纳瓦尔:AI时代,纯软件正迅速变得不可投资
来源:Naval Podcast / YouTube
嘉宾:Naval Ravikant(AngelList联合创始人兼董事长)
总时长:29分36秒
核心摘要
Naval Ravikant于2026/4/28分享自己在2025年12月Claude Opus 4.5发布后重返编码的经历。他构建了个人App Store——一个能一键将定制应用推送到iPhone的网页;用一句话prompt生成融合Tonal功能、Apple Health集成和肌肉图谱的健身应用;提出”vibe coding是比电子游戏更有趣的元宇宙”;断言纯软件不可投资(硬件、网络效应、AI模型才是风投标的);逐一点评Claude/ChatGPT/Gemini/Grok的差异化定位;揭示AI eager-to-please的本质缺陷(像猎狗——比你会抓鸭子,但指错目标会抓回一只鸟);预言Apple放弃AI将是本十年最大战略失误;最终以coding agents as 24/7 customer service reps的愿景收束——一人软件公司可扩展至数百万用户、数十亿美元收入。
一、A Return to Coding(00:15-03:08)
1.1 Claude Opus 4.5:拐点时刻
-
2025年12月Claude Opus 4.5发布,coding agents达到inflection point
“Around December of 2025, the coding agents in AI hit an inflection point with the release of Claude Opus 4.5.”
-
用户反馈:“这就像一个始终在线、速度极快、本质上免费且随时准备效劳的初级程序员”
“This is an agent that stays on track, can build apps soup to nuts, can solve thorny problems, and really feels like having a junior programmer at your disposal who’s fast, essentially free, and ready to please.”(这是一个始终按计划推进、能从头到尾开发应用程序、能解决棘手问题的助手,使用起来就像是随叫随到的一名初级程序员——他动作敏捷、几乎不用花钱,而且总是乐于助人。)
1.2 Naval重返编码
-
Naval有CS学位,理解计算机架构、网络、芯片、算法,但已decades没有认真编码
“I have a computer science degree; I understand computer architecture and networking, a little bit of chips, algorithms, et cetera. But I haven’t seriously coded in a long time.”
-
Activation energy被抹平:以前需要连接GitHub→Vercel/Firebase/Railway→学习大量jargon和工具,现在AI让这一切变得简单
“The activation energy to writing code is really high. You have to hook up all these different services to each other… And the AI now makes it really easy.”
-
使用Claude Code入门,也使用Codex解决更棘手的bug和深层问题 -
立即上瘾:“incredibly fun”
1.3 从coding assist到coding agent的范式转移
-
不再是”给出代码片段→复制粘贴到IDE”的模式 -
现在打开terminal,agent作为完整agent运行 -
几乎所有GitHub代码都是Unix based,text in text out——这正是这些模型擅长的
“Almost all code on GitHub is Unix based, which is all text in text out, and that is exactly what these models are built for.”
二、The Personal App Store(03:08-06:12)
2.1 个人App Store的构建
-
Naval构建了一个personal app store——一个网页,能一键将定制应用推送到他的iPhone
“I built a personal app store, a webpage that delivers custom apps straight to my iPhone with one click.”
-
Apple限制广泛分发只能在自己的设备上运行,但”在这个约束内,它真的很神奇”
“Apple blocks wide distribution so it only works for your own devices, but inside that constraint it is genuinely magical.”(苹果限制了该功能的大范围传播,因此它仅适用于你自己的设备,但在这一限制范围内,它确实堪称神奇。)
2.2 健身应用:一句话prompt生成
-
One-shot生成:一个prompt就生成了完整的custom workout app功能集成:
“I one-shotted a full custom workout app in one prompt combining Tonal features, Apple Health integration, a muscle body diagram, and strength scores pulled from scientific papers.”
2.3 晚餐桌场景:五分钟展示定制应用
-
核心场景
“You can literally be at dinner with someone, they describe an app they want, you describe it to Claude, and five minutes later you’re showing them that app on your phone.”
-
这展示了vibe coding的即时性和可及性——从想法到可交互产品的压缩时间从数周降至数分钟
💡 思考点:个人App Store的本质是”bypass Apple的审查和30%抽成”。当AI coding agents能one-shot生成完整应用并直接推送到手机时,App Store作为distribution gatekeeper的角色被架空。这是技术民主化的胜利,还是平台治理的灾难?
🔗 关联点:这与Epic Games vs Apple的诉讼形成对照——Epic用法律手段挑战Apple的垄断,而Naval用技术手段绕过它。后者可能更有效,因为它不需要等待法院裁决。
三、Vibe Coding Is a Video Game with Real-World Rewards(06:12-10:23)
-
3.1 Vibe Coding vs 电子游戏
-
电子游戏的反馈循环:保持你在能力边缘,不断给予反馈和奖励——但bounded和fake -
Vibe Coding的反馈循环:同样有即时反馈,但runs on a Turing machine——unbounded且actually matters
“Video games hook you by keeping you at the edge of your capability with constant feedback and rewards, but they are bounded and fake. Vibe coding has that same loop but it runs on a Turing machine so it is unbounded and actually matters.”(电子游戏通过持续的反馈和奖励,让你始终处于能力极限的边缘,从而让你沉迷其中,但它们是有限的且虚假的。Vibe编程虽然也有同样的循环机制,但它运行在图灵机上,因此是无限的,而且真正具有意义。)
3.2 Airchat重建案例
-
Airchat(语音社交平台)以前需要8-9名工程师、近一年时间开发 -
Naval完全靠自己重建,exactly the way he wanted it
“I rebuilt Airchat, which previously took a team of 8 or 9 engineers for nearly a year, entirely by myself and exactly the way I wanted it.”
-
Builder population expansion:能构建应用的人口比例从约0.1%提升到1-3%
“This raises the percentage of people who can build apps from roughly 0.1% to maybe 1 to 3%.”
3.3 为什么vibe coding比游戏更有趣
-
More productive(更有生产力) -
More constructive(更有建设性) -
Better feedback loops(更好的反馈循环) -
Build something you want(构建自己想要的东西) -
At the bleeding edge of technology(站在技术最前沿) -
甚至可能make money或career(虽然careers are kind of dead) -
通过实践学习大量计算机知识
💡 思考点:Naval将vibe coding定位为”比电子游戏更有趣的元宇宙”——这意味着游戏行业面临被”productive entertainment”替代的风险。当创造的乐趣大于消费的乐趣时,娱乐产业的商业模式是否需要重构?
🔗 关联点:这与Andrej Karpathy的”vibe coding已死,agentic engineering登场”形成时间线上的对照——Karpathy在2026年4月宣告vibe coding过时,而Naval在2026年4月底仍在热情推广。两人的分歧在于:Karpathy看重scalable software quality,Naval看重individual creative joy。
四、Pure Software Is Uninvestable(10:23-14:09)
4.1 核心论断:纯软件不可投资
-
Nivi引用Naval推文:“Pure software is rapidly becoming uninvestable”(纯软件正迅速变得难以获得投资) -
Naval纠正:那是watered-down version(阉割版),真正想说的是”pure software is uninvestable. Full stop.”
“That’s a watered-down version of what I really wanted to say, which is that pure software is uninvestable. I would just full stop right there.”
-
条件句:如果你的全部优势是”我在构建别人不知道怎么构建的酷软件”,那就是uninvestable
“If your whole advantage is like, ‘Hey, I’m building cool software that other people don’t know how to build,’ I think that’s uninvestable.”
4.2 两个原因
-
原因一:别人今天就能hack it together(拼凑出来)
“One is they can just hack it together today.”
-
原因二:coding agents进步速度如此之快,一年之内甚至不到一年就能构建架构良好的可扩展软件scalable software with good architecture
“The coding agents are getting better so quickly that within a year, or even less, they’ll probably be building scalable software with good architecture.”
-
结论:genie is out of the bottle(魔兽已出笼)
“That genie is out of the bottle.”
4.3 风投应该转向哪里
-
硬件(hardware) -
网络效应(network effects) -
AI模型(AI models) -
核心观点:training AI models is the new building software训练AI模型就是新的软件开发
“Training AI models is the new building software — for however long that lasts, until autoresearch and autotraining starts working.”
💡 思考点:Naval的”纯软件不可投资”论断与Jensen Huang的”从电子到Token”框架形成互文——真正的价值不在bits而在atoms + distribution + models。当 coding 被commoditized,风投的理性选择是上游(包含atoms–芯片,基础设施,能源 & 模型)和下游(分销/网络效应–平台、社区、品牌),而非中游(纯软件)。
🔗 关联点:这与Scott Galloway的”AI不会夺走你的工作,懂AI的人会”形成对照——Naval说的是投资逻辑,Galloway说的是就业逻辑,但两者共享同一个底层认知:纯code不再是壁垒。
五、A Place for Each Model(14:09-17:44)
5.1 同一模型多实例的陷阱
-
十个相同模型的实例互相交谈,只是对同一个问题投入十倍的token(throwing ten times more tokens at the same problem)。 -
不是actually smarter,只是更多计算资源被浪费
5.2 四模型生态位切分
|
|
|
|
|---|---|---|
| Claude |
|
|
| ChatGPT |
|
|
| Gemini |
|
|
| Grok |
|
|
💡 思考点:Naval提到的“十个相同模型的实例互相交谈,只是对同一个问题投入十倍的”,多agent协作的价值来自模型的异质性(diversity of models),不是同质性(同一模型的数量堆叠),他主张使用不同专长的模型进行协作而不是一个模型下构建多个Agent——不同的taste、不同的盲区、不同的strengths互相校验。
六、AI Is Eager to Please(17:44-21:58)
6.1 AI eager-to-please的本质缺陷
-
如果你push模型稍微接近答案了,它会找到那个答案——无论对错
“If you push these models even slightly toward an answer they will find that answer whether it is right or not.”(如果你稍微引导这些模型去寻找一个答案,它们就会找到那个答案,无论对错)
-
核心问题:模型太急于取悦,缺乏说”我不确定”的能力
6.2 Context window的限制
-
当代码库太大超过context window(约1M tokens),模型开始:
“As the codebase gets too large for the context window they start fixing the same bug five times, patching the surface when the problem is deeper, or just deleting the feature to make the bug disappear.”
6.3 猎狗隐喻
-
核心比喻:
“It’s a little bit like a dog. It’s better than you at catching the duck, but if you point it at a bird that’s not a duck, it might take that bird down instead.”
-
AI比人类更擅长执行,但如果你指错了方向,它会 enthusiastically 执行错误的方向 -
有效组合:human operator + coding model(人工操作员 + 编码模型),但你必须stay involved
“The combo that works is human operator plus coding model, but you have to stay involved.”
💡 思考点:Naval的猎狗隐喻揭示了AI代理的核心悖论——能力越强,方向性错误的风险越大。当agent可以one-shot生成完整应用时,”想要什么”的判断比”如何实现”的 execution 更重要。这是否意味着product taste正在取代coding skill成为最高杠杆技能?
🔗 关联点:这与Cat Wu的”product taste是AI时代PM必备技能”一致——当execution被automated,judgment成为瓶颈。
七、Why Math and Coding?(21:58-24:04)
7.1 为什么编码模型进步这么快
-
编码有tons of data,且verification是automatic的——code要么pass test要么fail
“These models excel at coding because there is tons of data and verification is automatic. Code either passes the test or it doesn’t.”(这些模型在编程方面表现出色,因为数据量庞大,且验证过程是自动进行的。代码要么通过测试,要么不通过。)
-
创意写作难以训练,因为没人能algorithmically verify什么是好的
7.2 顶级工程师反馈循环
-
编码模型近期进步最大的原因:top engineers开始使用它们,their taste fed back into training data
“The biggest reason coding models improved so fast recently is that top engineers started using them and their taste fed back into training data.”
-
这形成了一个正向循环:更好的模型→更多工程师使用→更多高质量feedback data→更好的模型
八、The Beginning of the End of Apple’s Dominance(24:04-27:43)
8.1 iPhone变成什么
-
当所有交互通过AI agent进行时,手机变成”just a screen with a battery and a connection”
“When all your interactions go through an AI agent, the phone becomes just a screen with a battery and a connection.”
-
Apple已经在使用Google的Gemini,此时iPhone和Android的区别是什么?
8.2 Apple的战略失误
-
核心论断:
“I think Apple giving up on AI will go down as the biggest strategic mistake in the tech industry of this decade.”
-
类比Microsoft错过mobile wave——Apple不会消失,但market cap会compress(市值会缩水) -
Apple将被迫competing on Samsung style margins, not monopoly margins(竞争三星风格的利润,而不是垄断利润)
8.3 Personal App Store对Apple生态的冲击
-
AI coding agents能将一次性定制应用直接发送到您的手机 -
这bypass了App Store的审查和30%抽成 -
用户不再需要从App Store下载应用,而是直接生成自己需要的 -
Apple的垄断地位被侵蚀:distribution control不再是护城河
九、Coding Agents As Customer Service Reps(27:43-29:36)
9.1 24/7自动修复bug
-
Naval在应用内构建bug reporting infrastructure -
用户遇到bug→点击按钮→logs自动上传→bug filed into server -
Claude每24小时review所有bug reports,自动修复,放入side branches供Naval review
“I have Claude go every 24 hours through all the bug reports and it just fixes them all, by itself, without my having to intervene. And it puts all the fixes into side branches for me to review.”
-
Naval的角色:final gate——ship it or don’t
“All I have to do is just review the fixes and say, ‘Ah, that wasn’t really a bug. That wasn’t a good fix. Don’t ship that.’ ‘Oh, that looks good. Makes sense. Ship it.’”
9.2 用户驱动的产品开发
-
未来模式:用户提出功能需求 → 对功能进行投票 → 云端意见领袖/维护者决定implement哪些
“The users will ask for features, they’ll vote on features, and then there’ll be some tastemaker or maintainer in the cloud who’ll look at that and say, ‘No, the users don’t know what they want.’ Or, ‘Oh, that makes a lot of sense. We should fix that or change that.’”
-
软件开发已成为与用户协作的过程,代理负责处理一切
9.3 一人软件公司愿景
-
Coding agents做完美的客户服务: -
核心预测:
“You truly can have one-person, two-person software companies now that can scale to millions upon millions of users and make billions upon billions of dollars.”
-
历史先例:这将从rare exception变成pattern
💡 思考点:如果一人软件公司真的能扩展到数百万用户和数十亿美元收入,”公司”的定义本身在崩塌。当创始人不需要hire anyone、不需要manage anyone、不需要raise capital时,venture capital的存在意义是什么?Naval的AngelList是否正在成为自己预言的受益者——或受害者?
🔗 关联点:这与Naval自己说的”pure software is uninvestable”形成内部一致性——如果一人公司不需要VC,那VC必须转向hardware/network effects/models。AngelList作为连接founders和capital的平台,可能正在见证自己的核心市场萎缩。
核心观点总结
关键数据
-
0.1%→1-3%:能构建应用的人口比例提升10-30倍 -
8-9名工程师→1人:Airchat重建的人力对比 -
5分钟:从描述idea到展示可交互应用的时间 -
24小时:Claude自动review和修复bug的周期 -
1M tokens:context window的当前上限 -
~30秒:从prompt到应用推送到iPhone的时间 -
4.5:Claude Opus版本号(拐点版本)
核心判断
-
Claude Opus 4.5是coding agents的拐点——从toy到tool的质变 -
个人App Store bypass App Store——Apple 30%抽成和审查被架空 -
Vibe coding是比电子游戏更有趣的productive entertainment——unbounded + real-world rewards -
纯软件不可投资——硬件、网络效应、AI模型才是VC标的 -
AI eager-to-please是feature也是bug——像猎狗,执行力强但方向依赖人类判断 -
Apple放弃AI是本十年最大战略失误——iPhone沦为screen+battery+connection -
一人软件公司可扩展至亿级用户——Notch/Satoshi模式从exception变pattern -
Top engineers的taste feedback into training data——编码模型飞轮的加速器
关键方法论
-
Personal App Store:webpage + one-click iPhone deploy,绕过Apple审查 -
Vibe Coding Workflow:describe to Claude → generate app → review → deploy -
Bug Fix Automation:user report → log upload → Claude 24h review → side branch → human gate -
模型选择策略:Claude for visual/ChatGPT for reliable/Gemini for data/Grok for unfiltered -
Context Window管理:codebase超过1M tokens时,agent开始architectural mistakes -
一人公司规模化公式:coding agent × product taste × distribution = millions of users
分析时间:2026-05-13
分析人员:有一只肥罗
夜雨聆风