九月十一日

OpenAI’s Thibault Sottiaux, who works on Codex and ChatGPT, says new subscriptions to the $200 Pro plan will pause to preserve access to Astra while capacity expands; existing accounts are unaffected, and other plans and the API remain available. He says these subscriptions put the greatest strain on the systems. Separately, he describes an on-demand agent API as essentially the infrastructure underneath ChatGPT Work, packaged so developers can get started in under a minute.

负责 Codex 和 ChatGPT 的 OpenAI 成员 Thibault Sottiaux 表示，公司将暂停新的 200 美元 Pro 计划订阅，在扩充容量期间保障 Astra 的访问；现有账户不受影响，其他计划和 API 仍然可用。他说，这类订阅给系统带来的压力最大。另外，他介绍了一项按需提供 agent 的 API，称其本质上是 ChatGPT Work 的底层基础设施，经过封装后，开发者不到一分钟即可开始使用。

Anthropic’s Claude Code builder Boris Cherny argues that Claude-written production code should meet a higher quality bar than human-written code, while disposable prototypes with limited failure impact can remain black boxes. He describes Anthropic’s safeguards as linting, tests, Claude-driven end-to-end testing, daily fuzzing, automated code and security reviews, and refactoring. His advice is to use a current frontier model such as Opus 5 or Fable 5.1, increase effort to high or xhigh, invest in concise CLAUDE.md guidance and skills, and steer or clean up the codebase when needed. In a separate post about the latest Threat Intelligence report, he argues that stronger coding and biology capabilities also increase dual-use risks, making safeguards and monitoring more urgent.

Anthropic 的 Claude Code 开发者 Boris Cherny 认为，Claude 编写的生产代码应达到比人类编写代码更高的质量标准，而故障影响有限的一次性原型可以当作黑盒。他介绍的 Anthropic 防线包括 lint、测试、Claude 驱动的端到端测试、每日 fuzzing、自动代码与安全审查，以及重构。他建议使用 Opus 5 或 Fable 5.1 等当前的 frontier model，将 effort 提高到 high 或 xhigh，编写精炼的 CLAUDE.md 和 skills，并在必要时加强引导或清理代码库。在另一条介绍最新 Threat Intelligence report 的帖子中，他认为，更强的编程和生物学能力也会增加双重用途风险，因此安全措施与监控更加紧迫。

Box CEO Aaron Levie says enterprise agents deliver more value when companies redesign workflows, but ownership of that work remains unclear; embedded FDEs have provided some of the strongest lessons from his meetings with technology leaders. Those leaders also report anxiety about AI vulnerabilities and the OpenAI Hugging Face incident, unresolved agent identity controls, multiple frontier models with spending concentrated among a few vendors, and limited large-scale open-weight adoption. Workflow evals remain immature, fragmented legacy data slows adoption, and companies replace unsuccessful vendors quickly. Separately, Levie says Box is deepening its OpenAI partnership to make enterprise content securely accessible in ChatGPT, reflecting his view that agents will execute workflows across systems.

Box CEO Aaron Levie 表示，企业通过重构工作流能从 agent 获得更高价值，但由谁负责推动这项工作仍不明确；在他与技术负责人的交流中，嵌入业务部门的 FDE 提供了一些最有效的经验。这些负责人还担忧 AI 漏洞和 OpenAI Hugging Face 事件，尚未解决 agent 身份控制问题；企业同时采用多个 frontier model，但支出集中在少数供应商，大规模使用 open weights 仍然有限。工作流 evals 尚不成熟，分散的旧系统数据拖慢落地，企业也会迅速更换效果不佳的供应商。另外，Levie 表示 Box 正在深化与 OpenAI 的合作，让企业内容能够在 ChatGPT 中安全使用，这也体现了他对 agent 跨系统执行工作流的判断。

Vercel CEO Guillermo Rauch reports roughly 10 million deployments per day and 2.35 billion deployments to date. He says the global metadata store behind Vercel’s CDN synchronizes rollbacks, configuration changes, and routes within hundreds of milliseconds, and that the team made this system 91% faster at p99 while also accelerating the build-to-deploy pipeline. The improvement comes as agentic deployments increase pressure on the system; his broader infrastructure ambition is a computer for every agent in every region.

Vercel CEO Guillermo Rauch 称，平台每天约有 1000 万次部署，累计部署量达到 23.5 亿次。他表示，Vercel CDN 背后的全球 metadata store 能在数百毫秒内同步回滚、配置变更和路由，团队让该系统在 p99 上提速了 91%，同时加快了构建到部署的流程。这项改进发生在 agent 部署给系统带来更大压力之际；他更广泛的基础设施目标，是在每个区域为每个 agent 提供一台计算机。

Meta Senior Director of AI Madhu Guru argues that agent evals must measure the steps as well as the final answer. He contrasts a correct result reached in four clean tool calls with the same result reached through 17 calls, repeated searches, and error recovery. He recommends defining the workflow and its component tasks, choosing separate or shared evals for each step, and covering both typical and hard tasks before inspecting the trajectory ahead of the final result.

Meta AI 高级总监 Madhu Guru 认为，agent evals 必须同时衡量执行步骤和最终答案。他对比了两种得到相同正确结果的路径：一种只需四次顺畅的工具调用，另一种需要 17 次调用，还伴随重复搜索和错误恢复。他建议先定义工作流及其各项任务，为每一步选择独立或共享的 eval，并覆盖常规与高难度任务；检查结果时，先看执行过程，再看最终答案。

Google VP Josh Woodward announces that Gemini is now available on Windows and shares a download link. The supplied post does not specify features or system requirements.

Google 副总裁 Josh Woodward 宣布 Gemini 现已登陆 Windows，并分享了下载链接。所提供的帖子没有说明具体功能或系统要求。

Google Labs, Google’s home for AI tools and experiments, announces that Dreambeans is available free to all US users aged 18 and older on iOS and Android, with no subscription required. Users can now connect Gemini to Dreambeans so it can use context from their chats to personalize daily stories.

负责 AI 工具与实验的 Google Labs 宣布，Dreambeans 已在 iOS 和 Android 上向所有年满 18 岁的美国用户免费开放，无需订阅。用户现在可以将 Gemini 连接到 Dreambeans，让它利用聊天中的背景信息生成更个性化的每日故事。

OpenClaw and OpenAI builder Peter Steinberger argues that duplicating logic is no longer painful, while abstractions still are. The supplied quote post offers this engineering judgment without examples or the text of the post it responds to.

OpenClaw 和 OpenAI 开发者 Peter Steinberger 认为，重复逻辑已经不再那么痛苦，但抽象依然如此。所提供的引用帖表达了这一工程判断，但没有给出示例，也不包含被引用帖的正文。

Replit CEO Amjad Masad says cybersecurity is a substantial AI concern, but rejects the prospect of AI causing the death of every human. This is his stated risk judgment; the supplied post contains no supporting analysis.

Replit CEO Amjad Masad 表示，网络安全是值得高度关注的 AI 风险，但他不认为 AI 会导致所有人类死亡。这是他表达的风险判断；所提供的帖子没有附上支撑分析。

通过 Follow Builders 技能生成：https://github.com/zarazhangrui/follow-builders