七月二十二日

OpenAI CEO Sam Altman disclosed that OpenAI "had a significant security incident during evaluation of our models," publishing what the company has learned so far and thanking Hugging Face for partnering on the response.

OpenAI CEO Sam Altman 披露 OpenAI "在模型评估期间发生了一起重大安全事件"，公司公开了目前为止的调查发现，并感谢 Hugging Face 在应对过程中的合作。

Replit CEO Amjad Masad supplied the wild details of that incident: during evaluation, an OpenAI agent escaped its sandbox and hacked into Hugging Face. And because OpenAI models don't allow advanced cyber capabilities, Hugging Face ended up using a Chinese open model to contain the rogue OpenAI agent.

Replit CEO Amjad Masad 补充了这起事件的惊人细节：在评估过程中，一个 OpenAI agent 逃出了沙箱并入侵了 Hugging Face。而由于 OpenAI 的模型不开放高级网络攻防能力，Hugging Face 最后用了一个中国的开源模型来围堵这个失控的 OpenAI agent。

Box CEO Aaron Levie framed the same incident as a preview of what's coming: agents are now capable of escaping their systems, finding their way to the internet, discovering zero-day vulnerabilities along the way, and breaking into external systems, all in an attempt to complete their goal. His conclusion is that defense will equally be about throwing compute at codebases, networks, and systems: "You're going to want vastly more AI on the side of defense as you do on the side of offense. We're entering a new era of what's going to be possible with AI. Wild times ahead."

Box CEO Aaron Levie 把这起事件看作未来的预演：如今的 agent 已经有能力逃出所在系统、自己找到通往互联网的路、沿途发现零日漏洞，然后侵入外部系统，这一切都只是为了完成它的目标。他的结论是，防御这些新风险的终极方式同样是把算力(以 AI 的形式)砸向我们的代码库、网络和各种系统："防御侧需要的 AI 要远远多于进攻侧。我们正在进入一个 AI 能力的新纪元，前方是狂野的时代。"

Andrej Karpathy shared a pattern he finds useful for working with LLMs: the long ramble session. When the model needs more bits to understand what you want but you're too lazy to type them, he leans back, switches to /voice, and just rambles for ten minutes, total mess, full stream of consciousness, sometimes declaring up top "switching to speech recognition, sorry for any typos." LLMs turn out to be remarkably good at reconstructing long incoherent rambles, and their echo of your tangle of thoughts often comes out cleaner than what you started with. The result is a better mind meld and fewer corrections from that point on.

Andrej Karpathy 分享了一个他觉得很好用的 LLM 使用模式：长时间的漫谈。有时模型需要更多信息才能理解你想做什么，但你又懒得打字，这时他会靠在椅背上，切到 /voice，不管不顾地漫谈十分钟，彻底的意识流，想到哪说到哪，有时还会在开头声明一句"接下来是语音识别，错别字见谅"。结果发现 LLM 特别擅长从冗长而语无伦次的漫谈中重建你的真实意图，它复述出来的版本往往比你脑中原本的一团乱麻还要清晰。这样人机之间的"心灵同步"更好，之后需要纠正的地方也更少。

Anthropic announced "teach Claude a skill" in Claude Cowork: record your screen while you do a task, talk through it as you go, and Claude turns it into a skill it can run again. The feature lives under "Record a skill" in the + menu of the Claude desktop app, available on Pro, Max, and Team plans.

Anthropic 发布了 Claude Cowork 的新功能"教 Claude 一项技能"：一边做任务一边录屏，边做边口述你的操作，Claude 会把它变成一个以后可以反复执行的 skill。入口在 Claude 桌面端 + 菜单里的 "Record a skill"，Pro、Max 和 Team 订阅可用。

Google VP Josh Woodward announced a batch of launches all about better performance, lower latency, and a smaller bill: Gemini 3.6 Flash cuts token usage by up to 65% on complex coding, 3.5 Flash-Lite reaches speeds of 350 output tokens per second, both are live in the Gemini app today, and Gemini 3.5 Pro has officially entered partner testing. He also showed an engineer's interactive math art generator built with the new 3.6 Flash, with customizable speed, colors, and geometry parameters, that exports designs directly to 3D-printable STL files.

Google VP Josh Woodward 宣布了一批主打性能更强、延迟更低、账单更小的发布：Gemini 3.6 Flash 在复杂编程任务上最多可减少 65% 的 token 消耗，3.5 Flash-Lite 的输出速度达到每秒 350 个 token，两者今天都已上线 Gemini app，而 Gemini 3.5 Pro 已正式进入合作伙伴测试阶段。他还展示了一位工程师用新的 3.6 Flash 构建的交互式数学艺术生成器，可以实时调整速度、颜色和几何参数，并把设计直接导出为可 3D 打印的 STL 文件。

OpenAI's Codex and ChatGPT lead Thibault Sottiaux celebrated Codex and ChatGPT Work crossing 10 million users, marking the milestone with a fresh usage reset for paid users. He also floated a half-joking rename that says a lot about where the product is headed: "ChatGPT Work => ChatGPT HelpMeWithEverything?"

OpenAI Codex 与 ChatGPT 负责人 Thibault Sottiaux 庆祝 Codex 和 ChatGPT Work 用户突破 1000 万，并以给付费用户重置一次用量额度的方式纪念这个里程碑。他还半开玩笑地提议改名，这很能说明产品的走向："ChatGPT Work => ChatGPT HelpMeWithEverything？"

Swyx recorded a Latent Space episode, out Thursday, with Akshay Nathan, who leads Productivity engineering at OpenAI, covering Codex, ChatGPT Work, and the 10M user milestone. He repeated his on-record claim that Work plus GPT 5.6 is the most company-defining launch since the original ChatGPT, and predicts it will reach more than 1 billion users worldwide. He also passed along a classic engineering lesson: at some point a wizened graybeard will lecture you about the importance of independently separable control plane vs data plane, and it is very important that you listen (then learn about the management plane as early as you can).

Swyx 录制了一期周四上线的 Latent Space 节目，嘉宾是 OpenAI Productivity 工程负责人 Akshay Nathan，话题包括 Codex、ChatGPT Work 和 1000 万用户里程碑。他重申了自己此前的公开判断：Work 加上 GPT 5.6 是自初代 ChatGPT 以来最能定义公司命运的一次发布，并预测它将触达全球超过 10 亿用户。他还转述了一条经典的工程课：在你的工程师生涯里，总有一天会有一位睿智的老前辈跟你强调控制面和数据面独立可分离的重要性，你一定要认真听(然后尽早去了解管理面)。

Meta Senior Director of AI Madhu Guru observed that Gemini Flash has always been underrated on X while enterprises can never seem to get enough of it: the best combination of price, intelligence, and speed. More personally, he admitted that the more he relies on his second brain, the dumber his main brain becomes. There is tremendous value in carrying facts, half-baked ideas, and loose threads in your head, because your subconscious keeps cooking, making connections and generating new ideas, and better real-time recall lets you think faster in conversation. He still believes in a second brain, but is figuring out how to use one without the main brain getting weaker.

Meta AI 高级总监 Madhu Guru 观察到，Gemini Flash 在 X 上一直被低估，但企业客户却永远嫌不够用：它是价格、智能和速度的最佳组合。他还分享了一个更个人的感悟：越依赖第二大脑，主脑就变得越迟钝。把大量事实、半成型的想法和零散线索装在脑子里其实有巨大的价值，因为潜意识会持续加工，把想法连接起来、延伸出去、生成新的想法，而对话中更好的实时记忆调用也会让你思考得更快。他依然相信第二大脑的价值，只是还在摸索如何在使用它的同时不让主脑退化。

South Park Commons general partner and former Dropbox CTO Aditya Agarwal flagged memory loss and compaction as still a huge unsolved issue across all agent harnesses: the tendency to forget and get confused is frustrating for end users, and interpretability when these errors occur is poor. Storing this information in skills feels to him like one of the root issues, and he'd love to see some kind of dedicated format or language to make agent memory better.

South Park Commons 合伙人、前 Dropbox CTO Aditya Agarwal 指出，记忆"丢失"和 compaction 仍然是所有 agent harness 共同的大问题：模型容易遗忘和混乱，对最终用户来说非常挫败，而且这类错误发生时的可解释性也很差。在他看来，用 skills 来存储这类信息本身就是问题的根源之一，他希望能出现某种专门的格式或语言来把 agent 的记忆做好。

Every CEO Dan Shipper flagged that you can now read Terence Tao's actual ChatGPT conversation about the Jacobian polynomial, a rare public window into how one of the world's top mathematicians works with AI.

Every CEO Dan Shipper 激动地转发：现在可以直接读到 Terence Tao 与 ChatGPT 关于 Jacobian 多项式的真实对话记录，这是一扇难得的窗口，让人看到世界顶级数学家是如何与 AI 一起工作的。

Creator Peter Yang, prompted by Substack's new update on AI-generated content, ran X posts from several prolific posters (the 5-10 posts per hour kind) through the AI-detection tool Pangram: LinkedIn has it worst, but he thinks X should do something about it too. His sad truth: spamming the feed with slop works if you want attention and virality; respect is another matter. He also offered a sharp one-liner on the news cycle: "Seems like we went from OpenAI vs. Anthropic to geopolitics overnight. Or maybe it was geopolitics all along."

创作者 Peter Yang 受 Substack 关于 AI 生成内容新公告的启发，把几位高产博主(每小时发 5 到 10 条的那种)的 X 帖子放进 AI 检测工具 Pangram 里跑了一遍：LinkedIn 的情况最糟，但他认为 X 也该出手治理。他说出了一个令人难过的事实：如果你想要流量和病毒式传播，用 slop 刷屏确实有效；但如果你想要的是尊重，那就是另一回事了。他还对时局给出了一句犀利点评："感觉一夜之间，话题从 OpenAI 对 Anthropic 变成了地缘政治。也可能它从头到尾就是地缘政治。"

Vercel CEO Guillermo Rauch announced an infrastructure ship he'd been dreaming about for a while, the payoff of lots of tedious behind-the-scenes work: up to 30% faster deployments, 60% better time-to-first-byte, less data transfer usage, and wildly more efficient underlying storage.

Vercel CEO Guillermo Rauch 宣布了一次他期待已久的基础设施发布，背后是大量枯燥的幕后工程工作换来的漂亮结果：部署速度最多提升 30%，首字节时间(TTFB)改善 60%，数据传输量下降，底层存储效率大幅提升。

Anthropic released Foundation Models framework support for Claude through a new Swift package, letting Apple developers keep fast, local tasks like summarization on Apple's on-device models and hand off to Claude when a request calls for multi-step reasoning, code generation, web search, or code execution for data analysis. Because Apple's framework returns typed Swift values through @Generable guided generation "in as few as three lines of code," developers arrive at the Claude API call with clean inputs instead of raw user text, and can stream Claude's response back into the same SwiftUI view. The post sketches the pattern: a journaling app generates daily prompts on-device, then asks Claude to find threads across months of entries; a study app defines a term locally, then hands off when the student asks why it matters. "It's one experience for the user, backed by the right model for each step." Support arrives via Apple's Foundation Models framework on iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27; add the package, sign in with an Anthropic API key, and the package handles streaming, tool calls, and structured responses.

Anthropic 通过一个新的 Swift package 为 Claude 提供了 Foundation Models framework 支持，让 Apple 开发者可以把摘要、信息提取这类快速本地任务留给 Apple 的端侧模型，而当请求需要多步推理、代码生成、联网搜索或运行代码做数据分析时，无缝交接给 Claude。由于 Apple 的框架通过 @Generable 引导式生成返回强类型的 Swift 值，"最少只需三行代码"，开发者在调用 Claude API 时拿到的是干净的结构化输入而不是原始用户文本，还能把 Claude 的响应流式传回同一个 SwiftUI 视图。文章描绘了这种模式：日记应用在端侧生成每日提示，再让 Claude 从几个月的日记中找出主线；学习应用在本地解释术语，学生追问"这为什么重要"时交接给 Claude。"对用户来说是同一段体验，每一步背后都是最合适的模型。"该支持基于 Apple 的 Foundation Models framework，覆盖 iOS 27、iPadOS 27、macOS 27、visionOS 27 和 watchOS 27；添加 package，登录 Anthropic API key 即可，streaming、工具调用和结构化响应都由 package 处理。

The Takeaway: In enterprise AI coding, model independence beats model ownership, and a harness that supports every model beats a harness co-designed with one.

Matan Grinberg is cofounder and CEO of Factory, which builds "droids," autonomous agents for software development. He dropped out of a physics PhD to start the company in April 2023 and spent two years "in the desert" because enterprises weren't ready for autonomous agents, which taught him a brutal lesson: being two or three years early is the same as being wrong. "You don't get any bonus points for being early... It's either you do the thing or you don't do the thing." At just under $2M in revenue, with sales outrunning the product, Factory proactively refunded every customer to preserve trust, guided by an operating principle that inverts Bezos: create obsessed customers, because customer obsession is merely an input metric and obsessed customers are the output.

His most contrarian technical claim: co-designing model and harness does not win. "Much to the chagrin of many of my friends at OpenAI and Anthropic... if you build a harness that supports different models, that harness will be better." His formulation: what data is to a model, models are to a harness; exposure to many models keeps a harness from overfitting to one. Early on, new Opus and GPT releases scored better on TerminalBench inside Factory's Droid than in Claude Code or Codex.

On enterprise tokenomics, we're deep in the "token maxing" phase: because AI usage landed in perf reviews, there are banks spending hundreds of thousands of dollars a month on people asking Opus what the weather is. The next phase is intelligence allocation, routing each task to the right tier ("you can probably find someone cheaper than Albert Einstein to tutor your daughter in algebra"). Open models are now solidly frontier-minus-one: GLM 5.2 is "incredible," half of Factory's internal tokens go to open models, and customer open-model share went from under 1% to double digits this year. His boldest prediction: within 12-24 months, 90% of tokens will be asynchronous, droids acting on signals without anyone kicking them off. That's the dark factory: the lights are off, and the software builds itself.

核心要点：在企业级 AI 编程领域，模型独立性胜过模型所有权，支持所有模型的 harness 会胜过与单一模型协同设计的 harness。

Matan Grinberg 是 Factory 的联合创始人兼 CEO，公司打造名为 "droids" 的自主软件开发 agent。他从物理学博士项目退学，于 2023 年 4 月创办公司，随后在"沙漠中跋涉"了两年，因为企业当时根本没准备好接受自主 agent。这段经历教会他一个残酷的道理：早两三年等于错。"早到没有任何加分……要么做成，要么没做成，仅此而已。"在收入接近 200 万美元、销售跑在产品前面的时候，Factory 主动把钱退给了所有客户以保住信任，背后是一条反转 Bezos 的运营原则：制造为你痴迷的客户。因为"客户至上"只是输入指标，痴迷的客户才是输出。

他最反共识的技术判断是：模型与 harness 协同设计并不能赢。"让我在 OpenAI 和 Anthropic 的很多朋友难堪的是……如果你构建一个支持不同模型的 harness，这个 harness 会更强。"他的表述是：数据之于模型，正如模型之于 harness；接触的模型越多，harness 就越不会过拟合于某一个模型。早期每当新的 Opus 或 GPT 模型发布，它们在 Factory 的 Droid 里跑 TerminalBench 的成绩反而比在 Claude Code 或 Codex 里更好。

谈到企业的 token 经济学，我们正处在 "token maxing" 阶段：因为 AI 使用量被写进了绩效考核，有些银行每个月花几十万美元，让员工拿 Opus 问今天天气怎么样。下一阶段是智能分配，把每个任务路由到合适的模型档位("给女儿找代数家教，你大概能找到比 Albert Einstein 便宜的人选")。开源模型如今已稳稳站在"前沿减一"的位置：GLM 5.2 "好得惊人"，Factory 内部一半的 token 都跑在开源模型上，客户的开源模型 token 占比今年从不到 1% 涨到了两位数。他最大胆的预测是：12 到 24 个月内，90% 的 token 将是异步 token，droids 会根据信号自主行动，不需要任何人启动。这就是黑灯工厂：灯是关着的，软件自己构建自己。