七月十一日

Peter Yang, who publishes practical AI tutorials for busy people, shared the concrete stack he used to kick off a new weekend project: use Fable to build a plan.html with design guidelines, use Claude Design to generate the components and screens, then hand off to GPT-5.6 to write the code. Alongside the build log, he raised a sharp question about the current rush of AI "taste" benchmarks, asking how anyone can actually benchmark something as subjective as design taste.

Peter Yang 专注于为忙碌的人做实用的 AI 教程。他分享了自己周末启动新项目时用的完整流程：先用 Fable 生成带设计规范的 plan.html，再用 Claude Design 做出组件和页面，最后交给 GPT-5.6 写代码。除了这份"搭建日志"，他还对当下扎堆出现的 AI"审美"benchmark 提出了一个犀利的问题：像设计品味这么主观的东西，到底要怎么做 benchmark？

Box CEO Aaron Levie argues the real story of enterprise AI is not getting employees to adopt yet another tool, but re-architecting the workflow underneath them. As organizations move from chat tools to agents, those agents have to be deployed against business processes that usually span multiple functions, which takes far more upfront work but is where the meaningful ROI actually comes from. He quotes a line that captures it: "Software asks the employee to adopt a tool, but infrastructure changes the operating layer underneath the employee... the value should not depend on someone remembering to use the AI every day." His prediction: the winners will be platforms with deep domain expertise, backed by FDE support, change management, well-organized data, and comprehensive evals for each workflow.

Box CEO Aaron Levie 认为，企业级 AI 的真正关键不是让员工再多用一个工具，而是重构员工背后的整个 workflow。当组织从 chat 工具走向 agent，这些 agent 必须被部署到通常横跨多个部门的业务流程上，前期投入要大得多，但真正可观的 ROI 也正来自这里。他引用了一句很到位的话："软件是让员工去采用一个工具，而基础设施改变的是员工脚下的整个运转层……价值不应该取决于某个人每天记得去用 AI。"他的判断是：最终的赢家会是那些具备深厚行业 know-how 的平台，同时配上 FDE 支持、变革管理、把数据整理好，以及为每条 workflow 建立完善的 evals。

Y Combinator President and CEO Garry Tan remixed YC's founding motto for the agent era with a single line: "Make something agents want." It is a compact reframing that positions agents, not just human end users, as the customers builders should now be designing and selling to.

Y Combinator 总裁兼 CEO Garry Tan 把 YC 那句经典口号改写成了 agent 时代的版本，只有一句话："Make something agents want（做出 agent 想要的东西）。"这是一个非常凝练的视角转换：把 agent，而不只是人类终端用户，当成 builder 现在就该去设计和售卖的对象。

Builder Zara Zhang observes that the line between "builder" and "creator" is blurring, and calls this the best moment yet for builders to start making content and for creators to start building products.

Builder Zara Zhang 观察到，"builder（做产品的人）"和"creator（做内容的人）"之间的界线正在变得模糊。她认为，对 builder 来说这是开始做内容的最好时机，对 creator 来说则是开始做产品的最好时机。

FPV Ventures partner Nikunj Kothari reported that Fable and Sol produced a fully working iOS app in just a few hours, built directly off his existing backend and web app. He framed it as a small but vivid example of how quickly AI tooling is collapsing the gap between shipping on the web and shipping native.

FPV Ventures 合伙人 Nikunj Kothari 分享说，Fable 加上 Sol 在短短几个小时内，就基于他已有的后端和 web app 做出了一个完全可用的 iOS app。他把这当成一个虽小但很生动的例子，说明 AI 工具正在多快地抹平"发 web 版"和"发原生 app"之间的差距。

South Park Commons General Partner and former Dropbox CTO Aditya Agarwal sketched where AI development is clearly heading, then asked why it isn't already here: run all your agents in the cloud, freely choose any model (frontier, open-source, Chinese, or American), pick any harness, get full tracing, and run recursive improvement loops. His frustration is that every piece feels inevitable, yet the integrated experience still doesn't exist.

South Park Commons 普通合伙人、前 Dropbox CTO Aditya Agarwal 勾勒出 AI 开发明显要走向的样子，然后反问为什么现在还没到位：把所有 agent 都跑在云上、可以自由选任意 model（frontier、开源、中国的、美国的）、可以选任意 harness、有完整的 tracing、还能跑 recursive improvement loop。他的"槽点"在于：每一块看起来都是必然会有的，但把它们整合到一起的体验却仍然不存在。

OpenAI CEO Sam Altman offered a notable admission: so far, he's "pretty sure AI has been net job-creating," and says this was not what he expected. He was less pessimistic than many, but still assumed that by this level of capability there would be some visible labor-market impact, and he leaves open that the trend could keep going. Separately, he highlighted a striking clinical result, quoting that "physicians found fewer flaws in GPT-5.6 responses than physician-written responses."

OpenAI CEO Sam Altman 给出了一个值得注意的"承认"：到目前为止，他"相当确定 AI 一直是净创造就业的"，并说这不是他原本预期的结果。他本来就没有很多人那么悲观，但仍然以为到了现在这个能力水平，就业市场会出现一些可见的冲击；他也留了个口子，说这个趋势有可能继续下去。另外，他还点出一个很惊人的临床结果，引用道："医生在 GPT-5.6 的回答里找到的问题，比在医生自己写的回答里找到的还要少。"

The Takeaway: The infrastructure for AI agents to buy and sell is largely built; the two things still holding it back are human trust and a brand-new form of fraud where thieves steal tokens instead of money.

Emily Sands, head of data and AI at Stripe, has a rare vantage point: Stripe now touches close to 2% of global GDP and a very large share of all AI buyers and sellers. She frames "agentic commerce" as a spectrum. At one end, fully autonomous agents discover a service and pay for it with no human in the loop, powered by Stripe's machine payments protocol. At the other, a person shops inside an AI surface like Gemini or ChatGPT and simply taps a buy button. The plumbing underneath is new either way: the Agentic Commerce Protocol (co-built with OpenAI, essentially "MCP for commerce"), a shared payment token that lets an agent pay without ever seeing the card and encodes exactly which merchants, amounts, and currencies are allowed, and Link as a guardrailed wallet for agents.

The most refreshingly specific insight: coding is no longer the bottleneck, deployment is. Agent traffic to Stripe's docs grew more than 10x in a year and is now about 40% of all docs traffic, and 70% of Stripe CLI requests now come from agents rather than people. Sands calls the new constraint "vibe deployment," and says Stripe Projects exists to orchestrate the signup, config, and integration work that used to be manual. She's equally blunt that per-seat SaaS pricing is breaking, because every inference call carries a real marginal cost, pushing companies like Lovable and Eleven Labs toward hybrid, usage-based billing and, for agents running at machine speed, real-time metering.

Her most under-discussed warning is token theft: "Fraudsters have figured out that in AI, you actually don't really need to steal money or credentials. You can just steal tokens." More than one in six signups at AI companies are multi-account abuse, free-trial abuse has more than doubled on Stripe in six months, and usage-billing "dine and dash" leaves the AI company eating the compute cost.

核心要点： 让 AI agent 买卖东西的基础设施基本已经搭好了；现在真正卡住它的是两件事：人类的信任，以及一种全新的欺诈方式，小偷偷的不再是钱，而是 token。

Emily Sands 是 Stripe 负责 data 和 AI 的高管，她的视角很稀缺：Stripe 如今触及接近全球 2% 的 GDP，也覆盖了绝大部分 AI 买家和卖家。她把"agentic commerce"看成一条光谱。一端是完全自主的 agent，自己发现某个服务、自己完成付款，全程没有人参与，背后是 Stripe 的 machine payments protocol；另一端则是人在 Gemini 或 ChatGPT 这类 AI 界面里购物，只是点一下"购买"按钮。不管在光谱的哪一端，底层的管道都是全新的：Agentic Commerce Protocol（和 OpenAI 一起做的，本质上就是"给 commerce 用的 MCP"）、一种 shared payment token（让 agent 在完全看不到银行卡的情况下付款，并且明确编码了允许哪些商家、哪些金额、哪种货币），以及作为 agent 专用、带护栏钱包的 Link。

最"具体到让人耳目一新"的一点是：写代码已经不再是瓶颈，部署才是。一年之内，访问 Stripe 文档的 agent 流量涨了超过 10 倍，如今占到全部文档流量的约 40%；Stripe CLI 的请求里有 70% 来自 agent，而不是人。Sands 把这个新瓶颈叫做"vibe deployment"，并说 Stripe Projects 就是为了把过去要手动做的注册、配置、集成这些活儿编排起来。她同样直白地指出，按席位（per-seat）收费的 SaaS 模式正在崩掉，因为每一次 inference 调用都有实实在在的边际成本，这把 Lovable、Eleven Labs 这样的公司推向了混合式、按用量（usage-based）计费；而面对以机器速度运转的 agent，则需要实时计量。

她最被低估的警告是 token theft（偷 token）："欺诈者已经发现，在 AI 里你其实根本不需要去偷钱或偷凭证，直接偷 token 就行了。"在 AI 公司的注册里，超过六分之一是多账号滥用；免费试用滥用在半年内于 Stripe 上翻了一倍多；而按用量计费下的"吃霸王餐"，最后让 AI 公司自己承担了那份算力成本。