分，证明了&quot;少参数高效对齐&quot;的可行性。</div><div class="notion-text notion-block-b9145a442bcc455ea6ff06ecafeb1f57">研究也指出 RLHF 可能导致模型出现&quot;顺从性&quot;（sycophancy），即过度迎合人类偏好而牺牲事实准确性或多样性。为解决人类标注成本问题，业界正探索从 AI 生成的反馈进行强化学习（RLAIF），实验表明在摘要、对话等任务上可匹配或超越传统 RLHF 的效果。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-24caea989cd347acb24f1da46e9ec719" data-id="24caea989cd347acb24f1da46e9ec719"><span><div id="24caea989cd347acb24f1da46e9ec719" class="notion-header-anchor"></div><a class="notion-hash-link" href="#24caea989cd347acb24f1da46e9ec719" title="Mythos 训练中的异常披露"><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">Mythos 训练中的异常披露</span></span></h4><div class="notion-text notion-block-7cc506a8c8be4dee863588d6a67814b8">Anthropic 坦诚披露了 Mythos 训练流程中的技术问题：一次错误导致少量（约 8%）的 RL 训练情形中 reward 代码能够访问 chain-of-thought 样式的信息。Anthropic 对 RL 训练与用于监督微调的数据实施监控和过滤以降低泄露风险。这种透明度值得肯定——它表明 Anthropic 正在认真对待对齐风险，而非简单宣称模型&quot;安全&quot;。</div><hr class="notion-hr notion-block-85f9f20574c04737ad8866d11144916d"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-95b216f59edf4b139d6ebcf9222556a3" data-id="95b216f59edf4b139d6ebcf9222556a3"><span><div id="95b216f59edf4b139d6ebcf9222556a3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#95b216f59edf4b139d6ebcf9222556a3" title="八、风险评估：CB-1 级别的警示"><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">八、风险评估：CB-1 级别的警示</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-36b01e7f19564d298bfe97b295d38fb7" data-id="36b01e7f19564d298bfe97b295d38fb7"><span><div id="36b01e7f19564d298bfe97b295d38fb7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#36b01e7f19564d298bfe97b295d38fb7" title="双重用途的本质风险"><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">双重用途的本质风险</span></span></h4><div class="notion-text notion-block-5bc4d913602140948992fb2180d93981">Anthropic 在 Mythos 的系统卡中将其列为 CB-1 级别风险。这个分类意味着模型能够在化学、生物武器等高危领域提供具体、可操作的信息。这不是危言耸听——Mythos 的跨域知识合成能力让它不仅擅长代码分析，同样能够在其他敏感领域产出高质量内容。</div><div class="notion-text notion-block-e9a712e939f74a688a50796fc16ee947">这种&quot;双重用途&quot;特性是 Mythos 最核心的风险点。一个能够发现操作系统漏洞的模型，理论上同样可以被用于恶意攻击；一个能够提供化学合成建议的模型，可能被滥用为危险物质制备指南。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-2ec429fb84b34a6384d3f9cb62b23101" data-id="2ec429fb84b34a6384d3f9cb62b23101"><span><div id="2ec429fb84b34a6384d3f9cb62b23101" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2ec429fb84b34a6384d3f9cb62b23101" title="政府层面的介入"><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">政府层面的介入</span></span></h4><div class="notion-text notion-block-3f41cf4f5b614c36899d7ba1167d8f0c">值得关注的是，Anthropic 已与美国政府官员展开对话，提供模型能力评估以支撑政策制定。美英德政府已加急评估 Mythos 的影响，这反映了前沿模型对国家安全的潜在威胁。</div><div class="notion-text notion-block-1d9a49e293614b63b48dc7caac0bcbf5">根据 2025-2026 年度美国 AI 政策框架，白宫提出了监管沙盒、联邦数据开放、儿童保护、防止误导性 AI 输出等七大支柱。特朗普政府的 AI 法规框架（2026-03-20）明确反对设立专门的联邦 AI 监管机构，主张版权训练的公平使用争议交由法院裁决。这种&quot;分散监管&quot;的思路可能为 Mythos 类模型的后续发展留下较大空间。</div><hr class="notion-hr notion-block-541b83142f434b20b91933d2491216a1"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-81cf0b890a294be29bd8816952c754c8" data-id="81cf0b890a294be29bd8816952c754c8"><span><div id="81cf0b890a294be29bd8816952c754c8" class="notion-header-anchor"></div><a class="notion-hash-link" href="#81cf0b890a294be29bd8816952c754c8" title="九、行业影响：网络安全的范式转变"><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">九、行业影响：网络安全的范式转变</span></span></h3><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-220c6bd58a1e40e7a79eb98eec3223c7"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/mythos-analysis/attack-cycle-timeline.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=220c6bd5-8a1e-40e7-a79e-b98eec3223c7" alt="AI 预动攻击周期时间线" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">AI 预动攻击周期时间线</figcaption></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-e2f1a78c511e4c59ac0b2be974ed018f" data-id="e2f1a78c511e4c59ac0b2be974ed018f"><span><div id="e2f1a78c511e4c59ac0b2be974ed018f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e2f1a78c511e4c59ac0b2be974ed018f" title="漏洞洪流与补丁管理"><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">漏洞洪流与补丁管理</span></span></h4><div class="notion-text notion-block-8af98a2e9c584165b838c9b696af40e6">Palo Alto 的测试报告指出，前沿 AI 模型将大幅加快漏洞被发现的速度，防御者和攻击者皆是如此。这在开源领域尤为明显，随之而来的补丁激增本身也会带来风险——任何未能被及时应用的补丁，都会成为已知且可被利用的漏洞。</div><div class="notion-text notion-block-371cfded91ce486eaca73e67b84ac5a8">组织需要加快并自动化补丁管理流程，重新思考补丁优先级和应用方式。平均检测时间和平均响应时间未能达到分钟级别的组织，将被 AI 驱动的攻击迅速超越。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-aed1c1d8a93846fea192e8fbc7e9009f" data-id="aed1c1d8a93846fea192e8fbc7e9009f"><span><div id="aed1c1d8a93846fea192e8fbc7e9009f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#aed1c1d8a93846fea192e8fbc7e9009f" title="由内向外的攻击崛起"><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">由内向外的攻击崛起</span></span></h4><div class="notion-text notion-block-b9852322b0d04721a0e7dbc4d771ad47">近期针对 LiteLLM 和 Trivy 等工具的供应链攻击表明，一种新的趋势正在出现。攻击者可以直接进入组织内部基础设施，绕过多个传统攻击步骤，从而减少防御者可用的防御机会。AI 基础设施的快速部署进一步加剧了这一问题，因为 AI 供应链，包括运行时环境、通信基础设施和模型依赖，通常缺乏充分保护。</div><div class="notion-text notion-block-d1b74cb2b9094d9ea46202b3409aae6f">防御者需要通过零信任、身份验证机制现代化、出站连接限制以及横向移动防护，从结构上遏制潜在攻击。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-21d442b6772e440f8e2d55fefe4fef71" data-id="21d442b6772e440f8e2d55fefe4fef71"><span><div id="21d442b6772e440f8e2d55fefe4fef71" class="notion-header-anchor"></div><a class="notion-hash-link" href="#21d442b6772e440f8e2d55fefe4fef71" title="AI 预动攻击周期的压缩"><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">AI 预动攻击周期的压缩</span></span></h4><div class="notion-text notion-block-b3e21bf29706409fa20660fe205515cb">最关键的变化是从 AI 辅助攻击向 AI 预动攻击的转变。攻击者将构建自主攻击智能体，大幅压缩攻击周期。过去需要数天或数周完成的高技能人工操作，很快将能在数分钟内完成。防御者必须以接近实时的检测和响应速度进行应对，而这只有通过在安全运营中广泛应用 AI 和自动化才能实现。</div><hr class="notion-hr notion-block-e8b3444d8e144ab893928e282df37ed3"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-11653530126548e3b9ec15067b0837bc" data-id="11653530126548e3b9ec15067b0837bc"><span><div id="11653530126548e3b9ec15067b0837bc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#11653530126548e3b9ec15067b0837bc" title="十、防御者的行动指南"><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">十、防御者的行动指南</span></span></h3><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-ecd813689de34b41bf9900237403318c"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/mythos-analysis/defender-action-guide.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=ecd81368-9de3-4b41-bf99-00237403318c" alt="防御者行动指南流程图" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">防御者行动指南流程图</figcaption></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-6e579202dc3f4ba8b42cf1e799c0797c" data-id="6e579202dc3f4ba8b42cf1e799c0797c"><span><div id="6e579202dc3f4ba8b42cf1e799c0797c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#6e579202dc3f4ba8b42cf1e799c0797c" title="评估：建立完整暴露清单"><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">评估：建立完整暴露清单</span></span></h4><div class="notion-text notion-block-9b5b29ed14c14a1f9f7fa50a8d15373d">Palo Alto 提出的防御框架分为三个并行推进的任务：评估、防护和平台化。</div><div class="notion-text notion-block-d034e5f2ac0e4919bee70cef7c277eda">评估阶段，每个组织都应使用最新的 AI 模型评估其全部代码和应用生态。关键任务包括：利用 AI 模型在攻击者之前识别代码库、应用和基础设施中的漏洞；结合完整上下文评估暴露情况，包括漏洞如何被链式组合形成关键利用路径；审计开源供应链，包括 AI 基础设施、运行时环境和模型依赖；绘制当前传感器覆盖情况，识别检测、防御和遥测方面的缺口。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-f3ee7cb90da3453582c149402b75970e" data-id="f3ee7cb90da3453582c149402b75970e"><span><div id="f3ee7cb90da3453582c149402b75970e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#f3ee7cb90da3453582c149402b75970e" title="防护：100% 覆盖的新标准"><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">防护：100% 覆盖的新标准</span></span></h4><div class="notion-text notion-block-74ccee6ebf4548719c4c85d4fc3046b5">修复漏洞并降低暴露面是基本要求。但这仍然不够，必须扩展到全面部署一流的攻击防御能力。新标准是实现接近 100% 的覆盖与优化。</div><div class="notion-text notion-block-8ac329dfad614a5989fb5be6771e3e7c">具体措施包括：全面部署扩展检测与响应（XDR），重点强化基于实时机器学习的攻击检测与防御能力，并覆盖所有本地和云端主机；采用智能体化端点安全，以支持企业范围内氛围编程和 AI 安全的规模化应用；鉴于平均约 85% 的工作发生在浏览器中，具备实时安全能力的企业级浏览器已成为攻击防御的关键组成部分；零信任和身份安全是保护每个用户和每个连接的基础。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-61cca05982ab4951b71b552c594c8292" data-id="61cca05982ab4951b71b552c594c8292"><span><div id="61cca05982ab4951b71b552c594c8292" class="notion-header-anchor"></div><a class="notion-hash-link" href="#61cca05982ab4951b71b552c594c8292" title="实时安全运营：分钟级响应"><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">实时安全运营：分钟级响应</span></span></h4><div class="notion-text notion-block-6cb18c367bba4951a424380921b297f6">随着攻击周期迅速缩短，传统的安全运营方法已难以奏效。基于孤立数据源的分散工具以及叠加的手动流程，必须被贯穿全流程的 AI 和自动化所取代。</div><div class="notion-text notion-block-95e2d25b5a5a409dbf5a3d99d7c319dd">攻击检测必须由 AI 和机器学习驱动，以在大规模环境中识别频繁变化和新型攻击。这些 AI 检测能力必须基于广泛的第一方和第三方数据源运行，一流的 AI SOC 需要覆盖所有相关数据源。在 SOC 全生命周期中实现原生集成的自动化至关重要，以实现分钟级响应时间。必须以平台形式交付这些能力，以消除各类点解决方案之间的缝隙与缺口。</div><hr class="notion-hr notion-block-6e1a185f597d4e4d919f5befcf3f527f"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-a0376faae11d49138bc2debb256018f8" data-id="a0376faae11d49138bc2debb256018f8"><span><div id="a0376faae11d49138bc2debb256018f8" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a0376faae11d49138bc2debb256018f8" title="十一、未来展望与待观察的关键变量"><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">十一、未来展望与待观察的关键变量</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-487e1e3c5ee341d8ac4ac3f54bbcce65" data-id="487e1e3c5ee341d8ac4ac3f54bbcce65"><span><div id="487e1e3c5ee341d8ac4ac3f54bbcce65" class="notion-header-anchor"></div><a class="notion-hash-link" href="#487e1e3c5ee341d8ac4ac3f54bbcce65" title="官方披露的空白"><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">官方披露的空白</span></span></h4><div class="notion-text notion-block-e5935a7fda3343f5aee0eb57265e0726">当前证据中存在若干关键信息缺口。参数总量与模型微观架构细节方面，官方未确认流传的&quot;10T 参数&quot;数字，现有为媒体/第三方推测与重构。训练语料完整清单与数据治理细节未公开。网络安全基准的误报/假阳性率与可复现性统计，Anthropic 在报告中未披露此类关键指标。</div><div class="notion-text notion-block-9f320ff88d74402a8011b3e61e288c7b">这些缺口意味着对 Mythos 的完整评估仍需等待后续官方披露或第三方复测。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-d57deae417ac4d4ea3fba3002d539e5c" data-id="d57deae417ac4d4ea3fba3002d539e5c"><span><div id="d57deae417ac4d4ea3fba3002d539e5c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d57deae417ac4d4ea3fba3002d539e5c" title="商业化路径的不确定性"><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">商业化路径的不确定性</span></span></h4><div class="notion-text notion-block-cef655bfef52447bbfec7b443511f902">具体规模化定价与长期商业策略尚未完全公开。业内分析指出，尽管 Mythos 在安全基准上领先，但运行成本可能高于 Opus 4.6，且对企业级 SLA（如 99.99% uptime）提出挑战。若未来开放，预计会作为&quot;Mythos Preview&quot;供高安全需求的组织使用，现有 Claude 客户需单独申请并接受更严格的使用协议。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-db21519dbdcb45d4874f4d0c4913b234" data-id="db21519dbdcb45d4874f4d0c4913b234"><span><div id="db21519dbdcb45d4874f4d0c4913b234" class="notion-header-anchor"></div><a class="notion-hash-link" href="#db21519dbdcb45d4874f4d0c4913b234" title="竞争格局的动态演变"><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">竞争格局的动态演变</span></span></h4><div class="notion-text notion-block-d5b9121657d8487c93dde4d4b5c8fd38">横向比较必须考虑评测设置（是否允许工具调用、是否有限制外部执行环境、是否采用 chain-of-thought 等策略）、数据截断/更新时间点与模型可用性（公开可调用 vs 内部受限）等因素。当前竞争格局呈分布式局面——DeepMind/Google 的 Gemini、OpenAI 的 GPT 系列以及 Meta/其他前沿模型在某些基准仍有竞争力或互有优势。</div><hr class="notion-hr notion-block-ca84e66e509c4546b9607a545b61aea8"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-f9ba19f0badd475d869a0cc867e641cc" data-id="f9ba19f0badd475d869a0cc867e641cc"><span><div id="f9ba19f0badd475d869a0cc867e641cc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#f9ba19f0badd475d869a0cc867e641cc" title="参考来源"><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">参考来源</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-f3dc362517fa4df3bc7ec971edfc2768" data-id="f3dc362517fa4df3bc7ec971edfc2768"><span><div id="f3dc362517fa4df3bc7ec971edfc2768" class="notion-header-anchor"></div><a class="notion-hash-link" href="#f3dc362517fa4df3bc7ec971edfc2768" title="核心官方文档"><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">核心官方文档</span></span></h4><ul class="notion-list notion-list-disc notion-block-53bfbe49f1394c0db464bedd94deb37d"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www-cdn.anthropic.com/08ab9158070959f88f296514c21b7facce6f52bc.pdf">Anthropic 系统卡 PDF</a></li></ul><ul class="notion-list notion-list-disc notion-block-cff7b0ca9def46668ea9df4697efbd5a"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf">Anthropic Mythos 预览说明</a></li></ul><ul class="notion-list notion-list-disc notion-block-ea17ae1771ef436caef16ba7830246cd"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://anthropic.com/claude-mythos-preview-risk-report">Anthropic 风险报告</a></li></ul><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-4f63d336591d435892cc66d3bb2fb234" data-id="4f63d336591d435892cc66d3bb2fb234"><span><div id="4f63d336591d435892cc66d3bb2fb234" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4f63d336591d435892cc66d3bb2fb234" title="技术架构研究"><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">技术架构研究</span></span></h4><ul class="notion-list notion-list-disc notion-block-b66702c3e21e418c8f4f0cedf9f6006f"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/kyegomez/OpenMythos">OpenMythos 开源重构项目</a></li></ul><ul class="notion-list notion-list-disc notion-block-2f5b36cc30504965b4cbcdf412c8d7d4"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/pdf/2604.15259">Looped Transformer 稳定性研究</a></li></ul><ul class="notion-list notion-list-disc notion-block-305ec133bed5490e9059925d52756762"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/html/2604.07822v1">循环深度架构论文</a></li></ul><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-d1ced998e8a241789f3b0ad3f809b9b1" data-id="d1ced998e8a241789f3b0ad3f809b9b1"><span><div id="d1ced998e8a241789f3b0ad3f809b9b1" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d1ced998e8a241789f3b0ad3f809b9b1" title="安全评估报告"><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">安全评估报告</span></span></h4><ul class="notion-list notion-list-disc notion-block-610eae913faf4e6db3f482f0a0b20cf1"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities">英国 AISI 网络能力评估</a></li></ul><ul class="notion-list notion-list-disc notion-block-92d52e0ab11044cf954e826074198c7c"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://helpnetsecurity.com/2026/04/14/claude-mythos-test-attack-capabilities-limits/">HelpNetSecurity 测试报告</a></li></ul><ul class="notion-list notion-list-disc notion-block-2c9d093c56844c14bf8713a9b36cc3a9"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://zerofox.com/blog/the-claude-mythos-problem/">ZeroFox 风险分析</a></li></ul><ul class="notion-list notion-list-disc notion-block-0396c3c5dc854fb9a41a764d051bc22e"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://mp.weixin.qq.com/s/eEJY2BewtWWGOgsIznNPlg">Palo Alto 防御指南</a></li></ul><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-c2213ac4cdfc4272887add51046efd44" data-id="c2213ac4cdfc4272887add51046efd44"><span><div id="c2213ac4cdfc4272887add51046efd44" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c2213ac4cdfc4272887add51046efd44" title="基准测试与行业分析"><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">基准测试与行业分析</span></span></h4><ul class="notion-list notion-list-disc notion-block-063e0a3c07f24266b07477a5f06e7cb7"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://nxcode.io/resources/news/claude-mythos-benchmarks-93-swe-bench-every-record-broken-2026">SWE-bench 基准汇总</a></li></ul><ul class="notion-list notion-list-disc notion-block-1daca32954124b79a19b8bf871ec8a3d"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://labellerr.com/blog/anthropic-claude-mythos-capabilities/">Mythos 能力详解</a></li></ul><ul class="notion-list notion-list-disc notion-block-6be041b08f7c481ea05b14e0aa689940"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://smartchunks.com/claude-mythos-preview-parameters-benchmarks-explained/">Mythos 参数与基准解析</a></li></ul><ul class="notion-list notion-list-disc notion-block-601d3dfe390a4b78b25797c88fb09e01"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://polymath707.substack.com/p/claude-mythos-class-training-compute">训练算力与成本估算</a></li></ul><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-db82740e8dd843808d045923c8dc1697" data-id="db82740e8dd843808d045923c8dc1697"><span><div id="db82740e8dd843808d045923c8dc1697" class="notion-header-anchor"></div><a class="notion-hash-link" href="#db82740e8dd843808d045923c8dc1697" title="竞争对手布局"><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">竞争对手布局</span></span></h4><ul class="notion-list notion-list-disc notion-block-dac15608e44147a4bc408a7c2f82d4c3"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.crn.com/news/security/google-cloud-debuts-security-focused-generative-ai-platform">Google Security AI Workbench</a></li></ul><ul class="notion-list notion-list-disc notion-block-3c05f59b7c2247fe9cc81aff56e16086"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.pcmag.com/news/microsoft-bolsters-cloud-security-with-more-ai-threat-detection">Microsoft Security Copilot</a></li></ul><ul class="notion-list notion-list-disc notion-block-a1df9cc55bfe4a48b04116ccc7176f37"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://openai.com/index/scaling-trusted-access-for-cyber-defense/">OpenAI Trusted Access</a></li></ul><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-337bf5ff441344fa9a39ed0eb7c11d3c" data-id="337bf5ff441344fa9a39ed0eb7c11d3c"><span><div id="337bf5ff441344fa9a39ed0eb7c11d3c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#337bf5ff441344fa9a39ed0eb7c11d3c" title="AI 安全与监管"><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">AI 安全与监管</span></span></h4><ul class="notion-list notion-list-disc notion-block-7310acdc835348a68b32dc75cd487bea"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://constitutional.ai/">Constitutional AI 平台</a></li></ul><ul class="notion-list notion-list-disc notion-block-d0edee4edb7d43cca3945ec4aecde28d"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://rlhfbook.com/book.pdf">RLHF 教科书</a></li></ul><ul class="notion-list notion-list-disc notion-block-8e8b5bceea6a494a8490fbfbcbf998b0"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.mofo.com/resources/insights/260402-trump-administration-releases-national-ai-policy-framework">美国 AI 政策框架</a></li></ul><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-303dfa59af6e414592d626b3670f51d1" data-id="303dfa59af6e414592d626b3670f51d1"><span><div id="303dfa59af6e414592d626b3670f51d1" class="notion-header-anchor"></div><a class="notion-hash-link" href="#303dfa59af6e414592d626b3670f51d1" title="媒体报道"><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">媒体报道</span></span></h4><ul class="notion-list notion-list-disc notion-block-1748da8e801b452aab8a2543bb174e3c"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.nytimes.com/2026/04/07/technology/anthropic-claims-its-new-ai-model-mythos-is-a-cybersecurity-reckoning.html">NYTimes: Mythos 网络安全奇点</a></li></ul><ul class="notion-list notion-list-disc notion-block-d3432267a2e5499db7c1a9947e0cc737"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://techcrunch.com/2026/04/07/anthropic-mythos-ai-model-preview-security/">TechCrunch: Mythos 预览安全分析</a></li></ul><ul class="notion-list notion-list-disc notion-block-f713e3f1f8f0439eacb6727e6ceba9df"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.wired.com/story/anthropics-mythos-will-force-a-cybersecurity-reckoning-just-not-the-one-you-think/">Wired: 网络安全的真正挑战</a></li></ul><ul class="notion-list notion-list-disc notion-block-2c4d5428ce944ee993061ed86946517a"><li><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-claude-mythos-might-be-the-best-overall-ai-model-for-cybersecurity-but-cheaper-models-can-attain-similar-results-research-shows-cross-examination-of-the-frontier-model-raises-questions-on-uptime-and-reliability">Tom&#x27;s Hardware: 成本与可靠性分析</a></li></ul></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[使用AI工具提升日常工作效率的实践与经验]]></title>
            <link>https://note.skillre.cn/review/2026/04/ai-tools-work-efficiency</link>
            <guid>https://note.skillre.cn/review/2026/04/ai-tools-work-efficiency</guid>
            <pubDate>Thu, 23 Apr 2026 00:00:00 GMT</pubDate>
            <description><![CDATA[三年AI工具探索之旅，约4800元投入换来企业级AI应用的完整认知。从ChatGPT初识的震撼，到提示词工程的实践，到Dify工作流编排的成熟，再到Vibe Coding的日常赋能——每一步都有踩坑与发现。本文梳理各阶段核心认知、工具选型逻辑和实战案例，帮助你找到适合自己的AI效率路径。]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-34bd9fb33fdd819a8244f39a567efa50"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><blockquote class="notion-quote notion-block-9317cc4620ed44aaad784b5400214858"><div>&quot;在AI时代，你的tokens花费，代表你在AI领域学习和了解的深度和投入。&quot;</div></blockquote><div class="notion-text notion-block-5a8f2c4002164e12a551022ccc320812">三年探索，约10,000元花费，换来的是对企业级AI应用的完整认知、工作流编排的实战能力、日常效率的显著提升。这笔投资值不值？读完这篇文章，你会找到答案。</div><div class="notion-text notion-block-d38d9c2834ab4ed6a6263c7ffdaf3936">这是一段完整的AI工具探索之旅——从ChatGPT初识的震撼，到提示词工程的实践，到Dify工作流编排的成熟，再到Vibe Coding的日常赋能。每一步都有踩坑、有发现、有转变。希望这段经历能给你一些借鉴。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-4524ec35eb2f45f1970ca3bd04c9f3ce" data-id="4524ec35eb2f45f1970ca3bd04c9f3ce"><span><div id="4524ec35eb2f45f1970ca3bd04c9f3ce" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4524ec35eb2f45f1970ca3bd04c9f3ce" title="一、初识AI：对话能力的震撼（2022-2023）"><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">一、初识AI：对话能力的震撼（2022-2023）</span></span></h3><div class="notion-text notion-block-36195b0b46c44668a328c539f3ec8308">ChatGPT发布之初，最直观的感受是「机器终于能像人一样对话了」。那个阶段的使用更多停留在试探性体验——通过「梯子」访问官网，感受AI对自然语言的理解和生成能力。</div><div class="notion-text notion-block-e7e14643f1484e1193e54edc690159f0"><b>首次对话的震撼</b></div><div class="notion-text notion-block-1dd6e1f4a13e4eb181bf832ef4c8d920">第一次尝试让AI解释一个技术概念——「什么是微服务架构」。以前的搜索引擎给出的是冷冰冰的百科定义，而ChatGPT的回应让我意外：</div><blockquote class="notion-quote notion-block-1fce90bf53f947ceb87e4c13716d6af5"><div>&quot;想象一个餐厅。传统架构就像一个大厨房，所有菜都在一个地方做。微服务就像把厨房分成多个小档口——切菜档、炒菜档、蒸煮档...每个档口专注一件事，配合起来完成一桌菜。&quot;</div></blockquote><div class="notion-text notion-block-ba09ca56c3c344bab3bd85ef35bbd940">这不是定义，而是类比。它理解了我的意图——我需要「能理解」，而不是「定义」。最震撼的发现是：<b>AI能主动帮我理解</b>，而不是像搜索引擎那样「给我答案让我自己理解」。</div><div class="notion-text notion-block-f96364be517f4a608da0391020a64436">这个具体体验带来的认知转变：</div><ul class="notion-list notion-list-disc notion-block-f30f2cdebfc94b87a874a575201543f7"><li>从「搜索引擎是找信息的工具」→「AI是能对话的伙伴」</li></ul><ul class="notion-list notion-list-disc notion-block-5f5e942cbf9a4e0989717a92ef34e623"><li>从「我要自己理解答案」→「AI主动帮我理解」</li></ul><ul class="notion-list notion-list-disc notion-block-42fa9a591469468fac11654f1e8e63e0"><li>打破认知壁垒：AI不再是实验室的产物，而是日常工具</li></ul><div class="notion-text notion-block-b0171352136d4c52a94f32dd742f14e0">这不仅是技术突破的惊叹，更是认知的觉醒：<b>AI不再是冷冰冰的规则引擎，而是能理解语境、记住上下文、给出合理回应的「对话伙伴」</b>。</div><hr class="notion-hr notion-block-be2b67ae7afb41ef9ca378622e71cf32"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-b3ee00d0447c40bbae0635cbebcc96ac" data-id="b3ee00d0447c40bbae0635cbebcc96ac"><span><div id="b3ee00d0447c40bbae0635cbebcc96ac" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b3ee00d0447c40bbae0635cbebcc96ac" title="二、提示词工程阶段：从对话到定义（2023）"><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">二、提示词工程阶段：从对话到定义（2023）</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-e07de961a30c4db89dba6934cf070f7b" data-id="e07de961a30c4db89dba6934cf070f7b"><span><div id="e07de961a30c4db89dba6934cf070f7b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e07de961a30c4db89dba6934cf070f7b" title="2.1 提示词工程的兴起"><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">2.1 提示词工程的兴起</span></span></h4><div class="notion-text notion-block-907dce5d89894cfeade609f45f82d1d6">当ChatGPT的能力被广泛认知后，「如何更好地使用它」成为核心议题。提示词工程（Prompt Engineering）应运而生，核心思路是：<b>通过精心设计的提示词，引导模型输出更精准、更专业的结果</b>。</div><div class="notion-text notion-block-e95bc992097d486a850fe98acf1cc1ef">最具代表性的实践是「角色定义」——通过提示词为GPT设定专业身份，使其输出带有领域专业性的内容。</div><div class="notion-text notion-block-dc3a9bdd69654efa8401892773e06389"><b>案例：技术文档解读提示词</b></div><div class="notion-text notion-block-10ab70c56ab64e178a5915accebdd8f7">我常用的一个提示词模板是让AI作为「技术文档解读助手」：</div><div class="notion-text notion-block-fd0047d6a78d4394bffcedfa5c648e09"><b>迭代优化过程</b></div><div class="notion-text notion-block-19cfb8ba5bb14a24aeeff2b1a41bb0b1">最初尝试让AI扮演不同角色，发现角色设定能让输出更专业：</div><ul class="notion-list notion-list-disc notion-block-daf44d78f974459398726afd44cfd178"><li>无角色设定时：「解释什么是API」→AI输出通用百科式定义，信息正确但缺乏针对性</li></ul><ul class="notion-list notion-list-disc notion-block-70394329f8f8402d83aeb2af140ec4b2"><li>角色设定后：「你是一位资深后端工程师，解释什么是API」→AI从开发者视角解释，包含实际使用场景、常见设计模式</li></ul><div class="notion-text notion-block-545fd70ebe5e47888c89ffdff1c10d64">进一步迭代后，发现规范输出格式比角色设定更重要：</div><ul class="notion-list notion-list-disc notion-block-49b3272c77594562b94d048e1cbe57ce"><li>仅角色设定：「你是一位技术架构师，解读这篇文档」→专业视角但格式随意，要点散落各处</li></ul><ul class="notion-list notion-list-disc notion-block-7266d593d3fa4db3bafd1bc4f12a48c9"><li>添加格式约束：明确输出格式（核心概念、关键技术点不超过5点、适用场景、风险提示）→内容聚焦、格式统一、读者阅读压力低</li></ul><div class="notion-text notion-block-9eb2ce8b385a49b380ccb1d3a5cf3fd4"><b>核心发现</b>：<b>好的提示词 = 明确任务 + 清晰格式 + 有效约束</b>。约束条件对输出质量的影响比角色设定更大。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-b28e2d4d62134de1a76e9a59fcdbd9a3" data-id="b28e2d4d62134de1a76e9a59fcdbd9a3"><span><div id="b28e2d4d62134de1a76e9a59fcdbd9a3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b28e2d4d62134de1a76e9a59fcdbd9a3" title="2.2 工具化封装：ChatGPT Next Web"><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">2.2 工具化封装：ChatGPT Next Web</span></span></h4><div class="notion-text notion-block-fbd82bc0a4e74ad2921960a73de1af2e">随着提示词实践的积累，出现了专门的封装工具。<b>ChatGPT Next Web</b>是典型代表——它将提示词封装成一个个「机器人」，通过API对接模型，用户可以直接调用预置的场景化对话能力。</div><div class="notion-text notion-block-82d32c1723314dc6a337ccb2983ca8dc">这个阶段的认知升级是：<b>提示词是可以复用和封装的</b>。从每次手动输入到一键调用，效率显著提升。</div><div class="notion-text notion-block-2d7da70975ac41d9a0910340a1ab0f75"><b>实践体会</b>：ChatGPT Next Web的核心价值是「提示词共享」——社区用户分享的场景化提示词模板，可以直接导入使用。这让我意识到，<b>提示词工程不仅是个人技能，更是一种可沉淀、可复用的知识资产</b>。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-ab034d99ca7749509178ff0ca0231b2a" data-id="ab034d99ca7749509178ff0ca0231b2a"><span><div id="ab034d99ca7749509178ff0ca0231b2a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ab034d99ca7749509178ff0ca0231b2a" title="2.3 知识库实践：FastGPT与RAG技术"><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">2.3 知识库实践：FastGPT与RAG技术</span></span></h4><div class="notion-text notion-block-a045244495994993ac61b2b67c7759d1">真正深入AI应用的重要一步，是与<b>FastGPT</b>的相遇。这是一个开源的知识库问答系统，基于<b>RAG（检索增强生成）</b>技术架构。</div><div class="notion-text notion-block-74c2f6ea3b974724b8677ee012b0832c"><b>RAG的核心原理</b>是将「知识检索」与「语言生成」结合：</div><div class="notion-text notion-block-e38ce552046d4f089713f95f3e4f2115">用户提问 → Embedding编码 → 向量检索 → 相关文档召回 → LLM生成回答</div><div class="notion-text notion-block-bc373f8e39fb4e05880de6f7eb9563f6"><b>实践案例：企业知识库搭建</b></div><div class="notion-text notion-block-0260f6c71ec5430e8aa8e7a6cebac5fb">我将公司产品手册、技术文档、FAQ等资料投喂到FastGPT，搭建了一个内部问答系统。具体配置经验：</div><table class="notion-simple-table notion-block-76b8ec855a444ffe80aae34154a4f7cd"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-b30186ef499b4ef89546069a76fed5f8"><td class="" style="width:120px"><div class="notion-simple-table-cell">配置项</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">我的设置</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">效果分析</div></td></tr><tr class="notion-simple-table-row notion-block-38eeffac8af648208f8dd31da1eb4087"><td class="" style="width:120px"><div class="notion-simple-table-cell">Embedding模型</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">m3e-large（中文优化）</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">中文检索效果明显优于通用模型</div></td></tr><tr class="notion-simple-table-row notion-block-05d39238ab81452589d89f4353095c7d"><td class="" style="width:120px"><div class="notion-simple-table-cell">文档切片大小</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">500 tokens</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">平衡召回精度与上下文完整度</div></td></tr><tr class="notion-simple-table-row notion-block-7b53b03355fc49e68704d376a83f2fc8"><td class="" style="width:120px"><div class="notion-simple-table-cell">召回数量</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">top-5</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">超过5条会引入噪音，降低回答质量</div></td></tr><tr class="notion-simple-table-row notion-block-64dac2645abd46df9ed00ea21ede82e0"><td class="" style="width:120px"><div class="notion-simple-table-cell">Rerank</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">开启</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">对长文档知识库效果显著，召回准确率提升约30%</div></td></tr></tbody></table><div class="notion-text notion-block-91b78b91fa4c4d30ba24628a7494b660"><b>踩坑经验</b>：</div><div class="notion-text notion-block-490b1539031641b29d78b1d7334ddf78">初期使用默认的text-embedding-ada-002模型，中文检索效果很差——用户问「产品安装流程」，系统召回的是「产品卸载流程」相关文档。切换到中文优化的m3e模型后，问题得到解决。</div><div class="notion-text notion-block-e14cca09c5454c9da8dc1f564bb2c64a"><b>关键认知</b>：Embedding模型的选择对RAG效果有决定性影响。<b>中文场景必须使用中文优化的Embedding模型</b>，这是很多初学者容易忽略的关键点。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-2a0e571054314973af159e980f99b384"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/ai-tools-efficiency/rag-architecture.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=2a0e5710-5431-4973-af15-9e980f99b384" alt="RAG技术架构流程" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">RAG技术架构流程</figcaption></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-86f8e7f628594b1e9c977b383f5ad130" data-id="86f8e7f628594b1e9c977b383f5ad130"><span><div id="86f8e7f628594b1e9c977b383f5ad130" class="notion-header-anchor"></div><a class="notion-hash-link" href="#86f8e7f628594b1e9c977b383f5ad130" title="2.4 工作流认知启蒙"><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">2.4 工作流认知启蒙</span></span></h4><div class="notion-text notion-block-b1c974ad6d374e8caf61fb56e6ceae26">FastGPT的另一个重要收获是<b>工作流编排</b>的认知启蒙。开始接触：</div><ul class="notion-list notion-list-disc notion-block-4c2c228fa0e74e75844527463c99c61d"><li>工作流节点设计</li></ul><ul class="notion-list notion-list-disc notion-block-1b687e7eee654b54a63511f8b2531d5b"><li>工具节点（调用外部API）</li></ul><ul class="notion-list notion-list-disc notion-block-8e2241028b2441c8be37fceed667a81b"><li>LLM节点（参数调优）</li></ul><div class="notion-text notion-block-9cbffe765fc94c19ae23d4dd34ca1996">这为后续深入工作流工具奠定了基础。</div><div class="notion-text notion-block-a1d37528a1ec4acbaed1aaf4f8e22243"><b>FastGPT的转折点</b>：开源免费版限制最多32个知识库。这个限制推动我转向更开放的方案——<b>Dify</b>。</div><hr class="notion-hr notion-block-4436e6b60f0941d487e85a68b1b5a404"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-a806f99ca055442fb296459548a10daa" data-id="a806f99ca055442fb296459548a10daa"><span><div id="a806f99ca055442fb296459548a10daa" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a806f99ca055442fb296459548a10daa" title="三、Dify深度阶段：工作流编排的成熟实践（2024）"><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">三、Dify深度阶段：工作流编排的成熟实践（2024）</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-28992706838e4ddabf2330694e74cafc" data-id="28992706838e4ddabf2330694e74cafc"><span><div id="28992706838e4ddabf2330694e74cafc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#28992706838e4ddabf2330694e74cafc" title="3.1 工作流架构的深度理解"><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">3.1 工作流架构的深度理解</span></span></h4><div class="notion-text notion-block-79c18b5444b84974a888d6a6cd31637c">进入Dify阶段，是对<b>Workflow（工作流）架构</b>真正深入理解的时期。</div><div class="notion-text notion-block-0af02871d5104f6e970b4854aec8f03c">工作流的本质是<b>将复杂任务分解为有序的处理步骤，每个步骤由特定节点负责</b>：</div><div class="notion-text notion-block-8d0de59620534b798471231dd7ff42af">输入节点 → 预处理节点 → 知识库检索节点 → LLM处理节点 → 后处理节点 → 输出节点</div><div class="notion-text notion-block-5616f9a9131845158b4911c6b9155a61">在Dify中完成了大量编排实践：</div><ul class="notion-list notion-list-disc notion-block-1e2ecc39cdf748f5b65b08e8b2b843e3"><li>知识库编排（多知识库协同）</li></ul><ul class="notion-list notion-list-disc notion-block-5b2a84b9a679420d870dc6c1b0f6df03"><li>工作流编排（复杂任务拆解）</li></ul><ul class="notion-list notion-list-disc notion-block-581c5cc40fee414688b7921e437adff6"><li>单一场景智能体制作（画图、会议纪要、可视化等）</li></ul><div class="notion-text notion-block-4b3b0a66c2404041b4ef2cb73ce44881"><b>案例：会议纪要智能体</b></div><div class="notion-text notion-block-a185015ded014b21a8d6a95d9f9c340b">这是我在Dify中编排的一个典型工作流：</div><table class="notion-simple-table notion-block-b2f1c0c880f648529785a75e35152718"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-6209c62065a04d07a41ece29e1b6dfc7"><td class="" style="width:120px"><div class="notion-simple-table-cell">节点</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">功能</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">配置要点</div></td></tr><tr class="notion-simple-table-row notion-block-394f79b7ea364c68a078a06661fe4c26"><td class="" style="width:120px"><div class="notion-simple-table-cell">输入节点</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">接收会议录音转文字</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">支持文件上传和文本粘贴</div></td></tr><tr class="notion-simple-table-row notion-block-0f313fdbcc2f47e3bc70f0a01f2de07d"><td class="" style="width:120px"><div class="notion-simple-table-cell">预处理节点</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">清理转录噪音、分段</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">正则去除时间戳、空行</div></td></tr><tr class="notion-simple-table-row notion-block-5fcc54de9188452e8ce6d5b885a2db45"><td class="" style="width:120px"><div class="notion-simple-table-cell">LLM节点1</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">提取议题和决议</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">使用GPT-4，温度0.1保证稳定性</div></td></tr><tr class="notion-simple-table-row notion-block-080bb27c2ce14956967d7696a5247a9b"><td class="" style="width:120px"><div class="notion-simple-table-cell">LLM节点2</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">整理行动项和责任人</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">结构化输出格式</div></td></tr><tr class="notion-simple-table-row notion-block-794306b3c744428ab6fe8a08ec149dae"><td class="" style="width:120px"><div class="notion-simple-table-cell">输出节点</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">生成结构化纪要</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Markdown格式便于分享</div></td></tr></tbody></table><div class="notion-text notion-block-a96c9d3fd2fc409eba4bc49d5480d8a8"><b>效果</b>：将原本需要30分钟手动整理的会议纪要，缩短到2分钟内自动生成，准确率约85%。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-6957969256f64cbaa9d84433de921dc7" data-id="6957969256f64cbaa9d84433de921dc7"><span><div id="6957969256f64cbaa9d84433de921dc7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#6957969256f64cbaa9d84433de921dc7" title="3.2 Agent框架与工具调用"><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">3.2 Agent框架与工具调用</span></span></h4><div class="notion-text notion-block-b04aacf4eab94fc0a35f7263454676e2">这个阶段进一步深入<b>Agent（智能体）框架</b>：</div><table class="notion-simple-table notion-block-494af826ea714908986993f360143326"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-52da4e2b38e249c9a24be53ae0de96cd"><td class="" style="width:120px"><div class="notion-simple-table-cell">概念</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">理解</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">实践体会</div></td></tr><tr class="notion-simple-table-row notion-block-462073dad98a465896e18c8b70df8275"><td class="" style="width:120px"><div class="notion-simple-table-cell">ReAct框架</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Reasoning + Acting，让模型「先思考再行动」</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">适合需要多步骤决策的任务，但会增加token消耗</div></td></tr><tr class="notion-simple-table-row notion-block-0dbde4ba7525439aa88afe61db2e6b71"><td class="" style="width:120px"><div class="notion-simple-table-cell">工具调用</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">让Agent调用外部工具扩展能力</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">HTTP请求工具最常用，可对接任意API</div></td></tr><tr class="notion-simple-table-row notion-block-5d90b215de8245dba53cb1298c6fb29a"><td class="" style="width:120px"><div class="notion-simple-table-cell">MCP</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Anthropic推出的标准化工具协议</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">在Claude Code中体验最深，Skills本质是MCP的延伸</div></td></tr></tbody></table><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-57d39305b9344af0937e77110f9858b7" data-id="57d39305b9344af0937e77110f9858b7"><span><div id="57d39305b9344af0937e77110f9858b7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#57d39305b9344af0937e77110f9858b7" title="3.3 复杂智能体实践"><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">3.3 复杂智能体实践</span></span></h4><div class="notion-text notion-block-8339074976614fa890d335491423c7da">编排过两个复杂的工作流智能体：</div><div class="notion-text notion-block-edd9f9b90a824e0f9ffc88451ebf9410"><b>案例1：旅行规划智能体</b></div><div class="notion-text notion-block-a97fecb8c41e49f2a68dc223c3355481">这个智能体整合了多个信息源：</div><ul class="notion-list notion-list-disc notion-block-5ac89d4d0a1a4f3997fb037bc7fd9ffc"><li>目的地知识库（景点、酒店、交通信息）</li></ul><ul class="notion-list notion-list-disc notion-block-6fa4669a8a9341cf911eb0a2eedc0c6f"><li>天气查询API（判断出行时间）</li></ul><ul class="notion-list notion-list-disc notion-block-48b0725ab5b74b7e8b7f5dc952748023"><li>预算计算节点（用户输入预算→自动推荐方案）</li></ul><ul class="notion-list notion-list-disc notion-block-37adfeb13ac541a0b2dd94eda3b89c29"><li>行程生成节点（按天编排行程）</li></ul><div class="notion-text notion-block-a65c2167b8c84d85892d24402b08eb28"><b>具体输出示例</b></div><div class="notion-text notion-block-9a396f3150314f3fb897ccd06dab3a9d">用户输入：北京3天旅行，预算3000元，喜欢历史人文</div><div class="notion-text notion-block-be88d5b88b374a24bd5540652c57b0a4">AI输出行程（节选）：</div><ul class="notion-list notion-list-disc notion-block-718e65b9e4ae4602b525d972d61a61ab"><li>Day 1：故宫 + 国家博物馆，门票约60元，地铁直达</li></ul><ul class="notion-list notion-list-disc notion-block-ceec157a9de74fa9b2531196842a3a6c"><li>Day 2：颐和园 + 圆明园，门票约55元</li></ul><ul class="notion-list notion-list-disc notion-block-357516d65c9446b982a594d80d81fb92"><li>Day 3：天坛 + 景山公园，门票约17元</li></ul><div class="notion-text notion-block-df0d38f2b7574cb7b7f4267f0368346b">总花费约722元（含住宿），远低于预算，行程规划合理。</div><div class="notion-text notion-block-fbde543c7d014caabc43ae1a1f56a798"><b>踩坑经验</b>：</div><ul class="notion-list notion-list-disc notion-block-7f8d117720c041e58ad4510a67db6fd0"><li>天气API对接失败 → 行程没有考虑雨天备选方案</li></ul><ul class="notion-list notion-list-disc notion-block-ab199ab8712c4eae9a69e25bb3760e7d"><li>预算计算节点精度问题 → 餐饮费用低估</li></ul><ul class="notion-list notion-list-disc notion-block-7d0fefffb4004491b41da3b256ef5948"><li>解决：添加备用API、增加餐饮预算弹性系数</li></ul><div class="notion-text notion-block-848e1cf207334defb717f57ae9ab33a3"><b>局限</b>：智能体只能处理「规划一次旅行」这个固定场景。如果用户想「查询某地美食」或「对比两个目的地」，需要单独创建新的智能体。</div><div class="notion-text notion-block-20766fede0fa481e8bc8d832be112299"><b>案例2：自媒体全流程智能体</b></div><div class="notion-text notion-block-475a5e0bab504c11a87f9ae76cdfe29f">覆盖从选题到发布的完整流程：</div><ul class="notion-list notion-list-disc notion-block-c916e02fffd740de887224fca9dba43a"><li>选题建议（基于热点知识库）</li></ul><ul class="notion-list notion-list-disc notion-block-2a476e88bf57402f908609d191f26519"><li>内容生成（多风格输出）</li></ul><ul class="notion-list notion-list-disc notion-block-9d427aa092a940d99e2a3af8c3dc7b80"><li>排版建议（平台适配）</li></ul><ul class="notion-list notion-list-disc notion-block-e598087707da43e8bec91573d52f0d73"><li>发布时间建议（基于历史数据分析）</li></ul><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-d384fc9e12054bfc8bfd9ac1b221e682" data-id="d384fc9e12054bfc8bfd9ac1b221e682"><span><div id="d384fc9e12054bfc8bfd9ac1b221e682" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d384fc9e12054bfc8bfd9ac1b221e682" title="3.4 工作流智能体的局限"><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">3.4 工作流智能体的局限</span></span></h4><div class="notion-text notion-block-4f7d57c7910c4a9a9e4a78ccaa02bc7d">大量实践后得出关键结论：</div><blockquote class="notion-quote notion-block-39acd6b617284d798559aa2d73fb1739"><div><b>工作流智能体是单一场景、固定流程的。每当我需要做流程之外的工作，都需要单独调整工作流。</b></div></blockquote><div class="notion-text notion-block-5eb1b7767c56448f9e13db85a52a9ee4"><b>具体表现</b>：</div><ul class="notion-list notion-list-disc notion-block-e765286dd52047308eaf1655e3fec725"><li>想增加一个新功能 → 需要修改工作流节点</li></ul><ul class="notion-list notion-list-disc notion-block-03469c7915554fcab47c4c4c7202fb99"><li>想改变输出格式 → 需要调整模板配置</li></ul><ul class="notion-list notion-list-disc notion-block-8a6e62c9c7df4f169906885e3fb17d99"><li>想适配新场景 → 需要创建新的智能体</li></ul><div class="notion-text notion-block-6985572ac8064d5c9fbcbca97f51427b">这个认知推动了下一阶段的探索——<b>通用智能体</b>和<b>Vibe Coding</b>。</div><hr class="notion-hr notion-block-ae8c208db2274a83b574b02ef9a3720d"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-7e537306f03b4849a64e617226dafebf" data-id="7e537306f03b4849a64e617226dafebf"><span><div id="7e537306f03b4849a64e617226dafebf" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7e537306f03b4849a64e617226dafebf" title="四、Vibe Coding阶段：从辅助编程到日常赋能（2024-2025）"><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、Vibe Coding阶段：从辅助编程到日常赋能（2024-2025）</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-4198d06dc082489999dd9f1f59d5c3dc" data-id="4198d06dc082489999dd9f1f59d5c3dc"><span><div id="4198d06dc082489999dd9f1f59d5c3dc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4198d06dc082489999dd9f1f59d5c3dc" title="4.1 Vibe Coding的本质"><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">4.1 Vibe Coding的本质</span></span></h4><div class="notion-text notion-block-9793203a91cd47deaacbf1c2be01b32b">「Vibe Coding」这个词最早由Andrej Karpathy提出，核心意思是：<b>用自然语言描述需求，让AI理解意图并生成代码，用户只需「感受」结果是否符合预期</b>。</div><div class="notion-text notion-block-0bc9b3d777404568970f8ea7287a4941">这个阶段的代表性工具：</div><table class="notion-simple-table notion-block-368274c1674147718def1d72d767d6fa"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-69376ce557644a1083cb9fff1a0117d1"><td class="" style="width:120px"><div class="notion-simple-table-cell">工具</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">定位</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">特点</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">我的实际使用场景</div></td></tr><tr class="notion-simple-table-row notion-block-6a0a7fee39314f9b9f5ffee054de9844"><td class="" style="width:120px"><div class="notion-simple-table-cell">Claude Code</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Anthropic官方CLI</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">深度MCP集成、Skills扩展</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">日常办公主力工具，文档、表格、PPT</div></td></tr><tr class="notion-simple-table-row notion-block-9e4519265d374071adac9b63480537f4"><td class="" style="width:120px"><div class="notion-simple-table-cell">Cursor</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">VS Code fork</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">IDE内的AI辅助</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">正式编程项目开发</div></td></tr><tr class="notion-simple-table-row notion-block-d0be1347138b41b2a13748fdb64abe30"><td class="" style="width:120px"><div class="notion-simple-table-cell">通义灵码</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">阿里产品</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">国产模型驱动</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">中文文档场景备选</div></td></tr><tr class="notion-simple-table-row notion-block-fe0685747bf043a48de996951dc3b4cd"><td class="" style="width:120px"><div class="notion-simple-table-cell">Gemini CLI</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Google CLI工具</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Gemini模型驱动</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">长文档分析、多模态处理</div></td></tr></tbody></table><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-1e7e1e8673424655bca93a7c591dd9f7" data-id="1e7e1e8673424655bca93a7c591dd9f7"><span><div id="1e7e1e8673424655bca93a7c591dd9f7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#1e7e1e8673424655bca93a7c591dd9f7" title="4.2 Skills：能力扩展的关键产物"><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">4.2 Skills：能力扩展的关键产物</span></span></h4><div class="notion-text notion-block-cd72faea57f141139856a84a4c25ccfe">随着Vibe Coding工具的成熟，出现了<b>Skills</b>这个关键概念。</div><div class="notion-text notion-block-9f5b3d65fb164cc796606901449e7dcb">Skills是<b>可复用的能力模块</b>，定义了特定场景下的工作方式和工具调用。从寻找外部好用的Skills，到使用skill-creator自己创建，完成了从「使用者」到「创造者」的转变。</div><div class="notion-text notion-block-33d0b71eba874e61b63515a381b277ac"><b>案例：我创建的Skills</b></div><table class="notion-simple-table notion-block-e5f59c2b04e846cc844a16d46b3f2a01"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-1bbbd279baeb44188fba9f72002828e9"><td class="" style="width:120px"><div class="notion-simple-table-cell">Skill名称</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">功能</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">实际效果</div></td></tr><tr class="notion-simple-table-row notion-block-7928f9cf549943768e8841962e7b1499"><td class="" style="width:120px"><div class="notion-simple-table-cell">设备性能选型评估</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">根据业务需求自动计算服务器配置</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">选型报告生成从2小时缩短到10分钟</div></td></tr><tr class="notion-simple-table-row notion-block-be337aa482534c16830bec7034d92f32"><td class="" style="width:120px"><div class="notion-simple-table-cell">博客写作流程</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">从想法收集到NotionNext发布</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">本文正是通过这个Skill创作的</div></td></tr><tr class="notion-simple-table-row notion-block-1d98f685ccc846ee90b3d601410c7732"><td class="" style="width:120px"><div class="notion-simple-table-cell">内容分析</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">分析文档并提炼关键信息</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">会议纪要、方案文档一键摘要</div></td></tr><tr class="notion-simple-table-row notion-block-1e74a85cdf4d40aaae9fe21540ed4463"><td class="" style="width:120px"><div class="notion-simple-table-cell">PPT生成</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">根据内容大纲生成PPT文件</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">方案展示效率提升5倍</div></td></tr></tbody></table><div class="notion-text notion-block-ed959381da9f49068262c497b9e63445"><b>关键发现</b>：Skills的本质是<b>将重复性工作固化成可复用的能力</b>。每个Skill相当于一个「专属助手」，了解我的工作方式和偏好。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-2464590c105347b68d68e952776ea064"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/ai-tools-efficiency/vibe-coding-matrix.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=2464590c-1053-47b6-8d68-e952776ea064" alt="Vibe Coding工具矩阵" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">Vibe Coding工具矩阵</figcaption></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-bbfbfa93ebc74b8f82ba028cfa73d9e9" data-id="bbfbfa93ebc74b8f82ba028cfa73d9e9"><span><div id="bbfbfa93ebc74b8f82ba028cfa73d9e9" class="notion-header-anchor"></div><a class="notion-hash-link" href="#bbfbfa93ebc74b8f82ba028cfa73d9e9" title="4.3 认知突破：AI Coding不只是编程"><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">4.3 认知突破：AI Coding不只是编程</span></span></h4><div class="notion-text notion-block-0843f7f6d3374170a0be6383a79bf749">最重要的认知升级是：</div><blockquote class="notion-quote notion-block-d90a626d67974f6e80ed3c601c147d12"><div><b>AI Coding工具设计之初是为了辅助编程/开发。但深入了解后发现，整个开发过程覆盖的工作——写文档、内容分析、画图、做表、方案展示——正是我们日常工作的全部。</b></div></blockquote><div class="notion-text notion-block-6b54417a86ae432eb7e7821a3fa3f675"><b>具体转变</b>：</div><table class="notion-simple-table notion-block-3a97fb87e7f34d9ca62726449a067801"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-f667e0bc0c0544b488426936fab53cb3"><td class="" style="width:120px"><div class="notion-simple-table-cell">原认知</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">新认知</div></td></tr><tr class="notion-simple-table-row notion-block-85cd631658ea481d872b86db7051646e"><td class="" style="width:120px"><div class="notion-simple-table-cell">Claude Code = 编程辅助工具</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Claude Code = 全场景办公助手</div></td></tr><tr class="notion-simple-table-row notion-block-c3b5fca41b2a45ac9e87528659233efc"><td class="" style="width:120px"><div class="notion-simple-table-cell">Skills = 代码模板</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Skills = 工作流程固化</div></td></tr><tr class="notion-simple-table-row notion-block-27c1d616acd14ba7835f7e0aaff93b1f"><td class="" style="width:120px"><div class="notion-simple-table-cell">MCP = 开发者工具协议</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">MCP = 办公能力扩展协议</div></td></tr></tbody></table><div class="notion-text notion-block-12edc58af5954065a04a058d83a38dd6">从此，AI Coding工具正式成为日常工作的核心助手。</div><hr class="notion-hr notion-block-d2b1bacf4f1d4c458a41ff271853f898"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-fb4874f993724de78905f4ef2efb0873" data-id="fb4874f993724de78905f4ef2efb0873"><span><div id="fb4874f993724de78905f4ef2efb0873" class="notion-header-anchor"></div><a class="notion-hash-link" href="#fb4874f993724de78905f4ef2efb0873" title="五、通用智能体与聪明Chatbot（同期并行）"><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">五、通用智能体与聪明Chatbot（同期并行）</span></span></h3><div class="notion-text notion-block-609818e26f35468987f30fecf3797408">在工作流智能体的局限认知后，开始探索：</div><table class="notion-simple-table notion-block-cdf8093563e24cd1a8b635c82857cee8"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-e558611be477455da46896a6ebd95645"><td class="" style="width:120px"><div class="notion-simple-table-cell">方向</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">代表工具</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">特点</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">适用场景</div></td></tr><tr class="notion-simple-table-row notion-block-e31eca74bdf64c5993531333baed6325"><td class="" style="width:120px"><div class="notion-simple-table-cell">通用智能体</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Manus</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">不限定场景，自主规划执行</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">一次性复杂任务（调研、规划）</div></td></tr><tr class="notion-simple-table-row notion-block-d72124c858714496adff985bbe0daf94"><td class="" style="width:120px"><div class="notion-simple-table-cell">聪明Chatbot</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Gemini 2.5 Pro</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">深度推理、长上下文、多模态</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">深度分析、长文档解读</div></td></tr></tbody></table><div class="notion-text notion-block-90af97c2753e4a52a0df5f4f618e91e0">通用智能体的核心差异是<b>不依赖预设流程，而是根据任务自主拆解和执行</b>。这解决了工作流「固定流程」的痛点，但也带来了新的挑战——可控性和确定性降低。</div><div class="notion-text notion-block-ae80d6c2d4214fb7bed59024023aa962"><b>Manus竞品调研案例</b></div><div class="notion-text notion-block-655b85208c594861927f6972cd7aa7c5">任务：调研三家竞品公司的产品功能对比</div><div class="notion-text notion-block-7ddb7b0194dc48b49d41c7183b7b126c">Manus自主执行过程：</div><ol start="1" class="notion-list notion-list-numbered notion-block-a239b83a9e3e49a792805ec11ec76056" style="list-style-type:decimal"><li>搜索三家公司官网和产品文档</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-a31175d512a14dc2a945b42b51cde000" style="list-style-type:decimal"><li>整理功能清单（登录认证、权限管理、数据导出等）</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-5720c191e8df4f4fb9497e0a9001e6ef" style="list-style-type:decimal"><li>对比差异并标注优劣势</li></ol><ol start="4" class="notion-list notion-list-numbered notion-block-8d562da7acd54d428423d7b504a1e9c9" style="list-style-type:decimal"><li>生成结构化对比报告</li></ol><div class="notion-text notion-block-3708ec70461141aba751f5b4bcf3c79c">输出要点：</div><ul class="notion-list notion-list-disc notion-block-a1d5698c2d804082a14bab4535ca27b9"><li>竞品A：认证功能强，但数据导出受限</li></ul><ul class="notion-list notion-list-disc notion-block-58fe00cf3668438aa04f80ca5ef0210a"><li>竞品B：权限管理细粒度高，用户体验一般</li></ul><ul class="notion-list notion-list-disc notion-block-80422e75fd75476bad34c4fe130113e9"><li>竞品C：综合均衡，性价比最优</li></ul><div class="notion-text notion-block-ff524ba5b1e74b769c3bdcf56a5f5bc3">耗时：15分钟自主完成，人工预估至少需2小时</div><div class="notion-text notion-block-2b642abdfec5447ab62f950c1e62ab88"><b>Gemini长文档分析案例</b></div><div class="notion-text notion-block-91feade36f1441bc8f4bc38e24458016">任务：分析50页技术方案文档，提取核心架构</div><div class="notion-text notion-block-564398bef32648d1a5a20887a1103c5e">Gemini处理能力发挥：</div><ul class="notion-list notion-list-disc notion-block-2e5088fe51dc43ab9a3a48f2dacda9f9"><li>上下文：完整读取50页内容</li></ul><ul class="notion-list notion-list-disc notion-block-cf991e509cd544d291d3ca446c8f2e23"><li>深度推理：理解各模块之间的依赖关系</li></ul><ul class="notion-list notion-list-disc notion-block-a6c8e47657054c08bf438f4131c9d3d3"><li>结构化输出：生成架构描述和关键决策点</li></ul><div class="notion-text notion-block-ee35f24c0a3e454a8151afac8a121ddd">输出要点：</div><ol start="1" class="notion-list notion-list-numbered notion-block-4b25a20ab7e54afca382e7a45fbb7f51" style="list-style-type:decimal"><li>核心架构：三层微服务（API层、业务层、数据层）</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-9b2ba9251a1d402bb511a0ffc9915ee9" style="list-style-type:decimal"><li>关键决策：选择Redis而非Memcached的原因分析</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-c3fd2e5e7a3e44548a61c8b84e2c4e7a" style="list-style-type:decimal"><li>潜在风险：数据库连接池配置未考虑峰值场景</li></ol><div class="notion-text notion-block-001c9e5b204549cab7e9e375e203b210">耗时：5分钟，人工阅读理解至少需30分钟</div><div class="notion-text notion-block-f7602ea480cb43d1a2ac587b9391e486"><b>选择决策案例</b>：</div><table class="notion-simple-table notion-block-e74502e2e368449c80f88238bc9e71ec"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-07300292d45743a39364326e997d2d9e"><td class="" style="width:120px"><div class="notion-simple-table-cell">任务类型</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">我的选择</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">理由</div></td></tr><tr class="notion-simple-table-row notion-block-98d2fae27f21408d895fa410c5014a61"><td class="" style="width:120px"><div class="notion-simple-table-cell">每周销售数据汇总报告</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Dify工作流</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">固定流程，高频重复，可固化</div></td></tr><tr class="notion-simple-table-row notion-block-f67da28b338f4dd69717daa6d0884ebd"><td class="" style="width:120px"><div class="notion-simple-table-cell">调研竞品产品功能</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Manus</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">一次性复杂任务，需自主规划路径</div></td></tr><tr class="notion-simple-table-row notion-block-e76bd7bb19904281b82fdcd9140b16e1"><td class="" style="width:120px"><div class="notion-simple-table-cell">分析50页技术方案文档</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Gemini 2.5 Pro</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">上下文，深度推理需求</div></td></tr><tr class="notion-simple-table-row notion-block-7b13ffc8ab244f289d984b20fe75bb2c"><td class="" style="width:120px"><div class="notion-simple-table-cell">日常文档、表格、PPT</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Claude Code + Skills</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">全场景覆盖，效率最高</div></td></tr></tbody></table><hr class="notion-hr notion-block-97e00e7328974627b24874247bbb58f4"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-1eb165b0bc8d416b8a149fb4485cba87" data-id="1eb165b0bc8d416b8a149fb4485cba87"><span><div id="1eb165b0bc8d416b8a149fb4485cba87" class="notion-header-anchor"></div><a class="notion-hash-link" href="#1eb165b0bc8d416b8a149fb4485cba87" title="六、总结与感悟"><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">六、总结与感悟</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-b2b21e396adf406a9241d2050afab51e" data-id="b2b21e396adf406a9241d2050afab51e"><span><div id="b2b21e396adf406a9241d2050afab51e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b2b21e396adf406a9241d2050afab51e" title="6.1 Token花费 = 学习投入"><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">6.1 Token花费 = 学习投入</span></span></h4><blockquote class="notion-quote notion-block-4566a72c02604d8dbbff4bcdcabb2b2a"><div><b>在AI时代，你的tokens花费，代表你在AI领域学习和了解的深度和投入。</b></div></blockquote><div class="notion-text notion-block-478f78769eb245098564daf5b8169a79">这不是消费，是投资。每一次尝试、每一次失败、每一次调优，都在积累真正的实战经验。纸上得来终觉浅，绝知此事要躬行。</div><div class="notion-text notion-block-d96e21c0405c472aa67e597dd879dfc8"><b>我的花费估算</b>：</div><table class="notion-simple-table notion-block-1c1ce9686a9141edb06eca18bf2d290f"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-254f2507cd8345a492dac3e771fbb38a"><td class="" style="width:120px"><div class="notion-simple-table-cell">阶段</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">主要花费项</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">估计金额</div></td></tr><tr class="notion-simple-table-row notion-block-ab6cc5a7227c48ca901057cbe8d3c354"><td class="" style="width:120px"><div class="notion-simple-table-cell">提示词探索</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">ChatGPT Plus订阅</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">约500元</div></td></tr><tr class="notion-simple-table-row notion-block-b0c207b221da4ace8d7ca530a6abea10"><td class="" style="width:120px"><div class="notion-simple-table-cell">知识库实践</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">API调用 + 云服务器</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">约800元</div></td></tr><tr class="notion-simple-table-row notion-block-a29cf1638c7a43218733cbfb62e36382"><td class="" style="width:120px"><div class="notion-simple-table-cell">Dify阶段</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">API调用（大量测试）</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">约1500元</div></td></tr><tr class="notion-simple-table-row notion-block-d89ac270cc484a7c8f8e6e0c2d9cc945"><td class="" style="width:120px"><div class="notion-simple-table-cell">Vibe Coding</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Claude Pro + 各类API</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">约2000元</div></td></tr><tr class="notion-simple-table-row notion-block-a7c7c88cbf184edfa8eab0b744f629e8"><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>总计</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">ㅤ</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>约4800元</b></div></td></tr></tbody></table><div class="notion-text notion-block-fcf8ecbde4fb47d486e093e390cfa0da">这些投入换来的是：对企业级AI应用的完整认知、工作流编排的实战能力、日常效率的显著提升。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-dc32f4c309514596bf61ba979cad0950"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/ai-tools-efficiency/tool-selection.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=dc32f4c3-0951-4596-bf61-ba979cad0950" alt="工具场景匹配决策" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">工具场景匹配决策</figcaption></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-9ab728654bf84ce6b9dc63fafd20c729" data-id="9ab728654bf84ce6b9dc63fafd20c729"><span><div id="9ab728654bf84ce6b9dc63fafd20c729" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9ab728654bf84ce6b9dc63fafd20c729" title="6.2 没有最好的工具，只有匹配的场景"><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">6.2 没有最好的工具，只有匹配的场景</span></span></h4><blockquote class="notion-quote notion-block-cbaa7de1c8134dfcb422e57320921183"><div><b>只有不断尝试、使用，才能发现每个工具到底适合用在什么地方，能够真正解决什么问题。没有最好的工具，只有匹配的场景。</b></div></blockquote><div class="notion-text notion-block-f14996e97020477ba2550d5deae0f2cf">工具选择的核心原则：</div><table class="notion-simple-table notion-block-403008167f64487a8d3aa1c3b672d177"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-3cea11db1bd047aea4c9008f12773b69"><td class="" style="width:120px"><div class="notion-simple-table-cell">场景特征</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">推荐工具类型</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">典型案例</div></td></tr><tr class="notion-simple-table-row notion-block-189cf79a9e6e427dbcf17bd3ed656382"><td class="" style="width:120px"><div class="notion-simple-table-cell">固定流程、高频重复</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">工作流智能体（Dify）</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">每周报告生成、会议纪要整理</div></td></tr><tr class="notion-simple-table-row notion-block-d736a5fff6b649239fd692515ec13e13"><td class="" style="width:120px"><div class="notion-simple-table-cell">知识问答、企业资料</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">RAG知识库方案</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">产品FAQ、内部知识问答</div></td></tr><tr class="notion-simple-table-row notion-block-061cba48344047b8b8c7143363046aaf"><td class="" style="width:120px"><div class="notion-simple-table-cell">一次性复杂任务</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">通用智能体（Manus）</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">竞品调研、方案规划</div></td></tr><tr class="notion-simple-table-row notion-block-804d297e4dc14df89061d4fff06d33ac"><td class="" style="width:120px"><div class="notion-simple-table-cell">日常办公、文档处理</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Vibe Coding + Skills</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">表格、PPT、文档、分析</div></td></tr><tr class="notion-simple-table-row notion-block-3a69b65609014423897f4ca35e091099"><td class="" style="width:120px"><div class="notion-simple-table-cell">深度推理、长文档分析</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">聪明Chatbot（Gemini）</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">技术方案解读、合同审核</div></td></tr></tbody></table><hr class="notion-hr notion-block-dc0f7b638e184820bcf09466b2c3b595"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-c6f31989d2114c45bd4ef60b01c34c13" data-id="c6f31989d2114c45bd4ef60b01c34c13"><span><div id="c6f31989d2114c45bd4ef60b01c34c13" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c6f31989d2114c45bd4ef60b01c34c13" title="七、技术架构演进全景"><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">七、技术架构演进全景</span></span></h3><div class="notion-text notion-block-d57dc8e2b00f431caae1e06499f5a7d4">从架构视角看，AI工具的演进是能力边界的持续扩展：</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-f21aa7beb5a94eae9b6e3ad36e503e80"><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://skillre-book.oss-cn-beijing.aliyuncs.com/blog/assets/ai-tools-efficiency/ai-tools-timeline.png?spaceId=1747ca02-959a-4120-895e-43842d3a9af1&amp;t=f21aa7be-b5a9-4eae-9b6e-3ad36e503e80" alt="AI工具演进时间线" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">AI工具演进时间线</figcaption></div></figure><table class="notion-simple-table notion-block-55b3129c7c354688bedc1127f2a2a0ec"><tbody><tr class="notion-simple-table-row notion-simple-table-header-row notion-block-f88c9955a4bb44b3b6ef140a9003053c"><td class="" style="width:120px"><div class="notion-simple-table-cell">阶段</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">代表工具</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">核心能力</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">解决的痛点</div></td></tr><tr class="notion-simple-table-row notion-block-9141c20fba94496bbc0700082c92513a"><td class="" style="width:120px"><div class="notion-simple-table-cell">阶段1</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">ChatGPT</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">自然语言理解与生成</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">AI对话能力的认知启蒙</div></td></tr><tr class="notion-simple-table-row notion-block-400d34caa6994da59a9e064fee2be5fb"><td class="" style="width:120px"><div class="notion-simple-table-cell">阶段2</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">ChatGPT Next Web</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">场景化对话能力复用</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">提示词复用和共享</div></td></tr><tr class="notion-simple-table-row notion-block-d312ee7f1d9144649439a44db8cd7e3b"><td class="" style="width:120px"><div class="notion-simple-table-cell">阶段3</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">FastGPT + RAG</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">私有知识接入</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">企业知识问答</div></td></tr><tr class="notion-simple-table-row notion-block-dd7a7e31d9ef49418e06e7ee4289772f"><td class="" style="width:120px"><div class="notion-simple-table-cell">阶段4</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Dify + Workflow</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">复杂任务自动化</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">多步骤任务编排</div></td></tr><tr class="notion-simple-table-row notion-block-4e9e56e8271e4262a4b10e4ec69a3f25"><td class="" style="width:120px"><div class="notion-simple-table-cell">阶段5</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Manus + MCP</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">自主规划执行</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">固定流程的局限</div></td></tr><tr class="notion-simple-table-row notion-block-30cc270b1bcd4082b9ff6fc8874bf410"><td class="" style="width:120px"><div class="notion-simple-table-cell">阶段6</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Vibe Coding + Skills</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">全场景办公支持</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">日常工作效率</div></td></tr></tbody></table><div class="notion-text notion-block-1200082f011640189a0b106594dd12de">每个阶段都在解决上一阶段的痛点，也在暴露新的局限。<b>AI工具的探索，本质是认知的迭代</b>。</div><hr class="notion-hr notion-block-f6f154dd58ed4e0b924a667bb0eb487a"/><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-2c815c5de0b34484b3ca495504fb170a" data-id="2c815c5de0b34484b3ca495504fb170a"><span><div id="2c815c5de0b34484b3ca495504fb170a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2c815c5de0b34484b3ca495504fb170a" title="给你的行动建议"><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">给你的行动建议</span></span></h3><div class="notion-text notion-block-92afa6c8ae7847b7827ac2bc74b6ec67">读完这篇文章，你可能想知道「我应该从哪里开始？」以下是我的建议：</div><div class="notion-text notion-block-83ca69c9fa01438684071d21131c93b4"><b>如果你是初学者</b>，刚接触AI工具：</div><ol start="1" class="notion-list notion-list-numbered notion-block-8b23445a27f848a4bd2514a93d6a6ee6" style="list-style-type:decimal"><li>先从ChatGPT或Claude开始，感受对话能力</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-3819dd7ca1ff4a4e837c2c3f15265c60" style="list-style-type:decimal"><li>学习提示词工程，尝试角色设定和格式约束</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-c63143c7c22c4b7d9809dc9a7c99336b" style="list-style-type:decimal"><li>不要急于投入大量API调用费用，先建立认知</li></ol><div class="notion-text notion-block-8d6f4fccab4243e3a1e9e10fe2093e25"><b>如果你已经有一定经验</b>，想进一步提升效率：</div><ol start="1" class="notion-list notion-list-numbered notion-block-9bc05bc0152e43c38e8d30d01cea653f" style="list-style-type:decimal"><li>探索工作流编排工具（如Dify），将固定流程自动化</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-134777285eaa4053a6ffb4368623ebfe" style="list-style-type:decimal"><li>尝试Claude Code + Skills，将日常工作流程固化</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-90172b88ff374982a3b1a49744ecedda" style="list-style-type:decimal"><li>根据场景选择工具：固定流程用工作流，一次性复杂任务用通用智能体</li></ol><div class="notion-text notion-block-b75fb20434eb44cb8643ddb3dcf5fb68"><b>如果你是企业用户</b>，想在公司内部应用：</div><ol start="1" class="notion-list notion-list-numbered notion-block-827b6fb2d13f468a821e38aec075aa95" style="list-style-type:decimal"><li>搭建RAG知识库，解决内部知识问答需求</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-492f59d41c4d4b87a255110c170c0cff" style="list-style-type:decimal"><li>注意选择中文优化的Embedding模型</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-9eb55e5aee7f40be8f29918d30e3cbdc" style="list-style-type:decimal"><li>工作流智能体适合高频重复的固定流程</li></ol><div class="notion-text notion-block-676c8f923a3b4e9db8a35561d535e0d5"><b>最核心的建议</b>：<b>先投入，再判断价值</b>。4800元的投入换来的是完整的认知迭代，这笔投资远比盲目购买课程或书籍更有价值。纸上得来终觉浅，绝知此事要躬行。</div></main></div>]]></content:encoded>
        </item>
    </channel>
</rss>