七月二十日

Box CEO Aaron Levie argued that gatekeeping frontier model access has become self-defeating: with strong open-source alternatives sitting just behind the frontier, even a full American ban on open models wouldn't stop other ecosystems from using them — against us, or simply to defend themselves better — while our own companies fall behind, so the calculus of AI regulation simply must change. He also took on a common economic fallacy: when AI costs drop, spend usually goes up, not down, because cheaper AI gets consumed for a far wider set of tasks — you write more code, review more of it for bugs and security, run agents on datasets you couldn't afford to process before — so for the foreseeable future anything that lowers token costs drives up inference demand. And he flagged a good post on AI diffusion: progress will ultimately be rate-limited by interaction with the real world, which is why coding was adopted so fast — one person can write, test, and run a near-infinite amount of code end to end without anyone in the outside world doing anything differently — while life sciences still needs years of drug testing.

Box CEO Aaron Levie 指出，对前沿模型搞封锁式管控已经变得自伤：当开源替代品的水平只落后前沿一点点时，即使美国能彻底禁掉开源模型，其他生态也不受这个禁令约束，他们照样可以拿这些模型来对付我们，或者用来更好地防御自己，而我们自己的公司反而会掉队，所以 AI 监管的算盘必须重新打。他还纠正了一个常见的经济学误区：AI 成本下降时，支出通常不降反升，因为更便宜的 AI 会被用在广得多的任务上，你会写更多代码、做更多 bug 和安全审查、让 agent 去跑以前根本处理不起的大数据集，所以在可预见的未来，任何降低 token 成本的东西都会推高推理需求。他还转发了一篇关于 AI 扩散的好文：AI 带来的进步最终会被它与真实世界的交互速度限制，这正是编程被最快采纳的原因，一个人就能端到端地写代码、测代码、跑代码并创造价值，外部世界谁都不用改变任何做法，而生命科学的新药研发依然需要多年的测试。

Vercel CEO Guillermo Rauch made the case that cybersecurity is one of the best benchmarks for superintelligence — "The 'IQ test' of software engineering." It's easy for a model to one-shot an XYZ clone and impress people on X, but that's not a good test: finding, patching, reversing, and exploiting demand a cognitive skill that transcends programming languages, runtimes, and frameworks — true reasoning power and "corner thinking" that very, very few humans excel at. The best engineers he's worked with in his career have usually had a deep background or interest in security.

Vercel CEO Guillermo Rauch 论证了为什么网络安全是衡量 superintelligence 的最佳基准之一，他称之为"软件工程的智商测试"。让模型一把梭出一个 XYZ 克隆、在 X 上惊艳众人其实很容易，但那不是好的测试：漏洞挖掘、修补、逆向和利用需要一种超越编程语言、运行时和框架的认知能力，需要模型真正的推理实力和"钻角思维"，而这恰恰是极少数人类才擅长的。他职业生涯中合作过的最强工程师，通常都有很深的安全背景或兴趣。

Anthropic's Cat Wu shared the exact prompt she uses to have Claude Cowork manage her calendar for the week: keep meetings under 20 hours, dedupe conflicting meetings, look at past weeks to learn which kinds of meetings she declines, don't count dinners against the cap, build a skill that refines itself over time, and ask before updating any invites.

Anthropic 的 Cat Wu 分享了她让 Claude Cowork 管理每周日程的完整 prompt：会议控制在 20 小时以内，去掉冲突的重复会议，翻看过去几周学习她通常会拒绝哪类会议，晚餐不计入 20 小时上限，构建一个能持续自我完善的 skill，并且在更新任何邀请之前先征求她的同意。

Replit CEO Amjad Masad offered a crisp explanation for why huge consumer subscription businesses are so rare: consumers spend their money on food, rent, entertainment, phone and internet, and shopping — software is typically bought by their companies. That's why, outside of Netflix and Spotify, it's hard to think of truly massive consumer subscription businesses.

Replit CEO Amjad Masad 给出了一个简洁的解释，说明为什么超大规模的消费者订阅生意如此罕见：消费者的钱花在食物、房租、娱乐、话费网费和购物上，软件通常是由他们的公司来买单的。这就是为什么除了 Netflix 和 Spotify 等少数几家，你很难想出真正巨大的消费者订阅业务。

Y Combinator CEO Garry Tan declared that Markdown files are universal and will survive millennia — unsurprisingly, they're a fantastic data format when the intelligence stack is in flux at this extreme pace.

Y Combinator CEO Garry Tan 宣称 Markdown 文件是通用格式，能流传千年。在智能技术栈以如此极端的速度动荡演进的当下，它毫不意外地成为了绝佳的数据格式。

Every CEO Dan Shipper reported that AI just passed some barrier: for the past week it has been able to automatically handle about 70% of the copy edits his team at Every would ordinarily do by hand — the first time that's happened, and he's been trying for years.

Every CEO Dan Shipper 报告说 AI 刚刚跨过了某道门槛：过去一周，它已经能自动完成 Every 团队通常需要手工处理的约 70% 的文字校对工作。这是多年来的头一次，而他为此已经尝试了很多年。

Builder Zara Zhang argued we need to get used to the idea that code and software can now be disposable, with examples all discarded after use: design playgrounds and modals you spin up just to fine-tune the look and feel of a design, HTML pages generated to help you understand a codebase, and throwaway dashboards created to inspect something once. She also shared her go-to advice for people starting to post content: take the top three questions friends and colleagues ask you most frequently and turn those answers into videos or posts, saying it exactly like you said it in real life — that's your positioning. Anything you've repeated more than three times is worth turning into a video; she started posting partly because she was tired of repeating herself.

Builder Zara Zhang 认为我们需要适应一个新观念：代码和软件现在可以是一次性的。她举的例子全都用完即弃：为了微调设计观感而临时搭建的 design playground 和弹窗、为了帮自己理解代码而生成的 HTML 页面、为了检查某个东西而临时拉起的一次性 dashboard。她还分享了给内容创作新手的标准答案：把朋友、熟人、同事最常问你的三个问题拿出来，把你的回答变成视频或帖子，用你在现实中原本的说法讲出来，这就是你的定位。任何你重复讲过三次以上的东西都值得做成视频；她当初开始发内容，一部分原因就是受够了在现实生活里不断重复自己。

The Takeaway: The payment rails for AI agents to spend real money are already built and live — what's missing now is human trust, and the fraud frontier has shifted from stolen credit cards to stolen tokens.

Emily Glassberg Sands leads data and AI at Stripe, which sees roughly 2% of global GDP and sits under basically every major AI company's billing stack — an unusually complete view of the machine economy. A year ago, agents-as-buyers was hypothetical; today the infrastructure is deployed: the Agentic Commerce Protocol (co-built with OpenAI) lets a business expose its catalog once and sell through every compliant agent, the shared payment token encodes exactly which merchants an agent may charge, up to what amount, in what currency, for how long — without the agent ever seeing card details — and Link, Stripe's 300-million-user consumer wallet, is becoming the wallet for agents, with budgets and per-transaction guardrails. Best Buy, Coach, Shopify, and Wix are on the supply side; Google, Microsoft, OpenAI, and Meta on the agent side. Stripe even keeps a self-driving-style L1-to-L5 autonomy scale for agentic commerce, and she puts consumers today at roughly level two.

Her sharpest warning is token theft, which she calls one of the most under-discussed topics in AI: "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 farming free credits, free-trial abuse has more than doubled on Stripe in six months, and usage-based billing enables a "dine and dash, but for tokens." SaaS could shrug off fraud because marginal cost was near zero; with real inference costs, a high enough fraud rate breaks a product's economics fast.

Meanwhile, vibe coding is solved and vibe deployment is the new bottleneck — agent traffic is now 40% of Stripe's docs traffic and 70% of its CLI's API requests, hence Stripe Projects, which lets agents provision Vercel, Supabase, Cloudflare, and Twilio straight from the command line. The macro tell: US business formation is accelerating entirely on solopreneurs — five million Americans now make their living solo, hundreds of thousands clearing $1M a year. Twelve months out, she expects agents as multifaceted economic actors: "It's not Emily permissioning an agent to buy on her behalf. It's Emily has an agent who's tasked with running a business."

核心要点：让 AI agent 花真金白银的支付轨道已经建好并上线了，现在缺的是人类的信任，而欺诈的前沿也已经从盗刷信用卡转向了偷 token。

Emily Glassberg Sands 在 Stripe 负责数据与 AI。Stripe 经手全球约 2% 的 GDP，几乎所有主要 AI 公司的计费系统都跑在它上面，这让她对机器经济有一个异常完整的视角。一年前，"agent 作为买家"还是纸上谈兵；如今基础设施已经落地：与 OpenAI 共建的 Agentic Commerce Protocol 让商家把商品目录暴露一次，就能通过所有兼容的 agent 卖货；shared payment token 精确编码了 agent 可以向哪些商家扣款、上限多少、什么币种、有效多久，agent 全程看不到卡号；而 Stripe 拥有 3 亿用户的消费者钱包 Link 正在变成 agent 的钱包，自带预算和逐笔交易的护栏。供给侧有 Best Buy、Coach、Shopify、Wix；agent 侧有 Google、Microsoft、OpenAI、Meta。Stripe 内部甚至有一套类似自动驾驶 L1 到 L5 的 agentic commerce 分级，她认为今天的消费者大致处在 level two。

她最尖锐的警告是 token 盗窃，她称之为当下 AI 领域最被低估的话题之一："欺诈者已经想明白了，在 AI 时代你其实不需要偷钱或偷凭证。你直接偷 token 就行。"AI 公司超过六分之一的注册是刷新号薅免费额度的多账号滥用，免费试用滥用在 Stripe 上六个月内翻了不止一倍，而按用量计费还催生了"吃霸王餐，只不过吃的是 token"。SaaS 时代边际成本近乎为零，欺诈无伤大雅；但在真实的推理成本面前，欺诈率一高，产品的经济模型很快就会崩掉。

与此同时，vibe coding 已经被解决，vibe deployment 成了新的瓶颈：agent 流量已占 Stripe 文档流量的 40%、CLI API 请求的 70%，这正是 Stripe Projects 的由来，它让 agent 直接从命令行开通并集成 Vercel、Supabase、Cloudflare、Twilio。宏观层面的信号是：美国新增企业注册的加速完全来自 solopreneur，500 万美国人靠单人公司谋生，其中数十万人年入超过 100 万美元。展望十二个月后，她期待 agent 成为多面手式的经济主体："不是 Emily 授权一个 agent 替她买东西，而是 Emily 有一个 agent，它的任务是经营一门生意。"