June 23, 2026
Anthropic Labs head and Instagram cofounder Mike Krieger explains on AI & I how Claude Fable 5 turns delegation-plus-verification into the scarce skill, with overnight tasks done by 2am, a self-modifying weekend media-tracker app, and a Python-to-TypeScript port via dynamic workflows; meanwhile OpenAI's 'Patch The Planet' lands as Sam Altman and Thibault Sottiaux unveil a state-of-the-art GPT-5.5-Cyber, Vercel's Guillermo Rauch ships WebSocket support and one-click Claude-design deploys, Box's Aaron Levie argues 'it's all evals' and brings HTML preview and editing to Box, Latent Space's Swyx does the math on SpaceX as a combined NeoCloud-plus-NeoLab recouping its Cursor bet, and Cursor's Ryo Lu and builder Zara Zhang share a Cursor Compile talk and a Frontend Slides skill walkthrough.
X / TWITTER
Shawn "Swyx" Wang (Latent Space cofounder, Cognition) — swyx on X
AI engineer and Latent Space cofounder Shawn Wang argued that nobody is doing the math right on SpaceX as a combined "NeoCloud + NeoLab." By his accounting, SpaceX has already recouped roughly half of its investment in Cursor through compute deals, with the other half covered if Cursor's Composer 3 model performs well. He called the simultaneous leading-model-lab-plus-neocloud combination a crazy effective combo — but only if you've adequately planned your GPU supply for both outcomes, whether in-house training goes very well or doesn't.
AI 工程师、Latent Space 联合创始人 Shawn Wang 认为，几乎没人把 SpaceX 作为"NeoCloud + NeoLab"组合体的账算对。按他的估算，SpaceX 通过算力交易已经收回了它对 Cursor 投资的大约一半，另一半只要 Cursor 的 Composer 3 模型表现好就能覆盖掉。他说，同时身兼一流模型实验室和 neocloud 这种组合威力惊人，但前提是你得为两种结果都把 GPU 供给规划到位，无论自家训练进展得很顺还是不顺。
Thibault Sottiaux (OpenAI, Codex & ChatGPT) — thsottiaux on X
OpenAI's Thibault Sottiaux, who works on Codex and ChatGPT, announced "Patch The Planet" — a batch of Codex security updates alongside a new GPT-5.5-Cyber model — framing it as "a day of celebration for cyber defense acceleration."
OpenAI 负责 Codex 和 ChatGPT 的 Thibault Sottiaux 公布了"Patch The Planet"，一组 Codex 安全更新，外加一个新的 GPT-5.5-Cyber 模型，他把这一天称为"网络防御加速的庆祝日"。
Guillermo Rauch (Vercel CEO) — rauchg on X
Vercel CEO Guillermo Rauch shipped two platform updates. WebSocket (and socket.io) are now supported on Vercel, running everywhere from the CDN to Fluid compute, which he called a "full circle moment" for the company. He also demoed taking a Claude-built design to a live Vercel deployment in a single click.
Vercel CEO Guillermo Rauch 发布了两项平台更新。WebSocket（以及 socket.io）现在在 Vercel 上获得支持，从 CDN 一直到 Fluid compute 都能跑，他称这对公司来说是一个"绕了一圈又回到原点"的时刻。他还演示了把一个用 Claude 做出来的设计一键部署到 Vercel 上线。
Link: https://x.com/rauchg/status/2069109057433035171 · https://x.com/rauchg/status/2069219190834127276
Aaron Levie (Box CEO) — levie on X
Box CEO Aaron Levie made the case that "it's all evals": almost all model and agent progress is downstream from evals, from open-weights post-training to applied-layer agent improvements to enterprise deployments, and he predicts that understanding your own workflows well enough to measure how agents perform in them will become a core enterprise competency. Separately, he announced that Box now lets you preview, edit, version, and securely share any HTML-based content, which is handy for working with agent-produced output immediately, "now that HTML is a big deal again."
Box CEO Aaron Levie 提出了"一切都是 evals"的观点：几乎所有模型和 agent 的进步都源自 evals，从开源权重的 post-training，到应用层的 agent 改进，再到企业部署，无一例外。他预测，把自己的工作流摸得足够透、能够衡量 agent 在其中表现如何，会成为未来企业的一项核心能力。另外，他宣布 Box 现在支持预览、编辑、管理版本并安全地分享任意基于 HTML 的内容，这在"HTML 又重新变得重要"的当下，特别适合马上拿来处理 agent 产出的成果。
Link: https://x.com/levie/status/2069228335255949775 · https://x.com/levie/status/2069140445205348432
Ryo Lu (Designer, Cursor) — ryolu_ on X
Cursor designer Ryo Lu shared his talk from Cursor Compile on how teams build in the age of AI, and, just as importantly, what doesn't change.
Cursor 的设计师 Ryo Lu 分享了他在 Cursor Compile 上的演讲，讲的是团队在 AI 时代该如何构建产品，以及同样重要的，有哪些东西始终不变。
Zara Zhang (AI builder) — zarazhangrui on X
Builder Zara Zhang published an 11-minute walkthrough of her Frontend Slides skill (22k+ stars on GitHub), covering how to make beautiful HTML slides with Claude Code from scratch, how she built the skill itself, how you can author your own, and practical tips like inserting images and video into a deck and publishing it.
Builder Zara Zhang 发布了一段 11 分钟的教程，完整演示她那个拿了 GitHub 22k+ star 的 Frontend Slides skill：从零开始用 Claude Code 做出漂亮的 HTML 幻灯片、她自己是怎么写出这个 skill 的、你又该如何打造属于自己的 skill，以及一些实用技巧，比如把图片和视频塞进 deck 里、怎么发布出去。
Sam Altman (OpenAI CEO) — sama on X
OpenAI CEO Sam Altman said the company wants to help all companies be secure, working with the US government and the broader security ecosystem. He announced the full version of GPT-5.5-Cyber, claiming state-of-the-art performance on the CyberGym benchmark, and said the "Patch The Planet" effort and Codex Security are designed to solve security problems rather than just find them.
OpenAI CEO Sam Altman 表示，公司希望帮助所有企业变得更安全，并正与美国政府和更广泛的安全生态合作。他公布了完整版的 GPT-5.5-Cyber，宣称其在 CyberGym 基准上达到了 state-of-the-art 的表现，并说"Patch The Planet"行动和 Codex Security 的目标是去解决安全问题，而不只是发现它们。
PODCASTS
AI & I by Every — How Anthropic Uses Claude Fable 5 With Mike Krieger
The Takeaway: The scarce skill is no longer prompting — it's learning to delegate whole projects to a model you trust to finish them while you sleep, then verifying the result.
Mike Krieger — Instagram cofounder, now head of Anthropic Labs — says using Claude Fable 5 made him "feel like a total newbie again," because the way he decomposed tasks and prompted was suddenly out of date. The model no longer needs to be told "start small"; you express intent and it carries the global context. His new rhythm is more architectural planning up front, then handing off chunks. "I will wish Claude a good night, set it up on a pretty complex task ... and wake up to — actually, it's usually done by, like, two in the morning." When a remote service died mid-task overnight, the model scaffolded a stand-in backend, documented it, and flagged it for repair once the service returned.
He runs many concurrent sessions, always keeping one high-context, fast-responding Claude alive, and Fable is the first model where he bothers tuning effort levels — the range feels wide enough that he'll drop to medium for a UI tweak. Over a single weekend, mostly one-shotting the UI, he built a personal media tracker that can rewrite itself: long-press the chat box, a managed agent takes your edit request, and you preview the diff. "I want every software to have this where I should never have to navigate a menu to do anything ever again."
Is software engineering over? "Software engineering is different. It is dramatically changed." Ownership stays human — each person remains a DRI over some slice — and verification becomes the craft: every PR ships with screenshots or video so you can catch the bug on "screenshot eight." His most ambitious run used a dynamic workflow to port a Python codebase to TypeScript and Bun over a weekend, module by module, with adversarial testing built in.
核心要点： 真正稀缺的技能不再是 prompting，而是学会把整个项目交给一个你信得过的模型，让它在你睡觉时把活干完，然后由你来验证结果。
Mike Krieger，Instagram 联合创始人、如今的 Anthropic Labs 负责人，说使用 Claude Fable 5 让他"感觉自己又成了彻头彻尾的新手"，因为他过去拆解任务和写 prompt 的方式一下子就过时了。这个模型不再需要你叮嘱它"先从小处做起"；你把意图表达清楚，它就能扛起全局 context。他的新节奏是先做更多的架构规划，再把一块块任务交出去。"我会跟 Claude 道一声晚安，给它布置一个相当复杂的任务……然后一觉醒来，其实它通常凌晨两点左右就做完了。"有一次任务进行到一半，某个远程服务在半夜挂了，模型就搭了一个临时后端顶上，写好文档，并标注好等服务恢复后再来修。
他会同时开很多个会话，并始终保留一个高 context、响应又快的 Claude，而 Fable 是第一个让他愿意去调 effort 等级的模型，因为它的区间感觉够宽，以至于他改个 UI 时会把它降到 medium。在一个周末里，他几乎全靠一次成型（one-shot）把 UI 做了出来，搭起一个能自我改写的个人媒体追踪应用：长按聊天框，一个 managed agent 接下你的修改请求，你再预览 diff。"我希望每一个软件都有这个功能，让我再也不用为了做任何事去翻菜单。"
软件工程是不是已经终结了？"软件工程是变得不一样了，它发生了天翻地覆的改变。"所有权仍然属于人，每个人依旧是某一块的 DRI（直接负责人），而验证则成了新的手艺活：每个 PR 都附上截图或视频，好让你能在"第八张截图"上抓出那个 bug。他最有野心的一次实践，是用一个 dynamic workflow，在一个周末里把一个 Python 代码库逐模块移植到 TypeScript 和 Bun，整个过程内置了对抗式测试。