主题综述

AI Native 产品到底长什么样 · AI-Native Products

主题综述

更新日志

主流共识

第一点:"GPT wrapper" 这个嘲讽不成立——AI-native 公司确实在跑出新的 ROI 曲线。(注:这条共识 2026-07 被 Accel 自己的复盘复杂化了,见"分歧·应用层可防御性"。)

"We've kind of come to the opinion that there is no AI. There's like a bunch of subspaces that are totally different that all require their own strategy. … GPT wrapper was this derogatory term. I think we've kind of come to the conclusion that's not even a thing."
「我们的结论是:不存在 'AI'——而是一堆完全不同的子空间,每个都需要各自的策略。……'GPT wrapper' 本是个贬义词,而我们得出的结论是:它根本不是一个东西。」
"The AI native companies are far outpacing their SaaS counterparts. And you can see it in terms of new companies blowing past this golden metric of time to a hundred million of ARR … you're seeing this 10x plus improvement in the customer experience as well, whereas SaaS 2.0, you generally saw a little bit more of an incremental improvement, call it 25, 50%."
「AI-native 公司正在大幅跑赢它们的 SaaS 同行——你能从新公司冲过'1 亿美元 ARR 用时'这个黄金指标上看到。……客户体验也出现 10 倍以上的提升,而 SaaS 2.0 通常只是略微渐进的改进,比如 25%、50%。」
Sarah Wang · The State of AI

第二点:AI-native 公司的优势之一是"没有 legacy 系统"——这是相对于 incumbents 的结构性优势

Jesse Zhang(Decagon)的观察是:AI-native 相对 incumbents 的结构性优势,在于没有 legacy 系统的包袱、因而能更敏捷地把方案重做。(podwise 仅存该点的三手转述,故此处转述、不作逐字引用。)

第三点:真正赢的 AI-native 案例大都解决了一个 incumbents 内部尝试失败的具体痛点

"The success rate of the ones where the enterprise went with an outside vendor like a Greenlight or a Tactile was much higher than the success rate of when they tried to build stuff themselves."
「企业用外部供应商(如 Greenlight、Tactile)的成功率,远高于自己内部建设。」
Jared Friedman / The Lightcone · Inside The MIT AI Study

第四点(2026-07 新增):"编码不再是瓶颈,验证/评审才是"——这条在两家 AI-pilled 组织里被独立说出。Anthropic 的 Fiona Fung 与 Sierra 的 Clay Bavor 用几乎相同的措辞描述了瓶颈的位移。

"when not only more people checking in code, but different disciplines, but also the throughput is so high, how do we think about verification? That's this other shift that I'm seeing."
「不仅更多的人在提交代码,还有不同的学科,而且产出率如此高——我们如何考虑验证?这是我看到的另一个转变。」
"what is the constraining factor? It used to be writing code. Now it's probably reviewing code. Pretty soon it will be deciding what is worth building …"
「约束因素是什么?以前是写代码,现在大概是评审代码,很快就会变成'决定什么值得造'。」

分歧在哪

模式 A · "捡没人想做的脏活"——Reducto 路径

Reducto(Adit Abraham 与联合创始人 Ronak)给的是 AI-native 案例里最反直觉、也最具体的范本。下面两条出自 Ronak(YC 访谈中 Diana Hu 以 "Ronak" 点名后由他作答的那位):

"None of these like AI application layer companies like want to be PDF processors. It's just not something that's like exciting to them."
「那些人工智能应用层公司都不想成为 PDF 处理器。这对他们来说没什么吸引力。」
"we made it a more interesting problem because we took a different approach than what a lot of other folks had been doing in the sense that We turned PDF processing, which is usually just writing a bunch of rules for how to process this type of file type and this other type of file type into a computer vision problem. So the insight we had is like, we are going to parse and understand these documents the way humans do."
「我们让问题变得更有趣,因为我们采取的方法与许多其他人不同,我们将 PDF 处理(通常只是为如何处理这种类型的文件和其他类型的文件编写一堆规则)变成了一个计算机视觉问题。我们的想法是,我们将像人类一样解析和理解这些文档。」

具体经济数据(来自 Reducto 同期访谈):录播客前五个月增长 ;处理 10 亿+ 页、却只烧了约 100 万美元资本;约 40% 客户用到 2 个及以上 API 端点。

"Every mistake that you make at ingestion is going to compound throughout your pipeline."
「每一个在 ingestion 阶段犯的错,都会在你整个 pipeline 上复利。」

Adit 自己也承认 Reducto 不是他们最初想做的

"I think if you asked us two years ago, would you be excited to work on PDF processing? The gut reaction would be no. It doesn't sound like a fun problem."
「如果两年前问我们,做 PDF 处理你兴奋吗?直觉反应是不。这听起来不是个好玩的问题。」

模式 B · "替代人工劳动而非增强软件"——Decagon / Sierra 路径

Jesse Zhang (Decagon) 把 AI-native 押到了预算来源这件事上。「从软件预算转向人工劳动预算」这个提法出自主持人 Harry Stebbings 的提问,Jesse 的回答是——这已经在发生:

"I would say that's already happening in our space. … when you're in the application layer, if you're building enough of a product, you're more benchmark against what is the business problem you're solving. And the business problem you're solving is going to be way bigger, right? … You're saving human labor. You're transforming the way that their customer experience works. … It's, in my opinion, one of the few markets right now that has true PMF with AI."
「我认为这已经在我们这个领域发生了。……当你在应用层时,如果你构建了足够多的产品,你就会更多地以你正在解决的业务问题为基准。你正在解决的业务问题会更大,对吧?……你正在节省人力。你正在改变他们的客户体验运作方式。……在我看来,这是目前为数不多的、拥有真正 PMF 与 AI 的市场之一。」

2026-07 新增:Decagon 的正面对手 Sierra 第一次上桌。 联合创始人 Clay Bavor(前 Google 18 年)给出的不是"替代 vs 增强"的口号,而是一条 AI-native 产品的技术栈判据——往下打多深就够了。Sierra 自建 agent 框架(首任研究负责人是写出 ReAct 论文的 Princeton 教授),但明确不做预训练、只在开放权重模型上微调:

"number one is a willingness to invest as far down the technology stack as you need in order to build the service and product that you want. … I think it's important that you are in control of your own destiny enough and that you don't tell yourself a story that you need to go further than you actually need to do."
「第一条是:愿意为了做出你想要的服务和产品,往技术栈下面打到需要的那一层为止。……重要的是你要足够掌握自己的命运,同时不要给自己编一个'我必须走得比实际需要更远'的故事。」

而"替代 labor"在 Sierra 的语境里是按领域分级的——不是所有场景都要顶配智能:

"You don't need Mythos to return a pair of shoes, right? … we've got some capability overhang, so to speak, for doing something like that. But in a range of domains, Coding, certainly, science, material science, legal, right where the stakes are very high … we're going to see effectively unbounded demand for greater levels of intelligence."
「退一双鞋不需要动用顶级模型,对吧?……做这种事我们其实是'能力过剩'的。但在另一批领域——编码、科学、材料科学、法律,这些风险极高的地方——我们会看到对更高智能近乎无上限的需求。」

Bavor 还给了一个成本结构的重定义——未来的资本分配里,token 预算会和薪资、股权并列成为"人头成本"的一部分:

"capital allocation will look more like how do we allocate OPEX and then headcount. And headcount will be both headcount for salaries and SBC and also tokens associated with headcount. Here's your salary. Here's your token budget. Have at it."
「资本分配会更像是:怎么分配 OPEX,然后是人头。而人头既包括薪资和股权,也包括这个人头对应的 token。这是你的工资,这是你的 token 预算,去用吧。」

这跟姊妹主题 saas-postmortem 里 Bret Taylor 的 "every company needs an agent in 2027" 是同一条逻辑——但 Decagon 跟 Sierra 在同一市场正面竞争,Bavor 自估这会长成 Uber/Lyft 式的双寡头、而非 AWS/Azure 式的多云并存。

模式 C · "野花园 / 专有数据复利"——AI Opportunity panel → DoorDash 实证

这个 panel 的三条主线(转述):(1) 传统软件向 AI-native 转型;(2) 平台超越 SaaS、把"劳动力预算"纳入;(3) "wild garden"业务靠专有数据形成复利优势。其中第三条他们给了一个很到位的说法:

"It's a blueprint for how to potentially deal with a world where the source of the raw material is actually what is rare."
「这是一个蓝图,说明在'原材料本身才是稀缺'的世界里该怎么办。」

Rampell 同时给了一个Camp B 的反话

"I know it sounds pithy to say software is eating labor, but really software is augmenting labor."
「我知道'软件正在吞噬劳动力'听起来很精辟——但其实软件是在增强劳动力。」

2026-07 新增:DoorDash 把"专有数据复利"从蓝图变成实证。 上一版"还想知道"里问过——有没有 incumbents 真正做成了 AI-native 重做的反例?DoorDash 就是。Stanley Tang 把护城河直接命名为数据:

"I think we just have such a huge advantage over everyone else because we have something that everyone else doesn't have. It's called DoorDash. We have 10 billion deliveries of data to extract from."
「我们对所有人都有巨大优势,因为我们有别人没有的东西——它叫 DoorDash。我们有 100 亿次配送的数据可以挖。」

而 DoorDash 的方法论恰恰是模式 C 与模式 A 的交点——从 use case 倒推,而不是先造技术再找问题,Stanley 借此直接批评了这轮 AI/机器人创业的主流做法:

"in the software world, when we went through YC, we're always taught to … build something people want. … But then when it comes to hardware and hard tech and AI and robotics, it's people just kind of do the opposite where they try to build a tech first, and not really think about the use case they're building towards."
「在软件世界,我们读 YC 时被反复教导要造人们想要的东西。……可一到硬件、硬科技、AI 和机器人,人们却反着来——先把技术造出来,根本没想清楚要服务的 use case。」

产品侧的 AskDoorDash(自然语言点单)也给了具体行为数据:餐厅侧 50% 的 AI 会话是从"没点过的店"下单(DoorDash 历史上最难撬动的指标),杂货侧客单 大 40%。Andy Fang 顺手抛出一条会长期发酵的观察——下一个 DoorDash 会是 agentic-first

"there's more agent traffic on the web than human traffic … if someone were to create DoorDash today … I think it would look very different, probably more agentic first."
「网上的 agent 流量已经超过人类流量……如果今天有人从头做一个 DoorDash,样子会非常不一样,大概率是 agentic-first 的。」

模式 D · "AI 作为网络效应的复利器"——Chris Dixon / Anish Acharya 路径

Chris Dixon 不把 AI 当作产品本身,而当作让网络效应更可持续的工具:

"I've always suspected in tech, in Silicon Valley, kind of we underestimate the power of just kind of brands and consumer inertia. And I think you're sort of seeing that today with ChatGPT of just like such a household name, like overnight almost, that even though it doesn't have in this sort of technical sense, maybe network effects, memory and things, but I mean, it's not, that's more stickiness network effects."
「我一直怀疑,在科技圈、在硅谷,我们低估了'品牌'和'消费者惯性'的力量。今天你能在 ChatGPT 身上看到——它几乎一夜之间家喻户晓,尽管在技术意义上它可能并没有网络效应、记忆这些东西——那更多是一种黏性式的网络效应。」

同场对谈的 Anish Acharya(a16z 消费合伙人)把这条线推到消费端付费与产品野心上:

"I mean, one of our, our sort of extreme views here is that the future of consumer disposable income will be like food, rent, software. And software is going to subsume a lot of the other areas of discretionary spend today."
「我的意思是,我们这里的一个极端观点是,未来消费者的可支配收入将用于食物、租金和软件。软件将取代今天许多其他可自由支配的支出领域。」
"maybe a controversial statement right now is that there are no marketing problems, only product problems, because the technology allows you to be so ambitious on behalf of your customer"
「现在可能存在争议的说法是,不存在营销问题,只有产品问题,因为技术允许你为了你的客户而变得雄心勃勃。」

模式 E · "内嵌于社交结构 + 熟人分发"——Wanaka 路径

张阳 (Wanaka) 给的是一个英文 AI-native 叙事里完全缺席的模式——AI 不是产品本身,是降低创作门槛的工具,但产品的真正动能来自社交关系结构

「天然有一些内容它就只能用社交关系去分发,就是 UGC 的内容。……比如说我的一个随拍,我今天可能在楼下拍一个路灯,或者拍了一只猫,这些东西对整个互联网里的人来说其实没有任何价值。但是他对我的朋友,因为这个身份的关系在,所以他就觉得这事情好像挺好玩,我要点个赞。」
— 张阳 · AI + 游戏 + 社交的新演绎(逐字稿仅中文)

「消费的是关系、不是内容」这句收拢,其实出自主持人曲凯,张阳当场认领并把它落到 AIGC 上:

曲凯:「对,因为他消费的其实是这个关系,消费的不是那个内容本身。」
张阳:「对,所以我觉得我们在看 AIGC 内容我觉得也是一样……但是大量的 AIGC 出来的这种普通人的内容,它其实只有一个去向,就是给你的朋友看。……它只对你的朋友价值。」
AI + 游戏 + 社交的新演绎(逐字稿仅中文)

创作工具决定上限这一点,张阳的原话是:

「这个事情还是回到创作工具的能力本身。就是你的创作工具如果能够支撑那些创作者做出好的东西,其实我觉得是有可能的。但是我们看现在这种上下滑的这种模态,我觉得它上下滑这种模态倒没什么,主要是它生产的能力太弱了,就你直接去套用大模型的指出的能力,这种方式是有问题的。」
— 张阳 · AI + 游戏 + 社交的新演绎(逐字稿仅中文)

这条线跟模式 D 的 Chris Dixon "AI as compounding force for networks" 在逻辑上接近,但张阳 的 implementation 是分发结构(熟人圈)而非产品功能——这是 Chris Dixon 没明说的版本。

模式 F · "startup-shaped holes / 实证派"——Lightcone / Garry Tan

Garry Tan / Lightcone 不讲哲学,讲市场结构:

"There's this startup shaped hole in basically every process or every sort of annoying system that should exist that doesn't exist yet."
「基本上每一个流程、每一个'应该存在但还不存在的烦人系统'里,都有一个 startup 形状的洞。」
Jared Friedman / The Lightcone · Inside The MIT AI Study
"The knock on effect for startups then is if you can actually build something that works, the enterprises will talk to you because they have no other options."
「这对 startup 的连锁效应是——如果你真能造出能跑的东西,企业会跟你谈,因为他们没别的选择。」
The Lightcone / Garry Tan · Inside The MIT AI Study

这条线跟模式 A(Reducto)天然契合——"没人想做但企业刚需"的事就是 startup-shaped hole 的具体例子。

分歧新增 · 应用层到底防不防得住(Casado vs Accel)

主流共识第一点说"GPT wrapper 不是 thing",但 Accel 增长团队 2026-07 的自我复盘正面顶了回去——他们承认自己在 AI 应用层"比同行慢",因为担心模型能力越铺越宽时,应用层公司太善变、护城河守不住,于是主动退回基础设施

"we were admittedly slower than our peers on the application layer for AI because we were very worried about how fickle they were when the weights of the models were getting wider and wider and we spent a lot of time in infrastructure …"
「坦白说,我们在 AI 应用层比同行慢,因为我们非常担心——当模型的能力边界越铺越宽时,这些应用层公司有多善变。所以我们花了大量时间在基础设施上……」
Accel 增长团队(Speaker 3)· Accel: The Quiet Firm Behind Facebook, Cursor, Nebius

Factory 的 Matan Grinberg 从应用层内部给出了自己的护城河答案——不是"贴着某个模型",而是"对所有模型保持独立",因为企业最怕单点绑定:

"The thing that enterprises are really caring about … they do not want anyone to kind of be their single point of failure. … something that really matters is model independence."
「企业真正在乎的是……他们不想让任何人成为自己的单点故障。……真正要紧的是模型独立性。」

Matan 甚至给了一个反直觉的技术论断,直接反驳"模型与 harness 协同设计更强"的实验室观点:

"much to the chagrin of many of my friends at OpenAI and Anthropic, this is not true. If you build a harness that supports different models, that harness will be better. … what data is to a model, models are to a harness."
「让我在 OpenAI 和 Anthropic 的很多朋友很懊恼——这(协同设计更强)不成立。如果你的 harness 支持多个模型,它会更强。……数据之于模型,就等于模型之于 harness。」

把三方并排:Casado 说 wrapper 不成立;Accel 用钱投票、退回 infra;Factory 说防御力来自"模型无关"而非"贴模型"。同一件事(应用层可防御性)三个截然不同的下注方向。

分歧新增 · 企业先行 vs 消费原生(Kevin Weil)

模式 D/E 押的是消费端 AI-native 产品,但 Kevin Weil(前 OpenAI CPO)指出这一轮的形态与以往互联网周期相反——是企业先行,消费端的"那个 eBay"还没出现

"Enterprise like B2B stands out because that is where we do the majority of our economically valuable work and models are getting increasingly good at doing economically valuable work. … models also cost money to use. … you start having costs right away as a business in a way that maybe you didn't have as much if you were like building a consumer social thing before."
「B2B 之所以突出,是因为绝大多数有经济价值的工作发生在那里,而模型正越来越擅长做有经济价值的工作。……用模型是要花钱的。……你一开张就有成本,而以前做一个消费社交产品可能没这么大成本。」

但 Weil 同时给了消费端一条全新的分发路径——绕过网站和 App,直接长在模型平台上:

"you can build a business in the future using this apps platform that doesn't have a website or a mobile app and is sort of entirely built around These kind of new platforms."
「未来你可以用这个 apps 平台造一门生意——它没有网站、也没有 App,完全围绕这类新平台构建。」

分歧新增 · 分发结构 = agentic influence / choke point(Accel)

Chris Dixon 讲"品牌黏性"、张阳讲"熟人分发"、Kevin Weil 讲"apps 平台",Accel 给了第四种、也是最具操作性的一种分发结构——agent 本身成了分发渠道,"AI optimization"取代 SEO

"I think in the last era, it was search engine optimization. In this era, it's AI optimization. These products are now the distribution mechanism for all the downstream products and services that you can use. And they tend to prioritize the tools that have the best developer and user experience."
「上一个时代是搜索引擎优化,这个时代是 AI 优化。这些产品(Claude、Codex、Cursor 这类)现在成了所有下游产品与服务的分发机制,而它们倾向于优先推荐开发者/用户体验最好的工具。」
Accel 增长团队(Speaker 1)· Accel: The Quiet Firm Behind Facebook, Cursor, Nebius
"all of a sudden, these Agents are making decisions about downstream workflows and downstream tool creation. And we just sort of said, we have to invest in every single company … that is like the choke point that's metering out these decisions."
「突然之间,这些 agent 开始替下游工作流和下游工具做决策。我们于是说:必须去投每一家处在这条流里的公司——尤其是那个'掐着这些决策口子'的 choke point。」
Accel 增长团队(Speaker 2)· Accel: The Quiet Firm Behind Facebook, Cursor, Nebius

(实证:Supabase 从投资时的不到 100 万开发者涨到近期 900 万,Accel 归因于"AI 最容易集成它";60% 的 YC 公司选它。)

暗流 · 速度

Philip Clark (Thrive Capital) 给了一个跨所有模式的共同观察——注意"象棋 vs 快棋"这个比喻是他转述 Michael Dell 的原话:

"I think one of the really wild and exciting things about the AI era is that the timelines are getting compressed in a really radical way. I had the privilege of speaking to Michael Dell actually a few weeks ago. … And his line to me was, in some ways feel very similar in magnitude of impact. But whereas the internet era was like playing chess, the AI era is like playing speed chess."
「我认为关于人工智能时代真正疯狂和令人兴奋的事情之一是,时间线正在以一种非常激进的方式被压缩。实际上几周前我有幸与 Michael Dell 交谈过。……他对我说,在某种程度上,感觉在影响的程度上非常相似。但互联网时代就像下象棋,人工智能时代就像下快棋。」
Philip Clark(转述 Michael Dell)· Inside Thrive Capital

2026-07 新增:Factory 给这条暗流加了一个反向注脚——"早"本身没有奖励。Matan 用两年"沙漠期"证明,真正的约束不是模型能力、而是买方(工程师/采购)的行为惯性:

"being two or three years early is the same as being wrong. … There's no consolation prize. It's either you do the thing or you don't do the thing."
「早两三年,跟做错了是一回事。……没有安慰奖。要么你做成,要么没做成,就这样。」

无论选哪种 AI-native 模式,决策窗口都比 SaaS 时代窄得多——但 Factory 提醒:窗口错在早的一侧,代价一样是致命。

都没说透的

我的看法

判断(不是事实):这次 6 篇最有价值的贡献,是把上一版"模式 A+F 组合最稳"的结论从孤例推向可复用——DoorDash 证明"从 use case 倒推、拿独家数据做复利"不是 Reducto 的偶然,而 Matan 的"早=错"和 Clay Bavor 的"别把栈打得比需要更深"共同指向同一条纪律:AI-native 的胜负手在'节奏与栈深的克制',不在'模型有多前沿'。我现在更相信,防御力这件事上 Accel 的悲观比 Casado 的乐观更值得当回事——"GPT wrapper 不是 thing"更像是 2025 的口号,而 2026 年真正在下注的钱(Accel 退回 infra、Factory 押模型无关、Sierra 只微调不预训练)都在用行动承认:纯应用层薄壳确实会被模型能力扩张吞掉,活下来的要么握数据(DoorDash)、要么握分发 choke point(Cursor)、要么握 model-agnostic 的编排层(Factory)。 最被高估的仍是"替代 labor"的 narrative——四个当事人连边界都划不出来,说明它离可验证的商业模式还有距离。

把握程度:中等偏上(较上一版升一档)。最强支撑是 DoorDash 与 Sierra 两个当事人一手数据把模式 C、B 坐实;最弱环节是"应用层会被吞掉"——Accel 是用"我们没投"来间接背书,属于缺席证据,不是正面证据。

还想知道什么

取材