Muse 拆解：同一套 Claw 内核，如何换一种方式与人相处。
作者丨Rhea
编辑丨李娜 岑峰
9 月 20 日前后，一条帖子在 X 上被大量转发。AI 应用联合创始人 Ansh Nanda 询问 Meta 新上线的个人 agent Muse，为什么它的系统文件与开源项目 OpenClaw 几乎相同。他为截图配上一句评价Muse is LITERALLY OpenClaw for normies。 Muse 就是给普通人用的 OpenClaw。“Muse is LITERALLY OpenClaw for normies”一帖在 X 上广泛传播，Friedman 次日回应。 雷峰网(公众号：雷峰网)开发者随后在 Muse 的工作区里找到了与 OpenClaw 同名的 SOUL.md、MEMORY.md。第二天 Meta 超级智能实验室产品负责人 Nat Friedman 就回应了：Muse 的代码从零构建，产品上深受 OpenClaw 启发。至于为什么沿用这些文件，他的解释是 OpenClaw 作者 Peter Steinberger “把这些东西做对了”。Sarah Perez 写的《Meta admits Muse's likeness to OpenClaw isn't a coincidence》我们在试用了Muse 一段时间后发现，这个 APP 最大的创新和突破就来自于那些技术圈看不上的东西，开发者往往能自己手动搭建很多工具，但他们却忽略了一个事实：大多数人其实不会为了一项能力先学会一整套工具。Muse 和 OpenClaw 用的是同一套个人 agent 范式，能力上其实没有什么代际差异。变化主要是在人和 agent 相处的方式上：OpenClaw 要人学机器，Muse 让机器学人。“学习人”有两层意思。一个是学会了用户，它从已有的社交账户和旧助手的记忆里认识这个人。另一个是学会了人和人相处的方式，用一段一直在的对话、一个看得见样子的角色、一句来得及时的回应，把之前那些终端、配置文件和日志都换掉了。这些东西过去都得用户自己先学会。本文结合笔者的实测，以及对 muse.ai 网页端公开资源和 Muse 安卓安装包的静态分析，看看 Muse 是怎样把这份学习成本挪走的。Muse 于 2026 年 9 月 8 日在美国上线，在 App、网页或 WhatsApp 里都能派活，App 关掉以后还在后台接着干。拆开内核看，Muse 和 OpenClaw 是同一套结构：agent 运行框架、能上网的浏览器、定时任务、用 Markdown 存的记忆，还有一份定义“它是谁”的人格文件。安卓版连手机的叫法也沿用了 OpenClaw，代码里把手机叫作“节点”，云端下命令，手机去执行。所以说这些零件早就有了，只是 Meta 有一个近乎天才的产品经理把它们全做对了。产品体验做得足够好就能吸引更多的人。大家熟悉的故事是上线两周，Muse 就登上了美区 App Store 免费榜第一。我们在未登录状态下读取了 muse.ai 网页端公开下发的页面、脚本和界面翻译表，没有调用鉴权接口，也没有运行客户端。其中的简体中文界面文案表共 6,350 条，近九成已经译好，也就是说 Muse 还没在中文地区上线，中文界面倒是基本备齐了。这份文案表就像产品的词典，看得出它打算怎么跟用户说话。后来笔者又拿到了 Muse 的安卓安装包（12.0.0.14.169），同样只做了静态分析，没安装也没运行。包里的中文也译了近九成，只是还没统稿，有意思的是有几处把 agent 译成了“客服”。右边是文案表的开头，一行一句界面文案。左边的 89%，就是从这 6,350 条里数出来的这份继承一直延伸到第一次见面的仪式。OpenClaw 的 BOOTSTRAP.md 让 agent 在对话里醒来，和用户商量名字、气质和签名 emoji，再写进 IDENTITY.md。Muse 的身份文件存的也是这几样，项目代号叫 Hatch，意为孵化。Muse 沿用了 OpenClaw 工作区的核心人格文件都是同样的流程，OpenClaw 是在终端里面搞定，Muse 却把它搬进了角色定制界面。Meta 留下了开源社区摸索出来的骨架，但是骨架之外的体验几乎全部重写。我们稍微总结一下，可以把重写的部分拆成五个从用户身上挪走的相处成本。OpenClaw 的第一道门槛是一台电脑，你得自己准备、自己配置、自己维护。扎克伯格的判断是，不会有几十亿人买一台 Mac Studio 回家自己配。Muse 给每个用户一台电脑但是不需要用户自己去运维（天知道 openclaw 的运维有多花时间！）。你只需要用 Meta 账号一登录，同一账户中心下的 Facebook、Instagram 和 Threads 就自动连上了。背后的云端电脑平时不露面，只在需要的时候冒出一张浏览器卡片，或者一个请用户接管的按钮。唯一需要用户做的就只是打开 muse.ai 网页，或者下载 muse APP.（当然国内用户需要的做的远远不止这些）云端电脑平时不可见，只在需要时以浏览器卡片或接管请求的形式出现其实能力越宽，空白输入框越难填。几乎所有 Agent 产品都会面临这个问题：第一件事到底该让 agent 做什么？Muse 的办法是自己先开口。“点子”页列出它能替用户办的事，点一下“Let's do it”就开工；“动态”是它按自己的节奏写的短帖和播客；“目标”留给长期的事，它先问清楚，再排计划、定期汇报。Muse 根据已掌握的信息，主动提出可以替用户完成的事动态页的冷启动产品体验也做得很好。用户第一次看到动态页，会先看到一段提示，说明这里是干什么的、有什么价值。什么用户信息都没有的时候去浏览，看到的是一些功能说明和状态提示，但它没有直接留白，而是巧妙地用未来会出现的帖子的形态来介绍功能。动态页还支持修改 prompt，也就是说用户可以看到这个动态页是怎么运作的，你甚至可以直接修改它的运作方式。动态页冷启动时不留白，用未来帖子的样子介绍自己，还把运作方式交给用户修改Meta 给用户安排的第一件事跟钱有关。首席 AI 官 Alexandr Wang 在 X 上发起 #MuseMoneyChallenge，让 Muse 帮用户省下 1,000 美元。用户不用先弄懂 agent 是什么，看得懂账单就行。Meta 把“帮用户省钱”设计为 Muse 的第一项任务我们在网页静态逆向中发现，其实用户还没开口，Muse 已经在创建虚拟机时给自己定好了一个默认目标。空白页的口吻也像交朋友：“随着我们彼此了解，会逐渐呈现。”Muse 设计总监 Alex Cornell 说，团队一直拿同一个问题来校准：如果用户是在跟另一个人说话，这件事该怎么运作？Agent 于是被放在屏幕顶部，像通讯录里的一个联系人。主对话只有一个，就像微信里跟一个人只有一个窗口。发出请求，Muse 先回一个 ?，再去后台干活。Muse 先确认收到，再进入后台执行，交互方式接近给朋友发消息这种取舍在用词上最明显。笔者统计了 6,350 条中文界面文案的关键词：据 6,350 条中文界面文案统计。“目标”已剔除“目标位置”等技术义项；统计反映界面用词，不代表使用频率开发者的概念都还在但只是被埋得越来越深了。MCP 一共只出现 6 次，全在第三方提交连接器送审的表单里，插件连一次都没有出现。需要的工具还不存在时官网的说法是“智能体会为你创建该工具”。安卓安装包里已经写好了这句话的样子。用户说一句“我想创建一个关于…的空间”，Muse 就去搭一个页面，搭好以后放进 App 的标签栏，引导语就写的是“打造你的专属应用”。今年 3 月 OpenClaw 最火的时候，笔者有过一个判断：skill、MCP 是开发者的概念，普通用户没必要懂。可 OpenClaw 和国内好几家 agent 产品，还是建起了技能商店。其实也可以理解，人们总会迫不及待的把熟悉的概念往新事物上套用，因为传统互联网的「商店」「生态」概念已经被这些大厂的产品经理玩的烂熟于心，但新的技术需要的是全新的思考方式，而不是一个粗暴的套用。笔者斗胆猜测，接下来会有很多模仿 muse 的产品出现，一定有人在里面还继续保留「商店」的概念，不管是 skill 商店还是 MCP 商店还是 Agent 商店。船大了难掉头，人的思维也是一样的，我们习惯在熟悉的东西里面找安全感，这也无可厚非，高手对局都是你下你的，我下的我的，反正市场会用脚投票。App Store 的前提是用户自己当组装者，agent 时代组装交给了 agent，再把技能商店摆到用户面前，“人学机器”就从后门溜了回来，好比餐馆在点菜前先请客人挑锅铲。Muse 面向普通用户的那一层连一个 skill 都没有，甚至用了半天我们才发现原来根本不用选择模型，或者类似「快速」「专家」这样的档位。不过 skill 并没有消失。安卓安装包里，管连接器的那部分代码就叫 skills，足足 647 个文件，到了界面上一律改叫“连接器”，“技能”“插件”“MCP”一次都没出现。锅铲一把没少，全收进了后厨。同样被收进后厨的还有操作。比如同样是让 agent 用手机发短信，OpenClaw 的文档要求管理员先把 sms.send 写进网关配置的白名单，Muse 安卓版只会在手机上弹一句“Muse 想用你的手机发送一条短信”，点一下就放行，也可以设成始终允许。再回到那场出生仪式。Muse 的默认形象 Jolly 是个奶油色、带腮红的毛绒角色，用户可以给它改名、换装、挑声音。Muse 让用户给 agent 起名、换装，配置过程被设计成一次角色定制在逆向中我们观察到选声音的流程用 Muse 的第一人称写成：“我们来找找我的声音吧！”一个配置动作，写成了角色在找自己。刚开始用 Muse 的时候，我被它引导着设置头像。它不仅做好了几个可以选择的风格，而且会默默在后台帮你把头像做成一个会动的动漫卡通形象，这个体验非常超预期。一次直出的效果已经非常非常惊艳了。大多数 agent 靠转圈动画和滚动的思考过程表示“正在处理”。Muse 把它分成三层：普通用户看表情，想多知道的看任务清单，工程师翻原始日志。最上面一层交给 Jolly，执行任务时它抱起小笔记本，生成媒体时捧起一颗光球。我们在逆向工程中观察到界面文案里还有“内部设置”和标着“Meta 专用”的声音选项，管语音识别后端、通话工具模型这类工程参数，还有调试工具。OpenClaw 认识用户靠一份 USER.md 模板，要么手填，要么在对话里慢慢攒。Muse 也留着这份 USER.md，安卓版的目标页就从里面读用户资料。我们在逆向工程中观察到 Muse 给搬家的用户备好了一段发给旧助手的提示词，请它“整理一份关于我的便携式记忆快照”，每条一句有依据的事实，拿不准的标“推断”。新助手替用户给旧助手写了一封交接信。Muse 替用户写好的搬家提示词，连该省略什么都交代了比认识用户更难的其实是分寸。同一段提示词还交代健康状况、财务细节、心理健康这类信息除非用户明确要求，否则不收。Muse 读入旧记忆时，也被要求“跳过任何短暂或敏感的内容”。分寸也写进了它的手机工具里面，安卓版拨号工具的说明写得很清楚：替用户开口、跟对方说话的那个通话工具只给商家用。打给家人和朋友的话 Muse 只负责拨号，开口的还是用户自己。发短信也是这样，群发或者带附件的，它只起草，最后那一下由用户自己按。Sensor Tower 9 月的数据里，Muse 次日留存接近 ChatGPT 的两倍，但是掉得也快，第 8 到 9 天就被反超。也就是说也许第一次被接管的体验很棒，但是人们的事情是零碎的，要想一直服务用户还是得看怎么运营 AI 自身的能力。浏览动态的时候，在各种我感兴趣的资讯里，呼和浩特出行的相关信息时不时就会更新出来，这就让我非常自然地刷到这些信息，而不是被动地接受一条无聊的通知。Muse 还会把这些信息直接做成可视化的卡片，虽然不怎么好看，但确实已经非常够用了。资讯和通知在这里巧妙地合并了，用起来非常舒服。动态页的节奏也掌握得很好，并不是你主动刷新，它就会给你一大堆没看过的新内容，显然这里的目的不是为了留存。出行信息混在感兴趣的资讯里自然出现，资讯和通知合成了一件事该打扰我的时候，它也不含糊。Muse 从我的邮箱里找到有必要通知我的信息，直接通知了我，甚至让我很诧异它怎么能通知得这么准。workbuddy 目前的布局是围绕办公场景来搭建国内日活最高的办公 agent 是腾讯 WorkBuddy（花旗研报估算超过 1,300 万）。笔者拆过它的安装包，内核文件以腾讯的编程 agent CodeBuddy 命名。Muse 的安卓安装包里，程序认的也还是 OpenClaw 那几个人格文件。两家都是从开发者工具长出来的。不同的是 WorkBuddy 主打你的职场超能力，把能力分成 Skill、专家、专家团三级，用户得当那个给项目配人的主管。在工位上这说得通，但换到家里很少有人愿意为了订一张票先当一回主管。两家走得最近的地方是关于怎么处理支付的问题，也就是说 Agent 迟早要帮助人消费或支付，所以这件事应该从一开始就考虑到，我们来看看两家的思路和行动。WorkBuddy 内置了微信支付 AI 专属卡，挂在用户的微信支付账号下，和零钱分开，支付指令交给确定性的程序处理，原始指令不给模型看。这套护栏在笔者横测过的三家办公 agent 里写得最认真。Muse 安卓安装包里也有一张 Link 虚拟卡，智能体并不会看到你的信用卡。碰到钱的时候，两家的都是选择单开一个账户。但要是真的去点一杯咖啡其实还难以做到。在早一点的时间我们测过 workbuddy，笔者直接问 WorkBuddy 能不能帮忙点咖啡，它想了很久，扣掉 14 个积分后回答做不了。笔者自己连上“美团生活助手”、绑好账号，又来回几轮，到手的是一张到店团购券，糖度和冰量这些日常非常方便选择的东西全靠打字来说。该连哪个服务、绑哪个账号，都得用户自己先知道，像走一遍采购流程。钱之外，两家的规矩就分开了。WorkBuddy 的规矩落在文件和钱包上。它跑在用户自己的电脑上，沙箱允许读整块硬盘，写只限临时目录，真实文件要逐个授权。Muse 多出来的规矩落在人身上。前文写过，对商家它能替用户开口，对家人朋友只拨号。WorkBuddy 照着一张组织架构图搭，Muse 照着一张亲疏远近的关系图搭。放回国内看，组织架构那张图早就是公共设施了，企业微信、钉钉、飞书谁都能接进去。在千问办公的自研代码里，企业微信出现的次数比阿里自家钉钉多出近一倍。据报道，能替用户在微信里给人发消息的“小微”6 月开始灰度，到 10 月中旬还没全量，蚂蚁“灵光”新版和字节“小豆”也都还没正式上线。笔者推断，国内办公 agent 已经上了三家、个人 agent 马上也会出。Muse 的成功精髓在于把 Agent 背后那些复杂的概念简化成了大众都可上手的地步，随着广大的用户低成本上手，越用越深入，往后个人 agent 的几个方向也许能管中窥豹一番：
- 第一，新的应用商店会长成“点子”的样子。货架上摆的会从一项项技能变成一件件替用户想好的事。谁能让点子页变成商家抢着上的位置，谁就握住了 agent 时代的分发入口。
- 第二，多 agent 协作会藏到一个联系人背后。用户要的是一个固定的说话对象并不是一堆助理。安卓安装包里已经有子智能体了，它们只在状态栏露一行字，比如“3 个子智能体工作中”。身体也一样，手机、眼镜、家里的小设备都在陆续接进来，跟用户说话的始终是同一个名字。
- 第三，个人 agent 会从我的走向我们的。界面文案显示，Meta 做过一套共享智能体，从用户的 Muse 里分出一个有边界的版本给别人用，如今新建入口已经停用。10 月 20 日，Muse 的分享链接也被暂停了，起因是安全公司 Claroc 发现，6,800 多个分享页能从公开的 DNS 记录里查到（据 Ars Technica）。门开过又关上了，关上的还不止一扇，可安卓安装包里又备好了新的，用户可以邀请别人查看自己的“空间”。需求是客观存在的，毕竟我们家里的事很少只关乎一个人。
- 第四，长期留存靠的是高频的日常小事。省下 1,000 美元的宣传比较偏向于运营层面，能换来一次转发，如果第九十天还愿意打开它，靠的肯定是无数件生活小事的体验提升。
对于一些自己跑了几个月 Hermes 或者 openclaw、opencode 的开发者说，Muse 做的事他一样都不新鲜，可他还是非常成功，这再次证明了新技术需要被翻译成大众产品，中间隔着非常大的鸿沟，而极少数人能率先跨过去。身处在这样一个技术巨大进步，大部分产品体验又稍显割裂的时代，meta 的确像一道光刺进来，照亮了一方天地。这让我们不免想到经典的看山三重境界：
- 看山是山：Muse 就是 OpenClaw，共享同一份 SOUL.md 和同一场出生仪式。
- 看山不是山：一个要求用户先学会它，一个先学会了用户。
- 看山还是山：Muse 依然是 OpenClaw，只是这一次，Meta 在山脚下修了一条普通人也能走上去的路。
那份 SOUL.md 就在 Muse 的身份页里，点开就能看、能改，但大多数用户这辈子都不会去点它。Meta 做对的或许就是这件事。
可独家畅览：
专家演讲PPT
大会报告全文
热门论文解读
学术新星访谈
扫描上方二维码
或点击「阅读原文」关注专区。
未经「AI科技评论」授权，严禁以任何方式在网页、论坛、社区进行转载！
公众号转载请先在「AI科技评论」后台留言取得授权，转载时需标注来源并插入本公众号名片。
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Muse Dismantling: The same Claw core, how to deal with people in a different way.
By Rhea
Editor of Lina Tsing Feng
Around 20 September, a post was transmitted in large numbers on X. The AI application co-founder, Ansh Nanda, asked Meta about the new individual on line, angent Muse, why its system file is almost identical to OpenClaw. He evaluates Muse is Literally OpenClaw for images for the previous sentence. Muse is OpenClaw for ordinary people. “Muse is Literally OpenClaw for recentes” widely distributed on X. Friedman responded the following day. Ray Peaks (WeChat Official Account:Ray Peaks Network) developers then found SOUL.md, MEMORY.md in the Muse workspace, also known as OpenClaw. The next day, Nat Friedman, the head of the Meta super smart lab product, responded: Muse's code was built from zero, and the product was inspired by OpenClaw. As to why these documents should be followed, he explained that OpenClaw author Peter Steinberger had "made these things right." Sarah Perez's likeness to OpenClaw isn't a coincidence, and after experimenting with Muse, it turns out that the biggest innovation and breakthrough of the APP comes from something that the technology community can't see, and developers can often build tools themselves, but they ignore one fact: Most people actually don't learn a set of tools for one ability. Muse and OpenClaw use the same personal anent paradigm, and there are no intergenerational differences in abilities. The change is mainly in the way people live with angent: OpenClaw needs to learn about machines, Muse makes robots learn about people. There are two levels of meaning to learn. One is to learn about users who know this person from the memory of existing social accounts and old assistants. The other is learning how to live with people, changing the previous terminals, configuration documents and logs with an ongoing dialogue, a visible role, and a timely response. These things used to be learned by the users themselves. This paper combines the author's findings with a static analysis of the public resources available on the muse.ai web page and the installation package in Muse Andre to see how Muse moved the learning costs. Muse went online in the U.S. on September 8, 2026, and was able to work in App, web page, or whatsApp, and then went backstage. Open the kernel to see that Muse and OpenClaw are the same set of structures: angent operating frames, web-accessible browsers, timed tasks, memory from Markdown, and a personality file that defines who it is. The Android version also uses OpenClaw, which is called "node" in the code, which is ordered by the cloud and executed by the cell phone. So these parts are already there, and it's just that Meta has an almost genius product manager doing them all right. The product experience is good enough to attract more people. The story you're familiar with is two weeks on the line, and Muse is in the United States. Free list number one. We read the pages, scripts and interface translation tables publicly distributed at the muse.ai page end without login status, without calling the AMA interface and without running the client. Of these, 6,350 English-language interface files have been translated, i.e., Muse has not been online in the Chinese region, but the Chinese interface is almost ready. It's like a product dictionary, and it shows how it's going to talk to users. The author later obtained the Andre installation kit (12.0.0.14.169) of Muse, which was also analysed staticly and was not installed or operated. The Chinese language in the bag is almost 90% translated, but it's just not available, and it's interesting how many places anent is translated into "arrival." Right is the beginning of the file sheet, with one line of interface case. On the left, 89 percent of this inheritance, which is counted from this 6,350 miles, extends to the first meeting ceremony. OpenClaw's BOOTSTRAP.md let angent wake up in a conversation and talk to the user about names, aromatics and signatures emoji, and then write in IDENTITY.md. This is also the case with the Muse identity files, the project codenamed Hatch, intended to hatch. Muse followed the same process for the core personality files of OpenClaw workspace, and OpenClaw was done in the terminal, and Muse moved it into the role custom interface. Meta left the skeletons that were found in open-source communities, but the experience outside the skeletons was almost entirely rewritten. Let us sum up a little, so that the rewritten part can be broken down into five living costs removed from users. OpenClaw's first threshold is a computer that you have to prepare, configure and maintain. Zuckerberg's judgment is that there won't be billions of people buying a Mac Studio home for themselves. Muse gives one computer to each user but does not need the user to transport it himself (God knows how long it takes for an openclaw to travel!). All you have to do is log in with the Meta account number one, and Facebook, Instagram and Threads under the same account center are automatically connected. The cloud-side computer behind it does not appear at all times, except for a browser card, or a button that the user is asked to take over. The only thing that needs to be done by users is to open the muse.ai web page, or to download the muse APP. (Of course, what domestic users need to do is much more) cloud-end computers, which are not visible at all times, and the broader the ability to appear in the form of browser cards or requests for take-over, the more difficult it is to fill in the blank input box. Almost all Agent products are facing this problem: what should the first thing be done by angent? Muse's way is to speak first. The “point” page lists what it can do for its users, and points out “Let's do it”; “dynamic” is its short posts and podcasts at its own pace; and “targets” is left for the long term, before it is planned and regularly reported. According to the information available, the cold starter experience of offering a dynamic page of events that can be done on behalf of users has also been good. For the first time a user sees a dynamic page and first sees a reminder of what it is and what it is worth. When no user information was available, it was viewed with functional descriptions and state tips, but it did not simply leave a message, but it was cleverly presented in the form of future posts. The dynamic page also supports changes That means that users can see how this dynamic page works, and you can even modify how it works. The first thing about changing Meta arrangements for users is money. Chief AI Alexander Wang launched #MuseMoneyCallenge on X to save $1,000 for the user. Users don't have to figure out what angent is, just understand the bill. Meta designed the first task to save money for the user, which we found in the static reverse of the web page, but the user hasn't said anything yet, and Muse has set himself a default target when creating the virtual machine. The blank page also talks like a friend: "As we know each other, it will come to pass." How does this work if the user is talking to someone else? Agent was placed at the top of the screen, like a contact in the address book. There's only one main conversation, just like WeChat with one person. Make a request. Muse acknowledges receipt and enters backstage for execution, and the trade-off is most obvious when it comes to communicating to friends. The author counted 6,350 Chinese interface words: 6,350 Chinese interface script statistics. “Target” has been removed from technical terms such as “target location”; statistically, reflecting the language of the interface, does not mean that the concept of a frequency developer is still in place but is just getting deeper and deeper. MCP appeared only six times, and none of the plugins appeared once in the forms submitted by third parties for review. The term “AI agent will create this tool for you” is the term of the Timenet. It's written in Andre's installation bag. And the user said, "I want to create a space about...," and Muse went to set up a page and put it on the App label, and the guide was, "Make your exclusive application." In March of this year, when OpenClaw was the hottest, the writer had a judgment: skill, MCP was the concept of the developer, and ordinary users didn't have to understand it. But OpenClaw and several of the country's angent products built a skill shop. Indeed, one can understand that one can’t wait to apply familiar concepts to new things, because the traditional Internet “stores” concept of “ecology” has become so familiar with the product managers of these large plants, but new technologies require a whole new way of thinking, not a crude one. The writer ventured to speculate that there would be a lot of mimuse- imitating products coming up, and that there would be people inside who would continue to maintain the concept of "store" whether it was skill or MCP or Agent. There's nothing wrong with us being used to feeling safe in things we're familiar with. You're the one with the best hands, mys, the market is the one with the feet. AppStore's premise is that the user is a assembler and the agent-time assembly is handed over to angent, then the skill store is put in front of the user, and the “human learning machine” slips back through the back door, as if the restaurant had asked its guests to pick a pot shovel before ordering. Muse, the floor for ordinary users doesn't even have a skill. It took us half a day. It was found that there was no choice of a model or a "quick" "expert" slot. But Skill didn't disappear. The part of Android's installation package, the code for the connector, is called skills, with a total of 647 files, which is always called "connector" on the interface, and "skilled" "plug-in" "MCP" has not appeared at all. All the shovels are in the back. It's the same thing as cooking. For example, it's the same thing that makes angent texting on his cell phone, OpenClaw's file requires the administrator to write the sms.send first into a white list configured for the gateway, and the Muse Andre only plays a "Muse wants to send a text message on your cell phone" on the phone, so it can be done by clicking and always allowed. Back to the birth ceremony. Muse's default image, Jolly, is a cream-coloured, gilded furful role that users can change, change and pick out. Muse lets the user name and change to angent, and the configuration is designed as a role custom for the process by which we see the sound in reverse, written by the first person of Muse, "Let's find my voice!" A configuration action, and a character is looking for itself. When I started using Muse, I was guided by it to set my head. Not only did it do a few of the options, but it did it on the backstage to help you make a moving cartoon image of your head, an experience that was so much more than expected. It's been really, really amazing. Most of the angents express themselves as being “worked out” by turning animated and rolling thinking processes. Muse divides it into three layers: ordinary users look at their faces, want to know more about the job list, engineers go through the original log. The top floor was given to Jolly, who carried out his mission with a small notebook and raised a light ball while generating the media. In reverse engineering, we observed that there were also “internal settings” and voice options marked “Meta” in the interface case, engineering parameters such as voice recognition backends, telephone tool models and debugging tools. OpenClaw recognises users by using a USER.md template, either hand-filling or slowly saving in conversations. Muse also keeps this USER.md, where the Android version of the target page reads user information. We observed in reverse engineering that Muse had prepared a paragraph of prompt for the mover's assistant, asking it to "assemble a portable memory snapshot of me" and to take an incorrect "inference" for each sentence. The new assistant wrote a letter to the old assistant for the user. Muse's mover prompt, which was written for the user, is a little bit more difficult than knowing the user. The same paragraph prompt also provides information on health status, financial details and mental health, which is not received unless explicitly requested by the user. When reading old memories, Muse is also asked to “jump anything short or sensitive”. It's also written in its cell phone tools, and the Anzor dial-up tool is clearly written: the phone that opens the user's mouth and talks to each other is used only by businesses. If you call a family or a friend, Muse will just dial, and the user will be the one who's talking. The same is true of text messages, which are sent in groups or with attachments, which are drafted only, and the last one is pressed by the user itself. In September, Sensor Tower's data, Muse kept nearly twice as much as ChatGPT the next day, but fell fast, and the 8th to 9th day was turned upside down. Which means maybe it's a good experience to be taken over for the first time. Users still have to see how to operate AI's own capabilities. When browsing through the dynamics, and in the information that I'm interested in, the relevant information about the travels of Houthi and Houthi are updated from time to time, it makes me very natural to brush that information instead of passively accepting a boring notice. Muse will also make this information a visualized card, which is not very good-looking, but which is really enough. Information and circulars have been ably combined here and used very well. The rhythm of the dynamic pages is well understood, and it's not your initiative to refresh it. It gives you a whole bunch of new things you haven't seen, and it's clearly not intended to be retained. Traveling information naturally appears in information of interest, and it is not ambiguous when information and notification combine something to disturb me. Muse found the information in my mailbox that I needed to be notified, and he even informed me, and I was surprised how it could be so informed. The current layout of workbook is the construction of one of the country's highest daytime offices around office scenes. The writer removed its installation package and the kernel file was named after Tencent program agent CodeBuddy. Muse's Andre's installation bag, the program still recognizes OpenClaw's character files. They both grew out of the developers' tools. The difference is that WorkBuddy, who beats you in the workplace, divides your abilities into three levels: Skill, expert, expert team, and the user is the director of the project partner. This makes sense in the workplace, but very few people in the family are willing to be supervisors for a ticket. The closest place for the two families is the question of how to deal with payments, which means that Agent has to help people consume or pay sooner or later, so this should be taken into account from the very beginning, and let's look at the thinking and actions of the two families. WorkBuddy has a WeChat payment AI exclusive card attached to the user ' s WeChat payment account, separate from the change, and the payment instruction is assigned to the certainty program, and the original instruction is not shown to the model. It's written most carefully in three offices across the pen. There's a Link virtual card in the Muse Andre installation bag, and AI agent won't see your credit card. When it comes to money, both families choose to open a single account. But it's hard to get a cup of coffee. We checked the workbook earlier, and the writer asked WorkBuddy if he could help with the coffee, and it was a long time ago, and he couldn't answer with 14 points. The author attached himself to the American Life Assistant, tied his account, and then rounded it back and forth, and arrived with a ticket to the shop, which is all typed in terms of sugar and ice, which are very easy everyday choices. It is up to the user to know which service and which account number to tie, as if through the procurement process. Besides the money, the rules of the two families are separated. WorkBuddy's rules lie on documents and wallets. It runs on the user's own computer, and the sandbox allows the entire hard drive to be read, with only a temporary catalogue, and the real document is authorized on a case-by-case basis. Muse more rules fall on people. As previously written, it can speak for users and dial only family and friends. WorkBuddy follows an organizational chart, and Muse follows an intimate map. Turning back to the country, the organizational chart was already a public facility, and the company WeChat, the nails, the flying book could be picked up by anyone. In the self-study code of the office, the company WeChat appears more frequently than Ali. Home nails almost double. It was reported that the “minus” that could inform the user in WeChat began in the middle of the month of October, and that neither the new version of the ant “Line” nor the byte “Bean” had been officially online. It's assumed that the national office is already in three, individual parties will be out soon. The essence of Muse's success is to simplify the complex concepts behind Agent to a point where the public can do it, and the more it goes, the deeper it gets, the more it gets, the more it gets.
- First, the new app store will look like an idea. The shelf moves from one skill to one that is good for the user. Whoever can turn the idea page into a shoplifter, holds the distribution portal of the age of angent.
- Second, more agent collaboration would hide behind a contact. The user wants a fixed speaking object, not a bunch of assistants. There are already sub-AI agents in Andre's installation kits, which show only one line in the status bar, such as "three sub-AI agent jobs". The same is true of the body, where mobile phones, glasses, and small equipment in the home are being picked up and the same name is always spoken to the user.
- Third, a person would go from me to us. The interface file shows that Meta has made a shared AI agent, a bordered version from the user's Muse, which is now disabled. On October 20th, Muse's shared link was also suspended because the security company Claroc found that more than 6,800 shared pages could be found in public DNS records (according to Ars Technica). The door was opened and closed, and more than one door was closed, and a new one was prepared in the Butler installation bag, where users could invite others to look at their “space”. Demand exists objectively, and, after all, very few of our family matters concern one person alone.
- Fourth, long-term retention is dependent on routine routines of high frequency. Saved by $1,000, the publicity is more in favour of the operational level, allowing for a forwarding, and if it is still open on the ninetieth day, it must be due to the increased experience of countless small things.
Hermes, or an openclaw, or an opencode developer, say that he's not as new as Muse, but he's very successful, which is yet another proof that new technologies need to be translated into mass products, with a huge gap between which a very small number of people can take the lead. In an era of great technological advances, when most product experiences are slightly fragmented, Meta does come in like a light, lighting up one side. This reminds us of the classic Three Mountain Frontiers:
- The mountains are mountains: Muse is OpenClaw, sharing the same SOUL.md and the same birth ceremony.
- The mountain is not a mountain: one asks the user to learn it first, one learns it first.
- Look at the mountains or the mountains: Muse is still OpenClaw, except this time, Meta built a road that ordinary people can walk under.
The SOUL.md is on the Muse identity page, and it can be seen and changed, but most users will never order it in their lifetime. Meta did the right thing and maybe that's it.
It's an exclusive view:
Expert presentation PPT
General Assembly
The reading of popular papers
Academic New Star Interview
Scan 2D above
This post is part of our special coverage Syria Protests 2011.
It is forbidden to reproduce it in any way on web pages, forums, and communities without authorization from the "AI Science and Technology Review"!
Reproduction of WeChat Official Account should be authorized by a message from the back of the "AI Technology Review" to mark the source and insert this WeChat Official Account card.
It's an original Ray Peak article, which is forbidden to reproduce without authorization. For further information, see the information below.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.