AI的现状：宏观、应用与消费者

过去几年，AI领域最大的问题就是哪个模型会胜出。

For the last few years, the biggest question in AI was which model would win.

下一阶段可能更少关注模型本身，而更多关注在它们之上构建的东西。

The next phase may be less about the models and more about what gets built on top of them.

在本期节目中，Jen Kha与Anish Acharya坐下来探讨AI的未来走向，从日益激烈的模型竞争格局，到将原始智能转化为人们实际使用的产品的应用爆发。

In this episode, Jen Kha sits down with Anish Acharya to unpack where AI goes next, from an increasingly competitive model landscape to the explosion of applications turning raw intelligence into products people actually use.

他们讨论了为什么AI模型不会变成商品，开放权重模型在哪些方面有优势，以及为什么Anish认为即使前沿实验室继续壮大，应用层也能捕获显著价值。

They discuss why AI models aren't becoming commodities, where open-weight models have an advantage, and why Anish believes the application layer can capture significant value even as frontier labs continue to grow.

然后他们转向消费级AI，个人代理开始替我们购物、管理收件箱并采取行动。

Then they turn to consumer AI, where personal agents are beginning to shop, manage inboxes, and take action on our behalf.

Anish解释了为什么这可能是消费级开发者复兴的契机，以及新一代创始人可能需要更宏大地思考AI带来的可能性。

Anish explains why this could be a renaissance for consumer builders and why a new generation of founders may need to think much bigger about what AI makes possible.

为了帮我解析这一切令人难以置信的丰富内容，我要请出Anish Acharya。

To help me break down all things around this incredible abundance, I'm gonna bring up Anish Acharya.

嘿，Anish。

Hey, Anish.

谢谢，Anish。

Thanks, Anish.

你好。

Hi.

太棒了。

Awesome.

太棒了。

Awesome.

太棒了。

Awesome.

嘿，Anish。

Hey, Anish.

这周早些时候，我和Anish参加了一个GP的场外会议，他告诉我他已经开始用GrackBot，还给它买了好多牛仔裤。

Anish and I were at a GP off-site earlier this week, and he shared with me that he's already running GrackBot and has purchased a bunch of jeans for him.

那么，Anish，你想分享一下你买了什么吗？

So, Anish, do you wanna drop what you purchased?

真的，没骗你。

True story.

真的，没骗你。

True story.

对。

Yes.

我要透露一个重要的受保护的秘密，那就是我主要穿Frame的牛仔裤。

I'm gonna reveal an important secret protected IP, which is that I mostly wear frame jeans.

Frame是个很好的品牌。

Frame is a great brand.

而且Grock Bots确实是个很棒的产品。

And Grock Bots is an awesome product, actually.

我觉得Grock Bots的一个显著特点就是它很会想办法。

I'd say the kind of defining characteristic of Grock Bots is sort of resourcefulness.

前几天晚上我睡觉前说，嘿，给我买一条受这些启发的牛仔裤。

I went to bed a few nights ago and said, hey, buy me a pair of jeans that are inspired by these.

我拍了一张我现在穿的牛仔裤的照片。

I took a photo of my current jeans.

我说，花超过500美元，把它搞定。

I said, spend more than $500 and get it done.

我早上醒来，它已经研究过了，找到了一条，版型一样，颜色不同，用我的信用卡买了，现在正在路上。

I woke up in the morning, and it had researched, found a pair, same fit, different wash, used my credit card, purchased them, and they're on the way.

所以我觉得这会是越来越多见的事情。

So I think that is gonna be something that we see more and more of.

我们已经有了这些能力，现在很多突破将来自于资源fulness，以及以大多数消费者能理解的方式呈现的产品架构。

We already have the capabilities, and now a lot of the kinda unlock will come from resourcefulness and also the kind of product architecture delivered in a way that most consumers can understand.

太棒了。

Awesome.

太棒了。

Awesome.

太棒了。

Awesome.

对。

Yeah.

我跟团队说了，我要让我的机器人终于处理掉那堆东西——我答应我丈夫要卖掉它们已经两年了。

I told my team that I'm gonna set my bot to finally take care of the pile of things I've been promising my husband that I'm gonna sell for the last two years.

这就是这个周末的项目。

That is the project for this weekend.

所以，Anish，我们之前问过一个问题。

So, Anish, we'd asked the question earlier.

今天的人工智能领导者中，哪一个会在三年后成为明确的赢家？

Which one of today's AI leaders will be the clear winner in three years from now?

你怎么看？

What's your take?

我觉得会有很多赢家。

I'm a many winners guy.

对。

Yeah.

我和你们很多人一样，都是看好多个赢家的。

I'm in good company with many of you.

你看过去两周发生的事，X AI 在模型方面从根本算不上真正的竞争者，一下子变成了三强之一。

I mean, if you look at what's happened in the last two weeks, X AI went from not even being a real contender on the model side to being one of three.

所以，我们非常意外地从两强争霸变成了三强鼎立。

So we extraordinarily went from a two horse race to a three horse race.

再往大了看，这一年里，Anthropic 之前感觉强得不行，好像做什么都不会错，结果现在 OpenAI 又连续三个月表现非常出色。

And even more broadly over the course of the year, we went from anthropic feeling like they were so dominant, they could do no wrong, to OpenAI who's just had an excellent three months.

新模型确实非常厉害。

The new models are exceptional.

新的 Codex 工具和 ChatGPT 桌面应用也做得非常好。

The new Codex harness and chat gbt desktop app is very well done.

而且我们看到这些实验室在朝着不同方向专业化发展，同时规模也都在疯狂增长。

And we're seeing the sort of specialization in different directions of these labs are both growing like crazy.

尽管彼此都在不断取得成功，X AI 以及它的开源权重模型也在快速崛起。

Despite each other's continued successes, X AI and his open weight as well.

所以我绝对是站在‘多方共赢’这一边的。

So I'm definitely in the many winners camp.

对。

Yeah.

有意思的是，X上的舆论风向也值得关注。虽然它不总是完美的未来风向标，但往往能提前反映出至少开发者群体的情绪。

It's interesting to see the sentiment also on X, which is not always a perfect, you know, weather vane for the future, but oftentimes a early indicator of at least where developer sentiment is.

而且最近似乎有很多来自云服务方面的反对声音，比如关于token使用量之类的。

And there's been a lot of pushback from cloud, it seems like recently on people in terms of token usage or etcetera.

所以开发者们在这些事情上往往是‘顺风倒’的。

And so developers tend to be fair weather fans on these things.

他们会涌向最新、最棒、最强的模型，尤其是在过去六到八周里。

They will go where the latest and greatest and very best model isn't in particularly the last six to eight weeks.

我觉得在活动流量方面，我们会看到一些非常有趣的趋势。

I think we're gonna see some very interesting traction in terms of the flow of activity.

但显然，Anthropic今年晚些时候要上市，而且大家对这件事也非常关注。

But obviously, Anthropics going public later this year, and also there's a lot of keen interest on this.

那么，这正好把我们带到了今天讨论的主题。

So with that, that actually brings us straight into the topic of discussion today.

那么，智能的下一个前沿在哪里，又是什么？

So where and what is next in the next frontier of intelligence?

太棒了。

Amazing.

谢谢你，Jen。

Thank you, Jen.

那么，让我为大家开场，如果你们有问题，随时可以插话。

So let me tee this up for everybody, and please hop in if you got questions.

我们先从宏观层面和市场动态说起。

So let's first cover the kind of macro and what's happening at a market level.

然后我们会进入应用层，大致聊聊为什么应用是智能原语的产品化。

Then we're gonna hop into the application layer broadly and sort of talk through why applications are the productization of the intelligence primitive.

最后，我们来谈谈消费者，比如Dropbox和其他一些产品的发布。

And then finally, let's talk about consumer, you know, with the launch of Dropbox and a few other products.

最近这两周在消费领域确实很有意思。

It's actually been a very fun couple of weeks in consumer.

好的。

Okay.

希望我们的好朋友利奥波德不会介意我拿他的情境感知开个小玩笑。

Hopefully, our dear friend, Leopold, doesn't mind me poking a little fun at him here with situational awareness.

请吧。

Please.

下一个。

Next.

好了。

Alright.

你看。

Look.

我觉得关于这是不是泡沫的讨论已经有点过度了，或者说已经讨论得很充分了。

I think that the kind of case for this being a bubble is over sort of discussed or at least fully discussed.

我觉得，实际上，一个较少被讨论的分布外话题是，如果我们不够乐观会怎样？

I think, actually, the out of distribution topic that's less discussed is what if we're insufficiently optimistic?

如果你看看一些底层指标，它们指向的本质上就是无限的需求和高度受限的供给。

And if you look at some of the underlying indicators, what they point to is essentially infinite demand and highly constrained supply.

比如B200，虽然不是最尖端的GPU，但其每小时价格却在上涨。

Things like b 200, which is a non sort of cutting edge GPU prices going up on a per hour basis.

这非常奇怪。

That is very strange.

通常，我们看到这些产品是高度通缩的，这似乎指向了非常紧张的供给和几乎无限的需求。

Normally, we see these things be highly deflationary, and it sort of points to very constricted supply and essentially infinite demand.

所以我们正在思考和讨论很多，基于这些二阶指标，乐观的合理依据是什么。

So we're thinking and talking a lot about what's the kind of informed case for optimism here given some of these second order indicators.

SaaS泡沫，或者说SaaS的起起落落，是市场心理的一个有趣缩影。

The SaaS bubble was a very or the SaaS sort of lip saw was an an interesting peek into market psychology.

二月份，当我们看到许多SaaS公司股价下跌30%到40%时，我们说市场对软件板块超卖了。

Back in February, when we saw this 30 to 40% drawdown on a bunch of SaaS names, we said that the market is oversold software.

果不其然，现在就是这样。

Lo and behold, here we are.

很多这些公司的股价又涨了40%。

Many of those names are back up 40%.

所以我不太确定我们大家一起到底完成了什么。

So I'm not quite sure what we collectively accomplished.

但我可以告诉你我们当时说的，现在依然成立，那就是企业软件支出占8%到12%。

But I'll tell you what we said then, which is still true today, which is for the enterprise software spend is eight to 12%.

这其实不是支出的大头。

It's just not a huge proportion of spend.

所以用AI写代码来做自己的工资单或客户关系管理，上升空间并不大。

So the upside to vibe code your own payroll or CRM is not particularly high.

但下行风险几乎是无限的。

The downside is essentially unlimited.

显然，如果工资单这类事情没做好，会有各种合规问题。

Obviously, there's all kinds of sort of compliance implications of not getting things like payroll right.

所以现在大多数企业软件都要求很高的精确度，而编程代理根本达不到这个要求。

So most enterprise software today demands a level of precision that just isn't afforded by coding agents.

不过有一件事确实发生了，就是潮水退去了。

The one thing that has happened though is the sort of tide has receded.

所以很多之前靠大量股权激励之类的东西掩盖了真实业绩的SaaS公司，现在问题都暴露得很明显了，它们要么加速转型，要么就完蛋。

So for a lot of SaaS companies that had a ton of SBC and things that distorted their economic performance, I think that's very much visible now, and they're going to have to sort of accelerate or die.

所以SaaS市场比我们前几个月想的要悲观一些，但还是有一些生死攸关的问题要解决。

So let's bleak for the SaaS sort of market than perhaps we all collectively thought for a few months there, but still some sort of existential questions to address.

关于护城河的讨论特别多。

There's been a huge sort of discussion of moats.

还有护城河吗？

Are there any moats?

已经没有护城河了。

There's no more moats.

这很有意思，因为如果你真的研究过护城河——最著名的就是《七力》这本书里总结的，那是我最喜欢的书之一——你会发现绝大多数护城河其实并不受低成本智能的影响。

And it's very funny because if you actually study moats, which I think are most famously codified in the book seven powers, it's one of my favorites, the vast majority of moats actually are not affected by abundant low cost intelligence.

你知道，当你想到网络效应、规模效应（体现在分销上）、品牌效应（我们在硅谷往往低估它），这些效应从来没有像现在这么好过。

You know, when you think about network effects, scale effects, which shows up in distribution, brand effects, which we tend to discount in Silicon Valley, these things aren't as good as they've ever been.

再多的编程代理也不会让耐克不再是耐克。

No amount of coding agents is gonna make Nike not Nike.

Instagram 的力量从来不是构建 Instagram 应用的复杂性。

The power of Instagram was never the complexity of building the Instagram app.

当然，它背后是那个网络。

Of course, it was the kind of network behind it.

所以我实际上认为大多数护城河和以前一样坚固，而且它们仍然是创造复利价值的关键。

So I actually think the majority of moats are as good as they've ever been and, of course, are still critical to building compounding value.

有几条护城河是暴露的。

There are a couple of moats that are exposed.

对我来说，集成模式是最明显的一个。

For me, the integration mode is the most obvious one.

SAP 的集成和退出集成是出了名的复杂。

SAP is so famously complex to integrate into and out of that.

哪怕只是从一个SAP版本迁移到下一个版本，都是一种生死攸关的风险。

It's a sort of existential risk to even migrate from one version of SAP to the next.

编码代理让这种情况大大改善。

Coding agents makes this dramatically better.

我觉得对于FIs和GSIs来说，其实存在一个根本性的问题：当它们历史上一直是这种集成节点时，它们的价值会变成什么？

I think there's a bit of an existential question actually for FIs and GSIs as to what will their value be when they've historically been this sort of point of integration.

所以我认为这个护城河确实有点风险，但其他传统护城河依然存在，而且和以前一样重要。

So I do think this moat is a little bit at risk, but for the other traditional moats, they persist and they're as important as they've ever been.

我觉得这是一个非常重要的概念。

And I think this is a really important concept.

当你开始思考企业里哪些职能岗位能创造超额收益时，通常是产品、销售、工程和研究。

As you start to think about what are the job functions in the enterprise that are alpha creating, it's typically product, sales, engineering, research.

反过来，企业里哪些职能岗位可能说行政有点太悲观，但它们是支持其他职能的，比如法务、人力资源、财务等等。

And conversely, what are the job functions in the enterprise that are sort of maybe administrative is too bleak, but they are supporting other functions, legal, HR, finance, etcetera.

我们真的认为，合理的架构，也是正在出现的架构是：对于像销售或产品这样有无限上升空间的岗位，你总是想用前沿模型。

We really think that the kind of rational architecture and the one that is emerging is that for jobs that have unlimited upside like sales or product, you always wanna use Frontier Tokens.

原因在于，你根本不知道新功能或拿下客户账户能带来多少价值。

And the reason for that is you just don't know what the value of the new product feature or closing a customer account is.

这实际上是无上限的，所以从经济角度看，为哪怕智商高一点的模型付任何价格都是合理的，比如你的Fable 5、Grok或GPT 5.6。

It's effectively unbounded, and therefore, it's economically rational to pay almost any price for a model that's even one IQ point smarter, your Fable five or your Grok or your GPT five six.

相反，谈到财务这类事情，最好的结账方式就是准确。

Conversely, when you talk about something like finance, the best way to close the books is accurately.

你不可能比准确还要好十倍。

You can't close it 10 x better than accurately.

所以，这就成了一个有上限的问题，在这种情况下，使用开源权重模型配合强化学习，在帕累托效率成本曲线上是合理的。

So as a result, you kinda have this bounded upside problem where it makes sense to use open weight models with reinforcement learning for the kind of Pareto efficient cost curve.

也许在我们深入这个话题之前，因为这是个有争议的话题。

And maybe before we we go at this one, because this is a a degree debate.

而且，几周前Kimmy发布时，关于这个话题有很多争议，因为相对成本是讨论的焦点。

And, when Kimmy dropped a few weeks ago, there was a lot of consternation about this topic just given the relative cost, which was the focus of of the topic of discussion.

但我们的创始人，来自Decagon的Jesse Zhang，发了一篇很棒的文章，指出在某些方面，对于像Decagon这样的很多公司来说，开源实际上是唯一的选择。

But our founder, Jesse Zhang from Decagon, dropped this great post around the fact that in some respects, and for a lot of companies like Decagon, open source is actually the only option.

这不仅仅是成本的问题。

It's not just cost.

而是他们实际上可以本地化部署、训练和微调它。

It's that they can actually localize it, train, fine tune it.

所以，也许可以稍微展开一下这个配置，聊聊其中的细微差别，以及为什么创业者们不必担心，尽管事实如此，这个话题其实有很多可用的资源。

And so maybe unpack a little bit of that configuration, talk through the nuances there and why folks shouldn't be concerned, even though that is the case for startups, that there's a lot in the way of abundance around this topic.

对。

Yeah.

我的意思是，我们看到的其中一个重要话题或趋势是，不同模型之间确实存在各自的比较优势。

I mean, one of the big topics that we're seeing or one of the big trends is that there are just one that are sort of comparative advantages of different models.

而且这些模型往往有各自的专注领域，这些领域之间甚至有些相互矛盾。

So and the models often have sort of areas of focus that are almost at tension with each other.

你会看到某些模型具有高度的神经质倾向。

You see a certain set of models that have a high degree of neuroticism.

比如，它们有点像自闭症模型。

Like, they're sort of autistic models.

GLM五二和GLM五三就是很好的例子，它们非常死板，只做你让它们做的事，别的什么都不做。

GLM five two and GLM five three are great examples of this where they're very literal and only do exactly what you told them to do and nothing more.

然后我们看到像k3这样的模型，它们就开放得多。

Then we're seeing models like a k three that are just much more sort of open.

它们很自以为是，而且很有创意。

And they're very presumptuous, and they're creative.

在组织里，这两种模型都有各自的用武之地。

And there are rules for both types of models in the organization.

而且，这些模型的思维方式往往互相冲突。

And often, the sort of shapes of those minds, if you will, are at odds with each other.

所以，这就是为什么你实际上需要多个模型的原因之一。

So that is like one reason you actually want to have multiple models.

强化学习确实是个非常重要的点。

The reinforcement learning is a really important point.

如果你真的有一个问题，可以通过推理轨迹让模型专门化，你就能在自己的领域里为客户群建立起一种复利优势，从而能够把智能塑造成比任何通用智能都更适合你的问题。

If you actually have a problem that you can specialize the model around with your reasoning traces, you can start to create this compounding advantage in your domain for your customer base, where you're able to kind of shape the intelligence to be better than any general intelligence for your problem.

哈维在这方面也取得了一些很好的成果。

Know Harvey's had some great results with this as well.

不过，这种强化学习的代价就是会失去通用性。

Now the trade off of that kind of reinforcement learning is you lose generality.

所以，如果你有一个专门针对法律问题优化的模型，它可能不太擅长解决理论数学问题，这对哈维的用途或者迪卡贡的客服来说都没关系。

So if you have the best sort of model that's fine tuned for solving legal problems, it may not be great at solving sort of theoretical math problems, and that's okay for Harvey's uses or in the case of Decagon customer support.

这种开放式的专业化特性非常独特，也是我们的初创公司选择它们的原因之一。

So this sort of open way specialization property is something that's very unique and one of the reasons our startups are selecting them.

这也是个大话题，自从一月份以来我们学到了很多，我们真的应该每个月都做一次，珍，或者，说实话，每周一次，变化实在太快了。

This is also a big topic, we've learned so much since January, we should really do this monthly, Jen, or Yeah, I mean, honestly, weekly, there's just so much changing.

在一月、二月的时候，有很多讨论，而且非常独特有趣。

So in January, February, there was a lot of discussion, and it's very idiosyncratic and interesting.

你知道，Anthropic Quad发布了一个所谓的法律插件。

You know, Anthropic Quad released what is called a legal plugin.

你知道，插件其实就是技能文件的集合。

You know, plugins are just collections of skill files.

你可以把它想象成一个技能文件的压缩包。

You can think of it as a zip of skill files.

技能文件其实就是提示词。

Skill files are just pros.

它们就是很长的提示词。

They're just long prompts.

然后当时就引起了一阵恐慌，突然之间，汤森路透和其他一些大型法律公司股价大幅下跌。

And there was this huge panic, and all of a sudden, Thomson Reuters and a bunch of other sort of, you know, big legal names traded down dramatically.

但那些其实只是提示词，当时有很多讨论，说实验室会不会垂直整合，向上整合到应用层。

But those were really just prompts, and there was a lot of discussion about if labs were going to integrate vertically integrate up into the application layer.

结果恰恰相反，我们看到的是，是的，他们在垂直整合，但他们是向下整合到推理和计算层面。

Instead, we've seen the very opposite, which is, yes, they are vertically integrating, but they're vertically integrating down into inference and compute.

现在回头看，这其实很合理，因为推理的工作负载非常同质化。

It's actually logical now in hindsight because the workloads for inference are very homogeneous.

所以你可以把价值链的某一部分做到极大的规模。

So you can build enormous scale in one part of the value chain.

而当你想到应用层的时候，你知道，有很多特殊情况和独特需求，比如定价、打包、产品化，以及市场想怎么购买。

Whereas when you think about the application layer, you know, you've got so many idiosyncrasies and unique needs in terms of pricing, packaging, sort of productization, how the market wants to buy.

所以，进入应用层实际上比进入推理层更具挑战性，而且运营支出也更重。

So it's actually a much more challenging and OpEx heavy proposition to move into the application layer versus moving down into the inference layer.

这就是我之前提到的那个点，也就是关于模型商品化的讨论。

And this is the point I alluded to earlier, which is sort of this discussion of model commoditization.

你知道，如果你每天都用这些模型，就像我一样，我对自己有个标准，就是每个新模型出来，我都要用它做点东西，不管大小。

You know, if you use the models every day, which I do, I sort of hold myself to a standard of making something either small or big with every model that comes out.

你开始意识到这些东西并不是商品，它们在领域层面有比较优势。

You you start to appreciate the fact that these things are are not commodities, that they have comparative advantage at a domain level.

一个很好的例子就是OpenAI，他们的新GPT模型在知识工作方面非常非常出色。

So a great example is open AI with their new GPT models are just so so good at knowledge work.

那个框架也为知识工作设置得很好。

The harness is also very well set up for knowledge work.

你知道，如果你用过ChatGPT桌面应用，你就明白我的意思了。

You know, if you've used the chat GPT desktop app, you know what I mean.

如果你还没装，赶紧去装一下。

If you haven't, please install it.

真的特别酷，也很有意思。

It's very, very cool and interesting.

我刚才说的“平台”，其实更像是一个产品容器，比如浏览器那样。

And it's the perfect sort of when I say harness, I kinda mean kind of product container like a browser.

它是做表格、幻灯片演示和文档写作这类工作的完美载体。

It's the perfect product container to do spreadsheets and slide presentations and written documents and all of that type of work.

如果你看看 Codeium，我相信在座很多人都在用，它的设计完全是面向软件工程的。

If you look at quad code, which many of you I'm sure have used, it's just so oriented towards software engineering.

它是基于终端界面的，从那些细微的设计决策到它的专长领域，比如代码规划和代码测试，都是专门为软件工程师打造的。

You know, it's in a terminal UI, everything from the small design decisions to the areas in which it specializes, like code planning and code testing, is oriented towards the software engineer.

这两款产品在各自的专精方向上都做出了不少取舍。

And there are many trade offs both products are making for that sort of respective specialization.

所以你看，这种针对特定领域的专业化分工其实已经出现了。

So one, you've kinda got this domain level specialization that's already occurring.

第二点，就像我之前提到的，有一种我称之为“五大性格特质”的东西，如果大家研究过的话。

And then two, as I mentioned earlier, you've got this sort of I think of it as the big five sort of personality traits if if folks have studied that.

你知道，一个人不可能既高度开放又高度神经质。

You know, you can't be both highly open and highly neurotic.

而且，有时候当你把智能应用到会计问题上时，你希望它更神经质一些。

And, you know, sometimes when you have an intelligence you're applying to an accounting problem, you want neuroticism.

当你把它应用到设计问题上时，你希望它更开放。

When you're applying it to a design problem, you want openness.

所以实际上，组织里需要这两种思维，这就是为什么你会选择像GLM-5.3而不是KEK-3这样的模型。

So you actually have a need for both types of minds in the organization, which is why you would select something like a g l m five three versus a k e k three.

所以在我们看来，它们绝对不是同质化的商品。

So definitely not commodities in our view.

这一点很重要。

This is an important point.

你知道，在很多产品类别中，模型聚合带来的效果会大于各部分之和。

You know, there are many product categories in which model aggregation delivers a greater than sum of parts outcome.

你知道，一个好的比喻就是 Expedia。

And, you know, a good metaphor for this is Expedia.

你知道，用 Expedia 比直接去美联航、达美或西南航空的网站要方便得多。

You know, it's so much more useful to use Expedia than it is to go to United than to go to Delta than to go to Southwest.

你只需要一个地方，能看到所有航空公司的库存。

You just want a single place where you can benefit from seeing every airline's inventory.

同样，在编程领域，我们在 Cursor 上看到很多这样的情况：比如你想用一个前沿模型来做规划，但执行时可以用一个较小的模型。

Similarly, you know, in coding, we're actually seeing this with cursor a ton where you want to do a very frontier model for planning, for example, but then you can use a lesser model for execution.

你真的需要一个产品框架或产品架构，让你能同时使用多个模型。

And you really need to have one product harness or sort of product architecture that lets you use multiple models.

创意工具是另一个很好的例子，有些模型专门擅长不同的模态。

Creative tools is another great example where you've got, you know, models that specialize in different modalities.

比如像 Eleven Labs 这样的工具，当然在语音和音乐方面非常出色。

So you've got something like an 11 labs, which, of course, is incredible at voice, music as well.

还有像 Black Forest 这样的工具，在视频和创意指导方面做得非常好。

And And you've got something like Black Forest, which is doing such an excellent job in kind of video and and creative direction.

正确的产品就是把所有这些整合到一个统一的界面里。

And the correct product is to bring all of these together into one shell.

最后，还有研究和决策。

And then finally, research and decisions.

我们经常看到这种情况，模型通常是用不重叠的数据集训练的。

We see this all the time where, you know, the models are trained with sort of non overlapping datasets often.

所以，你可以通过把同一个查询交给多个模型来对抗性地获取更多信息，然后让另一个模型帮你收敛结果。

So you're able to just get more information by running the same query through many models adversarially and then having a separate model sort of help you converge.

这正是应用层大放异彩的地方，因为实验室们既受激励驱动，又在结构上只能提供自家模型。

This is a place where the application layer really shines because labs, of course, are both incentivized and structurally only able to provide their own in house models.

而你作为应用层的聚合者，可以提供最优质的方案。

You as an application sort of aggregator can provide the best of breed.

好的。

Okay.

让我们进入应用层。

Let's jump into the apps layer.

应用层的关键在于，智能就像购买云服务一样，是一种基础能力。

Now the key point about the application layer is that, you know, intelligence is a primitive, just like buying cloud as a primitive.

Salesforce 是怎么做的呢？

And what does Salesforce do?

他们把 AWS 的云服务这种底层能力，转化成了 CRM 软件，为各类客户群体带来实际的经济价值。

It sort of takes the, you know, AWS cloud primitive and turns it into CRM software that delivers an economic outcome for all of their customer segments.

AI 应用层也是同样的道理。

The same thing is true of the AI application layer.

拥有基础的智能能力固然很好，但真正需要的是像 Harvey 这样的工具，把这些能力转化为法律行业的实际经济效益。

You know, it's great to have the raw intelligence primitive, but you really need Harvey to turn that into an economic outcome for the legal industry.

信用合作社的情况也类似。

Similar for somebody like credit unions.

这是一个非常有趣的细分市场，因为他们在购买产品的方式、产品的形态以及市场目标的设定上，都有着非常独特的偏好。

It's a really interesting market segment where they're so idiosyncratic in how they want to buy products, how they want the product sort of productized, and and the shape of the ambition for their market.

大多数信用合作社并不希望将员工人数减半。

You know, most credit unions don't wanna decrease their headcount by half.

他们想翻倍。

They wanna double it.

对，他们希望在保持业务盈利的同时实现翻倍。

Right, and they want to double it while having an economically performing business.

所以，这就是他们对智能基础能力在其市场领域中发挥作用的一种非常特定的看法。

So it's just a very specific way that they see the intelligence primitive playing out in their market segment.

而应用层的机会就在于成为那个实现这一点的角色。

And the application layers opportunity is to be the one that kind of delivers that.

这是一个有点高级的概念，但我认为很重要。

This is a bit of an advanced concept, but I think an important one.

如果你看看AI使用方式的演变，它已经从提示模型发展到将模型放入循环中，你知道，“代理”这个词被过度使用了。

If you look at the kind of way that the evolution of AI use has gone, it's gone from prompting models to putting models in loops, you know, the the term agent is overused.

但代理其实就是模型在循环中，加上工具、记忆和其他一些东西。

But agent is just a model in a loop with sort of tools and memory and a few other things.

一个很好的例子就是编程。

A great example of this is coding.

我们都见过软件公司这种情况：一个bug被报告，被复现，然后生成修复方案，再验证。

Know, we've all seen this from software companies, which is a bug gets reported, it gets reproduced, a fix gets generated, it gets verified.

如果是低风险的修复，就直接集成并发布了。

If it's a low risk fix, it gets integrated and shipped.

也许客户会收到一封邮件，说你的bug已经修复了。

And maybe the customer gets an email saying your bug was fixed.

如果是高风险的改动，可能会有人工审核。

If it's a high risk change, perhaps a human reviews it.

但这样一来，企业收到的每一个bug都能通过这个编码循环自动修复。

But that way, every bug that actually gets reported to the enterprise now gets autonomously fixed through this coding loop.

当你把这个思路应用到业务的其他部分，比如价格优化、采购，这些都是很自然的业务循环，完全可以由这些模型自动化处理。

As you start to take that idea and apply it to other parts of the business, things like price optimization, things like procurement, these are very natural sort of business loops that occur that can be fully automated by these models.

然后，可能最有野心的循环是业务循环，就是说，你做了一个对业务影响很大的改动，模型会回来说：我觉得我们需要在蒂华纳开个分公司。

And then perhaps the most ambitious type of loop is the business loop, which is, hey, you make a change that's very cross cutting to the business, and the model comes back and says, hey, I think we need to open a branch in Tijuana.

但模型不能自主完成这件事。

Now the model can't do that autonomously.

但它可以在整个业务的表层做出改变，这非常了不起。

But it can make a change at the sort of surface level of the entire business, which is extraordinary.

这就是企业自动化将通过AI实现的方式。

This is how enterprise automation is going to occur through AI.

对我来说，编程一次又一次地证明了这一点，法律是另一个很好的行业例子，而不是市场。

I think for me, coding has just been over and over again, an illustration, legal is another great area of industries, not markets.

这是马克常说的话，他说得非常对：如果你能把智能看作一种基础能力，那我们现在就把编程智能也看作一种基础能力。

This is something that Mark says, and he's so right, which is if you can look at intelligence as a primitive, let's think now about coding intelligence as a primitive.

所有这些产品都在各自的技术栈领域里发挥作用。

All of these products are working in the sort of respective areas of the stack.

你知道，Quad Code在把原始硬件暴露给开发者方面做得非常出色，一直到Replant，它为不熟悉代码的普通小企业主提供了一个很好的抽象层。

You know, quad code is such an excellent job of kind of exposing the raw hardware, so to say, to the developer, all the way up to replant, which is a great abstraction layer for the average small business owner that's unfamiliar with code.

这些都是对编程智能和智能本身在定价、产品化、包装等方面的不同变体，而且它们都在发挥作用。

These are variations of sort of pricing, producting, pack productization, packaging for the coding primitive and intelligence, and all of them are working as a result.

所以我认为我们一个很大的思维模式转变是，确保我们把这些当作行业来评估，而不一定是简单的市场。

So I think a big mental model shift for us is ensuring that we're assessing these as industries, not necessarily simple markets.

好的，消费者这边最近几周确实很精彩。你知道，我们三年来一直说这会是消费者的季度，但我觉得这次可能真的是消费者的季度了。

Okay, and consumer consumers had a really cool couple of weeks, you know, we've been saying for, you know, for three years that this is going be the consumers quarter, but I think that this might be consumers quarter.

我们深入聊聊吧。

Let's go into it.

到目前为止，阻碍消费者发展的其实有几个因素。

The things that have actually held back consumer so far have been a couple of things.

首先，消费者不太喜欢为软件付费。

You know, the first is consumers don't love paying for software.

这个教训我们反复学到了。

We've learned this lesson over and over again.

而且不幸的是，和过去软件的魔力不同，AI软件在分发和用户互动上有边际成本。

And unfortunately, unlike the sort of magic of software in the past, AI software has marginal costs of distribution and engagement.

这些边际成本有时候会非常高。

And the marginal costs can sometimes be very high.

你知道吗，我建了一个应用来帮我浏览我的X时间线，每拉一个新用户就要花250美元。

You know, I built a an app I use to help you browse my x timeline, and it costs $250 to onboard a new user.

所以如果我是初创公司的创始人，看到这个，即便初始成本是零，250美元的成本也很难让一个面向大众市场的免费产品运转起来。但现在情况正在改变，因为开放权重模型大幅降低了成本，性能也更强了。

So if I'm a startup founder looking at that looking at a kind of $250 even with a $0 tack onboarding cost, it's very hard to make a mass market free product work that is changing now because of open weight models dramatically cheaper and more performant.

第二点是，我们从来没有过AI原生的分发渠道，没有AI的App Store。

You know, the second is we've never had an AI native distribution channel, there's no App Store for AI.

所以这个消费产品的实际周期更像Web 2.0时代，你必须同时构建渠道和产品，而不是像移动互联网那样，整个生态系统有一个集中的分发点。

So this actual product cycle for consumer looks more like web two point o, where you have to kind of build the channel alongside the product, and less like mobile, where you actually have the central point of distribution for the entire ecosystem.

最后一点，我觉得很重要，命令行界面——我们基本上处于AI的DOS时代。

Then the final point, think is an important one, you know, command line is we're sort of in the DOS era of AI.

要让这项技术及其能力被消费者完全接受，我们需要一个所谓的Windows。

And for this technology and its capabilities, it sort of fully be embraced by consumers, we're going to need the windows, so to say.

所以我认为在产品设计和工艺方面还有很多工作要做，以确保消费者知道如何使用这些神奇的新功能。

So I think there's just a ton of work to be done around product and design craft to ensure that consumers know how to consume all this magical new capabilities.

有两件事正在起作用。

Two things are working.

编码代理非常出色，我知道我们一直在讨论，想想它们如何为消费者工作很有趣。

So coding agents are extraordinary, I know have been discussing, it's interesting to think about how they work for consumers.

现在，如果你想想这个数字原生企业家的概念，如果你不是程序员，历史上这种情况通常表现为你是个YouTube创作者。

Now, if you think of this concept of the digitally native entrepreneur, if you're not a programmer, way that's historically shown up is you're a YouTube creator.

而且十年前我们有过一阵道德恐慌，说孩子们想当YouTube创作者，而不是宇航员。

And there was a whole moral panic that we had, you know, ten years ago about the kids want to be YouTube creators, not astronauts.

但我会反过来理解，那些在互联网上长大的孩子其实想在互联网上创业，而唯一的方式还是当创作者。

But I would interpret that instead as the kids actually who grew up on the Internet wanna build businesses on the Internet, and the only way to do it again is being a creator.

现在有了编程代理，你可以构建一个软件产品，每年产生10万美元的收入，甚至100万美元的收入。

Now with coding agents, you can build a software product that generates a $100,000 of revenue a year, a million dollars of revenue a year.

这些不是风险投资支持的企业，但这是一个类似夫妻店的SaaS机会，正在兴起，我觉得对国家来说非常非常酷。

Now these are not venture backable businesses, but it's a sort of mom and pop SaaS opportunity, which is emerging and I think very, very cool for the country.

个人代理，我们在1月份对Open Claw有过集体兴奋的时刻，那是一个非凡的原始组件组合，但它从未真正跨入消费者领域。

Personal agents, we had this collective moment of excitement around open claw in January, and it was an extraordinary sort of composition of primitives, but it never really crossed over into consumer.

你知道，那更像是面向开发者的东西，更像是家庭酿造计算机俱乐部的能量。

You know, was sort of a developer oriented thing, more of the homebrew computing club kind of energy.

随着GrockBot和ChatGPT工作的出现，我们开始看到个人代理被转化为消费者可以使用的软件。

We're starting to see with the emergence of GrockBot and ChatGPT work, personal agents being turned into software that consumers can use.

Anish，其实在我们看城镇演示之前，你能先停一下吗？

Anish, actually, do you mind just pausing on this before we go to the town demo?

因为，你知道，你是在上一代消费应用体验时代创业的创始人。

Because, you know, you you were a founder building in the last era of the the consumer app experience.

而且我甚至一想到这个，就觉得，天哪，现在到底怎么定义消费者？

And when I even think about it, I like, gosh, how do you even define consumer today?

因为，你知道，现在水管工用Grock Bot来彻底改造他们的整个业务。

Because, you know, the plumber that utilizes now Grock Bot to completely turn around their business end to end.

这算消费者还是企业？

Like, is that consumer or is that enterprise?

因为，这很像，几乎像是PLG驱动的。

Because, like, it's very like it's almost like a like PLG led.

对。

Yeah.

这种趋势，但它是从消费者开始的，然后跨界到企业。

Movement, but but it's coming from as a consumer that then crossover into enterprise.

而且特别是，消费应用的最后一个时代更偏向于以娱乐作为变现方式。

And and particularly, like, the last era of consumer application is more towards entertainment as a way to to monetize.

所以也许可以展开讲讲这部分，特别是你在这方面投入的时间。

And so maybe unpack some of that and and particularly where you've been spending time as a part of that.

我的意思是，我们的简单规则是，如果你无法通过销售来证明获取客户的合理性，这通常意味着15K的年度合同价值，那你就必须通过营销来获取他们。

I mean, our simple rule is if you cannot justify acquiring a customer through sales, which usually means a 15 k ACV, you have to acquire them through marketing.

我们把他们视为消费者，而大多数小企业主就是这样的。

We think of them as a consumer, which is most small business owners.

所以我认为，在我们的投资思维中，水管工绝对是消费者。

So I think that the plumber is definitely the consumer in our sort of investing mind.

娱乐领域非常巨大。

Entertainment is huge.

而且将会出现一大批AI原生的娱乐公司。

And there's going to be a bunch of AI native entertainment companies.

你知道，我会说Character某种程度上就是一家娱乐公司。

You know, I would argue character was kind of an entertainment company.

短剧现在是个很大的趋势，主要在亚洲，现在也开始传到这边来了。

There's been a huge trend around short form drama, mostly in Asia, and that's starting to come over here.

其中很多都是生成式的，或者说有生成式辅助的。

Many of those are generative or sort of generative assisted.

所以你看，我觉得娱乐这块会非常庞大。

So look, I think entertainment is going to be massive.

大多数人想的是怎么打发时间，而不是节省时间，消费者对生产力没那么感兴趣。

Most people want to spend time not save time, like consumers not that interested in productivity.

所以这肯定会发生，而且可能值得专门深入探讨一下。

So that's definitely going to happen, and probably worth a separate deep dive.

好的，我觉得对于用过Town的人来说，那体验真的很神奇。

Okay, I think town for folks who have used it, it's just such a magical experience.

你知道，这是我给所有人——朋友、家人、行业里的人——的头号建议，就是请一定要去用这些产品，因为当你看到它们每天的变化时，很容易就能建立起直觉。而Town作为一项投资，是我们的合伙人Alex Rampel做的，它真的是一个非常出色的生产力产品。

And you know, this is like the number one sort of piece of advice I give to everybody, friends, family, folks in the industry is like, please just use the products, Because it's so easy to build intuition when you see how they change day to day and town as an investment, our partner Alex Rampel made, it's really extraordinary productivity product.

你会看到产品通过记忆优势，随着时间推移不断叠加改进。

And you sort of see how the compounding improvement of the product through memory advantages over time.

所以第一天你使用一个产品时，它对你还不怎么了解。

So the first day you use a product, it doesn't know you that well.

它就像一个刚入职的新员工，还在慢慢上手。

It's sort of like an employee, a new hire who's just getting up to speed.

到了第30天，它就能很好地替你做出判断了，因为它已经吸收了三十天的上下文、记忆和技能。

By day 30, it's able to make excellent assumptions on your behalf, because it just has soaked in thirty days of sort of context, memory, and skills.

而且我们越来越多地看到这种模式，就是给最终用户带来的价值在不断增加，这体现在业务的留存率上，也体现在每个客户的定价能力上。

And this is a pattern that we're seeing more and more, the sort of compounding value being delivered to the end customer, showing up as retention in the business, and sort of showing up as pricing power on a per customer basis.

对。

Yeah.

这个例子很好，因为大家可以把Town用在个人场景里，而且它是免费试用的。

This is a great one because folks can utilize Town for their personal use case, and it's a free, you know, trial.

他们一开始会给你大概40个积分，差不多这个数。

They give you, I think, something like 40 credits to start or something around there.

所以一旦你把它接入个人邮箱，你就能看到它到底有多高效。

And so you can kind of see it once you plug into your personal email how productive it actually is.

在工作方面，我始终保持收件箱清零。

On the professional front, I'm always inbox zero.

但在个人方面，我的收件箱里有两万封邮件。

On the personal front, my inbox is, 20,000.

大卫·乔治听到这估计心里在犯嘀咕，觉得这太不像话了。

David George is is probably cringing on the inside here just because it's unacceptable.

不过确实如此。

However True.

你知道，个人生活里的事情很常见。

You know, personal life things are common.

所以如果你给我的个人邮箱发邮件，我永远不会回复你。

So if you email me on my personal, I will never respond to you.

不过，我把Town接入了个人邮箱，现在我都懒得去看了。

However, I plugged town into it, and, like, I don't even check it anymore.

如果有重要的事情，Town会主动提醒我。

If there's something important, town will surface it to me.

而且，它还会帮你清理订阅之类的，优化各种事情，现在已经开始自我改进了。

And, also, it does all the scrubbing of, like, subscriptions and all the things that it can optimize, And it's starting to now self improve upon itself.

所以它会给你发邮件，比如提醒你，这个订阅花了多少钱。

So like, it'll send you emails where it says like, hey, this routine is costing this much.

然后告诉你，这样操作可以帮你省下积分。

Like, here's how you could actually save your credits by this.

所以这算是解锁了生产力方面的一些功能。

So it's sort of this this unlock into what starts on the productivity side.

而且你说得对，可能个人用户不会为此付费。

And to your point, maybe people won't pay for that personally.

但一旦它开始被接受，然后扩展到更多领域，你就会觉得，好吧。

But once it starts to get locked in and then expand in terms of the remit, you're like, okay.

我愿意付那几十块钱，因为它能帮我管理生活，让我省心。

I'll pay the whatever x bucks, you know, just because it helps to manage my life and I can put it on autopilot.

对。

Yeah.

珍，你说得太对了。

It's it's such a great point, Jen.

而且，我对这个的理解就是，一个经验丰富的员工，一个老员工，对比一个新员工，你知道，新员工可能很聪明，甚至可能比老员工成本还低。

And, like, my mental model for this is just an experienced employee, a tenured employee versus a new hire, you know, the new hire may be brilliant, and may even cost less than tenured employee.

但我们都知道老员工的价值，他们能代表组织和你自己做出很好的判断。

But we all know the value of a tenured employee, they're just able to make great assumptions on behalf of the organization and you.

而且，你知道，这可能有点哲学，但我觉得这就是未来的方向。

And, you know, may this is a little philosophical, but I think this is where it all goes.

就像我们刚才谈到的企业里的编码循环和业务循环，我们认为还有一套非正式定义的循环，它们真正勾勒出了消费者的生活。

Just as we talked about kind of coding loops and business loops for the enterprise, we think there's a set of loops that are informally defined that really sort of lay out a consumer's life.

想想家庭、友谊、金钱、健康。

Think of family, friendships, money, health.

这些都是你在不断改变信息、做决定、采取行动、执行，然后循环继续的领域。

These are all areas where you're sort of changing information, decisions, agency, execution, and then the loop continues.

所以我们开始看到一些这样的循环在自我提升方面出现，健康和金融是OpenAI关注的两个领域。

So we're starting to see some of these sort of loops emerge around self improvement, kind of health and finance are the two areas that open AI is focused on.

我们看到很多初创公司在做购物相关的项目。

We've seen a bunch of startups working on shopping.

但我们认为，最终的结果会是消费者生活质量的大幅提升。

But we think that like the kind of way that this ends up playing out is a dramatic quality of life improvement for the consumer.

这其实跟过去的产品周期很像，80%的收益都流向了最好的市场。

That really follows the shape of past product cycles where 80% of the surplus is delivered to the best market.

Anish，你觉得所有这些——抱歉。

Do you think Anish that all of these sorry.

也许我们回到上一张幻灯片，这里有个问题。

Maybe just going back to the last slide, there's a question here.

当你想到这些个人代理的例子，不管是town还是ethos等等，都指向一个助手拥有上下文，但似乎有很多不同的选择。

When you think about these personal agent examples, whether it be town or ethos, etcetera, all point to, you know, kind of one assistant having context, but it seems like there's many different options.

你觉得最终会不会像时间管理那样，在这个个人生活方面出现一个主导平台？

Do you think it'll end up being, you know, sort of one dominant platform for this personal aspect of your life as as time management?

还是说它会像一个操作系统，有很多代理在后台互相沟通和配置？

Or will it be like an operating system where do you have many kind of talking to each other and kind of configuring on the back end?

我觉得这里可以用比较优势的观点来看。

I I think comparative advantage point kinda comes to mind.

你知道，你希望你的财务顾问具备的特点，和你希望你的派对策划人具备的特点是不一样的。

You know, I think the the characteristics you want from your CFA are different from the one that you want from your sort of party planner.

而且这个领域太广了，所以我觉得，确实会有一些重叠的背景信息。

And then just the surface area is so broad that I think that, yes, there's overlapping bits of context.

我觉得Rockbots在展示这一点上做得很好，他们有很多方向略有不同的机器人，但都协调起来，最终实现一个全局最优的结果。

And I think Rockbots has done a nice job of kind of illustrating this in product, or you have many bots that are pointed in slightly different directions that all coordinate to deliver a globally optimum optimal outcome.

有几个问题。

There's a few questions.

我想回到你之前讲过的那些话题。

I'm gonna go back to topics you've covered earlier.

如果应用层能获得经济收益，那你怎么看待来自模型公司的竞争？他们会允许价值在下游积累吗？

So if the application layer captures economic outcomes, How do you think about the competition from the model companies, and what will they allow value accretion to happen downstream?

而且，应用层的公司能跟那些前沿实验室竞争，去争夺那个特定市场吗？

And and, you know, are companies at the app layer able to compete with the frontier labs going after that particular market?