AI Builders Digest — 2026-08-28

2026-08-28

AI Builders Digest — 2026-08-28

X / TWITTER

Anthropic Claude Code engineer Thariq

Thariq shipped a SendFeedback tool for Claude Code: users can now ask Claude to draft a feedback report and approve it, removing the friction of manually invoking /feedback. He also warned that fraudulent usage requests are targeting customers, highlighting the tension between abuse prevention and preserving access for legitimate users.

Thariq 为 Claude Code 推出了 SendFeedback 工具：用户现在可以直接让 Claude 起草反馈报告并确认发送，不必再手动调用 /feedback。他同时提醒，多家客户正在遭遇欺诈性用量请求，反映出平台必须在防滥用与保障正常用户访问之间持续权衡。

Vercel CEO Guillermo Rauch

Guillermo Rauch highlighted infrastructure built for agent workloads: global compute across multiple regions, failover, up to 10,000 concurrent sandboxes, and a default ramp of 5,000 vCPUs per minute. Vercel is also shipping a security dashboard and vercel security check CLI so agents can improve security posture with human approval or run checks on a cron schedule.

Guillermo Rauch 展示了面向 agent 工作负载的基础设施：多区域全球计算、故障切换、最高 10,000 个并发 sandbox，以及默认每分钟扩容 5,000 个 vCPU。Vercel 还将推出安全仪表盘和 vercel security check CLI，让 agent 在人工确认下修复安全问题，或通过 cron 定期检查。

Box CEO Aaron Levie

Aaron Levie says enterprise AI is increasingly a context and governance problem, not merely a model-capability problem. Box's Q2 revenue reached $321.1 million, up 9% reported and 11% in constant currency, with demand driven by enterprises trying to expose unstructured content to agents securely. His applied-AI thesis is that companies need model-swappable platforms, granular access controls, audit logs, and real-time alerts around the contracts, research, financial documents, and other content that agents use.

Aaron Levie 认为，企业 AI 越来越是上下文与治理问题，而不只是模型能力问题。Box 第二季度营收达 3.211 亿美元，同比报告口径增长 9%、固定汇率口径增长 11%；增长动力来自企业希望安全地把非结构化内容开放给 agent。他对 applied AI 的判断是：企业需要支持随时切换模型的平台，并围绕合同、研究资料、财务文档等 agent 所用内容建立细粒度权限、审计日志和实时告警。

FirstMark Capital VC Matt Turck

Matt Turck argues that NVIDIA's Nemotron work and Hugging Face relationship place it at the center of open-source AI, while giving Hugging Face a strategically aligned home. His broader market view is that apps are turning into labs faster than labs are turning into apps, venture outcomes are following a “hyper power law,” and investors should stop underestimating China while distinguishing an AI bubble from a duration-mismatch problem.

FirstMark Capital 投资人 Matt Turck 认为，NVIDIA 的 Nemotron 布局及其与 Hugging Face 的关系，使 NVIDIA 成为 open-source AI 的中心，同时也为 Hugging Face 提供了战略一致的归宿。他更广泛的市场判断是：应用公司变成实验室的速度快于实验室变成应用公司；风险投资结果正在呈现“hyper power law”；投资者还应停止低估中国，并区分 AI 泡沫与期限错配问题。

SPC General Partner Aditya Agarwal

Aditya Agarwal says the AI frontier will be determined by post-training. He announced DeepCogito's $43 million Series A for research into large-scale reinforcement learning and recursive self-improvement, centered on iterated distillation and amplification (IDA), which the company says it has demonstrated publicly across models from 3B to more than 600B parameters.

SPC General Partner Aditya Agarwal 判断，AI 前沿竞争将由 post-training 决定。他宣布 DeepCogito 完成 4,300 万美元 Series A，用于研究大规模 reinforcement learning 与 recursive self-improvement；其核心方法是 iterated distillation and amplification（IDA），公司称已在 3B 至 600B+ 参数模型上公开验证。

Anthropic's Claude

Anthropic announced a built-in browser inside Claude Cowork. When a task requires a website, Claude can open a browser in Cowork's side panel, navigate, fill forms, and complete the workflow; the browser is isolated from the user's existing logins. Claude in Chrome remains available for paid users who prefer automation inside their already authenticated browser.

Anthropic 宣布在 Claude Cowork 中内置浏览器。当任务涉及网站时，Claude 可在 Cowork 侧栏打开浏览器、导航、填写表单并完成流程；该浏览器与用户现有登录状态隔离。对于希望在已登录浏览器中自动化操作的付费用户，Claude in Chrome 仍然可用。

OpenClaw and OpenAI builder Peter Steinberger

Peter Steinberger's concise signal is that Codex's visualization capability has improved markedly. The attached demonstration is the substantive artifact; his other post merely asks whether the market may be a bubble and provides too little standalone context to support a reliable summary.

OpenClaw 与 OpenAI builder Peter Steinberger 给出的明确信号是：Codex 的可视化能力已经显著提升，实质内容主要在所附演示中。他另一条关于“是否存在泡沫”的帖子缺乏足够上下文，因此不做延伸解读。

Every CEO Dan Shipper

Dan Shipper flagged the arrival of Codex native apps and framed the current moment as a golden age for polymaths and philosophers. His point is that increasingly capable tools reward people who can connect disciplines and reason across domains, offering an antidote to epistemic despair.

Every CEO Dan Shipper 关注到 Codex native apps 已经推出，并把当下称为博学者与哲学家的黄金时代。他的核心判断是：能力持续增强的工具正在奖励那些能够连接不同学科、跨领域推理的人，也为“知识无从把握”的焦虑提供了一种解药。

Google VP Josh Woodward

Josh Woodward shared an experiment in “multiplayer AI,” pointing toward AI experiences designed for multiple simultaneous participants rather than the standard one-user, one-assistant interaction. The post is brief, so the linked demonstration is the primary source.

Google VP Josh Woodward 分享了一项“multiplayer AI”实验，指向一种不同于单用户与单助手对话的多人同步 AI 体验。帖子文字很短，具体信息以链接中的演示为准。

PODCASTS

AI & I by Every: A $10B Hedge Fund’s AI Playbook (Best of the Pod)

The Takeaway: Walleye's AI transformation works because CEO and CIO Will England treats adoption as an operating mandate backed by leadership behavior, incentives, training, and measurable workflows, not as a side experiment.

England runs a hedge fund with roughly $10 billion in assets and about 400 employees. His memo is deliberately blunt: “Using ChatGPT is not cheating,” and refusing useful AI because it is imperfect is comparable to refusing the internet in 1995. Walleye began an internal program two years earlier after an analyst demonstrated tools intended to automate much of his own work. That example shaped the firm's approach: find domain experts close to real workflows, give them access and institutional support, then build agents that augment judgment rather than merely produce demos.

The contrarian point is that AI adoption is primarily a leadership and incentive-design problem. England expects people to use AI proudly, leads from the front, and treats failure to prepare employees and investors as a breach of responsibility. He still sees an edge for humans in rare, fuzzy situations without enough historical precedent, while machines excel at finding nonlinear patterns beyond human comprehension. His personal workflow applies the same philosophy: he uses AI to turn quick voice notes or bullets into a daily journal covering family, work, and health, making longitudinal self-review nearly frictionless.

核心结论： Walleye 的 AI 转型之所以有效，是因为 CEO兼CIO Will England 把采用 AI 视为由领导行为、激励机制、培训和可度量工作流共同支撑的经营要求，而不是边缘实验。

England 管理着一家资产规模约 100 亿美元、约 400 名员工的对冲基金。他在内部备忘录中直言：“使用 ChatGPT 不是作弊”；因为 AI 尚不完美就拒绝使用，类似于 1995 年因为互联网不完美而拒绝上网。两年前，一名分析师展示了旨在自动化自己大部分工作的工具，Walleye 随即启动内部 AI 项目。这也奠定了公司的方法：找到最接近真实业务流程的领域专家，给予工具与组织支持，再构建增强判断力的 agent，而不是只做演示。

最反直觉的一点是，AI adoption 首先是领导力与激励设计问题。England 要求员工坦然使用 AI，自己带头实践，并把未能帮助员工和投资者迎接变化视为失职。他仍认为，在缺少历史先例、模糊且低频的情境中，人类具有优势；而机器擅长发现人脑无法理解的非线性模式。他的个人工作流也遵循同一逻辑：每天用语音或要点记录家庭、工作和健康，再由 AI 整理成日志，使长期自我复盘几乎没有摩擦。

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders