AI Builders Digest — 2026-07-10
Stats: xBuilders 20, totalTweets 40, podcastEpisodes 1.
X / TWITTER
Swyx, AI engineer and Latent Space co-host
Swyx called out a design signal in Theo's keynote: hand-drawn Excalidraw/tldraw-style slides can feel more human and thoughtful than polished AI-generated presentation aesthetics. He also noted that many agent labs avoid discussing Chinese model usage because of government and defense sales concerns, while Cognition appears to have done the harder production work: multilingual propaganda/censorship evals, post-training correction, and cheap high-throughput serving.
Swyx 点出了一个产品表达信号：Theo keynote 里手绘风格的 Excalidraw/tldraw slides，比常见的 AI 生成式“精致商务感”更有人味、更可信。他还提到，很多 agent lab 因为要卖给政府和国防客户，会回避使用中国模型；但 Cognition 看起来做了更难的工程化工作：多语言宣传/审查 eval、post-training 修正，以及低成本高吞吐 serving。
Links: https://x.com/swyx/status/2074953099748450346, https://x.com/swyx/status/2074919183947808881
Josh Woodward, VP at Google Labs / Gemini
Josh Woodward asked users what they are surprised Gemini still cannot do well and should have fixed long ago. This is a useful product signal: Gemini is actively fishing for the "obvious broken expectations" that matter more than flashy demos.
Josh Woodward 直接向用户询问：Gemini 哪些本该早就做好的能力仍然不够好。这是一个值得关注的产品信号：Gemini 在主动收集“用户以为理所当然、但产品还没做到”的落差，而这类问题通常比炫技 demo 更影响留存。
Link: https://x.com/joshwoodward/status/2074847444823674883
Boris Cherny, Claude Code at Anthropic
Boris Cherny announced /checkup for Claude Code. It cleans unused skills, MCPs, and plugins, deduplicates local and checked-in CLAUDE.md files, splits root instructions into nested CLAUDE.md files and skills, disables slow hooks, updates Claude Code, enables auto mode, and pre-approves commonly denied read-only commands, with confirmation before changes.
Boris Cherny 发布了 Claude Code 的 /checkup。它会清理未使用的 skills、MCPs、plugins，去重本地和仓库里的 CLAUDE.md，把 root CLAUDE.md 拆成嵌套 CLAUDE.md 和 skills，关闭慢 hooks，更新 Claude Code，开启 auto mode，并预批准常被拒绝的只读命令；所有变更前都会确认。
Link: https://x.com/bcherny/status/2074997570317779038
Thibault Sottiaux, Codex and ChatGPT at OpenAI
Thibault Sottiaux signaled an active shipping week at OpenAI, pointing people to a 10am livestream after a reset week. The late-night office food post was noise, but the useful signal is cadence: OpenAI's Codex/ChatGPT side appears to be in release mode this week.
Thibault Sottiaux 表示 OpenAI 本周进入 shipping 节奏，并引导大家关注 10 点 livestream。深夜外卖本身不是重点，真正的信号是节奏：OpenAI 的 Codex/ChatGPT 线这周明显处在发布状态。
Link: https://x.com/thsottiaux/status/2074885402918601082
Nan Yu, Head of Product at Linear
Nan Yu framed product marketing as a story designed to be retransmitted, not just received: PMM to sales, customer PMM, customer, peer, user, buyer, champion, and org. That is a compact way to think about B2B messaging: the story must survive multiple handoffs.
Linear 的 Head of Product Nan Yu 把 product marketing 定义成“要被继续转述的故事”，而不是只给眼前听众看的材料：PMM 到 sales、客户 PMM、客户、同行、用户、buyer、champion、组织。对 B2B 来说，这是很实用的判断：信息必须能跨多次转手仍不失真。
Link: https://x.com/thenanyu/status/2074907752829223043
Cat Wu, Claude Code and cowork at Anthropic
Cat Wu previewed a walkthrough from single-player Claude Code to multiplayer Claude Tag. The product direction is explicit: AI moves from completing sentences, to writing features, to monitoring team channels, doing proactive work, accepting steering from the whole team, and remembering prior instructions.
Cat Wu 预告了从单人 Claude Code 走向多人 Claude Tag 的 walkthrough。方向很清楚：AI 从补全句子，到写完整功能，再到监听团队 channel、主动做事、接受整个团队 steering，并记住上周说过的话。
Link: https://x.com/_catwu/status/2074925531519468012
Thariq, Claude Code at Anthropic
Thariq argued that rewrites can now be good, cheap, and fast, and that this should update how engineers think about software work. His caveat matters: Bun is unusually testable and verifiable, but models are improving at filling the verification gap for less clean systems.
Anthropic 的 Thariq 认为，现在 rewrite 可以又好、又便宜、又快，这应该改变软件工程师的模型。他也保留了关键限定：Bun 这类项目特别容易测试和验证，但模型会继续补齐那些不那么干净的系统里的验证缺口。
Link: https://x.com/trq212/status/2074993112217461020
Amjad Masad, CEO of Replit
Amjad Masad pushed back on comparing autonomous agents to hand-written code, arguing that compilers are not judged against engineers writing assembly by hand. He also floated CAD/3D modeling as a possible Replit expansion, hinting that browser-native creation environments may move beyond software into design and modeling.
Replit CEO Amjad Masad 反对一直拿 autonomous agents 和手写代码做比较：就像我们不会拿 compiler 去和工程师手写 assembly 对比。他还抛出了 Replit 是否应该加入 CAD/3D modeling 的问题，暗示 browser-native creation environment 可能从软件扩展到设计和建模。
Links: https://x.com/amasad/status/2075080984211624154, https://x.com/amasad/status/2075003156745089264
Guillermo Rauch, CEO of Vercel
Guillermo Rauch argued that AI will make all software native, with stronger platform affinity and uncompromising performance. He also announced Grok 4.5 availability for Vercel customers and showed enthusiasm for an agent stack coming together for personal productivity agents.
Vercel CEO Guillermo Rauch 认为，AI 会让所有软件都变得 native，更强调平台亲和性和极致性能。他还宣布 Grok 4.5 已面向 Vercel customers 可用，并表示 agent stack 的拼图正在合拢，将用于个人 productivity agents。
Links: https://x.com/rauchg/status/2075018147330232707, https://x.com/rauchg/status/2074920996201796067, https://x.com/rauchg/status/2074874713143460150
Aaron Levie, CEO of Box
Aaron Levie said the newest AI models are getting much better at complex knowledge-worker tasks, especially in legal, professional services, healthcare, and other sophisticated work domains. His read on Grok 4.5 is cost plus performance, and the larger enterprise implication is more leverage over business documents and domain-specific data.
Box CEO Aaron Levie 认为，最新 AI models 在复杂知识工作任务上进步很快，尤其是 legal、professional services、healthcare 等专业领域。他对 Grok 4.5 的判断是 cost + performance；更大的企业级含义是，AI 将进一步释放文档和垂直领域数据的价值。
Link: https://x.com/levie/status/2075073587015516228
Ryo Lu, Design at Cursor
Ryo Lu called Grok 4.5 in Cursor the start of a new era and asked users to try it. The signal is less the wording and more the distribution: frontier model releases are now instantly becoming UX events inside coding tools.
Cursor 的 Ryo Lu 称 Cursor 中的 Grok 4.5 是“新时代的开始”，并邀请用户试用。这里真正的信号不是措辞，而是分发方式：frontier model release 正在立刻变成 coding tool 里的 UX event。
Link: https://x.com/ryolu_/status/2074951992884244606
Zara Zhang, builder and OpenClaw founder
Zara Zhang highlighted a team-culture failure mode of agent-heavy work: a founder bought Codex Max for everyone, people got most work done by talking to Codex all day, and human collaboration declined. Her conclusion is that enterprise AI needs to move beyond single-player human-agent work into human-human-agent collaboration. She also said frontend design is the single thing preventing her from using Codex more frequently.
Zara Zhang 提醒了 agent-heavy 工作方式的团队文化风险：一个 founder 给所有成员买了 Codex Max，大家整天和 Codex 对话完成工作，结果人和人之间反而不怎么沟通，会议取消，协作变少，团队文化变差。她的结论是，enterprise AI 不能停在 single-player human-agent collaboration，必须走向 human-human-agent collaboration。她还直说，frontend design 是阻止她更频繁使用 Codex 的单一最大问题。
Links: https://x.com/zarazhangrui/status/2075004775436005687, https://x.com/zarazhangrui/status/2075003007520096416
Nikunj Kothari, partner at FPV Ventures
Nikunj Kothari argued that polished output is becoming correlated with AI slop, pushing the market back toward raw and human expression. He also noted the rapid developer pendulum swing between Codex and Claude Code, with each release flipping sentiment from "we are dead" to "we are back."
FPV Ventures 的 Nikunj Kothari 认为，“精致感”正在快速和 AI slop 绑定，市场可能会重新回到更 raw、更有人味的表达。他还观察到开发者在 Codex 和 Claude Code 之间快速摇摆，每次发布都会让情绪从“我们完了”切到“我们又回来了”。
Links: https://x.com/nikunj/status/2075033190708961675, https://x.com/nikunj/status/2074878958525657452
Peter Steinberger, OpenClaw and OpenAI
Peter Steinberger clarified that OpenAI hired him, not OpenClaw, and that the OpenClaw Foundation remains independent with sponsors rather than owners. He also showed an agent interaction pattern where nameplates provide additional context when user input is needed.
Peter Steinberger 澄清：OpenAI hired him, not OpenClaw；OpenClaw Foundation 仍然独立，是 sponsors 而不是 owners。他还展示了一个 agent interaction pattern：当需要用户输入时，用 nameplate 提供额外上下文。
Links: https://x.com/steipete/status/2075046949896736835, https://x.com/steipete/status/2074969319042363808
Dan Shipper, CEO of Every
Dan Shipper teased a prompt and open-source repo for Every subscribers, tied to a content workflow that raised the stakes for AI-assisted publishing. The signal fits Every's current positioning: packaged AI workflows are becoming part of paid media products, not just free blog posts.
Every CEO Dan Shipper 预告会给 Every subscribers 发布一个 prompt 和 open-source repo，围绕的是一套提高 AI-assisted publishing 上限的内容工作流。这个信号很符合 Every 的定位：成体系的 AI workflow 正在成为付费媒体产品的一部分，而不只是免费博客文章。
Link: https://x.com/danshipper/status/2074882061869961585
Aditya Agarwal, General Partner at South Park Commons
Aditya Agarwal urged founders not to waste the current moment and to be maximally ambitious. His sharper point: pure software may no longer capture value the same way it did five years ago, and SPC wants founders working with atoms, hardware, science, biohacking, and heretical software theses.
South Park Commons 的 Aditya Agarwal 提醒 founders 不要浪费当前窗口，要“maximally ambitious”。更尖锐的判断是：pure software 不再像五年前那样捕获价值；SPC 更想要接触 atoms、hardware、science、biohacking 的 founders，或者拥有足够 heretical software thesis 的人。
Links: https://x.com/adityaag/status/2074892507306238235, https://x.com/adityaag/status/2074892952233705956
PODCASTS
AI & I by Every: How a Writer Uses AI Without Losing His Voice
The takeaway: the most advanced AI workflow may require deliberately protecting the non-AI parts of your mind. The guest describes using AI heavily for building software around a membership business, including replacing expensive newsletter infrastructure and hardening an older Flask app, while also setting strict barriers against internet and phone dopamine so writing remains connected to a deeper creative self. The most useful business lesson is brutal: if legacy SaaS stops innovating and keeps squeezing customers, motivated users can now rebuild the workflow with frontier models, cheap infrastructure, and enough taste.
核心 takeaway：真正高级的 AI workflow，可能反而需要刻意保护你脑子里“不使用 AI 的部分”。嘉宾一方面大量用 AI 为会员业务构建软件，包括替代昂贵的 newsletter infrastructure、加固旧 Flask app；另一方面又严格隔离手机和互联网的 dopamine，避免写作失去和深层创作自我的连接。最值得创业者注意的是这个现实：如果 legacy SaaS 长期不创新还持续榨取客户，今天有动机的用户已经可以用 frontier models、便宜基础设施和自己的品味把 workflow 重建出来。
Link: https://www.youtube.com/watch?v=7ND0lQmLJlA
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders