AI应用风向标（公众号：ZhidxcomAI）
作者｜毕伟豪
编辑｜漠影
如果有这样一个AI，它有一部手机和一台电脑在云端24小时运转等你发号施令，能登录你的淘宝、高德、大麦，能帮忙挑选商品，还能在云端电脑上开发游戏，甚至可以一句话做出应用，你会拿它来做什么？
我们最近就遇到了一款这样的产品——千问Go。
我们可以让它打开大麦，帮忙找到想看的演唱会并完成收藏；让它调用云端电脑，下载并替换喜欢的壁纸；甚至直接让它操作电脑写代码，开发一款城市建造模拟游戏。
甚至！我只发了一句话：帮我去淘宝买个好点的4k显示器，接Macmini用。
它在云手机里筛选出了三款显示器，然后让我选，选好以后，它直接给我创建了订单！虽然不能直接付款，但点开淘宝就能看到待付款订单了。
智东西11月3日报道，昨天，我们拿到了阿里千问Go的体验资格。
先给不清楚的小伙伴简单讲一下背景，近几个月由于Muse的爆火，把Personal Agent这个概念的热度引爆了，随后，国内外大小厂纷纷跟进这个主动式解决个人生活中具体需求的产品形态。
而千问Go，正是阿里正在邀测的Personal Agent产品。
智东西在拿到体验资格后，也是第一时间从各个角度进行了实测。从结果来看，千问Go已经能够在已适配的应用中执行多种操作，云手机、云电脑也展现出了一定的实用性，用户还可以通过自然语言随时生成小应用。
当然，实测中也能看到一些有待完善的地方，包括数据源覆盖范围、任务执行延迟、操作流畅度等。但总体而言，千问Go已经展现出较为完整的任务执行能力，实际体验也有不少亮点。
一、自带手机和电脑， 能连淘宝，上大麦
千问Go是一个内置在千问APP里的模式，用户获得体验资格以后，应用上方将会出现千问Go的选项。点击进入后，一个简洁的页面映入眼帘，对话框上方有一些功能区域，最核心的是云端和我的应用这两部分。
点击“云端”选项卡，能看到有云手机、云电脑以及本地电脑三个选项。其中，云手机和云电脑分别对应着一台跑在云端的虚拟设备，这些设备里安装了目前千问Go能够支持操作的几款应用，手机端有淘宝、大麦等，电脑有VS Code、Chrome、夸克等。
这些云端设备，都是支持实时控制的。比如我们让千问Go在云手机的淘宝里搜一下狼人杀桌游套装然后添加购物车，它就会帮忙打开淘宝并进行搜索然后添加购物车，并且用户还可以随时接管这台手机来自行操作。
当然，操作淘宝需要用户提供手机号进行登录，用户可以选择把手机号和验证码发给千问Go或者手动登录。
同理，大麦也可以在云手机里操作，比如我们让千问Go帮忙找一下黄霄云的演唱会然后添加一下“想看”。千问Go会直接在大麦里搜索并执行添加的操作。
值得注意的是，受平台安全限制，大麦是不可以由Agent自动登录的，用户需要手动在云手机里进行登录。
云电脑也很有趣，千问Go会提供一个运行在云端的PC主机，里面有一些内置的应用，比如我们可以打开夸克，然后让它下载一些《画江湖之不良人》的壁纸保存到网盘里。
夸克应该是千问Go适配程度比较高的应用，在任务执行过程中，如果用户手机上有夸克，是可以直接关联的。不过，想要在云电脑上查看夸克中的具体内容时，仍需要通过浏览器重新扫码登录，操作衔接不够顺畅。
二、复盘年度消费，帮忙抢演唱会门票
看完了基础功能和操作，我们来让千问Go去尝试做一些比较复杂的任务，比如拉取淘宝、高德的数据来对年度消费进行复盘。提示词如下：
请根据我淘宝、高德的历史购物、出行和生活记录，制作一份我的年度生活审计报告。不要只统计金额，还要从消费习惯、兴趣偏好、生活轨迹等角度分析我。找出三个最意外的发现，并给我颁发一些有趣的年度奖项，例如最冲动消费奖、最值得投资奖、重复购买冠军等。所有结论必须有真实数据支撑。
千问Go通过云手机抓取了我淘宝和高德的数据，当然，如果大家日常使用飞猪的话，应该还可以抓取出行信息。千问Go根据我的信息，对我26年的消费进行了细致的盘点，还提示我购物车正在腐烂（下架商品过多）。
不过值得注意的是，千问Go目前可以抓取的数据源比较有限，同时当操作涉及资金安全边界的时候平台会有风险提示。
可能上面这个还不太直观，我们换一个测试，正好试试看千问Go的核心功能之一“我的应用”。据介绍，在我的应用模式下，千问Go会连接阿里的AI应用开发平台秒悟，用户可以通过需求描述快速生成应用。
比如，我们让千问Go去抓取大麦的六位歌手的演唱会信息，然后帮我制作一个演唱会抢票应用。提示词如下：
我想看演唱会，法老、弹壳、杨和苏、毛不易、徐佳莹、黄霄云这些吧，然后帮我制作一个盯票、抢票系统，提前通知我抢票，实时更新内容。系统要结合高德和飞猪，做一个演唱会路线版块，规划每张票从北京出发看演唱会的路线与成本。
千问Go从大麦上抓取了相关信息，然后生成了抢票应用，并且还设置了一个定时任务，每天定时去抓取这些歌手的演唱会票务信息，然后实时更新这个应用。
虽然现在千问Go还不能替我下单、抢票，但比对平台商品价格、查询演唱会票务信息这些事情，显然它已经能够很好地完成了。
我们还让千问Go做了一个文字模拟人生的小游戏，采用了前面抓取的淘宝和高德数据以及我输入的开支信息，每天有两点行动力，可以选择上班、加班、摸鱼、约会等选项，通过这些选项来模拟人生进度。
说实话，用这个来预测一下什么时候升职加薪还挺有意思的。
三、手机里也能玩电脑，云端开发模拟建造游戏
云手机都测试了，怎么能不试试云电脑呢。
千问Go里的云电脑中有Chrome浏览器、夸克以及VS Code，一些比较简单的指令，例如用云电脑登录GitHub抓取热门开源项目、搜索图片然后更换桌面壁纸，都可以比较轻松地完成。
接下来我们用两个比较大的任务来测试一下云电脑的性能，首先我们让千问Go去检索全网信息并制作全球AI产业数据可视化大屏，这个任务需要检索和编程两步进行，提示词如下：
千问Go检索了非常多的信息之后，用VS Code编写了一套可视化大屏，整个过程都可以在云电脑上实时预览。能看到这块屏幕效果相当不错，包含公司、数据和可交互的产业地图。
因为是云端环境开发，所以千问Go还特意打了离线可用的压缩包，解压后即可打开运行。
我们还让千问Go去开发了一款城市建造模拟游戏，虽然是一个比较常规的模型测试任务，但非常考验千问Go调用云主机中编程工具的能力，提示词如下：
实际体验中，千问Go的开发效率超出了预期。尽管云端桌面环境存在一定的帧率下降，但从整体开发过程来看，它对任务的执行比较顺畅，最终呈现的效果也不错。
整体实测下来，云主机的模式在千问Go上已经能够看到效果了，不管是云手机还是云电脑，都能够相对精确地完成我们提交的任务，同时在已有应用的操作上也比较流畅。
不过，在一些具体环节上，还是存在一些小问题，比如云手机和云电脑在用户接管模式下都会存在响应迟缓的情况。就云电脑而言，接管模式下手机操作其实是不太方便的，如果后续产品和开发团队能够优化接管体验感，例如增加横屏和一些操作上的便捷方式等，相信这部分产品功能会更吸引人一些。
价格上，千问Go的免费额度有限，需要订阅来获取更多额度，分为高级、精英、旗舰和智尊四档，连续包月价格分别为19元、49元、128元、399元，最高额度是免费版的50倍。
写在最后：
从实测过程中看，Personal Agent似乎还有很长的路要走，而且这条路上的阻碍也将会是前所未有的多。
具体而言，操作第三方应用这件事，一直让Personal Agent处在一个相对尴尬的位置，以至于千问Go的解法依然是云手机和云电脑的“暴力”方式 ，而这种方法无疑会带来风控的问题。
如何让Personal Agent不会因为第三方应用而苦恼，将会在未来很长一段时间内，成为各家思考的焦点。
目前，千问Go所采用的机器半自动登录以及自动化抓取相结合的方式，具备一定的合理性，但也存在延迟高和用户接手情况频繁的问题。
如何获取信息、如何“无痛”操作第三方平台、如何让Personal Agent操作更流畅，任重而道远。
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。
AI Application Wind Cursor (WeChat Official Account:Zhidxcom AI)
The author of the book, Bee Wai Ho.
Edit Indigo
What would you do with an AI that has a mobile phone and a computer operating 24 hours a day in the clouds that you can order, log in to your treasures, gords, barley, help select goods, develop games on cloud computers, or even use them in one sentence?
We've recently met a product like this — a thousand questions for Go.
We can open barley, help find the concert and finish the collection; get it to the cloud computer, download and replace the wallpaper that you like; and even let it operate computer writing codes and develop a city-building simulation.
Even! I just said, "Get me a better 4k monitor for Macini."
It filtered three monitors on the cloud phone, and then let me choose, and when it was finished, it just created an order for me! Although payments could not be made directly, the order of outstanding payments could be seen when the treasures were opened.
On 3 November, Jigwan reported that yesterday we had acquired the qualifications of Ali thousands asking Go.
A brief background was given to the unsure partner, who in recent months, as a result of the Mouse fire, detonated the heat of the Personal Agent concept and followed up on this proactive product pattern of individual life, both domestic and international.
And it's the Personal Agent product that Ali is asking for.
When intellectuals are qualified for experience, it is also the first time that they are measured from all angles. As a result, Go has been able to perform a variety of operations in a suitable application, and cloud cell phones and cloud computers have shown some practicality, and users can readily generate small applications in natural languages.
There are, of course, some areas to be improved in the field, including coverage of data sources, delays in task execution, operational fluidity, etc. Overall, however, Go has demonstrated a more complete capacity to carry out tasks, and there are many highlights of practical experience.
I, with my own phone and computer, can hunt treasures and go to barley.
Go is a model that is embedded in APP, and when users are qualified for experience, the application above will have the option of asking Go. When you click, a simple page is shown in the curtains, and there are some functional areas above the dialogue box, at the core of which are the clouds and my applications.
Click on the Cloud tab to see three options: cloud cell phones, cloud computers and local computers. Among them, a cloud cell phone and a cloud computer correspond to a cloud-run virtual device, which contains several applications currently available in order to support the operation of Go, with a mobile phone carrying treasures, barley, VS Code, Chrome, Quake, etc.
These clouds support real-time control. For example, if we ask Go to search the cloud phone for a table-knife suit and add a shopping cart, it will help open the shop and search and then add a shopping cart, and the user can take over the phone from the business at any time.
Of course, there is a need for users to provide mobile phone numbers to log in, and users can choose to send cell phone numbers and authentication codes to Go, or to log in manually.
By the same token, barley can also operate on a cloud cell phone, for example, when we ask Chiu Go to help look for a yellow cloud concert and add "want to look." Thousands ask Go to search barley directly and perform added operations.
It is noteworthy that, under platform security restrictions, barley is not automatically recorded by Agent and users need to log in manually on cloud phones.
The cloud computer is also interesting, and Go will provide a PC host running in the clouds, with some built-in applications, such as Quake, which can then download some of the wallpapers from The Bad People of the River and the Lake to be stored on the web.
Quake is supposed to be a relatively high-suitability application for Go, which can be directly linked in the course of a mission if the user's mobile phone contains Quake. However, in order to view the specific content of Quake on the cloud computer, there is still a need to recode the login through the browser, which does not function smoothly.
ii. Recycle annual consumption to help rob concert doors Tickets.
We've finished the basics and operations, and we've come to ask Go to try some of the more complex tasks, such as collecting data from treasure hunts and Goth to reset annual consumption. prompt reads as follows:
Please produce a report on my annual living audit based on my treasure hunt, my history of shopping, travel and my life record. It's not just about the amount, it's about the consumption habits, the preferences, the trajectory. Find out the three most unexpected discoveries and give me some interesting annual awards, such as the most impulsive consumer prizes, the most worthwhile investment prizes, and the repurchase of champions. All conclusions must be supported by real data.
He asked Go to retrieve my data on treasure hunts and Gautts via the cloud phone, and of course, if you use flying pigs on a daily basis, you should be able to capture travel information. According to my information, Go made a detailed inventory of my 26-year consumption and also suggested that my shopping car was rotting.
It is worth noting, however, that Go currently has limited access to data sources and that the platform has risk tips when operations involve financial security boundaries.
Maybe it's not intuitive, so let's try one of Go's core functions. It was described that under my application model, Go would connect to Ali ' s AI application development platform, and users could quickly generate applications through demand description.
For example, let's ask Go to retrieve the concert message from the six barley singers and then help me make a concert robbery application. prompt reads as follows:
I want to see the concerts, pharaohs, shells, Yang and Sue, Mao, Seo Jiaxuan and Huang Xiao Yun, and then help me create a ticket-ticking and ticket-taking system, notify me in advance of the robbery and update the content in real time. The system combines Gode and flying pigs to make a section of the concert's route and to plan the route and cost of each ticket from Beijing to watch the concert.
He asked Go for information from the barley, then generated a robbery application and set up a time-bound mission to retrieve the concert tickets from the singers every day and update the application in real time.
Although it is not possible to ask Go for a ticket for me now, it is clear that it can be done better than the price of the platform's commodities and the search for ticket information.
We've also made a little video-image game for Go, using the treasure-hunting and the Gold. The data and the information I've entered about spending, which is two things per day, can be used to model life progress by choosing options such as work, overtime, fishing, dating, etc.
To tell you the truth, it's interesting to use this to predict when a promotion is due.
iii. Computers can also be played on mobile phones.
It's been tested. How can we not try the cloud computer?
Among the cloud computers in Go are Chrome browsers, Quake and VS Code, easier instructions, such as entering GitHub with a cloud computer to capture hot-door open-source projects, searching for pictures and replacing desktop wallpapers.
So we're going to test the performance of the cloud computer with two larger missions, the first of which we're going to ask Go to retrieve information on the entire web and to create a visual screen of global AI industry data, which will need to be retrieved and programmed in two steps, as follows:
Following a very large amount of information retrieved by Go, a video screen was developed using VS Code, which allows real-time previews on cloud computers. It's pretty good to see it, with companies, data and interactive industrial maps.
Because it was cloud-based environmental development, Go was also specifically involved in the compression package available to offline, which could be opened once the pressure was lifted.
We have also asked Go to develop a city construction simulation game, which is a more routine model test mission, but which tests very well the ability of Go to access the cloud host programming tool, as follows:
In practical experience, Go was more efficient than expected. Although the cloud desktop environment has experienced a certain decline in frame rates, the overall development process has shown a smoother implementation of the mission and, ultimately, a good performance.
The overall picture is that the cloud host model can already be seen on Go, whether it's a cloud phone or a cloud computer, to perform the tasks that we've submitted with relative precision, and to be operational in the application that already exists.
However, there are some minor problems at specific points, such as the slow response of cloud phones and cloud computers under the user-take-over mode. As far as cloud computers are concerned, mobile phone operations under the takeover model are not convenient, and it is believed that this part of the product will be more attractive if follow-up products and development teams are able to optimize the experience of taking over, such as additional screens and some operational conveniences.
In terms of price, Go’s free quota is limited and requires subscriptions to obtain more quota, which is divided into four classes: senior, elite, flagship and intellectual. The monthly price of successive parcels is 19 yuan, 49 yuan, 128 yuan, 399 yuan, with the highest quota being 50 times the free version.
At the end:
In fact, it appears that Personal Agent still has a long way to go and that the obstacles on this road will be unprecedented.
Specifically, the operation of third-party applications has kept Personal Agent in a relatively awkward position, so that Go’s solution remains the “violence” approach to cloud phones and computers, which undoubtedly poses a problem of control.
How to keep Personal Agent from suffering from third-party applications will be the focus of thinking for a long time to come.
Currently, there is a reasonable combination of machine semi-automatic log-in and automated grabs used by Go, but there are problems of high delays and frequent user takeovers.
How to access information, how to operate a third-party platform “without pain” and how to make the operations of Personal Agent more fluid is a long way to go.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.