七月二十三日

Guillermo Rauch says his favorite metric for AI progress is "WTFs/day" — and it's accelerating. This week Fable found a 15-30% memory efficiency improvement in Turbopack / Next.js "nearly autonomously," three days after Sol helped Vercel find novel vulnerabilities in some of the most audited code in the world; he also mentions internal work shrinking binaries by 10-20x. "It's one thing to read benchmarks or stories online. It's another to watch these machines pull engineering feats every day."

Guillermo Rauch 说他最喜欢的 AI 进展指标是每天的 "WTF 时刻" 数量，而且这个数字还在加速上升。本周 Fable 几乎全自主地在 Turbopack / Next.js 里找到了 15% 到 30% 的内存效率优化，就在三天前，Sol 还帮 Vercel 在一些全球审计最严格的代码里发现了全新的漏洞；他还提到内部有项目把二进制体积缩小了 10 到 20 倍。他说："在网上看 benchmark 和故事是一回事，亲眼看着这些机器每天完成工程壮举是另一回事。"

He also announced a Vercel partnership with Shopify, calling the two companies "two sides of the same coin: entrepreneur-first, developer-led, web-obsessed," with plans to bring Vercel's infrastructure for apps and agents to millions of businesses.

他还宣布了 Vercel 与 Shopify 的合作，称两家公司是 "一枚硬币的两面：创业者优先、开发者主导、痴迷于 Web"，计划把 Vercel 的应用与 agent 基础设施带给数百万商家。

Anthropic released the Claude Security plugin for Claude Code in beta: scan your changes for vulnerabilities before you commit, or run a full scan across your codebase, all from the terminal on the Claude inference you already run.

Anthropic 发布了 Claude Code 的 Claude Security 插件 beta 版：可以在 commit 之前扫描改动中的安全漏洞，也可以对整个代码库做全量扫描，全部在终端里完成，用的就是你已有的 Claude 推理额度。

The Anthropic Economic Index is also now available as a connector: turn it on from the connectors menu and it works in any conversation, with the full datasets still free to download.

Anthropic Economic Index 现在也上线了 connector：在 connectors 菜单里开启后可以在任何对话中使用，完整数据集依然可以免费下载。

Aaron Levie amplified a post from Anthropic's Head of Economics finding that jobs have been less negatively impacted by AI than expected. His read: most jobs can't be fully automated, only certain tasks within them, and automating those tasks often increases output and sustains demand for the job. He quotes the research: "So far, AI is both skill-biased and labor-augmenting. It complements domain expertise. It relies on humans in the loop to direct and evaluate the most complex work." His conclusion: "Jevons paradox is alive and well."

Aaron Levie 转发了 Anthropic 首席经济学家的文章，研究发现 AI 对就业的负面冲击比预期要小。他的解读是：大多数工作无法被完全自动化，只能自动化其中的部分任务，而这些任务被自动化后往往会放大产出、维持甚至提升岗位需求。他引用研究原文："到目前为止，AI 既偏向高技能者，也在增强而非替代劳动力。它是领域专长的补充，最复杂的工作仍需要人类在环中指挥和评估。" 他的结论是：杰文斯悖论依然成立。

Thibault Sottiaux, who leads Codex at OpenAI, posted a teaser that racked up 6,000+ likes overnight: "Unbelievable excited for what's coming together. Tomorrow is feeling codexy." Expect a Codex announcement imminently.

OpenAI Codex 负责人 Thibault Sottiaux 发了一条预告，一夜之间收获 6000 多个赞："难以置信地兴奋，一切正在成形。明天感觉很 codexy。" Codex 大概率马上有新发布。

Amjad Masad fired a shot at model routers: "If you're incentivized to push certain models, your router is merely a facade" — a pointed take on platforms whose routing decisions are shaped by their own economics rather than user outcomes.

Amjad Masad 对模型路由开炮："如果你有推某些特定模型的利益动机，那你的 router 只是个幌子。" 矛头直指那些路由决策被自身经济利益而非用户效果驱动的平台。

He also teased Replit's next move on mobile: "We invented mobile coding — now we're taking it to the next level."

他还预告了 Replit 在移动端的下一步动作："我们发明了移动端编程，现在我们要把它带到下一个层级。"

Thariq, who works on Claude Code at Anthropic, said he regrets not trying /design sooner: working on frontend with Claude Design plus Claude Code together is "actually so good." The post resonated, pulling in nearly 1,500 likes.

在 Anthropic 做 Claude Code 的 Thariq 说，后悔这么晚才试 /design：用 Claude Design 加 Claude Code 一起做前端 "真的太好用了"。这条帖子引发共鸣，收获了近 1500 个赞。

Madhu Guru, formerly of Google's Gemini, Veo, and Nano Banana teams, pushed back on data-security fears around Chinese models: "Using a Chinese-trained LLM doesn't mean they get your data." An LLM is a giant file of numbers; with open-weights models you download that file and run it in your own cloud, at which point the model trainer is out of the picture and your data stays wherever you run it.

曾在 Google 主导 Gemini、Veo 和 Nano Banana 的 Meta AI 高级总监 Madhu Guru 反驳了围绕中国模型的数据安全担忧："用中国训练的 LLM 并不意味着他们能拿到你的数据。" LLM 本质上是一个装满数字的大文件；开放权重模型可以下载后跑在你自己的云环境里，这时训练方已经完全不在链路中，数据留在你运行它的地方。

Peter Yang's /no-ai-slop writing skill hit 1,000 GitHub stars in a single day — "people are sick of slop!"

Peter Yang 的 /no-ai-slop 写作 skill 一天之内在 GitHub 上拿下 1000 个 star，他的评价是："大家真的受够了 AI 味儿的水文！"

He also demoed the skill working in reverse: when a friend complained about slop on LinkedIn, he used it to generate a reply packed with every AI cliché in existence. The skill is free to grab.

他还演示了这个 skill 反向运行的效果：朋友抱怨 LinkedIn 上的水文太多，他就用它生成了一条集齐所有 AI 陈词滥调的回复。这个 skill 可以免费获取。

Zara Zhang shared a prompting philosophy in five words: "Thick context, thin prompt." She often describes her problem without specifying a solution or spec, and the model frequently surprises her with a better solution than she could have thought of.

Zara Zhang 用五个词分享了她的 prompt 哲学："厚上下文，薄指令（Thick context, thin prompt）"。她经常只描述问题本身，不给出方案或规格说明，而模型常常给出比她自己能想到的更好的解法，带来惊喜。

She also noted that one of the most exciting things about starting a company today is that you get to invent not just a new business, but a new type of organization.

她还提到，今天创业最令人兴奋的一点是：你不仅在发明一门新生意，还在发明一种全新的组织形态。

Josh Woodward announced that Flow, Google's AI filmmaking tool, is giving everyone 50 free credits per day through August 31.

Josh Woodward 宣布 Google 的 AI 电影创作工具 Flow 在 8 月 31 日之前每天送所有用户 50 个免费 credit。

Garry Tan published a new YC Request for Startups: "We want to defend America" — an explicit call for defense-focused founders.

Garry Tan 发布了新的 YC Request for Startups："我们想要保卫美国"，明确向国防方向的创业者发出招募信号。

Aditya Agarwal, former Dropbox CTO, argued that culture always eats strategy for breakfast, and that startup culture stems from three sources: the founder's personality and DNA, the early team's, and the early product's. Some of it is top-down and some emergent — and it goes wrong either when a weak founder fails to institutionalize it, or when disagreement fragments the company into pockets of different cultures.

前 Dropbox CTO、South Park Commons 合伙人 Aditya Agarwal 提出，文化永远把战略当早餐吃掉，而创业公司的文化来自三个源头：创始人的个性与 DNA、早期团队的个性与 DNA、早期产品的个性与 DNA。其中一部分自上而下，一部分自然涌现。两种情况会出问题：创始人太弱、无法把文化制度化；或者内部分歧太大，公司裂成一块块文化飞地。

The Takeaway: The teams getting the most out of AI aren't using agents to do their jobs — they're using them to clear the mechanical work off the table so every remaining human hour goes to ideas and judgment.

核心要点： 从 AI 中获益最多的团队并不是让 agent 替他们干活，而是让 agent 清空所有机械性工作，把剩下的每一小时人类时间都投入到创意和判断上。

Every, Dan Shipper's AI media and products company, just launched All Access, a $625-per-year tier built around the Builder Pack: 10 exclusive partner benefits from the AI tools the team uses daily, including $1,000 in Codex credits, free Cursor access, plus Anthropic, Render, PostHog, Notion, Framer, and Flora perks. It generated the largest subscription revenue increase in Every's history. The team behind it, Yash (who runs Sparkle and works on growth), Austin (growth), and creative director Douglas, broke down how they actually work.

Dan Shipper 的 AI 媒体与产品公司 Every 刚上线了 All Access，一个每年 625 美元的会员层级，核心是 Builder Pack：来自团队每天在用的 AI 工具的 10 项独家合作权益，包括 1000 美元的 Codex credit、免费 Cursor 使用权，以及 Anthropic、Render、PostHog、Notion、Framer 和 Flora 的福利。它创造了 Every 历史上最大的一次订阅收入增长。操盘这次发布的团队成员 Yash（负责 Sparkle 并参与增长）、Austin（增长）和创意总监 Douglas 拆解了他们的真实工作方式。

The best story is Austin's. The night after launch, teammates suggested urgency emails targeting users who showed checkout intent but didn't convert. Austin screenshotted the Slack thread, dropped it into Codex, ran the team's /LFG compound-engineering loop (brainstorm, plan, work, review, compound), and went to the gym. By the time he got back, the agent had built four targeted cohorts in Kit, drafted audience-specific emails through Spiral's MCP to match their approved writing style, mined past sends for what drove opens and clicks, embedded a proven share image, scheduled the sends, and posted the plan to Slack for team review. He spent ten minutes tweaking headlines. Those emails drove over $25,000 in revenue by mid-morning. His reaction: "I wasn't like, holy shit, that's so cool. I was like, that's basically what it should do."

最精彩的是 Austin 的故事。发布次日晚上，同事提议给有结账意向但未转化的用户发一波限时催单邮件。Austin 直接截图 Slack 讨论串扔进 Codex，跑了团队的 /LFG 复利工程循环（头脑风暴、规划、执行、审查、沉淀），然后去了健身房。等他回来时，agent 已经在 Kit 里建好四个精准分组、通过 Spiral 的 MCP 按团队认可的文风为每组写好邮件、翻遍历史邮件找出最能带来打开和点击的写法、嵌入一张验证过的分享图、排好发送时间，还把整个方案发到 Slack 供团队审核。他只花了十分钟改改标题。第二天上午，这几封邮件就带来了超过 25000 美元的收入。他的反应是："我没有觉得天哪太酷了，我的感觉是，它本来就该做到这样。"

The operating framework, credited to colleague Kieran: "You're at the top and bottom of the AI sandwich. You decide what the problem is, you frame the idea, and then you review the work." Yash applies it by automating his entire A/B testing pipeline with Claude and PostHog because setting up experiments is "fake work" — the interesting part is deciding what to test. Douglas frames the shift as moving from playing an instrument to conducting an orchestra, and notes you can conduct several at once. Their advice for new builders: dupe a product you love as an MVP, build something you don't think you can, and if you're a creative, rebuild your portfolio with the full stack.

他们的工作框架来自同事 Kieran 的说法："你处在 AI 三明治的顶层和底层。你决定问题是什么、框定想法，最后审查产出。" Yash 的实践是用 Claude 加 PostHog 把整条 A/B 测试流水线自动化，因为搭实验本身是 "假工作"，真正有意思的是决定测什么。Douglas 把这场转变形容为从亲自演奏乐器变成指挥交响乐团，而且你可以同时指挥好几个乐团。他们给新手 builder 的建议：把一个你热爱的产品复刻成 MVP、去做一件你觉得自己做不到的事，如果你是创意工作者，就用整套工具栈重做你的作品集。