Writing · 2026年9月8日 · 9 分钟阅读
AI 应用会活下来，但不会以现在的方式
模型会继续吞掉今天的功能，但基础设施不会自动发明属于 AI 时代的产品。
本文基于我最初发布在 Bonjour 的中文动态整理。 查看 Bonjour 原文 ↗
很多人说，从长远来看，AI 应用公司仍然拥有巨大的机会。更长期来看，我们都会死。所以真正有用的问题，不是无限期地乐观或悲观，而是在一个不短也不长的时间尺度上判断：AI 应用会不会被模型公司吞掉？
我的结论是：AI 应用会活下来，但不会以现在的方式。
今天的应用为什么看起来如此脆弱
短期来看，大量 AI 应用确实很难建立稳定边界。模型能力持续扩张，推理成本不断下降，平台公司也在把搜索、写作、编程和 Agent 能力直接整合进基础产品。许多建立在单一模型能力或单次调用之上的产品，会随着底层更新迅速失去差异。
但这并不等于“应用层没有价值”。它只说明，能够被一次模型更新替代的东西，本来就还没有成为真正的应用。
如果把 Cloud 和算力看成油田与电网，把模型公司看成炼油厂与发电设备，那么很多所谓 AI 应用仍然只是输油管、电灯泡，或者给旧机器换上的一台电动机。它们证明了新基础设施有用，却还没有重新定义人如何工作和生活。
基础设施不会自动发明产品
拥有油田和炼油能力的公司，并没有因此创造宝马或奔驰。汽车之所以成为产品，不只是因为汽油存在，也因为驾驶体验、制造体系、产品设计、品牌和对用户的理解被组织在了一起。
电力也是一样。电网建成、电路接入，并不会自然产生冰箱、空调、洗衣机、电视和电脑。电力只是让这些产品成为可能；至于人们会用电做什么，仍然需要应用公司在真实生活中一点点发现。
AI 也会经历类似的过程。模型提供智能，Token 成为一种基本消费单元，Agent 开始执行任务——这些变化非常重要，但它们仍然只是产品出现的前提。
从基础设施，到 AI+，再到 AI Native
我倾向于把当前演进分成三个阶段。
1. 基础设施阶段
市场关心谁能训练出更好的模型、拥有更多算力，以及谁能把推理成本压得更低。人们探索的是这种新型基础设施的能力边界。
2. AI+ 阶段
大量公司把 AI 接入已有的软件、岗位和流程：给原有产品增加生成能力，把任务交给 Agent，用自然语言补充或替代部分 GUI。
这些改造能够产生真实价值，但很多仍像早期工厂把蒸汽机换成电动机——动力来源变了，厂房布局、生产流程、组织方式和人的角色却没有改变。
3. AI Native Applications 阶段
真正的 AI Native 产品不是给旧产品加一层 AI，也不是给模型套一个垂直场景的外壳。它以“智能持续存在”为前提，重新设计人的行为、关系、工作流和生活方式。
移动互联网时代的原生产品并不是把 PC 网站缩进手机。AI 时代的原生产品，也不会只是把更多按钮换成聊天框。
To B：重构生产，而不只是替换工具
工厂刚接入电力时，最直接的做法是用电动机替换中央蒸汽机，但仍保留原来的动力轴、机器排列和工作节奏。真正的变化发生在每台机器都拥有独立电机之后：厂房不再围绕中央动力轴设计，生产线和组织方式得以重新布局。
今天的 Coding Agent 已经开始进入生产流程，改变人与代码、软件和任务之间的关系。但如果一个组织只是把原有步骤逐项交给 Agent，它仍然处于“替换电动机”的阶段。
更大的机会在后面：当智能成为普遍基础设施之后，公司如何重新设计协作、决策、责任与交付？
To C：AI 时代还没有自己的冰箱和空调
消费级 AI 目前大多停留在聊天框中。聊天框很重要，就像电灯一样会成为日常；但电灯并不是电力时代的全部。
真正塑造现代生活的，是后来出现的冰箱、空调、洗衣机、电视、电脑和手机。它们共享同一套电力基础设施，却创造了此前无法被清楚表达的需求，也改变了家庭、娱乐、沟通与工作的方式。
AI 的消费级产品也可能如此。未来重要的形态很可能与今天的 Context 工具、Claude Code 或通用聊天产品完全不同。它未必以对话开始，也未必以一次任务完成结束。它可能进入人的创造、表达、沟通、记忆和决策，与人建立长期关系。
模型公司当然可以创造重要应用，也可能占据一些巨大的品类。但拥有最强模型，并不等于天然知道 AI 时代的“空调”“电脑”或“手机”应该是什么。
真正的问题
今天的许多 AI 应用可能会消失。这只能说明第一批应用还不成熟，不能证明应用层最终没有价值。
我们已经拥有了 AI 时代的电力，却还没有发明出它的冰箱、空调和电脑。
因此，下一代重要的应用公司未必拥有最强的模型。它们更可能率先回答一个产品问题：
当智能、Agent 和 Coding Agent 成为基础设施，当 Token 成为基本消费单元时，我们究竟会以什么方式工作、生活和创造？
谁能影响这种方式，谁才真正开始定义 AI 时代的应用。
People often say that AI application companies still have enormous long-term potential. Over an even longer horizon, all of us will be dead. So the useful question is neither permanent optimism nor permanent pessimism. It is whether, over a horizon long enough to matter but short enough to reason about, AI applications will simply be absorbed by model companies.
My answer is that AI applications will survive — just not in their current form.
Why today’s applications look so fragile
In the near term, many AI products do have weak boundaries. Models keep expanding their capabilities, inference keeps getting cheaper, and platform companies keep integrating search, writing, coding, and agentic behavior into their core products. A product built around one model capability or one API call can lose its differentiation after a single infrastructure update.
But that does not mean the application layer has no value. It means that something replaceable by one model release had not yet become a real application.
Imagine cloud and compute as oil fields and electrical grids, and model companies as refineries and generators. Many things we call AI applications today are still pipelines, light bulbs, or electric motors attached to old machines. They demonstrate the usefulness of the new infrastructure without redefining how people live or work.
Infrastructure does not invent products by itself
Oil companies with fields and refineries did not automatically create BMW or Mercedes-Benz. A car became a product not only because fuel existed, but because driving experience, manufacturing, industrial design, brand, and an understanding of users were assembled into one system.
Electricity followed the same pattern. Building the grid did not automatically produce refrigerators, air conditioners, washing machines, televisions, or computers. Electricity made those products possible; application companies still had to discover, inside real life, what people would do with it.
AI will move through a similar process. Models supply intelligence, tokens become a basic unit of consumption, and agents begin to execute work. These changes are profound, but they are still preconditions for products rather than finished products themselves.
From infrastructure, to AI+, to AI-native applications
I find it useful to think in three stages.
1. The infrastructure stage
The market focuses on who can train the strongest models, secure the most compute, and reduce inference costs. The central question is the capability boundary of the new infrastructure.
2. The AI+ stage
Companies add AI to existing software, roles, and workflows. They introduce generation into established products, delegate tasks to agents, and use natural language to supplement or replace parts of the GUI.
This creates real value. But much of it resembles an early factory replacing a steam engine with an electric motor while leaving the shafts, floor plan, production process, and organization unchanged.
3. The AI-native application stage
A genuinely AI-native product is not an old product with an AI layer, nor a model wrapped in a narrow vertical interface. It assumes that intelligence is continuously available and redesigns behavior, relationships, workflows, and ways of living around that fact.
Mobile-native products were not merely desktop websites squeezed onto smaller screens. AI-native products will not be interfaces where more buttons have been replaced by chat boxes.
B2B: redesigning production, not merely replacing tools
When factories first adopted electricity, the obvious move was to replace the central steam engine with an electric motor while keeping the old power shafts, machine arrangement, and working rhythm. The deeper transformation came when every machine could have its own motor. Factory layouts no longer had to orbit a central shaft, so production lines and organizations could be redesigned.
Coding agents are already entering production and changing the relationship between people, code, software, and tasks. But an organization that merely hands each existing step to an agent is still in the motor-replacement stage.
The larger opportunity comes later: once intelligence is ordinary infrastructure, how should a company redesign collaboration, judgment, accountability, and delivery?
Consumer AI: we still have not invented the refrigerator
Most consumer AI still lives inside a chat box. Chat will matter and may become as ordinary as electric light, but the light bulb was not the whole electrical age.
Modern life was shaped by the refrigerator, air conditioner, washing machine, television, computer, and phone. They shared the same electrical infrastructure, yet each created needs that people could not have clearly articulated beforehand. Together they changed domestic life, entertainment, communication, and work.
Consumer AI may follow the same path. Its most important future forms may look nothing like today’s context tools, Claude Code, or general chat products. They may not begin with a conversation or end when one task is completed. They may enter creation, expression, communication, memory, and decision-making, building a durable relationship with the person using them.
Model companies can certainly build important applications and may dominate some enormous categories. But owning the strongest model does not mean automatically knowing what the air conditioner, computer, or phone of the AI era should be.
The question that matters
Many of today’s AI applications may disappear. That tells us the first generation is immature; it does not prove that the application layer will remain unimportant.
We already have the electricity of the AI era. We have not yet invented its refrigerator, air conditioner, or computer.
The next important application company may not own the strongest model. It may simply answer a product question earlier and better than everyone else:
When intelligence, agents, and coding agents become infrastructure, and tokens become a basic unit of consumption, how will we work, live, and create?
Whoever shapes that answer will have begun to define the application layer of the AI era.