OpenAI 如何将代码工具打造成常驻个人助理 dots
英文标题：How OpenAI turned its coding tools into dots, an always-on personal assistant
OpenAI 近期推出可在后台常驻运行的个人 AI 助理 dots，该产品源自其此前构建的代码工具 Codex，目前面向每月 100 美元起的 Pro 级别及以上订阅用户开放。通过连接用户的日历、邮件和本地计算机环境，dots 能够在无需提示词触发的情况下主动预测需求并处理复杂事务，代表了深度学习应用从被动响应转向主动 Agent 的趋势。然而，主动式 Agent 在深度获取用户数据时的隐私边界、对底层现实基础设施的冲击，以及未来整合支付交易权限的可行性，仍是整个行业面临的严峻考验。
3.1 KB
389 B
33.1 KB
6.7 KB
53.0 KB
248 B
42.6 KB
中文
常驻后台的个人 AI 助理——即使在你没有主动下达指令时也在后台持续工作——成为了今年关于人们究竟将如何使用深度学习模型的关键答案。初创公司 Instinct 一夜成名，而 Meta 也全力投入其助理产品 Muse 的营销之中。
OpenAI 自然不甘落后。该公司在 9 月下旬的开发者日上发布了 dots，这也是其交出的助手级产品答卷。负责 OpenAI 企业产品团队的 Alexander Embiricos 本月参加了 TechCrunch Disrupt 大会，深入阐述了该产品的诞生过程——首先探讨的问题是：为什么如此多的公司似乎都在采用软萌可爱的吉祥物，来充当这些万亿参数矩阵乘法机器的虚拟化身？
“在打造 AI 助理的探索中，其实正在发生一种趋同进化，”Embiricos 在谈到构建 AI 助理的努力时表示，“在研发这款产品的过程中，我们意识到，如果能将其界定为一个实体和一次持续的对话，而不是无数个分散的会话……整个体验就会开始变得直观起来。”
该助手目前已在 ChatGPT 桌面端和移动端应用上线，面向 Pro 级别或更高阶的订阅用户开放，套餐起售价为每月 100 美元。一旦连接到你的日历、电子邮件、计算机以及其他工具，该 Agent 就会尽力协助你完成工作。
Embiricos 指出，利用 OpenAI 今年早些时候发布的编程工具 Codex 应用，其实就能实现 dots 的大部分功能，但这需要更多的操作和逐步引导。他回忆道，当看到妻子开始整夜开着电脑以便让她的 Agent 继续在后台工作时，他感到由衷的高兴。
“在产品团队中，我思考的是：我们如何让它对人们来说既直观又实用？”他在 Disrupt 大会上对与会者表示，“这在很大程度上并不是要在产品中塞入某种全新的技术，而是在于你如何向用户呈现它，以及它带来的实际体验如何。”
让技术变得直观，意味着需要将公司模型的能力重新包装成更容易获取的形式；在之前基于为软件工程师构建的 Agent 式编程工具孵化出 ChatGPT Work 时，OpenAI 就已经走在了这条路上。甚至这个名字——源于 OpenAI 标志内部负空间的巧妙构思——也是为了“让人觉得它就是某种最终可能会无处不在的东西”。（当被问及何时得知竞争对手、AI 企业家埃隆·马斯克（Elon Musk）拥有 dot.com 域名时，Embiricos 说：“我很久之后才发现。”）
常驻 Agent 的核心理念在于它不需要即时 Prompt 提示词触发。对 OpenAI 而言，这意味着更高的 Token 消耗量；对用户而言，则意味着获得了一个更具前瞻性的助手，“它在某种程度上会为你着想，并思考如何为你提供帮助”，正如 Embiricos 所描述的那样。
在我们的测试中，该助手预先思考的能力显而易见且通常很有帮助，但目前还不够完善。它能够标记日历冲突和即将到来的截止日期，但不一定总能把握最及时的时机。
“在早期开发阶段，它第一次向我提醒某个日程冲突时，是在事情发生前三分钟才告诉我的，”Embiricos 谈到他的 dot 时说道，“它就像是在说：‘嘿，这件事情马上就要到了，你可能得处理一下。’可我只剩下三分钟了！”
Embiricos 预计随着时间推移，时机把控会逐步改善，但主动介入的工作方式正是该产品的核心。“我曾听说一家无人机配送初创公司的工程师提到：‘我们遇到了系统宕机故障，而当我刚察觉到宕机发生时，我的 dot 早就已经注意到了这次宕机，对故障原因进行了分析，并带着有助于解决宕机的实用信息主动来找我了。’”
向 dot 交付系统权限
与大多数 AI Agent 的应用场景一样，AI 模型为了发挥效用所需的个人信息量，与你究竟向一家软件公司暴露了多少个人数据之间，存在着天然的张力。Meta 的 Muse 就曾因其构建的个人档案（包括你的家人和朋友信息）而饱受批评；该软件利用这些档案来推测你可能想和某位朋友去什么样的餐厅，但这同时也可能为 Meta 的广告业务带来难以估量的价值。那么 dots 是否也在收集类似的数据？
“是的，我们在 dots 中内置了一个我们称之为‘主动研究’（proactive research）的功能，”Embiricos 表示，“它会定期根据对你的了解、以往的对话记录以及你开放给它的权限内容进行思考，然后去调研：‘我能做些什么来更好地协助你？’接着，它基本上就会主动提出相应方案。”
然而，要将 Agent 的数据需求与其母公司的数据诉求完全剥离开来十分困难。正在开发自身个人 AI 助理的初创公司 Hark 就明确指出，他们不会像 OpenAI 和 Meta 那样为了推动广告业务而收集数据。
针对 dots，OpenAI 表示该 Agent 拥有与 ChatGPT 本身相同的隐私控制机制：用户可以关闭模型训练（阻止 OpenAI 使用其对话数据来改进 AI），并控制其访问计算机内部资源的范围和方式。尽管如此，这些控制目前还不够细致，且该应用有时会表现得比较强势。例如，它目前拒绝接受对云盘的“只读”访问权限。
“我认为我们需要做的是构建极其细腻、极其颗粒化的控制机制，”Embiricos 说，但他同时指出，模型的智能水平和自主判断力是这枚硬币的另一面。
“这其实是我们在产品刚发布时受到一些批评的地方，”Embiricos 表示，“我们在 Agent 能够采取的行动上设置了非常保守的默认策略，它会频繁地请求权限许可。我记得当时就预料到会收到这类反馈，但我觉得‘这没关系，随着时间推移我们可以逐步进行微调’。”
个人 AI Agent 能否实现规模化？
关于 AI Agent 的另一个核心疑问在于它们能否实现规模化扩展。当只有少数人拥有它们时，体验确实很有趣，诸如 Agent 成功抢到热门预约的故事也不绝于耳。然而，企业并不一定乐于与 Agent 打交道，甚至可能根本不具备相应的承接能力。人们也肯定不喜欢与它们对接（顺便提一句：别再让你们的 AI Agent 来向我推销了），Embiricos 本人也是持这种态度的人之一。他透露自己身边依然配有一名人类助理，因为 OpenAI 的企业合作伙伴希望与真人互动，他还补充道：“我个人绝不打算和别人的 AI Agent 开会。”
如果 AI 行业认定的“Agent 是我们的未来”这一观点是正确的，我们能否在不压垮现有商业体系、网站以及支付网络等基础设施的前提下达成这一愿景？
“在 OpenAI……在推出 Agent 与外部世界交互的新功能方面，我们并不总是走得最快的，”他说，“当我们推出新能力时，必定会与整个生态系统保持极其紧密的合作。在这里我可能不便透露过多，但我们正在与多家支付处理机构展开深入的交流与沟通。”
通过与支付公司合作让 AI 助理具备代为花钱的能力，对 OpenAI 来说可能会是一笔潜在利润丰厚的首创业务，但这对消费者而言也是一个巨大的心理门槛。允许助手访问我的电子邮件收件箱已经让人感到要求过高了，而让它接入我的银行账户在目前看来依然是一条不可逾越的底线。至少目前如此。
English / 原文
The always-on AI personal assistant, which keeps working in the background even when you aren’t asking it to, is this year’s answer to the question of how people will actually use deep learning models. The startup Instinct became an overnight sensation, and Meta put the full force of its marketing behind its assistant, Muse.
OpenAI wasn’t going to be left out. The company released dots, its assistant offering, at its developer day in late September. Alexander Embiricos, who leads OpenAI’s enterprise product team, joined us at TechCrunch Disrupt this month to explain where the product came from—starting with the question of why so many companies seem to be turning to cutesy mascots as avatars for their trillion-parameter matrix multiplication machines.
“There’s kind of this convergent evolution happening,” Embiricos said of the efforts to build AI assistants. “As we were working on the product, we realized if we could frame this as, it’s an entity and it’s a single conversation, not many different conversations….it starts to become intuitive.”
The assistant is available in the ChatGPT desktop and mobile apps to subscribers on the Pro tier or higher, with plans starting at $100 a month. Once it is connected to your calendar, email, computer and other tools, the agent tries to help you get your work done.
Embiricos points out that you could achieve much of dots’ functionality with Open AI’s Codex app, its coding tool, released at the beginning of this year. But that requires more work and hand-holding. He recalled celebrating when his wife started leaving her computer on all night so that her agents could keep working.
“On the product team, what I think about is, how do we make this intuitive and useful to people?” he told Disrupt attendees. “A lot of that isn’t building some new piece of tech into the product; it’s how you frame it to the user and how it feels.”
Making the technology feel intuitive meant repackaging the functionality of the company’s models into something more accessible, a path it was already on as it spun ChatGPT Work out of the agentic coding tools it had built for software engineers. Even the name—a play on the negative space inside of OpenAI’s logo—is intended “to feel like just something that could eventually be ubiquitous.” (Asked when he learned that rival AI entrepreneur Elon Musk owns the dot.com URL, Embiricos said, “I found out very late.”)
The idea of a persistent agent is that it doesn’t need prompting. For OpenAI, means more token usage, and for users, a more pro-active assistant “that’s sort of thinking about you and thinking how to help you,” as Embiricos put it.
In our testing, the assistant’s ability to think ahead is apparent and often useful, but it isn’t quite there yet. It will flag calendar conflicts and upcoming deadlines, but not necessarily in the most timely fashion.
“The first time it told me about one of those [conflicts]—this is like early in development—it told me three minutes before the thing,” Embiricos says of his dot. “It’s like, ‘Hey, this thing is coming up. You might want to do something about it.’ I only have three minutes!”
Embiricos expects the timing to improve over time, but the proactive approach is key to the product. “I was hearing about an engineer at a drone delivery startup, and they were saying, ‘We had this outage, and by the time I noticed that the outage was going on, my dot had already noticed the outage was happening, had already analyzed the outage, and had come to me with useful information for how we could resolve the outage.’”
Giving dot the keys
As with most AI agent use cases, there’s a tension between the personal information an AI model needs to be useful and exactly how much personal data you are exposing to a software company. Muse has been criticized for the profiles it builds of your family and friends, which the software uses to know, say, what kind of restaurant you might want to go to with a specific pal, but could also add incredible value to Meta’s advertising business. Does dots collect similar data?
“Yes, we have a thing that we built into dots, which we call proactive research,” Embiricos said. “On a cadence, it will think about what it knows about you, what is the conversation that it’s had, and then what have you given it access to that it can think about, and then it will go and research, ‘What could I do to be more helpful to you?’ And then, basically, it will propose that.”
It’s difficult to separate the data needs of the agent from that of its parent corporation. Hark, a start-up developing its own AI personal assistant, is explicit in pointing out it that it doesn’t collect data to fuel an advertising business the way that OpenAI and Meta do.
With dots, OpenAI says that the agent has the same privacy controls as ChatGPT itself: Users can turn off model training (stopping OpenAI from using their conversations to improve its AI), and control what resources it accesses within your computer and how. Still, those controls aren’t granular and the app can be pushy. For example, it refuses to accept read-only access to cloud drives.
“What I think we need to do is basically build out super, super, super granular controls,” Embiricos said, but he pointed to the model’s intelligence and discretion as the other side of this coin.
“This is actually something we kind of got some flack about when we launched,” Embiricos said. “We had really conservative defaults over what the agent would do, and it would ask for permission a lot. I remember knowing that we were going to see that feedback, but thinking, ‘This is okay, we can tune that over time.’”
Can personal AI agents scale?
The other big question about AI agents is whether they can scale. They’re fun when only a few people have them, and stories of agents snagging hard-to-get reservations abound. But businesses don’t necessarily like dealing with agents and might not be equipped to. People certainly don’t like dealing with them (note: stop having your AI agents pitch me), and Embiricos is among those people. He also has a human assistant because OpenAI’s corporate partners want to interact with a person, he said, adding, “I personally never intend to have a meeting with someone else’s agent.”
If the AI industry is right that agents are our future, can we get there without overwhelming the infrastructure that our world currently runs on, from businesses and websites to payment networks?
“At OpenAI…we’re not always moving the fastest in terms of shipping the next ability for the agent to work with the world,” he said. “When we do, we’re doing it really closely in collaboration with the rest of the ecosystem. I probably can’t say too much here, but we’re deep in lots of conversations with payment processors.”
Letting an AI assistant spend money through partnerships with payment companies would be a potentially lucrative first for OpenAI, but also a big ask for consumers. Giving an assistant access to my email inbox already feels like a big ask, but my bank account is off limits. For now.