Read_935 - 人工智能将吞噬应用软件

是的，人工智能很重要，但认为人工智能将摧毁垂直和功能型软件商业模式的结论完全说不通。

Yes, AI is a big deal, but the conclusion that AI is going to kill the vertical and functional software business model simply makes no sense.

事实上，人工智能根本不会摧毁软件公司。

The truth is that AI simply isn't going to kill software companies.

等这场恐慌过去之后，我们会发现，人工智能是软件行业有史以来最好的事情。

After all this panic has passed, we'll see that AI is the best thing that ever happened to the software industry.

最懂比特币的人打造了Audible。

The best in Bitcoin made Audible.

我是盖·斯旺，这是Bitcoin Audible。

I am Guy Swann, and this is Bitcoin Audible.

大家最近怎么样？

What is up, guys?

欢迎回到Bitcoin Audible。

Welcome back to Bitcoin Audible.

我是盖·斯旺，这个世界上没人比我读过更多关于比特币的内容。

I'm Guy Swann, the guy who has read more about Bitcoin than anybody else you know.

今天我们有一个有趣的AI话题，我觉得这个话题触及了当前普遍存在的悲观情绪——认为软件行业即将消亡，人类即将被取代。

And we've got a fun AI episode today, that I think hits on there's a lot of doom and gloom and this idea that software is all about to die, and humans are about to be replaced.

我知道我之前在好几个不同集里零零散散地聊过这个话题。

And I know I've talked about it off and on in a bunch of different episodes here.

但这篇文章实际上来自a16z.news，不过它是他们的Substack专栏。

But this this article actually from the a16za16z.news, but it's it's their substack.

这看起来像是Substack。

This looks like substack.

这篇文章由亚历克斯·伊默曼和圣地亚哥·罗德里格斯撰写，非常精彩，虽然没有直接讨论经济学，却点出了几个至关重要的经济原理，我认为这些正是大多数人完全忽视的关键点。

But there are it's an article by Alex Immerman and Santiago Rodriguez, and it's a really good one that, without actually talking about economics, hits on a handful of really critical economic principles that I think are some of the things that are just totally totally missed on people.

为什么有那么多人声称我们即将迎来超级智能，人类将全部被取代，所有人都将失业，诸如此类的说法。

And why there's so many people who claim, like, we're about to hit super intelligence and all humans will be replaced, and nobody's gonna have a job anymore and all of this stuff.

我一直认为，而且多次在节目中说过，人们之所以相信这些荒谬的观点，正是因为无法理解基本的经济原理。

I've always thought, and I've said that multiple times on this show is that the failure to understand basic economic principles are why people believe these ridiculous things.

这些说法根本就是错的。

They're simply not true.

这些情况不会发生，背后有根本性的原因可以解释为什么。

They're not going to happen, and there's fundamental reasons to understand why.

我认为这篇文章实际上触及了其中一些要点，但又没有完全点明。

And I think this article actually hits on a few of them without actually hitting on a few of them.

我觉得它为我提供了一个很好的基础，他们所探讨的是对价值主张、护城河仍将存在之处、以及哪些主要护城河将崩溃或发生巨大变化的理性而冷静的分析，这些变化正深刻影响着我们当前对市场经济、软件经济和互联网经济的理解。

And it I think it gives me a good foundation, which I I think what they get into is a very sensible and sober look at what the value propositions are and where the moats will still be and what the major moats are that will be collapsing or changing dramatically in the way that we currently see the market economy and how the software economy and the Internet economy work.

所有这些即将被人工智能削弱或彻底改变的行业。

All of these various industries that are going to be undermined or changed dramatically by AI.

但这其实是正常的创造性破坏，他们可以指出一些具体事例，并提出一些框架来清晰地说明这一点。

But that this is normal creative destruction, and there's a number of things that they can point to and framings that they can suggest to make it clear.

然后，我认为，把这些观点与经济原理联系起来，尤其是价值与判断在经济中的根本性质，能够更完整地描绘出一幅清晰的图景。

And then I think taking those things and explaining how they connect to the economic principles, the very fundamental nature of value and judgment in an economy, I think, can complete paint a much more complete picture there.

所以我认为，这是一篇非常适合在节目中阅读并分享我看法的文章。

So I thought this would be a really good one to read on the show and give my thoughts on.

顺便向人权基金会致意，别忘了他们现在正在销售奥斯陆自由论坛的门票，活动在六月举行。

A quick shout out to the Human Rights Foundation, and don't forget that they have tickets on sale now to the Oslo Freedom Forum, June.

该链接将在节目笔记中提供。

That link will be done in the show notes.

另外，我还有很多其他不错的链接，比如如果你想在river.com上购买比特币，我有一个推荐链接，能给我带来大约5美元的收益。

Also, I have a lot of other great links, like if you wanna buy bit coin on river.com, I have an affiliate link, and it sends me like $5 or something like that.

我不确定。

I don't know.

但这是一种完全免费的支持节目的好方式。

But it's a really great way to help out the show that's totally free.

我非常喜欢River。

I love River.

它们是我主要使用的平台之一。

They're one of my main ones to use.

我还有一个关于Fold和桌游Hold'l Up的推荐链接。

I also have an affiliate for Fold and for the the board game, Hold'l Up.

我也有Coin Kite、BitBox和Blockstream Jade的推荐链接。

I have one for Coin Kite and for BitBox and Blockstream Jade.

我整理了这些产品和服务供你们使用，因为它们都是我真心喜爱的。

I have a bunch of these that I am making available for you guys because these are products and services that I really, really like.

而且人们总是问我：你用什么？

And people are always asking me, what do I use?

这件事你应该怎么做？

What should I do for this?

所以我把它们集中起来，同时也提供联盟链接，这样你们就能支持我的工作，帮助维持这个节目运行，并资助我正在做的 Pear Drive 和其他开发项目。

And so I tried to collect them together and also just give affiliate options so that you can support my work and help keep this show running and help fund Pear Drive and the other development projects that I'm working on.

任何形式的支持我都非常非常欢迎。

Any and all support is highly, highly welcome.

感谢所有为此付出的人。

And I thank you all for for those who do.

特别感谢 Audio Naughts，他们已经逐渐形成了一种‘氛围编程’的对话和社群，因为大家都在努力构建东西，我觉得这太棒了。

And a shout out to the audio naughts who have basically become a vibe coding conversation, vibe coding group because everybody's trying to build stuff and I just find that awesome.

向这些朋友以及所有支持节目的人致意。

So shout out to those guys and everybody who supports the show.

好了，让我们进入今天的阅读内容。

With that, let's get into today's read.

文章标题是《好消息，AI将吞噬应用软件》，作者是亚历克斯·伊默曼和圣地亚哥·罗德里格斯。软件行业正在经历一场恐慌。

And it's titled, good news, AI will eat application software by Alex Immerman and Santiago Rodriguez The software industry is having a panic attack.

自2026年初以来，公开软件公司的ETF下跌了30%，抹去了自ChatGPT发布以来的所有涨幅。

Since the start of 2026, ETFs for public software companies have fallen by 30%, erasing all the gains since the launch of ChatGPT.

像Salesforce、Adobe、Intuit、ServiceNow和Veeva这样的行业标杆企业，在短短几周内股价下跌了25%至30%。

Companies like Salesforce, Adobe, Intuit, ServiceNow, and Veeva bellwethers that have compounded investor capital for a decade or more are down 25% to 30% in a matter of weeks.

热门的Substack文章描绘了一个企业软件客户群被掏空、标普指数陷入多年大幅下跌的世界。

Viral substack posts imagine a world where the customer base for enterprise software is hollowed out and the S and P enters a massive years long drawdown.

他们称之为SaaSpocalypse（SaaS末日）。

They're calling it the SaaSpocalypse.

这正迅速成为市场的共识。

It's rapidly become the market consensus.

AI将摧毁软件行业。

AI is going to kill the software industry.

是的，AI确实很重要，但认为AI将摧毁垂直和功能型软件商业模式的结论完全说不通。

Yes, AI is a big deal, but the conclusion that AI is going to kill the vertical and functional software business model simply makes no sense.

事实是，AI根本不会摧毁软件公司。

The truth is that AI simply isn't going to kill software companies.

当这场恐慌过去之后，我们会发现，AI其实是软件行业有史以来最好的机遇。

After all this panic has passed, we'll see that AI is the best thing that ever happened to the software industry.

为什么会这样呢？

Why is that?

这个基本论点建立在对软件公司实际销售内容的误解之上。

The bare case rests on a basic misunderstanding of what software companies actually sell.

市场把软件当作一种大宗商品投入品来对待，仿佛软件公司的价值在于其代码，而更便宜的代码意味着更激烈的竞争，因此公司价值更低。

The market is treating software as though it were a commodity input, as if the value of a software company resided in its code, and cheaper code meant more competition and therefore cheaper companies.

但代码从来都不是价值的来源。

But code is never where the value has lived.

如果价值真的在代码上，这些公司根本不可能变得如此庞大。

If code is where the value was, these companies would have never gotten so big in the first place.

它们早在多年前就会被开源软件或发展中国家廉价的软件工程劳动力击败。

They would have been killed years ago by open source software or by competition from cheap software engineering labor in developing countries.

如今看空的观点通常可以归为四类。

The bearish arguments today usually fall into one of four categories.

也许基础模型公司会向上游延伸，掌控每一个功能特定的应用程序；或者企业会自行开发代码来替代其内部工具，至少利用这种可能性来削弱软件公司的定价能力；或者现有玩家会利用人工智能大幅扩展产品范围，彼此竞争。

Maybe the Foundation Model Labs will move up the stack and own every function specific application, or maybe enterprises will vibe code replacements for their internal tooling, or at least use the option of doing that to reduce software businesses' pricing power, Or maybe existing players will use AI to massively expand their product breadth rubbing up against each other.

或者，大量新进入者——著名的单人十亿美元公司——会以低价压垮现有企业。

Or maybe a flood of new entrants, the famous single person billion dollar company, will undercut incumbents on price.

再加上，智能代理不会在意品牌忠诚度或熟悉的名字，只会选择任何特定任务中最便宜的选项。

Pile on top of this, agents won't care about brand loyalty or familiar names, only the cheapest options for any particular task.

人工智能可能会增加竞争，但也会极大地扩展软件公司能做什么、能多快完成，以及它们服务的市场能变得多大。

AI might increase competition, but it'll also dramatically expand what software companies can do, how fast they can do it, and how large the markets they serve can become.

最终结果不会是利润率压缩至零，软件行业将变得更大，那些赢得竞争优势的公司仍将拥有持久的护城河。

The end result won't be margin compression to zero software will be a much bigger industry with durable competitive advantages for the companies that earn them.

真正重要的护城河并不会消失。

The moats that matter aren't going away.

关于商业护城河的经典当代著作是汉密尔顿·海尔默的《七种力量》。

The classic contemporary book on business moats is Hamilton Helmer's Seven Powers.

他列出了企业建立强大竞争优势的七种不同方式：规模效应、网络效应、反定位、转换成本、品牌、垄断资源和流程优势。

He lists seven distinct ways in which companies develop robust competitive advantages: scale, network effects, counter positioning, switching costs, brand, cornered resources, and process power.

让我们逐一分析它们。

Let's go through them.

转换成本或许是唯一会发生变化的护城河。

Switching costs are perhaps the one moat that really is going to change.

确实，人工智能正在改变更换供应商时的摩擦和成本效益分析。

It's definitely true that AI is changing the friction and the cost benefit analysis associated with switching vendors.

智能代理可以协助完成许多过去令人头疼的迁移工作。

Agents can assist with a lot of migration work that used to be a headache.

因此，那些拥有‘人质’而非‘客户’的legacy公司——借用我们同事亚历克斯·兰普尔的说法——将感受到比以往更大的压力。

So it means legacy companies with hostages, not customers, to borrow a phrase from our colleague Alex Rample, will feel a lot more pressure than they're used to.

但这对整个软件行业来说是一件好事。

But that is a good thing for software as a whole.

当公司必须通过赢得客户的忠诚度，而不是仅仅依赖供应商锁定时，结果将是更好的产品、更快的创新，以及一个增长更快、为客户创造更多价值的健康竞争生态系统。

When companies have to earn their customers' loyalty instead of relying just on vendor lock in, the result is better products, faster innovation, and a healthier competitive ecosystem that grows faster and delivers more value to its customers.

我们预计人工智能将促使一些客户转向新的赢家，但不会削弱大型企业的行业利润池，企业只会变得更好。

We expect AI will shift some customers to new winners, but it won't impair industry profit pools at large companies will just get better.

网络效应是一种经典的护城河，而且不会消失。

Network effects are a classic moat, and they aren't going away.

我们通常将网络效应应用于社交媒体平台或市场平台——网络中的节点越多，加入其中的吸引力就越大。

We tend to invoke network effects for social media platforms or marketplaces the more nodes in the network, the more attractive it is to be on it.

但同样的原理也适用于展现出生态系统、协作和数据网络效应的应用软件产品。

But the same applies to application software offerings that exhibit ecosystem, collaboration, and data network effects.

表面上看，Salesforce 是一个客户关系管理数据库，但任何在企业环境中工作过的人家都知道，Salesforce 也是一个生态系统。

On the surface, Salesforce is a CRM database, but anyone who has worked in an enterprise setting knows that Salesforce is also an ecosystem.

当所有人都使用同一个平台时，这个网络就会自我强化。

When everybody uses one platform, the network becomes self reinforcing.

你使用 Salesforce，是因为所有人都在使用 Salesforce。

You use Salesforce because everyone uses Salesforce.

使用 Salesforce 的公司越多，基于 Salesforce 构建的第三方应用程序生态系统以及 Salesforce 平台管理员和专家就越有价值。

And the more companies use Salesforce, the more valuable the ecosystem of third party applications built on top of Salesforce and platform administratorsexperts in Salesforce.

近年来，Figma 也遵循了类似的模式。

In recent years a similar thread is true for Figma.

每个设计师，接着是工程师、产品经理、市场人员，都购买 Figma，因为所有人都在那里协作。

Every designer, then every engineer, product manager, marketer buys Figma because everyone is collaborating there.

去参加年度 Config 大会，亲身体验这个生态系统的价值。

Go to the annual Config conference and witness the value of the ecosystem first hand.

同样的动态正在 AI 原生一代中出现。

And the same dynamic is emerging in the AI native generation.

Harvey 和 Hebia 正在构建金融和法律协作平台，将服务提供商、客户，以及未来的智能代理，整合到同一个系统中。

Harvey and Hebia are building finance and legal collaboration spaces that connect service providers and clients, and soon their agents, on a single system.

使用这些平台的人和智能代理越多，平台的价值就越大。

The more people and agents who use these platforms, the more valuable the platforms become.

Elise AI 的维护产品是一个多边网络，每增加一个用户或供应商，其价值就提升一分。

Elise AI's maintenance product is a multi sided network that becomes more valuable with every unit and vendor added.

随着迁移变得更容易，聚合也变得更容易。

As migration gets easier, aggregation gets easier.

但在软件免费的世界里，这些网络效应并不会消失。

But these network effects simply don't go away in a world where software is free.

事实上，由于人工智能使网络更强大，你现在能用网络做的事情比以前多得多。

In fact, insofar as AI makes the network more powerful, you can just do much more with a network than you could before.

我们应该预期，人工智能会使这些网络效应比以往更强大。

We should expect to see AI make these network effects more powerful than they were before.

规模从来不是软件中的核心护城河。

Scale was never the defining moat in software.

对于Salesforce来说，规模的重要性不如云服务商或工业企业那么高，但在某些情况下，对于AI应用而言，当计算支出超过人力成本时，规模可能会带来优势，因为大规模使用令牌的用户能获得单位成本优势。

It's just not as important for Salesforce as it is for a cloud provider or an industrial company, but to some extent it may matter for AI applications where compute spend exceeds labor costs, driving a unit cost advantage to the larger consumers of tokens.

此外，在某些领域，规模仍能通过规模经济发挥作用，将维护负担集中到一处，因为专业化带来的生产率提升在人工智能时代依然有效。

In addition, there are places where scale will still help as a straightforward economy of scale to concentrate that maintenance burden in one place since productivity gains from specialization don't go away in an AI world.

Stripe强调了集中式基础设施对其所有客户的价值。

Stripe highlights the value of centralized infrastructure benefiting all of its clients.

其合规基础设施承担了跨越数十个国家的监管合规成本，使个别企业无需自行应对。

Its compliance infrastructure absorbs the cost of navigating regulations across dozens of countries so that individual businesses don't have to.

其支付优化算法通过路由和重试交易来最大化授权率，随着交易量的增加而不断改进，并能将这些节省的成本传递给客户。

Its payment optimization algorithms, which route and retry transactions to maximize authorization rates, improve with every dollar of volume, and they can pass those savings on to their customers.

最后，规模将继续惠及那些处于数字与实体交汇点的公司。

Finally, scale will continue to benefit companies at the intersection of bits and atoms.

Anduril、Flock Safety 和 Waymo 在提高其硬件产品产量的同时，将继续降低单位成本。

Anduril, Flock Safety, and Waymo will continue to see lower unit costs as they produce higher volume of their hardware offerings.

品牌经久不衰。

Brand endures.

无论好坏，‘买IBM不会被开除’依然是大多数企业中的现实。

For better or worse, no one got fired for buying IBM remains a fact of life in most enterprises.

如果每个行业都变得更加拥挤，如果突然涌现出大量兜售‘氛围编码’ERP的临时创业者，我们应预期强大品牌的力量会增强。

And if every industry gets more crowded, if there's suddenly an explosion of fly by night solopreneurs selling vibe coded ERPs, we should expect the power of strong brands to increase.

品牌是在无限选择的世界中传递可靠性的信号。

Brand is how you signal reliability in a world of infinite optionality.

任何新兴公司都不可能立即复制像Stripe、Shopify或ServiceTitan那样建立起来的信任与知名度。

No upstart is going to instantly replicate the trust and recognition that companies like Stripe or Shopify or ServiceTitan have built.

你越接近核心业务功能，人们在支付处理方面就越不愿意尝试创新，品牌效应也就越强大。

The closer you sit to business critical functions, people really don't want to get creative when it comes to payment processing the more powerful brand effects will be.

如果你是一家初创公司并向客户收费，你默认会选择基于Stripe构建。

If you are a start up and you charge customers, you build on Stripe by default.

我们承认，随着更多决策被交由AI代理处理，而这些代理只关注价格优化，忽略人类所重视的软性因素，品牌的力量可能会发生变化——这是代理驱动的增长。

We do acknowledge the power of brand might change as more decisions are delegated to AI agents that optimize for price without the soft considerations that humans have agent led growth.

但只要这些代理仍需向担心被解雇的人类汇报，'买IBM不会被开除'这一原则依然成立。

But as long as they report to humans who have to worry about getting fired, the no one got fired for buying IBM principle still holds.

像高质量的专有数据这样的稀缺资源，也不会停止其重要性。

Cornered resources, like high quality proprietary data, aren't going to stop mattering either.

如果摩擦成本降为零，仅仅将公开数据整合成可用界面的价值就会降低，因为任何人都能做到这一点。

If friction goes to zero, simply consolidating publicly available data into a usable interface becomes less valuable because anyone can do it.

但如果AI能让人们用高质量数据做以前无法做到的更多事情，那么那些难以获取的资源就会变得极其宝贵。

But if AI enables doing much more with high quality data than you could before, then the stuff that you can't get easily becomes extremely valuable.

我们观察到彭博社的实时市场数据、Open Evidence庞大的医学文献库以及Vilex的法律数据库所具有的强大优势。

We have observed the power of Bloomberg's live market data abridges millions of clinical conversations, Open Evidence's vast medical library and Vilex's legal database.

而在这个新时代，或许最强的护城河是流程能力，正如Hebia的George Sivulka所称的‘流程工程’。

And perhaps the strongest moat of all in this new era is process power, or as George Sivulka of Hebia calls it, process engineering.

应用软件可以被看作是一种被存储的流程。

Application software can be thought of as a stored process.

它编码了关于组织职能应该如何运作的观点，这些观点经过数年乃至数十年的使用逐渐固化，成为与组织本身密不可分的一部分。

It encodes opinions about how the function of an organization should operate and those opinions calcify over years and decades of use into something that is inseparable from the organization itself.

成功的应用软件公司是那些能够与客户围绕这些工作流程共同演进的公司。

Successful app software companies are the ones that co evolve with their clients around these workflows.

随着这些工作流程不断深入组织内部，流程工程对挑战者而言变得越来越重要，也愈发难以复制。

As those workflows penetrate ever deeper into an organization, process engineering only becomes more important and more difficult for challengers to replicate.

以Harvey为例。

Consider Harvey.

如果Harvey深刻理解某家律师事务所如何组织其工作，那么即使编码成本为零，新进入者也无法在一夜之间复制这种能力。

If Harvey deeply understands how a particular law firm structures its work There is simply no way a new entrant can replicate that overnight, even with the cost of coding being zero.

这种嵌入式工作流程知识随着软件从记录系统转变为行动系统而变得更强大，而非更弱，因为你能够利用这些知识做更多事情。因此，随着底层模型的改进，Harvey的编排层——即引导模型输出通过特定专业工作流程的框架——其价值会呈复合增长。

That kind of embedded workflow knowledge becomes more powerful, not less, as software moves from a system of record to a system of action because you can just do much more with that knowledge, so as the underlying models improve, Harvey's orchestration layer, the scaffolding that routes model output through specific professional workflows, compound in value.

更好的模型并不会让应用层变得更薄，反而让它更强大，因为真正的难点从来不是原始智能，而是知道如何运用它。

Better models don't make the application layer thinner, they make it more capable because the hard part was never raw intelligence it was knowing what to do with it.

平台转型会催生新的赢家和新的护城河。

Platform shifts create new winners and new moats.

但作为投资者，我们发现还有一个最终的、持久的竞争优势来源尤其令人兴奋，那就是反向定位。

But there's one final source of durable competitive advantage that we find particularly exciting as investors, and that is counter positioning.

反向定位是一种新进入者可以召唤并运用的力量。

Counter positioning is a kind of power that can be summoned and wielded by new entrants to a market.

当一家新公司拥有某种商业模式，而该模式由于某种原因对现有企业缺乏吸引力时，就形成了反向定位。

It's when the new company has a business model which, for whatever reason, is unattractive for the incumbent company to compete against.

克莱·克里斯滕森的颠覆理论是一种经典的反向定位，但这种差异化定位并不总是以低成本为特征。

Disruption theory from Clay Christensen is a classic type of counter positioning, but it doesn't always have to be low cost as the differentiated counter position.

在软件领域，一种新的技术栈可能为初创公司开辟空间，使其创造出 incumbents 难以复制的新产品和商业模式，例如 Databricks 及其 Lakehouse 模型。

In software, a new technology stack could create the opening for a start up to create new kinds of products and business models that are difficult for incumbents to replicate, like Databricks and their Lakehouse model.

代理模型在完成工作和取代任务方面，无疑为新创公司提供了对抗现有企业的反向定位机会。

The agent model of doing work and replacing tasks is certainly going to create some counterposition opportunities for new startups to challenge incumbents.

关于代理型新创公司用基于价值的定价取代按席位收费模式，已经有很多讨论了。

There's been a lot of ink spilled on the disruption of per seat pricing at the hands of agentic upstarts with value based pricing.

让我们以客户服务为例。

Let's take customer service as an example.

Decagon 的客户服务产品按处理的对话次数收费，而不是按代理席位收费，并最终将按解决的问题数量收费。

Decagon prices its customer support product per conversation handled, not per agent seat, and will eventually price per resolution achieved.

这从根本上实现了供应商与客户之间激励机制的更好对齐。

That's fundamentally a better alignment of incentives between vendor and buyer.

像 Zendesk 这样的现有企业很难做出同样的转变，因为这会蚕食其现有的按席位收费收入。

An incumbent like Zendesk can't easily make that same move without cannibalizing its own seat based revenue.

就像百视达无法在不摧毁自身商业模式的情况下匹配奈飞的订阅模式，或者 PeopleSoft 无法在不颠覆其盈利模式的情况下匹配 Workday 的 SaaS 模式一样。

Just as Blockbuster couldn't match Netflix's subscription model without destroying its existing economics, or PeopleSoft couldn't match Workday's SaaS model without upending its monetization.

从新商业模式起步的公司不会面临这种困境，这也是平台转型如此可靠地催生新赢家的核心原因。

Companies that start with the new business model don't face that dilemma, and it's the core reason why platform shifts so reliably produce new winners.

但你猜怎么着？

But guess what?

市场上最终定价权的总量并不一定减少了。

The total amount of end state pricing power in the market didn't necessarily decrease.

这意味着客户现在可以选择他们想订阅的商业模式，更好的那个会胜出。

It just means customers now have a choice of business models they'd like to subscribe to and the better one will win.

这就是竞争性市场一直以来的运作方式。

That's how competitive markets have always worked.

AI并不是第一次出现创造性破坏浪潮，重新排列市场并改变游戏规则。

AI is not the first time that a wave of creative destruction has rearranged markets and shifted the playing field.

但关键是，由此产生的商业模式几乎总是在总机会规模上远远超越旧模式。

But here's the thing, the business models that result almost always dwarf the old ones in the scale of the total opportunity.

伟大的软件分化即将到来。

The great software bifurcation is coming.

所以是的，AI肯定会改变垂直和功能型软件，但它不会看起来像一场大屠杀。

So yes, AI will definitely change vertical and functional software, but it won't look like a massacre.

也许毛利率会进入一个新的稳定状态，也许由于切换成本使采购团队在供应商谈判中拥有更多话语权，导致定价能力下降，但AI也通过更高效地利用劳动力支持利润率的扩张。

Maybe gross margins settle into a different steady state, maybe pricing power is diminished because switching costs give procurement teams more leverage in vendor negotiations, but AI also supports margin expansion due to a much more efficient use of labor.

但无论利润率最终如何，我们都预计规模将大幅扩张，因为我们的同事甘什·阿克哈里亚常说，世界仍然缺乏软件。

But no matter where margins end up, we expect that scale will expand dramatically, because as our colleague Ganesh Akharya likes to say, the world is still short software.

我们远未达到全球对高质量软件需求的饱和点，随着代码成本降低，我们理应看到市场需求进一步增长。

We are nowhere near saturating the world's demand for high quality software, and as code becomes cheaper, we should just expect to see the market demand more.

在这场AI转型之后，我们将面对一个规模更大、为客户创造更多价值的软件产业。

On the other side of this AI transition, we'll be looking at a much bigger software industry that provides much more value to its customers.

企业将能够服务更多客户，进入相邻市场，并自动化那些过去过于复杂或成本过高而被搁置的工作流程。

Companies will be able to serve more customers, enter adjacent markets, and automate workflows that were previously far too complex or too expensive to touch.

过去ACV过低的客户，突然间有了具有吸引力的经济模型，那些曾经被归为‘太难处理’的构想，如今变得有趣且可行。

Customers that were previously too low ACV will suddenly have attractive economics ideas that would once have gone into the too hard pile suddenly become interesting and feasible.

护城河依然存在，只要还有护城河，就完全有理由相信，成功且持久的企业将继续生存并繁荣发展。

There will still be moats, and as long as there are moats, there's plenty of reason to expect hugely successful and highly durable businesses to survive and thrive.

AI不会摧毁软件产业，它将把该产业一分为二。

AI isn't going to destroy the software industry it's going to split it into two parts.

确实会有一些软件公司类别面临真正的压力。

There really will be some categories of software companies that face genuine pressure.

那些主要作为商品功能薄封装、仅以稍更便捷的格式展示数据的前端工具，是脆弱的。

Front end tools that serve primarily as thin wrappers around commodity functionality and do relatively little beyond presenting data in a slightly more convenient format are vulnerable.

那些仍使用过时界面、却每年提价的现有记录系统应当感到担忧。

Incumbent systems of record that still operate on archaic interfaces but raise prices every year should be worried.

那些定价模式陈旧、价值主张远逊于AI原生竞争对手的软件公司也应警惕。

So should software companies that have an outdated pricing model and value proposition that's just inferior to what AI native competitors can offer.

在这个环境中获胜的公司将是那些提供真正价值的公司，而不是那些为客户群筑起最高围墙的公司。

The companies that win in this environment will be the ones delivering genuine value, not the ones that built the highest walls around their customer base.

但这只是创造性破坏。

But that's just creative destruction.

这些公司面临以前从未有过的压力，对整个行业来说是好事。

It's great for the industry that these companies are facing pressure that they weren't facing before.

其中一些公司会找到解决办法并变得更强大。

Some of them will figure things out and get stronger.

其他人则不会，将会倒闭。

Others won't and will die.

这很好。

That's good.

其余的软件生态系统，那些致力于为客户创造真实价值的公司，将迎来巨大增长。

The rest of the software ecosystem, the companies that are committed to delivering real value for their customers, is set to grow massively.

所以，是的，一些公司会失败，但整个行业将获胜，而且会大获全胜。

So yes, some individual companies will lose, but the industry will win and win big.

SaaSpocalypse 并不是软件的终结。

The SaaSpocalypse isn't the death of software.

而是更大事物的开端。

It's the start of something much bigger.

好吧。

Alright.

那么我们就说到这里。

So that wraps us up.

向这篇文章的作者致敬。

Shout out to the authors of this one.

我有这篇文章的链接。

I have the link to the article.

这应该是来自A16Z的Substack。

This is from I think it's the A16Z Substack.

但这篇文章中有大量非常出色、坦率地说，极其简单而深刻的经济学见解。

But there's just a ton of really really good, honestly, very simple and very good economics in this article.

也许无意中，他们在讨论的广泛概念以及对市场演变的论述，与奥地利学派的理论高度契合，尽管他们可能并没有刻意如此。

Maybe unintentionally, just kind of in the the broad concepts that they discuss and how they talk about the evolution of markets and stuff that are deeply aligned with Austrian theory, probably without intending to be.

但最能体现这一点的例子，或许是杰文斯悖论：当某样东西变得更易获得、生产成本更低时，对其的需求必然会增加。

But probably the best example or the best thing to connect this to is Jevon's paradox, is that when you make something more available and cheaper to produce, the demand for it inevitably increases.

所以很多人会认为，这再也不能作为一项盈利业务了，或者会因为太容易被复制而被彻底取代。

So so many people will think they're like, oh, well, it won't be viable as a business anymore, or it will just be totally replaced because it's too easy to create.

但实际上，这篇文章中有一句话我特别喜欢，它很好地阐释了这一理念：客户将能够自行服务，抱歉。

When in fact, and there's actually a line in this that I really liked, that kind of demonstrated the the concept behind this, is that customers will be able to serve excuse me.

公司能够服务更多客户，进入相邻市场，并自动化那些过去过于复杂或成本过高而无法触及的工作流程。

Companies will be able to serve more customers, enter adjacent markets, and automate workflows that were previously far too complex or too expensive to touch.

过去ACV过低的客户，突然间会变得具有吸引力的经济价值。

Customers that were previously too low ACV will suddenly have attractive economics.

所以他们在这里谈论的‘ACV过低’，指的是年度客户价值或年度合同价值，类似这样的概念。

So what they're talking about here is, like, too low ACV is is annual customer value or annual contract value, something like that.

他们基本上在讨论的是，每个新客户能带来多少收入，相对于你为该客户提供服务所付出的成本。

And what they're basically talking about is, like, how much will each new customer bring in versus the cost of sustaining whatever it is that you're providing that customer.

所以，如果你在业务中新建了一个领域或功能，只带来10个客户，而你只能向这些客户每月收取5美元，那么你必须确保这个功能的年成本低于600美元，否则就是在浪费钱。

So if you build a new area of your business or a new feature that only brings in 10 customers, well then and and you're only expected to or it only makes sense to be able to charge those customers $5 a month, well then you'd be you better be able to provide that feature for less than $600 a year or you're wasting money on it.

这正是为什么某些功能、软件、定制化选项或偏好在市场中得不到回应的原因——因为它们是小众问题。

This is exactly why certain features or software or customizations or preferences don't get answered in a market because they it it's the it's the niche problem.

对吧？

Right?

这同样解释了为什么广播媒体中，你不会在有线新闻或有线电视上看到专门讲解比特币钱包的教程、热门产品的开箱视频，甚至像《黑镜》这样高风险的内容。

It's the same reason why broadcast media, while you wouldn't see on cable news or cable TV, you wouldn't see a show that was just tutorials about Bitcoin wallets or unboxing videos for popular products or even something kind of crazy and high risk like Black Mirror.

你只会看到它出现在某些非常特定的平台或制作公司，它们确切知道这些内容如何为受众带来益处。

You would only ever see it on very specific very specific platforms or production companies that knew exactly why or how it would benefit their audience.

但现在YouTube和流媒体出现了，改变了内容的提供方式，以及吸引特定利基受众或服务网络中一小部分群体或某种文化小众的ACV（客户价值）与成本收益比。

But now YouTube and streaming come along, and it changes the nature of how that content is provided and what the ACV, what the the cost per benefit ratio of catching or attracting certain niche audiences or being able to serve some small subset of a network or some small group or some cultural niche.

突然间，所有这些内容都变得触手可及。

And suddenly, all of these things are available.

突然间，Netflix敢于冒险制作《黑镜》这样的节目，因为只要能留住那五万或十万无法在别处找到这类内容的用户，提供多样化的节目类型、评级、时长和严肃程度与低俗内容，Netflix实际上从中受益了——因为他们只卖一份订阅，而大多数用户只需要两到三部喜欢的剧集来 binge 观看或探索，就足以证明继续订阅的价值。

Suddenly, Netflix can take the risk on something like Black Mirror because keeping the 50,000 or a 100,000 users that can't find that content somewhere else on that platform, having a variety of content, types and ratings and length and seriousness versus stupidity, you know, all of the various types of content that might serve the huge swath of customers, Netflix has actually benefited because they sell one subscription and they simply need most people only need two to three good shows that they might wanna binge watch or explore to justify keeping that subscription.

此外，现在内容的分发成本也更低了。

And then in addition, the delivery of that content has a lower overhead now.

再想想YouTube，有多少人仅靠做产品评测就能生存或获益，因为他们可以自由选择、拓展领域。

And then think about YouTube, the number of people who could survive or benefit from just doing reviews, just doing reviews of products and being able to branch out because people can simply pick and choose.

我的意思是，整个付费点播的概念。

I mean, just the whole idea of pay per view.

对吧？

Right?

这在有线电视和广播媒体中曾是一个巨大的概念，那就是所谓的‘高级’之后，你可以按客户付费观看。

Like, that was a huge thing in cable and broadcast media that there was this concept after after, you know, quote, unquote, advanced, and you could get it per customer that you could pay per view.

这曾经是个大事。

That was like a big deal.

现在，基本上所有内容都变成了你可以随时选择自己想看的内容，不再有固定播出时间。

Now, basically, everything is some form of you just watch the content that you specifically select at the time that you specifically want to, and there's no schedule.

你不必等到下午五点才看你的教学视频。

You don't have to wait until 05:00 in the afternoon to watch your tutorial videos.

你不必纠结是否要等《黑镜》的下一集。

You don't have to decide whether like, wait for the next episode of Black Mirror.

如果你愿意，你完全可以一天之内一口气看完全部三季或五季内容。

You'd literally get to binge watch binge watch the entire three seasons or five seasons, whatever it is, in one day if you just feel so inclined to do that.

直到提供服务的经济模式、付费方式、平台分发以及网络连接发生变化，这种情况才成为可能。

This wasn't this wasn't even possible until the economics around providing the service, around paying for the service, around the platform delivery, and, and the connection or not connection.

随着互联网的发展，观众的规模和覆盖面发生了巨大变化。

The exposure to the size and breadth of the audience changed dramatically with the nature of the Internet.

如今的媒体内容比九十年代更少了吗？投入制作、提供和内容创作的资金也更少了？

And is there less media today or less money going into production or provision and content creation today than in the nineteen nineties?

因为现在做起来更容易了。

Because it's easier to do.

我认为，现在出现了一个全新的经济活动子集，叫做内容创作者，我也在其中，如果没有这种平台转变和经济模式的变化，这个播客根本不可能存在。

I think the fact that there's now this entire subset of economic activity that is called content creator, myself included, this podcast doesn't exist without that platform shift, without the change in those economics.

在九十年代的电视和传统媒体上，你根本不可能持续播出一档名为《Bitcoin Audible》的节目，内容涉及经济学、哲学、工程学和互联网历史的技术变革。

You couldn't possibly sustain a show about, you know, economics and philosophy and engineering and technological shifts in Internet history called Bitcoin Audible on television, on broadcast media back in the nineteen nineties.

我的意思是，当然，那些东西当时根本不存在，但这个受众非常小众，甚至可以说是小众中的小众。

I mean, obviously, none of that stuff existed, but it's this is a very niche audience, and it's kind of a niche within a niche too.

现在比特币和加密货币已经是一个相当庞大的领域了，但我服务的并不是广大受众。

There's, like, Bitcoin and crypto is a pretty big environment now, and I do not serve the broader audience.

我谈论的很多内容对他们来说完全毫无意义，而他们大多数人只想交易垃圾币。

There's a lot of stuff that I talk about that wouldn't make any damn sense to them, and most of them actually wanna trade garbage.

对吧？

Right?

他们参与其中是因为想着：我怎么才能赚钱？

They wanna they're they're in it because they're like, how can I make money?

我什么时候该卖出这个代币，买入那个代币？

And which when do I sell this token and buy this token?

我的观众不是这样的。

That is not what my audience is.

我的节目也不是这样的。

That is not what my show is.

我讨厌这种说法，这完全是垃圾。

I hate that that's total garbage.

我还不如直接做个关于赌场的节目。

I might as well just have a show about casinos.

但新的技术和平台降低了进入门槛和服务提供成本，从根本上改变了参与这一市场的经济模式。

But new technologies and new platforms that lower the barrier to entry and lower the barrier of service provision fundamentally change the economics of the ability to actually participate in that market.

具体来说，结合杰文斯悖论的背景，它使得为更广泛、更多样化的小众群体、更小的网络和更次要的功能提供服务成为可能——这些在过去成本过高时毫无经济意义，如今却突然变得可行。

And it specifically, in the context of Jevon's paradox, it specifically opens it up to be able to apply or provide services for a vastly greater array or variety and, again, niche audiences, smaller networks and smaller features, less important ones that did not make economic sense when the costs were too high that now suddenly become available.

同样的事情也会发生在软件上。

And that same thing is gonna happen to software.

同样的事情其实已经在软件领域发生了。

That same thing is kind of already happening to software.

我认为这主要还停留在极客或独立创业者层面，人们只是在为自己构建东西。

I think it's just largely in, more isolated in kind of the tinkerer or the solopreneur sense is that people are building things for themselves.

这正是我所做的。

That's exactly what I've done.

事实上，我所接触或聆听过的每个人，比如斯特凡·洛韦拉和马蒂·本特，所有在播客和相关讨论中的人，所有在比特币播客圈子里的人，都创建了自己定制的工作流程和设置，并且都在使用人工智能来加速实现过程，以更低的成本和更短的周期从零发布一集内容。

And in fact, everybody that I've talked to or listened to when it comes to, like, Stefan Lovera and Marty Bent, everybody in the podcast and sessions, everybody in kind of the pod Bitcoin podcasting circles, is they've all created their own custom workflows and setups, and they're all using AI to make their implementations faster and and get from zero to episode published with lower cost and lower turnaround.

而这正是这篇文章中我特别喜欢、非常认同的另一个观点。

And this is where another thing that they specifically mentioned in this article that I really, really liked, and and I love this kind of distinction.

我认为人们经常忽略的一件最重要的事是，价值发生了转移。

One of the one of the biggest things that I think people constantly miss is that what is that value shifts.

价值转移到了新的挑战所在之处。

Value shifts to where the new challenge is.

当你让某件事变得可自动化，或者让某种智能能够完成某项任务时，这背后其实是对经济运作方式或其不同层级的极度简化。当你解决了底层的某个环节，上层的某些部分就会变得更加清晰。

When you make something automatable, or you make something, you know, you you get intelligence to to be able to do something, it's the idea there's this drastic oversimplification of how economies work or the different layers of it, and some of the layers become much more clear when you solve a layer underneath it.

当底层变得更容易理解时，上一层的关注点就会变得突出。

When it it becomes far simpler to understand how the next layer up becomes the focus.

这稍微有些简化，但我认为从概念上讲是个不错的例子。

This is slightly oversimplified, but I think it's a good example just conceptually.

成为一名优秀的画家，以创作出一幅优秀的作品。

Being a good painter for getting a good piece of art.

没错，我们常常把做一件事的过程、行为或使用特定工具的技能，与这件事本身的价值混为一谈。

Right, is we often conflate the process, the act or the skill of doing a thing a particular way or with a particular tool with the value of the thing itself.

对吧？

Right?

一幅伟大的艺术作品之所以伟大，是因为它极其困难，还是因为它有价值，是因为制作出如此精美的东西非常困难，你必须是个好画家，这真的很难，你要掌握各种画笔，还得弄清楚所有这些事情？

Is a great piece of art is great because it's very very difficult, or it's valuable because it's very very difficult to, a, make something this pretty, and you have to be a good painter, it's really hard, you get all these brushes, you gotta figure all this stuff out.

或者吉他和音乐。

Or guitar and music.

对吧？

Right?

我有价值，是因为我花了这么多时间和精力去学习如何弹吉他吗？

Is it, oh, I'm valuable because I have spent all of this time and investment into learning how to pluck the guitar.

我能够很好地弹奏，凭借记忆和情感记忆，还能即兴发挥。

And I can do it very well, and I can do it with memory and with my emotion memory, and I can do it offhand.

我非常擅长用吉他创作音乐。

I'm very skilled at creating creating music with the guitar.

所以我的价值在于掌握工具本身，但工具会变化。

So my value is in my ability to master the tool itself, but then the tool changes.

当工具发生变化时，会发生什么？

And what happens when the tool changes?

我们意识到，如果不懂音乐理论，或者没有创作好音乐的感知能力，会弹吉他其实并不重要，真正有价值的是音乐本身。

We recognize that being able to play the guitar doesn't really matter if there's no good music, or you don't understand music theory, or you don't understand you don't have the sense of creating good music because the real value is the music.

真正有价值的是绘画或艺术品本身的美。

The real value is the beauty in the painting or the the artwork itself.

从这个角度看，这段内容非常出色，它真正揭示了关键点。我在其他几篇文章和一些人的讨论中也发现过类似观点：在这项工作中，价值主张最核心的部分在于过程。

And in this sense, there's a really good there's a really strong section in this that I think really uncovers, and I've I've found this with a few other articles and other people talking about this as to where is the strongest part of the value proposition in doing this work, it's in the process.

让我重新读一下这一段。

So let me reread this section.

它说：或许在这个新时代中最强大的动力就是流程能力。

It says perhaps the strongest mote of all in this new era is process power.

或者，正如Hebiah的George Svocalca所称的，流程工程。

Or Or as George Svocalca of Hebiah calls it, process engineering.

应用软件可以被看作是一种被存储的流程。

Application software can be thought of as a stored process.

它编码了关于组织运作方式的种种观点，而这些观点经过数年乃至数十年的使用，逐渐固化为与组织本身密不可分的一部分。

It encodes opinions about how the function of an organization should operate, And those opinions calcify over years and decades of use into something that is inseparable from the organization itself.

成功的应用软件公司，是那些能够与客户围绕这些工作流程共同演进的公司。

Successful app software companies are the ones that co evolve with their clients around these workflows.

随着这些工作流程不断深入组织内部，流程工程对后来者而言变得愈发重要，也愈发难以复制。

As those workflows penetrate ever deeper into an organization, process engineering only becomes more important and more difficult for challengers to replicate.

那么，我们所说的这个到底是什么意思？

So what what do we mean by this?

我相信我读过这篇文章，我知道我读过，但我想我是在节目中读到的。

And I believe I read I know I read the article, but I think I read it on the show.

我得去翻一下订阅的节目列表。

I'll I'll have to dig into the the feed.

这不会是很多集之前的事，因为我们最近刚讨论过AI编程和 vibe 编程的内容，关于一个纯粹的过程究竟是如何被创造出来的，或者说是通过实际行动才真正浮现出来的。

It wouldn't have been that many episodes back because we've talked about AI coding and the vibe coding stuff, recently, but about how it is that a process only is actually created, or a process only actually emerges through the act of doing.

随着时间推移，当你面对现实，面对需求、判断或这个事物真正提供的价值时，这个工具本身才真正显现其意义。

It is over time, and as you run into reality, as you run into the need or the judgment or the value that thing has actually provided, that the tool itself is actually providing.

我不认为我确实读过这篇文章。

And I don't I don't think I did read it.

我觉得这是来自彼得·斯坦伯格的，我想那是他的姓，他是 OpenCLOA 的创始人。

I think it was from Peter Steinberg, the I think that's his last name, the guy who created OpenCLOA.

我觉得这像是我读过的一篇很长的推文之类的东西。

I think this was like a long Twitter post or something that I read.

我很难在记忆中定位它。

I'm having a hard time placing it in my memory.

但文章讲的是，当他同时管理多个代理运行，并且每月向 GitHub 提交多达六千次时，他拥有大约十五个代理在执行不同的任务，他实际上是在管理整个流程，引导这支‘交响乐团’该往哪里走，如何思考我们学到的东西，并适应这些新变化。

But it was about the fact that what he is managing when he's managing multiple agents running at the exact same time and he's publishing, you know, he's making 6,000 commits to GitHub every single month, and he's got, like, 15 agents or whatever doing different jobs, is he ends up managing the process and directing this orchestra of where to go and how to think about what we learned and adapt to those new things.

这意味着，目标、判断力和长期记忆成为了你能提供的价值，因为你不断调整和引导这些进程向前发展。

Which means essentially, goals, judgment, and long term memory become the value that you can provide because you tweak and guide things as they move forward.

你知道，记忆并不仅仅是试图找到正确的信息片段，或围绕某一件事收集所有上下文。

You know, memory isn't simply a task of trying to find the right pieces or having all of the context around one particular thing.

你知道，你不能只是读上十亿本书，就变得更聪明或对世界有更深刻的理解。

You know, it's like, you know, you can't just read a billion books and then you're smarter or you have a better sense of the world.

你还需要筛选哪些是好书，否则你很容易只是把一堆垃圾塞进脑子里，因为你读了十亿本错误的书。

You also have to filter what a good book you can eat just as easily just fill your head with garbage because you've read a billion of the wrong books.

这其实更重要，尤其是在一个内容无限、创作内容的成本几乎为零的世界里。

It matters just as actually, it matters far more, especially in a world where there's an infinite amount of content, where the the price to create content is essentially zero.

筛选并判断什么是有价值的、什么值得占用你的记忆，这才是关键。

It matters to filter and judge what is valuable and what is a good use of your memory.

这正是那种情况。

And that's exactly the sort of thing.

任何长期使用过智能代理的人都知道，记忆和流程是两大难题。

Anybody who's worked with an agent for long enough knows that memory is one of the biggest problems, and process is one of the biggest problems.

因为你经常会误入歧途，或者未能正确识别某个本应被掌握的教训其实并未被内化，又或者你的模型脱离了上下文，导致你破坏了一堆东西，然后不得不想办法拉回正轨、清理上下文，以便将最关键的因素重新置于优先位置？

Because the number of times that you can go down the wrong rabbit hole or not properly learn recognize that a a lesson that's supposed to have been learned has not actually been internalized or got out of context with your model, and so you broke a bunch of things and you have to figure out how to reel it back in and clear your context so that you can put those things at the forefront of the most critical things to consider?

这正是你需要考虑或调整你的结构、网关或你实际所处系统的方式——你可以通过你的AI代理编写代码，强制确保这些要素永远不会在上下文中丢失。

Well, this is exactly how you would what you would have to keep into account or modify your structure, like kind of your gateway or the system that you're actually working in, which you can code with your AI agent to basically insist that those things are never lost in context.

这意味着你必须做些与LLM智能代理本身所做不同的事情，因为问题不仅仅在于权重和上下文，更在于流程。

That means that you have to do something different than what the agent with the LLM itself is actually doing, because it's not just in the weights and the context, it's in the process.

而在于你如何实际应用这些要素，并围绕你所重视的或认为最重要的判断来构建定制化的工作流程。

It's in how you actually apply those things and then create a custom workflow around what you value or what you think are the most important judgments.

因为你可能拥有某种软件，它只关注你想要实现的功能，以及为客户提供有趣的小工具之类的东西。

Because you might have software that just values or just is concerned most most concerned with the features that you want to implement and having fun gadgets and stuff for your customers.

或者你可能拥有某种软件，其目标是建立信任。

Or you might have software where your goal is trust.

你的目标是可靠性。

Your goal is reliability.

你的目标是建立一个无法被破坏的基础。

Your goal is having a base that cannot be f'd with.

如果在最终状态或前端有一些功能没有直接或立即提供，或者需要做出某种权衡，但只要你能确定用户随时可以使用你的应用程序完成它必须完成的那一件事，这种权衡就是值得的。

And if a little bit of features on the end, at the end state or on the front end aren't directly or immediately provided or, you know, have some sort of a trade off, that's worth it if you can know that someone can always go to your application and do the one thing that you need that application application to do.

这些是不同的考量。

Those are different considerations.

这些是不同的判断。

Those are different judgments.

这些需要人类参与并做出决策。

Those are things that humans must be involved in to make calls on.

因为判断本质上是人类特有的能力。

Because judgment is specifically a mortal thing.

判断是价值的延伸，而价值只能来自真实世界中的生活与风险，也许有一天你走出去就再也回不来了。

Judgment is is a extension of value, and value only comes from living in the real world and having a risk, and maybe that maybe you walk out one day and die.

关于人工智能将取代所有人、将没有工作、这次情况不同的悲观论调太多了，让我惊讶的是，人们竟然还如此盲目，认为事情根本不是这样的。

And there's so much of this doom and gloom that AI is gonna replace everybody, and there aren't gonna be any jobs, and this time is different, and it's incredible to me how, god, how blind people still are to this, that that's just not how it works.

并不是说，哦，我们突然就达到了一个它开始这样运作的点。

And it's not that, oh, we just got to a point where suddenly it does work that way.

它从根本上就不是这样运作的。

It just fundamentally doesn't work that way.

我来告诉你一个简单的原因，帮助你理解为什么。

And I'll tell you a simple reason to understand why.

大语言模型就像一张地图。

LLMs are a map.

它们是一张地图。

They're a map.

它们是基于人类判断、人类评估以及人类与现实世界联系而得出的权重。

They are they are weights derived from human judgment, human assessment, and human connection to the real world.

然后，所有这些信息被汇总在一起，从这组数据中确定出各种模式。

And then all that information is aggregated together, and patterns are determined within that set of data.

但如果这些数据并非来自现实世界的体验，也没有真正面对宇宙中的矛盾，那么它们就不再有价值了。

But if that data is not generated from real world experience and basically clashing up against the contradictions of the universe, then they aren't valuable anymore.

它们无法真正创造出一致的体验，也无法在现实世界中持续存在。

They won't actually create a consistent experience and be of something that will sustain itself in the real world.

它必须具备某种互动性。

It has to have some sort of interaction.

它必须有某种方式来判断什么是好的输出，而且不能是循环的。

It has to have some sort of means to judge what a good output is, and it can't be circular.

你不能使用那个输出。

You can't use that output.

你不能用那个输出来做判断，然后把它重新放回输入中，以为这样就能得到更好的系统。

You can't judge that output, and then put it back into the input, and think that you're gonna get a better system.

我们已经看到，当大语言模型试图用自身的输出数据进行训练，或者图像模型试图用自身最佳输出进行训练时，系统就会变得越来越糟糕，这背后是有原因的。

There's a reason why we've seen this fundamentally when LLMs try to train themselves on LLM output data or when image models try to train themselves on the best output of the image model, that it gets retarded.

这不是我们能够通过工程手段解决的问题。

This isn't something that we're going to engineer around.

这是关于这些事物本质的普遍真理。

It's it's a universal truth about the nature of what these things are.

它们是衍生品。

They are derivatives.

它们不可能优于其来源。

They cannot then be better than what they are derived from.

它是一张智能的地图，而智能只有在与现实世界和对生命体（终有一死之物）的真实价值判断相对应时才有意义。

It is a map of intelligence, and intelligence only matters if it corresponds to the real world and real value judgments for something that is alive and can die one day.

你无法因为拥有更好的流程，就从之前的地图中绘制出更准确的地图。

You cannot draw a more accurate map because you've got a better process from the previous map.

你可以将这个流程应用到现实世界中，通过更好的系统或更好的绘制方法来画出更准确的地图，但地图永远不可能像现实本身那样复杂、多样，也无法更好地反映地面实际情况。

You can take that process into the real world, and you can draw a better map because you have a better system or a better way to draw it, but the map will never ever ever be as complex or varied or be able to better account for the reality of the situation on the ground than reality itself.

现实永远比任何对其的描绘都更加复杂和深邃。

Reality will always be infinitely more complex and infinitely deeper than any map of it.

事实上，地图一旦画出就是错误的，因为一切都在变化，而地图在确定模式的准确性或精确性方面，所能深入的程度是有限的。

In fact, the map is wrong as soon as it's drawn because everything changes, and there's only so there's only so deep of a degree that a map can go in actually determining the accuracy or the precision of a pattern.

试图构建的这种地图或模式权重系统越大、越通用，它在完成任务时的效率就越低。

And the bigger and more general the attempt of making that map or that pattern of that pattern weighting system is, the less efficient it is at doing the job.

这其实就是基本的经济学原理。

And this and this is basic economics.

如果我们试图创造一个能做所有事情的超人AI，而且是一个庞大的单一模型，它最终一定会失败。

If we try to create an a superhuman AI that can do everything, and it's one big giant model, it's just gonna fail.

它会因为基本的经济学原理而失败。

It's gonna fail to basic economics.

运行它的成本会太高，而为了在该系统本身的权重上获得边际改进，其成本远高于创建成千上万个小型专用模型，并由一个通用代理来调度上百个其他通用代理协助管理，而这一切都由一个人类主导，他构建一系列真正针对其目标和要完成任务的具体流程与结构。这些流程会变得越来越深入、越来越复杂、越来越具体，直至其复杂性呈指数级增长，使得更大的模型在试图实现或解答这些问题时反而更加困难。

It will cost too much to run it, and it will cost too much to get marginal improvements in the weighting of that mount of that system itself than it will be to create thousands of tiny specific models and have one general agent that can spin up a 100 other general agents to help manage while it's all conducted by a human who creates a series of processes and structures that are actually specific to the thing they are trying to do and the goal they are trying to achieve, and these process will get processes will get deeper, and more complex, and more specific to everything that we are trying to accomplish to the point that everything the complexity itself becomes exponential, such that the bigger models have even worse time trying to achieve it or answer it.

而所有这些实际上只在实际执行的过程中逐步展开。

And all of this actually only unfolds in the process of the doing of the thing.

你不能只是拍张照片，然后画个地图，就以为它能永远生效，事情并不是这样发展的。

You can't just take a picture of it, and then map it, and then suddenly, it works forever, and that's how everything goes.

这确实需要时间、回应和互动才能发展起来；随着它的演进，它会改变环境，进而迫使新的发展和变化，因为所有这些因素都在相互作用，没有任何东西是静止不变的。

It it literally requires time, and response, and interaction for it to be developed, and as it's developed, it changes the environment that then forces new development and new change, because all of these things are interacting, and nothing is static.

这就像试图通过知道空气中每一个分子的位置、方向和速度，然后计算这些分子在未来六周内相互碰撞后会发生什么来预测天气。

It's like trying to predict the weather by having a knowing where every single molecule is in the air and knowing every direction and velocity, and then trying to calculate what's going to happen in the next six weeks by watching all of these molecules bounce into each other.

这甚至不是一个智能问题，而是一个纯粹的计算问题。每当我们实现某种更自动化或更好地提供这些功能的技术时，一切都会变得更加复杂，因为你的输出现在成为了输入的一部分。

It's not even it's not an intelligence problem, it's a sheer compute problem, and every time we do something that can more automate or better provide any of those things, everything gets more complex because your output is now part of the input.

这正是我们在经济学中所看到的情况。

This is exactly what we see in economics.

一旦衡量标准本身变成了目标，它就不再是一个好的衡量标准。

As soon as the measure itself becomes the goal, it ceases to be a good measure.

一旦市场上出现某种模式，我们觉得‘这绝对是每个人赚钱的方式’，那么这种模式就会崩溃，所有人和所有事物都会失败，模式也就随之消亡。

As soon as we have a pattern in the market where we're like, oh, this is definitely the way that everybody's gonna make money, well, it breaks down and everybody and everything fails, and the pattern the the pattern dies.

为什么？

Why?

因为只有当你以极小、极简的方式操作时，才能真正隔离并利用这种模式。

Because you can actually only isolate and take advantage of that pattern if you're doing it in a tiny minimal sense.

而一旦你规模足够大，或者有足够多的资金试图利用这种模式，模式就会失效，因为你正在预测这个模式。

And as soon as you're large enough or there's enough volume or liquidity trying to take advantage of that pattern, the pattern stops working because you're predicting the pattern.