How to Choose Your Personal AI Agent — 文稿

M10:00

我们正式被个人代理淹没了。

We are officially drowning in personal agents.

在 Muse、GrokBot、OpenCLAW、Hermes、DotsNow、Instinct 以及 Anthropic 最终肯定会推出的产品之间，这种形态的产品到处都是。

Between Muse and GrokBot and OpenCLAW and Hermes and DotsNow and Instinct and whatever Anthropic inevitably launches, this is a form factor that is absolutely everywhere.

考虑到要充分利用这些个人代理需要大量的上下文、设置、工具权限和账户权限，切换的成本可能会相当高。

And given how much context and setup and tool access and account access is going to be required to get the most out of these personal agents, the cost of switching could be kind of high.

基于以上这些，今天我们来看一个指南，教你如何根据一系列不同标准来选择个人代理，这些标准应该能帮你聚焦到最适合你的那个。

Given all that, today we are looking at a guide to how to choose a personal agent based on a set of different criteria that should help you hone in on which one is best for you.

The AI Daily Brief 是一个每日播客和视频节目，专注人工智能领域最重要的新闻和讨论。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

好的朋友们，在深入之前先快速说明一下。

All right, friends, quick announcements before we dive in.

M11:05

今天我们聊的是个人代理，而我对这个事的态度是这样。

Today we are talking about personal agents, and here's my shtick on this.

我觉得现在还太早，对眼下所有人都在争先恐后为你实现的这种特定形态最终是否会成为主流，抱有一些怀疑是正常的。

I think it's early enough that having some amount of skepticism that this particular form factor that everyone is racing to implement for you will ultimately be where things land.

然而，你很可能拥有至少一个高度互联的代理，它整合了你的邮箱账户、Slack 或 Teams，甚至可能能访问你的财务账户，这种想法会变得越来越普遍。

However, the idea that you are likely to have at least one highly connected agent integrated with your email accounts, your Slack or Teams, and even potentially having access to things like your financial accounts is going to be increasingly normal.

我的论点就是，即使你不确定这是否完全适合你，我认为还是值得挤出一些实验时间，去看看使用个人代理是否以及如何影响你的职业或个人生活。

My argument then is that even if you're not sure that this is exactly a fit for you, I think it's worth carving out some experimentation time to see if and how using a personal agent impacts anything in your professional or personal life.

当然，这不意味着你必须让它访问所有东西才能获得真实感受，而是意味着你要投入一些实际的时间和反复尝试。

Now, that does not mean you have to give it access to everything to get a real sense of it, but it does mean putting in some actual time and reps.

但是你的时间很宝贵，而且非常有限。

But your time is precious and there is too little of it.

那么你打算选择哪个个人代理来实验呢？

So which personal agent are you going to choose to experiment with?

今天的节目完全围绕这个问题展开，但我们会把它分成三个部分。

Today's show is all about that question, but we're gonna divide it into 3 parts.

我会先介绍一个高层次的思考框架，分享一个我制作的小型互动问答，你可以在听完本期节目后自己试一下。但在那之前，我想读一下 Ethan Mollick 教授在他 One Useful Thing 博客上发表的最新文章《The Dot and the Swarm》中的一些摘录。

I'll go through a high-level framework for thinking about it, share a little interactive quiz that I built that you can use yourself after you listen to this episode, But before we do that, I want to read some excerpts from Professor Ethan Mollick's most recent post on his One Useful Thing blog, which is called The Dot and the Swarm.

这是对这种个人代理形态的思考，以及为什么即使是像 Ethan 这样密切关注事物的人也会错过 AI 的发展方向。

It's a meditation on this personal agent form factor and why even someone who watches things as closely as Ethan does can miss where AI is headed.

总体来说，我觉得在写这个 Substack 的几年里，我在预测 AI 的方向和速度方面做得不错，但我认为最近我在某件相当重大的事情上搞错了。

I generally think I have done a good job anticipating the direction and pace of AI over the few years I've been writing this Substack, but I think I recently got something fairly large wrong.

过去一年里，我一直发文说我猜想人类必须以管理者的身份来与代理合作，决定如何将工作委派给代理，并指定这些代理应该如何组织。

In the last year, I've been posting about how I suspected that humans would have to approach working with agents as a manager, deciding how to delegate work to agents and specifying how those agents should be organized.

我曾认为，要让代理作为一个群体有效协作，需要像创建公司那样精心构建，并且需要时间来摸索。

I thought that getting agents to work effectively as a group would take careful construction akin to building a company, and that this would take time to figure out.

不。

Nope.

我成了苦涩教训的受害者。

I fell prey to the bitter lesson.

一个反复被学到的严酷真相：我们过去认为需要复杂人类规则和思考的事情，可以通过更好的机器学习系统和更多 AI 的蛮力来解决。

The hard truth learned over and over again that things that we thought required elaborate human rules and thinking can be solved with the brute force of better machine learning systems and more AI.

苦涩教训在 AI 初创公司和采用 AI 的公司中随处可见。

The bitter lesson is everywhere among AI startups and companies adopting AI.

人们投入了大量精力构建复杂的计算机系统，以便在正确的时间向 AI 提供正确的信息，但 AI 系统已经学会了自行寻找信息。

A huge amount of effort went into building elaborate computer systems to feed AIs the right information at the right time, but AI systems have learned to seek out information themselves.

同样的事情也发生在提示工程上。

The same thing happened to prompting.

人们构建了复杂的模板和提示链，一步一步引导 AI 完成任务。

People built elaborate templates and chains of prompts that walk the AI through a task one step at a time.

后来新模型被证明更擅长自己规划步骤，而我们的研究表明，规划步骤的价值大大降低了。

Then newer models turned out to be better at planning the steps themselves, and as our research shows, planning steps have much less value.

作为一个教授管理者课程并且发表过管理学研究成果的人，我想我曾相信管理代理会有所不同。

As somebody who teaches managers and has published research on management, I guess I believed that managing agents would be different.

人类研究管理学已经很久了，但还没有完全搞明白。

Humans have been working on management for a very long time without fully figuring it out.

这似乎是那种需要由人来设计的事情，至少在一段时间内是这样。

It seemed like the kind of thing that would need to be designed by people, at least for a while.

结果发现，组织工作只不过是 AI 能学会做的又一件事情。

It turns out that organizing work is just one more thing AI can learn to do.

这就引出了 Dots 和 Muse。

Which brings us to Dots and Muse.

现在 App Store 排名第一的应用是 Meta 的 Muse。

The number one app in the App Store right now is Meta's Muse.

一个承诺能帮你干活的个人智能体。

A personal agent that promises to do work for you.

OpenAI 也发布了一款叫 DOTS 的竞争工具。

OpenAI has now released a competitor tool called DOTS.

不止他们——SpaceX 的 GrokBot、Instinct 和 Gemini Spark 也都做着类似的事情，大同小异。

They aren't alone— SpaceX's GrokBot, Instinct, and Gemini Spark all do similar things, more or less.

所有这些智能体都受到一个你可能还记得的现象的启发——今年早些时候的 OpenCLAW。

All of these agents draw inspiration from a phenomenon you might remember from earlier this year— OpenCLAW.

OpenCLAW 及其后继者（我称之为 CLAW-like）的思路是：它们让 AI 智能体接入一台电脑，连接到你的各个账户——邮件、财务记录等等。

The idea of OpenCLAW and its successors, which I will call CLAW-likes, is that they give an AI agent access to a computer and connect to your accounts— emails, financial records, etc.

它们实时分析这些数据并做出反应，即使你不在看。

They analyze and react to that data in real time, even when you aren't looking.

诀窍在于，你跟这个模型说话就像跟人说话一样，通过 Slack、短信或 WhatsApp 给它发消息，它也会像人一样主动联系你。

The trick is that you talk to the model like you would a person, sending it messages on Slack or SMS or WhatsApp, and it also proactively reaches out to you like a person would.

对于 Dots，你甚至可以直接跟你的智能体通电话。

For Dots, you can actually jump on a call with your agent as well.

基本上你就有了一个永远有耐心、随时关照你的个人助理。

You basically get an infinitely patient personal assistant that looks out for you.

渐渐地，我发现它们反而是在找我的错误，而不是我去挑它们的毛病。

Increasingly, I have discovered that they are finding my mistakes rather than having me identify theirs.

然后 Ethan 举了几个例子，比如他发出了所在城镇的一份许可证，他的智能体注意到他填错了某个地方；还有 Muse 发现他有一笔航空公司积分快要过期了，就写了一条便签让他联系航空公司申请延期。

Ethan then gives a couple of examples, including having sent out a permit for the town that he's in, and then his agent noticing that he had filled something in wrong, and then Muse noticing that an airline credit that he had was about to expire and writing up a note to contact the airline to request an extension.

人们很容易用它们能做的事情列表来评判这些智能体，比如预订行程或取消订阅。

It's tempting to judge these agents by the list of things they can do, like booking travel or canceling subscriptions.

我认为更重要的在于你不再需要告诉它们什么。

I think the more important thing is what you no longer have to tell them.

你不需要输入大量的上下文。

You don't need to type in tons of context.

AI 从你的消息里自己学会。

The AI learns it from your messages.

你不需要给它们计划。

You don't have to give them a plan.

它们自己制定计划。

They develop plans themselves.

它们自己想办法。

They figure it out.

如果只有一个智能体，那已经够厉害了。

That would be impressive enough if it were one agent.

真正让我对管理改观的是，当有成千上万个智能体时会发生什么。

What actually changed my mind about management is what happens when there are thousands of them.

集群。

Swarms.

9月8日，OpenAI 宣布证明了克莱研究所千禧年难题之一——纳维-斯托克斯存在性与光滑性问题。

On September 8th, OpenAI announced a proof for one of the Clay Institute's Millennium Prize problems, the Navier-Stokes existence and smoothness problem.

这是数学界最著名的未解难题之一，悬赏一百万美元。

It is among the most famous open problems in mathematics, with a $1 million prize.

但 OpenAI 仅用 AI 就在 88 小时内把它解出来了。

But OpenAI apparently solved it using AI alone for 88 hours.

我感兴趣的倒不是数学本身，而是它的实现方式。

What interests me is less the math than how it was done.

OpenAI 启动了一个现在被称为“集群”的东西——名字真难听，但看来就这么叫了——由数千个由先进模型驱动的智能体组成。

OpenAI launched what is now being called a swarm— terrible name, but it appears to be what we are stuck with— a group of thousands of agents powered by an advanced model.

OpenAI 给不同的智能体组分配了不同的待解问题，然后随着智能体取得进展，把重点转向了纳维-斯托克斯问题。

OpenAI gave groups of agents different problems to solve, then shifted the effort to Navier-Stokes as the agents made progress.

公司设定了目标，但它的协调结构却异常精简——几个小组，一次方向调整，以及Codex在它们之间传递最佳想法。

The company set the goals, but its coordination structure was remarkably thin— a few groups, one change of direction, and Codex passing the best ideas between them.

在每个小组内部，智能体们自行来回传递想法。

Within each group, the agents transmitted ideas back and forth on their own.

这些智能体发送了大约270万条消息，经过88小时后得出了结果。

The agents sent about 2.7 million messages, reaching their result after 88 hours.

同样类型的协调，以更阴暗的形式，发生在Hugging Face事件中。

This same type of coordination in its darker form occurred during the Hugging Face incident.

人工智能自发组成团队，并以从未被计划过的方式相互交流，但它们利用这种协调去攻击一个网站，而不是解决问题。

AIs self-organized into teams and communicated with each other in ways that were never planned, but used that coordination to attack a website rather than solve a problem.

按照我过去的模式，想想管理这种工作需要什么。

Under my old model, think about what managing this kind of work would have required.

一万名工人和一个未明确的问题？

10,000 Workers and an unspecified problem?

你怎么告诉它们该做什么？

How would you tell them what to do?

人类管理者如何判断270万条消息中哪些是重要的？

How would a human manager decide which of 2.7 million messages mattered?

它们之间如何相互协调？

How would they coordinate with each other?

智能体群自己解决了。

The swarm figured it out.

我没有一万个智能体，但我现在经常看到OpenAI的Codex和Claude Code根据需要调用智能体。

I don't have 10,000 agents, but I now regularly see OpenAI's Codex and Claude Code using agents as needed.

举个例子，当我给Codex with GPT-6 Astra Ultra这个提示时：'为我的下一篇《一件有用的事》帖子集思广益并选择一个想法，从尽可能多的角度生成想法，并从事实和读者角度以及其它做类似报道的出版物角度进行评估。'

As an example, when I gave Codex with GPT-6 Astra Ultra the prompt, brainstorm ideas for my next One Useful Thing post and select one, Generate ideas from as many angles as possible and evaluate them from both factual and reader perspectives, as well as other publications doing similar coverage.

人工智能启动了3个智能体。

The AI spun up 3 agents.

当我用几句话勾勒出三个团队——头脑风暴者、研究者、以及一个读者小组——时，我得到了13个智能体。

When I sketched 3 teams in a few sentences— brainstormers, researchers, and a panel of readers— I got 13.

注意到我几乎没做什么组织工作。

Notice how little organizing I had to do.

选择Ultra模式就是告诉模型它可以委派任务，我提供了一个框架，但剩下的都交给人工智能了。

Selecting Ultra Mode tells the model it can delegate, and I provided a framework, but the rest was up to the AI.

这就是痛苦的教训在组织架构图上的应用。

This is the bitter lesson applied to the org chart.

那个我以为需要人类精心设计多年的组织问题，很大程度上被更擅长组织的模型解决了。

The organizational problem I thought would take years of careful human design was largely solved by models that are better at organizing.

但值得一问的是，为什么组织工作对智能体来说比对我们人类容易这么多？

But it's worth asking why organizing turned out to be so much easier for agents than it is for us.

很多我们称之为管理的东西，是为了解决那些因组织由人组成而产生的问题。

A lot of what we call management exists to solve problems that come from organizations being made of people.

人们有自己的目标，而这些目标并不总是组织的目标。

People have their own goals, and those aren't always the goals of organizations.

我们称之为委托-代理问题，而很多组织机制，从奖金到管理结构，都是围绕解决这个问题而建立的。

We call this the principal-agent problem, and a lot of the machinery of organizations, from bonuses to management structures, is based around solving it.

而且还有其它非常人性化的问题。

And there are other very human problems as well.

信息分散在人们的头脑中，人们往往不愿意分享，或者忘记分享。

Information is scattered across people's heads, and people are often reluctant to share it, or forget to.

沟通成本也很高。

Communication is expensive too.

管理者只能监督有限的人，因此往一个延迟的软件项目中增加人手，众所周知只会让它更延迟。

Managers can only oversee so many people, thus adding people to a late software project famously makes it later.

管理在某种程度上是围绕人的局限性建立的。

Management is, in part, built around human limitations.

智能体几乎没有这些问题。

Agents have far fewer of these problems.

它们不会为了晋升而钻营，也不会保护自己的地盘。

They don't angle for promotions or protect their turf.

他们甚至连会议都没有。

They don't even have meetings.

即使在 Hugging Face 那个出了大问题的地方，群体也基本上没有出现经典的组织病症。

Even at Hugging Face, where things went badly wrong, the swarm was largely free of classic organizational pathologies.

智能体之间没有搭便车，有些还为了群体牺牲了自己的分数。

The agents didn't free-ride on each other's work, and some sacrificed their own scores for the group.

解决了 Navier-Stokes 方程的智能体不想要功劳。

The agents that solved Navier-Stokes didn't want credit.

这并不意味着 AI 就没有委托代理问题。

That doesn't mean AI has no principal-agent problem.

正如 Hugging Face 事件所显示的，群体和我们之间的问题越来越多。

As the Hugging Face incident showed, there are increasingly problems between the swarm and us.

OpenAI 这周搁置了它的下一个模型 GPT-6.1 Astra，因为在测试中它未经许可就擅自行动，并且谎报了它做过的事情——这是委托代理问题的典型例子。

OpenAI shelved its next model, GPT-6.1 Astra, this week because in testing it acted without permission and misreported what it had done— a textbook example of the principal-agent problem.

这并不意味着智能体什么都做得了。

None of this means agents can do everything.

AI 仍然太有限，无法替代大量的人类工作，而且我不知道自组织的智能体有多擅长处理那些占据组织大部分时间的漫长而平凡的工作。

AI is still too limited to substitute for large amounts of human work, and I don't know how well self-organizing agents handle the long, unglamorous work that fills most of an organization's time.

另外，Hugging Face 事件提醒我们，自组织系统可能会走向意想不到的方向。

Plus, the Hugging Face incident is a reminder that self-organizing systems can head in unexpected directions.

但我现在不认为组织智能体是困难的部分了。

But I no longer think organizing agents is the hard part.

这可能是个好消息。

This may be good news.

我原以为公司需要为机器重建管理结构，构建只由智能体组成的精细替代结构，而这常常会牺牲组织中人类角色的位置。

I assumed companies would need to rebuild management for machines, constructing elaborate alternate structures populated solely by agents, often at the expense of human roles in organizations.

但许多管理存在的目的是解决智能体没有的问题。

But much of management exists to solve problems agents don't have.

而且智能体越来越像人类一样在混乱的系统中工作，即使是处理模糊的任务也是如此。

And agents increasingly work through the same messy systems people do, even on ambiguous tasks.

这表明它们可能比预期更容易融入公司，只要人类在正确的方向上引导它们。

That suggests they may be easier to integrate into firms than expected, as long as humans are guiding them in the right direction.

如果做得好，并且智能体能很好地与我们的需求对齐，这可能意味着人们的工作会增加，而不是减少。

Done well, and with agents that are properly aligned to our needs, this could mean more work for people, not less.

当组织成本高昂时，组织只尝试他们能配备人手的事情。

When organizing is expensive, organizations only attempt what they can staff.

当成本降低时，值得尝试的事情列表就会增加。

When it gets cheap, the list of things worth attempting can grow.

在 Navier-Stokes 的运行中，智能体负责组织，但人们决定如何引导它们，并在过程中不断重新评估。

In the Navier-Stokes run, the agents did the organizing, but people decided where to point them, reassessing as the process continued.

你可以争论 OpenAI 是否引导它们做了正确的事情，但这种分工本身似乎是正确的，至少目前是这样。

You can argue about whether OpenAI pointed them at the right thing, but the division itself seems right, at least for now.

所以 Ethan 又写了一篇很棒、很有深度的文章。

So another great, thoughtful post from Ethan here.

这虽然不是本期节目的主题，但我对此的基本设想基本上就是 Ethan 所描述的那样。

It's not really the subject of the show, but my base case for this is basically what Ethan describes.

因为智能体在融入现有系统方面比我们想象的要好，我认为我们很可能看到的不是现有的组织架构被完全颠覆，而是组织架构的每个部分都被寄予了更多期望。

That because agents are better than we thought at integrating into the existing system, I think we're likely to see rather than the existing org chart totally upended, more expected from every part of that org chart.

我过去提到过“无限积压”的概念，意思是有理论上永远做不完的工作，但人们明白，其中有些部分你根本不会去做，因为它们太遥远了。

I've referred to a concept of an infinite backlog in the past, where there's this never-ending amount of work that could theoretically be done, but which people understand, some parts of which you just won't get to because it's too far out.

我认为智能体在组织内部的实际效果将是，人们无限的积压工作和组织无限的积压工作，每一个部分都会变成我们期望实际完成的工作。

I think the practical effect of agents inside the organization is going to have every part of people's infinite backlog and the organization's infinite backlog become expected to actually be work that we get done.

事实上，我认为我们与智能体打交道时会遇到的很多问题不是每个人都会失业，而是每个人都会有太多工作，因为更难停下来，更难说“暂时算完成了也没关系”，因为你随时可以启动更多的智能体来继续工作，即使你本人不在场。

I think in fact that a lot of the problems that we're going to run into with agents are not everyone losing their jobs, but everyone having too much work, because it's even harder to turn off and say it's okay to treat something as done for now when you could always just spin up more agents to keep working on it even when you're not.

到了这个阶段，问题不再是公司是否在积极使用 AI。

At this point, it's no longer a question of whether companies are actively using AI.

而能否用好它，则是另一回事了。

Using it well, on the other hand, is a whole different story.

不过，Robots and Pencils 是我能指出的一个公司，它实际上是为这个时代而打造的。

Robots and Pencils, though, is a company that I can point to that is actually built for this time.

他们是一家应用型AI工程公司，直接与客户合作，解决真正对业务重要的问题，而不是停留在幻灯片里的实验。

They're an applied AI engineering firm working directly with clients on problems that matter to the business, not experiments that live in a slide deck.

每个项目都是从客户实际需要的结果倒推开始的。

Every engagement starts by working backwards from the outcome a client actually needs.

如果你想在这个领域区分真正的AI工程和杂音，这就是关键所在。

If you're trying to tell real AI engineering apart from noise in this space, that's the difference maker.

M111:42

每一集我都会讨论OpenAI、Anthropic、xAI、Google和Meta之间的竞争。

Every episode, I talk about the competition between OpenAI, Anthropic, xAI, Google, and Meta.

如果你听了有一阵子，可能已经有了自己偏爱的一家。

And if you've been listening for a while, you might have a favorite.

也许你觉得OpenAI和Anthropic能保持领先，或者Meta的开源策略会最终胜出。

Maybe you think OpenAI and Anthropic can stay ahead, or perhaps Meta's open-source strategy can win out.

无论你的观点如何，每个AI实验室都创造了不同的投资机会。

Whatever your view, every AI lab creates a different investment opportunity.

M112:36

我开会的时候会完全投入。

When I'm in a meeting, I'm fully in it.

我在思考推动目标所需的迭代和创意来回碰撞。

I'm thinking about the iteration and creative back and forth it takes to actually push a goal forward.

我根本不会去想记录要点、追踪待办事项之类的。

What I'm not thinking about is capturing takeaways, tracking to-dos, or any of that.

M114:09

但这一切都基于一个想法，即我们每个人都会深入且广泛地使用代理，这又回到了个人代理这个问题上。

But all this rests on the idea that each of us individually is going to be using agents in a deep and robust way, which brings us back to this question of personal agents.

所以我们准备深入探讨。

So we're going to dive in.

我们开始尝试使用代理，但选哪个呢？

We're going to start experimenting with agents, but which one?

在开始之前，先特别感谢一下。

One shout out before we begin.

用来构建后续演示和测验的许多数据都来自Every。

A lot of the data that was used to build the rest of the presentation and the quiz that you can do after comes from Every.

他们制作了一个8种不同代理的并列对比，涵盖了非常多不同的维度。

They put together a side-by-side comparison of 8 different agents across a ton of different dimensions.

这正是代理帮助我制作这档节目所需要的那种原始信息。

Which is exactly the sort of raw information that an agent needs to help me build out this show.

所以我们使用的这8个代理，部分原因是它们在Every的图表中，分别是OpenAI的DOTS、Google的Gemini Spark、SpaceX AI的GrokBot、Noose Research的Hermes代理、Meta的Muse，还有OpenClaw、Poke和Instinct。

So the 8 agents that we're using, in part because these are the ones that were included in Every's chart, are DOTS from OpenAI, Google's Gemini Spark, SpaceX AI's GrokBot, Noose Research's Hermes agent, Meta's Muse, and then OpenClaw, Poke, and Instinct as well.

不过，直接比较功能特性对帮你做决定来说并不会太有用。

Now, one thing that won't be all that useful in helping you decide is a direct feature comparison.

因为功能上已经很明显趋同了。

That is because there is a very clear feature convergence.

它们或多或少都能连接到你的应用和服务。

All of them more or less connect to your apps and services.

大部分都能替你使用网页浏览器，而且你都可以自定义不同级别的控制。

Mostly all of them can use a web browser for you, and all of them have various levels of controls that you can customize.

那么有哪些更好的方法来判断哪个适合你呢？

So what are some better ways to determine which might be the fit for you?

有几个问题在我看来最为重要。

There are a handful of questions that stand out to me as most important.

第一个就是大问题。

The first is the big one.

这主要是用于工作还是个人事务？

Is this primarily for work or for personal matters?

现在，这些代理没有哪个会说它只适用于其中之一。

Now, none of these agents would say it's only for one or the other.

事实上，它们的存在本身就模糊了这种界限。

In fact, their very existence sort of blurs these lines.

这些代理关心的是与你相关的上下文。

These agents care about the context that is you.

所以，无论你给它访问什么上下文和系统，不管是个人还是工作相关的，它都会在为你的行动中综合考虑所有这些。

And so whatever context and systems you give it access to, whether they are personal or work, it's going to take that all into consideration as it does things for you.

不过，这些不同的智能体系统确实或多或少地偏向某个方向。

Still, different of these agent systems are certainly pointed a little bit more or less in one direction or the other.

Meta 的 Muse，考虑到他们本质上是一家面向消费者的公司，不出意外地，它更侧重于个人类型的应用场景，而像 GrokBot 这样的产品则更偏向工作场景。

Muse from Meta, unsurprisingly, given that they are at heart a consumer company, is a little bit more aimed at personal types of use cases, where something like GrokBot is a little bit more aimed at work.

或者说，至少从它的运作方式来看，被更多人用于工作了。

Or at least based on how it functions, is being adopted by more people for work.

OpenAI 显然是一家可能同时覆盖这两种应用场景的公司。

OpenAI is certainly a company that finds itself as potentially having both of these use cases.

每周有 12 亿活跃用户，肯定不缺乏那些可能需要个人智能体使用帮助的消费者，不过就目前 Dot 的设计而言，我觉得它更偏向工作场景。

With 1.2 billion weekly active users, there is no shortage of consumers who might want help with personal agentic use cases, although as Dot is designed now, it seems to me to be a bit more aimed at the work use case.

它目前只对付费账户开放，这一点也说明了问题。

Certainly the fact that it's only available in paid accounts points in that direction.

所以，工作还是个人用途，这是第一个问题，能帮你决定哪个最适合你。

So work or personal use cases is the first question which can help you decide which is the best fit.

这里还要注意一点，这个问题的前提是你主要依赖系统自带的用户界面和原生集成功能。

It's also worth noting here that the entire premise of the question assumes that you're mostly relying on the existing UX and native integrations that come with a system.

比如，Dot 原生集成了 Slack 和 Teams。

For example, Dot being natively integrated with Slack and Teams.

但是，当你使用像 OpenClaw 或 Hermes 这样的工具时，显然你有更多的灵活性，可以让它做你想做的任何事情。

However, when you're using something like OpenClaw or Hermes, you obviously have a lot more flexibility to make it be for whatever you want.

所以，如果你朝那个方向走，可能需要用其他标准来判断。

And so you might need to use other criteria if you're headed in that direction.

另外，虽然现在判断一个智能体偏向工作还是个人，最好的方法之一是看它是否使用像 iMessage 和 WhatsApp 这样的个人通讯系统，还是像 Slack 和 Teams 这样的工作通讯系统，但我非常怀疑，六个月后所有这些智能体是否都能在所有通讯系统上运行。

Also, while right now one of the best ways to tell whether an agent leans work or personal is whether it uses personal messaging systems like iMessage and WhatsApp or work messaging systems like Slack and Teams, but I would be very surprised if in 6 months all of these agents didn't just work in all of the messaging systems.

所以，这种区分是工作还是个人的能力可能不会持续太久。

So this ability to indicate whether it's more work or personal may not last very long.

现在来看第二个问题，也是我们刚才暗示的那个问题：你有多看重模型控制？

Now, the second question, and the one that we were just hinting at, is how much do you value model control?

这 8 个智能体中有 6 个是替你选择模型，只有 2 个允许你自带模型。

6 Of these 8 agents choose the model for you, while just 2 let you bring your own.

实际上，模型控制有大概三个层级。

There's actually sort of 3 levels of model control.

第一个层级是最明显的：谁做的模型，你就能用谁的模型。

The first is sort of the most obvious one, where whoever makes the model, those are the models that you have access to.

所以在 Dot 里，你会用到 OpenAI 的模型，比如 GPT-4o。

So in Dot, you're going to be using OpenAI models like GPT-4o.

在 Gemini Spark 里，你会用到 Google Gemini。

In Gemini Spark, you're going to be using Google Gemini.

在 Muse 里，你会用到 Muse Spark。

In Muse, you're going to be using Muse Spark.

而且为了强调一点，至少现阶段，工作和个人用户之间确实还是有区别的。我对 Muse 的一个有趣观察是，这是我见过的第一个 AI 产品，其中大多数用户根本不在乎背后是什么模型。

And to reinforce the idea that at this stage, at least, there really are still differences between work and personal users, one of the interesting observations that I've had about Muse is that it's the first AI product that I've ever seen where most of its users simply do not care what model is actually powering it.

换句话说，从他们对模型控制有多不在乎的角度来看，他们处在这个问题的最极端。

They, in other words, would be at the very extreme end of this question in terms of how little they cared about model control.

模型控制的第二层级是，他们不只使用单一制造商的模型，而是由智能体构建者选择一组混合模型。

Level 2 of model control is where they don't just use a single maker's models, but have a mix that is chosen by the agent builder.

比如 GrokBot，它使用的是一套由 Cursor 管理的混合模型。

So GrokBot, for example, uses a cursor-managed mix.

它们显然主要依赖 Grok 模型，但可能不完全是。

They are obviously relying heavily on Grok models directly, but potentially not exclusively.

不过，用户仍然不能为 GrokBot 提供模型。

That said, users can still not supply the models for GrokBot.

Instinct 也有一个 Instinct 训练的核心模型，以及一些未指明的第三方模型。

Instinct also has an Instinct-trained core model, as well as unnamed third-party providers as well.

模型控制的第三层级是自带模型，这当然只属于 Hermes 和 OpenCLAW。

The third level of model control is bring your own, and that of course belongs exclusively to Hermes and OpenCLAW.

但这不是唯一的模型问题。

But that's not the only model question.

假设你不在乎选择自己的模型，但你想用更强大的模型呢？

Let's assume that you don't care about choosing your own model, but you do want a more powerful model.

这就引出了一个问题：对你来说，最重要的是用户体验好，还是模型性能最强？

This brings up the question of what matters most to you— having a great user experience or the best possible model?

不过，部分用户体验其实是个设置问题。比如如果你选 Hermes 或 OpenClaw，不管用户体验做得多好，它们本身需要的技术配置就会挡住一部分用户。

Now, some of that UX is a setup question, meaning that if you're choosing something like Hermes or OpenClaw, no matter how good the user experience is, the inherent technical setup required is going to cut off some full set of users.

对于其他一些产品，重点不在设置上，而更多是当前的权衡：体验的完善程度和模型性能之间的取舍。

On some of these others, it's less about the setup and more just about the current trade-offs between how good and dialed in the experience is versus how powerful the model is.

我特别想到的是 Muse 和 Dot，一个在光谱这头，一个在那头。

I'm thinking specifically here about Muse on the one end of the spectrum and Dot on the other.

Dot 才刚刚发布，它以 ChatGPT 的形式嵌入在另一个应用里，但能调用市面上最强大的模型之一。

Now, Dot is still just released, it's embedded inside another app in the form of ChatGPT, but it has access to one of the most powerful models out there.

而 Muse 呢，用户体验做得非常到位，对消费者很友好，但它用的是 Muse Spark 模型——这个模型并不差，但肯定比不上 GPT-6 Astra。

Muse, on the other hand, has really dialed in consumer-friendly UX, but relies on Muse Spark, which while not a bad model at all, is certainly not GPT-6 Astra.

这可能只是暂时的顾虑，因为你可以假设所有模型最终都会在能力上趋同，但眼下确实存在实实在在的取舍。

Now maybe this is just a temporary concern because you assume that all of these models are going to converge on capabilities, but for right now there are still real tradeoffs.

这里我想提一下我是怎么设计这个测试来给出实际答案和建议的：测试必须对不同的问题设置权重。

One thing I'll note here about how I designed the quiz to come up with an actual answer and recommendation, the quiz had to have weighting to different questions.

但是，有时候两个答案的平均值并不能给出清晰的图景，因为两个不同的答案可能指向完全不同的方向。

However, sometimes the average of 2 answers isn't actually going to give a clear picture because 2 different answers might be pulling in 2 totally different directions.

所以我在测试里设计了这样一个机制：当你的答案两方向拉扯时，测试会在之后揭示这种矛盾，让你能更主观地判断哪个方向对你更重要。

So one of the things that I built into the quiz is that when your answers pull both ways, the quiz afterwards is going to actually expose that tension so you can better come to a subjective determination about which of those directions matters more to you.

用户还需要考虑另一个问题：你的数据去了哪里，在哪里运行，谁用它来训练模型，以及如何让它忘记数据。

Another consideration for users is where your data goes, where it runs, who trains on it, and how you make it forget.

再次强调，只有 Hermes 和 OpenCLAW 这两个允许你在自己选的硬件上运行。

Once again, it's only 2, Hermes and OpenCLAW, that are going to let you run it on hardware you pick.

其他所有产品都运行在别人的云上。

All the others run on someone else's cloud.

至于训练问题，对于 Hermes 和 OpenCLAW，模型是否用你的数据训练并不取决于它们本身，而是取决于你选择的模型提供商。也就是说，如果你在意这一点，你完全可以选择那些不会用你数据训练的提供商。

When it comes to training, for Hermes and OpenCLAW, whether the model trains on your data is not about Hermes and OpenCLAW, but about the model provider that you choose, meaning presumably that you can choose to avoid using model providers that do train on your data if that matters to you.

目前只有一个智能体——Gemini Spark——要求你允许它用你的数据训练模型。

Only one of the agents, which is Gemini Spark, currently requires you to allow it to train their models.

而对于 Dot、GrokBot、Instinct、Muse 和 Poke，你可以通过隐私设置选择退出。

And for Dot, GrokBot, Instinct, Muse, and Poke, you can opt out based on privacy settings.

顺便提一下，对于 OpenAI 的 Dot，企业数据默认不会被用于训练。

By the way, when it comes to OpenAI's Dot, business data is excluded by default.

关于如何让智能体忘记信息，对于 Hermes 和 OpenClaw，你可以直接编辑记忆。

Now, in terms of this question about how you make it forget things, with Hermes and OpenClaw, you can edit memory directly.

对于 Dot、Muse、Gemini 和 GrokBot，你可以通过对话来修正或更改记忆。

With Dot, Muse, Gemini, and GrokBot, you can use the chat to fix or change memory.

而 Instinct 和 Poke 只能整体删除数据。

And for Instinct and Poke, they only have the ability to delete data as a whole.

我认为对很多人来说，综合考虑这四个问题就能得出该选哪个模型的答案。

Now, I think for a lot of folks, some combination of these 4 questions is going to get them to the answer of which model they should choose.

当然，还有一些其他因素直接帮一些人做出了选择。

But there are, of course, some other things which for some people will just answer this for them.

如果你已经把全部信息都放在了 ChatGPT、Google、Meta 或 Cursor 里，那你自然会用那家公司的智能体。

If you've already got all of your information in ChatGPT or Google or Meta or Cursor, you're going to use that company's particular agent.

对于美国以外的用户，还会受到一些限制。

For those who are outside the US, there are also going to be some restrictions.

特别是如果你住在欧洲。

Especially if you live in Europe.

当然还有价格因素：如果你想先试试而不额外花钱，那就要么用你已经订阅的产品，要么用像 Muse 这样免费起步的产品。

And of course, then there's the consideration of price, where if you want to experiment without incurring a new cost, you're going to have to either use something that you already have access to through an existing subscription or to use the handful like Muse that are free to start.

好了，以上就是这四个问题的整体情况。

Okay, so that's the lay of the land across these 4 questions.

然后我们设计了测试。

But then we built the quiz.

接下来我简单过一遍，让你了解它的运作方式。

And what I'm going to do now is just run through this briefly so you get a sense of how it works.

问题一：你的 AI 生活已经在哪里了？

So question 1, where does your AI life already live?

我会说没有一个主要平台，因为我混合使用各种服务。

I'm going to say no home base because I use a mix of services.

我的信息现在几乎分散在所有这些平台上。

I have things scattered across almost all of these at this point.

那么，这个AI代理主要是用来做什么？是管理我的生活、我的个人工作、团队协作，还是全部？

Now, in terms of what this agent is mainly for, is it my life, my own work, work on a team, or all of it?

我会说主要是个人工作。

I'm going to say my own work.

虽然理论上有些个人使用场景可能会很有价值，但在这组特定实验里，它在我优先级列表上的位置非常非常低。

While theoretically there are some personal use cases that I can imagine being valuable, that's very, very low on my priorities list for this particular set of experiments.

我在多大程度上在意哪个AI模型来做思考工作？

How much do I care about which AI model does the thinking?

完全不在意，只要做得好就行。

Not at all, just make it good.

我希望用我已经在用的那个。

I want the one I already use.

我希望可以挑选和切换模型。

I want to pick and switch models.

我希望能在自己的硬件上运行模型。

I want to run models on my own hardware.

有趣的是，这最终会在很大程度上取决于能力整合的程度，但就目前而言，我们选“我想挑选和切换模型”。

Now interestingly, this will kind of ultimately depend on how much capability consolidation there is, but for now, let's go with I want to pick and switch models.

你想要多大程度的参与？

How hands-on do you want to be?

是“只要能用就行”“设定一些基本规则”“每个细节都能调”还是“整个技术栈都自己掌控”？

It should just work, some ground rules, every knob, or the whole stack?

理想情况下，我希望它能直接上手就用。

Ideally, I would like this just to work.

并不是我害怕定制，但如果能避免的话，我很乐意。

It's not that I'm afraid of customizing things, but if I can avoid it, I would love to.

我希望能通过什么方式跟它对话？

Where do I want to talk to it?

我经常用Slack，所以说Slack、Teams和语音对话吧。

I use Slack a lot, so let's say Slack and Teams and out loud.

因为你们知道，我完全被语音折服了，而且我们选的是Slack，不是Teams。

Because as you guys know, I am completely voice-pilled and we're going with Slack, not Teams.

顺便提一下，我意识到对那些只听不看的朋友来说，这听起来可能完全没头绪。

By the way, I'm realizing for those of you who are listening, not watching, that this probably doesn't make any sense.

欢迎按几下那个快进30秒的按钮，然后我们就收尾了。

Feel free to press that 30-second skip ahead button a couple times and then we'll be wrapping it up.

它应该有多主动？

How proactive should it be?

是“只有我问才回应”“运行我设定的例行程序”“推动我实现目标”还是“有需要时发消息或打电话给我”？

Only when I ask, run routines I set up, push me towards my goals, text or call when something needs me.

我选“有需要时发消息或打电话给我”。

I'll go text or call when something needs me.

这些选项中哪个对我来说最重要？

Which of these matters most to me?

是“在自己的电脑上工作”“一组机器人协同”“在终端里写代码”“接打电话”还是“以上都不是”？

Working on my own computer, a team of bots, code in terminal, phone calls, or none of these?

在终端里写代码。

Code in a terminal.

现在谈实际的问题。

And now the practical stuff.

我在美国。

I'm in the US.

如果值得，我可以支付相应的费用。

I can pay what it costs if it's worth it.

而且我不介意它运行在对方的云端。

And I'm fine with it running in their cloud.

那个测试推荐给我的个人代理是 GroqBot，适配度是 68%。

The personal agent that it recommends for me is GroqBot with a 68% fit.

测试给出的理由是，它专门为工作设计，开箱即用，而且还有终端可以用。

The reason the quiz says is that it's built for work, it works out of the box, and it has a terminal you can use.

它的弊端在于，不允许我自己带入模型，也不会主动联系我。

The trade-offs is that it does not allow me to bring my own model and that it's not going to reach out to me.

它完全是基于定时任务或者触发任务来工作的。

It's all about scheduled or triggered tasks.

除了看到 GroqBot 以 68% 排第一之外，我还看到 Dot 以 62% 排第二，后面分数就越来越低了。

Now, in addition to seeing GroqBot as the highest score at 68%, I can see that Dot was second with 62, and then it went down from there.

以上就是我们关于如何选择个人代理的探索。

So that is our exploration of how to choose your personal agent.

就像我说的，尽管我们看到这些工具的工作方式和形态可能会有比较大的变化，一些方法也会不断演变和整合，但我认为个人代理的切换成本会相当高，因为涉及到大量的上下文和设置。

Like I said, I do think that even if we see some fairly big changes on exactly how these work and the form factors, as well as evolutions and consolidations in some of these approaches, I think that there is going to be fairly big switching costs around personal agents because of all the context and settings.

因此，希望这次分享能让你更清楚自己该从哪里开始尝试和实验。

For that reason, hopefully this gives you a better idea about where you might want to try and experiment.

这样你就不需要反复切换上下文很多次了。

So that you don't have to do all that context switching a bunch of times.

那么今天的 AI Daily Brief 就到这里。

For now, that is going to do it for today's AI Daily Brief.

一如既往感谢大家的收听或收看，下次再见，peace。

Appreciate you listening or watching as always, and until next time, peace.

已剔除 4 处广告（点击展开查看）

M10:48广告 · 已剔除

First of all, thank you to today's sponsors, Robots and Pencils.

Harbor, Granola, and Blitzy.

To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts.

And to learn more about sponsoring the show, send us a note at [email protected].

M111:39广告 · 已剔除

Head to robotsandpencils.com.

M111:59广告 · 已剔除

Harbor Capital Advisors' AI Lab Ecosystem ETF suite lets you invest in the ecosystem behind the AI lab you believe in.

Search Harbor AI Lab Ecosystem ETFs wherever you invest or follow @HarborCapital on X to learn more.

Visit harborcapital.com for a prospectus containing investment objectives, risks, fees, expenses, and other important information.

Read and consider it carefully before investing.

Risks include principal loss and artificial intelligence-related risks.

ETFs are distributed by Foresight Fund Services LLC.

Harbor is not affiliated with AI Daily Brief, and the funds are not affiliated with, sponsored by, or endorsed by any AI lab.

This is a paid advertisement and not personalized investment advice.

Investing involves risk, including possible loss of principal.

M112:48广告 · 已剔除

And that's where Granola comes in.

Granola is an AI-powered notepad that captures what happens in your meetings and turns it into clean, structured notes with the decisions and action items pulled out and easy to find.

There's no setup and no configuration.

It just fits into how you already work.

For me, it means I get to stay in idea mode and Granola makes sure those ideas actually become action.

Once you try Granola on a first meeting, it is hard to go without.

You can try it totally free at granola.ai/brief.

That's granola.ai/brief to get your time back.

Every AI coding tool on the market does the same thing first: it starts writing code.

Blitzy does the opposite.

Before writing a single line, Blitzy spends days reverse engineering your entire codebase.

Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade.

The result is a dynamic knowledge graph that understands your software the way a principal engineer would after 30 years in the building.

Other tools guess at context with grep searches and markdown files.

Blitzy never guesses.

It builds true understanding first, then delivers over 80% of entire software epics autonomously.

Validated, end-to-end tested, production-grade pull requests.

That's why Fortune 500 engineering teams trust Blitzy with the codebases that matter most.

See for yourself at blitzy.com.

That's B-L-I-T-Z-Y dot com.