Top Signals
ChatGPT Work turns a phone prompt into an end-to-end workflow
Sam Altman described giving ChatGPT Work a single prompt from his phone: use his chat history to propose three long-weekend destinations for nine friends, build a full-stack coordination site, reach group agreement, make reservations, and draft the invitation email in Gmail. He said the entire chain “just worked.” OpenAI's Thibault Sottiaux added that Work handles at least 20 tasks for him each day, including negotiating an internet bill, unsubscribing from spam, and finding deals. The consequential shift is from generating an answer to operating across research, software creation, group decisions, transactions, and communication from one request.
Sam Altman 描述了自己如何在手机上只给 ChatGPT Work 一条 prompt：利用他的聊天记录，为九位朋友提出三个周末旅行目的地，搭建一个用于协商选择的 full-stack 网站，在团队达成一致后完成预订，并在 Gmail 中起草邀请邮件。他表示，整条任务链“就这样成功了”。OpenAI 的 Thibault Sottiaux 补充说，Work 每天至少替他处理 20 件事，包括协商网络账单、退订垃圾邮件和寻找优惠。真正重要的变化，是 AI 正从生成答案升级为通过一次请求完成调研、软件创建、集体决策、交易和沟通。
The applied AI layer grows as models improve
Box CEO Aaron Levie argued that intelligence alone cannot transform enterprise processes. Useful systems must connect models to company data and software, insert human decisions through the right UX, create feedback loops that improve data and models, and satisfy regulatory and compliance constraints. The implementation for bank onboarding is fundamentally different from legal contract review or manufacturing. Levie's contrarian point is that stronger models expand rather than erase this applied layer: better capabilities make more ambitious workflows automatable, increasing the need for industry-specific companies that can bridge AI to real operations.
Box CEO Aaron Levie 认为，仅有 intelligence 并不能改造企业流程。真正有用的系统必须把模型连接到企业数据和软件，通过合适的 UX 在关键节点引入人工决策，建立能够持续改善数据与模型的反馈循环，并满足监管与合规要求。银行客户 onboarding 的实现方式，与法律团队的合同审查或制造业流程有根本区别。他提出的反直觉观点是，更强的模型不会消灭 applied AI layer，反而会扩大它：模型能力越强，可自动化的 workflow 就越有野心，也就越需要能够把 AI 接入真实业务的垂直公司。
Trust, not token supply, is the adoption bottleneck
Peter Yang said that outside AI-heavy circles, the leading concern is not running out of tokens but whether people trust ChatGPT enough to connect Gmail, Calendar, Google Workspace, or Microsoft Office. Zara Zhang identified a second adoption problem: a general chat product presents a blank box, so users freeze because they do not know what to ask. She proposed measuring AI adoption by the time from a user need arriving to the resulting feature shipping, not by tokens consumed. Together, the posts suggest that permissioning, product guidance, and completed outcomes are more meaningful adoption tests than raw usage volume.
Peter Yang 表示，在 AI 氛围没那么浓的圈子里，人们最担心的并不是 token 不够用，而是能否放心让 ChatGPT 连接 Gmail、Calendar、Google Workspace 或 Microsoft Office。Zara Zhang 指出了另一个 adoption 难题：通用聊天产品只给用户一个空白输入框，很多人会因为不知道该问什么而停住。她建议不要用消耗了多少 token 衡量 AI adoption，而应衡量从用户需求出现到对应功能上线需要多久。这几条观点共同说明，权限信任、产品引导和实际完成的结果，比原始使用量更能反映真实采用程度。
Vercel compiles its TypeScript CLI to a 1.28 MB native binary
Vercel CEO Guillermo Rauch compiled the Vercel CLI's TypeScript to a fully static native binary with `scriptct`, using code translated by GLM 5.2 Fast. He reported a 1.28 MB binary, 1.5 ms mean startup overhead, and a 2.94-second mean compile time while retaining readable TypeScript and support for Node modules including HTTPS, filesystem, path, OS, and crypto. The result embeds neither V8 nor QuickJS, showing a concrete path from an existing TypeScript tool to a tiny deployable native artifact.
Vercel CEO Guillermo Rauch 使用 `scriptct` 把 Vercel CLI 的 TypeScript 编译成了 fully static native binary，其中代码由 GLM 5.2 Fast 转译。他给出的结果是：binary 仅 1.28 MB，平均启动开销 1.5 ms，平均编译时间 2.94 秒，同时保留可读的 TypeScript，并支持 HTTPS、filesystem、path、OS 和 crypto 等 Node module。产物既不嵌入 V8，也不嵌入 QuickJS，展示了一条把现有 TypeScript 工具转化为轻量、可部署 native artifact 的具体路径。
Open weights gain another industry signer as security incentives stay contested
Vercel signed the Open Weights and American AI Leadership letter, with Guillermo Rauch arguing that open source, data, protocols, and research underpin technological progress and that open weights are the next logical frontier. Replit CEO Amjad Masad surfaced a counterpoint to the assumption that open models are attackers' default choice: he cited a former Anthropic employee's claim that hackers prefer massively subsidized AI-lab subscriptions for attacks. The juxtaposition keeps the debate grounded in both ecosystem openness and the economic incentives shaping real misuse.
Vercel 签署了 Open Weights and American AI Leadership letter。Guillermo Rauch 认为，open source、数据、协议和研究共同支撑了技术进步，而 open weights 是顺理成章的下一步。Replit CEO Amjad Masad 则对“攻击者默认会选择 open model”这一假设提出了反例：他转述一名前 Anthropic 员工的说法，称黑客更偏好使用由 AI lab 大幅补贴的订阅服务发起攻击。两条动态把讨论同时落在生态开放，以及决定真实滥用行为的经济激励上。
AI product impact is moving from adjacent features to net-new software
Meta Senior Director of AI Madhu Guru framed current AI deployment as phase one: companies with distribution are using AI to move rapidly into adjacent problem areas and build functionality, such as virtual clothes try-on, that previously required substantial custom software. He expects phase two to produce far more net-new features and innovation, making AI's effect on the shape of the software ecosystem undeniable. His thesis explains why product-level impact can lag rapid internal execution: companies are still learning the playbook before the broader ecosystem visibly changes.
Meta Senior Director of AI Madhu Guru 把当前的 AI 落地定义为第一阶段：拥有 distribution 的公司正在借助 AI 快速进入相邻问题领域，并构建过去需要大量 custom software 才能实现的功能，例如虚拟试衣。他预计第二阶段会出现更多真正全新的功能和创新，让 AI 对软件生态形态的影响变得无可争议。这个判断也解释了为什么产品层的外部影响会落后于企业内部飞速提升的执行速度：公司仍在摸索 playbook，整个生态的变化尚未完全显现。
Podcast
The MAD Podcast with Matt Turck: OpenAI’s Compute Chief: We Can’t Build Fast Enough | Sachin Katti
The Takeaway: OpenAI's compute strategy is becoming full-stack because demand is already consuming everything the company can bring online, while power and physical supply chains, not model ambition, set the pace.
核心结论：OpenAI 的 compute strategy 正在走向 full-stack，因为所有新上线的算力都会立刻被消耗，而真正决定扩张速度的，是电力与实体供应链，而不是模型层面的野心。
OpenAI Head of Industrial Compute Sachin Katti, previously Intel CTO, described AI data centers as giant liquid-cooled supercomputers that turn electrons into tokens. He said a directional estimate of $50 billion in OpenAI compute spending this year sounded right. OpenAI funds new generation, transmission lines, transformers, and substations when connecting facilities to the grid; where grid capacity reaches its limits, it is also exploring on-site generation, currently including gas turbines. Katti called nuclear the densest clean energy source and said it “can't come soon enough.”
OpenAI Head of Industrial Compute Sachin Katti 曾任 Intel CTO。他把 AI data center 描述为把电子转化为 token 的巨型 liquid-cooled supercomputer，并表示 OpenAI 今年投入约 500 亿美元 compute 的估算大方向上合理。建设设施接入电网时，OpenAI 会为新增发电能力、输电线路、变压器和变电站提供资金；在电网容量触顶的地区，公司也在探索现场发电，目前包括燃气轮机。Katti 称核能是最密集的清洁能源，并直言它“越快到来越好”。
The custom Jalapeno chip targets tokens per watt, using OpenAI's knowledge of its own models and workloads to co-design more efficient inference hardware. Katti said inference may already consume the majority of compute and that even training increasingly consists of inference-heavy work such as synthetic-data generation and post-training. He rejected overbuilding as OpenAI's main risk: demand far exceeds supply, every increment comes online and is consumed immediately, and AI doing more AI research could multiply the number of experiments beyond the limit imposed by scarce human researchers. The danger, in his view, is slowing down while factories and supply chains remain unable to add capacity fast enough.
定制芯片 Jalapeno 的核心指标是 tokens per watt，它利用 OpenAI 对自身模型和 workload 的了解，共同设计更高效的 inference hardware。Katti 表示，inference 可能已经消耗了大部分 compute，而且训练本身也越来越依赖 inference，例如生成 synthetic data 和 post-training。他认为 OpenAI 的主要风险不是过度建设：需求远超供应，每一份新上线的算力都会立即被用掉，而 AI 参与更多 AI research 后，实验数量可能突破稀缺人类研究员原本构成的上限。在他看来，真正的危险是放慢脚步，而工厂和供应链仍无法足够快地增加产能。