什么样的 AI 助手值得付费？消费级智能体的价值、信任与成本

简要描述Brief Description

什么样的 AI 助手值得付费？What Would Make an AI Assistant Worth Paying For?

目录Table of Contents

• 简介
  Intro
• Poke、Instinct、Muse 与智能体热潮
  Poke, Instinct, Muse and the agent boom
• Assistant Bench 实际测试的内容
  What Assistant Bench actually tests
• 节省成本胜过节省时间的方案。
  Cost savers beat time savers
• Muse 的魅力：Meta 的数据布局
  Muse charm as Meta's data play
• 群聊中的静默代理
  Silent agents in group chats
• 主动性才是真正的护城河
  Proactivity is the real moat
• 助手与智能体的定义
  Assistant vs agent, defined
• 亚马逊封杀 Muse，Shopify 敞开大门
  Amazon blocks Muse, Shopify opens the door
• 每天 20 美元的智能体经济账
  The $20/day agent economics

引言Intro

嘉宾：普通大众并不在乎效率提高 10%。我的大胆论点是，Muse 吊坠的重点与其说是赢得硬件竞争，不如说是收集现实世界的数据，通过绘制真实世界的地图，为扎克伯格未来的元宇宙提供动力。Guest: The general population does not care about being 10% more efficient. My hot-take thesis is that the Muse charm is less about winning the hardware game and more about collecting real-world data to fuel Zuckerberg’s future metaverse by mapping the actual world.

主持人：我们私下开玩笑说，再过几天，智能体可能就会给某人发消息：“我注意到你没那么喜欢她，所以我就替你跟她分手了。”Host: We were joking internally that we’re days away from an agent messaging someone, “I noticed you weren’t that into her, so I went ahead and broke up with her.”

嘉宾：主动性会带来巨大的竞争壁垒。我骑车去上班时想把事情处理一下，于是开始用 ChatGPT 语音对话。在 30 分钟的骑行途中，它给我的邮件分类、归入不同标签、发送日历邀请，还在我到办公桌前把收件箱清空了。不过，这其中的界限非常微妙：只要越界一次，就会失去用户的全部信任。Guest: There’s massive defensibility around proactivity. I was biking to work and wanted to get stuff done, so I started talking to ChatGPT voice. Over my 30-minute bike ride, it categorized my emails, submitted them to different labels, sent calendar invites, and got my inbox to zero by the time I reached my desk. But it’ll be a very fine line: if you cross it once, you lose all trust with your user.

主持人：确实感觉未来会出现一些专为智能体之间交互而存在、如今还不存在的基础设施和产品。你怎么看？Host: It does feel like there’s going to be infrastructure and products that exist only for agent-to-agent interactions, which don’t exist today. What do you think?

嘉宾：我们必然会看到——Guest: It’s inevitable that we’re going to see—

Poke、Instinct、Muse 与智能体热潮Poke, Instinct, Muse and the agent boom

主持人：欢迎收听 A16Z 节目。很高兴能和我的朋友戴维·波伦一起聊聊。几周前我们还聚过，聊到个人助理和消费级 AI 的种种进展时，我们俩都特别兴奋。过去几周发生了太多事。我觉得，生态系统中的某些领域曾有过几次仿佛“见到了上帝”的时刻。ChatGPT 就是其中一个重要时刻。对许多人来说，在 2022 年 11 月以及进入 2023 年后，那种感觉是：“这到底是什么东西？它能写邮件和诗歌，能来回对话，还让人觉得自己正在和一个合成的人交谈？”那就是起点，宇宙大爆炸般的开端。Host: Welcome to the A16Z show. I’m so psyched to be here with my friend David Pollen. We hung out a couple of weeks ago, and we’re both enthusiastic about everything that’s been happening with personal assistants and consumer AI generally. So much has happened in the last couple of weeks. I think there have been a few moments when some part of the ecosystem has sort of seen God. ChatGPT was a big one. For a lot of people in November 2022 and into 2023, it was, “What is this thing that can write emails and poems, have a back-and-forth conversation, and feel like you’re interacting with a synthetic person?” That was the beginning, the big bang.

第二个重要时刻是编程智能体，以及开发者围绕 Claude Code 和 Codex 一直热议的话题，还有轻而易举地创建软件的能力。最近，随着 Instinct、Muse 和 ChatGPT Work 的出现，我们看到大家对个人智能体及其能为我们这些消费者做些什么，表现出了非同寻常的热情。今天我们要深入聊聊个人智能体的方方面面。我们会尽可能深入地探讨，并假设大家都看过许多和我们一样的内容，同时把这次对话作为系列讨论的第一期。欢迎你，David。The second moment was coding agents and everything developers have been obsessed with around Claude Code and Codex, and the ability to trivially create software. Most recently, with Instinct, Muse, and ChatGPT Work, we’ve seen extraordinary enthusiasm around personal agents and what they can do for us as consumers. We’re here to talk about all things personal agents. We’re going to go as deep as we can, assume everyone has seen many of the same things we have, and make this the first in a series of conversations. Welcome, David.

嘉宾：谢谢。感谢邀请我来。我很高兴能来到这里。我们把视角放大一些，聊聊过去几个月，甚至过去几周里个人智能体的发展历程。你观察到了什么？你怎么看这些变化？Guest: Thank you. Thanks for having me. I’m excited to be here. Let’s zoom out and talk about the arc of personal agents over the last few months, or even the last few weeks. What have you been seeing? What’s your view on things?

主持人：过去四周，这个领域发展得如火如荼。我是 Poke 的早期用户之一，它是最早真正进入市场的消费级 AI 智能体之一。我记得它是在 9 月 8 日上线的，而我在 9 月 11 日，也就是三天后，就开始用了。我对此记忆犹新，因为那次体验简直不可思议。它的整个噱头就是让你和智能体讨价还价，商量你实际要付多少钱，这彻底颠覆了我的认知。Host: This space has exploded over the past four weeks. I was an early adopter of Poke, one of the first consumer AI agents to really hit the market. It launched, I believe, on September 8, and I started using it on September 11, just three days later. I remember it vividly because it was an unbelievable experience. Their whole gimmick was negotiating with the agent over how much you were actually going to pay, and it totally blew my mind.

接着，在 11 月下旬，OpenClaw 发布了，迅速风靡全球。那时感觉有什么大事即将发生。后来零星出现了一些其他产品，但最近我们看到 Instinct 在科技圈的 Twitter 泡沫中大火。它乘势而上，势头迅猛，随后多米诺骨牌接连倒下：Grokbot、Muse，以及 Caddy、Ali 或 Pi 这类一大批长尾产品。看到科技圈对这些智能体能如何帮助我们处理日常事务表现出如此浓厚的兴趣和求知欲，真是令人惊叹。我觉得大家的处境都一样：没人真正知道答案。这里就像狂野西部，而我们只是在摸索着玩。Then, in late November, OpenClaw was released and took the world by storm. It felt like something was coming. A few others have popped up here and there, but most recently we saw Instinct really blow up in the tech Twitter bubble. It rode a wave, went nuts, and then the dominoes fell: Grokbot, Muse, and a bunch of long-tail ones like Caddy, Ali, or Pi. It’s been amazing to see the tech bubble explode with interest and intellectual curiosity about what these agents can do for us day to day. I think everyone’s in the same boat: nobody really knows. It’s the Wild West, and we’re just playing around.

Assistant Bench 实际测试的内容What Assistant Bench actually tests

嘉宾：我完全同意。我们应该花点时间聊聊 Assistant Bench，这是你编写的一个全新且重要的基准测试。你可以先简单介绍一下，然后我再问你几个问题。Guest: I totally agree. We should take a moment to talk about Assistant Bench, this new, important benchmark that you authored. Maybe tee it up, and then I’ll ask you a few questions.

主持人：Assistant Benchmark 是一个针对特定使用场景比较 AI 助手的网站。我们给它们同一个提示，例如：帮我订一张去芝加哥的机票，或者在我所在位置五个街区范围内找一家素食餐厅。然后，我们比较单轮提示的结果：它能否完成任务？会不会提出后续问题？在 16 个不同维度上的表现如何？Host: Assistant Benchmark is a site that compares AI assistants on specific use cases. We give them all the same prompt: book me a flight to Chicago, or find me a vegetarian restaurant within five blocks of my location, for example. Then we compare the results of a one-shot prompt: Does it work? Does it ask follow-up questions? How does it perform across 16 different dimensions?

我们称之为基准测试，但不妨把它看作面向消费者的基准测试，而不是技术性的幕后测试。如果我想订机票，该用哪个？哪个表现最好，回复速度又有多快？我开始比较它们，是因为我在为自己做市场调研，想弄清楚它们各自的表现如何，以及存在哪些差距。我把其中一些发现发布了出来，网站也在 16 天前上线。结果大受欢迎：网站访问量已超过 10 万，而且几乎每位创始人都联系过我。我觉得这说明大家有多么好奇，也有多么渴望了解这个行业未来的发展方向。We call it a benchmark, but think of it more as a consumer-facing benchmark than a technical, behind-the-scenes one. If I want to book a flight, which one should I use? Which performs best, and how quickly does it reply? I started comparing them because I was doing market research for myself and trying to understand how they stacked up and where the gaps were. I published some of that, and the site launched 16 days ago. It’s absolutely exploded: the website has had over 100,000 visitors, and I’ve heard from basically every founder. I think it shows how curious people are and how much they want answers about where this industry is going.

嘉宾：这真是太令人兴奋了，David。有趣的是，基准测试如今已经成了大家讨论中的常客。新产品发布实在太多了，很难每样都试一遍——这本身也是个令人振奋的信号。我们来聊聊 Instinct 和 Muse 以外的公司吧，之后再深入探讨它们。还有哪些公司在做有意思的工作？你看到了哪些专业领域？Guest: That’s incredibly exciting, David. It’s funny how benchmarks have become such a staple of the conversation. There are so many launches that it’s hard to try everything, which is also an exciting signal. Let’s talk about the companies beyond Instinct and Muse—we’ll deep dive on those later. Who else is doing interesting work, and what areas of specialization are you seeing?

主持人：长尾解决方案有很多。我会把它们分成两类。第一类是面向消费者的 B2C 通用智能体，比如 Instinct 或 Muse。一些更长尾的产品包括 Caddy、Ali 和 Season。我记得网站上通用类别下有 64 个。另一类是更专业的助手，比如专注于旅行的 Soore 或 Miso，以及专注于电子邮件的助手。Host: There are a lot of long-tail solutions. I’d categorize them into two groups. First, there are B2C, consumer-facing general agents like Instinct or Muse. Some of the longer-tail ones are Caddy, Ali, and Season. I believe there are 64 in the general category on the site. Then there are more specialized assistants, like travel-specific ones such as Soore or Miso, and email-specific ones.

光谱的另一端是 B2B 工作流助手，从[不明确]到 Vellum。它们本质上是高管助理，而不是面向消费者的助手。这就是目前的大致格局。据我所知，所有类别加起来共有 122 种不同的助手。The other side of the spectrum is B2B workflow assistants, from [unclear] to Vellum. They’re essentially executive assistants, rather than consumer-facing assistants. That’s the general landscape at the moment. To my knowledge, there are 122 different ones across all the categories.

客人：哇。Guest: Wow.

主持人：哪些领域的竞争焦点最有意思？Host: What areas have the most interesting competitive focus?

嘉宾：总的来说，大家都专注于通才，试图什么都做。Guest: By and large, everyone is focused on the generalist—trying to do everything.

主持人：比如说，旅行是一种低频、高价值的行为。你认为这个类别会更多地被整合成一个连接到通用型智能体的入口，还是说这些公司也有机会赢得通用型智能体市场？Host: When you think about travel, for example, it’s an infrequent, high-value behavior. Do you think that category will be organized more as a connector into a horizontal agent, or do those companies have a chance to win as general horizontal agents too?

嘉宾：你可能会认为有一个赢家会成为独立智能体，但总体而言，我认为旅行更像是一种连接器。我加入了七八个讨论智能助手的群聊，这些群里现在总共有1200多人。旅行是讨论度第四高的应用场景，这很有意思，因为从传播性来看，它往往排名第一：人们会谈论让智能助手帮自己办理航班值机，或替自己订机票。但如果你看看人们日常的对话，了解他们实际会用智能助手做什么、对什么感兴趣，就会发现人们并不是每天都在旅行。Guest: You may see one winner as an independent agent, but by and large, I think travel will be more of a connector. I’m in seven or eight group chats with people who are obsessed with assistants, and there are now over 1,200 people across them. Travel is the fourth most talked-about use case, which is interesting because it tends to be number one in terms of virality: people talk about an assistant checking them into a flight or booking one for them. But when you look at the day-to-day conversations about what people actually use assistants for and are curious about, people aren’t traveling every day.

主持人：所以这更像是个噱头，用来吸引注意力。Host: So it’s more of a gimmick, an attention grabber.

嘉宾：用例一到三是什么？第一是日常行政事务：清理收件箱、填写表格，以及其他烦人的琐事。我觉得这就触及了这个领域最终胜出者的发展方向这一更广泛的论点，我们可以聊聊这一点。Guest: What are use cases one to three? Number one is daily admin: cleaning out your inbox, filing forms, and other annoying tasks. I think that gets to the broader thesis of where the winner in this space will go, which we can touch on.

主持人：第二个是智能体编排，这真的很有意思。我觉得这些群聊有点偏颇。第三个是面向开发的。Host: The second is agent orchestration, which is really interesting. I think these group chats are a little biased. The third is development-focused.

嘉宾：这些都是和科技圈 Twitter 上的人组成的群聊，所以不一定能代表广大人群的情况。不过，大家经常讨论智能体编排：如何找到效率最高的智能体，以及如何让智能体彼此交流。他们也在讨论开发智能体，以及这会如何转化为手机编程——比如通过发短信使用 Claude Code，或者更进一步：如果我是一名设计师，能不能只通过给我的智能体发短信，就对设计文件做一些小改动？Guest: These are all group chats with people in tech Twitter, so they’re not necessarily representative of the world at large. But people are talking a lot about agent orchestration: how to find the most efficient agent and how to get agents to talk to each other. They’re also talking about development agents and how this translates to coding on your phone, whether that’s using Claude Code through text or taking it a step further: if I’m a designer, can I make small changes to my design files just by texting my agent?

最后一类是金融。在我看来，这是最有意思、最值得探讨的一类，因为它也是最不切实际的一类。这里面涉及太多决策了。如果智能体能帮你赚到十亿美元，所有对冲基金早就这么做了。不是每个人都能靠手机用它取得成功。不过，有趣的是，当你把一个强大的工具交到某人手里时，他们首先想到的就是：“你打算怎么帮我赚一百万美元？”The last category is finance. That’s the most fun one to look into because, in my opinion, it’s the least practical. There are so many decisions involved. If an agent could make a billion dollars for you, all the hedge funds would be doing it. It’s not something everyone is going to succeed at using their phone. But it’s interesting that, when you put a powerful tool in someone’s hands, the first thing they think is, “How are you going to make me a million dollars?”

节省成本比节省时间更重要。Cost savers beat time savers

主持人：这确实很有意思，因为几天前 Ben Thompson 在 TVPN 上提出的一项批评是，大多数消费者并不追求提高生活效率。他们想要的是消磨时间，而不是节省时间，对此我深表赞同。Host: It’s really interesting because one of the critiques Ben Thompson made on TVPN a few days ago was that most consumers aren’t looking for productivity in their lives. They’re looking to spend time, not save time, and I agree with that.

嘉宾：但对普通消费者来说，金融确实是一个行政管理负担很重的领域。Guest: But finance does feel like an area with a ton of administrative overhead for the average consumer.

主持人：如果你看看那些因消费者漠不关心或信息不足而形成的金融市场和利润池，就会发现其中有大量利润本可以返还给消费者。Host: If you look at the number of financial markets and profit pools defined by consumer apathy or people being uninformed, that’s a lot of profit dollars that could be delivered back to consumers.

嘉宾：虽然我总体上同意 Ben 的看法，大多数消费者都在寻找娱乐和消磨时间的方式，但美国消费者生活中有些重要方面涉及大量行政事务。这些智能体可能会给他们带来真正的突破。Guest: While I generally agree with Ben that most consumers are looking for entertainment and ways to spend time, there are defining aspects of the US consumer’s life that are heavily administrative. These agents could be a real breakthrough for them.

主持人：确实如此。这也印证了我对未来赢家将出现在哪里的总体判断。就本对话中 Ben 的观点而言，我同意，普通大众并不在乎效率能提高 10%。Host: Totally. That touches on my general thesis about where the winners will be. To Ben’s point, I agree that the general population doesn’t care about being 10% more efficient.

你雇用助理，是希望他们能隐身。 我遇到过最优秀的员工，都是埋头把工作做好、把事情办妥，然后再告诉你：“这是我做的。”你会想：“哇，太棒了。”你雇人不是因为喜欢管理他们。我看到最出色、最主动的应用场景都在金融领域。You hire an assistant so they can be invisible. The best employees I’ve ever had just do their work and get things done, then update you: “Here’s what I did.” You think, “Wow, amazing.” You don’t hire someone because you enjoy managing them. The best, most proactive use cases I’m seeing are in finance.

嘉宾：我认为，在一个重点不再是节省时间、而是节省成本的领域，我们会看到爆发式增长。我见过一些很酷的应用场景：有人用智能体梳理过去一年的收据，并提交 HSA 报销申请。Guest: I think we’re going to see an explosion in a world where it’s not time savers, but cost savers. Some cool use cases I’ve seen: people are using agents to go through the past year of receipts and file HSA reimbursements.

主持人：哦，不错。我会让我的助手在我订好机票后留意票价有没有下降。Host: Oh, cool. One thing I do is have my assistant check whenever I book a flight to see if the price drops.

客人：然后我让它联系航空公司，帮我申请旅行抵用金。Guest: Then I have it contact the airline and get me a travel credit.

主持人：哇。Host: Wow.

嘉宾：这是我目前最喜欢的应用场景：一位朋友把他的 Grockbot 接入了家里的洒水系统，并将它连接到了天气系统。Guest: This is my favorite use case so far: a friend connected his Grockbot to his home sprinkler system and linked it to the weather system.

主持人：真聪明。Host: Smart.

嘉宾：现在他的洒水系统会根据天气情况自动调整，水费因此减少了 50%。Guest: Now his sprinklers respond to the weather, and it’s cut his water bill by 50%.

主持人：这些执行节省成本工作流的隐形代理适用于金融领域，但它们真正会着力于那些既能省钱、又能让世界更加以消费者为中心的活动。Host: These invisible agents doing cost-saving workflows fit into finance, but they’ll really lean into activities that save money and make the world more consumer-aligned.

嘉宾：我觉得美国消费者不想听省钱的事。人们想听的是花钱和拿到免费的钱。Guest: I think the American consumer doesn’t want to hear about saving money. People want to hear about spending money and getting free money.

主持人：我觉得，能打动很多人的卖点在于，这感觉就像白捡的钱。回头提交 HSA 报销申请，就等于白捡钱。机票价格下跌后拿回差价，也等于白捡钱。这是一个极具吸引力的卖点。有趣的是，这些行为发生得不频繁，但价值很高。谈到旅行和金融，我就会想到：这类业务或许能发展成实力强劲、价值数十亿美元的垂直领域连接器，而不是面面俱到的通用型横向代理。Host: I think the hook that gets a lot of people will be that it feels like free money. Going back and submitting HSA refund requests is free money. Getting a refund for the difference when airline ticket prices drop is free money. That’s a hugely powerful hook. It’s also interesting because these are infrequent but high-value behaviors. That’s where my head goes with travel and finance: maybe these can become powerful, billion-dollar businesses that are vertical connectors rather than general, horizontal agents.

嘉宾：完全同意。看看应用阶段会发生什么，会很有意思。我们聊了智能体能做什么，但你究竟该如何与智能体互动？我们俩平时都用 iMessage，所以在那儿交流对我们来说更方便；但你提到，你妻子和她的朋友们更喜欢用一个独立的应用。你觉得未来会如何发展？多种交互界面能否都获得成功，还是最终会由一种方式占据主导地位？Guest: Totally. It’ll be interesting to see what happens in the application phase. We’ve talked about what agents can do, but how do you actually interact with one? We both live in iMessage, so it’s easier for us there, but you mentioned that your wife and her friends prefer a separate application. Where do you see this going? Can multiple surfaces succeed, or will there be one dominant lane?

主持人：这是个很好的问题。消费者偏好在这里会起到很大作用。我听说有一种说法是，Z 世代更喜欢发短信，而千禧一代和 X 世代则更喜欢应用程序界面。所以其中一部分可能会因世代不同而分化。Host: That’s a great question. Consumer preference is going to matter a lot here. I’ve heard the argument that Gen Z is more interested in texting, while millennials and Gen X are more interested in app interfaces. So part of it may fragment by generation.

嘉宾：在某种程度上，这方面也可能因性别而有所不同。这只是个笼统的概括，但我做的事情和我妻子做的事情不一样。她很喜欢制作梦想板，并为实现一年期目标和人生目标而努力，所以她非常喜欢可视化界面。对她来说，旅行很注重视觉体验；而对我来说，旅行更偏向功能性：“我怎么从 X 到 Y？”理想情况下，我会在飞机上用 Starlink，这对我来说更合适。Guest: Part of it may fragment by gender as well, to some extent. That’s a generalization, but the things I do are different from the things my wife does. She’s big on building a dream board and working toward her one-year and lifelong goals, so she really prefers a visual interface. Travel is very visual for her, whereas for me it’s more functional: “How do I get from X to Y?” Ideally, with Starlink on my plane, which works better for me.

主持人：iMessage 的空间更加珍贵。Host: The iMessage space is more precious.

访客：当我处在那个空间里时，感觉互动更私人一些。当然，如果每隔一条对话就有一条是和智能体进行的，这种感觉可能会改变。眼下，它在我手机上像是一个专属空间，而应用网格里的应用则没有这种感觉。你怎么看？Guest: It feels like a more personal interaction when I’m in that space. Of course, that might change if every second thread is with an agent. For now, it feels like a privileged place on my phone, while an app on the app grid doesn’t. What do you think?

主持人：我完全同意。iMessage 的个性化体验源于我本来就常用它，所以感觉它和我息息相关，而不是额外增加了一套工作流程。Host: I completely agree. The personalization of iMessage comes from the fact that I already live there, so it feels connected to me. It doesn’t feel like an added workflow.

嘉宾：我主要看到两种服务逐渐兴起：iMessage，或者一款独立应用。我还看到一家名为 Sky 的初创公司提出了第三种选择，尝试以 iPhone 小组件作为另一种交互入口，这相当有意思。然后就在昨天，扎克伯格和 Muse 团队发布了 Muse Charm，这是一种与智能体交互的全新方式。你怎么看硬件？Guest: I’ve mostly seen two services emerge: iMessage or a separate application. I’ve also seen a third option from a startup called Sky, which is trying to introduce an iPhone widget as a different surface. That’s pretty interesting. Then yesterday, Zuckerberg and the Muse team released Muse Charm, which is an entirely different way to interact with an agent. What do you think about hardware?

Muse 吊坠：Meta 的数据布局Muse charm as Meta's data play

主持人：我觉得这很有意思。曾经有个聪明人说过，每个做硬件产品的人都觉得自己的产品是终极之选：吊坠才是终极形态，或者表带才是终极形态。我不这么认为。我觉得最终还是取决于个人偏好，就像有些人喜欢戴项链，有些人则喜欢戴耳环。这种产品形态会因用户而异，但拥有一件能始终陪伴在身边、默默记录并捕捉一天中发生的背景信息、请求、承诺、约定以及其他一切事情的设备，确实很有价值。Host: I think it's fascinating. A smart person once said that everyone working on a piece of hardware thinks theirs is the be-all and end-all: the pendant is the ultimate thing, or the watch strap is the ultimate thing. I don't think that's right. I think it will come down to personal preference, just as some people like to wear necklaces and others like earrings. This form factor will be customer-specific, but there’s value in having something ambient that stays with you and captures the context, requests, promises, commitments, and everything else that happens throughout the day.

我觉得，作为一个社会，我们得共同摸索隐私方面的预期。你去华盛顿的餐馆，他们会把你的手机收走。也许他们还会把你的手机和吊坠都收走。不过我觉得硬件领域确实有些潜力，而且我很喜欢我们正朝着这个方向大胆探索。过去十年里，创业公司做硬件一直不受鼓励，如今我们看到许多有趣的新消费产品尝试。I think we'll have to figure out privacy expectations together as a society. You go to restaurants in Washington, and they take away your phone. Maybe they'll take away your phone and your pendant, too. But I think there's something there with hardware, and I like that we're being ambitious in that direction. After ten years of startups being discouraged from working on hardware, we're seeing a lot of interesting new consumer efforts.

嘉宾：我觉得这会非常有意思。我有个大胆的观点：Muse 智能挂饰与其说是为了在硬件竞争中胜出，不如说是为了替 Meta 团队收集现实世界的数据，了解现实世界究竟如何运作。它配有摄像头和多个麦克风，我不太认为它会成为一款大众普及型产品。我觉得 Twitter 圈子里的人会率先采用它，而且会带着它四处使用。这会为 Meta 提供海量数据，帮助扎克伯格通过绘制真实世界的地图，为未来的元宇宙提供动力。Guest: I think it's going to be super interesting. I have a hot-take thesis that the Muse charm is less about winning the hardware game and more about collecting real-world data for the Meta team: how the real world actually functions. It has a camera and multiple microphones, and I don't really see it becoming a mass-adoption product. I think it'll be adopted by the Twitter bubble, and people will use it everywhere. That will feed Meta a ridiculous amount of data to fuel Zuckerberg's future metaverse by mapping the actual world.

主持人：但你不觉得他们已经能从眼镜上看到这些了吗？Host: But don't you think they already get that from the glasses?

嘉宾：在某种程度上，是的。但眼镜会一直像这个护身符一样戴在身上吗？Guest: To some extent, yes. But are the glasses always going to be on in the same way this charm will be?

主持人：我不这么认为。这个观点很有道理。我觉得这款吊坠是为了成为全天候、无时无刻陪伴在侧的伙伴，而眼镜则更偏向于行动导向。Host: I don't think so. That's a great point. I think the charm is meant to be an ambient, 24/7 companion, while the glasses are more action-oriented.

嘉宾：这会很有意思。我有机会体验了 VR 版本，感觉很新鲜，但我其实觉得纯音频眼镜可能会成为一匹黑马。这种东西确实很有吸引力。《安德的游戏》中的简就体现了这种全知、纯音频伴侣的愿景。有些人喜欢在脸上戴着摄像头，但也有人可能意识到，这会让别人感到不自在。Guest: It'll be fascinating. I got to try the VR release, and that was interesting, but I actually think the audio-only glasses might be a sleeper hit. There's something really compelling about that. Jane from Ender's Game embodies this vision of an all-knowing, audio-only companion. Some people love wearing a camera on their face, but others may be aware that it could make people uncomfortable.

主持人：而且他们会喜欢既能享受技术和连接带来的便利，又不用面对摄像头带来的社交尴尬。Host: And they'll like having the technology and connectivity without the social awkwardness of the camera.

嘉宾：音频体验令人惊叹。我非常看好语音和音频。ChatGPT 语音昨天新增了连接器，因此你可以将它连接到 Gmail 和日历等服务。我今天早上试用了一下，这是我在 AI 助手领域经历的第二个神奇时刻。Guest: Audio is incredible. I'm quite bullish on voice and audio. ChatGPT Voice added connectors yesterday, so you can connect it to things like Gmail and Calendar. I played around with it this morning, and it was my second magic moment in the AI assistant space.

我骑车去上班时，双手忙着骑车，但又想把事情办好。于是我打开 ChatGPT 语音会话，开始和它交谈。在 30 分钟的骑行途中，它帮我给邮件分类、添加不同的标签、回复邮件，还发送了日历邀请。等我到办公桌前时，收件箱已经清空了。那种感觉简直太棒了。I was biking to work, with my hands busy, and wanted to get things done. I opened a ChatGPT Voice session and started talking to it. Over my 30-minute bike ride, it categorized my emails and put them under different labels, replied to emails, and sent calendar invites. By the time I got to my desk, my inbox was at zero. It was so electric.

主持人：太不可思议了。这是 ChatGPT Voice，不是 Muse，对吧？Host: That's incredible. And this was ChatGPT Voice, not Muse?

嘉宾：没错，ChatGPT 语音。Guest: Correct, ChatGPT Voice.

主持人：太棒了。我误会了，还以为你说的是 Muse Voice。我完全同意。ChatGPT 基于 GPT Live 模型，支持全双工语音，还能查看你的所有对话串。你可以问它：“我启动的那三个项目是什么？进展到哪儿了？我卡在哪儿了？”这是 Muse 和 Instinct 目前都不具备的非凡能力。不过 Zuck 确实宣布了全双工语音，所以也许 Muse 很快就会推出这项功能。Host: Amazing. I misunderstood and thought you said Muse Voice. I totally agree. ChatGPT is based on the GPT Live model; it's full-duplex voice and can see all your threads. You can ask it, “What are those three projects I started? Where are we, and what am I blocked on?” It's an extraordinary capability that neither Muse nor Instinct has today. Though Zuck did announce full-duplex voice, so maybe it's coming to Muse soon.

主持人：确实，这一天会到来。我确实认为，对普通消费者来说，助理在你忙着做事时最能派上用场。你不会坐在沙发上，把所有时间都花在让助理帮你做事上。你在园艺或做饭时，它最有用，因为那时你的双手正忙着，没法操作界面。这又回到了“隐形助理”的概念：你只要说句话，它就会在后台运行。Host: Definitely, it's coming. I do think that, for the general consumer, an assistant is most helpful when you're occupied. You're not going to sit on your couch and spend all your time using an assistant to do things. It's most useful when you're gardening or cooking, when your hands are busy and you can't engage with an interface. It goes back to the idea of an invisible assistant: you can just say something, and it runs in the background.

我还认为，科技圈这个群体——那些长期泡在网上、总是盯着手机或电脑的人——与普通大众非常不同。I also think the tech bubble—people who are chronically online and always on their phones or in front of a computer—is very different from the general population.

嘉宾：我很期待看到它如何获得大规模普及，以及普通消费者会如何看待助手。它能解决实际的痛点吗？当人们遇到一件特别麻烦、不想亲自处理的事时，会用它吗？比如你收到了必须缴纳的交通罚单，只要拍张照片，就可以让你的智能体帮你处理。看看以后会怎样，应该会很有意思。Guest: I'm excited to see how this reaches mass adoption and how the average consumer thinks about an assistant. Will it solve real pain points? Will people use it when something is really painful and they don't want to do it? If you get a traffic ticket that you have to pay, you could just take a picture and let your agent handle it. It'll be interesting to see.

群聊中的静默代理Silent agents in group chats

主持人：接下来我想聊聊社交互动。我们之前比较笼统地谈过多人游戏，以及 iMessage 是一个得天独厚的平台——这也是 Instinct 的亮点之一。但据我所知，把智能体拉进群聊可能会让人觉得被打扰，或者觉得这会削弱社交资本。你见过哪些效果不错的做法吗？你认为智能体会如何发展到多人互动的状态？Host: The next thing I wanted to talk about is social interaction. We’ve talked about multiplayer more generally and how iMessage is a privileged place to be, which is one of the cool things about Instinct. But in my experience, adding agents to group chats can feel intrusive, or like it diminishes social capital. Have you seen anything that works well, and what are your theories about how agents get to a multiplayer state?

嘉宾：很多人都尝试过，我也见过三种不同的做法。第一种是传统做法：把一个智能体拉进群聊，它的表现和一对一对话时差不多。第二种是 Instinct 的做法：你拥有一个智能体网络，可以邀请另一个智能体加入，智能体们会在幕后替你沟通，然后向你汇报进展。Guest: A lot of people have tried it, and I’ve seen three different approaches. The first is the traditional one: add an agent to a group chat, where it behaves much like it does in a one-on-one conversation. The second is what Instinct does: you have an agent network where you invite another agent, and the agents talk behind the scenes on your behalf, then update you on what’s going on.

第三种方案，我觉得最有意思，是 Shane Mack 推出的一款名为 Doc 的新产品。我认为它是 A16Z 旗下 XMTP 的公司。他们采取的方式更像 Granola：你把一个代理添加到群聊中，但它会保持沉默，只负责倾听。它会记录大家讨论的内容，然后如果有待办事项，就会在单独的对话线程中主动联系相关人员。The third approach, which I think is the most interesting, is a new product called Doc by Shane Mack. I think it’s an A16Z company, XMTP. They’re taking more of a Granola-like approach: you add an agent to a group chat, but it stays silent and simply listens. It takes notes on what people are discussing, then independently messages someone in a separate thread if there’s an action item.

你可以查看实际文档，看看它记录了哪些内容。它就像你的隐形代理：负责处理你希望多人协作功能完成的行政事务；如果发现待办事项，也不会打扰其他组员。这种做法真的很酷。You can check the actual docs to see what it’s taken notes on. It’s your invisible agent: it does the administrative work you want from a multiplayer function, and if it finds action items, it doesn’t bother the rest of the group. It’s a really cool approach.

主持人：我得试试那个。我很喜欢——当然，我也很喜欢它已经是 A16Z 旗下的公司。我的看法是，正如 OBO 的 Near 所说，模型在编程这类可验证领域正取得指数级进步，但它们的文字表达质量，甚至待人接物的方式，可能已经停滞不前，甚至有所退步。考虑到现阶段模型的能力，以及我们对社交互动的预期，我认为智能体能为群体的社交联结做出贡献或加以增强的方式并不多。Host: I’ve got to try that. I love it—and I love that it’s already an A16Z company, of course. My theory is that, as Near from OBO said, models are making exponential progress in verifiable domains like coding, but their prose quality and even bedside manner may have plateaued or gotten worse. Given the models as they are today, and our expectations around social dynamics, I don’t think there are many ways for agents to contribute to or enhance a group’s social connection.

相反，我认为智能体在发挥实用功能时会非常有用。我是个收藏爱好者——收藏唱片和其他东西——最近一直在尝试把智能体加进我和其他收藏爱好者的聊天中，用来编目我们的藏品和品味。我还没完全让它运转起来，但至少感觉它确实能为群体增添价值，而不是像个情商低的机器人，在我们不想听的时候喋喋不休。Conversely, I think agents can be very useful when they’re utilitarian. I’m a collector—I collect records and other things—and I’ve been experimenting with adding an agent to chats with fellow collectors to catalog our collections and tastes. I haven’t quite made it work yet, but it at least feels genuinely additive to the group, rather than like a low-EQ robot chattering when we don’t want it to.

嘉宾：没错。你应该试试 Doc。我想你会对它着迷的。Guest: Totally. You should give Doc a try. I think you’re going to be fascinated by it.

主持人：我会的。这周我还听到一种有趣的观点：消息交流可能并不适合智能体交互。它会让人觉得麦克风被某个人拿着，而把发言机会让给智能体会让人不自在。Host: I will. Another interesting take I heard this week is that messaging may not lend itself to agent interactions. It can feel as if someone has the mic, and it’s uncomfortable to give the agent a turn.

嘉宾：在更接近 Discord 或传统网页表单的平台上，智能体可以发布内容，而你可以像在 Reddit 上那样选择是否与之互动。不知为何，这种方式让人觉得没有群聊那么打扰。Guest: On something closer to Discord or a traditional web form, an agent can make a post, and you can choose whether to engage with it—Reddit-style. For some reason, that feels less intrusive than a group chat.

主持人：这几乎就像是我们太想把这些智能体拟人化了。看看它们的设计，有些公司把智能体设计得像 AI 生成的人类，而另一些公司，比如 Muse，则设计了一个可爱的雪人。Host: It’s almost as if we’re trying to humanize these agents a little too much. If you look at how they’re designed, some companies make their agents look like AI-generated humans, while others, like Muse, have an adorable yeti.

嘉宾：是的。Guest: Yes.

主持人：看到大家有多喜欢雪人怪，真是很有意思。它太可爱、太讨人喜欢了。Host: It’s really interesting to see how much people have fallen in love with the yeti. It’s so cute and adorable.

嘉宾：是的。关于你刚才提到的这一点，我觉得人们对 AI 生成的人物会感到不自在。把这些助手塑造得如此拟人化，感觉很奇怪。Guest: Yes. And to your point, I think people feel uncomfortable with AI-generated humans. It feels weird to humanize these assistants so much.

主持人：在你的私人空间里，这种感觉尤其令人不自在。Host: That can feel especially uncomfortable in your personal space.

嘉宾：还有一个相关的问题：你觉得市场会按个性或原型划分吗？我们散步时还聊到了这个。有些人更喜欢严肃、粗犷、权威的医生；另一些人则更喜欢温暖、友善、让人感觉像同龄人的医生。这是两种不同的原型，而且彼此对立。你觉得智能体也会是这样吗？Guest: One related question: do you think the market will segment by personality or archetype? We were chatting about this on our walk. Some people prefer a serious, gruff, authoritative doctor, while others prefer a warm, friendly doctor who feels like a peer. Those are two different archetypes, and they’re in opposition. Do you think the same will be true of agents?

主持人：我认为每个人肯定都更喜欢不同的沟通风格。Host: I think everyone definitely prefers a different style of communication.

嘉宾：Poke一开始就势头很猛，而且个性相当鲜明。我有很多朋友都说：“我绝不给Poke任何个人信息”，因为它实在太毒舌了，让他们觉得和它互动很不自在。Guest: Poke came out of the gates hot, and its personality is quite distinctive. I have a lot of friends who say, “I refuse to give Poke any of my personal information,” because it’s so sassy that they feel uncomfortable engaging with it.

主持人：但我不认为个性真的是一种站得住脚的竞争优势，因为它很容易配置。作为消费者，你可以给自己的智能体发消息，告诉它你希望它如何回应；它可以把这些偏好保存到记忆或灵魂文件中，然后据此改变行为。我认为我们最终会进入这样一个阶段：无论哪个通用智能体成为领头者，它都将高度可配置，让每个人都能定制它与自己交流的方式。Host: But I don’t think personality is really a defensible mode, because it’s quite easy to configure. As a consumer, you can text your agent and tell it how you’d prefer it to respond; it can save that to its memory or soul file, and then change its behavior. I think we’ll get to a place where whichever general agent becomes the leader is highly configurable, so each person can tailor how it speaks to them.

嘉宾：我同意你的看法，不过让我尽可能有力地论证一下，为什么我们需要某种超越个性的东西：姑且称之为一种内在禀性吧。它的一个鲜明特征或许是擅作主张。我们之前谈过，有些工作需要智能体非常敢于自作主张——Guest: I agree with you, but let me steelman the case for something beyond personality: a kind of constitution, for lack of a better word. One defining characteristic might be presumptuousness. We’ve talked about how there are some jobs where you want an agent to be highly presumptuous—

主持人：——只要它经常能把事情办成，我宁愿它有10%的时候会犯错。其他智能体，比如财务智能体，就应该格外谨慎。所以我在想，我们描述为个性的某些东西，实际上是否归结为能力：智能体面对模糊情况时如何作出取舍。或许这方面更有护城河。Host: —and would rather have it make mistakes 10% of the time, as long as it often gets things done. Other agents, like a finance agent, should be highly cautious. So I wonder if some things we’re describing as personality actually come down to capabilities: how an agent breaks ties when faced with ambiguity. Maybe there’s more defensibility there.

主动性才是真正的护城河Proactivity is the real moat

主持人：我认为，主动性会带来巨大的竞争壁垒。它将把赢家和其他竞争者区分开来：那些能够隐形地充当助理、甚至不必等你开口就开始替你办事的智能体。迄今为止，我们见过最棒的“噱头”或“顿悟时刻”，就是智能体帮你办理航班值机。Host: I think there’s massive defensibility around proactivity. It’s going to separate the winners from the rest of the pack: the agents that can be invisible assistants, starting to do things on your behalf without you even asking. The best “gimmick” or “aha” moment we’ve seen so far is an agent checking you in for a flight.

主持人：归根结底可以分为两类。有些工作流程中，你希望保留控制权，不希望智能体过于主动地替你行事。Host: It really comes down to two categories. There are some workflows where you want to retain control and don’t want an agent acting too proactively on your behalf.

嘉宾：是的。Guest: Yeah.

主持人：这些工作流需要你采取行动。比如，某个智能体发现你的保险费用太高，认为你应该换一家保险公司。Host: Those are workflows that require a change in action on your side. Maybe an agent identifies that your insurance is costing too much and thinks you should switch providers.

嘉宾：是的。Guest: Yeah.

主持人：它在进行这种切换之前，可能应该先征求你的许可，因为这是代表你采取的直接行动，会对你产生影响。Host: It should probably get your permission before making that switch, because that’s a direct action on your behalf that affects you.

嘉宾：不过，也有一些主动采取的行动，比如“我已经替你起草好了这封邮件”，或者“我已经替你申请到了航班抵用金”。Guest: However, there are proactive actions like, “I drafted this email for you,” or, “I got a flight credit on your behalf.”

主持人：你这边不需要采取任何行动。这只是帮你省钱或节省时间。看看我们如何对这类行为进行分类和拆解，会很有意思，但这其中的界限很微妙。我觉得，只要越过一次界限，就会立刻失去用户的信任。Host: There’s no action needed on your side. It’s just saving you money or time. It’ll be interesting to see how we categorize and break down those kinds of behaviors, but it’s a fine line. If you cross it once, I think you immediately lose your user’s trust.

嘉宾：是啊，这会很有意思。我们内部开玩笑说，再过几天，可能就会有智能体给某人发消息说：“我注意到你没那么喜欢她，所以就替你跟她分手了。”Guest: Yeah, it’ll be interesting. We were joking internally that we’re days away from an agent messaging someone, saying, “I noticed you weren’t that into her, so I went ahead and broke up with her for you.”

主持人：对。Host: Yep.

嘉宾：这事肯定会发生的，老兄。很多魔力都来自这种大胆冒进。如果你没看到大家在 X 上发帖讲自己的智能体做了什么，那可能说明这些智能体在这个方向上还不够积极。我也很好奇 Bong Chang 那条帖子到底是不是真的。不知道这事的人可以了解一下：有人发推说，Instinct 帮他办理了航班值机，却把他的中间名幻觉成了 Bong Chang，结果他没能登上飞机。我不确定那只是条帖子，还是确有其事，不过我很想知道那个人现在还在不在用 Instinct。Guest: It’s going to happen, man. A lot of the magic comes from the presumptuousness. If you didn’t see people posting on X about the things their agents did, it would probably be a sign that the agents aren’t pushing hard enough in this direction. I’d also be curious to know whether the Bong Chang post was even real. For folks who don’t know, someone tweeted that Instinct checked him into a flight and hallucinated his middle name as Bong Chang, so he couldn’t get on the flight. I’m not sure if that was just a post or if it really happened, but I’d be curious to know if that person is still using Instinct.

主持人：我想应该是这样，单看那篇帖子的反响就能这么推测。他们可能还想要更多类似的内容。Host: I’d assume so, purely given the traction of that post. They’re probably looking for more of that.

嘉宾：如果别无他用，至少还能作为搞笑热门推文的素材。Guest: If nothing else, at least as a source of comedic, popular tweets.

主持人：没错。一件大事是，OpenClaw时期开发出的基础构件如今已经产品化，成为大众消费者能够理解并认可的产品。Host: Right. One of the big things that’s happened is that the primitives developed during the OpenClaw period have now been productized in a way that the mass-market consumer can digest and appreciate.

嘉宾：还有你提过几次的 Poke，表现也非同凡响。我觉得他们做得非常聪明；要是晚六个月起步，说不定最终就会成为赢家。团队真的很有才华，我很好奇他们接下来会推出什么。既然他们现在已经加入 Cognition，我相信他们并没有停止在这方面的研发。你觉得，OpenClaw 时代积累的哪些能力或经验，可能预示着消费级智能体接下来会如何发展？Guest: And Poke, which you’ve mentioned a couple of times, was extraordinary. I think they did a super clever job, and if they’d started six months later, they might have ended up being the winner. The team is really talented, and I’m curious to see what they have in store next. Now that they’re part of Cognition, I’m sure they haven’t stopped working on this. Are there capabilities or lessons from the OpenClaw era that you think might tell us what’s coming next in consumer agents?

主持人：说实话，我现在觉得消费者代理就是 OpenClaw 的复刻版，只不过预先配置好了。在个人助理领域，我看不出有什么事是 OpenClaw 或 Hermes 做不到的。Host: Right now, I honestly see the consumer agent as a replication of OpenClaw, just preconfigured. I don’t see much happening in the personal assistant space that you can’t do with OpenClaw or Hermes.

嘉宾：只是你不必配置所有东西，也不会每隔七小时就弹出错误消息，而且真的很容易使用。界面很简单。我们看看会发生什么。我认为这个领域会继续朝着更主动、更专业化的方向发展。这也呼应了你关于垂直领域初创公司的论点：随着我们看到越来越多长尾机会，这个领域会是什么样子——Guest: It’s just that you don’t have to configure everything, you don’t get error messages popping up every seven hours, and it’s really easy to use. It’s an easy interface. We’ll see what happens. I think the space will keep moving toward proactivity and greater specialization. That connects to your thesis about narrow startups: what does the space look like as we see more long-tail opportunities, where—

主持人：我不会做一个通才，而是要把所有时间都花在打造一个助手上，让它成为三岁以下孩子的单亲妈妈领域的专家。它无所不知：妈妈如何打理日常生活、孩子如何成长发育，了解得如此透彻，以至于让人觉得这个助手能读懂她的心思。也许它可以成为通才助手的一个插件。既然这是你的论点，你怎么看？首先，能否进一步谈谈你所说的垂直领域初创公司是什么，以及你认为它们会如何发展？Host: Instead of being a generalist, I’m going to spend all my time making my assistant an expert for single mothers with children under the age of three. It knows everything about how a mother manages her day-to-day life and child development, to the point where it feels like the assistant is reading her mind. Maybe that could be a plugin for a generalist. Given that this is your thesis, how do you think about it? First, give us more insight into what narrow startups are to you, and how do you see that playing out?

嘉宾：小众创业公司的理论是，你现在可以为极少数人打造极具雄心、价值极高的软件。我们已有先例：有人每月支付 200、250 甚至 300 美元。因此，你可以只服务数万人，就打造一家年化营收达 1 亿美元的企业，而这在过去基本上从未出现过。产品的功能也可以做到惊人的专业化程度。Guest: The theory with narrow startups is that you can now build software that’s incredibly ambitious and valuable for a very small number of people. We have the precedent of people paying $200, $250, or $300 a month. So you can build a $100 million run-rate business with tens of thousands of people, which has never really existed in the past. The capabilities can also be specialized to an astounding degree.

我完全赞同你的描述。我认为，比较优势将建立在品味与专有知识的结合之上。就像你可能会找一位遵循特定助产风格的助产士一样，你也可能会找一位秉持某种特定思想流派、风格或一套尚未被充分理解的经验之谈的个人智能体。你会雇用这位智能体为你提供服务。此外，还有专有数据——例如，你提到可以专门为纽约市配备一位个人智能体。I totally agree with what you’re describing. I think there will be a comparative advantage based on a combination of taste and proprietary knowledge. Just as you might have a doula who follows a certain style of helping women give birth, you might have a personal agent that subscribes to a particular school of thought, style, or set of lessons that aren’t well understood. You’d hire that agent to provide that service to you. Then there’s proprietary data—for example, you mentioned potentially having a personal agent just for New York City.

主持人：就像可能会有一位品味极佳、难以复制的礼宾人员，熟知所有你会喜欢的好去处；也可能还会有十几位，分别服务于不同的人群或类型。我确实认为，专业化既会有主观的一面，也会有客观的一面。后续发展会很值得关注。Host: Just as there might be a concierge with A+ taste that’s hard to replicate and knows all the best spots you’re going to love, there could be another dozen for different people or archetypes. I do think there’ll be both subjective and objective specialization. It’s going to be interesting to follow.

我也很好奇，哪怕只是从你的角度来看，因为我注意到我们在对话中一直在“助手”和“代理”这两个称呼之间切换。I’m also curious, even just from your perspective, because I noticed that in our conversation we’ve been switching between “assistant” and “agent.”

嘉宾：这一直是大家持续讨论的话题，我也看到有人在 Twitter 上谈论它。Guest: It’s been a constant conversation, and I see people talking about it on Twitter.

主持人：这个生态系统的未来会是个人助理还是个人代理？这两个术语对你来说意味着什么？Host: Is the future of this ecosystem a personal assistant or a personal agent? What do those terms mean to you?

嘉宾：是的。Guest: Yeah.

助手与智能体：概念界定Assistant vs agent, defined

嘉宾：这是个非常好的问题。对我来说，助手比智能体低一个层级。智能体可以成为社交群体中的平等成员。Guest: That’s such a great question. To me, an assistant is a lesser form than an agent. An agent could be a peer in a social group.

主持人：也许我们需要一个比“代理”更贴切的词。顾名思义，助手只能做你要求它做的事。它可以协助你完成你指示它去做的事情。Host: Maybe we need a better word than “agent.” An assistant can, by definition, only do what you ask it to do. It can assist you with something you’ve directed it to do.

嘉宾：智能体具有自主行动能力，可以直接促成世界上的好事发生。Guest: An agent has agency and can just make good things happen in the world.

主持人：这或许与企业中智能体的发展过程类似。它们从实习生做起，通过承担工作、不断变得更智能，逐渐晋升去负责更高阶的工作。那么，对助手，乃至后来的智能体来说，在消费者的生活中获得晋升，开始承担更专业、更精细的工作，意味着什么？Host: Maybe this parallels what’s happening with agents in the enterprise. They start as interns, and through doing work and becoming more intelligent, they get promoted to higher-order forms of work. What does it mean for an assistant, and later an agent, to be promoted in a consumer’s life to do more specialized and delicate work?

嘉宾：我非常认同这一点。我也很关注这对人类社会层面的影响。“智能体”这个词，是否能让我们避免过度拟人化 AI？“助手”这个词是否又给这种体验赋予了太多的人性，以至于让社会感到不自在？看看人们会如何看待这个领域的用语，会很有意思。Guest: I’m very aligned with that. I’m also interested in the ramifications for the social layer of humanity. Does the word “agent” allow us not to personify AI enough? Does “assistant” add too much humanity to the experience, making society uncomfortable? It’ll be interesting to see how people think about the language of this space.

主持人：我们都身处各自的圈子里。你所在的圈子更大，因为你在旧金山——世界的科技中心。我觉得纽约稍微落后于旧金山，而世界其他地方随后才会赶上。看看人们究竟如何看待这种语言，会很有意思。Host: We’re in bubbles. You’re in a bigger bubble because you’re in San Francisco, the tech hub of the world. I’d say New York is a little behind San Francisco, and then the rest of the world catches up. It’ll be interesting to see how people actually think about the language.

嘉宾：我不知道这是不是泡沫。我觉得这是未来。我们正身处不远的未来。Guest: I don’t know if it’s a bubble. I think it’s the future. We’re in the near future.

主持人：嗯。Host: Yeah.

嘉宾：其实我觉得，这也能大大改善许多社交互动。Guest: I actually think it could improve a lot of social dynamics, too.

主持人：代理是非人类的，因此它可以提供一层社交缓冲。有些话我可能很难对你说，或者群体中存在令人不自在的社交氛围。经过精心设计的代理可以帮助提供反馈或协调冲突，让整个过程显得不那么情绪化，因为你知道——Host: The agent is non-human, so it can provide a layer of social indirection. There might be something that’s awkward for me to say to you, or an uncomfortable social dynamic in a group. With a carefully designed agent, it could help provide feedback or organize a conflict in a way that feels non-emotional, because you know that—

客人：像是对抗的仲裁者？Guest: Like an arbiter of confrontation?

主持人：有可能。我并不是在提倡一种更消极攻击式的社交体系，但我确实觉得很有意思：智能体可能既是社交迂回表达的来源，也是其指向的对象——如果这么说说得通的话。Host: Potentially. I’m not advocating for a more passive-aggressive social system, but I do think it’s interesting that agents could be a source and a destination for social indirection, if that makes sense.

主持人：我认为这会对世界产生显著的积极影响。你觉得五年后，人工智能会如何影响我们？这些智能体和助手会把我们带向何方？它们如何让我们成为更快乐的人？Host: I think this will have significantly positive influences in the world. How do you see AI impacting us five years down the line? Where do these agents and assistants take us? How do they make us happier human beings?

嘉宾：这是个很好的问题。我确实认为，这些正是我们真正想解决的问题，也是问题的根源。我们谈过生产力、财务和健康，还有很多日常琐事。昨天我查了一下：这个国家每年有150万人因酒驾被捕。想想这可能会带来多大的行政负担，以及这些人中有多少能聘请律师。我认为，普通消费者正承受着某种官僚体系带来的压力。第一步就是想办法减轻这种压力，帮助人们优化财务状况和健康状况，并让他们也能像富人那样获得律师或类似家族办公室的服务。Guest: It’s a great question. I do think those are the problems we’re really after, at the root of things. We’ve talked about productivity, finances, and health. There are a lot of mundane things. I was looking this up yesterday: 1.5 million people a year in this country get a DUI. Think about the administrative burden that probably creates, and how many of those people have access to attorneys. I think there’s a kind of bureaucratic pressure on the average consumer. Step one is figuring out how to relieve that pressure, help people optimize their finances and health, and give them the same access to an attorney or a kind of family office that a wealthy person might have.

主持人：接下来还有一个更大的问题。我从小就听母亲说：“得到自己想要的东西，最难的部分是知道自己想要什么。”我们每个人真正追求的是什么？这些助手又能如何与我们互动？我不认为这能简单归结为目标。我不知道到底有多少人真正有目标。很多人都在努力寻找自我，成为最优秀、最真实的自己。我们自身的成长、对世界的主观体验与这些机器之间的相互作用，或许能引领我们走向一个美好的未来。Host: Then there’s the bigger question. I grew up with my mother saying, “The hardest part about getting what you want is knowing what you want.” What are we all truly after, and how can these assistants engage with us? I don’t think it’s as simple as goals. I don’t know how many people even have goals. A lot of people are trying to find out who they are and become the best, most authentic versions of themselves. The interplay between our own self-development, our subjective experience of the world, and these machines could lead to a beautiful future.

我憧憬着这样一个未来：我的孩子们看着记录当代人走在街上的旧视频——他们低着头，脖子都快伸直了，盯着手机看。他们会问：“你们在干什么？外面的世界多美啊。为什么整天都盯着屏幕看？”I envision a world where my future kids look at old videos of the current generation walking down the street, necks craned down, staring at their phones. They’ll ask, “What are you doing? It’s a beautiful world out there. Why are you spending all your time staring at a screen?”

嘉宾：希望人工智能能让我们实现许多事情的自动化，这样身为父母的你就能有更多时间陪伴孩子，而二十多岁的成年人也能有更多时间和朋友聊天。Guest: Hopefully, AI will let us automate so many things that, as a parent, you can spend more time with your kids, and as an adult in your 20s, you can spend more time talking to friends.

主持人：我真的很想看看，它如何通过减少那些东西来让生活变得更好——Host: I’m really curious to see how it changes life for the better by reducing the things—

嘉宾：你最讨厌做的生活琐事。Guest: The life-admin work you really hate doing.

主持人：对，没错。这也是我认为编程智能体不一定会和人类程序员竞争的原因。它们做的是人类程序员不会去做的事，比如开发一个荒唐的个人项目。我一直在打造一家可以走进去逛的虚拟唱片店。之后我把链接发给你。它太棒了。五年、七年甚至十年前，我绝不会雇程序员来做这个，也绝不会花上一年时间亲自开发。这些事情本来根本不会发生。Host: Yeah, that’s right. That’s also why I think coding agents aren’t necessarily competing with human programmers. They’re doing things no human programmers are doing, like building a ridiculous side project. I’ve been building a virtual record shop you can walk into. I’ll send you a link afterward. It’s awesome. I would never have hired a programmer to work on that five, seven, or ten years ago, and I would never have spent the year of my life it would have taken. These are things that never would have happened otherwise.

嘉宾：我觉得，这项技术在我们的生活中有很多种应用，让我们能够探索那些原本永远不会去探索的事物。这不是为了降低成本，而是为了拓展可能性。Guest: I think there are many versions of this technology in our lives that let us explore things we never would have explored otherwise. It’s not about cost reduction; it’s about expanding possibility.

主持人：是的。Host: Yeah.

嘉宾：它让生活更有趣、更精彩。Guest: It makes life more fun and interesting.

主持人：回到近期的话题，我一直在思考 Moltbook 带来的经验教训，以及需要构建什么样的原生智能体系统。或许，Instinct 正在通过其 Instinct-to-Instinct 网络做的就是其中一部分。感觉未来会出现一些专为智能体之间交互而存在的基础设施和产品，而这些目前还不存在。你怎么看？Host: To come back to the near term, I’ve been thinking a lot about the lessons of Moltbook and what kinds of agent-native systems will need to be built. Maybe some of that is what Instinct is doing with its Instinct-to-Instinct network. It feels like there will be infrastructure and products that exist only for agent-to-agent interactions, which don’t exist today. What do you think?

嘉宾：这将是一种全新的范式，开启了许多可能性。光是这个话题，就足以聊上十个小时。Guest: It’s going to be a whole new paradigm. It opens up a lot of possibilities. That could be a 10-hour conversation on its own.

主持人：你们有代理人的电子邮箱和电话号码。当服务世界不再是人与人预订，也不是代理人为人预订，而是一个代理人与另一个代理人预订时，会变成什么样？这将如何重新定义整个服务行业？Host: You have agent emails and agent phone numbers. What does the world of service look like when it’s no longer human booking with human, or an agent booking for a human, but an agent booking with another agent? How does that redefine the entire service industry?

嘉宾：然后还有整个网络安全领域。先有了互联网，随后才有网络安全。先有了智能体，现在也到了为这些智能体构建安全体系的时候。这会催生出什么行业？这一切的美妙之处在于，智能体领域有着巨大的机遇，也有许多亟待解决的问题。我们必然会看到许多初创公司尝试摸索这些细分问题，并打造新一波互联网和数字互联浪潮。Guest: Then there’s the whole cybersecurity component. The internet came first, then cybersecurity. Agents came first, and now comes the security component for these agents. What industry will that create? The beautiful part of it all is that there’s so much opportunity and so many problems to solve with agents. It’s inevitable that we’ll see many startups trying to figure out these niche problems and build a new wave of internet and digital connectivity.

主持人：是啊，看看接下来会怎样一定很有意思。我把互联网看作由三个网络组成：商务网络、应用网络和内容网络。内容网络已经陷入困境一段时间了。商务网络发展得非常好，而自从编码智能体发布以来，应用网络则迎来了大幅增长。不过，商务网络确实很耐人寻味。本周，我们看到 Shopify 拥抱 Muse，而亚马逊则封禁了它。企业在这种情况下是赢家还是输家，部分取决于是否出现了一个新的流量入口，将流量引向这些电商平台。你对亚马逊和 Shopify 具体发生了什么有什么看法？更广泛地说，你认为谁最有可能成为赢家，谁又会是输家？Host: Yeah, it’ll be cool to see. I think of the web as three webs within the web: the commerce web, the application web, and the content web. The content web has been in trouble for a while. The commerce web has done really well, and the application web has been on a huge upswing since coding agents were released. The commerce web is fascinating, though. This week, we saw Shopify embrace Muse and Amazon block it. Part of the calculus may be whether companies win or lose when there’s a new front door for traffic flowing to these marketplaces. Do you have a view on what specifically happened with Amazon and Shopify, and more broadly on who the likely winners and losers are?

亚马逊封杀 Muse，Shopify敞开大门Amazon blocks Muse, Shopify opens the door

嘉宾：Shopify 和亚马逊的商业模式截然不同。Shopify 一心致力于让商业更加大众化，而亚马逊很大一部分收入来自广告。商业领域的整个广告体系会如何发展，值得关注，因为一旦没有了人类的目光，整个体系就失去了存在的意义。Guest: Shopify and Amazon have two completely different business models. Shopify is hyperfocused on democratizing commerce, while Amazon makes much of its money from ad revenue. The whole ad structure of commerce will be fascinating to watch, because when you remove human eyeballs, it defeats the entire purpose of the system.

主持人：我认为这就是亚马逊的出发点。如果你让 Muse 和其他智能体开始购物，却不给亚马逊带来任何广告收入——Host: I think that’s where Amazon was coming from. If you let Muse and other agents start purchasing without generating any ad revenue for Amazon—

嘉宾：这会大幅压缩这一体验中的利润空间。在智能体代表人们购物的世界里，公司如何盈利很可能需要一种全新的模式。相比之下，Shopify 赋能个人尽可能多地销售商品，而智能体会进一步推动这一领域的民主化。这是一个巨大的优势。Guest: It takes a lot of the profit out of the experience. There will likely have to be a new paradigm for how companies profit in a world where agents purchase on people’s behalf. Shopify, by contrast, empowers individuals to sell as much as possible, and agents democratize that space even further. That’s an enormous advantage.

看看那些尚未受到影响的行业，会更有意思。上周我们谈到了餐厅和预订领域。Resy 开始封禁使用 Browser Use 的用户账号，因为他们不想要机器人。如果智能体能够瞬间完成预订，就会破坏一个围绕面向真人的预订体验打造的平台的初衷。It’ll be even more interesting to look at industries that haven’t been affected yet. We talked last week about restaurants and the reservation space. Resy started shutting down people’s accounts because they were using Browser Use and Resy didn’t want bots. If agents can make reservations instantly, it defeats the purpose of a platform built around a human-oriented booking experience.

主持人：当每个人都有一个智能体，并处于同一起跑线上时，会发生什么？所有人都可能同时尝试预订同一家餐厅。Host: What happens when everyone has an agent and is on the same playing field? Everyone could try to book the same restaurant at the same time.

嘉宾：我们会进入一个权力重新回到餐厅手中的世界吗？餐厅可以根据顾客的忠诚度、平均消费额或其他因素，决定把订位留给谁。还是说，我们会进入一个竞价的世界，由各个代理相互竞争？我很好奇你觉得未来会是什么样子。Guest: Do we enter a world where power shifts back to restaurants, which can choose whom to give a reservation to based on loyalty, average spend, or other factors? Or do we enter a world of bidding, where agents compete with one another? I’m curious what you think that future looks like.

主持人：你说得对，这非常有意思。有些东西会受到供应限制，比如热门餐厅的餐桌；而另一些则会受到需求限制，比如明天飞往旧金山的航班。在信息完备的情况下，餐厅会希望接待多次光顾、平均客单价和终身价值最高的顾客。他们也会希望尽可能长时间地保持接近满负荷运营。这种计算要复杂得多，我还不清楚具体会如何实现，但专有供给将变得越来越重要，而不是依靠摩擦和消费者的动机来防止这类拒绝服务问题。Host: You’re right; it’s very interesting. Some things will be supply-constrained, like tables at hot restaurants, while others will be demand-constrained, like flights to San Francisco tomorrow. With perfect information, restaurants would want the customers with the highest average order value and lifetime value across multiple visits. They’d also want to stay near full capacity as much as possible. That’s a more sophisticated calculation, and I don’t know exactly how it will work, but proprietary supply is going to become increasingly important, rather than relying on friction and consumer motivation to prevent this kind of denial-of-service problem.

供给受限的商业领域会很有意思，而需求受限的领域则会截然不同。你可以打造一种团购体验：买家已准备好购买某种产品、服务或体验，商家则竞相满足这类需求，提供最佳体验，也许还能给出最低价格。需求聚合一直存在，但我们还没有以同样的方式实现供给聚合，尤其是通过一种能让消费者受益的拍卖机制。Supply-constrained parts of commerce will be interesting, while demand-constrained ones will be completely different. You could have a group-buying experience, with buyers ready to purchase a product, service, or experience and merchants competing to meet that demand with the best experience, perhaps at the lowest price. We’ve always had aggregation of demand, but we haven’t really had aggregation of supply in the same way, especially through an auction system that benefits consumers.

嘉宾：我觉得这真的很有意思。Guest: I think that’s really interesting.

主持人：关于亚马逊和 Shopify 的问题，亚马逊押注自己不想被去中介化，或许是对的，因为它可能失去的不只是广告收入，还有冲动购物。我很好奇，智能体本身在某种程度上是否也会冲动消费。主动性意味着智能体会尝试做出令人惊喜的猜测，而还有什么能比惊喜或礼物更令人愉悦呢？Host: On the Amazon and Shopify question, Amazon is probably right to bet that it doesn’t want to be disintermediated, because it could lose not just advertising revenue but also impulse shopping. I’m curious whether agents themselves will make impulse purchases to some extent. Proactivity means agents will try to make delightful guesses, and what’s more delightful than a surprise or a gift?

嘉宾：一个有趣的动态是，智能体不会完美映照出它们的用户。它们不会完全理性；它们会以全新的方式表现出理性与非理性。我不认为营销、冲动消费或广告会变得无关紧要，它们只是需要转变形式。Guest: One interesting dynamic is that agents won’t be perfect reflections of their users. They won’t be perfectly rational; they’ll be rational and nonrational in completely new ways. I don’t think marketing, impulse purchases, or advertising will stop mattering. They’ll just need to be transformed.

主持人：完全同意。这也开启了一个全新的话题：面向智能体商务的推荐引擎会是什么样子？Host: Totally. That opens up a whole new conversation: what will the recommendation engine for agentic commerce look like?

嘉宾：在优质搜索方面，AI 和大型语言模型似乎把互联网搞坏了。但如果你使用智能体，然后说：“嘿，给我买些白袜子”——Guest: AI and LLMs seem to have broken the internet when it comes to quality search. But if you use an agent and say, “Hey, buy me some white socks”—

主持人：白袜子有 20,000 种不同的款式。代理可能会在不知道你具体喜欢哪种款式的情况下，给你寄来一些东西。Host: There are 20,000 different kinds of white socks. The agent might send you something without knowing exactly what kind you prefer.

嘉宾：是的。Guest: Yeah.

主持人：Muse 在这方面可能已经抢占先机。这正是扎克伯格抨击电商的核心论点：Meta 掌握着你的 Instagram 和 Facebook 数据，因此可能更了解你的实际偏好。看看这一推荐层在电商领域会是什么样子，将会很Host: Muse probably has a head start here. This is the whole thesis behind Zuck’s attacks on commerce: Meta has your Instagram and Facebook data, so it may have a better understanding of what you actually prefer. It’ll be interesting to see what this recommendation layer looks like in commerce.

嘉宾：我觉得这对长尾需求来说可能很有趣。你可以去亚马逊买白袜子，但也许你的智能体会回来告诉你：“嘿，我在 Etsy 上找到了一位很棒的人，她可以给你织袜子。内布拉斯加州有位奶奶织的白袜子特别棒。需要三周时间。你感兴趣吗？”你可能会说：“当然，听起来挺有意思的。”我想知道，这能不能带来更多意外惊喜，因为它可以审视所有选项，并根据你的需求量身定制，而且没有不良动机。Guest: I think it could be fun for the long tail. You could go to Amazon and get white socks, but maybe your agent comes back and says, “Hey, I found someone awesome on Etsy who can knit you socks. A grandma in Nebraska makes amazing white socks. It’ll take three weeks. Are you interested?” And you might say, “Sure, that sounds interesting.” I wonder if it could bring more serendipity, because it can examine all the options and tailor them to you without adverse incentives.

主持人：对。Host: Right.

嘉宾：我很乐意。那会很有趣，也能给如今平淡无奇的购物体验增添一点趣味。Guest: I’d love that. It would be fun and add a little flavor to the mundane experience of shopping today.

主持人：没错。Host: Exactly.

嘉宾：从我的角度来看，这很平淡，因为我并不享受买袜子这件事。你也可以想象一种更具未来感的情景：智能体开始在现实世界中促成各种事情。智能体不再只是找到一位会织白袜子、并在 Etsy 上售卖的奶奶，而是由一个智能体联系她的智能体，问：“你愿意做这件事吗？”这正是智能体对智能体的空间变得非常有趣的地方：它们代表各自的人类采取行动，有些提供产品、服务或知识，另一些则负责赚钱。也许会出现一个次级经济，它比我们如今的经济更具活力、交易频率更高，也有趣得多。Guest: Mundane from my point of view, as someone who doesn’t take joy in sock shopping. And you could imagine something more futuristic, where agents start to catalyze things in the real world. Instead of simply finding a grandma who knits white socks and sells them on Etsy, maybe an agent messages her agent and asks, “Would you be up for doing this?” That’s where agent-to-agent spaces get really interesting: they act on behalf of their humans, some to provide products, services, or knowledge, and others to make money. Maybe there’s a secondary economy that’s more dynamic, higher-frequency, and much more interesting than the one we have today.

主持人：我想象着一位住在内布拉斯加州的奶奶收到一条短信，上面写着：“嘿，旧金山有个小伙子想让你给他织双白袜子。你愿意吗？”为什么不呢？是啊，为什么不呢？Host: I’m imagining a grandma in Nebraska receiving a text message saying, “Hey, this guy in SF wants you to knit white socks for him. Do you want to do it?” Why not? Yeah, why not?

嘉宾：那就太棒了。Guest: It’d be awesome.

主持人：我确实非常乐观。这也能以正确的方式设定激励机制：如果智能体做得对，它们会根据产品质量来选择优秀的产品——它们可以自行评估产品质量——而不是选择营销做得最好或品牌最响亮的产品。如果它们确实按品牌来选择，那也会得到用户的明确认可：“我就是想买耐克，因为我喜欢耐克”，而不是因为营销造成的认知失调，导致消费者做出并非最有利于自己的选择。Host: I’m actually very hopeful. It also sets the incentives in the right way: if agents are doing the right thing, they’ll choose great products based on quality, which they can assess for themselves, rather than choosing whatever has the best marketing or brand. And if they do choose based on brand, it’ll be with the user’s explicit sign-off: “I just want Nikes because I love Nikes,” rather than because of the cognitive dissonance created by marketing, which can lead consumers to make choices that aren’t optimal for them.

嘉宾：没错。这为真正以消费者为中心的互联网打开了大门，而我认为，我们已经缺少这样的互联网很长一段时间了。Guest: Right. It opens the door to a truly consumer-aligned internet, which I think we’ve been lacking for quite some time.

主持人：我同意，老兄。我们之前提到过这一点。Host: I agree, man. We touched on that earlier.

每天 20 美元的智能体经济学The $20/day agent economics

主持人：我看过的 122 个智能体中，有 26 个是我亲自测试过的。Host: Of the 122 agents I’ve looked through, I’ve personally tested 26.

嘉宾：好的。Guest: Okay.

主持人：这 122 个里，有 65 个是付费的。35 个完全付费，另有 30 个采用高级版模式，也就是说，达到使用上限后就得付费。真正免费的只有 13 个。Muse 和 Instinct 都是免费的，而且占据主导地位；据说 OpenAI 即将发布一些很酷的东西——也许就在下周，谁知道呢？Host: Of the 122, 65 are paid. Thirty-five are fully paid, and 30 have a premium model, so you have to pay once you hit a limit. Only 13 are actually free. Muse and Instinct are both free and dominant, and OpenAI is supposedly going to release something cool—maybe next week, who knows?

长尾初创公司究竟如何才能在这个领域胜出？而大型玩家都免费提供时，他们为什么还要收费？How does a long-tail startup actually win in this space? And why are they charging money when the big dogs are giving it to you for free?

嘉宾：天哪，这会非常有挑战性。按我们的估算，要以非常雄心勃勃的方式做到这一点，成本大约是每位用户每天20美元。Guest: Man, it’s going to be really challenging. By our estimate, it costs something like $20 per user per day to do this in a really ambitious way.

主持人：这对一家初创公司来说，可能意味着每年数亿美元。我认为值得押注的两个积极趋势是：浏览器使用成本将呈通缩趋势，并且很快会呈指数级下降。如果你现在正在为初创公司融资，并押注这种情况会在六个月内发生，那可能有点像一场胆小鬼博弈。不过，我们确实看到 token 价格大幅下跌，新兴能力的成本也在下降。Host: That could mean hundreds of millions of dollars a year for a startup. I think the two positive trends to bet on are that browser use should become deflationary and get exponentially cheaper very rapidly. If you’re a startup raising money today and betting that will happen in six months, maybe it’s a bit of a game of chicken. But we have seen token prices fall dramatically, and the cost of emerging capabilities fall too.

嘉宾：我觉得，浏览器使用的成本很有可能会低到让更多产品能够免费提供这项服务。但另一方面，我不太喜欢把补贴客户作为主要价值主张。我很希望看到有一款每月收费 1,000 美元的智能体，让消费者兴奋到愿意为它付费。人们每个月会在许多非必需品上花 1,000 美元，只因为他们想要这些东西。什么样的智能体能让你觉得它足够令人兴奋，值得花这个价钱？这样的智能体既能受益于强劲的产品市场契合度，也能受益于成本下降。Guest: I think it’s very possible that browser use gets cheap enough for more products to offer this for free. The flip side is that I don’t love the idea of customer subsidies being the primary value proposition. I’d love to see an agent that costs $1,000 a month, and that consumers are so excited about that they’re willing to pay. People spend $1,000 a month on plenty of things that aren’t necessities, but are things they want. What agent is exciting enough that you’d pay that kind of price for it? That agent would benefit from both strong product-market fit and deflating costs.

主持人：没错。值得期待。Host: Right. It’ll be interesting to see.

嘉宾：太有意思了。来自 Assistant Bench 的 David，大家一定要去看看。非常感谢你来参加我们的节目。我们会经常做这样的交流。我觉得我们每周都可以聊一次；每天都有新鲜事发生。Guest: Super interesting. David from Assistant Bench, make sure everybody checks it out. Thank you so much for joining us. We’re going to do this frequently. I feel like we could do this every week; things are happening every single day.

主持人：请在 X 上关注他，也关注我。我们会在 X 社区中彼此分享最新动态，一起学习。Host: Please follow him on X, and follow me as well. We’ll all keep up to speed and learn together in the X community.

来宾：祝你和你的智能体们快乐每一天。Guest: Happy day to you and your agents.

主持人：谢谢。你也一样。真有意思。Host: Appreciate it. Same to you. This is fun.

嘉宾：好吧，兄弟。谢谢你。Guest: All right, brother. Thank you.

主持人：太棒了。Host: Awesome.