六月二十六日

Box CEO Aaron Levie (levie on X) argues that AI regulation has quietly become reality — "We now have de facto AI regulation" — and frames the choice ahead as a prisoner's dilemma at insane scale. His logic: if the US stays at the frontier and slows model releases under heavy review, it gains an economic and geopolitical edge by controlling who gets frontier intelligence and when; but if a rival like China keeps pace without slowing, US delays end up advantaging their models and tech stack, and banning those models only puts the US at a steeper disadvantage. He reads open-weight models as the clear winner, becoming the base layer for the sovereign-AI strategies other countries will now pursue to avoid depending on US access. The whole bet, he says, hinges on one question: can closed models stay at the frontier in perpetuity, or will someone catch up?

Box CEO Aaron Levie（X 上的 levie）认为 AI 监管已经在不知不觉中成为现实——"我们现在已经有了事实上的 AI 监管"——并把接下来的局面描述为一场超大规模的囚徒困境。他的逻辑是：如果美国始终保持在前沿，并在严格审查下放慢模型发布，它就能通过控制谁、在什么时候获得前沿智能来赢得经济和地缘政治优势；但如果像中国这样的对手不放慢脚步、又能跟上节奏，美国的拖延反而会让对方的模型和技术栈占便宜，而封禁这些模型只会让美国处境更糟。他判断 open-weight 模型是明显的赢家，会成为各国为了不依赖美国而推行的主权 AI 战略的底层基础。他说，整个赌局归根结底就一个问题：闭源模型能否永远守在前沿，还是终将有人追上来？

URL: https://x.com/levie/status/2070310706369712272 · https://x.com/levie/status/2070370225271251161

South Park Commons general partner and Bevel Health co-founder Aditya Agarwal (adityaag on X) offers a contrarian warning to founders: "It is stupidly hard to start a pure software company right now." His point is that customers no longer want software, they want outcomes — so you end up doing a lot of custom work, every customer wants to deeply customize, and anything horizontal is "on a ticking clock." Opportunities still exist, he says, but only if you hold a genuinely heretical two-to-three-year view of where the world is heading.

South Park Commons 普通合伙人、Bevel Health 联合创始人 Aditya Agarwal（X 上的 adityaag）给创业者抛出一个反共识的警告："现在创办一家纯软件公司难得离谱。"他的观点是，客户要的不再是软件，而是结果——所以你最后会被迫做大量定制化工作，每个客户都想深度定制，而任何横向（horizontal）的东西都"在倒计时"。他说机会依然存在，但前提是你对未来两三年的走向，得有一个真正离经叛道的判断。

URL: https://x.com/adityaag/status/2070179913647485344

Vercel CEO Guillermo Rauch (rauchg on X) shared two builder-relevant signals. First, on tooling: he detailed how Vercel imbues coding agents with its design standards and praised Next.js's new "Ways to fix this" errors with one-click "Copy prompt" buttons as "a work of (agentic) art" — small UX moves that feed agents clean context. Second, on the model market: he reports Grok Imagine Video has become the clear leader in Vercel's AI Gateway, now accounting for roughly 50% of all videos generated by developers on the platform.

Vercel CEO Guillermo Rauch（X 上的 rauchg）分享了两个对 builder 有价值的信号。其一是工具层面：他详细讲了 Vercel 如何把自家的设计规范注入到 coding agent 里，并称赞 Next.js 新的 "Ways to fix this" 报错配上一键 "Copy prompt" 按钮是"一件（agentic 的）艺术品"——这些小的 UX 改动其实是在给 agent 喂干净的上下文。其二是模型市场：他透露 Grok Imagine Video 已经成为 Vercel AI Gateway 上的明显领头羊，如今占了开发者在平台上生成视频量的大约一半。

URL: https://x.com/rauchg/status/2070241572416078161 · https://x.com/rauchg/status/2070215849970119090

OpenAI's Thibault Sottiaux (thsottiaux on X), who works on Codex and ChatGPT, highlighted how far Codex has spread inside OpenAI — "Codex for everything at OpenAI" — and pointed to the February 2nd release of the Codex App as an inflection point that visibly pushed adoption beyond engineering teams. The thread frames Codex less as a developer tool and more as a general-purpose agent now used across the company.

OpenAI 的 Thibault Sottiaux（X 上的 thsottiaux）负责 Codex 和 ChatGPT，他强调了 Codex 在 OpenAI 内部扩散之广——"OpenAI 里什么都用 Codex"——并指出 2 月 2 日 Codex App 的发布是一个拐点，明显把使用范围推到了工程团队之外。整条 thread 把 Codex 描绘得越来越不像一个开发者工具，而更像一个如今全公司都在用的通用 agent。

URL: https://x.com/thsottiaux/status/2070205520552886305 · https://x.com/thsottiaux/status/2070205719501254860

AI tutorial creator Peter Yang (petergyang on X) shared a concrete glimpse of where agentic browsing is headed. Planning a December trip to Japan, he asked Codex to drive a browser through Google Flights and individual hotel websites, find and save all the prices, return direct booking links for his target dates, and compile everything into a doc — joking that he should just tell it to book the whole thing next. It's a vivid example of agents doing real, multi-step web research rather than chat.

AI 教程创作者 Peter Yang（X 上的 petergyang）给出了一个 agentic 浏览未来走向的具体画面。为了规划 12 月去日本的行程，他让 Codex 自己操作浏览器去逛 Google Flights 和各家酒店网站，把所有价格找出来并保存，针对他想要的日期返回直接预订链接，再把这一切整理进一个文档——他调侃说下一步干脆让它直接把全部都订了得了。这是 agent 做真正多步骤网页调研（而不是单纯聊天）的一个生动例子。

URL: https://x.com/petergyang/status/2070353698140958818

Builder Zara Zhang (zarazhangrui on X) pushes back on a common founder misconception — that building an audience is somehow at odds with building a company. In her view they're increasingly the same thing: storytelling, to users, investors, and candidates alike, is arguably the founder's most important job. Her sharpest line captures why: "Building is getting cheaper, attention is getting more expensive."

Builder Zara Zhang（X 上的 zarazhangrui）反驳了创业者中一个常见的误解——以为"做内容、攒受众"和"做公司"是对立的。在她看来，这两件事正越来越变成同一件事：无论是对用户、投资人还是候选人，讲好故事都可以说是创始人最重要的工作。她最犀利的一句话点出了原因："Building 越来越便宜，而注意力越来越贵。"

URL: https://x.com/zarazhangrui/status/2070188279941738952

Anthropic is releasing a new Swift package that lets Apple developers call Claude through Apple's own Foundation Models framework — handing off to Claude when a request needs multi-step reasoning, code generation, web search, or code execution, while keeping fast local tasks (summarization, extraction) on Apple's on-device models. Because the framework returns typed Swift values from `@Generable` annotations, developers reach the Claude API call "with clean inputs instead of raw user text," and the package handles streaming, tool calls, and structured responses straight back into a SwiftUI view. The pattern lets one user experience be backed by the right model at each step: a journaling app can generate daily prompts on-device, then ask Claude to find threads across months of entries; a study app can define a term locally, then hand off to Claude when the student asks why it matters. Support arrives on iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27 — add the package, sign in with an Anthropic API key, and pass typed outputs into a Claude request.

Anthropic 发布了一个新的 Swift package，让 Apple 开发者可以通过 Apple 自家的 Foundation Models framework 来调用 Claude——当一个请求需要多步推理、代码生成、联网搜索或执行代码时就交给 Claude，而把快速的本地任务（摘要、信息抽取）留给 Apple 的端上模型。由于该框架会通过 `@Generable` 注解返回带类型的 Swift 值，开发者在调用 Claude API 时拿到的"是干净的输入，而不是原始的用户文本"，这个 package 还会负责 streaming、工具调用，并把结构化结果直接送回 SwiftUI 视图。这种模式让同一套用户体验在每一步都由最合适的模型来支撑：一个日记应用可以在端上生成每日提示，再让 Claude 在几个月的记录里找出线索；一个学习应用可以在本地解释一个术语，等学生追问"这为什么重要"时再交给 Claude。该支持将登陆 iOS 27、iPadOS 27、macOS 27、visionOS 27 和 watchOS 27——添加 package、用 Anthropic API key 登录，然后把带类型的输出传进一个 Claude 请求即可。

URL: https://claude.com/blog/claude-for-foundation-models

The Takeaway: The internet's 28-year, ad-funded business model is breaking, because machines now outnumber humans online and bots don't click on ads.

Matthew Prince, co-founder and CEO of Cloudflare — which sits in front of a huge share of global internet traffic — drops a genuine milestone: bot traffic passed human traffic in the first half of 2026, years ahead of his own team's 2027 forecast. He treats "agent," "bot," and "crawler" as synonyms, and the driver is AI. Where you might visit five sites to research a digital camera, "my agent might visit 5,000," and he'll take the over bet that within five years the internet carries a thousand times more bot traffic than human. That snaps the load-bearing assumption of the web: "the business model of the Internet historically has been ads and bots don't click on ads."

The most useful thread is how Cloudflare keeps turning its own pain into products — DDoS mitigation, a registrar, a developer platform (Workers) — all built because they needed them, then sold to everyone. On security, his forecast is vivid and specific: for the next two years expect "a Log4j-like vulnerability every single week" as AI models get scarily good at finding bugs, followed by a head-fake where software ends up far more reliable. Cloudflare already runs an agent, trained on ten years of incidents, that reviews every code and config change — and watched its background rate of incidents fall off a cliff the day it shipped.

Then the line that names the episode: with 93% of R&D now using AI coding tools, Prince warns "almost every company is going to go through some exercise like this where they're gonna cut a bunch of their team" — and frets most about the mid-career engineers being told, mid-game, that the rules just changed.

核心要点： 互联网用了 28 年的、靠广告养活的商业模式正在崩塌，因为机器在网上的数量已经超过了人类，而 bot 不会去点广告。

Matthew Prince 是 Cloudflare 的联合创始人兼 CEO——这家公司挡在全球互联网相当大一部分流量的前面——他抛出了一个真正的里程碑：bot 流量在 2026 年上半年就超过了人类流量，比他自己团队 2027 年的预测提前了好几年。他把 "agent"、"bot"、"crawler" 当作同义词，而背后的驱动力就是 AI。你为了研究买哪台数码相机可能会逛五个网站，而"我的 agent 可能会逛 5000 个"，他还押注：五年内互联网承载的 bot 流量会是人类流量的一千倍。这一下就动摇了网络赖以支撑的前提："互联网历来的商业模式就是广告，而 bot 不会点广告。"

最有价值的一条线索，是 Cloudflare 如何不断把自己踩过的坑变成产品——DDoS 防护、域名注册商、开发者平台（Workers）——全都是因为自己需要才做出来，然后卖给所有人。在安全方面，他的预测既生动又具体：未来两年，随着 AI 模型在找漏洞上强得吓人，要做好"每周都冒出一个 Log4j 级别漏洞"的准备，但随后会出现一个假动作——软件最终反而会变得可靠得多。Cloudflare 已经在用一个基于十年事故数据训练出来的 agent，审查每一次代码和配置变更——上线那天，它眼看着事故的"背景噪声"水平断崖式下跌。

接着就是给这一集起名的那句话：如今 93% 的研发人员都在用 AI 编程工具，Prince 警告说"几乎每家公司都会经历这样一道工序，去裁掉一批团队成员"——而他最担心的，是那些处于职业中期、被告知"游戏规则在中途变了"的工程师。

URL: https://www.youtube.com/watch?v=UN47z_opfmo