OpenAI 的 Dot agent 是企业软件，也能帮你订晚餐

这是一个仿佛从上周就开始讲起的故事：OpenAI 的新 agent 平台 Dots，里面有一群能听你差遣的可爱小家伙。

但不同于极其亲民的 Meta Muse，Dots 给人的感觉非常像一套碰巧也能替你点墨西哥卷饼的工作场所软件——重点在于工作。

OpenAI 本周早些时候发布了 Dots。和 Muse 一样，Dots 拥有软乎乎的拟人化头像以及可自定义的名字。OpenAI 表示，未来你将能拥有多个 Dot，但目前只能获得一个。我给自己的 Dot 起名叫 Dotty McDotface。

其界面看起来与 Muse 类似：你在一个窗口中与 agent 对话，在另一个窗口中查看它在虚拟机里点击操作的过程。但 Dots 并不那么像私人购物助手，尽管它们可以帮你买东西。它们更像同事：能够使用其他软件的商业软件。你的 Dot 虚拟机一开始便可访问 Blender、GIMP 等应用。你还可以通过桌面版 ChatGPT 应用授予它访问自己电脑的权限；如果需要把事情用口头方式理清，也可以给 Dot 打电话。OpenAI 首先向最高等级账户的用户提供 Dots，其中包括我为了测试而报销的每月 100 美元 Pro 账户。Muse 及其颇受关注的竞争对手 Instinct 不需要花钱——至少现在如此，至少在 Meta 将其真正推向市场、并且可能开始通过广告等方式从中获利之前如此。不过，OpenAI 的受限发布方式很能说明这款产品是为谁打造的，尽管它有可爱的头像。

我一直在测试其他 AI agent，让它们处理日常个人事务，结果好坏参半；我很好奇 OpenAI 的机器人表现会如何。尽管它更具企业软件气质，OpenAI 表示，凭借云端电脑和应用访问权限，Dots「几乎什么都能做」，因此我让它尝试了一些更日常的使用场景。

首先，我让 Dot 为我家新的互联网服务提供商预约安装时间（听到了吗，Comcast？你要小心了）。它几乎成功了，甚至在我批量删除的邮件里找到了 100 美元的促销折扣，但最后卡在了一个「人类验证」环节。这对机器人而言显然是个问题。它说：「这项特定验证需要持续按住鼠标，而我的浏览器控制功能不支持。请在这里打开我的浏览器并按住『Press & Hold』直到验证完成，然后告诉我，我就会继续处理安装日期。」我照做了，但感觉有点像便利店外的几个青少年请我替他们买葡萄酒饮料。

我想知道 Quantum 为了在这个弹窗里加入「（而不是机器人）」开了多少场会。 截图：The Verge / OpenAI

凭借我这个人类按住按钮的能力，Dot 一路完成到了结账页面，并要求我输入银行账户信息。事实证明，它并没有像 Muse 那样顺畅的 Stripe 集成，可以保存你的信用卡信息并使用虚拟卡结账。而且，无论如何，这家新服务商要求使用我的 银行账户，才能让我的月度账单获得 5 美元折扣。Dot 暂停下来，让我在虚拟浏览器中输入号码；我还不够信任它，不愿输入自己的支票账户号码，于是决定在自己的电脑上全部手动完成。这并不是一个多么出色的开局。

我又交给 Dot 几项个人任务，结果同样好坏参半。它找不到一家我一直想尝试的共享办公空间提供免费参观和入场的选项，反而提出要替我注册一张 35 美元的日票。我让同样在测试中的 Instinct agent 查找试用选项；它找到了，并在几分钟内安排好了预约。事后，我通过电话质问 Dot：「你为什么错过了这个选项？」它回答：「我得出了错误的结论。」我挂断了电话。

我准备开始把这称为照烧鸡测试。 截图：The Verge / OpenAI

后来，我让它从我的 Ikea 账户中找出一件我买过的悬挂式花盆的信息。它报告称，自己无法登录我的 Ikea 账户，因为陷入了循环的安全验证。在互联网的另一个角落，附近的照烧餐厅也不允许 Dot 进入其点餐页面。Dot 的某些特征似乎使它比 Muse 或 Instinct 更常在安全验证上受阻；Dot 提供了 OpenAI 关于网站拦截其云端浏览器流量的页面作为解释。无论原因是什么，它需要的手动输入和浏览器接管都比其他 AI agent 更多，可能是因为它本来就不是为点照烧餐或处理 Ikea 退货而专门打造的。

当我授予 Dot 访问一个完全由我控制的对象——我的个人网站——的权限时，我才真正看到它进入状态。这个网站过时得令人尴尬，我让 Dot 整理出一套更简洁的新设计。它做到了；当我坐下来打出希望在下一版本中修改的内容时，我想：「去他的，我直接给它打电话吧。」我花了 10 分钟，按栏目逐一讲清自己想要的修改，然后离开去洗碗；不久后我收到提示，说下一次迭代已准备好供我审阅。并不完美，但更好了。AI 再次证明：当你给它一大批数据——例如我连续唠叨 10 分钟——并让它转化成有用成果时，它最能发挥价值。之后，我又进行了几轮迭代：在电脑浏览器窗口中与 Dot 聊天，同时在家里收拾 Brio 火车轨道和毛绒玩具时，用手机应用的语音转文字功能继续交流。

Dot 的某些特征似乎使它比 Muse 或 Instinct 更常在安全验证上受阻

当我授予 Dot 访问电脑的权限后，它就显得合理得多了。在那里，我给了它两项任务：把桌面上几段我参与的热门电视节目视频文件合成为一个尺寸适合社交媒体的片段；以及将我们在原型中讨论完成的个人网站改动部署上线。我不得不授予它多得惊人的权限，但它成功了。我看着它在网站后台到处点击，逐个区块重建原型。我离开电脑后，手机上的 ChatGPT 应用会向我推送更新——合成视频片段已经完成，我一边喝咖啡一边审阅了它。这是我人生中最有硅谷感觉的一刻。

我当然也可以自己想办法把所有这些片段拼在一起。事实上，我第一次就是自己做的，但成品看起来很粗糙，而且文件太大，无法上传到任何社交媒体网站。Dot 做得好得多，几分钟内我就得到了可以分享的视频片段。它也完成了网站更新。新设计很基础，但相比我直到昨天还在使用的复杂而过时的版本，已经进步了太多。不得不佩服 Dotty McDotface。

一边帮你更新网站，一边替你点晚餐——这就是我们曾被承诺的 AI agent 未来。 截图：The Verge / OpenAI

如果将 Dot 视为并使用成「面向普通人的 Codex」，它就合理得多：一个可爱的软团，能在你做其他事情时处理部分业务。Muse 这类更偏个人化的 AI agent，感觉像是高级版自动填表工具和通用线上购物车；而 Dot 则像是未来工作方式的先兆。

或者说，_一种_未来的工作方式。可以推测，因为你是预先付费的，它不会像 Muse 和 Instinct 最终可能面临的那样承受变现压力；这很符合企业软件的逻辑。没有人希望自己工作时被 AI 唠叨。当然，前提是你从事的工作确实能由 AI 协助完成，并且你信任 OpenAI 来做这些事。这是两个很大的前提！

它确实帮我完成了一些任务；老实说，如果没有它，我大概根本不会抽时间去做。这相当方便。但我不认为它方便到值得每月 100 美元，至少对我这个行业而言如此。我能理解，对于像我们的朋友 Casey Newton 这样正全力创办一家新企业的人，它会有用得多。

而且，在协助我登录 Uber Eats 账户后，它最终确实帮我点到了照烧餐。

It’s a tale as old as last week: OpenAI’s new agent platform, called Dots, is full of cute little guys who can do your bidding.

But unlike the ultra-approachable Meta Muse, Dots feel very much like using workplace software that happens to be able to order you a burrito — emphasis on work.

OpenAI announced Dots earlier this week. Like Muse, Dots have blobby, anthropomorphic avatars and customizable names. In the future, OpenAI says you’ll be able to have multiple Dots, but right now you get one. I named mine Dotty McDotface.

The interface looks similar to Muse’s; you chat with the agent in one window and follow its work in another as it clicks around on a virtual machine. But Dots are less personal shoppers, even though they can help you buy stuff. They’re more like coworkers: business software that can use other software. Your Dot’s virtual machine can access apps like Blender and GIMP right off the bat. You can also give it access to your own computer through the desktop ChatGPT app, and you can call your Dot if you need to work through things out loud. And OpenAI is offering Dots first to users on its highest-tier accounts, including the $100-per-month Pro account I expensed to test it. Muse and its buzzy competitor Instinct cost you nothing — at least now, and at least until Meta can get it off the ground and presumably start using it to serve you ads, etc. But OpenAI’s gated rollout says a lot about who the product is for, cute avatars notwithstanding.

I’ve been testing other AI agents with day-to-day personal tasks — yielding mixed results — and I was curious how OpenAI’s bot would compare. Despite its more enterprise-y vibes, OpenAI says Dots “can do nearly anything” with its cloud computer and access to apps, so I gave it a shot with some more pedestrian use cases.

To start, I had Dot try to schedule an installation appointment for a new internet service provider at my house (you hear that, Comcast? You’re on notice). It almost got there, and even found a $100 promotional discount in my email that I had mass deleted, but got stuck at a “human check.” Which is problematic for a bot. “This particular check needs a sustained mouse hold that my browser controls don’t support,” it said. “Please open my browser here and hold ‘Press & Hold’ until it completes, then tell me and I’ll continue with installation dates.” I obliged, but it felt a little bit like some teenagers outside of the convenience store asking me to buy them some wine coolers.

I want to know how many meetings Quantum had about adding “(and not a bot)” to this pop-up. Screenshot: The Verge / OpenAI

With my human ability to click and hold a button, Dot got all the way to the checkout screen, where it asked me to input my bank account details. Turns out there’s no slick Stripe integration like Muse has to save your credit card info and use a virtual card for checkout. And in any case, the new provider needed my bank account to get a $5 discount on my bill every month. Dot paused to let me enter the numbers into the virtual browser; I didn’t trust it well enough to punch in my checking account number and decided to just do it all myself on my own computer. Not exactly a spectacular start.

I gave Dot a few other personal tasks, and the results were mixed. It couldn’t find the option for a free tour and entry at a coworking space I’ve been meaning to try out, and instead offered to sign me up for a $35 day pass. I had Instinct’s agent, which I’ve also been testing, look for the trial option; it found it and set up an appointment within minutes. “Why did you miss this option?” I questioned Dot over the phone after the incident. “I jumped to the wrong conclusion,” it said. I hung up.

I’m going to start calling this the Teriyaki Test. Screenshot: The Verge / OpenAI

Later, I had it try to retrieve some information in my Ikea account about a hanging planter I bought. It reported that it was unable to log in to my Ikea account because it got stuck in a looping security check. Elsewhere on the internet, the neighborhood teriyaki spot wouldn’t allow Dot onto its ordering page. Something about Dot’s footprint seems to get it tripped up on security checks more often than Muse or Instinct; Dot offered a link to OpenAI’s page about websites blocking its cloud browser traffic as an explanation. Whatever the reason, it required more manual input and browser takeovers than other AI agents, possibly because it wasn’t purpose-built for ordering teriyaki or handling your Ikea returns.

I really saw Dot come into its element when I gave it access to something fully within my control: my personal website. It is embarrassingly outdated and I asked Dot to pull together a new, more streamlined design. It did, and when I sat down to type out some changes I want made for the next version, I thought, “To hell with it. I’ll just give it a call.” I hashed out the changes I wanted section by section over 10 minutes, walked away to do some dishes, and got a ping that the next iteration was ready for me to review. Not perfect, but better. Once again, AI proves most helpful when you give it a large dataset (me rambling for 10 minutes) it can turn into something useful. I carried out a few more rounds of iterations chatting with Dot in a browser window on my computer and using voice-to-text on the phone app as I walked around the house cleaning up Brio train tracks and stuffies.

Something about Dot’s footprint seems to get it tripped up on security checks more often than Muse or Instinct

Dot made a lot more sense once I gave it access to my computer. There, I gave it two tasks: Put together a handful of video files on my desktop featuring a TV hit I did into one clip sized properly for social media, and deploy the changes to my personal website that we had worked through in a prototype. I had to give it a shocking number of permissions, but it worked. I watched it click around the website backend and re-create the prototype block by block. I walked away from the computer and got updates on the ChatGPT app on my phone — the consolidated video clip was ready, which I reviewed while I drank a coffee. It was the most Silicon Valley I’ve ever felt in my life.

I absolutely could have figured out how to put all those clips together myself. In fact, I did the first time around, but it looked janky and was too big to upload on any social media website. Dot did a much better job and I had a clip to share within minutes. It handled the website update, too. The new design is basic, but it’s miles ahead of the complicated, out-of-date version I’d been rocking until yesterday. Gotta hand it to Dotty McDotface.

Helping you update a website while ordering dinner is the AI agent future we were promised. Screenshot: The Verge / OpenAI

Dot makes a lot more sense when you think of it and use it as “Codex, but for regular people” — a cute little blob that can take care of some part of your business while you do other things. Where the more personal AI agents like Muse feel like glorified autofill and a universal online shopping cart, Dot seems like a harbinger of the future of work.

Or a future of work. Presumably, since you pay for it upfront, it won’t be subject to the kind of monetization pressure as Muse and Instinct eventually will, which makes sense for enterprise software. Nobody wants to be nagged by AI while they work. That is, assuming you do the kind of work that AI can help you with, and you trust OpenAI to do it. Big caveats!

It helped me out with some tasks that, if I’m being honest, I would have just not gotten around to doing. That’s pretty handy. But I don’t think it’s $100-a-month handy, at least in my line of work. I can see how it would be much more useful to someone like our friend Casey Newton, who’s in the thick of launching a new business.

And with some help logging into my Uber Eats account, it did actually order my teriyaki in the end.

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

• Allison Johnson

• AI

• Hands-on

• OpenAI

• Reviews

• Tech