我们使用AI的7种方式正在改变

我们使用AI的方式正在发生变化。

The way we use AI is changing.

一开始，我们是提示工程师，然后是上下文工程师，接着是框架工程师，再后来是循环工程师。

First, we were prompt engineers, then context engineers, then harness engineers, then loop engineers.

虽然把这些都看作是一个接一个的流行词很诱人，但现实是，我们都在同一条路上，一起摸索如何最好地使用这项全新的技术，它既在改变我们现有的工作方式，也在为我们能做什么工作开辟新的可能性。

And while it's tempting to view all of these things as just one silly buzzword giving way to the next, the reality is that we are all together on this journey figuring out just the best way to use a completely new technology that is changing both how we do our current work and opening up new possibilities of what work we can even do.

今天，我们来聊聊我们使用AI的方式正在变化的七个方面。

Today, we're talking about seven ways that how we use AI is changing.

《AI每日简报》是一个每日更新的播客和视频节目，内容涵盖AI领域最重要的新闻和讨论。

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

好了，朋友们，在开始之前，先快速说几个通知。

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

首先，感谢今天的赞助商：Blitzy、Robots & Pencils、Harbor和Hyperagent。

First of all, thank you to today's sponsors, Blitzy, Robots & Pencils, Harbor, and Hyperagent.

想要收听无广告版本的节目，请访问patreon.com/aidailybrief，或者你也可以在Apple Podcasts上订阅。

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

想了解更多关于赞助节目的信息，请给我们发邮件到sponsors@aidailybrief.ai。

To learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai.

今天的节目讲的是很多模式上的变化，我觉得这些变化正在改变我们日常与AI互动的方式。

Today's episode is about a bunch of patterns that are shifts, I think, in how, on average, we interact with AI on a daily basis.

从ChatGPT诞生之初，我们与AI的互动模式就从来没有长时间保持不变过。

One thing that's been true since the very beginning of ChatGPT is that our patterns of interaction with AI have never been static for long.

一开始，人们发现各种新的提示方式，效果有好有坏，然后通过大量的文章、帖子、YouTube视频、播客和课程传播开来。

At the beginning, it was people discovering new ways to prompt the AI that had better or worse results, which was then communicated through hundreds of articles, posts, YouTube videos, podcasts, and courses.

但随着时间的推移，互动模式的变化变得更加显著和实质性。

But over time, the interaction pattern shifts have gotten more significant and more substantive.

在过去一周左右，我看到很多公告，确实强化了这些互动模式中的一些重大转变。

Over the last week or so, I saw a number of announcements that really reinforce some fairly big shifts in these interaction patterns.

虽然我不认为我们使用AI的方式会真正一成不变，但我确实觉得其中一些转变，也就是我将在本期节目中讨论的，是相当根本性的，而且很可能会持续下去，所以如果大家想高效利用时间，这些转变可能更值得关注。

And while I don't think that how we use AI will ever be truly static, I do think that some of these shifts, which I'm going to cover in this show, are fairly fundamental and likely to stick, and thus, of course, if one is trying to be efficient with their time, may be more worthy of attention.

所以首先是一个互动模式，这个模式与其说是我们在改变与AI互动的方式，不如说是产品设计决策在改变我们与AI互动的选项。

So first up is an interaction pattern that is not so much us changing how we interact with AI, but product design decisions that are shifting the options that we have to interact with AI.

很明显，现在大家都在朝着简化、整合和用户体验的方向发展，让最终用户少做决策，而不是多做决策。

Very clearly, there is a push towards simplification and integration and user experiences that have less rather than more cognitive decision-making for the end user.

坐在我这个位置上，我觉得有趣的是，早期的人工智能往往违背了硅谷的传统逻辑，即除非你提供极其简单明了的体验，否则用户不会来。

One thing that's been interesting for me, sitting in this seat that I have, is the extent to which early AI has tended to defy traditional Silicon Valley logic that users won't show up unless you give them an incredibly simple and clear experience.

实际上，恰恰相反，很多早期采用者，不仅仅是科技圈的人，还有更广泛的知识工作者，他们不但没有抗拒复杂性，反而一头扎了进去。

In fact, quite to the contrary of that, a lot of the early adopters, not just from tech circles, but from broader knowledge work circles, have, rather than resisting complexity, thrown themselves in headfirst.

一个很好的例子是，今年早些时候我推出的免费Claw Camp项目，有9000人报名，其中大多数人甚至不是工程师，也不怎么懂技术，他们基本上是在玻璃上爬行，试图弄清楚如何构建自己的智能体。

A great example of this for me was when 9,000 people, most of whom weren't even close to engineers or really technical in any way, signed up for the free Claw Camp program I put out earlier this year, which basically had them crawling through glass to figure out how to build their own agents.

然而，看起来，即使愿意做这种艰苦工作的人在AI领域比过去其他科技领域要多，仍然有很多人想要那种简单性。

And yet it appears that perhaps even if the portion of people who were willing to do that sort of hard work has been higher in AI than in other past tech sectors, there are still a lot of folks who want that simplicity.

当然，简单是新的Meta Muse产品的关键词之一，即使是那些完全有能力处理复杂性的高级用户也在称赞这个产品的这一点。

Certainly, simplicity is one of the watch words for the new Meta Muse product and something that even the power users who fully have the capability of dealing with complexity are lauding in the product.

How I AI的Claire Vo在很多地方都讨论过，她认为Muse的设计考虑得非常周到，特别是关于它向用户隐藏了什么，没有隐藏什么。

How I AI's Claire Vo has discussed in numerous places how well thought out she thinks the Muse design is, specifically around what it does and doesn't obfuscate from the user.

Node.js的创造者Ryan Dahl，他当然也不缺乏处理复杂性的能力，也称赞Muse在简单性方面做得很好。

Node.js creator Ryan Dahl, who again, certainly has no shortage of complexity capability, also shouted out Muse for having nailed the simplicity.

技术专家汤姆·古德温写道：“Muse的妙处在于，他们用一种非常容易上手的方式把它介绍给了大家。

Wrote technologist Tom Goodwin, "The beauty of Muse is that they've managed to introduce it in a very accessible way.

别告诉我你怎么工作的，直接教我怎么做事就行。”

Don't tell me how you work, just show me how to do things."

大卫·赫尔曼写道：“我觉得Muse会赢得休闲AI的竞争。

David Herman wrote, "I think Muse is going to win the casual AI race.

我给我爸看了，他快70岁了，一下子就明白了。

I showed my dad, who is nearly 70, and it clicked.

他特别惊讶，我居然能让它帮我起草一份完整的合同，还能在上面改东西，最后通过短信发给客户。

He was amazed that I could have it create an entire contract for me, update things on it, then email and send it to a customer all through texts.

Claude和OpenAI都很棒，但到目前为止，我觉得真正理解这些模型的人不到1%。

Claude, OpenAI are great, but so far I'd say less than 1% of the population gets what these models are.

Meta让这一切变得超级简单。”

Meta just made it all super easy."

很明显，这是他们有意为之的。

And it's clear that this was intentional.

Meta的首席AI官亚历山大·王写道：“团队在打造Muse时，重点之一就是让它开箱即用。

Meta Chief AI Officer Alexander Wang wrote, "A big part of the team's focus when building Muse was to make something that just worked.

为了让产品简单易用，团队在细节上下了很多功夫。”

So much work went into all the little details to make the product simple and easy to use."

当然，让这成为一种趋势，而不仅仅是Meta团队的设计决策的，是这种简化正在各处出现。

And of course, the thing that makes this a trend, not just a design decision from the Meta team, is that this sort of simplification is showing up everywhere.

这周最大的产品更新之一来自Anthropic，他们宣布Claude Cowork和Chat不再分开。

One of the biggest product updates that we got this week came when Anthropic announced that Claude Cowork and Chat are no longer separate things.

Claude正在走向一种体验，不再是碎片化的、用户还没开始就要做很多选择，而是统一的单一体验。

Claude is careening towards an experience where instead of it being fragmented and full of decisions for the user before they even get started, instead a single unified experience.

而且，关键词还是简单、简化。

And again, the watch word was simplicity, simplification.

顺便说一句，正如Claude Code的创造者Kat Wu指出的，这种简单不仅体现在消除Cowork和Chat的区别上，还体现在整合Claude Design，让标准的核心Claude也具备制作幻灯片、设计、文档等所有能力，而无需切换到专门的体验中。

And by the way, as pointed out by Claude Code creator Kat Wu, this simplicity extends not only to eliminating the difference between Cowork and Chat, but also to integrating Claude Design so that your standard core Claude also has all of the capabilities to do things like make slides, designs, docs, without having to switch into some specialized experience.

而且，Anthropic团队表示，这同样是基于用户的直接反馈。

And again, the team at Anthropic reports that this was based on direct user feedback.

Anthropic实验室负责人、Instagram前联合创始人迈克·克里格写道：“我们最常听到的是，人们不确定该从哪个产品开始。”

Anthropic Labs lead and former Instagram co-founder Mike Krieger wrote, "The most common thing we hear, people aren't sure which product to start with.

这种摩擦阻碍了人们充分利用这些模型的能力。”

That friction gets in the way of getting the best from what these models can do."

正如我本周早些时候在节目里提到的，我个人并不完全认同这一点，因为我喜欢更具体的控制，但天哪，我似乎和大多数人的想法不太一样。

Now, as I mentioned in an earlier episode this week, I'm not totally sold on this on a personal level because I feel like I like having more specific control, but boy, do I seem out of sync with the average here.

克里斯蒂安·塞利格转发了公告并说：“能不能对Codex也这么做？拜托了。”

Christian Selig reposted the announcement and said, "Can we do this with Codex too, please?

这两种模式就像两个父母在吵架。”

The two modes are like having two parents fighting."

沙努·马修总结道：“人们不想在桌面上同时操作两三个文本代理、两三个聊天机器人代理、两三个终端。”

Shanu Matthew summed up, "People don't want to interact with two to three text agents, two to three chatbot agents, two to three terminals on desktop.

你想要一个简单统一的界面，让代理们协调完成工作。

You want a simple, unified interface and want the agents to coordinate how it gets done.

你只需要提问和接收结果。”

You simply want to ask and receive.

这个平台应该协调和委派各个代理，为任务分配模型，处理是本地、云端还是虚拟机的问题，并且要主动和被动地应对任务。

The platform should coordinate and delegate agents, route models for tasks, handle whether it's local cloud or VM, be proactive and reactive about tasks.

Instinct、Muse 和 Grokbot 真的引起了共鸣，而且越来越明显的是，在我看来，我们又处在一个范式转变的时刻。

Instinct, Muse, and Grokbot really hit a nerve, and it's increasingly clear we're in another paradigm shift moment in my opinion."

现在，如果第一个变化是关于我们使用 AI 的界面简化，那么第二个变化本身也是一种简化，它关乎我们如何控制 AI，这就是向单线程发展的趋势。

Now, if the first change is about the simplicity of the interfaces through which we use AI, the second change is itself also a version of simplicity that has to do with how we control AI, and this is the trend towards monothreads.

单线程的理念是，你不需要切换上下文窗口，也不需要为每个新需求或新提示单独解释，而是使用一个持续运行的单一长线程，它继承之前对话的上下文，从而管理整个体验。

The idea of a monothread is that instead of having to switch context windows and explain new needs and new prompts separate from the previous things that you've done, instead, you use a single long continuously running thread that inherits the context from past conversations as a way to manage the entire experience.

以前，这受到上下文窗口的限制。

Now previously, this was limited by the context window.

到了一定程度，AI 就会“累”，也就是说它的上下文窗口满了，然后它就开始做一些奇怪的事情，比如出错、丢失指令等等。

At some point, the AI would get tired, i.e., its context window would get filled, and it would start doing weird things, making errors, losing instructions, et cetera.

我去年年底和今年年初做的很多早期课程和教程，都是关于如何编写好的交接文档，以便在不同线程之间传递上下文。

Many of the earlier courses and lessons that I did from late last year and early this year were all about how to build good handoff documents to move context between different threads.

但今年这种情况开始改变，尤其是 Codex 团队专注于压缩技术，改变了人们通过单一统一线程与 Codex 交互的方式。

That started to shift this year, though, especially as the team at Codex really focused on compaction to change the way that people could interact specifically with Codex through a single unified thread.

Codex的Nick Baumann在四月份写道：“很多编码智能体的设计都建立在这样一个假设上，即突破上下文窗口并进行压缩会导致结果越来越差。”

Codex's Nick Baumann wrote back in April, "So much coding agent design is built on the assumption that breaching context windows and compacting context yields progressively worse results.

“当你放弃这个假设时，它开辟的产品方向非常令人兴奋。”

When you drop this assumption, the product direction it opens up is very exciting."

他是在回应开发者Anthony Kroger，后者写道：“我用Codex时真的从不担心上下文窗口。”

He was responding to developer Anthony Kroger, who had written, "I literally never worry about context windows using Codex.

“它可以压缩大约三次，但模型仍然能记住细节。”

It can compact like three times, and the model still remembers the details somehow."

大约一周后，Codex团队的Nick在一篇题为《我的Codex线程是活的》的文章中进一步阐述了他的想法。

In a post about a week later, Nick from the Codex team expanded on his thoughts in a piece he called "My Codex Threads Are Alive."

他在文中写道：“我现在最有用的Codex线程是我过去三周一直在用的那个。”

In it, he wrote, "The most useful Codex thread I have right now is the one I've been using for the last three weeks.

“每小时，它都会检查我的Slack、Gmail以及我写的或正在关注的PR。”

Every hour, it checks my Slack, Gmails, and PRs I wrote or I'm watching.

“每小时，它都会把噪音变成我可以采取行动的清晰信号。”

Every hour, it turns the noise into clear signal I can act on.

我现在用Codex的方式变了，以前是开很多短对话，现在是保持少量几个线程，围绕重复出现的工作流一直用下去。

My Codex usage has shifted from starting lots of short-lived chats to keeping a smaller number of threads alive around recurring work streams.

我还是会经常开新线程，但有些工作不应该每次我问个问题就重置。

I still start fresh threads constantly, but some work should not reset every time I ask a question."

总之，他写道：“我已经彻底迷上单线程了。”

In sum, he writes, "I have become monothread pilled."

Nick接着说：“我最近说过，Codex团队在压缩方面的改进，正在为编程代理开辟新的产品方向。”

Nick continues, "I recently said that the compaction improvements made by the Codex team are opening up new product directions for coding agents.

很多代理产品的设计都假设长线程最终会退化。

A lot of agent product design assumes long threads eventually degrade.

如果真是这样，产品就应该引导用户去创建新线程，并写上下文摘要。

If that is true, the product should push users towards creating new threads and writing context summaries.

Codex团队在压缩上做了很多工作，这些工作削弱了这个假设。

The Codex team has been doing a lot of work on compaction, and that work weakens the assumption.

当一个线程在多轮对话中还能保持有用时，我就希望同一个线程能继续处理那个重复出现的任务。

When a thread can stay useful across many turns, I want the same thread to keep working on the recurring task.

简单来说，有了好的上下文压缩，一个线程的价值会随着时间增长。

Put simply, with good context compaction, a thread's value increases over time."

尼克是在四月份写的那篇文章，而我在这一年里确实越来越发现自己的行为模式也在向那个方向转变。

Now, Nick wrote about that back in April, and I have very much found myself throughout the year seeing more and more of my behavior shift to that pattern.

比如，我在AI每日简报网站上做的所有事情，目前都是通过一个单一的Claude Code线程来管理的。

Everything that I do on the AI Daily Brief website, for example, is managed at the moment from a single Claude Code thread.

当我进去做修改或更新时，我不会重新开始那个线程，而是直接接着上次的进度继续。

When I pop in to make a change or an update, I'm not starting that thread all over, I'm just jumping into where we left off.

我用Claude从节目生成视频片段的那条社交流程是另一个线程，而且我甚至发现这种行为开始影响我在Claude其他部分（不仅仅是Claude Code）中的交互方式。

The social pipeline through which I use Claude to generate video clips from the show is another thread, and I've even found this behavior start to influence how I interact in other parts of Claude outside of just Claude Code.

比如，在过去的几个月里，在Projects中，我发现对于我经常做的事情，比如更新或写新的广告文案，我不再每次需要做这个任务时都开一个新的线程或对话，而是直接跳到上次做类似任务的地方继续。

Over the past couple of months, for example, inside Projects, I find that when it's things that I do frequently, such as updating or writing new ad copy, instead of starting a fresh thread or conversation every time I need to do that task, I just jump in at the end of the last time I did a similar task.

结果发现，不只是我一个人开始这样改变自己的行为。

Well, it turns out I haven't been the only one who has started to shift their behavior in this way.

事实上，从这周开始，这已经正式成为Claude Projects的工作方式了。

In fact, this is now, as of this week, officially how Claude Projects work.

周四，Claude账号发帖说：“项目现在从Claude Code中的一个对话开始运行。”

On Thursday, the Claude account posted, "Projects now run from one conversation starting in Claude Code.

你描述需要做什么，Claude会指挥并行线程，在你合上笔记本电脑后继续工作。”

You describe what needs doing, and Claude directs parallel threads that keep working after you close your laptop."

换句话说，并不是没有多个线程。

In other words, it's not that there aren't multiple threads.

而是这些多个线程由Claude管理，你通过一个线程与Claude交互。

It's that those multiple threads are managed by Claude, and what you do is interface with Claude through a single thread.

这和我刚才描述的模式非常相似。

This is very similar to the pattern that I had just described.

顺便说一句，在Claude推出这个功能的前一周，Cursor也推出了类似的功能。

And by the way, a week before this came to Claude, it had also come to Cursor.

实际上，当Cursor推出他们的项目版本时，这个逻辑就已经内置了。

In fact, when Cursor introduced their version of Projects, it had this logic built in.

在他们的公告中，他们写道：“不是为每个任务创建一个聊天，而是你与一个协调代理在单个持久线程中协作。”

In their announcement, they write, "Rather than creating a chat for every task, you work with a coordinator agent in a single persistent thread."

这个转变的影响还挺大的。

And the impacts of this shift are pretty big.

Claude Code的创始人Boris Cherny写道：“Projects不仅改变了我与Claude互动的方式，也改变了我写代码的方式。

Claude Code creator Boris Cherny wrote, "Projects have changed not only how I interact with Claude, but how I code.

我不再管理会话了。

I stopped managing sessions.

我只要有想法就直接发过去，Claude会把这些想法拆分成不同的线程，而且项目会记住我的工作方式。”

I just send thoughts as they come, Claude splits them into threads, and the project remembers how I work."

我觉得这其实触及到了我们与AI协作方式上的一个相关转变，那就是聊天机器人越来越像是智能体舰队的管理者。

And this, I think, gets to a very related shift in how we work with AI, which is that chatbots are increasingly agent fleet managers.

看看我刚才举的那些例子，变化的其实是我们与协调型AI交互的界面，但那个协调型AI仍然在幕后调度着许多不同的子智能体去完成各个部分的工作。

If you look at all of those examples that I just gave, what was changing was the place that we interact with a coordinator AI, but that coordinator AI was still spinning up lots of different sub-agents to do different parts of the work.

我觉得有意思的是，在一段时间里，我们很多人可能都以为聊天机器人界面本身只是一个过渡性的界面，虽然方便熟悉，但最终会被别的东西取代。

What's interesting about this to me is that I think for a while, many of us would have assumed that the chatbot interface itself was perhaps an intermediate interface, something that was convenient and familiar for the time but might give way to something else.

有趣的是，当你把单线程和聊天机器人作为智能体舰队管理者这两者结合起来看，就会发现，模型公司和硬件公司并没有让我们抛弃聊天机器人，而是悄悄改变了聊天机器人的本质，但并没有把界面这个基础从我们脚下抽走。

What's interesting is that when you take the combination of monothreads and chatbots as agent fleet managers, what has happened is that instead of shifting away from chatbots, the model and harness companies have quietly shifted what chatbots actually are, but without pulling the interface rug out from under our feet.

我们仍然能享受到那种和聊天机器人互动的舒适体验。

We still get that same comfortable experience of interacting with a chatbot.

实际上，甚至更简单了，因为我们可以通过一个单一对话来完成。

In fact, one that's even more simplified because we can do it from a single thread.

但那个聊天机器人现在做的事情，和一年前相比，已经大不相同了。

But what that chatbot is doing is very, very different than what it was doing a year ago.

Ethan Mollick写道：“Claude Projects之所以这么有意思，是因为它处理多代理团队的能力非常强。

Ethan Mollick wrote, "What makes Claude Projects so interesting is that it handles teams of agents really well.

你和一个主协调代理对话，它会启动各种专业代理。

You talk to a main orchestrator agent, and it spins up specialists.

基本上，它会创建一个组织来解决你的问题，根据你的偏好混合使用昂贵和廉价的代理。”

Basically, it creates an organization to solve your issues, mixing expensive and cheap agents depending on your preferences."

顺便说一句，这对我们其他的软件服务也有影响，这些服务可能会越来越像我们主协调代理的上下文数据库。

This also, by the way, has implications for the rest of our software services, which might increasingly start to look like context databases for our main coordinator agents.

Signal写道：“我所有的iMessage和邮件现在都通过一个单一的代理界面来处理。

Signal wrote, "All of my iMessages and email now run through a single agent interface.

我还在发短信，但我把手机上的所有邮件客户端都删了，也退订了所有邮件订阅，删掉了所有邮件应用。

I still text, but I delete every email client from my phone and canceled every email subscription or app I had.

现在对我来说，Gmail基本上就是个数据库。

Gmail is basically just a database to me now.

谷歌的其他服务也越来越像这样了，也就是说，我再也不想手动写邮件、创建文档、做幻灯片、安排日程或者管理文件了。

Increasingly so are the rest of Google services, i.e., I never want to manually write an email, create a doc, build slides, organize a calendar, or manage files again.

我只想表达我的意图，然后让底层的记录系统正确更新。

I just want to express intent and have the underlying systems of record updated correctly.

我觉得现在你会看到应用很快消失。

I think now you'll see apps disappear quickly.

很多应用最好的情况是变成数据库，最坏的情况就是完全过时。

Many will become databases at best or entirely obsolete at worst.

智能体会成为你个人和职业生活的默认界面。

The agent will be the default interface to your personal and professional life."

即使这不会完全实现，或者需要比像Signal这样有远见的人想象的更长的时间，可以肯定的是，你与之互动的聊天机器人正在做一些与过去不同的事情。

And even if that doesn't fully happen, or if it takes a lot longer than someone as forward-thinking as Signal might think, it is certainly the case that the chatbots that you interact with are doing something different than they used to.

市面上所有AI编程工具做的第一件事都一样。

Every AI coding tool on the market does the same thing first.

它们会直接开始写代码。

It starts writing code.

但Blitzy恰恰相反。

Blitzy does the opposite.

在写一行代码之前，Blitzy会花好几天时间逆向分析你的整个代码库。

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

成千上万个代理会处理数百万行代码，梳理出每一个依赖关系、每一个未记录的约束条件，以及过去十年里做出的每一个架构决策。

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

最终形成一个动态知识图谱，就像一位在系统里深耕了三十年的首席工程师那样，对你的软件了如指掌。

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

其他工具靠grep搜索和markdown文件来猜测上下文。

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

而Blitzy从不猜测。

Blitzy never guesses.

它先建立真正的理解，然后自主交付超过80%的完整软件史诗。

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

经过端到端验证、测试、生产级质量的拉取请求。

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

这就是为什么财富500强的工程团队把最重要的代码库交给Blitzy。

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

你自己去blitzy.com看看就知道了。

See for yourself at blitzy.com.

就是B-L-I-T-Z-Y点com。

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

最好的团队不是靠一个明星扛着所有人。

The best teams don't have a single star carrying everyone else.

他们了解自己的长处，也清楚彼此的短板，并且会利用这两点。

They know their own strengths and each other's weaknesses, and play to both.

这就是Robots and Pencils刻意打造的团队。

That's the team Robots and Pencils has built on purpose.

那里没有人为了凑人头数而埋头做琐碎的工作。

Nobody there is grinding through busywork to pad a headcount number.

人们是为了解决难题而来的，他们留下来是因为周围每个人都在同时进步。

People come for the hard problems, and they stay because everyone around them is leveling up at the same time.

在一个满是只想快速招人的公司市场里，这值得一看。

In a market full of companies that are just trying to hire fast, that's worth a look.

去 robotsandpencils.com/careers 看看吧。

Check out robotsandpencils.com/careers.

每期节目，我都会聊到 OpenAI、Anthropic、SpaceX AI、谷歌和 Meta 之间的竞争。

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

如果你听了一段时间，可能已经有了自己偏爱的公司。

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

也许你觉得 OpenAI 和 Anthropic 能保持领先，或者 Meta 的开源策略能胜出。

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

不管你怎么看，每个人工智能实验室都创造了不同的投资机会。

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

Harbor Capital Advisors的AI实验室生态系统ETF系列，让你能够投资于你相信的AI实验室背后的生态系统。

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

在你投资的任何地方搜索Harbor AI实验室生态系统ETF，或者在X上关注@HarborCapital了解更多信息。

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

访问harborcapital.com获取招股说明书，其中包含投资目标、风险、费用、支出及其他重要信息。

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

投资前请仔细阅读并考虑这些信息。

Read and consider it carefully before investing.

风险包括本金损失以及与人工智能相关的风险。

Risks include principal loss and artificial intelligence-related risks.

Harbor ETF由Foresight Fund Services, LLC分销。

Harbor ETFs are distributed by Foresight Fund Services, LLC.

Harbor与AI Daily Brief无关，这些基金与任何AI实验室也没有关联、赞助或背书关系。

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

这是一则付费广告，并非个性化投资建议。

This is a paid advertisement and not personalized investment advice.

投资有风险，可能损失本金。

Investing involves risk, including possible loss of principal.

本期AI每日简报由Hyperagent赞助播出，Hyperagent让你运行多组智能体，团队可以共同管理。

This episode of the AI Daily Brief is brought to you by Hyperagent, where you run fleets of agents your team can manage together.

别再依赖本地智能体和聊天工作流，等着你的笔记本电脑来提示了。

Forget local agents and chat workflows waiting on your laptop to be prompted.

Hyperagent在云端部署全天候运行的智能体，在你团队已经使用的工具中完成实际工作。

Hyperagent deploys always-on agents in the cloud, doing real work across the tools your team already uses.

营销智能体把竞争对手的动态变成落地页。

Marketing agents turn competitor moves into landing pages.

销售智能体丰富潜在客户信息，起草邮件，并更新CRM。

Sales agents enrich leads, draft emails, and updates the CRM.

运营智能体处理文书工作，跟踪预算。

Ops agent chases the paperwork and tracks the budget.

每个智能体都能访问共享上下文，并遵守你关于范围和审批的规则。

Every agent has access to shared context and follows your rules about scope and approvals.

是时候让你的智能体像队友一样了。

It's time you add agents that feel like teammates.

去Hyperagent招一个吧。

Hire yours at Hyperagent.

在hyperagent.com/aidailybrief注册，就能获得100美元的额度。

Get $100 in credits at hyperagent.com/aidailybrief.

顺便说一句，我们节目里经常提到一个简单但影响深远的变化，那就是你与聊天机器人的互动方式，会越来越多地转向语音。

And by the way, one simple but profound change, which we've talked about a lot on this show, is that the way that you're interacting with that chatbot is increasingly going to be through your voice.

我今年发布的每一个免费学习项目，一开始都会强烈建议大家要么用ChatGPT的原生语音控制，要么就装个类似Whisperflow的工具，因为语音能带来更高的效率和更丰富的上下文。

Every free learning program I've released this year starts by imploring people to either use native ChatGPT voice control or to get something like Whisperflow set up for the sheer efficiency and expanded context that voice is going to enable.

而且，这越来越成为智能体产品的基本配置了。

And increasingly, this is just table stakes for agent products as well.

这周，GrokBot也加入了，推出了自己的原生语音集成。

This week, GrokBot joined the party with its own native voice integration.

SpaceX AI的Danny Graziosi写道：“用语音模式快速给你的幕僚长布置任务，同时它还能协调你的机器人团队，这感觉就像魔法一样。”

SpaceX AI's Danny Graziosi wrote, "Using voice mode to rapid fire tasks to your chief of staff while it orchestrates your bot team feels like magic."

同一团队的帕克·康拉德写道：“GrokBot已经在向更无缝的交互模式迈出了一大步。”

Parker Conrad from that same team wrote, "GrokBot already took such a leap towards a more seamless interaction model.

“我很有信心，语音会让它更进一步。”

I'm bullish on voice taking that even further."

加拉·比宗指出，关键不只是语音本身，而是语音现在遇到了更强大的底层能力。

Gara Bizon pointed out that it's not just voice alone, but the fact that voice is now meeting expanded underlying capabilities.

他写道：“今年我在很多智能体产品上试过语音，但直到智能体真的能去做事情之前，它总感觉像个噱头。”

He wrote, "Tried voice on a bunch of agent products this year, and it always felt like a gimmick until the agent could actually go do things.

“语音加上真正能干活，这才是组合拳。”

Voice plus real work is the combo."

而且，即使是那些已经拥抱语音的公司和项目，现在也变得更加投入。

And even the companies and projects that were already voice-pilled are getting even more so.

OpenAI刚刚发布了一个新的Codex广告，他们团队的两名开发者在锻炼时，用语音和Codex里的语音智能体对话，这个智能体由他们新的GPT Live 1模型驱动。

OpenAI just released a new Codex ad with two of their team's developers talking to the voice agent in Codex, powered by their new GPT Live 1 model while they're working out.

谷歌也在做同样的事情，这周发布了Gemini 3.8 Live和3.8 Live Extended Thinking，并写道：“智能和并行推理方面的重大升级，让用语音协作和执行复杂任务变得更加直观。”

Google is also on the same tip, releasing Gemini 3.8 Live and 3.8 Live Extended Thinking this week, writing that, "Major upgrades in intelligence and parallel reasoning make it more intuitive to collaborate and execute complex tasks using your voice."

上周，就像Codex在健身广告里宣传的那样，我们推出了Devin Voice，让使用Devin的开发者也能通过语音接入新的GPT Live模型来编程。

Last week, similar to the thing that Codex was advertising in that workout ad, we got Devin Voice, which allowed developers using Devin to also tap into the new GPT Live model to code by talking.

阿里·K·米勒总结得很好，她说：“每当我看到一个新的语音优先界面，我内心就会有个小小的声音在欢呼。”

And Ali K. Miller summed it up when she said, "Whenever I see a new voice-first interface, a tiny voice inside of me screams with joy.

“语音是现在和未来的界面。”

Voice is the interface of the present and future."

今年，除了单线程和更强大的代理协调器（它们能管理自己的代理群并接受语音指令）之外，我们给AI下达指令的方式也发生了变化。

Now, alongside mono threads and more capable agent coordinators that manage their own fleets and voice-based instructions, the nature of the instructions that we give to our AI has also changed this year.

总结一下这个重大转变：我们不再只是给AI提示，而是越来越多地给它们设定目标。

Summarizing the big shift, instead of prompting our AIs, we increasingly set goals for them.

这自然给了AI更多的自由度和自主权，让它们自己决定如何完成工作，并反复循环直到任务完成。

Inherently, this creates more latitude and agency on the part of the AIs for figuring out how that work should get done, and for working in recurring loops until that work gets done.

这始于Claude Code和Codex中引入的像/goal这样的原语。

This started with the introduction of primitives like /goal in Claude Code and Codex.

这是对这种交互模式的一种正式化，让人们可以指定目标，而不仅仅是输入提示。

It was a formalization of this interaction pattern that allowed people to specify goals instead of just dropping a prompt.

后来，这个想法扩展成了整个循环工程的概念，我们在整个夏天都在讨论这个，大概一个月前还办了一次网络研讨会。

This would later expand to the whole idea of loop engineering, which is something we've talked about throughout the summer and even did a webinar on about a month ago.

循环这个概念真正火起来，是因为 OpenClau 的创始人 Peter Steinberger 写道：“你不应该再给编程代理下提示了。”

The idea of loops really got a big lift when OpenClau creator Peter Steinberger wrote, "You shouldn't be prompting coding agents anymore.

你应该设计循环，让循环来给你的代理下提示。”

You should be designing loops that prompt your agents."

循环这个概念其实很简单，就是除了定义目标之外，你还要为这个目标定义成功标准，让代理可以自己衡量，这样它就能根据你给出的可量化标准，持续循环运行，直到达成目标。

This idea of a loop is simply that in addition to defining the goal, you define success criteria for the goal that the agent can measure itself against, so that it can run in a continuous loop until it achieves said goal based on the quantifiable criteria given.

关键是，为什么值得在这个节目里提，是因为虽然循环和循环工程在我们这些早期采用者的圈子里，听起来很像那种我们喜欢的流行词，但在其他人看来可能会翻白眼，但实际上，在思考如何最大化我们与 AI 的互动效果时，这确实是一个相当关键的思维转变，这也是我们在这里花这么多时间讨论它的原因。

And the thing is, and the reason that it's worth mentioning in this show, is that although loops and loop engineering has all of the hallmarks of the type of buzzy terms that we love over in our early adopter circles, but which reasonably elicit eye rolls among the rest This is actually a fairly critical mindset shift when it comes to thinking about how to get the most out of our interactions with AI, which is why we spend so much time on it here.

我们与 AI 互动方式的第六个变化，我想可以称之为多模态，也就是说，在个人技术栈中使用多种不同的模型，不仅仅是在组织层面，甚至是在单个用户的工作范围内。

The sixth change in how we interact with AI is, I guess you might call multimodality, i.e., the using of multiple different models in a personal stack, not just on an organizational level, but even within the domain of an individual user's work.

我还没想好用什么确切的语言来描述这种多模态，我的很多 Patreon 支持者都喜欢指出这一点。

I have not yet settled on the exact right language to describe this multimodality, as a number of my Patreons love to point out.

但是，一方面，很多高级 AI 用户一直把模型选择看作是一种获取好处的方式。

But while on the one hand, many advanced AI users have always viewed model selection as a way to get benefits.

换句话说，不只是局限于某一个模型，因为那是你订阅的那个，而是要找到最适合你特定任务的模型。

In other words, not just being constrained to a single model because that's the one you pay for a subscription for, but finding the model that does the best with your particular task.

现在的不同之处在于，工作的复杂性、难度、深度和持续时间，再加上最先进模型的成本，以及这些模型在标准套餐中补贴的变化，意味着这不再只是早期采用者的小众玩法，而是现在成本和效率上的基本要求了。

The difference now is that the complexity, difficulty, depth, and duration of the work, combined with the cost of the most advanced models and shifts in the subsidy of those models in the standard plans, means that this is not just an early adopter alpha thing anymore, but basically a cost and efficiency requirement now.

而且，随着时间推移，那些稍微落后于最先进水平的便宜模型，现在也能做到几个月前还算是顶尖水平的事情，这也让这种多模型使用变得更加可行。

It's also enabled by the fact that the more time that goes on, the more that cheaper models that are just outside the state of the art can do things that would've been state of the art just a couple of months ago.

Fable 5.1 发布的时候，这种情况就特别明显。

You saw this a lot when Fable 5.1 dropped.

当时X平台和其他AI社区里，到处都是人们分享的各种技巧，教你怎么在不花大钱的情况下充分利用这个模型。

You had people all over X and the other AI communities sharing their hacks on how to get the most out of the model without breaking the bank.

而且，你可以看到，这正在从事后补救的技巧，逐渐变成一种实际的使用系统和纪律。

And increasingly, you can see this moving from ex post hacks to actual usage systems discipline.

这种交互模式的转变，和把聊天机器人当作代理舰队管理者的想法非常契合。

And this interaction pattern shift dovetails very closely with the idea of chatbots as agent fleet managers.

很多人注意到的一件事是，至少在目前，像Fable 5.1这样的模型，默认情况下不仅会用最先进的模型作为协调者，还会用它来驱动所有子代理。

One of the things that a lot of folks have noticed is that, at least at current, the default for a model like Fable 5.1 is to not just use the most advanced model as the coordinator, but also to use it for all the sub-agents.

我们很多人都遇到过这种情况：给Fable一个我们认为很普通的任务，结果回来一看，当前的使用额度已经用掉了一大块，因为它派出去做研究的五到十个子代理全都用了最先进的Fable模型，而实际上这个任务根本不需要这么高的配置。

Many of us have had the experience where we give Fable what we think is a fairly standard task, only to come back and see a major percentage of our current usage cap tapped because it turns out that the five or 10 sub-agents that it spun up to go do some sort of research were all using the most advanced Fable model as well, even though that was completely not required for the task at hand.

简单来说，搞清楚哪些任务需要什么类型的模型和能力，以及如何在管理AI时合理分配，这将会越来越成为一项重要的技能。

TLDR, understanding where different types of models and model capabilities are needed and how to normalize that in your management of your AI is going to increasingly be an important discipline.

顺便说一句，这个趋势如果不小心的话，可能会和那个追求简单化的趋势产生冲突。

And by the way, this is one trend that, if not careful, could kind of run at odds with that trend towards simplicity.

随着各个模型和工具实验室努力让普通用户体验变得更简单，在很多情况下，这意味着会隐藏甚至取消那些更精细的控制选项，比如决定哪个任务用哪个模型。

As all of the model and harness labs try to make the average experience more simple, in many cases, that will mean obscuring or even doing away with the more fine-grained controls for determining things like which tasks get which model.

理论上，实验室和模型公司会尝试在底层构建自动化系统来帮你搞定这些细节，比如自动识别出研究任务所需的子代理并不需要最先进的模型。

Now, theoretically, the labs and model companies will try to build automated systems underneath that can figure out those specifics for you, i.e., to realize that the sub-agents needed for a research task are not state of the art.

但我不确定他们总能做对，所以我觉得这仍然是人们需要自己掌控的事情。

But I'm not sure that they're always going to get it right, and so I think that this will still be something that people need personal agency around.

最后一个交互模式的转变，我最近已经谈了很多，所以就不深入了，那就是共享代理的概念。

The last interaction pattern shift is one that I've talked about a lot recently, so I won't go too deep on, but is the idea of shared agents.

2026年会被视为代理真正落地的一年，经过这么多年的期待，终于成真了。

2026 will be looked back at as the year that agents became real after years and years of excitement.

但在今年前三分之二的时间里，几乎所有的智能体行为都是个人独立使用的。

But for the first two-thirds of the year, almost all of that agent behavior was siloed to individuals.

我们每个人都在各自构建自己的Open Claude团队，改变自己的工作方式，但不一定改变了整个团队的工作方式。

It was each of us individually building our open Claude teams, changing how we do work, but not necessarily changing how our teams do work overall.

现在你开始看到这种转变了。

You're starting to see that shift.

一个例子是今年夏天Claude推出了Claude Tag，取代了旧的Claude in Slack系统。旧系统里每个人可以调用自己的Claude实例，而新系统是基于团队的，每个频道都有自己的Slack，可以在该频道的上下文中工作，不绑定任何个人。

An example of that came this summer when Claude introduced Claude Tag, which replaced the old system of Claude in Slack, where each individual could call upon their own Claude instance with a new team-based system where each channel had its own Slack that could work across the context of that channel, not tied to any one individual.

现在，这个想法很明确。

Now, the idea here is pretty clear.

我们的一部分工作确实是由个人完成的，但很多工作是由团队完成的，我们可能需要的是存在于团队协作中的智能体，而不仅仅是存在于个人小世界里的智能体。

Some amount of our work gets done, yes, as individuals, but a lot of it gets done in teams, and we're probably going to need agents that live at the intersection of our teams, not just agents that live in our own little individual worlds.

现在你开始看到很多这方面的努力了。

And you're now starting to see tons of efforts towards these ends.

Mio XYZ这周发布了，号称是“整个团队共享的第一个AI员工”。

Mio XYZ launched this week as the, quote, "first AI employee your whole team shares."

在他们的公告文章里，他们写道：“从七月开始测试以来，Mio为团队节省了数千小时，完成了数万个任务。”

In their announcement post, they wrote, "In beta since July, Mio saved teams thousands of hours and completed tens of thousands of tasks.

你不需要另一个代理。

You don't need another agent.

你需要的是一个和团队共享的代理。”

You need a shared one with your team."

我觉得每天都能看到至少五六个类似的例子，尽管多人AI的具体形态还在发展中。

On any given day, I feel like I see a half dozen or more examples of this same impulse, although exactly what multiplayer AI means is still very much in development.

本月初，Ethan Mollick写道：“多人AI，即组织中许多人可以一起使用AI来达成目标，目前仍然是使用AI时最大的非技术性问题之一。”

Earlier this month, Ethan Mollick wrote, "Multiplayer AI, where many people in an organization can use AI together to accomplish goals, remains one of the biggest non-technical problems in using AI right now.

现有的方法通常比较原始，基本上就是把AI当作群聊里的一个人。

Approaches tend to be pretty primitive and based around AI as a person in your group chat.

这很局限。”

That is limiting."

Aria Bachtani写道：“我们仍然没有一个一致的思维模型来界定多人代理系统中的主体是谁。”

Aria Bachtani writes, "We still don't have a consistent mental model for who the principal is in a multiplayer agentic system.

是你吗，用户？

Is it you, the user?

智能体是你的代理，继承你的背景、权限和权威。

Agent is your proxy inheriting your context, access, and authority.

是智能体本身吗？

Is it the agent itself?

智能体有自己的身份、关系、记忆和访问权限。"

Agent has its own identity, relationships, memory, and access."

Aria继续说，区别在于"智能体文明与人类文明在智能体速度上的差异。

It's the difference, Aria continues, "between agent civilizations versus human civilizations at agent speed.

实际上，答案两者都有，但适用于不同的系统和用例。"

Realistically, the answer is both, but for different systems and use cases."

Atlan的Rishi Goyal Bhatnagar写道："多人AI将成为2027年最大的设计问题之一。

Atlan's Rishi Goyal Bhatnagar writes, "Multiplayer AI will be one of the biggest design problems of 2027.

这个短语本身，我遇到的每个人理解都不一样。

The phrase alone means something different to every person I talk to.

是共享记忆、上下文、组织设计，还是完全别的东西？

Is it shared memory, context, org design, or something else entirely?"

有意思的是，里希接着展示了为什么这些观点单独看时都很有说服力。

And what's interesting is that Rishi then goes on to show why each of these sounds compelling when you look at it in isolation.

他写道：“从记忆的角度看，五个代理在同一个项目上，却有五个不同版本的事件经过，而且彼此之间互不沟通。”

"The memory lens," he writes, "five agents on one project, five different versions of what happened, none of them talking to each other.

这简直一团糟。

It's a mess.

解决办法就是用一个共享记忆代替五个私有副本，这才是关键所在。

Fixing it, one shared memory instead of five private copies is the whole game.

但接下来还有上下文的角度。”

But then there's the context lens."

里希继续说：“我接触过的几乎每个人都为自己代理构建了专属的上下文层。”

Rishi continues, "Nearly everyone I talk to has built their own little context layer for their own agents.

这对他们来说很有效。

It works for them.

但放到整个团队里，就撑不住了。

Stretch it across the whole team, and it falls apart.

个人层面的上下文感觉已经解决了。

Personal context feels solved.

但组织层面的还完全空白。

Org wide is wide open."

然后，当然，还有组织变革这个维度。

And then, of course, there's the organizational change dimension.

Rishi总结道：“多人AI最难的部分可能根本不是技术问题。”

Rishi concludes, "The hardest part of multiplayer AI might not be technical at all.

让整个团队接受一种共享的工作方式，这更像是组织设计，而不是工程问题。

Getting a whole team to adopt one shared way of working is more org design than engineering.

没人愿意承认这一点。”

Nobody wants that to be true."

很明显，这个领域会有大量的工作要做。

What's clear is that there is going to be a ton of work done in this area.

乔什·罗森写道：“我们还没就多人AI的架构模式达成一致，就已经决定让AI走向多人协作。”

Josh Rosen writes, "We've decided AI is going multiplayer before we've agreed on the architectural patterns for multiplayer AI."

然后，他当然接着试图解释这一切，让我们得到一个稍微更清晰的答案。

And then, of course, proceeds by trying to explain all of that and get us to a slightly more coherent answer.

如果你对这个话题感兴趣，我想花点时间介绍一下我们最新推出的免费自学培训项目，那就是团队多人AI冲刺营。

Now, if this is something that you are interested in, I will show for a moment the latest free self-directed training program that we've dropped here, which is the Multiplayer AI Sprint for teams.

这个项目并不假设多人AI已经成熟，或者说，坦白讲，也不假设那些重要问题已经有了答案。

This does not assume that multiplayer AI is figured out, or frankly, almost any of the answers to those important questions.

这个冲刺营要做的是，让大家开始为多人AI——在我们看来，其实就是共享代理——如何在你的团队中运作打下基础。

What the sprint does is it gets people to start to build the foundations to how multiplayer AI, which is really just another way, in our case, of saying shared agents, can function inside your team.

在四个冲刺阶段中，我们按周划分，但你不一定非得按周来，我们会先盘点你们团队目前实际使用的AI工具，或者如果你还比较新，也许可以请组织里某个在AI方面更资深的高手或专家，来帮你们了解在你们具体组织环境下的最佳实践。

Over the course of the four sprints, which we've divided into weeks but don't have to be split up that way, we are going to inventory first what your team is actually running right now for AI, and/or to the extent that you are even more new, perhaps bring in some champion or expert from your organization who is a little bit farther in their stack to help you understand best practices in your specific organizational's context.

一旦你完成了这样的盘点，就进入第二阶段，这个阶段主要关注上下文，也就是写下你的AI对你了解什么，以及它应该了解你什么。

And once you have that sort of inventory, you move then into sprint two, which is all about context, writing down what your AI knows about you and what it should know about you.

第三阶段是查看工作流程和上下文中的重叠部分，这会帮你找到可能构建共享代理的地方。

Sprint three is looking at the overlap in work streams and in context, which is going to give you your possible places to build shared agents.

第四个冲刺阶段就是真正动手构建的时候。

And sprint four is going to be where you actually do that building.

当然，你可以多次重复第三和第四个冲刺阶段，尽可能多地试验不同的智能体，只要对你有帮助。

And of course, you can then repeat sprint three and sprint four multiple times to experiment with multiple agents as much as it helps.

很快就会有大量的产品体验和优秀的产品思维涌入多人AI或共享智能体领域。

There are going to be tons and tons of product experiences and great product thinking coming to the multiplayer AI or shared agent space in very short order.

所以，如果你想再等等，让行业和规范再稳定一些，我完全理解。

And so to the extent you want to kick that can down the road just a little bit and let the industry and norm settle a bit more, I totally understand.

不过，我确实认为，对于那些想深入探索的团队来说，抢先一步肯定能获得一些优势，希望《多人AI冲刺》——当然在节目笔记里有链接——能帮到你。

That said, I do think for teams who want to dive in, there is certainly some alpha to be had by getting out ahead of that, and hopefully that's what the Multiplayer AI Sprint, which is of course linked in the show notes, can maybe help you do.

以上就是我们使用AI的方式正在改变的七种方式。

So that is the seven ways that how we use AI is changing.

希望这能给你一些有趣的新方向去探索，以及一些思考你与AI关系的新方式。

Hopefully, this gives you some interesting new directions to explore and some new ways to think about your relationship with AI.

今天的《AI每日简报》就到这里。

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

感谢你的收听或观看，一如既往。下次再见，祝一切安好。

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