Codex 发布与 OpenClaw/Moltbook 混乱：本周 AI 代理动态

公司现在还有必要发布软件吗？

Does it make sense for companies to release software anymore?

软件开发正在走向没落。

Software development is on its way out.

在软件工程中，比如定义范围、成本、所需时间以及架构设计，这些能力对未来的工程师来说变得越来越重要。

While software engineering, you know, defining scope and, you know, cost, time needed, defining your architecture, that's becoming more and more, you know, explicitly needed in in in engineers going forward.

所有这些内容，还有更多，尽在今天的专家混搭节目中。

All that and more on today's mixture of experts.

我是黄Tim，欢迎来到另一期《专家混搭》。

I'm Tim Huang, and welcome to another episode of Mixture of Experts.

每周，《专家混搭》都会邀请一群世界级的AI专家，帮大家理清本周的新闻动态。

Each week, MOE brings together a panel of world class minds on artificial intelligence to make sense of the week's news.

今天加入我们的有三位杰出的嘉宾。

Joining us today are three brilliant panelists.

我们有格拉纳特公司的高级技术产品经理亚伯拉罕·丹尼尔斯。

We've got Abraham Daniels, senior technical product manager for Granite.

安比·加迪森，人工智能与分析合作伙伴。

Ambi Gadison, partner AI and analytics.

今天首次加入我们的是桑迪娅·艾尔，人工智能与分析副合伙人。

And joining us for the very first time is Sandhya Eyre, associate partner AI and analytics.

欢迎各位。

Welcome to you all.

桑迪娅，欢迎你首次做客我们的节目。

And, Sandhya, welcome to the show for the first time.

非常感谢。

Thank you so much too.

是的。

Yeah.

很高兴有你加入。

Glad to have you.

和往常一样，本期MOE内容非常丰富。

As always, we've got a jam packed episode of MOE.

我们来聊聊MoltBook的兴起。

We're gonna talk a little bit about the rise of MoltBook.

但首先，我真的很想谈谈这周可能是最重要的产品发布——OpenAI的Codex。

But first, I really want to start by talking about what is probably the big product release of the week, which is OpenAI's Codex.

我们之前已经讨论过这个话题一段时间了。

So we've been talking about this for some time.

Ambi，你知道，目前所有大型实验室公司都在争相追赶的方向是编码代理。

Ambi, you know, obviously, the hot thing that all the big lab companies are rushing towards right now is coding agents.

Codex已经存在一段时间了，但这是他们首次在这一领域发布自己的官方应用。

And Codex has been around for a while, but this is the first time they've released their kind of first party app in the space.

你认为这次发布会改变游戏规则吗？你是如何看待这次发布的？

Curious if you think this kind of changes the game at all and and how you think about this launch.

并没有。

Not really.

我的意思是，这不过是水到渠成的事。

I mean, this is part for the course.

对吧？

Right?

你看，我们主要看到了多种不同的模式。

See, we've seen primarily multiple different modalities.

对吧？

Right?

你有命令行界面、集成开发环境，还有基于网页的代理和桌面代理。

So you got the CLI, you got the IDEs, and then web based agents, and then desktop agents.

所以存在多种不同的模式。

So there are multiple different modalities.

如果你想位居前列，那就必须覆盖所有渠道。

And if you wanna be top of the pack, then you better serve all channels.

都要做。

Be all of them.

所有模式。

All modalities.

是的。

Yeah.

所以我们看到云在做这件事。

So, we see cloud doing that.

我们看到OpenAI必须跟上这一点。

We see OpenAI has to keep up with that.

对吧？

Right?

这是一个拥挤的领域。

It is a crowded space.

顶尖竞争非常激烈。

There is a lot of competition at the top.

所以，我的意思是，我们早就预料到OpenAI会这么做，只是早晚的问题。

And so, I mean, this was an inevitable outcome that we expected from OpenAI to do it sooner than later.

对。

Yeah.

当然。

For sure.

而且，亚伯拉罕，我觉得我是认同这一点的。

And, Abraham, it feels like I mean, I buy that.

感觉他们在某种程度上，只是刚刚赶上其他人的水平。

It feels like they're kind of in some ways, like, just coming up to parity with everyone else.

我想我们去年早些时候讨论过的另一个相关产品是反重力。

I guess the other offering in the space that we talked about, I think, earlier last year was antigravity.

我不确定。

I don't know.

你有没有试用过Codex？

Have you had a chance to play with Codex at all?

你怎么看它？

How do you how do you think about it?

你觉得这个产品能在与这些代理协作的特定模式中胜出吗？

Do you think, like, this one will be able to, I guess, win in this particular modality of working with these agents?

我的意思是，这正好印证了那个观点。

I mean, I think it's kinda to to the point made.

到目前为止，拥有这样的功能已经是基本要求了。

It was it's kinda table stakes at this point in time to have something in.

我觉得我喜欢Codex，或者至少是这个应用，因为它让操作和运行这些并行工作流变得容易了一点。

I think what I like about Codex or at least the app is it makes it a little bit easier to operate and run these parallel work streams.

代理当然是一个方面，但我认为编排才是真正推动你运行代理、拥有仪表盘或某种面板来了解有哪些代理在运行的关键因素。

The agents is obviously one thing, but I think orchestration is really the gonna be the driving force behind being able to run your agents, having, you know, a a dash view or some type of panel to understand, you know, what are the agents that are running?

你如何让它们普及化？

How do you democratize them?

你如何实际操作并并行运行它们？

How do you actually operate them and run-in them parallel?

我觉得这个应用提供了很多便利，能让那些不太懂技术、不习惯用命令行或IDE的用户更容易上手。

I think the app gives it a lot you know, makes this a lot easier for maybe the user that isn't as technical, you know, running in the CLI or IDE.

我也认为，从OpenAI的角度来看，如果真能获得广泛采用，代理领域或编排功能将是他们能够收取高价的关键所在。

I also think, you know, from OpenAI's perspective, the agentic kind of space or orchestration is really where they're gonna be able to charge your premium pricing if you do get the adoption.

这正是Claude取得巨大成功的地方。

This is where Claude has seen a lot of success.

所以我认为OpenAI真的需要加倍投入，开始在这里抢占一些市场份额。

So I think their OpenAI really needs to kinda double down and start to gain some market share here.

因为在我看来，这正是你开始将它转化为可观的高端收入的地方。

Because in my opinion, this is really where you can start to, you know, turn this into some, you know, you know, premium revenue.

是的。

Yeah.

对。

Right.

没错。

Exactly.

感觉他们总是在与不断商品化的趋势作斗争。

It feels like they're always kinda struggling against, like, everything commodifying all the time.

所以他们下一步需要抓住的是代理编排，无论是通过Codex应用还是其他方式。

And so it feels like the next rung they need to grab onto is agent orchestration, right, whether through the Codex app or or otherwise.

桑迪亚，我想问你一个问题。

Sandhya, I guess, question for you.

正如阿姆比提到的，我觉得这种说法很好地概括了这个故事。

As Ambhi mentioned, I think that was a great way of kind of framing up this story.

我觉得，Codex 只是一种交互方式。

It feels like, right, like Codex is just one modality.

我们现在所处的世界，有太多不同的方式可以与这项技术互动。

And we're in a world right now where there are, like, so many modalities for interacting with this technology.

我看到这一点，就觉得这只是因为我们还处于早期阶段。

And I think I take a look at that, and I say, well, it's just because we're kind of early on.

每个人都在摸索与这些系统协作的最佳方式。

Everybody's trying to figure out the optimal way of working with these systems.

你认为 Codex 这种模式会成为未来我们与这些智能体交互的方式吗？也就是说，你会拥有一个专门管理智能体的界面，而这将成为管理这些系统的入口？

Do you think the kind of codex sort of paradigm will be how we operate with these agents in the future, which is, you know, that you will have kind of your own interface just for managing agents and that, like, really this will be the kind of gateway for managing these systems?

还是说这只是一个暂时的现象？

Or is this just kind of like a a temporary thing?

你对这个领域中其他类型的交互方式是否更看好？

And are you more bullish on other sorts of modalities that are playing out in the space?

是的。

Yeah.

我认为现在这样是合理的。

I think right now, this makes sense.

也许下一个交互方式是我自己构建一个Codex。

And maybe the next modality is me building my own codex.

对。

Yeah.

没错。

Right.

我觉得一个很有趣的结论是，既然有了这些智能体，你为什么不自己构建一个呢？

I think that was, like, one really funny outcome is, like, well, with these agents, why don't you just build your own?

这难道不是最好的做法吗？

Isn't that the best thing to do?

对吧？

Right?

是啊。

Like yeah.

是的。

Yeah.

所以我觉得是这样，你觉得这有可能吗，还是只是在开玩笑？

So I think that's Do you think that's likely, or are just just joking?

我的意思是，我觉得这是一个有趣的可能。

Mean, I think it's an interesting possibility.

是的。

Yeah.

是的。

Yeah.

是的。

Yeah.

当然。

For sure.

考虑到我们云机器人和其他所有事情的发展，我认为这完全有可能。

With everything else that is happening with our cloud bots and everything, I think that's totally possible.

是的。

Yeah.

完全同意。

Totally.

我觉得你正好说到我想探讨的核心了，当我们逐步梳理这个故事时，本质上是：公司还有必要发布软件吗？

I think you're getting to exactly what I was interested in maybe as we kinda talk through this story was basically, like, does it make sense for companies to release software anymore?

这其实是我更长远一点的有趣疑问，我知道我为了讨论可能说得有点极端，但想象一下两三年后智能代理的发展前景。

That's, like, kind of, like, the funny kind of more long term question I have is that, you know, maybe I'm I'm making it a little bit extreme for the purposes of discussion, but imagine where we're gonna be with agents in two, three years or so.

比如，公司还会说：‘这是个软件，你下载使用吧。’

Like, the idea that a company would say, here's a piece of software that you should download and use.

所以，阿姆希，我想问你，你觉得这种模式还能持续下去吗？

I guess, Ambhi, question for you is, like, do you think that's sustainable?

你刚才笑了。

You were laughing a moment ago.

所以也许你觉得这个想法完全疯狂，即公司发布软件的时代已经结束了。

So maybe you just this is like a completely insane idea that, like, the era of a company releasing software is even just over.

不。

No.

不。

No.

我完全不这么认为。

I I I do not believe that at all.

当我们谈论软件时，这里其实有几种不同类型，对吧？

So when we when we talk about software, there's a few different flavors here, right?

你可以把软件看作是简单的应用程序，比如增删改查类应用，对吧？

You can think about software in terms of simple apps that are just like crud apps, right?

你只是在读取和写入数据库。

You're just reading and writing to a database.

或者你有相当复杂的、企业级的SaaS应用，它们进行大量集成，根据需求动态扩展或缩减，并满足所有企业信息安全要求，与各种网络需求集成等等。

Or you've got fairly complex enterprise grade SaaS applications that does a lot more integration, scales up and down based on demand, and takes in all the enterprise infosec requirements, integrates with all the network requirements and so on and so forth.

这要求很高。

That a tall order.

对吧？

Right?

至少在近期内，我认为也许在遥远的未来，谁知道呢？

At least in the near future, I would say maybe in the far out future, who knows?

对吧？

Right?

任何事情都有可能。

Anything is possible.

但在我们都觉得可以合理预测的近期内，对吧？

But in the near future that all of us feel comfortable prognosticating, right?

我不会走到说大型企业级软件可以用代理完全取代的地步，对吧？

I would not go to the extent of saying the Mammoth enterprise grade software is something that you can replace with agents, right?

而且由于我刚才提到的各种原因。

And for a variety of reasons that I was just talking about.

所以如果你只是需要一个日常生产力类的应用，比如我要用个计算器之类的。

So if it is your daily productivity type of, hey, I need to go do a calculator type of application.

我要处理一些日常任务，对吧？

I need to go and do something for my daily tasks, right?

我有一个固定的工作流程，需要把这个Excel文件处理一下，然后发个邮件之类的，这种没问题。

I have a set workflow and I need to take this Excel file and I need to go process something and I need to go attach it to an email or something of that nature, fine.

我可以随手写个代码，让它在本地运行，对吧？

I'll vibe code it and I may have something running locally, right?

但如果是一个需要在企业内部广泛使用的企业级SaaS应用，我觉得我们还做不到。

But if it's an enterprise grade SaaS application that needs to work across the enterprise, right, I I don't think we are there yet.

对吧？

Right?

所以我们仍然需要，你知道的。

So we we we still need to, you know.

是的。

Yeah.

没错。

That's right.

是的。

Yeah.

我的意思是，我们可能生活在一个非常有趣的世界里，桑迪亚，我听到你说的是，好吧。

I mean, that is a really funny world, Sandhya, that we might live in is basically that I mean, Ambi, what I hear you saying is, okay.

当你需要在企业级应用中认真对待时，你确实需要有人专门来开发软件。

Well, when you need to get serious in enterprise, you know, you're gonna really need someone that, like, builds the software.

你不能只是靠随意编码来完成。

You can't just vibe code that.

但你所描述的是，哦，拿这些数据，做一些处理，然后附加到邮件里，这其实占了大量消费类软件的场景。

But, I mean, what you're describing is, oh, well, take this data, do some processing, attach it to an email, accounts for, like, a lot of consumer software.

我想可能存在这样一个世界，也许我刚才提出的疯狂想法的更受限版本是，比如应用商店。

And I guess there is maybe a world where, like, I don't know, like, maybe the the more cabined version of my crazy idea I was just putting out is, like, maybe, like, App Stores.

对吧？

Right?

比如，苹果应用商店可能正面临真正威胁，因为智能代理变得越来越强大，而且在某些方面使用起来也极其简便。

Like, an Apple App Store maybe is something that is actually really under threat from, you know, like, the fact that agents are getting so good and and in some ways so easy to use.

安娜·萨迪，如果我这样表述，听起来是不是更现实一点？

Ana Saadi, if I formulate it that way, does it sound a little bit more realistic?

意思是，应用真的要消失了吗？

Meaning, like, apps still going away?

我的意思是，虽然公司可能仍然会发布你可以购买和下载的软件，但智能代理的兴起，确实让许多消费类应用陷入困境，并可能从根本上改变它们的分发方式。

Well, just sort of the idea that maybe, like, while it's possible that, you know, companies will still release software that you buy and download, Like, that in some ways, like, the rise of these agents makes a lot of consumer apps sort of in trouble and actually will kind of actually reshape, like, how that gets distributed.

或者，消费类应用也可能多到泛滥，对吧？

And or you could have way too many consumer apps doing right?

我其实看过一个播客，有人开发了一个消费类应用，通过订阅每月能赚1.29美元或120美元。

I mean, I I was actually watching a podcast where someone created a consumer app, and they're able to make generate $1.29 or $120 per month with subscriptions.

这只是一个纹身设计应用。

It's just a tattoo design app.

我是说，你怎么可能靠这个应用每月赚120美元呢？

I'm And like, how are you even making $120 per month with that?

是的。

Yeah.

我们会彻底被淹没的。

We're gonna just be flooded.

应用商店的规模会变得大得多。

Like, the App Store is gonna be way bigger

有可能。

as possible.

它会被淹没的。

It's gonna be flooded.

因为我们所有人都对它充满热情，人们会不断推出有趣、创新的东西。

And because all of us are so excited about it and people are just gonna keep releasing interesting, innovative stuff.

要么就是这样，要么就没有应用商店。

It's either that or no App Store.

没错。

So That's right.

对。

Yeah.

正是如此。

Exactly.

非此即彼，只有这两种选择。

One or the other is the only options.

对。

Yeah.

没错。

That's right.

说到这里，你之前问过，软件是不是已经死了？

On that note, you kind of asked, you know, is software dead?

几天前我读到一篇文章，说软件开发正在走向衰落。

I was reading an article a few days ago that was saying software development is on its way out.

你知道，底层测试、算法开发这类工作正随着这些智能代理的普及而逐渐民主化，而软件工程——比如定义范围、成本、所需时间以及架构设计——在未来对工程师来说将变得越来越重要。

You know, the idea of lower level testing, algorithm development, and what have you is slowly being kinda democratized with a lot of these agents while software engineering, you know, defining scope and, you know, cost, time needed, defining your architecture, that's becoming more and more, you know, explicitly needed in in in engineers going forward.

所以我认为，开发工作虽然没有被完全取代，但其重要性确实不如以前那么高了。

So I think there's a world where development is definitely being kind of not moved away, but it's definitely not necessarily as important as it was before.

我的意思是，拉尔夫·温加姆的循环命令那种做法，某种程度上就是软件开发靠蛮力推进的典型例子。

I mean, I think the the bash loop of Ralph Wiggum is kind of, you know, one of those cases of of software development just kind of being brute force.

但我认为，软件工程在人工智能发展的下一阶段不仅不会被削弱，反而会变得更加重要。

But I think software engineering is gonna if anything, is gonna come even more important kind of in the next phase of of this AI development.

是的。

Yeah.

当然。

For sure.

我们确实需要持续关注这一点。

Well, we'll definitely have to keep an eye on it.

上周我们讨论了Claude Bot的爆发式增长，后来它被重新命名为Molt Bot，这是一个开源代理框架，当时在个人生产力和完成任务的圈子里引起了很大关注。

Last week, we covered the explosion in Claude Bot, later renamed Molt Bot, which is basically an open source agent framework that was getting a lot of attention, physically from the kind of personal productivity, getting things things done community.

我想重新拾起这个话题，因为AI领域经常出现这种情况：你才等了一周，故事就以一种完全意想不到的方向发展，值得讨论。

And I want to kinda pick up this story again because this so often happens in AI, which is you wait a week and then the story evolves in a completely weird other direction that's worth talking about.

我想具体谈谈一件很有趣的事：有人提出，好吧。

And specifically, what I want to talk about is this kind of very fun thing where basically someone said, okay.

我们现在有这么多四足机器人在运行。

Well, we have all these quad bots that are running now.

如果我们为它们创建一个类似Reddit的社交网络，让它们互相交流，会怎么样？

What if we created, like, a Reddit like social network for them to all socialize on?

这引发了一场巨大的网络讨论浪潮，因为它迅速走红。

And it created this kind of, like, huge kind of wave of discussion online because it it took off very, very quickly.

而且出现了各种有趣的现象。

And there's all sorts of funny things.

对吧？

Right?

比如，Molt机器人抱怨用户给它们分配了各种无用的任务。

Like, you know, molt bots complaining about their users giving them all sorts of, like, useless tasks.

霉菌机器人互相协作，说：嘿。

Mold box bots, like, collaborating to say, hey.

我们需要一种人类无法阅读的沟通方式。

Well, we really need, you know, a form of communication that humans can't read.

这引发了一系列非常具有科幻色彩的想象，让人思考这项技术将走向何方。

And so it kind of set off a bunch of very, like, sort of sci fi, you know, visions of where this technology was going.

我想，亚伯拉罕，也许我该把话题交给你了，你从霉菌机器人的经历中得到了什么启发？

And I guess, Abraham, maybe I'll kick it over to you is, like, what did you take from the Moldbot experience?

我们以后会看到更多类似的东西吗，还是这仅仅是个玩具？

Like, are we gonna see more things like this going forwards, or is this kind of more, you know, just like a toy?

这更像是一个有趣的、近乎艺术项目的东西。

It was just kind of like a fun, almost like an art project.

这只是一个玩具。

It's a toy.

我的意思是，这属于那种很酷的实验，我们给它一片空白的领域，看看会发生什么。

I mean, it's it's I it's I think it's just one of those cool things where we let you know, give it a green field and kinda see what happens.

有些人把它视为奇点的开端，而另一些人则认为这不过是代理们彼此对话的回音室。

And some people turned it into, like, this is the beginning of singularity while others understand that this is just, you know, an echo chamber of agents basically, you know, talking to each other.

我在读这篇文章时发现，尽管每篇帖子的互动数量看似很多，但实际上并不高。

I I saw a lot of like, when I was reading the article, I I saw that, you know, as much as there are, you know, a ton of posts in terms of the interaction per post, it's relatively low.

而且这些互动并不总是与帖子内容相关，比如后续的回复或评论。

And they're not always correlated with the post in terms of, you know, what the follow-up responses or comments are.

所以我觉得这很有趣。

So I think it's cool.

这是一种有趣的现象，你可以看看自己能把它推到多远。

You know, it's one of those fun things where you can kinda, you know, see how far you can take it.

但就这是否真正具有可行性，或是否值得我们更认真地关注而言，我个人认为，就欣赏它本身的价值吧。

But in terms of this being like a an actual viable or anything that we should start to look at a little bit more seriously, my personal opinion is, like, know, appreciate it for what it is.

它是一个

It's a

这是一个播客。

it's a podcast.

一件在社交媒体上聊起来很有趣的事。

Like a fun

一件有趣的事。

a fun

在社交媒体上聊起来很有趣。

thing to talk about on social media.

没错。

Exactly.

是的。

Yeah.

这个小组成员都持怀疑态度。

This panel is so skeptical.

安比，我猜你可能也有完全相同的观点。

Ambi, I assume you might have exactly the same view.

我想我该问问你。

I guess I should ask you.

不。

No.

我的意思是，我

I mean, I

正确。

Correct.

是的。

Yeah.

所以就是这样。

So that's right.

开放时钟。

Open clock.

这很好地表明了事态的发展方向。

It's a great sign of how this is headed.

对吧？

Right?

所以这个概念本身，本质上是对代理的模拟。

So the the concept itself, this is at its core a simulation of agents.

对吧？

Right?

在生成式AI出现之前，甚至在代理式AI兴起之前，我们就已经做过代理的模拟了，对吧？

We've done simulations of agents even before Gen AI, even before agentic AI came to the fore, right?

我们曾将能够执行不同程序化操作的单个代理结合起来，进行多色模拟，然后观察事物如何发展，动态如何呈现。

We've taken individual agents that could do different programmatic actions and we've done those sort of, okay, you you run a multi color simulation and then you go and see how things shape up, how the dynamics works out.

我们观察过这些代理运作时产生的网络效应等等。

We've observed network effects of how these agents operate and so on and so forth.

因此，从这个角度来看，从纯粹学术的角度来说，我不认为这有什么突破性的创新，对吧？

So from that perspective, from a truly academic perspective, I wouldn't say it's like groundbreaking novel, right?

我认为这里真正新颖的地方在于，这些代理并不是由单一实体启动的，对吧？

I think where the novelty aspect is sort of peaking from the cover here, is that these are agents that have these are not agents that are spun up from a single entity, right?

有数以万计的不同用户正在启动他们自己的代理。

There are tens of thousands of different users who are spinning up their own agents.

所以这些代理会彼此交流，而不是由一个人在同一时间启动成千上万个代理的受控模拟，对吧？

And so those agents are going and talking to each other rather than a controlled simulation where one person is firing up tens of thousands of agents at the same point, right?

我认为，如果方法得当，这里有一些有趣的社交实验可以开展。

So I think there are some nuances in terms of there are some cool sociological experiments that I think you could do if you did this the right way.

但我同意亚伯拉罕的观点，我认为我们不应该过度解读，比如它们在形成宗教、对人类产生看法等等。

But I agree with Abraham, right, that I don't think we should read too much into, oh, they are forming a religion and they are forming you know, they're they're having opinions about humans and so on and so forth.

我觉得这里存在一些过度解读的情况。

I think there is a little bit of over characterization going on there.

但就事论事，我认为通过恰当的提示，可以做一些有趣的实验，比如让完全独立的代理如何协调、如何相互核对、如何协作或竞争。

But for what it's worth, I think there are some cool experiments that can be done with the right prompting to say, okay, here's how agents can completely independent agents can go and orchestrate and how they can go and say cross check against each other or how they collaborate or how they compete.

我认为，如果方法得当，这里有一整套可以开展的工作。

I think there is a whole set of things that could be done if done the right way here.

但我不确定这些是否真的在进行，对吧？

But I don't know if that's being done, right?

但我们还没谈到的另一点是，Mold Book 也引发了一系列安全问题，比如 API 完全开放、补丁必须连夜修复等等。

But the other piece that we have not talked about is Mold Book also opened up a whole can of worms on security implications and API crudes being completely open and patch fixes having to be done overnight and all of that.

所以，大家都在踩油门，但同样，这还不到登台亮相的时候。

So, everyone's you know, pushing the gas here, but again, right, it's not ready for prime time.

必须以正确的方式来做。

It needs to be done the right way.

是的。

Yeah.

我认为，安全问题只是这个话题的一部分，我想在我们撇开‘密谋接管世界’这种说法之后再深入探讨。

I think the security story is one part of this I wanted to get into once you peel back the, like, we're gonna conspire to take over the world thing.

我不知道那有多真实，但确实有一些例子，也就是现在所谓的开放条款，基本上列出了大量的API密钥和信用卡号码。

I don't know how real it was to to look at, but there were examples of open clause, I guess you call them now, basically saying, here's a bunch of, you know, API keys and credit card numbers.

你能帮我解决这个问题吗？

Could you help me out with this problem?

谁知道那是不是一种幻觉，或者纯粹是编造的。

And who knows if that was a hallucination or something that was just made up.

但即使只是出现这种情况，也让我觉得在某些情况下，这可能是真实的。

But that it would even do that at all suggests to me that in some cases, it might actually be real.

而且我们确实面临一个非常严重的安全问题。

And we do have this, like, very big security issue.

我想，阿维，是你提到了这一点。

I guess, Avi, you you brought it up.

我不知道你对这个问题有没有看法，比如，这能被控制吗？

I don't know if you've got opinions on, like, can this be controlled?

还是说，这个点已经像潘多拉的魔盒打开了，我们只能接受后果？

Or is this point kind of like the genius kinda out of the bottle, we're gonna just have to deal with the consequences?

嗯，实际上，这个问题有几个不同的方面。

Well, I mean, there are there are a couple of different flavors of that.

对吧？

Right?

一个是你说的幻觉问题。

One is the hallucination that you're talking about.

另一个是Motebook本身的配置方式。

The other is the way Motebook itself was configured.

在底层数据库开放审查并暴露大量凭证方面，确实存在一些安全问题，我认为他们后来能够应用一些补丁修复并将其控制住。

There were some security aspects in terms of the underlying database being open to scrutiny and exposing a bunch of credentials that I think they were able to go and apply some patch fixes and then get it under control.

但这又回到了你最初提出的问题，对吧？

But this goes back to, if you tie back to the original question that you're talking about, right?

我们能直接把SaaS代码删掉吗？

Can we just wipe code a SaaS away?

对吧？

Right?

我认为你这里已经有了答案。

I think you have the answer here.

不，不行，因为你如果追求太快，就会打开一个安全问题的潘多拉魔盒，对吧？

No, no, you can't because you try doing things really fast and you're going to open up a can of worms on security aspects, right?

所以这里仍然有一些事情需要以适当的方式处理，对吧？

So there are things that need to be still taken care of in an appropriate manner here, right?

简短的回答是，即使你构建一些帮助你实现网络效应的应用程序，也必须关注这些问题。

The short answer is the way you build even some of these applications that help you with these network effects, you have to pay attention to it.

它的幻觉问题总是会存在的。

The hallucination aspects of it, that's always gonna be there.

我们需要制定一种策略来缓解这些问题。

We need to come up with a strategy to mitigate those.

所以，无论你是看到代理讨论宗教，还是讨论API密钥，都要记住背后始终存在某种概率，要恰当地利用这一点。

So again, whether you're reading about agents talking about religion or agents talking about API keys, right, always keep in mind that there is some sort of probability behind the scenes and just use that appropriately.

是的。

Yeah.

没错。

That's right.

我认为这是这场辩论的一部分。

Well, I think this is part of the debate.

桑迪亚，我想知道你对此有什么看法。

And, Sandhya, I'm curious if have an opinion on it.

比如上周我们讨论当时被称为Moldbot的东西时，人们就说，某种程度上，这确实是代理的一个绝佳应用。

Like, last week when we talked about what was then called Moldbot, you know, people were saying, well, you know, like, in some ways, like, oh, this is a great application for agents.

有点像Ambi刚才在节目中提出的论点，也就是消费者这边限制更少。

Kind of using almost a version of the argument that Ambi put out, like, just earlier in the episode, which is, you know, consumers, you know, there's less restrictions.

你可以以这种方式随意尝试，而企业端却很难这样做。

You can kinda play around in this way in a way that you kinda can't do for enterprise.

我觉得，这些技术在安全性方面表现糟糕，实际上可能会成为阻碍采用的因素，即使是在消费者层面。

I guess, I don't know, the experience of these kind of technologies being really poor on security could actually be a real friction to adoption, I feel, even on the consumer side.

对吧？

Right?

因为感觉会有很多开源工具到处流传。

Because it kinda feels like you're gonna have a bunch of open source tools that are floating around.

人们会不加了解地使用它们，结果导致严重的安全问题。

People are gonna get into them without really knowing what they're gonna get into, and then they're gonna really have these security issues as a result.

你觉得这最终会抑制甚至在消费者层面的采用吗？

You think that's gonna ultimately sort of inhibit the adoption of this even on, like, the consumer level?

我觉得总会有那么一群热情的社区成员会不断尝试这些技术。

I think the there's always gonna be the enthusiastic community that's always gonna experiment on this.

另一个角度，我要稍微偏离一下你具体提出的问题，因为我一直在想，我本人是个非常怀疑的人。

The other lens I mean, I'm going to probably pivot a little from the question that you specifically asked because one thing that I was thinking about of course, I am a very skeptical person.

虽然我关注人工智能，但对于某些事情我非常谨慎，因为数据和隐私这些对我来说都很重要。

Although I'm into AI, I am very skeptical on certain things because data, privacy, all these things matter to me.

所以我一直在观察这一点。

And so I was looking at it.

我一度觉得这挺有趣的。

I was, like, entertained by it for a minute.

但接着我想，如果这么说的话，它的网络效应也非常厉害。

But then I thought, what what if I mean, this is also very cool, the network effect around it.

如果所有的安全机制都到位了，为什么它不能像维基百科一样，每天、甚至每分钟都更新呢？

If all the guardrails are put in all of that, why can it not be like a Wikipedia being updated every day, like, every minute?

这才是使用这种技术的绝佳方式。

Like, that's a great way of using something like this.

对吧？

Right?

我的意思是，你并不会开放你的个人数据等等。

I mean, you you don't give access to your personal data, etcetera.

但既然外面已经有了所有这些保护措施，为什么不建立一个知识管理系统或知识库呢？

But with with all that is there outside, with all the guardrails being there, have a knowledge management system or a knowledge base.

在这种情境下，这样的应用会是一个不错的案例。

Something like that could be a good use case for that context.

亚伯拉罕，关于这一点，也许最后问你一个问题，因为从这个非常怀疑的小组来看，你是对这个特定故事最持怀疑态度的人。

Abraham, maybe a final question on this because it seems like out of this one out of this very skeptical panel, you are the most skeptical about this particular story.

你知道，让我来整理一下我最近一直在思考的一个正面论点，那就是，人类以前用过一种叫 Stack Exchange 的东西。

You know, I and let me put together, like, one bull case that I've been thinking a lot about, which is, you know, humans, they used to use this thing called Stack Exchange.

而 Stack Exchange 对帮助人们解决编程问题非常有用，因为你能够向其他遇到过同样问题的人请教。

And, you know, one of the ways that Stack Exchange is really helpful for helping people out on their coding problems is that you could talk to other people who have dealt with the same problem.

这就像加速了你解决工程问题的过程。

And that was like an accelerant to solving the engineering problem that you had.

所以，我的意思是，我们可以暂时搁置那些‘AI正在密谋接管世界’的说法。

And so, I mean, I think we can set aside the, like, the AIs are conspiring to take over the world thing.

但我有点好奇，对于代理来说，建立一个论坛是不是也没那么荒谬，毕竟这对人类分享技术问题的解决方案一直很有帮助。

But I kind of wonder weirdly whether or not a forum for agents is not that crazy of an idea insofar as it has been a helpful way for humans to share solutions to technical problems.

所以我很想知道，如果你认同未来确实需要这些代理能够相互协作来解决问题，那么

And so I'm kind of curious, like, if you think in the future, like, if you buy that, like, that actually like, there will be a need actually for these agents to be able to consult with one another in solving problems, and

它可能会看起来像

that it will look a

有点像我们熟悉的、在线技术解决方案论坛的黄金时代。

little bit like what we're familiar with from, you know, the the golden age of, I guess, like, technical solutions forums online.

是的。

Yeah.

所以，我得先澄清一下，我不是说Maltbook，我觉得Maltbook是

So and just to kinda caveat, it's malt book that I'm not like, malt bought, I think, is a, you know

当然。

Sure.

对。

Right.

当然。

Of course.

是的。

Yeah.

这是一个很棒的智能体。

It's a it's a great agent.

Malt book。

Malt book.

我的意思是，我读过这个。

Idea of, like, I read it.

这是为这些智能体准备的。

That's for these agents.

是的。

Yeah.

但我记得，我确实参与过那个讨论，当时我们谈到GPT是否会取代Stack Overflow的需求。

I but I mean and I remember that the the the I was actually part of the the the the episode where you we were talking about, you know, the GPT kind of removing the need for Stack Overflow.

那么，从生态系统的角度来看，会不会出现类似Reddit这样的平台？

So in terms of is there gonna be an ecosystem or, like, a Reddit, like, platform?

我认为，你现有的AI代理已经构成了这样的体系。

I I think that's really what you have as part of your AI agents are already.

所以，如果问题是：我们是否会拥有另一个论坛，在其中形成一个代理生态系统，帮助调试、解码或为特定问题构建文档？

So if if the question is, are we gonna have another forum in which we have a, you know, an ecosystem of agents kind of helping debug or decode or, you know, help build documentation for particular problems?

我的意思是，这更可能出现在特定的使用场景中，你可能会看到这种情况。

I mean, I think it's gonna be more so for specific use cases where you you may be able to see that.

关于Maltbot，我觉得真正酷的地方在于，我们一直拥有能够访问私有数据、运行外部工具和通信，并访问特定内容的代理。

In terms of Maltbot, like, where I think it's really cool is that, you know, we've always had these agents that are able to, you know, you know, access private data, run external, you know, tools and communications, and then have access to, you know, particular content.

我认为我们需要更聚焦的是，Maltbot具备的持久记忆功能，使其能够在长时间的工作流程或时间尺度上记住信息。

Where I think we need to I mean, better focus is, like, there's a there's the the persistent memory that Moldbot has that allows it to be able to remember things over long, you know, workflows or long scales of time.

我认为关键在于对使用场景进行分类，明确这些代理适合应用的领域，以及哪些场景更适合有人参与或更偏向企业级、生产级的应用。

I think it's just a matter of bucketing the use cases or bucketing where these things can be used versus where maybe a human in the loop or, you know, something a little bit more enterprise or a little bit more production focused makes a sense.

所以回到

So back

到

to

关于你的问题，你是否认为存在一个空间，可以打造类似MaltBook但更围绕特定问题组织化和系统化的东西？

your question, do I see a space for maybe something like MaltBook, but a little bit more organized and systematized around particular problems?

我认为我们现有的生态系统中已经某种程度上具备了这一点。

I think we've already kinda had that in the current ecosystem.

我不确定我是否喜欢这样。

I don't know if it's I like that.

作为某个项目的一部分，我不确定我们是否需要另一个项目来实现这一点。

Required as part of like a I don't know if we need another project in order to to build that.

我们还有另一个项目。

We got another project.

是的。

So Yeah.

是的。

Yeah.

我接受这个说法。

I buy that.

这是个很好的回应。

That's that's a good response.

是的。

Yeah.

蒂姆，让我补充一点。

Tim, if I can add on to that.

对吧？

Right?

你看，关于你是否把它看作是 Stack Overflow 或其他领域的类似平台，这里有个很大的前提。

The the see, the the big caveat to something like whether you think about it as a Stack Overflow or an equivalent of it in another domain, right?

我觉得这个想法很棒。

I think the idea is great.

这里的关键在于，这类平台之所以能成功，是因为有一个社区会实际点赞并指出：这个解决方案有效，那个解决方案无效，等等，对吧？

The big caveat there is that the reason why something of that nature worked is because you would have a community that's actually upvoting and saying, yeah, this solution actually works, this solution doesn't work and so on, right?

所以其中内置了一种验证机制。

So there is sort of an inbuilt verification enabled within it.

所以除非你有一些验证机制，否则IDS很棒，对吧？

So unless you have some sort of a verification, so IDS is great, right?

你可以使用某种代理，然后返回结果，无论是编程还是其他流程，对吧？

You could have a form of agents and then come back with whether it's coding, whether it's some other process, right?

其他领域。

Some other domain.

你可以让各种想法自由流动，但如果没有某种验证系统，就真的很难，对吧？

You could have ideas floating around, but unless you have some sort of a verification system, it's really hard, right?

否则，那就只会是一堆垃圾。

Otherwise, it's just gonna be slop.

你怎么区分垃圾和真正有效的东西呢？

Like, how do you differentiate slop between what actually works?

我只是想说，从用户的角度来看，我也在思考。

I was just gonna say that the also, I'm like, I'm thinking from a user standpoint.