Meta Muse、Grok Bot、OpenAI Dots：你应该使用哪个 AI 智能体？

过去三年里，AI一直是我们对话的对象。

For the last three years AI has been something that we talk to.

我们输入文字，它回复文字，然后我们拿着这些信息去做有用的事情。

We would type to it and then it types something back to us and then we would take that information and go do something useful with it.

但在过去八周里，这种情况发生了巨大变化，变化的形式就是所谓的AI代理。

But over the last eight weeks that has changed pretty dramatically and it comes in the form of this thing called an AI agent.

这个词虽然出现在很多新闻标题里，但没人真正搞清楚AI代理是什么。

Now this is a term that has been floated across many different headlines but no one quite understands what an AI agent is.

最好的理解方式是，它是一个能替你办事的AI模型。

The best way to think about it is it's an AI model that has the means to do things for you.

就像给AI模型装上胳膊和腿，它能到处跑，帮你完成各种任务。

It's like giving an AI model arms and legs and it's able to kind of run around and do various different tasks for you.

比如下载或订购你列在清单上很久的书，方便你阅读；帮你报税；或者在聊天里组织朋友间的聚餐。

Examples could be download or order that book that's been on your list for ages so that you can read it, filing your tax forms or organizing dinners between friends in your text chat.

AI代理能做的事情远不止这些。

And there's so much more that these AI agents can do.

所以你可能会以为最受欢迎的AI智能体来自Anthropic，甚至可能是OpenAI，但事实并非如此。

So you might expect that the most popular AI agents come from the likes of Anthropic and maybe even OpenAI but that's not actually true.

最受欢迎的智能体其实来自Meta，比如MetaMuse，还有埃隆·马斯克的GrokBot。

The most popular agents come from the likes of Meta in the form of MetaMuse, you've got Elon Musk with GrokBot.

OpenAI上周刚发布了Dots，我一直在试用，测试它的效果。

You've got OpenAI that released Dots just last week, I've been playing around with it, testing it out.

最后还有一个叫Instinct的应用，但奇怪的是，它其实并不是一个真正的应用。

And finally, have this app called Instinct, which weirdly isn't actually an app.

它几乎无处不在，只要你接触数字界面，它就在那里。

It just kind of lives everywhere and anywhere that you touch a digital surface.

这个应用是由一个23岁的年轻人创建的，现在价值数十亿美元。

It was created by this 23 year old and it's worth billions of dollars.

所以过去几周我一直在玩这些智能体。

So I've been playing around with these things for the last couple of weeks.

数百万其他人也下载了它们，做了同样的事情。

Millions of other people have downloaded it and been doing the same.

我一直在试图回答两个非常重要的问题。

And I've been trying to answer two very important questions.

第一，这些AI代理之间有什么区别？

One, what are the differences between these AI agents?

我该用哪一个？

Which one should I use?

第二，一开始使用AI代理到底有什么意义？

And two, what's the point of even using an AI agent in the first place?

我是不是应该从使用大语言模型或聊天机器人，真正转向使用AI代理？

Should I materially move from using an LLM or a chatbot?

它真的能为我的生活增添什么吗？

Does it actually add anything to my life?

我真的那么在意生产力吗？

Do I even care about productivity that much?

我为什么要用AI代理？

Why should I use an AI agent?

所以今天这期《In the Loop》节目，我们会深入探讨所有这些问题，还有更多内容。本节目由高通公司赞助播出。

So we're going dig into all of that and more on today's episode of In the Loop brought to you by Qualcomm.

不过在我开始之前，我想先跟所有《无限》的听众们聊聊大家最关心的问题。

But before I do that, I want to address the elephant in the room for all of you limitless listeners out there.

你们可能在想，《In the Loop》到底是什么？eJazz又去哪儿了？

You might be wondering what on earth is In the Loop and where are you eJazz?

好，我们先回答第一个问题。

Well, me start with the first question.

《In the Loop》是这个节目改版后的新名字。

In the Loop is the rebranded title of this show.

很明显，我们少了一位非常重要的搭档主持人。

It is obviously missing a very important co host.

Josh，如果你在看这个节目，我想你。

I miss you, Josh, if you're watching this.

但现在这个节目由我来主持，这算是一次品牌重塑，也是对我们在这里想做的事情的一次全新升级。

But it now currently involves me and it's kind of like a rebrand, a refresh of what we're trying to do here.

你们都知道，我对人工智能和前沿科技这些事儿特别上心。

I'm deeply passionate, as you guys already know, about all the stuff that is happening on AI and Frontier Tech.

这个新节目呢，我们会每周一到周五，一天一期，保证工作日一天都不落下。

And with this new show, we're gonna bring it to you five times a week, Monday to Friday, so we don't skip any weekdays.

所有最新动态，你都能第一时间掌握。

You'll be kept up to date with anything and everything that's going along.

每期大概二十到二十五分钟，短小精悍，你上下班路上听，或者在家做饭的时候听，都特别合适。要是你喜欢听我们的节目，或者喜欢这类内容，那就更好了。

Bite sized chunks, twenty to twenty five minutes so that you can listen to it on your commute to work or if you're cooking at home or whatever, if you just like or take pleasure in listening to our episodes or our kind of content.

对，这其实就是节目改了个名字。

And, yeah, it's just the rebranding of the show, I guess.

所以我不想让大家觉得这是个谜。

And so I don't want that to be a mystery.

如果你还在想这到底是怎么回事，是不是我和Josh闹掰了，我可以放心地告诉你，没有的事儿。

If you're still wondering what on earth has happened, whether Josh and I have broken up, I can happily say that we have not.

Josh是去追求更好的发展了，而且我们昨天还专门出了一期节目聊这事儿。

Josh has moved on to bigger and better things and we actually put out a whole episode about this yesterday.

所以一定要去我们的频道看看。

So definitely go check it out on our channel.

然后回答第二个问题，我现在在一个相当新潮的录音室里，感觉特别棒。

And to answer the second question, I'm in a pretty new swanky studio and it feels pretty awesome.

这比我们刚开始录节目时的卧室可升级多了，那时候大概是在第234期左右吧。

This is a major upgrade from the bedroom that we started recording the show around like two thirty four episodes ago or whatever that might be.

不过我还是带了一点家的感觉过来。

But I did bring a piece of home with me.

这幅画陪了我很久，所以能把它带到这里来真好。

This piece of art has been with me for a while and so it's lovely to bring it here.

所以，新录音室，新节目，我很期待开始。

So yeah, new studio, new show, excited to get going.

好，首先，AI代理到底是什么鬼？

Okay, so firstly, what the hell is even an AI agent?

我用Claude做了一个有趣的小动画来给你解释这个。

And I created this fun little animation with Claude to kind of describe this to you.

通常我们在网上互动时，得一个个去访问不同的网站。

So typically when we're interacting with the web, we have to navigate to individual different websites.

每个应用都得滚动、点按钮，费劲地找产品、服务或者那个按钮在哪，然后点进去，这整个过程挺老套也挺乱的。

Every single app we need to kind of like scroll, navigate through buttons, try and figure out where the product is or where the service is or where that button is, click on it and it's kind of an archaic whole mess.

我想用这个动画展示的是，AI代理能帮你把所有这些整合起来。

What I'm trying to demonstrate with this animation or what AI agents do specifically is it kind of coagulates all of that for you.

它就像成了你的中心枢纽，或者是你跟数字世界其他部分互动的媒介。

It kind of like becomes your center nexus or your medium of interaction with the rest of the digital world.

所以最后就变得没那么乱，也更容易互动，不用再去找网站、下载应用或者找按钮，你直接跟AI代理说，要什么它就神奇地出现或者帮你搞定。

And so it ends up becoming, I guess, a lot less messy and much easier to interact with instead of having to find a website or navigating an app to download it or finding that button, you just speak to an AI agent, ask for what you want and it magically appears or it magically gets done for you.

这个动画里没展示的另一个酷点是，AI代理能直觉地预测你可能想完成什么。

The other cool thing about this that isn't shown on this animation is the AI agent can intuitively predict what you might want to get done.

我觉得这解决了AI的一个大问题，就是99%的人不知道拿AI模型来干嘛。

And I think this solves a very big problem in AI, which is 99% of people don't know what to use their AI model for.

我肯定也属于那类人，看到新模型不断发布，就在想，这能为我做什么呢？

I'm certainly in that category where I see all these new models getting released and I'm wondering, what can this do for me?

有什么真正的新东西？

What's the net new thing?

如果我不会编程，也不是天生的设计师，作为一个普通人，我该怎么用它来让生活更高效，或者只是帮我腾出更多时间？

If I don't code, if I am not a designer at heart, how can I use this as a normal person and create a more productive lifestyle for me or just kind of free up more time for me?

个人代理或者我们即将讨论的这些新AI代理，它们的理念不仅仅是专注于提高生产力。

The idea of a personal agent or these new AI agents that we're about to get into is that it's not just focused on productivity.

有些像GrokBot和OpenAI的Dots。

Some of them are in the likes of GrokBot and OpenAI's Dots.

但当你看看MetaMuse，看看Instinct，它们更像是一种生活方式产品，我觉得很多人会从中受益。

But when you look at MetaMuse, when you look at Instinct, it becomes a sort of lifestyle product and that's something that I think a lot of people would benefit from.

不过这些先不说了，我们直接来看看这些代理的优缺点。

But enough of this, let's get into the actual agents pros and cons.

第一个AI代理恰好也是我目前最喜欢的个人助理。

So the first AI agent also happens to be my current favourite personal assistant.

它是所有应用商店里排名第一的代理，叫做MetaMuse。

It's the number one agent in all app stores and it's called MetaMuse.

这是Meta推出的，Muse是一种非常直观、非常有真人感觉的AI代理。

It's by Meta and Muse is this kind of very intuitive, very human feeling AI agent.

你下载一个应用，设置大概五秒钟就完成了。

You kind of download an app and the setup takes around five seconds.

就像添加你的号码、输入你的名字一样简单，然后你就会遇到一个奇怪又可爱的、像团块一样的角色。

It's as easy as adding your number, stating your name, and then you're introduced to this weird, wonderful looking blob like character.

我，我在这里找一个演示来展示它。

I'm I'm trying to find a demo here that demonstrates it.

哦，对。

Oh, yeah.

在这里。

Over here.

我们看到扎克伯格随意地靠在架子上，然后一个像外星泰迪熊一样的奇怪团块生物迎接我们，它跟你说话非常自然。

Where we got Zuck kind of like casually leaning against the shelf, and we're greeted by this kind of weird blob alien teddy bear looking creature and it talks to you very normally.

我下载这个AI代理时，立刻有几件事让我印象深刻。

Now a few things instantly hit me when I downloaded this AI agent.

第一，设置非常简单；第二，它没有用一堆文字轰炸我。

One, setup was super easy, but two, it wasn't bombarding me with text.

它几乎就是那种简短的形式，比如‘嘿，怎么了？’

It was almost just kind of short form kind of 'hey what's up?

‘我能帮你什么？’

How can I help you?

‘你有什么心事吗？’我喜欢它的一点是，它能主动预测我可能需要什么。

Is there anything that's on your mind?' And what I liked about it is it could proactively kind of predict what I might want.

它能做到这一点的方式之一，就是它能访问我的短信。

Now one way that it was able to do this is it got access to my text.

不过有一点很重要，你要明白，你肯定不希望AI代理随便去给你的朋友或家人发短信，或者做类似的事情。

Now it's very important to point out that, listen, you don't want AI agents randomly going out to text your friends or family or whatever that might be.

MetaMuse做不到那些。

MetaMuse can't do that.

没有你的明确批准，它很多事情都做不了，而且你可以决定在什么时候、什么情况下让它这样做最合适。

It can't do a lot of things without your required approval, and you can determine when and where that makes the most sense.

但是因为它能掌握那个上下文，它知道我在和未婚妻讨论想在亚马逊上买的东西。

But because it was able to have the context of that, it knew that I was discussing with my fiance something that I wanted to buy on Amazon.

然后它来找我，说，嘿，我看到你在找一些东西。

And it came to me and it was like, hey, like, I saw that you were looking for some things.

我其实帮你找到了一个更好的选择。

I actually found a better option for you.

那真的是它发给我的第一条消息，我当时想，好吧，这有点奇怪。

That was literally the first message that it sent me and I thought, okay, that's weird.

感觉有点侵入性，但我愿意多信任它一点，因为它来自像Meta这样信誉良好的公司。

It kind of feels a little bit invasive, but I'm willing to trust this a little bit more because it comes from a reputable firm such as Meta.

我知道Meta过去有过隐私问题，但扎克伯格特别强调了一点，尤其是在Meta Connect上宣布这个产品时我给你看的这段视频里，那就是隐私是这个整个产品的核心。

Now I get that Meta has had privacy issues in the past and one thing that Zuck has made very clear, especially in this clip that I'm showing you here when they announced this at Meta Connect, is that privacy is at the core of this entire product.

我一会儿会具体讲它适用于哪些地方。

And I'll get into where exactly that applies in a second.

但首先，我们来看看这个AI代理背后的实际数据。

But first, the actual stats behind this AI agent.

它的下载量已经……

It's been downloaded...

我觉得这个数字其实已经过时了。

I think this figure is actually outdated.

它的下载量已经超过了350万次。

It's been downloaded over 3,500,000 times.

而且就在两天前，我看到一个数据，基本上是说他们有200万日活用户，这挺惊人的。

And as of two days ago, I saw a stat that basically said that they had 2,000,000 daily active users, which is pretty insane.

当然，Meta的用户规模比这大得多。

Now granted Meta's user magnitude is much larger than that.

我觉得他们在Instagram、WhatsApp和Facebook上大概有35亿用户。

I think they have about 3,500,000,000 users across Instagram, WhatsApp, and Facebook.

但能在这样一个全新的AI产品上获得这么大的使用量，真的很酷。

But the fact that they're able to get this much usage on a new novel AI product is pretty cool.

他们还受到算力的限制。

They're also constrained by compute.

这也是为什么它目前只在美国上线。

That's partly why it's only rolled out in America.

但如果你现在住在美国，你就能用上这个东西，可以随便玩玩。

But if you're living in America right now, you get access to this thing and you can play around with it.

那它到底能帮你做什么呢？

So what exactly can this do for you?

嗯，我就拿我自己的亲身经历来说吧，这条推文当时还火了。

Well, I'll speak personally from my own anecdote on this tweet that actually went viral.

我当时正要去一个地方，结果我的第二班航班延误了七个小时。

I was basically on my way to a destination and my flight, my second flight got delayed seven hours.

所以我被困在机场里，那真是最让人抓狂的经历。

So I was stranded in this airport and it was the most frustrating experience ever.

我无聊得很，就打开Muse说，听着，我的航班延误了七个小时。

So I was bored and I brought up Muse and I said, listen, my flight's been delayed seven hours.

你能帮我申请赔偿吗？

Can you file for compensation for me?

我完全不知道自己有没有资格获得赔偿。

I had no idea whether I qualified for compensation.

我对航空公司的政策和规定一无所知。

I had no idea about my airline's policies or guidelines.

我就跟Muse说，嘿，我的第二班航班延误了。

I just told Muse, Hey, my second flight's been delayed.

你能帮我看看怎么拿到赔偿吗？

Is there anything you can do to help me out and figure out getting compensation?

结果它做了件特别神奇的事。

And it did a miraculous thing.

就在五分钟内，从我写下那句话开始，五分钟后，我的账户里就多了250美元的达美航空积分，可以用来以后坐飞机，它还帮我重新订了一班新的航班，让我比等着坐延误的那班飞机快得多地到达了目的地。

In exactly five minutes, so five minutes later from writing that prompt, I had $250 worth of Delta credits for future flights in my account and it even helped me rebook a new flight which got me out to my destination much faster than I would have if I had to wait for my delayed flight.

这里有几件事要说明一下。

Now a few things to point out here.

第一，我没有给它任何关于我航班的信息，但因为它能读取我的短信，它看到我一个小时前收到达美航空关于这次延误的消息或通知。

One, I gave it no context about my flights, but because it was able to read my text, it saw that I had a message or notification from Delta an hour earlier about this particular delay.

所以它很清楚我乘坐的是哪个航班，哪个航班或航空公司被延误了，以及我被改签到哪里，或者我本来想去哪里。

So it knew exactly what flight details I was on, which flight or airline carriage got delayed and where I was routed to or where I was trying to get to.

而且它没有把这些信息原封不动地复述给我。

And it didn't regurgitate all of that back to me.

它只是好像心里有数，很自然地回应我，说，听着。

It just kind of knew and responded intuitively to me and said, listen.

我懂你。

I got you.

我会，我会帮你搞定的。

I'll I'll figure this out.

第二，它能够自己导航到达美航空的网站。

Number two, it was able to navigate to the Delta website.

它并没有连接我的任何达美登录账号或达美应用。

So it didn't connect to any of my Delta logins or my Delta apps.

它只是直接去了官网。

It just went to the main website.

它找到了正确的联系方式。

It found the right contact details.

它自动生成了一封邮件。

It spun up an email.

它加上了我自己的邮箱地址，然后写完了整封邮件，并提交了整个表单。

It added my own email address and then it wrote out the entire thing and filed the entire form.

它加上了我的SkyMiles会员号。

It added my SkyMiles number.

它做了一大堆事情，这些事基本上要花我，我想说，三十到四十分钟。

It did a whole load of things that would have basically taken me, I wanna say thirty to forty minutes.

光是找到那个页面，我自己就得花三十分钟。

It would have taken me thirty minutes alone to find the actual page.

那个页面藏得很深，要翻好几层才能找到。

It's like hidden under several different pages.

Muse通过启动自己的桌面浏览器，完全自主地完成了这一切。

Muse did it all autonomously by spinning up its own desktop browser.

所以这就引出了这个东西到底是怎么运作的。

So this leads me to how this thing actually works.

说到AI模型，通常你只是跟它对话，它会根据已有的知识给你一个回答。

When you think of an AI model, typically you just talk to it and it gives you back a response built on the knowledge that it already has.

而像Muse这样的AI代理，它有自己的电脑。

With an AI agent like Muse, it is given its own computer.

这被称为虚拟机或虚拟电脑。

It's known as a virtual machine or a virtual computer.

你可以把它想象成你面前的笔记本电脑、台式机，或者你手机上的屏幕。

Think of this as like your laptop or your desktop that you see in front of your screen or on your phone right now.

但这个电脑是给这个AI大语言模型的，它基本上能做你能做的所有事情，只不过是在它自己的实例上。

But it's given to this AI LLM that can basically do the exact same things that you can, but on its own instance.

这里要注意的是，它跟你用的不是同一个桌面。

Now it's important to note here that it doesn't share the same desktop as you.

它看不到你的个人文件。

It can't see your personal files.

它看不到你自己电脑上的任何东西。

It can't see anything that you have on your own computer.

这是一个完全独立的东西，但它能做你能做的所有事情。

This is a completely separate thing, but it can do all the things that you can do.

它可以下载各种不同的插件。

It can download various different plugins.

它可以在浏览器上导航。

It can navigate on a browser.

它可以像你一样打开标签页，并且以闪电般的速度自主搞定这一切，因为它是AI，是软件。

It can open up tabs in the same way that you do and figure out all of this autonomously at lightning speed because it's AI and it's software.

所以它就这样反复地做，而且从不觉得无聊，这真的很棒。

So it just kind of like repeatedly does that and it never gets bored, which is amazing.

而这只是我遇到的众多例子中的一个。

And this is just one of very many examples that I've had.

我见过一个人，每天早上散步时拍下从树上或裸露电线上垂下来的电话线，然后让他的Muse代理向Verizon提交投诉，派技术人员去修理。

I've seen a guy take a picture on his walk every morning of telephone wires that were dangling down from trees or from exposed wires and ask his Muse agent to file a complaint to Verizon to send a technician out there to go and fix it.

然后它就完全自主地完成了，一个小时后他收到邮件说已经安排好了，你不需要到场，技术人员会在接下来的四十八小时内修好。

And it just did it completely autonomously and he got an email an hour later saying it's been scheduled, you don't need to be there, technician will fix this over the next forty eight hours.

我还有一个个人经历，最近我遇到高通的一位高管，她说，我不明白大家为什么对AI代理这么兴奋。

I have another personal anecdote where I met this exec at Qualcomm recently and she was like, I don't get the fuss around AI agents.

我觉得这不会是我个人生活中会用到的东西。

I don't think this is something that I might use in my personal life.

然后我说服她试试Muse，因为她有很多照片散落在Facebook、WhatsApp、iPhotos这些地方。

And I managed to convince her to play around with Muse because she had a bunch of photos scattered around Facebook, WhatsApp, her iPhotos and stuff like that.

几天后她给我写了一封很长的邮件，基本上说这彻底改变了我的生活，我现在用它做很多事情。

And she wrote me a very lengthy email a few days later basically saying this has completely changed my life and I'm using it for so many different things.

她有一个读书俱乐部。

She has a book club.

她让它帮忙组织给大家订书，寄到每个人的不同地址。

She gets it to organize ordering books for everyone, sending it to their various different addresses.

她也用同样的方式处理电影推荐。

She does the same thing for movie recommendations.

所以我想说的是，我会把Muse形容为一种生活方式的改变，一个由AI驱动的生活方式产品。

So the point here is I would describe Muse as a lifestyle change, a lifestyle product that is powered by AI.

AI并不是最显眼的部分。

AI isn't front and center.

它就是一个非常好的产品。

It's just a really good product.

而且实际上，当你揭开它的面纱，你会发现，为你做这一切的甚至不是前沿模型，这大概是下一个有趣的点。

And actually, when you uncover the hood, it's not even a frontier model that is doing all of this for you, which is the next interesting thing, I guess.

它用的是Muse Spark 1.2或1.3，也就是Meta最新的模型。

It is Muse Spark 1.2 or 1.3, whatever the latest Meta model is.

所以，看过之前几集的朋友都知道，Meta的模型并没有CloudFable 5.1或GPT 6.1 Astra那么聪明，那些才是领先的前沿模型。

So for any of you who have watched the previous episodes, Meta's models aren't as smart as CloudFable 5.1 or GPT 6.1 Astra, which are the frontier leading models.

但它的产品体验却比那两家公司打造的智能体要好。

And yet it has a better product experience than the agents that those two companies have built.

所以这就很神奇，Meta的优势在于，它拥有你Instagram、WhatsApp消息里的所有数据，它们几乎能直觉地知道你想要什么。

So it's this crazy thing where Meta has an advantage of having all the data built off of your Instagram, WhatsApp messages, and they kind of intuitively know what you want.

他们知道怎么跟每个人说话。

They know how to speak to individuals.

他们知道怎么打造一个真正好的消费产品，这就是我对MetaMuse的描述。

They know how to build a really good consumer product, and that's how I would describe MetaMuse.

如果你还没试过这个，我强烈推荐。

If this is something that you haven't tried yet, I would definitely recommend.

说到这儿，你可能在想，这东西免费吗？

Now you might be wondering on that note, is this thing free?

我还需要再付一个订阅费吗？

Do I need to pay for another subscription?

还是说，我直接打开就能用，它能连上我现有的任何账户，然后我就能直接用了。

Or is this something that I can just kind of flip on, it connects to any of my existing accounts and I can just use it.

是后者。

It's the latter.

就是我刚才给你的那个答案。

It's that answer that I just gave you.

第一，完全免费使用，直到每周大概一亿个token的上限。

Number one, it's completely free to use up till I think a 100,000,000 token cap per week.

我跟你说，如果你每周在这个AI代理上用到一亿个token，那你肯定是在干别的事。

Now let me tell you, if you're hitting a 100,000,000 tokens worth of usage on this AI agent per week, you are doing something else.

你肯定在经营好几家公司，普通人根本不需要达到那个上限。

You are running multiple businesses and there's no other reason that the average person needs to be hitting that cap.

但如果你真的特别喜欢Muse，它渗透到你生活的方方面面，比如帮你列购物清单、给你做饭、控制灯光什么的，那你可能会达到一亿的上限。

But if you end up do loving Muse so much and it hits every single aspect of your life and I don't know it's automating your grocery list, it's cooking for you, it's controlling your lights or whatever that might be, you might hit that 100,000,000 cap.

如果真的到了，那也只要每月20美元，虽然这跟一般AI订阅的价格差不多，但能让你用到大概五亿个token，这量真的非常充足。

And if so, it only costs you $20 a month, which granted is the cost of an average AI subscription, but it gets you to, I think it's like 500,000,000 tokens, which again is so plentiful.

我简直没法强调这有多划算。

I cannot emphasise how much this is.

你完全不用担心。

You don't need to worry.

我觉得免费版就完全够用了。

I think the free version is all good.

而且最酷的是，你可以直接把它连接到你的ChatGPT订阅或者Clothes订阅，然后就能让这个代理帮你做任何事，所有事都能搞定。

And the cool part is you can literally give it access to your ChatGPT subscription or your Clothes subscription, and you can just get this agent to do anything and everything for you.

最后，再提一下我之前说过的隐私问题。

And finally, a last point on the privacy thing that I mentioned earlier.

听着，Meta在隐私方面的名声一直不太好。

Now listen, Meta's had a bad rap when it comes to privacy.

大家都没忘记剑桥分析那件事，所以用Meta的产品时心里都有点防备。

People haven't forgotten the Cambridge Analytica stuff and so they're kind of wary when they use Meta's products.

他们预测，或者说我猜大家预测，Meta会用这些信息来优化自己的产品，让产品更赚钱或者更让人上瘾，这样你就会对它们的产品越来越依赖。

They predict or I would assume people predict that Meta would use this information to benefit their products so that they make it more lucrative or stickier to use, which gets you kind of addicted to their products.

不过，扎克伯格在发布Muse的时候明确说过，这款产品是以隐私为核心的。

Now one thing Zoc made very clear when he announced or launched Muse is that it's built around privacy.

而且他们用一种很特别的方式做到了这一点。

And they've done this in a very particular way.

我不打算深入讲它的架构，但基本上可以这样理解。

I'm not going get into the architecture, but it basically can be thought of as follows.

想象一下一家酒店。

Imagine a hotel.

这是一家豪华酒店，有很多房间，成千上万个房间。

It's a luxurious hotel and it has many rooms, hundreds of thousands of rooms.

每个房间都专门对应一个特定的智能体。

Each room is dedicated to a particular agent.

所以当你启动一个Muse智能体，打开应用并注册后，你就拥有了自己的智能体，它有自己的房间。

So when you spin up a Muse agent, when you open up the app and you sign up, you have your very own agent that has its own room.

那个房间其他任何人都无法进入。

Now that room can't be accessed by any other individual.

Muse的其他用户都无法以任何方式访问你的专属房间。

No other user of Muse can get any kind of access to your particular room.

在那个房间里，有一个相当于酒店保险箱的东西。

And in that room, there's equivalent of a hotel safe.

所有密码和凭证都会保存在那个保险箱里并加密，如果你不希望Meta访问，连Meta也无法获取。

And in that hotel safe is where all your passwords and credentials get saved and encrypted so that not even Meta can get access to it if you deem it.

所以，如果你想让你的代理访问你的邮箱，或者想让它访问某个需要密码的账户，你可以放心地这么做，不用担心任何人能接触到它。

So if you want to give your agent access to your email or if you want to give it access to, I don't know, some account that's secured by a password, you can do so confidently without worrying that anyone could get access to it.

至于房间本身，你知道，它就像一家酒店。

Now as for the room itself, you know, it's a hotel.

所以还有其他代理在同一个系统里四处活动。

So you have other agents wandering around the same kind of system.

Meta 相当于有一个礼宾员或哨兵守卫，基本上监控着你与 Muse 代理互动中进出的所有内容。

Meta has the equivalent of a concierge or a sentry guard, which basically monitors anything and everything that comes in and comes out of your interaction with your Muse agent.

如果它发现任何问题、恶意垃圾信息或其他什么，它会立刻切断。

And so if that flags any kind of issue or malicious spam or whatever that might be, it instantly cuts it out.

到目前为止，这是一种非常强大的防御攻击的方法。

And so far it's been an extremely robust method of defending against any kind of attacks.

我还没遇到过任何问题，实际上我也没见过其他人遭遇类似事件。

I haven't come across anything, I actually haven't seen any other instance of people suffering from incidents like that.

所以重点是，它是私密的，而且可以用来做一些非常酷的事情。

So the point is it's private, and it can be used for some really cool things.

最后我还想补充一点，从商业角度来看，用这个代理能做的事情简直太疯狂了。

The last thing actually I'll add about this is from a business perspective, it is crazy what you can do with this agent.

我见过一些例子，有人在Instagram上投放广告，或者他们的竞争对手在Instagram上发布社交媒体内容。

I've seen examples of people who run ads on Instagram or who have competitors that post social media content on Instagram.

然后他们问自己的Muse：嘿，你能去看看这个人的生意吗？

And they've asked their Muse, hey, could you go check out this person's business?

你能不能看看他们内容的风格，然后帮我模仿一下？

Could you kind of like look at the style of their content and maybe recreate the same for me?

或者，你去看看他们的粉丝列表，然后给所有粉丝发私信，说：嘿，我们给你提供五折优惠。

Or actually, you check out their follower list and DM all of them and say, hey, we're offering you a 50% discount.

去用我们的产品吧。

Go use our product.

然后Muse代理就会直接去执行这些操作。

And Muse agent will just go out and do that.

所以它既是一种生活方式的东西，又是一个非常实用的商业工具，但作为AI代理，它并不会让人感到害怕。

So it's this weird mixture of like a lifestyle thing and a really good business application, but it's not too intimidating as an AI agent.

我不知道。

I don't know.

这真的很酷。

It's very cool.

好的。

Okay.

接下来要讲的是我们的救世主，我们最崇拜的人，埃隆·马斯克。

So the next one is from our lord and savior, our favorite man ever, Elon Musk.

今天这期《In the Loop》节目，由高通公司赞助播出，我们会深入探讨所有这些话题，还有更多精彩内容。

So we're gonna dig into all of that and more on today's episode of In the Loop brought to you by Qualcomm.

马斯克最近可忙了。

Now Elon's been busy.

他一直在为SpaceX的AI项目疯狂采购计算资源。

He's been busy acquiring all the compute that he can for SpaceX AI.

我听说那个项目叫……

I actually heard that it's called.

它被重新命名为SpaceX SI，也就是超级智能，这个改名把之前的都盖过了。

It's being rebranded to SpaceX SI, superintelligence, the rebranding that trumped it.

这事儿挺大的。

It's a whole thing.

但他一直在筹钱买算力。

But he's been raising money to buy compute.

他搞了有史以来最大的AI IPO之一，比传闻中几周后要来的Anthropix还早。

He's IPO ed one of the biggest AI IPOs to date prior to Anthropix that's reportedly coming in a few weeks.

而且SpaceX实际上在前沿AI模型上还是落后了。

And SpaceX has, for all intents and purposes, still fallen behind on frontier AI models.

Grok 4.7发布了。

Grok 4.7 got released.

听着，这是个很棒的模型。

And listen, it's a great model.

它的单位智能成本非常低，但绝对算不上前沿。

It's very cheap per unit cost of intelligence, but it's by no means frontier.

埃隆·马斯克对此非常坦诚。

Elon Musk has been very upfront about this.

他说Grok 5最终会在年底前推出，并且将能与我们现在看到的Fable相媲美。

He said that Grok five will eventually come out by the end of the year, and that will be competitive to something that we're seeing with Fable right now.

但说到消费产品，我觉得SpaceX，或者更确切地说是X，确实有真正的优势，因为他们在全球有数百万用户，用X来转账、像银行一样存钱。

But when it comes to consumer products, I would say that SpaceX or X specifically has a real advantage because they have millions and millions of users all over the world that uses X to transact money, to store money like a bank.

X Money大概几个月前推出了，用来在社交媒体网站上分享世界上最重要的新闻和信息。

X Money launched, I think, a few months ago and to share the world's most important news and information on the actual social media site itself.

所以你可以说，它可以构建类似Meta的Muse那样的东西。

So you could argue that it could build something similar to what Meta has with Muse.

而实际上他们就是这么做的。

And that's actually what they ended up doing.

现在GrokBot在几个方面很独特。

Now GrokBot is unique in a few different ways.

GrokBot不仅仅是一个单一的AI代理，而是更大代理群体中的一员。

GrokBot can be thought of as not just a single AI agent, but a single agent collectively part of a larger swarm of agents.

这里有很多技术术语，但基本上就是说，不是只有一个智能体为你工作，而是有成百上千个智能体在为你工作。

Now there's a lot of technical jargon there, what that basically means is instead of one agent working for you, there are hundreds, maybe even thousands of agents working for you.

而且它们都能像MetaMuse那样访问电脑。

But they all get access to a computer in the very same way that MetaMuse is able to do that.

GrokBot的重点不一定在生活方式上，虽然根据我的个人经验，它在这方面也很擅长，但它更侧重于生产力。

And the focus of GrokBot isn't necessarily around lifestyle, although it's very good at doing that in my personal experience, but it's more focused around productivity.

即使在你面前这个演示中，GrokBot基本上也能访问你在个人生活或企业工作中可能使用到的任何高性能工具。

Even in this demo that you see in front of you, GrokBot basically gets access to any high powered tools that you might use on personal life or productive enterprise work kind of life.

比如Slack，比如管理客户或业务的CRM系统。

So think about Slack, think about CRMs to manage your customer or your business.

还有Notion、Jira、Atlassian，以及你日常可能用到的所有工作生产力应用。

Think about Notion, Jira, Atlassian, all the work productivity apps that you might use on a day to day basis.

你可以让GrokBot访问这些，它会安全加密存储你的密码，不会泄露任何信息。

You can give GrokBot access to, stores your password safely encrypted so that it doesn't leak any of that stuff.

而且它确实能做好知识型工作。

And it is able to actually conduct really good knowledge work.

根据我的经验，当我在为特定节目做研究、创建演示、想和你们聊什么话题的时候，它真的是一个很好的同事。

Now in my experience, when I'm doing sorts of research around particular episodes that we're going to do, creating demos, coming up with ideas of what to talk to you guys about, this has been a really good colleague.

我是故意用“同事”这个词的。

And I mean that term intentionally.

你可以把它看作是一个工作伙伴，而不是像朋友、导师或者像MetaMuse那样帮你节省时间的东西。

It can be thought of as like a work partner and less of something that can be like a friend, a mentor or something that can kind of like free up your time like MetaMuse.

GrokBot最酷的功能是，它可以自主监控你做的任务，然后最终独立地复制你做那些任务的过程。

Now the coolest feature by far with GrokBot is that it can monitor tasks that you do autonomously and then eventually replicate you doing those exact tasks on its own separately.

让我再说一遍。

Let me repeat that.

假设一个AI对某个任务一无所知，而这个任务你非常熟悉，可能是你花了好几年研究才搞明白的，或者因为你已经在网上混了几十年，所以自然而然就会做的。

Let's say that an AI has no idea about a task that you know very well-to-do that, that you've studied years trying to figure out or that you just do naturally because you've been on the internet for decades at this point.

GrokBot或者大多数AI模型通常缺乏这种背景知识。

GrokBot or most AI models in general usually lack the context.

你通常得手把手地教它。

You usually have to spoon feed it.

你得解释一下，嘿。

You need to explain, hey.

这就是我在做什么，以及我为什么这么做。

This is what I'm doing and this is why I'm doing it.

GrokBot 会观察你。

GrokBot sort of watches you.

它会记录你的屏幕，并理解你为什么做你正在做的事情。

It sort of records your screen and kind of understands why you're doing what you're meant to be doing.

所以这里的想法是，最终它会成为你日常可能想做的任务的完美模拟或预测器。

And so the idea here is that eventually it becomes the perfect emulation or predictor of tasks that you might want to do on a daily basis.

所以最终它会让你腾出时间去做其他事情，它可以把所有这些从你的日程中拿走。

So eventually it frees you up to go and do other stuff and it can just kind of take all of this off your plate.

我觉得这真的很酷。

I I think that's really cool.

现在说到价格，它和 Cursor 捆绑在一起。

Now in terms of pricing, it comes bundled with Cursor.

Cursor 是 SpaceX 在 IPO 后不久以 600 亿美元收购的 AI 编程平台。它不是免费使用的。

Cursor is the AI coding platform that SpaceX acquired shortly after their IPO for $60,000,000,000 It's not free to access.

你至少得付钱，我觉得是每月 20 美元，每周使用量很有限或者不算太有限，但有上限。

You have to pay at minimum, I think it's $20 a month for very limited or not so limited weekly usage, but it's capped.

它不像 MetaMuse 那样免费提供 1 亿个令牌。

It's not as free as MetaMuse with 100,000,000 tokens.

我觉得还是少一点，但如果你每月花 60 美元升级到 Cursor Pro Plus，就可以提高这个上限。

I think it's still a little less than that, but you could increase that limit if you forked out $60 a month with Cursor Pro Plus.

所以这更像是一个捆绑服务。

So it's more of a bundle thing.

当然不是免费的。

It's certainly not free.

如果你打算用这个平台，我觉得你应该在工作上先试试看。

And if you're gonna end up using this platform, I think it's something that you should trial out for your work stuff.

这是我个人的看法。

That's my personal take.

我觉得如果你想玩玩AI代理，真正理解它背后的神奇之处，但又没有明确的使用场景，甚至有点困惑自己到底想用它做什么，那就去用MetaMuse吧。

I think if you want to play around with an AI agent and really understand the magic behind it and you don't have an explicit use case, you kind of are confused about what you might want to use it for, go out and use MetaMuse.

那是目前最好的选择。

That's the best thing that's available right now.

但关于GrokBot，我最后想说的一点是，同时启动多个代理帮你干活，效率比我预想的要高得多，但确实也有点让人摸不着头脑。

But the final thing I'll say about GrokBot is spinning up multiple agents to do your work for you was way more productive than I thought it would be, but it's also a little confusing.

我喜欢像Muse或Instinct这类应用的地方，就是它们只有一个简单的界面。

Now what I like about, I guess an app like Muse or Instinct that I'm just about to talk about is that it's just a single interface.

就一个代理。

It's a single agent.

这样你可以和那个代理建立一种关系。

So you could build a sort of relationship with that agent.

但用GrokBot的话，你做不到这一点。

You can't really do that.

我用GrokBot的时候没有那种感觉。

I don't get that sense from using GrokBot.

这更像是跟一支团队一起工作。

It is more like working with a workforce.

他们有点机械和死板，但效率极高。

They're a little more robotic and mechanical, but extremely productive.

所以再说一遍，如果你是一家企业，正在考虑用AI代理还是模型更高效、更省钱，那就去试试吧。

So again, if you're an enterprise that's trying to figure out whether using AI agents versus models are more productive and cheaper for you, go try it out.

现在第三个AI代理有点奇怪，它好像没有固定的归属。

And now the third AI agent is one that weirdly doesn't have a home.

它叫Instinct，是一家隐形创业公司，我觉得几个月前才公开亮相，但下载量已经超过50万次，而且严格来说还在内测阶段。

It's called Instinct and it was this stealth startup which I think came out of stealth a few months ago but it's already been downloaded over 500,000 times and technically it's still in private beta.

它是由一个23岁的年轻人创立的，他在构建AI产品方面经验丰富。

It was started by a 23 year old that has a lot of history in building AI type products.

他叫Noah Shin，是个绝对的天才。

His name is Noah Shin and he's an absolute genius.

他以前在Sierra AI工作，那是一家专注于客服的AI公司。

He used to work for this company called Sierra AI, which was a customer service focused AI.

但真正让我印象深刻的是他写的那篇叫《Reflexion》的论文，我记得是2023年写的。

But what I'm really impressed by is this paper called Reflexion that he wrote I think back in 2023.

它是有史以来被引用最多的AI智能体研究论文。

And it is the most cited research paper for AI agents ever.

我看了这篇论文，它的核心思想是，如果你能指导AI或AI智能体在文档中写下它犯错或无法完成任务的根本原因，然后让它重新执行该任务，它会做得更好。

And so I looked at this paper and the core idea behind it is if you can instruct an AI or an AI agent to write down on a document the core reason why it made a mistake or why it couldn't complete a task and then get it to rerun that task, it does a much better job.

这个方法在GPT-4中得到了著名的应用，当时它的编码通过率是80%。

And this was famously used and seen in GPT-four which had a coding pass rate back then of 80%.

因为这篇论文，因为Noah Shin，通过率提升到了91%。

And because of this paper, because of Noah Shin, it got pushed to 91%.

所以Noah在最重要的两件事上有着丰富的背景和信誉：开发前沿AI，以及客户服务。

So Noah has a lot of background and credibility in two of the most important things, building bleeding edge frontier AI and also customer service.

也就是打造人们真正想用的产品。

So building a product that people actually want to use.

而这正是Instinct这个产品的真正体现。

And that's really the manifestation of this product called Instinct.

首先，我之前提到过，它没有家，因为它真的没有家。

Now, firstly, I mentioned earlier, it doesn't have a home because it literally doesn't have a home.

当我收到这个的邀请链接时，我只是点了链接，给了我的号码，然后它就突然出现了。

When I got an invite link to this thing, I just tapped the link, I gave it my number and then suddenly it just kind of appeared.

它通过我的iMessage给我发短信。

It texted me on my iMessage.

于是我打开iMessage，回复说你好，它就说，嘿，我是Instinct。

And so I opened up my iMessage and I responded saying hello and it said, hey, I'm Instinct.

你需要帮忙吗？

What do you need a hand with?

然后它开始变得更加主动，就像MetaMuse一样。

And it then started to be a lot more proactive in the same way that MetaMuse was.

与Muse相比，它的酷之处在于，没有单独的应用程序需要我去跟踪或打开。

Now what's cool about this compared to Muse is that there is no separate app that I need to kind of keep track of or open.

它只是一个非常强大的AI代理，完全隐形。

It is just kind of this very imposing AI agent that is completely invisible.

我知道这听起来完全是个矛盾的说法，但它确实就是这样。

And I understand that that's completely an oxymoron, but it really is that.

它无处不在，但又不会在我面前晃来晃去，也不会试图控制我做的每一件事。

It is anywhere and everywhere that I am without being in my face and imposing and trying to control everything I'm trying to do.

它就是一种环境式AI，很像托尼·斯塔克电影里的贾维斯，钢铁侠粉丝们懂的，而且用起来特别简单。

It's just this kind of ambient AI very much like a Jarvis from Tony Stark interview, Iron Man fans out there, and it's really easy to use.

所以，它其实是我最喜欢用的产品之一，但目前还在内测阶段。

So actually, it's one of my favorite products to use, but it is in private beta.

关于这个，我确实有一些看法。

And I do have a few comments on this.

最主要的一点是关于服务条款和隐私的。

The main one being around the terms and agreement and privacy.

我刚才解释了MetaMuse是如何围绕隐私来构建的，并且从过去的事件中吸取了很多反馈。

Now I just explained to you why MetaMuse has been built around privacy and has taken a lot of feedback from previous past incidents.

但Instinct的情况就不是这样了。

This hasn't been the case with Instinct.

当你签署条款协议并允许它访问你的手机，然后假设你删除了账户，它仍然拥有它之前访问过的所有数据，这些数据已经被索引，可以用来训练和改进模型。

When you sign the terms and agreements and you give it access to your phone and let's say you delete your account, it still has all the data that it had access to, that it's indexed, that it's able to train on and build a better model.

这也是很多AI领域的知名人物没有签约使用这个产品的一个关键原因，因为它相比其他任何产品都更具侵入性。

And that's been a key reason why a lot of very prominent figures in AI themselves haven't signed on to this thing because it's quite invasive compared to any other product that you could go out and use.

另外，我不知道背后是什么模型在驱动，也许你最终不在乎，但我确实想知道我的数据是如何被背后那个模型使用的。如果它是一个中国的开源模型，我可能会有不同的看法，不管它是中国的还是西方的顶尖模型。

The other thing is I don't know what model is powering it behind the hood and maybe you don't care at the end of the day but I do like to know how my data is being used for whatever model behind the hood and you know if it's a Chinese open source model I might have different thoughts about it whether it's that or versus a Western frontier model.

但所有这些可能仍然值得，因为有几个不同的功能。

But all that might still be worth it for a few different features.

第一，价格。

Number one, the price.

它完全免费，而且被补贴得很厉害。

It is completely free and being subsidised to high heaven.

当Instinct被要求评论他们的商业模式时，他们的回应是：我们的目标基本上是收取你通过Instinct促成的每笔交易的一定比例。

Now when Instinct was pushed to comment on their business model, their response was, our aim is to basically take a percentage of every transaction that you might facilitate using Instinct.

所以我的理解是，他们会大力推动通过Instinct代理进行电子商务。

So the idea I get from there is there's going to be a heavy push on e commerce via the Instinct agent.

所以你应该很容易就能去买东西，或者使用某个产品或服务。

So it should be very easy for you to go out and buy stuff or engage with a product or service.

然后他们就从交易中抽成，这挺合理的。

And they'll just take a percentage cut from that, which makes sense.

但我感觉他们不可能永远补贴下去，而且人们也不会完全信任AI代理去帮他们购物。

But I have a feeling that they can't subsidise this forever and people aren't gonna be fully trusting of an AI agent to go and do its shopping.

他们连基本的购物想法都提不出来。

They can't even come up with basic ideas of what to buy in the first place.

所以我觉得在这方面有点牵强。

So I feel like that's a bit more of a stretch when it comes to this.

所以我觉得不会永远免费，但现在试试还是挺酷的。

And so I don't think it'll be free forever, but certainly cool to try out for now.

第二件事是我亲眼看到Instinct打电话给一个服务，比如餐厅，帮我订了位。

The second thing is I have watched Instinct literally call up a service, in this case a restaurant, and book a reservation for me.

我特别好奇这是怎么运作的，它接入了11 labs的语音API，完全自主地完成了整个过程，看着挺奇怪也挺超现实的，但确实很酷。

I was super curious about how this would work and it got access to 11 labs, voice API and was able to just do the whole thing completely autonomously, which was very weird and surreal to watch, but cool nonetheless.