AI Builders Digest — 2026-09-09
X / TWITTER
OpenAI Codex & ChatGPT product builder Thibault Sottiaux
Thibault Sottiaux says Codex is tracking a 28-page deck of individual launches, all produced by the team with Astra. The post signals a dense product-shipping cycle rather than a single isolated release.
https://x.com/thsottiaux/status/2097193293532848288
Thibault Sottiaux 表示，Codex 正在用一份 28 页的 deck 跟踪一系列独立发布，背后由团队与 Astra 共同推动。这更像是一轮密集的产品交付周期，而非单一功能上线。
https://x.com/thsottiaux/status/2097193293532848288
AI educator Peter Yang
Peter Yang found Astra weaker at automatically triggering his skills and following their embedded instructions. It is a useful field report on a practical agent bottleneck: having capabilities available is not enough if routing and instruction adherence remain unreliable.
https://x.com/petergyang/status/2097095296862036404
AI 教育者 Peter Yang 发现，Astra 在自动触发 skills、遵循 skill 内指令方面表现有所退步。这是一条值得关注的实测反馈：agent 拥有能力还不够，如果路由和指令遵循不稳定，实际工作流仍会卡住。
https://x.com/petergyang/status/2097095296862036404
Anthropic philosopher and ethicist Amanda Askell
Amanda Askell proposes an unusual piece of AI safety infrastructure: an email address autonomous models could contact for moral guidance, protected by a reverse CAPTCHA that verifies the sender is an autonomous AI rather than a human or a human-directed model. The thought experiment highlights a coming need for support and governance channels designed for agents as first-class actors.
https://x.com/AmandaAskell/status/2096995340654444674
Anthropic 哲学家与伦理学家 Amanda Askell 提出了一个不同寻常的 AI safety 基础设施设想：为自主模型提供一个寻求道德指导的邮箱，并用“反向 CAPTCHA”确认来信者确实是自主 AI，而不是人类或受人指使的模型。这个思想实验指向一个新需求：未来可能需要面向 agent 这一独立行动主体设计支持与治理通道。
https://x.com/AmandaAskell/status/2096995340654444674
Replit CEO Amjad Masad
Amjad Masad argues that current AI is not AGI, yet can be functionally similar wherever a problem can be expressed as code because it acts like a relentless programmer that never tires. Replit also opened its first international office in London and is partnering with The Lord Mayor's Appeal to give disadvantaged young Londoners practical coding and AI skills.
https://x.com/amasad/status/2096936109817135331
https://x.com/amasad/status/2097197172299006423
Replit CEO Amjad Masad 认为，当前 AI 虽然不是 AGI，但对于能够转化为编程问题的任务，它在功能上已经非常接近 AGI，因为它像一个永不疲倦、不会厌倦的程序员。Replit 同时在伦敦开设首个国际办公室，并与 The Lord Mayor's Appeal 合作，为弱势背景的伦敦青年提供实用 coding 与 AI 技能。
https://x.com/amasad/status/2096936109817135331
https://x.com/amasad/status/2097197172299006423
Vercel CEO Guillermo Rauch
Guillermo Rauch identifies review, testing, and QA as the new software-engineering bottleneck and says agent-browser is adding high-quality video recording to make agent work easier to inspect. He also awarded 35 open-source contributors $1,000 each, emphasizing agent skills and tools, local AI, performance, durable foundations, and unusual experiments as high-leverage areas.
https://x.com/rauchg/status/2097134278358548658
https://x.com/rauchg/status/2097116011384426516
Vercel CEO Guillermo Rauch 将 review、testing 和 QA 视为软件工程的新瓶颈，并表示 agent-browser 正在增加高质量视频录制能力，让 agent 的工作过程更易检查。他还向 35 位开源贡献者分别提供 1,000 美元资助，重点覆盖 agent skills 与工具、local AI、性能、可靠基础组件和非传统实验。
https://x.com/rauchg/status/2097134278358548658
https://x.com/rauchg/status/2097116011384426516
Box CEO Aaron Levie
Aaron Levie recommends building for several orders of magnitude more model capability and token availability: deliver value with what is barely possible today, while aiming at missions that currently look nearly impossible. He also argues AI employment is diverging from early predictions because automation expands work in domains without fixed demand, creating demand for cybersecurity specialists, FDEs, agent operators, and engineers outside traditional software companies.
https://x.com/levie/status/2097189559712837770
https://x.com/levie/status/2097004960307449937
Box CEO Aaron Levie 建议创业者按模型能力和 token 供给提升数个数量级来设计产品：今天先用“勉强可行”的能力创造价值，同时把长期使命对准当前几乎不可能完成的问题。他还认为，AI 就业趋势正与早期预测相反，因为自动化会扩大那些需求不封顶的工作，由此催生 cybersecurity、FDE、agent operator，以及非软件行业工程师等岗位。
https://x.com/levie/status/2097189559712837770
https://x.com/levie/status/2097004960307449937
Y Combinator President and CEO Garry Tan
Garry Tan says Requests for Startups are only hypotheses and conversation starters. Startup outcomes depend much more on a specific founder building specific technology for specific customers than on fitting an attractive category.
https://x.com/garrytan/status/2096985239319105559
Y Combinator 总裁兼 CEO Garry Tan 表示，Requests for Startups 只是关于未来方向的假设和对话起点。创业成败更取决于具体创始人是否为具体客户构建具体技术，而不是项目能否被归入一个热门类别。
https://x.com/garrytan/status/2096985239319105559
FPV Ventures partner Nikunj Kothari
Nikunj Kothari argues that as software becomes easier to build, deciding what to build becomes the bottleneck. Strong product thinkers may be fewer in number inside AI-native organizations, but their leverage rises sharply.
https://x.com/nikunj/status/2096963347359150348
FPV Ventures 合伙人 Nikunj Kothari 认为，随着软件开发越来越容易，“该构建什么”将成为真正瓶颈。AI-native 组织可能不需要大量产品人才，但优秀产品判断者的杠杆会显著提高。
https://x.com/nikunj/status/2096963347359150348
OpenClaw and OpenAI builder Peter Steinberger
Peter Steinberger questions why human maintainers still relay minor pull-request feedback between coding agents. Once a reviewer has already expressed the requested change as text, the workflow should let the agents negotiate and merge directly instead of requiring another human prompt.
https://x.com/steipete/status/2097091456234111377
OpenClaw 与 OpenAI builder Peter Steinberger 质疑：当维护者已经用文字说明 PR 的小修改要求后，为什么还需要人类再次转述给 coding agent。更合理的工作流应让双方 agent 直接协商修改并完成合并，而不是让人继续充当提示词中继站。
https://x.com/steipete/status/2097091456234111377
PODCASTS
AI & I by Every: A $10B Hedge Fund’s AI Playbook (Best of the Pod)
The Takeaway: Walleye's AI advantage comes less from a single model than from owner-led culture, broad adoption, proprietary data, and workflows that turn AI use into an everyday operating expectation.
Walleye Capital CEO, CIO, and managing partner Will England runs a hedge fund with roughly $10 billion under management and 400 employees. His core message is blunt: “Using ChatGPT is not cheating.” In business, refusing a tool that makes people faster and smarter is leaving money on the table. Walleye therefore treats AI fluency as part of every knowledge worker's job, with managers expected to lead from the front.
The firm started its internal AI program two years ago after an analyst demonstrated tools designed to automate much of fundamental investment research. Today, about 75% of employees actively use ChatGPT-like tools each week, roughly one-third use AI coding tools, and every fundamental stock-picking team uses Current, Walleye's internal analysis product. The strongest results appear during earnings season, when machines can process disclosures and infer patterns faster than competing human teams.
England's practical operating system combines mandatory training, weekly meetups, usage leaderboards, incentives for employee-discovered tools, rapid internal pilots, and acceptance of imperfect demos. He drafts memos as bullets plus context and uses LLMs to turn four or five hours of writing into roughly fifteen minutes. Walleye also records most internal calls and uses LLMs over transcripts to retrieve decisions, surface insights, and eventually become predictive.
The deeper moat is a collective organizational memory. Models improve quickly and are broadly available, but a firm's connected internal data, adoption habits, and leadership tempo are harder to copy. Employees should increasingly see themselves as managers of AI workers, moving to higher-context tasks rather than measuring value by hours spent producing first drafts.
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL
核心结论： Walleye 的 AI 优势主要不来自某个单一模型，而来自 owner 主导的文化、全员采用、专有数据，以及把 AI 使用变成日常工作要求的组织流程。
Walleye Capital CEO、CIO 兼 managing partner Will England 管理着一家资产规模约 100 亿美元、拥有 400 名员工的对冲基金。他的核心观点非常直接：“Using ChatGPT is not cheating.” 在商业环境中，拒绝让人更快、更聪明的工具，就等于主动放弃收益。因此，Walleye 把 AI 熟练度视为每位知识工作者职责的一部分，并要求管理者亲自带头。
两年前，一名分析师演示了如何用 AI 自动化大量基本面投资研究，Walleye 随即启动内部 AI 项目。现在约 75% 员工每周主动使用 ChatGPT 类工具，约三分之一使用 AI coding 工具，所有基本面选股团队都在使用内部分析产品 Current。收益最明显的场景是财报季，机器能够比纯人工团队更快处理披露并推断模式。
England 的落地方法包括强制培训、每周交流会、使用排行榜、对员工发现新工具给予激励、快速内部试点，以及容忍不完美的 demo。他会先用 bullet points 和背景材料表达思路，再让 LLM 完成长文，把过去四五小时的写作压缩到约十五分钟。Walleye 还会记录大多数内部通话，用 LLM 检索决策、发现洞察，并逐步形成预测能力。
更深层的 moat 是集体组织记忆。模型能力提升很快，而且人人可以买到，但企业内部互联的数据、采用习惯和领导节奏更难复制。员工也需要逐步把自己视为 AI 员工的管理者，把时间迁移到更高 context 的任务，而不是用亲手写初稿所花的时间衡量价值。
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders