AI Builders Digest - 2026-07-28
X / TWITTER
Thibault Sottiaux, Codex & ChatGPT at OpenAI
Thibault Sottiaux says OpenAI feels unusually focused right now, and his more practical message is that ChatGPT has crossed from "answer box" into delegated phone-native work. His examples are mundane but important: negotiating an internet bill, cleaning spam subscriptions, finding a deal, and handling repeated personal tasks from a single prompt. The signal is that OpenAI builders are pushing ChatGPT as an action layer, not just a reasoning layer.
Thibault Sottiaux 认为 OpenAI 当前处在高度聚焦的状态。更值得注意的是，他把 ChatGPT 描述成手机上的任务执行层：谈网费、清理垃圾订阅、找优惠、处理日常重复任务，都可以从一个 prompt 开始。这个信号说明 OpenAI 内部正在把 ChatGPT 推向“替你做事”的产品定位，而不只是“回答问题”。
Links:
- https://x.com/thsottiaux/status/2081534792903147881
- https://x.com/thsottiaux/status/2081444811647963244
Peter Yang, AI tutorials and interviews
Peter Yang points to the next adoption blocker: trust, not token limits. Outside the AI-native bubble, mainstream users are less worried about burning tokens and more worried about whether ChatGPT should be allowed into Gmail, Calendar, Google Workspace, Microsoft Office, and other private work surfaces.
Peter Yang 提醒了一个更真实的 AI 采用门槛：不是 token 成本，而是信任。对于非 AI 原生用户来说，关键问题不是“会不会用完额度”，而是“我是否信任 ChatGPT 进入 Gmail、Calendar、Google Workspace、Microsoft Office 这些高隐私工作场景”。
Link:
- https://x.com/petergyang/status/2081555286817648738
Madhu Guru, Senior Director of AI at Meta
Madhu Guru argues that AI's product impact is still in phase one. Distribution-heavy companies are using AI to expand into adjacent workflows and ship features that previously needed lots of custom software, but the larger ecosystem shift is not yet fully visible. His phase two prediction: net-new product forms and visibly changed software categories.
Madhu Guru 判断，AI 对产品形态的影响仍处在第一阶段。有分发能力的公司正在用 AI 快速扩展相邻场景，做出过去需要大量定制软件才能实现的功能，但生态层面的改变还没完全显现。他认为第二阶段会出现更多真正全新的功能与产品形态，届时 AI 对软件生态的改变会更明显。
Link:
- https://x.com/realmadhuguru/status/2081437850466451736
Amjad Masad, CEO of Replit
Amjad Masad highlights a security angle from a former Anthropic employee: attackers may prefer heavily subsidized frontier-lab subscriptions over open models. The practical implication is that abuse prevention has to treat consumer and prosumer AI products as attractive attack infrastructure, not only monitor open-weight model misuse.
Replit CEO Amjad Masad 提到一个安全侧信号：攻击者可能更偏好使用被大幅补贴的前沿实验室订阅服务，而不是开源模型。实际含义是，AI 滥用治理不能只盯 open-weight model，也要把消费级和专业级 AI 产品当作潜在攻击基础设施来管理。
Link:
- https://x.com/amasad/status/2081576172656456076
Guillermo Rauch, CEO of Vercel
Guillermo Rauch put Vercel behind the Open Weights and American AI Leadership letter, framing open weights as the next frontier after open source, data, protocols, and research. He also shared a technical experiment compiling the Vercel CLI TypeScript codebase to a native binary with scriptc: 1.28 MB binary size, 1.5 ms mean startup overhead, and a fast compile path using GLM 5.2 Fast. The builder signal is a mix of policy stance and practical pressure toward lighter, faster AI-assisted developer tools.
Vercel CEO Guillermo Rauch 明确支持 Open Weights and American AI Leadership letter，把 open weights 视为 open source、data、protocols、research 之后的下一道开放边界。他还展示了把 Vercel CLI 的 TypeScript 编译成 native binary 的实验：1.28 MB 二进制、平均 1.5 ms 启动开销，并使用 GLM 5.2 Fast 辅助代码转换。这里的信号是两层：政策上支持开放权重，工程上追求更轻、更快、更贴近本地体验的 AI-assisted developer tooling。
Links:
- https://x.com/rauchg/status/2081546513885622760
- https://x.com/rauchg/status/2081517519303737559
Aaron Levie, CEO of Box
Aaron Levie says the biggest AI opportunity is not raw model intelligence but the applied AI layer that connects models to real workflows. Enterprise automation needs systems integration, data access, human decision points, UX, compliance, and feedback loops. His contrarian point: as models improve, the need for this applied layer may grow because companies will attempt more ambitious workflow automation.
Box CEO Aaron Levie 认为，AI 最大机会不在模型智能本身，而在把模型连接到真实工作流的 applied AI layer。企业自动化需要系统集成、数据接入、人类决策节点、UX、合规和反馈闭环。他的反直觉判断是：模型越强，应用层机会不一定变小，反而可能变大，因为企业会尝试自动化更复杂、更高价值的流程。
Link:
- https://x.com/levie/status/2081491621162668207
Garry Tan, President and CEO of Y Combinator
Garry Tan closed YC Startup School 2026 with Sam Altman as the anchor guest, then reduced the founder lesson to two words: "Be earnest." The useful signal is cultural rather than technical: YC is still pushing founders toward authenticity, directness, and seriousness over performative startup theater.
Y Combinator CEO Garry Tan 以 Sam Altman 作为 YC Startup School 2026 的收官嘉宾，并把给 founder 的提醒压缩成一句话：真诚，不要表演创业。这个信号偏文化层面：YC 仍在强调 founder 的真实、直接和认真，而不是创业姿态。
Links:
- https://x.com/garrytan/status/2081602195292864532
- https://x.com/garrytan/status/2081586567211348432
Zara Zhang, builder
Zara Zhang proposes a better metric for AI adoption: measure the time from user need to shipped outcome, not tokens burned. She also points out why generic chat products create so many tutorials: a blank box is cognitively hard, and many users do not know what to ask. Her own posting workflow is similarly low-friction: publish thoughts when they occur, often from things already said out loud.
Zara Zhang 提出一个更好的 AI adoption 指标：不要看消耗了多少 token，而要看从用户需求出现到实际交付结果用了多久。她还指出，通用聊天产品之所以催生大量教程，是因为空白输入框本身有认知负担，用户经常不知道该问什么。她自己的内容方法也很轻：想到就发，很多内容来自已经在现实对话中说过的话。
Links:
- https://x.com/zarazhangrui/status/2081627581997269192
- https://x.com/zarazhangrui/status/2081627109299310684
- https://x.com/zarazhangrui/status/2081304884469809295
Nikunj Kothari, partner at FPV Ventures
Nikunj Kothari's compressed prediction is that "proof of prompt" may replace "proof of work." Read narrowly, this is about prompts becoming evidence of intent, taste, process, and capability in AI-native work. Read broadly, it suggests a future where how someone asks, directs, and evaluates agents becomes a professional artifact.
FPV Ventures partner Nikunj Kothari 的一句话判断是：“proof of prompt” 可能会取代 “proof of work”。窄义上，这是说 prompt 会成为意图、品味、过程和能力的证据；广义上，它指向一种新工作范式：一个人如何提问、指挥和评估 agent，会变成职业能力的一部分。
Link:
- https://x.com/nikunj/status/2081383934928068619
Dan Shipper, CEO of Every
Dan Shipper is spending a week writing what he calls the definitive history of how Codex happened, based on deep interviews with OpenAI insiders. This is worth watching because Codex has become one of the clearest examples of AI moving from chat into software execution, and first-party process history may reveal how the product actually formed.
Every CEO Dan Shipper 正在写一篇关于 Codex 诞生过程的深度历史，材料来自与 OpenAI 内部人士的访谈。这值得关注，因为 Codex 是 AI 从 chat 走向软件执行的代表性产品之一；如果这篇文章够扎实，它可能会揭示这类产品是如何真正形成的。
Link:
- https://x.com/danshipper/status/2081412243388788988
Sam Altman, OpenAI
Sam Altman amplified two signals: he wants "a new kind of computer," and he says ChatGPT Work can coordinate a multi-person trip end to end from a phone prompt, including planning options, building a full-stack coordination site, reaching group agreement, making reservations, and drafting email. The claim is not just that AI can answer planning questions, but that work products, coordination surfaces, and external actions are merging into one interface.
Sam Altman 放大了两个信号：他想要“一种新的计算机”，并称 ChatGPT Work 已经能从手机 prompt 出发，端到端协调 9 人周末旅行，包括规划方案、生成全栈协作网站、推动集体决策、预订、起草邮件。这里的重点不是 AI 能回答旅行问题，而是工作产物、协作界面和外部行动正在融合进同一个入口。
Links:
- https://x.com/sama/status/2081513071135346814
- https://x.com/sama/status/2081396796174282900
PODCASTS
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 an industrial infrastructure strategy, because model demand now consumes every unit of compute the company can bring online.
Sachin Katti, OpenAI's head of industrial compute and former Intel CTO, describes AI data centers as factories that turn electrons into tokens. The most important shift is physical: liquid cooling, power generation, grid upgrades, transmission lines, substations, and even behind-the-meter generation are now part of the AI roadmap. His sharpest point is that slowing infrastructure buildout has repeatedly looked wrong in hindsight because demand keeps outrunning supply.
Katti also frames compute as a recursive loop: AI will increasingly help design the chips, systems, and infrastructure needed to train and serve the next generation of AI. That makes compute less like a procurement function and more like a core product and research capability.
Podcast link: https://www.youtube.com/watch?v=wEZBlmvxx4o
核心判断：OpenAI 的 compute 战略正在变成工业基础设施战略，因为模型需求会吃掉公司能上线的每一份算力。
Sachin Katti 是 OpenAI 的 head of industrial compute，曾任 Intel CTO。他把 AI data center 描述成“把电子变成 token 的工厂”。最关键的变化是物理世界：liquid cooling、电力供应、电网升级、输电线路、变电站，甚至 behind-the-meter 发电，都已经成为 AI 路线图的一部分。他最尖锐的判断是：过去每次认为“算力建设可以慢一点”，事后看几乎都会错，因为需求持续超过供给。
Katti 还把 compute 描述成一种递归循环：AI 会越来越多地帮助设计下一代 AI 所需的芯片、系统和基础设施。这意味着 compute 不再只是采购能力，而是 OpenAI 的核心产品能力和研究能力。
播客链接：https://www.youtube.com/watch?v=wEZBlmvxx4o
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders