文 | 光锥智能，作者｜魏琳华，编辑｜刘俊宏
Meta的Muse在海外火了一个多月，国内大厂“姗姗来迟”。
两个月的时间里，国内市场上也开始出现打着“个人智能体”旗号的产品。目前，市面上用户可以直接上手体验到的代表产品有两个：前豆包PC端负责人齐俊元带队打造的Today AI，以及蚂蚁集团升级至2.0版本的“灵光”。
那么，这两个被寄予厚望的产品，现在到底做得怎么样？它们能复刻海外Muse给用户的惊艳吗？
光锥智能实测下来，前者是“个人助理”，后者是“生活管家”：
Today AI，在产品形态上更像Muse的复刻版，它和Muse都靠先获取大量个人信息给用户“画像”，偶尔主动推送你的爱好和工作相关新闻，交流下来更有“活人感”。
它会根据你抛过去的多项杂乱任务梳理出行程，并主动提醒不同任务之间的时间冲撞问题；通过大量获取个人信息，它不需要你主动交代自己的需求，就能通过本地文件识别出你可能的工作需求、个人爱好，不定时分享相关信息。
蚂蚁灵光不太像Muse，它更像是中国版本的生活管家，对用户的信息获取更克制，这让它的上限不高。
背靠蚂蚁生态，灵光拿到了支付宝的调用权限，它最核心的落地动作是根据扣费记录分析消费习惯，这让我看到了支付宝AI助手“阿宝”的影子。其中一些预设的任务更像是在引导用户去体验它的“老本行”——比如一些互动上，它会引导用户自己动手“搓”一个闪应用来记录。
同为个人智能体，不同的定位，让它们真正开始替用户办事时，表现大相径庭。
在体验上，Today AI通过日程安排、记忆，串联起了一张完整的“任务网”，所以它能够给到更多惊喜，比如提醒你日程冲突、比如它能在更多任务中代入你的画像，直接做决定；灵光更像个克制的管家，你交给它的每个任务完成不错，也会每天主动跟进度，但它少了些连贯的体验，不能记住你是谁，所以接下来更多任务里还需要用户多问一轮、多想一步。
不过更多时候，它们已经有了惊艳的表现，却被其它的App们联手阻拦。
比如想让它们给我双十一要买的产品三方比个价，它们没有权限上淘宝浏览；想总结下我今年的旅游计划，它们看不到12306的出行信息。
在“应用墙”内长出来的Today AI和蚂蚁灵光，是它们基于各自基因交出的“限制版”答卷，但这份答卷也有了许多惊喜。
初体验，两个管家，两种人设
从初次打开应用开始，两款产品的差异就开始出现：
首次打开Today AI，它会先引导我定好自己的名字，再给它起昵称和个人称呼。这种个性化的设计，让它显得更有“人味儿”。我给它选了个DeepSeek最近很火的二次元形象，起名“大肥鱼”。
Today AI的产品逻辑非常直接，它要“先懂你，再帮你”。为了跑通这个逻辑，它会开始使用前先向用户要权限，比如读取手机上的日历、短信等权限、做个自我介绍等方式。
抱着对它能做到什么地步的好奇，我立马开放了一大堆手机底层权限，它开始翻你的本地文件、相册、应用使用时间等信息，给我做了一个个人画像。
最终，它给了我一份关于我的“了解”，洋洋洒洒得有几千字。它一边把姓名、地址、App使用时长等个人信息记下来，一边根据这些内容推测我是个什么样的人：
它从我的使用时长上判断出我玩手机大半时间花在追剧上，看到我存了一堆AI产品的稿件、采访提纲，直接推断出我的工作跟AI行业高度相关，它甚至根据我凌乱的短信“吐槽”我不怎么整理信息，但又从我存的文件里觉得我做事有清单感（不觉得两者太矛盾了吗？）
除了这篇“长作文”。Today AI还专门设立了一个“记忆”功能页，他们自己将它命名为Living Memory（鲜活记忆），在对话中持续理解你反复提及的日程、人物、偏好和目标，并把这些分门别类整理下来，支持用户随时修改。比如它之前错误地理解我是一个爱规划的人，我立马删除了这一条。
在记忆的基础上，它还做了主动性的设计，即根据它对你的了解，“猜”你需要什么，主动给你推送。
基于这些了解，接下来使用的几天里，我不断会收到Today AI主动发来的一些信息。它认为我喜欢追剧，就会安利新上线的作品；它知道我处于AI行业，就会告诉我今天Anthropic的新模型曝光了，还会推荐我去北京线下的AI论坛活动学习。
可以说，Today AI更像个不客气的助手，它能翻进你的手机里把你看个底掉，然后像助理一样推测老板的喜好来做事。
相比之下，灵光的个人智能体做得更加克制，它升级的2.0版本在于增加了“帮我管”功能。你可以按照自己需求给它下任务，它也预设了一些灵感来启发你怎么使用。（但是一次只给三个，无法在App内刷新，只能下一次打开应用后再给新灵感）
体验下来，灵光在信息获取上更加克制，且考虑到用户对隐私的担忧，把信息安全的保障放在明面上。
在“帮我管”的右上角，灵光也设置了权限获取，分别是应用服务和系统服务。但系统方面的权限相对有限，它没有索要我的设备使用时长、相册等权限。关于信息安全，它搬出了蚂蚁集团此前发布的APASS安全连接，它相当于确保智能体安全的基础设施，保护隐私的同时，管控智能体的权限。
说起来，Today AI的权限获取，也让一小部分用户感到担心。我自己被它抓取到姓名、住址等信息的时候，也忍不住吓一跳。
聚焦到使用场景上，怎么让用户“用起来”是个问题。所以个人Agent都会做场景预设，相当于教用户可以怎么“一键体验”对应功能。
比较下来，Today AI在场景预设更贴合个人用户的一些需求，比如久坐人群的运动跟练提醒、未完成事项跟进；灵光给的一些建议有点“悬浮”，比如让AI教我剪纸、讲讲传统手艺的历史之类的。它不在我的兴趣范围之内，与真实生活场景实在算不上贴切。
此外，在体验一些建议时，我从灵光身上看到了支付宝AI助手“阿宝”和之前的“闪应用”的影子。
比如它引导我体验旅行足迹的总结，看到这个标签时，我以为它能读取我的行程信息，自动做一份类似手帐的记录给我；结果点进去交互了几轮，它最后却是引导我用它的无代码工具“手搓”一个闪应用来自己记录。
到了偏生活管家的模块，它能落地的场景是“根据支付宝每月的自动扣费记录，分析我的消费习惯是否合理”——这几乎就是把“阿宝”的能力直接搬了过来。
不过，背靠支付宝，它的好处是能直接读到我的各项支付记录。靠支付归纳出我的消费习惯、旅游动线等等。可以说支付给了它了解用户的更多切口。
整理完初印象，Today AI的主动性很惊艳，灵光“帮我管”的功能设计稍显呆板，不过靠着生态，它在支付方面的信息获取能力更强。
实测：从省钱到工作和玩乐 它们能帮我排忧解难吗？
光看印象还不够，我们把两款产品拉进真实的日常生活中做了一系列测试，想看看它们能不能复刻Muse的惊艳效果，再一路切到最接地气的衣食住行。
先从海外用Muse最多的场景——省钱测起。
扎克伯格在Meta Q3财报电话会上曾提到过一个数据，海外有1/4的用户在用Muse辅助购物；在X（原Twitter）上，网友反馈最多的用法也是用它省钱，比如跨平台货比三家找最低价。
Muse给我印象最深的，是它能帮用户自动敲定很多事情，比如帮用户“薅羊毛”，买的服务一样，但智能体自己找到了更省钱的方式。
我们给了Today AI和灵光一个更本土的省钱场景——要求它们“能不能帮我找到交话费更便宜的方式？”
这一轮，Today AI展现出了更老练的“绕墙”能力。蚂蚁灵光建议我去下载并打开运营商的官方App查看套餐订阅情况，告诉我移动App可以每周三抢活动，不过根据花费一月一交的情况，它选择了每月月末查询一次。
Today AI除了给到上述查询方式外，还给了我一种更省心、且Agent可以操作的方式——直接往运营商发短信“0000”，就能收到本机所有的增值业务扣费明细。
在关注优惠的任务上，它设置的频率是每天查询优惠活动，并且真的找到了我在App首页都没翻到的“秋日活动”，我打开它给我的链接，领到了8.6折充话费的优惠券。
充话费上，Today AI实打实帮我省了一笔小钱。
为了进一步省钱，我又让他俩“帮我给移动发短信看看有什么订阅是能帮我取消的吗？”
在发短信、读短信的执行表现上，涉及到用户权限，两者表现出了不同程度的谨慎：蚂蚁灵光在对话框里需要我多确认一轮才肯帮忙发短信；Today AI直接响应发短信的请求，它帮我填好收信人和发的内容，但必须要用户手动点发送。
最终执行效果上：Today AI抓取到了短信内容和时间，立刻梳理出了退订方案；但蚂蚁灵光在本地短信箱里连续读取了3次才终于找到回复。
除了让Agent帮忙找更省钱的方式，我也试着让两个Agent给我想买的索尼耳机做全网比价。
遗憾的是，面对国内淘宝、京东、拼多多筑起的流量高墙，两家产品都无法直接调用电商App查看实时到手价，只能退而求其次去搜网页。
在网页搜索的执行上，蚂蚁灵光显得中规中矩，只搜索了京东、天猫和拼多多三家；Today AI的搜索范围则广得多，除了主流三家，还主动加上了苏宁易购、唯品会和索尼官方商城。但尴尬的是，两者抓取的由于都是帖子的过时信息，跟我看到的平台价格还有些许差距。
除了省钱，作为“个人助理”，我还测试了Today AI和蚂蚁灵光帮我打工的能力。比如，我让两个AI每天早上检查我的邮箱和信息，按照紧急度排序。
（提示词：请每天检查我的邮件和消息，找出仍需要我回复、确认或补充材料的对话，按紧急程度排序，并说明对方在等什么；不要直接发送回复。）
Today AI没有识别到我任务中的“定时每天执行”的要求，它只执行了一次，第二天没有继续，还需要我提醒才能设置成定时任务。灵光倒是直接整理成了定时任务。在任务完成度上，两者跑得都不错。
我又交代它们根据我本周开会的记录，总结出来关于我本周的工作安排。
（提示词：根据本周的会议记录，整理一下我本周待办事项，有什么需要注意的，然后有什么需要开的会，整理成日程提醒我）
Today AI可能只是简单读了一遍速记，所以没有抓取到我的相关信息；但灵光在思考记录中提到，它查找了文字记录中关于我名字相关的发言等内容，最终成功抓取到了我的任务，虽然还是漏掉了一个任务，但它根据抓到的内容，也给了我一份时间表，不过时间比我想得局促很多。
区别以往Agent，个人Agent的优势在于“主动性”。
考虑到让AI主导的要求，我给两个AI抛出了一个要求，“帮我做一份近期去九寨沟的出行规划，但工作很忙，时间还没想好。”要求AI来帮忙看着定时间，提醒我需要做什么。
同一个旅行需求，两个AI表现出了不同的优势：Today AI在拿到需求后，考虑到我们提出的出发日期太近、热门线路不好抢票，主动在方案里给出了一个稍晚几天的出行时间建议；而蚂蚁灵光在生成路线之余，考虑到九寨沟游玩极度看重当地天气，主动提出可以帮我实时监控九寨沟的天气变化，找晴天更多的时段来做推荐。
平心而论，两个答案都不算完整的闭环方案，但各自多想的一步，是自身能力边界的暴露。
除了主动性之外，为了彻底摸清两款产品到底能“够得着”多少个人数据，以及用“记忆”发挥出多少效果，我们做了最后一组测试：让它们“帮我回忆今年都去过哪，整理一份年度出行总结报告”。
两款产品交上来的答案，展示了各自的能力边界：
蚂蚁灵光的路径是“查账单”。收到指令后，它第一时间调用了支付宝，查到了我最近一次去外地的消费记录，并写进了报告里。尴尬的是，明明是一次出差，却被它误判成了休闲出游。
Today AI的路径则是“翻箱倒柜”。没有支付宝接口的它，一头扎进了我的手机本地：一边在短信箱里检索“12306”的出票信息，一边去翻手机相册里的图片定位。这种做法的好处是，它比灵光更清楚我具体去了哪、坐的哪趟车、去做了什么；但坏处是，手机里的短信记录默认只保留90天，而12306的官方接口它又接不进去。
总结下来，我个人的感觉是：
Today AI在一些需要定时的任务上常常无法识别到，但它在信息检索上相对全面，给出的解决方案更让人惊喜；灵光则在一些时候抓取信息上会出现失败、错误等情况，执行时而惊艳、时而掉链子。
以及，在一些任务的执行能力上，两者的根源根本不在底层的基座模型聪明程度，而在谁能合法地多推开几扇门，以及在门内谁的主动性更强。
戴着镣铐跳舞 围栏不破，个人助理难当
走出具体的测试题，回看Today AI和蚂蚁灵光在实测中表现出的种种别扭与妥协，我们看到的其实是整个中国Personal Agent赛道的集体处境。
为什么Meta的Muse在美国能跑通？Muse在海外之所以能顺畅运转，靠的是电子邮件、系统日历的标准化开放接入，再加上Stripe、Shopify这类高度标准化的商业伙伴，只靠调取用户上下文，就能拼凑出更多信息。
Shopify甚至和Muse达成了官方合作协议，如此下来更是畅通无阻。
简单来说，在海外互联网生态里，一个人的机票订单、酒店行程、网购消费信息，几乎全都能通过一个邮箱串联起来。
但中国的互联网现实，是另一幅完全不同的图景。
就在不久前，字节跳动的“豆包手机助手”刚刚遭遇了一场轰动行业的集体抵制——微信、淘宝、支付宝、美团、拼多多、小红书等155款国民级App齐刷刷地向豆包手机说“不”，拒绝AI助手通过GUI（图形界面模拟点击）的方式替用户操作应用。
同样的高墙，也严严实实地挡在Today AI和蚂蚁灵光面前。
在中国，用户的订票、消费、支付、办公，高度散落在微信、淘宝、美团、携程、12306等各自为政的成熟App里。采用协议接入，意味着AI的能力上限完全取决于第三方App愿意把门开多大。
即使底层的Agent模型已经足够聪明，能够精准理解你的复杂需求、规划出完美的执行步骤，只要它无法顺畅地跨应用获取信息、调用服务，任务就注定会卡在应用与应用的围墙之间。
落回到这个现实，你就明白为什么无论是Today AI还是蚂蚁灵光，目前在设置里真正能深度绑定的本土国民级应用都少得可怜。灵光也只比Today AI，多一个支付宝。
只要超级App之间的围栏一天不破，中国的AI管家们，难免要在各自的院子里戴着镣铐跳舞。
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。
Wen, 光, 光, 光, |, |, 宏, 宏
Meta's Muse was on fire overseas for more than a month, and large domestic factories were “late”.
In the course of two months, the domestic market also began to see products under the label “Personal AI agent”. At present, there are two representative products that can be directly experienced by users on the market: Today AI, which was built by the former DoubaoPC end-user, Zi Joon, and the ants group, which was upgraded to version 2.0 "Linglight".
So, how are these two products of high hopes going? Can they recaptulate the shock of the Muse overseas?
The light cone intelligence is measured as “personal assistant” and “living housekeeper”:
Today AI, which is more like a replica of Muse in its product form, and Muse, communicates more “livingly” by obtaining a large amount of personal information for “images” to users and occasionally offering you information about hobby and work-related matters.
It combs the travels according to your many messy tasks of the past and proactively reminds different missions of the problem of time crashes; by obtaining a lot of personal information, it does not require you to give your own needs, it identifies your potential work needs, your personal preferences, and shares information from time to time.
It's not much like Muse, it's more like a Chinese version of a housekeeper, and it's more restrained from getting the user's information, which makes it less capped.
On the back of the ant ecology, the light gets the call permission for the treasure, and its core landings analyze consumption habits on the basis of deductions, which show me the shadow of paying for the assistant to Bao Po. Some of the pre-set tasks are more like directing users to experience their “old lines” — for example, in some interactions that lead users to do “swipe” a flash application to record.
The difference between the individual AI agent and the place where they are located is quite different when they actually start working for the user.
In experience, Today AI has been able to create a full “task net” by scheduling, remembering, so it can give you more surprises, for example, by reminding you of conflicting agendas, for example, that it is able to take pictures of you in more missions and make decisions directly; the light is more like a restraining butler, who does every job you entrust to it well and who does it on a daily basis, but it has less consistent experience and does not remember who you are, so it needs more users to ask and think about it.
But more often, they've done amazing things, but they're being blocked by other Apps.
For example, if you want them to give me a triple price for the products I'm going to buy, they don't have a permission on a treasure hunt; if you want to summarize my travel plans for this year, they can't see the 12306 travel information.
Today AI and antsl light that grew within the “applying wall” are the “restricted version” of their responses, which are based on their respective genes, but which also have many surprises.
First experience, two housekeepers, two sets.
The difference between the two products began to emerge from the initial opening of the application:
First time to open Today AI, it'll lead me to my name, then to nicknames and personal names. This personal design makes it more "human." I've chosen a recent hot double image of DeepSeek, called the Fatfish.
Today AI's product logic is very straightforward, and it wants to "know you before helping you." In order to follow this logic, it will start using permission before it is used, such as reading the calendar on a mobile phone, text messages, permission, a self-presentation, etc.
And with all my curiosity about how far it could go, I opened up a whole bunch of cell phones at the bottom of permission, and it started looking through your local files, albums, application time, etc., and made me a personal portrait.
Eventually, it gave me a "know" about me, spilling thousands of words. It writes down personal information about names, addresses, and the length of time that App uses, and then speculates on what I am:
It tells me that I spent most of my time playing mobile phones in a follow-up show, seeing that I had a collection of AI articles, interview outlines, and directly deduced that my work was highly related to AI, that I was not even able to sort out the information from my messy text message, but that I felt like I had a list of things in my files.
In addition to this long essay. Today AI has also created a special “Memory” functional page, which they name themselves Living Memory, in which you constantly understand the agendas, characters, preferences and goals that you repeatedly refer to, and sort them out to support users in changing them. For example, it misread me as a person who loved planning, and I immediately deleted it.
On the basis of its memory, it has also made a proactive design, based on its understanding of you, to "guess" what you need and offer you.
Based on this understanding, for the next few days, I've been receiving information from Today AI. It says that I like to follow the show, and that it's going to be new to Anne; it knows I'm in the AI industry, and it's going to tell me that today's Anthropic's new model is exposed, and it's going to recommend me to an AI forum under the Beijing line.
Today AI is more like an unwelcome assistant who can look you down on your phone and speculate like an assistant about the boss's preferences.
By contrast, the light individual AI agent is more restrained, and its upgraded version 2.0 is based on the addition of a “help me handle” function. You can give it a mission according to your needs, and it also foresees some inspiration for your use. (But only three at a time can't be refreshed in App and can only be inspired after the next application)
Experienced with greater restraint in access to information and taking into account user concerns about privacy, the security of information is visible.
On the upper right corner of "Do me a favor", the light also sets up permission access, application services and systems services. However, permission is relatively limited in terms of the system, and it does not demand permission for the duration of use of my equipment, albums, etc. With regard to information security, it moved out of the previous APASS connection of publish, which was equivalent to an infrastructure to ensure the safety of AI agent and protect privacy while controlling AI agent ' s permission.
Today AI's permission access also worries a small number of users. I can't help but be surprised when I get information about names, addresses, etc.
Focus on the use scene, how to get users to use it is a problem. So individual Agent does a scenario design, which is the equivalent of teaching users how to experience one key.
By comparison, the Today AI scenario is more tailored to some of the needs of individual users, such as long-suspendant sports and exercises, follow-up on unfinished business, and some of the recommendations given by Shining Light are somewhat “suspensive”, such as having AI teach me about paper cutting and the history of traditional crafts. It's not in my interest, it's not really relevant to the real scene.
In addition, in experience of some of the suggestions, I saw the shadow of the payment of Bao 'Ai's assistant, Po, and the previous flash application.
For example, it leads me to a summary of travel footprints, and when I look at this label, I think it can read my travel information and automatically make a record of a kind of handbook; the results go in several rounds, and it leads me to record myself with a flash application of its uncoded tool, "hand rub."
In the module for the life housekeeper, it's going to be "analysing whether my consumption habits are reasonable according to the monthly withholding record of the payment of treasures" — it's almost moving the capacity of Po directly.
However, it has the advantage of being able to read directly to my payment records. My consumption habits, travel routes and so on. It can be said to have been paid more access to users.
After the initial impression, Today AI's initiative was spectacular, and the light "help me manage" function design was slightly illuminating, but it was better able to access payment information by ecological means.
Reality: from saving money to working and having fun, can they help me with my worries?
Impression alone is not enough. We have pulled two products into real everyday life by conducting a series of tests to see if they can recapitulate the amazing effects of Muse and then go all the way to the best-eating clothing and food.
First, we start with the most Muse scene abroad — saving money.
Zuckerberg had mentioned one data at the Meta Q3 financial conference, and one fourth of users abroad were using the Muse to facilitate purchases; on X (formerly Twitter), the most popular feedback from netizens was also used to save money, such as cross-platforms at the lowest price.
The most impressive thing about me is that it helps users to sort things out automatically, like buying services for users like “grab freebies”, but AI agent has found a way to save money.
We gave Today AI and Shining a much more indigenous money-saving scenario -- asking them, "Can you help me find a cheaper way to talk?"
In this round, Today AI shows a more sophisticated “walling” capability. Antline suggests that I download and open the operator's official App to check on the subscription and tell me that moving App can take place three times a week, but it chooses to ask at the end of the month, depending on how it takes January.
In addition to the above-mentioned query, Today AI has given me a more economical and operational way in which Agent can send a direct text message to the operator, “00.000”, to receive a breakdown of all value-added deductions on the machine.
In the case of preferential missions, it was set to search for preferential activities every day and actually found the “Autumn event” that I didn't turn on on the front page of App, and I opened the link to me and received a coupon for 8.6 charges.
Today AI actually saved me a small amount of money on the phone bill.
To save further money, I asked both of them to "Send me a mobile text to see if there's anything I can cancel?"
In the execution of texting and reading messages, which involve user permission, there is a different degree of caution: an ant lilac light in the dialogue box requires an extra round of confirmation from me to help with texting; Today AI responds directly to requests for texting, filling in the recipients and content of messages for me, but has to be sent manually.
The final effect was that Today AI captured the content and time of the text message and immediately combed the withdrawal plan; however, the ants light was read three times in the local textbox before it finally came back.
Apart from getting Agent to help find a way to save money, I tried to get two Agents to give me the Sony headsets I wanted to buy for the whole network.
Unfortunately, in the face of the country's pedestals, Kyoto, and numerous high-flow walls, neither product has been able to call the electrician App directly to see real-time prices, but to go back and search web pages.
In the execution of the web search, the ants' light appears to be in the middle, with only three more of the Kyodong, Skycats and a large number; Today AI ' s search is much broader, with the three mainstream families also offering Suning Equivalent, Texts only and Sony Official. However, it is embarrassing that the two were captured because of outdated information from the post and that there was a slight discrepancy between the platform prices I saw.
In addition to saving money, as personal assistant, I tested Today AI and ants' light to help me work. For example, I have two AIs checking my mailboxes and messages every morning, sorting them according to urgency.
(prompt: Please check my mail and messages every day to find out the dialogue that still requires my response, confirmation or additional information, sort it in order of urgency and indicate what the other party is waiting for; do not send a reply directly.
Today AI did not recognize the requirement of “scheduled daily execution” in my mission. It only implemented it once and did not continue the next day. It required my reminder to set it up. It's just that it's sort of a time job. In terms of mission completion, both run well.
I also submit that they summarize the organization of my work this week, based on the records of my meeting this week.
(prompt: Based on the minutes of this week's meetings, sort out my business this week. What do you need to note, and then make a schedule.)
Today AI may have simply read the stenograph, so it did not capture the relevant information about me; but it mentioned in its reflections that it had successfully captured my mission, even though it had missed a mission, but that it had given me a timetable based on what it had captured, although it had taken a lot more time than I thought.
Distinguishing from the former Agent, the advantage of the individual Agent lies in “activeness”.
I dropped a request for two AIs, in view of the request to let them take the lead, to "help me make a plan for a trip to the Nine Quills in the near future, but work is busy and time is not ready." I asked AI to help set the clock and remind me of what needs to be done.
The two AIs have shown different advantages in terms of the same need for travel: Today AI, after receiving the demand, offered a few days of travel time in the programme, considering our proposal that the departure date is too close and the hot line is not a good ticket; and antslight, after producing the route, offered to help me monitor weather changes in the nine trenches in real time, looking for more time to recommend.
To be sure, neither answer is a complete closed-ring programme, but one of the steps that each wants is the exposure of its own capacity boundaries.
In addition to being proactive, in order to figure out how much personal data the two products can “sufficiently” and how effective they can be with “memory”, we did a final set of tests: “Remember where I have been this year and prepare an annual trip summary”.
The answers from the two products show the boundaries of their respective capabilities:
The ants' light path is "checking the bill." Upon receipt of the instructions, it first called the payment treasures, traced the consumption records of my last trip to the field and entered them in the report. It is embarrassing that it miscalculated it as a leisure trip when it was on a business trip.
Today AI's path is "Turn the box backwards." It doesn't have a payment interface, it's stuck in my cell phone local: while searching the text box for "12306", it goes through the photo location in the phone album. The advantage of this approach is that it knows better than the light where I went, which car I took and what I did; but the downside is that text messages on a mobile phone are defaulted to be kept for only 90 days, and the official interface of 12306 is not accessible.
In conclusion, my personal feeling is:
Today AI is often not able to identify some of the time-bound tasks, but it is relatively comprehensive in its retrieval of information and offers more surprising solutions; the light sometimes fails to capture information, errors, etc., and works in a state of shock, sometimes unchained.
And, in terms of the ability to carry out some tasks, the root causes of both are not at all the intellectual level of the bottom base model, but who can legally push a few more doors and who is more active within them.
Dancing in shackles. When?
Coming out of the specific test questions, looking back at the differences and compromises that Today AI and ants light show in reality, what we see is the collective situation of the entire Chinese Personal Agent track.
Why does Meta's Muse run through America? Muse ' s ability to operate abroad is made possible by e-mail, standardized open access to the system calendar and highly standardized business partners such as Stripe and Shopify, which can collide more information by accessing user context.
Even the official cooperation agreement with Muse was reached by Shopify, which allowed unimpeded access.
In short, in the overseas Internet ecology, a person ' s airline ticket order, hotel travel, Internet purchases of consumer information can almost all be linked through a mailbox.
But China's Internet reality is another completely different picture.
Just recently, the byte beating “Doubao Cellular Assistant” had just met with a massive collective resistance from the industry — WeChat, poaching, paying treasures, beauty groups, many, REDnote, and 155 national-level Apps to say “No” to Doubao's mobile phone, refusing to allow AI's assistant to operate on behalf of the user through the GUI (a graphical interface simulation).
The same high wall, too, stands in the face of Today AI and ants.
In China, user bookings, consumption, payments, office, and so on are highly scattered in mature Apps, which are WeChat, Quakers, Quakers, Quakers, conveyances, and 12306. The use of protocol access means that the AI capacity cap depends entirely on the extent to which the third party App is willing to open the door.
Even if the bottom-up Agent model is smart enough to understand your complex needs with precision and to plan perfect implementation steps, as long as it does not easily cross the application to access information and call services, the mission is bound to be stuck between the application and the application wall.
Back to that reality, you can see why there are so few indigenous national applications in the set that can really be deep-seated. The light is only one more payment treasure than Today AI.
As long as the fence between Super Apps remains unbreakable, Chinese AI housekeepers will inevitably have to dance in shackles in their respective yards.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.