AI Builders Digest — 2026-06-19

X / TWITTER

Thibault Sottiaux, Codex and ChatGPT at OpenAI

Thibault Sottiaux highlighted two practical Codex updates: Codex App, CLI, and SDK can be used with open-source models, not only OpenAI models, and Codex users received a double reset, including one reset banked for later use. The important signal is not the reset itself, but OpenAI positioning Codex as a harness that can sit above multiple model providers instead of only as a closed OpenAI-model experience.

Thibault Sottiaux 强调了两个 Codex 实用更新：Codex App、CLI 和 SDK 可以配合 open-source models 使用，不只限于 OpenAI models；同时 Codex 用户获得了一次 double reset，其中一次还可以留到之后使用。真正重要的信号不是 reset，而是 OpenAI 正在把 Codex 定位成一个可接入多模型的 coding harness，而不只是封闭的 OpenAI 模型前端。

Links: https://x.com/thsottiaux/status/2067181377028538431, https://x.com/thsottiaux/status/2067399435009622521

Guillermo Rauch, CEO of Vercel

Guillermo Rauch argued that Vercel AI SDK is becoming more relevant as model competition intensifies. He pointed to GLM 5.2, an open model, surpassing Opus 4.8 in Vercel's Next.js Evals, then framed the bigger need as practical infrastructure for building and deploying agents, similar to how React and Next.js became the practical layer for web apps.

Guillermo Rauch 认为，随着模型竞争加剧，Vercel AI SDK 的重要性正在上升。他提到 open model GLM 5.2 在 Vercel 的 Next.js Evals 中超过 Opus 4.8，并把更大的问题定义为：行业需要一套真正可落地的 agent 构建与部署方案，就像 React 和 Next.js 曾经为 web app 提供了实用工程层。

Link: https://x.com/rauchg/status/2067242482190979186

Aaron Levie, CEO of Box

Aaron Levie pushed back on the idea that applied AI will be a thin wrapper over LLMs. His view: enterprise agentic workflows are complex enough that the application layer can build real moats through workflow depth, integrations, permissions, reliability, and domain-specific execution. For AI SaaS builders, this is the strongest business-model signal in today's feed.

Aaron Levie 反驳了“应用层只是 LLM 薄壳”的判断。他的观点是：enterprise agentic workflows 足够复杂，应用层可以通过 workflow 深度、集成、权限、可靠性和垂直领域执行能力形成真实护城河。对 AI SaaS 创业者来说，这是今天最重要的商业模式信号。

Link: https://x.com/levie/status/2067455756795039957

Garry Tan, President and CEO of Y Combinator

Garry Tan estimated that banning a frontier AI coding tool could cost roughly $12M per working hour in lost productivity, based on 5M AI-coding daily active developers and a modeled 2.7% throughput hit. He also argued that AI narrows the founder skill gap: technical founders gain access to business thinking, while business founders gain access to technical thinking, producing more startups that can actually work.

Garry Tan 估算，如果禁用某个 frontier AI coding tool，可能每个工作小时损失约 1200 万美元生产力。他的模型基于 500 万 AI coding 日活开发者和约 2.7% 的产出损失。他还认为 AI 正在缩小创始人的能力差距：技术创始人获得商业思维，商业创始人获得技术思维，因此会出现更多真正能跑起来的 startup。

Links: https://x.com/garrytan/status/2067366749411176831, https://x.com/garrytan/status/2067308407603048774

Zara Zhang, builder and AI product thinker

Zara Zhang warned against using AI for writing before developing taste and voice: if the model produces slop, you may not recognize it. She also made a sharp product point about vibe-coded personal apps: building may take a day, but discovering whether you will actually use the app takes a week, because most products assume a user who remembers, clicks correctly, and repeats the workflow every day.

Zara Zhang 提醒，不要在形成自己的 taste 和 voice 之前就依赖 AI 写作，因为如果模型产出 slop，你可能根本识别不出来。她还指出 vibe-coded personal apps 的关键问题：做出来可能只要一天，但验证自己是否真的会用需要一周，因为多数产品假设用户每天都记得打开、点对按钮、按流程执行，而真实的人经常懒、忘、断。

Links: https://x.com/zarazhangrui/status/2067423674689638652, https://x.com/zarazhangrui/status/2067313780724551853

Claude by Anthropic

Claude announced Claude Design beta for all paid plans on web and desktop. The stronger workflow signal is that Claude Design and Claude Code now connect both ways: users can hand designs off to build, or start from Claude Code and sync design projects from the terminal, with export paths to PDF, PowerPoint, and other tools.

Claude 宣布 Claude Design beta 面向所有 paid plans 开放，支持 web 和 desktop。更强的 workflow 信号是 Claude Design 和 Claude Code 开始双向联动：用户可以把设计交给 coding workflow，也可以从 Claude Code 出发，在 terminal 中同步 design projects，并导出到 PDF、PowerPoint 和其他工具。

Links: https://x.com/claudeai/status/2067325894268428560, https://x.com/claudeai/status/2067325893001826552, https://x.com/claudeai/status/2067325891781226581

Dan Shipper, CEO of Every

Dan Shipper said Every occasionally invests in AI founders they consider excellent, pointing to Tacit and its mission. He also linked the current excitement around fast AI progress back to his 2023 essay "Against Explanations," suggesting that AI may change how scientific progress happens.

Dan Shipper 表示 Every 会不定期投资他们认为优秀的 AI founders，并特别提到 Tacit 及其使命。他还把当前对 AI 快速进展的兴奋，连接回自己 2023 年的文章 “Against Explanations”，暗示 AI 可能改变科学进步的方式。

Links: https://x.com/danshipper/status/2067386342661624055, https://x.com/danshipper/status/2067386395283345808

Nan Yu, Head of Product at Linear

Nan Yu clarified that "taste" is not just aesthetic preference, just as "design" is not just visual design. The useful product-management read: taste is judgment across tradeoffs, behavior, and system quality, not surface polish.

Nan Yu 澄清，“taste” 不只是审美偏好，就像 “design” 不只是视觉设计。对产品管理更有用的理解是：taste 是对取舍、行为和系统质量的综合判断，而不是表面好不好看。

Link: https://x.com/thenanyu/status/2067327619897446721

Nikunj Kothari, Partner at FPV Ventures

Nikunj Kothari criticized tranched rounds because they can hurt the next employees who join: option strike prices may reflect headline valuations that preferred investors did not fully pay. His quick test: if a startup raises a large amount at under 10% dilution of the published valuation, it is likely a tranched round.

Nikunj Kothari 批评 tranched rounds，因为它们可能伤害下一批加入的员工：员工期权的 409A 可能按 headline valuation 走，但 preferred investors 实际并没有一次性按这个估值买入。他给出的快速识别方法是：如果一家 startup 宣布大额融资，但稀释比例低于 published valuation 的 10%，很可能是 tranched round。

Links: https://x.com/nikunj/status/2067399657639285150, https://x.com/nikunj/status/2067397092981772501

PODCASTS

AI & I by Every: GitHub’s COO Explains Why AI Hasn’t Replaced Developers

The takeaway: GitHub sees AI agents not as a replacement for developers, but as a force multiplying developer activity so aggressively that the platform itself has to evolve.

Kyle Daigle, COO of GitHub and Microsoft's CMO of Developer, described a sharp shift in who builds software and how much code gets created. GitHub is seeing non-traditional developers across legal, finance, and customer organizations use Copilot-style tools to build small apps and internal assets. At the same time, open-source maintainers are facing a flood of AI-generated pull requests, so GitHub is investing in controls, agentic code review, and agentic merge flows rather than imposing one universal moderation standard.

The most concrete numbers were striking: GitHub previously talked about one billion commits in a year, but Daigle said linear growth would put 2026 on track for 14 billion commits, with 17 million agent-created pull requests in March alone. His framing is that the world is moving from one developer building alone to one developer orchestrating many agents using that person's skills, resources, and context. Pricing and business models are still unresolved, but GitHub expects heavy agent users to need different limits from ordinary individual developers.

核心 takeaway：GitHub 并不认为 AI agents 会简单替代 developers，而是会把 developer activity 放大到平台本身必须重构的程度。

GitHub COO、Microsoft Developer CMO Kyle Daigle 描述了软件构建者和代码产量的双重变化。GitHub 看到法律、财务、客户团队等非传统开发者，也开始用 Copilot 类工具构建小应用和内部资产。同时，open-source maintainers 正被 AI-generated pull requests 淹没，所以 GitHub 更倾向于提供 controls、agentic code review 和 agentic merge flows，而不是替所有社区强行规定一套标准。

最具体的数字很有冲击力：GitHub 之前谈的是一年 10 亿 commits，但 Daigle 说如果线性增长，2026 年可能达到 140 亿 commits；仅 3 月就有 1700 万个 agent-created pull requests。他的核心判断是，世界正在从“一个 developer 独立写代码”转向“一个 developer 调度多个 agents，复用自己的 skills、resources 和 context”。定价和商业模式还没完全确定，但 GitHub 已经意识到 heavy agent users 与普通个人开发者需要不同的 limit 体系。

Link: https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL

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