最应该毫不犹豫接入的平台型初创公司，是那些直接提供现实世界服务的企业（如 Uber、DoorDash 等）。 The main startups that should probably integrate without thinking twice are those that do real world things (Uber / DoorDash etc).
Anish AcharyaAI Apps investing, A16Z
目录
观点原文
最应该毫不犹豫接入的平台型初创公司，是那些直接提供现实世界服务的企业（如 Uber、DoorDash 等）。
The main startups that should probably integrate without thinking twice are those that do real world things (Uber / DoorDash etc).
ABAB AI 解读
一、一句话结论
凡是“离线履约、线上只是入口”的公司，接入 ChatGPT 不是让渡核心资产，而是放大分发能力。
Uber / DoorDash 这一类，恰好符合这个条件。
二、核心分界线：你的“护城河”在哪？
这句话的真正判断标准不是大不大，而是：你的价值，是在聊天里完成，还是在现实世界完成？
1️⃣ 现实世界履约型（该接入）
- Uber：司机网络 + 实时调度
- DoorDash：骑手网络 + 商家履约
- Airbnb：房源审核 + 线下住宿体验
👉 ChatGPT 无法替代：
- 司机
- 骑手
- 仓配网络
- 实体资源
它只能：
- 帮用户“下单更快”
- 帮你“少做 UI”
- 这对你是纯增量分发。
2️⃣ 纯数字能力型（应谨慎）
- 写文案
- 总结信息
- 做推荐
- 基础分析
👉 这些正是 LLM 的核心能力区
一旦你把场景交出去：
- OpenAI 可以直接内建
- 你会被“能力吸收”
三、为什么现实世界公司“不怕被去中介化”？
因为去中介化在这里不成立。
想象一个极端场景：ChatGPT 想“取代 Uber”
它需要：
- 数百万司机
- 法律合规
- 保险体系
- 城市级调度
- 客服与纠纷处理
👉 这不是软件问题，是基础设施问题。
所以：AI 只能是入口，不可能是履约者。
四、对现实世界平台来说，ChatGPT 只是“更聪明的搜索框”
本质上，这些公司把 ChatGPT 当作：
- 新一代 Siri
- 新一代 Google Search
- 新一代 App Launcher
而不是：平台主人
所以接入的博弈是：
- 如果 AI 成为主要入口 → 你必须在
- 如果 AI 失败 → 你几乎没损失
- 典型的“无脑正期望值”决策。
五、为什么这类公司“现在不接入反而危险”？
因为一旦用户形成习惯：“在 ChatGPT 里打车 / 点外卖 / 订酒店”
而你不在：
- 订单会直接流向在场的对手
- 用户不会“多想一步再去 App”
- 入口迁移一旦发生，很难逆转。
六、这句话真正的“潜台词”
作者其实是在给创业者一个极冷静的分层建议：
❌ 不要幻想 ChatGPT App Store 是“公平分发市场”
✅ 把它当成“入口战争的新战场”
而：
- 现实世界履约型企业 = 天然强势方
- 纯软件 / 纯 AI 初创 = 天然弱势方
七、一句话总结
凡是把价值兑现留在现实世界的公司，AI 入口只会放大你；凡是把价值完全留在对话里的公司，AI 入口迟早会吞掉你。
AI
相关观点
- 软件正在迎来属于自己的「YouTube 时刻」 去年我写过一篇文章，讲 YouTube 如何很好地预示了编程世界将会发生什么。2005 年 YouTube 上线时，并没有填补一个显而易见的内容空白；但二十年后，它已经成为一家 5500 亿美元市值的巨头，其文化影响力甚至超过了传统电视。 现在，同样的“长尾创作浪潮”正在软件领域发生。 如果你最近几周在 X 上活跃，你一定已经看到了这些变化： Tobi 做了一个自定义的 MRI 仪表盘——按专业人士的说法，这原本需要商业软件才能查看； Marc 用 Wabi 生成了一部“技术乐观主义”电影和小说推荐系统； Levelsio 和 Joe Weisenthal 正在直播为观众实时架构应用； 有人直接在 CLI（命令行） 里完成整套广告活动的投放。 如果你以前还有“我为什么没做点东西”的借口，那这些借口现在已经变得非常站不住脚了。 在 LLM 出现之前，这些应用中的大多数可能根本不会存在。 谁有时间去啃陌生的 API 文档？ 谁会去写定制的医学影像软件？ 谁愿意投入可能需要数周时间，只为了做一个也许永远跑不出 localhost 的项目？ 尤其是如果你不是程序员，这些创造门槛几乎是不可逾越的： 你首先得学会写代码； 然后还得跟上层出不穷的框架； 一旦你停下脚步，整个行业就已经把你甩在身后。 而 Cursor、Codex、Claude Code、Replit、Wabi 这类工具，把“从想法到可用应用”的时间，从几周压缩到了几小时。 如果你多年前写过代码，但没跟上 Next.js 的更新——已经不重要了。 如果你从没写过代码——也不重要了，你只需要在文本框里输入，看会发生什么。 今天，任何人都可以发布一个应用。 YouTube 给我们的启示 把视频创作当作一种媒介来思考。在它的大部分历史中，导演需要“绿灯”和资金支持。上世纪 80、90 年代，门槛开始松动：PTA、Soderbergh、Tarantino 都是用小预算、非专业团队起步的。到了 2000 年代，每个人口袋里都有了摄像头，再加上 YouTube 这样的平台，什么算娱乐被彻底重写了。 YouTube 从猫视频进化到 MrBeast 的挑战节目，花了几年时间。但完成进化之后，我们就进入了今天的媒体生态：来自 YouTube 创作者的娱乐内容，和传统电视台同样常见。 软件也正在走同样的路径： 好莱坞时代：深度专业、动辄数千万美元才能做点东西； 90 年代独立导演时代：YC 创始人（如 Chesky、Armstrong）这样的“外来者”开始破局； YouTube 时代：LLM 人人可用。 软件的三大变化 互联网（部分得益于 YouTube）催生了人格驱动型媒体：VC、记者、创始人、政治人物——所有人都被鼓励多发帖、上播客、找受众、打造个人帝国。但直到最近，软件还不是一种个人表达或帝国构建的方式。这即将改变，原因有三： “建造者”的潜在市场被极大扩展 过去，只有一小撮人对软件感兴趣：刷 Tech/VC 的 X 动态、听同样的播客、读 Paul Graham 的文章。现在，你不需要“对软件感兴趣”也能做软件，你只需要对好点子感兴趣。 软件成为表达媒介，而不仅是工具 比如 Riley Walz，通过发布极具时代气息的软件项目建立了个人品牌：可视化旧金山停车执法人员、渲染 Epstein 文件等。我预计 LLM 会让这种“即时、好玩”的软件变得极其容易做出来。人们做一个有趣的软件，所需精力将和发一条有趣的 X 帖子差不多。 软件的价值会复利，内容的价值会衰减 YouTube 创作者必须不断产出视频才能保持影响力；大多数内容的“折旧期”是 2–5 年。软件却可以只发布一次，然后随着用户增长无限积累价值——这是一个被严重低估的优势。 模仿效应与“孩子们没问题” 我们的决策往往是模仿性的。 很多创作者说，他们开始创作内容，是因为看到别人也在这么做。如果有人告诉你想辞职去做视频，你可能会翻白眼——不是因为这很蠢，而是因为这太容易开始了，那为什么你自己还没做？ 现在，同样的模仿能量正在冲击软件领域。 你看到朋友在做东西，你也想做。软件正在成为一种病毒式传播的媒介，很快，每个人都将再也找不到不去做它的借口。 如今关于“孩子们”的焦虑很多。 我没有焦虑，只有一点点羡慕。AI 极大地民主化了杠杆和生产力，尤其是对有创造力的人来说。对于一个有好想法的年轻人而言，从未有过比现在更好的时代。 我相信，这一代孩子会过得很好。 Software's YouTube Moment is Happening Now Last year, I wrote about how YouTube is a pretty great herald of what we can expect to happen to coding. YouTube launching in 2005 didn't seem to fill an obvious content gap, and yet two decades later it's a $550bn business that's more culturally relevant than traditional TV. We're now seeing the same long-tail creation wave for software. If you've been on X the past few weeks, you've seen it happening: Tobi built a custom MRI dashboard that (according to the professionals) should’ve required commercial software to view. Marc is generating a techno-optimist movie and novel recommender with Wabi. Levelsio and Joe Weisenthal are architecting apps live for an audience. People are shipping entire ad campaigns from the CLI. If you previously had excuses for not building something, they just got a lot less convincing. Before LLMs, all of these apps might have never existed. Who had the time to dig through unfamiliar API documentation, or write custom medical imaging software, or allocate what might take weeks of time to build software that may never leave localhost? Especially if you weren’t a programmer, there were seemingly insurmountable barriers to creation. You had to learn how to code, for one. And then you had to stay up-to-date with a whole catalogue of frameworks, and if you stopped paying attention, the industry would move on without you. Tools like Cursor, Codex, Claude Code, Replit and Wabi have compressed the time from idea to working app from weeks to hours. If you programmed years ago but aren't current on Next.js releases, it doesn't matter anymore. If you've never programmed, you can type into a text box and see what happens. Today, anyone can ship an app. https://x.com/pmarca/status/2010858340088479886 What YouTube Teaches Us Think about video creation as a medium. For most of its history, directors needed a green light and funding. In the late 80s and early 90s, the reins loosened: directors like PTA, Soderbergh, and Tarantino all got started with small budgets and non-professional crews. In the 2000s, barriers collapsed further when everyone got a camera in their pocket and platforms like YouTube changed what could qualify as entertainment. It took a few years for YouTube to evolve from cat videos to Mr. Beast challenges. But after that evolution, we entered the media ecosystem we're in now, where it's as common for entertainment to come from creators on YouTube as from traditional broadcasters. Software is following the same path: the Hollywood era (deep expertise, tens of millions to build anything) gave way to the 90s indie director era (YC founders like Chesky and Armstrong as outsiders breaking in) which led to the YouTube era (an LLM at everyone's fingertips). Three implications for software The internet (partially thanks to YouTube) has given rise to personality-driven media: pretty much everyone, from VCs to journalists to founders to politicians are encouraged to write more posts, appear on all the hot podcasts, find an audience, and create an empire. But until recently, software hasn’t been a form of personal expression or empire-building. This is going to change for three big reasons: The addressable market for "builders" just expanded dramatically. For a while, software excited only a small slice of the population: people keyed into Tech/VC X feeds, consuming the same podcasts and Substacks, and quoting Paul Graham essays. Now you don't need to be "interested in software" to build software. You just need to be interested in good ideas. Software becomes a medium for expression, not just utility. Think about developers like Riley Walz who've built a name shipping zeitgeisty projects, like visualizing SF's parking cops, or rendering the Epstein files. I expect LLMs will make this kind of entertaining, of-the-moment software much more accessible. People will create funny software with about as much effort as a funny post on X. Software value compounds; content value decays. Creators on YouTube need to keep churning out videos to stay relevant: most content follows a 2-5 year depreciation schedule. Software can launch once and accrue value indefinitely as users come onboard. This is an underappreciated advantage of the medium. Our decisions are mimetic. Many creators say their drive to produce content online comes from seeing others do it. If someone tells you they want to quit their job and make videos, you'd roll your eyes, not because it's dumb, but because it's so easy to just start, so why haven't they? The Kids Are Alright Some people worry: shouldn't kids aspire to be astronauts or doctors? I have a different read. The reason we have so many influencers is because it's what you do if you're entrepreneurial and want to take destiny into your own hands. Maybe it's agency they're emulating, not YouTubing specifically. The same mimetic energy is now hitting software. You see your friends building things, and you want to do it too. Software is becoming a viral medium, and soon no one will have any excuses not to make it. There's a lot of angst about "the kids" these days. I have no angst, but maybe a little envy. AI has massively democratized leverage and productivity for creative people. There has never been a better time to be a young person with great ideas. I think the kids are gonna be alright.Anish Acharya · 2026/01/16
- Hannah Stolberg 与 Team OS：把团队隐性知识变成 AI 可调用的组织操作系统Hannah Stulberg · 2026/04/17
- Evan Spiegel：拒绝被收购，不做流量生意，只押注下一代计算平台Evan Spiegel · 2026/04/16
- AI产品不是做模型，是控制行业入口Winston Weinberg · 2026/04/15