The technology world is now firmly in the AI era. Over the last few years, we have seen a ton of innovation at the foundational model layer — beyond OpenAI, we saw the emergence of Mistral, Anthropic, Meta’s Llama, Google Gemini and many more. This year, the focus appears to have shifted from foundational models to the application layer. a16z now regularly publishes a list of top 50–100 GenAI apps, and it’s hard to go a day without a new fundingannouncement for an AI application. Are we seeing a new wave of consumer winners, or merely early experiments that will pave the way for more successful companies down the line?
科技界现在正坚定地处于人工智能时代。在过去的几年里，我们在基础模型层看到了大量的创新——除了 OpenAI 之外，我们还看到了 Mistral、Anthropic、Meta 的 Llama、Google Gemini 等的出现。今年，重点似乎已经从基础模型转移到应用层。a16z 现在定期发布前 50-100 名 GenAI 应用程序的列表，很难有一天没有 AI 应用程序的新资金公告。我们是否看到了新一波的消费者赢家，或者仅仅是早期的实验，将为更成功的公司铺平道路？
AI = Software* (with some caveats)
AI = 软件*（有一些注意事项）
Before we dig deeper, it may be useful to summarize where we are — AI is a new evolution of software. Software by itself is a commodity and is not defensible. Successful software companies rely on a handful of mechanisms to generate defensibility — economies of scale, switching costs, and most importantly, network effects. I’ll touch on the first two in this post, but will focus on the third. If you’ve followed my writing before, this shouldn’t surprise you. 🙂
在我们深入挖掘之前，总结一下我们所处的位置可能会有所帮助——人工智能是软件的新演变。软件本身就是一种商品，是不可防御的。成功的软件公司依靠一些机制来产生防御性——规模经济、转换成本，以及最重要的网络效应。在这篇文章中，我将介绍前两个问题，但将重点介绍第三个问题。如果你以前听过我的文章，这不应该让你感到惊讶。🙂
The biggest economic difference between AI companies and regular old software companies comes down to marginal costs, i.e. the cost to produce an additional unit of the product for a new customer. Software (e.g. OG Microsoft Windows or Lotus 1-2-3) had zero marginal costs — once the product was developed, you could produce infinite copies for new customers. The internet made this even more effective, by making it easier for customers to access these products. It did create some additional costs to serve every customer (media storage and compute), but these were typically not significant as a % of revenue (e.g. Google).
人工智能公司与普通老软件公司之间最大的经济差异归结为边际成本，即为新客户生产额外单位产品的成本。软件（例如OG Microsoft Windows或Lotus 1-2-3）的边际成本为零 - 一旦产品开发出来，您就可以为新客户生产无限的副本。互联网使客户更容易访问这些产品，从而使这一点变得更加有效。它确实为服务每个客户（媒体存储和计算）带来了一些额外的成本，但这些成本通常占收入的百分比并不重要（例如Google）。
AI companies, on the other hand, have non-zero marginal costs because of steep costs involved in training and compute. Training a model requires an upfront investment in the order of tens of millions. In addition, a query to a large language model (LLM) via a product like ChatGPT costs 10x more than a standard keyword search on Google (excl. training costs). This dynamic introduces economies of scale at the layer of foundation models. In simple words, the steep cost of training and running these models creates an entry barrier for new competitors. Of course, the efficacy of economies of scale for new companies is questionable at best — for a couple of reasons. First, many of these models may not be economically viable, i.e. they cost more money than they make. Large scale companies like Google, Meta, Salesforce, Snowflake, etc. have an advantage here as they can subsidize these models with existing cash flow. Second, lighter and less computationally intensive models might be good enough for the majority of use cases.
另一方面，由于训练和计算涉及高昂的成本，人工智能公司的边际成本不为零。训练一个模型需要数千万美元的前期投资。此外，通过 ChatGPT 等产品查询大型语言模型 （LLM） 的成本比在 Google 上搜索标准关键字的成本高出 10 倍（不包括培训成本）。 这种动态在基础模型层引入了规模经济。简而言之，训练和运行这些模型的高昂成本为新的竞争对手创造了进入壁垒。当然，规模经济对新公司的效果充其量是值得怀疑的——原因有几个。首先，这些模型中的许多可能在经济上不可行，即它们的成本高于它们赚的钱。像谷歌、Meta、Salesforce、Snowflake等大型公司在这里有优势，因为他们可以用现有的现金流补贴这些模式。其次，对于大多数用例来说，更轻量级、计算密集度更低的模型可能就足够了。
At the application layer, economies of scale aren’t a viable mechanism for defensibility either — because computational costs are an order of magnitude lower compared to the model layer. Your ability to pay for an OpenAI’s API or compute for your app isn’t a sustainable advantage over a future competitor. Some applications like Character.ai have attempted to avoid this problem by vertically integrating, i.e. building their own customized models. Here again, the efficacy of economies of scale is questionable for the same reasons I mentioned earlier.
在应用层，规模经济也不是一种可行的防御机制，因为与模型层相比，计算成本要低一个数量级。与未来的竞争对手相比，您支付 OpenAI 的 API 费用或为您的应用程序计算的能力并不是一个可持续的优势。一些应用程序（如 Character.ai）试图通过垂直集成来避免这个问题，即构建自己的定制模型。在这方面，规模经济的效力也是值得怀疑的，原因与我前面提到的原因相同。
This leaves network effects and switching costs as the only realistic modes of defensibility for most applications. Let’s take a look at how they apply across three broad categories of AI applications, borrowing terminology from the the last three decades of the internet.
这使得网络效应和转换成本成为大多数应用程序唯一现实的防御模式。让我们来看看它们如何应用于三大类人工智能应用，借用互联网过去三十年的术语。
1. AI Singleplayer Apps: AI 1.0
1. AI 单人游戏应用：AI 1.0
These applications enable a direct interaction between a user and an AI product, i.e. software. This includes writing assistants like Quillbot, avatar generators like Lensa or Remini. They represent the earliest wave of AI applications after the emergence of foundational models. I call this category AI 1.0, as users interact directly with AI (created by companies), along the lines of Web 1.0 where users accessed information (created by companies).
这些应用程序实现了用户与人工智能产品（即软件）之间的直接交互。这包括像 Quillbot 这样的写作助手，像 Lensa 或 Remini 这样的头像生成器。它们代表了基础模型出现后最早的人工智能应用浪潮。我把这一类称为AI 1.0，因为用户直接与AI（由公司创建）交互，就像Web 1.0一样，用户在Web 1.0中访问信息（由公司创建）。
There is no possibility of network effects here. Network effects exist when the addition of a user increases the value of the product for all users — this requires a multiplayer interaction, i.e. one between two or more users. Interactions here are singleplayer, i.e. users are interacting with software. So by definition, network effects cannot exist.
这里不存在网络效应的可能性。当用户的加入增加了所有用户的产品价值时，就存在网络效应——这需要多人互动，即两个或多个用户之间的互动。这里的交互是单人游戏，即用户正在与软件交互。因此，根据定义，网络效应是不存在的。
Users interacting with a bot cannot create a network effect
用户与机器人交互无法产生网络效应
What about switching costs? If you can embed your application deeply into user workflows, it can be very difficult to replace — Salesforce being the most obvious example. So far, I haven’t seen any obvious cases of this with singleplayer AI apps. Instead, what we’re seeing in this category of AI applications (particularly consumer) is a wide variety of rapidly commoditized applications. Some of them may go viral (e.g. Lensa) because of the novelty (and clear utility) of AI, but that is rarely sustainable.
转换成本如何？如果你能将你的应用程序深深地嵌入到用户的工作流程中，它可能很难被取代——Salesforce就是最明显的例子。到目前为止，我还没有看到任何明显的单人 AI 应用程序出现这种情况。相反，我们在这类人工智能应用（尤其是消费者）中看到的是各种各样的快速商品化应用。由于人工智能的新颖性（和明显的实用性），其中一些可能会传播开来（例如Lensa），但这很少是可持续的。
This leaves vertical integration — building your own, customized model. As I mentioned earlier, the sustainability of that is an open question as well.
这就留下了垂直整合——构建您自己的定制模型。正如我前面提到的，其可持续性也是一个悬而未决的问题。
2. AI-Enabled Multiplayer-lite Apps: AI 1.5
2. 支持 AI 的多人轻量级应用程序：AI 1.5
There is another category of AI applications where the product is a bit more than “pure software/AI”. Instead, users create an AI-generated content or supply unit, which is then exposed to other users. Crucially, the process of creating AI-generated content is frictionless and takes just a few taps/clicks or a simple prompt. Character.ai is a great example, where creating a new bot takes minutes. AI song generators like Suno and Udio are also good examples, where users create AI-generated songs that others can discover.
还有一类人工智能应用，其产品不仅仅是“纯软件/AI”。取而代之的是，用户创建一个 AI 生成的内容或供应单元，然后将其暴露给其他用户。至关重要的是，创建 AI 生成内容的过程是顺畅的，只需轻点几下/点击或一个简单的提示。Character.ai 就是一个很好的例子，创建新的机器人只需几分钟。像 Suno和 Udio 这样的 AI 歌曲生成器也是很好的例子，用户可以在其中创建其他人可以发现的 AI 生成的歌曲。
Network effects do exist here… in theory. User adoption leads to more AI-generated content which other users can interact with. The challenge is that the friction for creating supply (AI chatbots or songs) is very low, and so is the barrier to reaching a critical mass of content. And of course, everyone knows the content is AI-generated, so the resulting emotional value of the content is also low — the identity of the person who created that content unit is completely irrelevant. You only need so much user+AI generated content before the product becomes “good enough”. This is a relatively low barrier for competition and the defensibility granted by these network effects (or their “strength”) is minimal. This is why I think of this category as AI 1.5 — halfway between AI 1.0 (early, obvious use cases) and true enablers of human interaction.
网络效应这里确实存在......在理论上。用户采用会导致更多 AI 生成的内容，其他用户可以与之交互。挑战在于，创建供应（人工智能聊天机器人或歌曲）的摩擦非常低，达到临界内容质量的障碍也非常小。当然，每个人都知道内容是人工智能生成的，因此内容产生的情感价值也很低——创建该内容单元的人的身份完全无关紧要。在产品变得“足够好”之前，你只需要这么多用户+人工智能生成的内容。这是一个相对较低的竞争壁垒，这些网络效应（或它们的“强度”）赋予的防御性是最小的。这就是为什么我认为这个类别是AI 1.5——介于AI 1.0（早期、明显的用例）和人类互动的真正推动者之间。
The ease with which users can create AI-generated content makes the barrier to critical mass very low and weakens the value of the resulting network effects
用户可以轻松创建人工智能生成的内容，这使得临界质量的门槛非常低，并削弱了由此产生的网络效应的价值
There is a possibility for switching costs to exist here, perhaps in a different form. You could make an argument that Character.ai users form an emotional bond with the bots they create (think something along the lines of an AI boyfriend/girlfriend), which could introduce psychological switching costs. That’s certainly a possibility, but I have not yet seen any evidence to back up that claim — strong long-term retention would be a good starting point. All we have seen so far, is data that shows high session frequency and time spent. This is always good to have, but it does not tell us anything about switching costs or defensibility.
这里可能存在转换成本的可能性，也许以不同的形式存在。你可以提出一个论点，即 Character.ai 用户与他们创建的机器人形成了情感纽带（想想类似于人工智能男朋友/女朋友的事情），这可能会带来心理转换成本。这当然是一种可能性，但我还没有看到任何证据来支持这一说法——强大的长期保留率将是一个很好的起点。到目前为止，我们所看到的只是显示高会话频率和花费时间的数据。拥有这总是好的，但它并不能告诉我们任何关于转换成本或防御性的信息。
3. True AI-Enabled Multiplayer Networks: AI 2.0
3. 真正的 AI 支持的多人网络：AI 2.0
So how do you create real, meaningful network effects with an AI application? Based on what we’ve talked about so far, the app needs to have an AI-enabled multiplayer interaction which involves a lot more than AI-generated output created by another user. These applications use AI to enable a (higher friction) multiplayer interaction that was previously impossible. My definition of the AI-enabled marketplace on the Speedinvest blog is a subset of this.
那么，如何使用 AI 应用程序创建真实、有意义的网络效应呢？根据我们到目前为止所讨论的内容，该应用程序需要具有支持 AI 的多人游戏交互，这涉及的不仅仅是由另一个用户创建的 AI 生成的输出。这些应用程序使用 AI 来实现（更高摩擦力）多人游戏交互，这在以前是不可能的。我在Speedinvest博客上对AI市场的定义就是其中的一个子集。
One (very early-stage) example of AI 2.0 is Haz — a recent portfolio companyof mine — which feeds email receipts into an AI model to create a social feed based on users past purchases. The model helps Haz filter out purchases that are not relevant for the social feed (e.g. a hammer) and extract or find relevant product information to fill out the feed. This content unit then becomes the basis for interactions between users — from “likes" to bidding on what your friends own.
AI 2.0 的一个（非常早期）的例子是 Haz——我最近的一家投资组合公司——它将电子邮件收据输入到 AI 模型中，以根据用户过去的购买情况创建社交源。该模型可帮助 Haz 过滤掉与社交信息源不相关的购买（例如锤子），并提取或查找相关产品信息以填充信息源。然后，这个内容单元成为用户之间互动的基础——从“喜欢”到对你的朋友所拥有的东西进行出价。
The transition to AI 2.0 will require a more involved multiplayer interaction that is assisted by AI
向 AI 2.0 的过渡将需要在 AI 的协助下进行更复杂的多人交互
If this sounds vaguely familiar, it should. This mirrors the transition from Web 1.0 to Web 2.0, i.e. from applications that allowed users to access information to those that allowed users to both create and consume content. Web 2.0 enabled multiplayer interactions, as we see today on social networks like Facebook and marketplaces like Airbnb. While network effects have existed in the technology world for hundreds of years (starting with the original telephone), Web 2.0 led to a significant expansion the number of companies built on them — and consequently, value creation.
如果这听起来有点熟悉，它应该。这反映了从 Web 1.0 到 Web 2.0 的过渡，即从允许用户访问信息的应用程序到允许用户创建和消费内容的应用程序。Web 2.0 实现了多人交互，正如我们今天在 Facebook 等社交网络和 Airbnb 等市场上看到的那样。虽然网络效应在技术世界中已经存在了数百年（从最初的电话开始），但 Web 2.0 导致了建立在它们之上的公司数量的显着增加——从而创造了价值。
Can AI 2.0 bring about a similar expansion in AI applications with network effects? It’s certainly possible. We’re still very early, so I can’t point to any large scale successes of this yet. However, we are starting to see some early product concepts form around this idea, including Haz and a few others. Let’s just say this is a space I’m paying attention to very closely.
AI 2.0能否带来与网络效应类似的AI应用扩展？这当然是可能的。我们还为时过早，所以我还不能指出这方面有任何大规模的成功。然而，我们开始看到一些早期的产品概念围绕着这个想法形成，包括 Haz 和其他一些产品。可以说，这是一个我非常关注的领域。
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
