AI Builders Digest: 2026-06-25
X / TWITTER
Andrej Karpathy
Andrej Karpathy framed Claude Tag as the third major LLM UI paradigm: first, a website you visit; second, a desktop app; third, a persistent, asynchronous teammate with organization-wide tools, context, memory, permissions, and compute. His point is that the hard part is not the chat surface, but the under-the-hood engineering that lets an AI safely join normal team workflows.
Andrej Karpathy 把 Claude Tag 定义为 LLM UI 的第三次范式变化：第一代是你主动打开的网站，第二代是装在电脑上的 app，第三代是一个有组织级工具、上下文、memory、权限和计算环境的异步团队成员。他强调真正难的不是聊天界面，而是让 AI 安全进入团队日常工作流的底层工程。
Source: https://x.com/karpathy/status/2069547676849557725
Swyx
Swyx highlighted Zai's rise after its Hong Kong IPO, noting that GLM has moved from a model few people around him used to a leading open model that is now returning to the San Francisco AI builder scene "open for business." The signal is that Chinese open models are no longer peripheral in the frontier model conversation.
Swyx 提到 Zai 在香港 IPO 之后的上升势头，指出 GLM 已经从身边少有人使用的模型，变成了顶级 open model，并正在以更强姿态回到旧金山 AI builder 圈层。这个信号说明：中国 open model 已经不再是 frontier model 讨论里的边缘变量。
Sources: https://x.com/swyx/status/2069598378191941835, https://x.com/swyx/status/2069665232822366577
Josh Woodward
Google Labs VP Josh Woodward shared two Gemini ecosystem signals: a consumer image-template campaign inside Gemini App, and Florida State University's report that NotebookLM changed student study habits quickly after campus adoption. The more interesting builder signal is that Google is pushing both playful consumer creation and serious education workflows through the same Gemini surface.
Google Labs VP Josh Woodward 分享了两个 Gemini 生态信号：Gemini App 里的消费级图片模板活动，以及 Florida State University 反馈 NotebookLM 在校园落地后快速改变学生学习习惯。更值得看的 builder 信号是：Google 正在用同一个 Gemini 入口同时推进轻量创作和严肃教育工作流。
Sources: https://x.com/joshwoodward/status/2069408025362714957, https://x.com/joshwoodward/status/2069406832523624696
Boris Cherny
Anthropic's Boris Cherny described Claude Tag as a Slack-native coding agent: tag Claude in a channel, give it a task, and it spins up an isolated sandbox that can clone repos, write code, test, compile, and then disappear. He also uses Claude to monitor channels proactively, answer questions, draft PRs, and react with status emojis when threads are resolved.
Anthropic 的 Boris Cherny 把 Claude Tag 描述成 Slack 原生的 coding agent：在频道里 tag Claude，给它一个任务，它会启动隔离 sandbox，clone repo、写代码、测试、编译，完成后销毁环境。他还让 Claude 主动监控频道、回答问题、起草 PR，并用状态 emoji 标记线程处理结果。
Sources: https://x.com/bcherny/status/2069474691010707486, https://x.com/bcherny/status/2069474689819480394, https://x.com/bcherny/status/2069474688619958517
Thibault Sottiaux
OpenAI's Thibault Sottiaux posted quick Codex product-loop notes: Codex had a bug, the team fixed it, and they want more feedback. The useful signal is not the specific bug, but the fast public feedback loop around coding-agent reliability.
OpenAI 的 Thibault Sottiaux 发了几条 Codex 产品反馈闭环：Codex 有 bug，团队修了，并继续要更多反馈。这里的重点不是某个 bug 本身，而是 coding agent 可靠性正在通过公开、高频反馈快速迭代。
Sources: https://x.com/thsottiaux/status/2069579993588625574, https://x.com/thsottiaux/status/2069592160966733853
Peter Yang
Peter Yang asked the design question that matters for agentic software: what is design when the user of your product may be an agent looking for an API or CLI? He also connected human-agent interaction to managing a highly capable employee, which points toward new product primitives like task reviews, permissions, and status check-ins.
Peter Yang 提出了 agentic software 的关键设计问题：当访问产品的不是人，而是寻找 API 或 CLI 的 agent 时，设计到底是什么？他还把 human-agent interaction 类比成管理一个高能力员工，这会导向新的产品原语，比如任务 review、权限、状态同步。
Sources: https://x.com/petergyang/status/2069603490524254473, https://x.com/petergyang/status/2069530765352907180
Madhu Guru
Madhu Guru argued that the token-pricing debate is really a fight over where value accrues: model labs, application layer, distribution, or other layers. His view is that enterprises, agent startups, data providers, and labs are all still discovering business models, moats, value exchange, and execution playbooks in real time.
Madhu Guru 认为 token pricing 的争论，本质是价值会流向哪里：model lab、应用层、分发渠道，还是其他层。他的判断是，企业、agent startup、数据提供商和大模型实验室都还在实时摸索商业模式、护城河、价值交换和执行打法。
Source: https://x.com/realmadhuguru/status/2069455097193697393
Cat Wu
Anthropic's Cat Wu said Claude Tag is already changing how Anthropic ships products internally, claiming their internal version merges 65% of product PRs. She emphasized that Claude Tag is multiplayer and proactive, and shared guides for configuring permissions and common workflows.
Anthropic 的 Cat Wu 说 Claude Tag 已经改变了 Anthropic 内部发产品的方式，内部版本合并了 65% 的产品 PR。她强调 Claude Tag 的两个关键词是 multiplayer 和 proactive，并分享了权限配置与常见工作流指南。
Sources: https://x.com/_catwu/status/2069473118742331608, https://x.com/_catwu/status/2069484330938998993, https://x.com/_catwu/status/2069486403696869555
Thariq
Anthropic's Thariq shared operational patterns for Claude Tag: keep a pinned status message for many simultaneous threads, ask Claude to react with clear status emojis, and even create a scheduling channel where Claude helps find calendar time. This is the practical layer of agent management, not just model capability.
Anthropic 的 Thariq 分享了 Claude Tag 的实际运营方法：多线程时让 Claude 维护 pinned status message，用清晰的 emoji 表示任务状态，甚至建一个 scheduling channel 让 Claude 帮忙找会议时间。这是 agent 管理的实践层，而不只是模型能力展示。
Sources: https://x.com/trq212/status/2069474339679052144, https://x.com/trq212/status/2069474342220820657, https://x.com/trq212/status/2069474343512617390
Amjad Masad
Replit CEO Amjad Masad reacted to a launch by saying it starts with a prompt, but it takes a lot of work to get there. The useful builder reminder: prompt-first products can look magical from the outside, but the actual product is the accumulated engineering behind the prompt.
Replit CEO Amjad Masad 对一个发布的评论是：一切从 prompt 开始，但走到这里背后有大量工作。这个 builder 提醒很重要：prompt-first 产品外观看起来很神奇，真正的产品壁垒是 prompt 背后累积出来的工程系统。
Source: https://x.com/amasad/status/2069588152285794373
Guillermo Rauch
Vercel CEO Guillermo Rauch is recruiting serious agent builders into a private X Chat group for feedback on eve, especially people with high-quality critical feedback or concrete requirements. He also amplified Vercel + Cursor, reinforcing Vercel's positioning around AI coding and deployment workflows.
Vercel CEO Guillermo Rauch 正在邀请深度 agent builder 加入一个 X Chat 群，为 eve 提供高质量批评和需求反馈。他还转发了 Vercel + Cursor，继续强化 Vercel 在 AI coding 与部署工作流中的位置。
Sources: https://x.com/rauchg/status/2069590431646769472, https://x.com/rauchg/status/2069513849578082474
Alex Albert
Anthropic researcher Alex Albert said the new Claude workflow feels less like using a tool and more like managing a team. That sentence captures the core product shift across the day: the interface is moving from command-response to delegated, persistent work.
Anthropic researcher Alex Albert 说新的 Claude 工作方式不再像使用工具，而更像管理一个团队。这句话抓住了今天的核心产品变化：界面正在从命令-响应，转向可委派、可持续运行的工作系统。
Source: https://x.com/alexalbert__/status/2069470389391241314
Aaron Levie
Box CEO Aaron Levie sees a barbell forming in AI pricing: expensive frontier models on one side, cheap but good open or closed-weight models on the other. He argues the applied AI layer wins by routing across models, tuning for real workflows, adding evals, setting up data well, and using domain-specific FDEs. He also pointed to Claude Tag in Box as headless software that turns enterprise content into a portable knowledge base from Slack.
Box CEO Aaron Levie 认为 AI pricing 正形成杠铃结构：一端是高成本 frontier model，另一端是便宜但足够好的 open 或 closed-weight model。他判断应用层的机会在于跨模型路由、围绕真实 workflow 调优、做 eval、整理数据，并用 domain-specific FDE 深入客户流程。他还把 Box 里的 Claude Tag 看作 headless software，让企业内容可以在 Slack 中变成可调用知识库。
Sources: https://x.com/levie/status/2069639600310767616, https://x.com/levie/status/2069596515560267891
Garry Tan
Y Combinator CEO Garry Tan highlighted Linzumi as "Codex but actually multiplayer" and called it magical for teams. He also argued that AI will make usable data grow exponentially, pressuring storage products like Dropbox to rethink plan limits.
Y Combinator CEO Garry Tan 重点提到 Linzumi，称它是“真正 multiplayer 的 Codex”，对团队很有魔力。他还指出 AI 会让可用数据指数级增长，像 Dropbox 这样的存储产品需要重新思考容量限制。
Sources: https://x.com/garrytan/status/2069474420113146355, https://x.com/garrytan/status/2069434452628185241
Nikunj Kothari
FPV Ventures partner Nikunj Kothari launched Plug That Shop, built with Conductor, Anthropic Opus 4.8, Shopify UCP CLI, and Railway. The tool creates contextual embedded shops for websites, turning Shopify's command-line product access into a one-click buyer experience.
FPV Ventures partner Nikunj Kothari 发布了 Plug That Shop，技术栈包括 Conductor、Anthropic Opus 4.8、Shopify UCP CLI 和 Railway。这个工具把 Shopify 的命令行商品访问能力，转成网站上的上下文嵌入商店和一键购买体验。
Sources: https://x.com/nikunj/status/2069534712763490668, https://x.com/nikunj/status/2069547206504566980
Peter Steinberger
Peter Steinberger amplified the Google Workspace CLI story, joking that Google fired the person who made it because they made it. Under the joke is a serious builder signal: internal CLI and agent-facing tools are becoming strategically important, even when large organizations do not yet know how to value them.
Peter Steinberger 转发了 Google Workspace CLI 的故事，调侃 Google 因为员工做出了 Google Workspace CLI 而开除了他。玩笑背后是一个严肃信号：面向内部、面向 agent 的 CLI 工具正在变得有战略价值，但大组织还未必知道如何评价这类工作。
Source: https://x.com/steipete/status/2069594195522941059
Aditya Agarwal
South Park Commons partner Aditya Agarwal shared SPC's investment in Longshot Space, a company building a cannon that fires payloads into orbit. This is outside pure AI, but relevant as a deep-tech founder signal: AI-era infrastructure investing is not only software, it also includes extreme physical-world leverage.
South Park Commons partner Aditya Agarwal 分享了 SPC 对 Longshot Space 的投资，这家公司正在做把 payload 打入轨道的 cannon。它不属于纯 AI，但作为 deep-tech founder 信号仍有价值：AI 时代的基础设施投资不只在软件，也包括高杠杆的物理世界技术。
Sources: https://x.com/adityaag/status/2069464862556619263, https://x.com/adityaag/status/2069464865568166180
Claude
Claude announced Claude Tag beta for Slack on Claude Enterprise and Team plans. The product lets a single Claude participate in a channel, build shared context, take initiative through ambient behavior, follow up on stale threads, and surface relevant work across channels and tools.
Claude 官方宣布 Claude Tag 在 Claude Enterprise 和 Team 计划中进入 Slack beta。这个产品让一个 Claude 进入频道、建立共享上下文，并通过 ambient behavior 主动跟进沉寂线程、从多个频道和工具中提取相关信息。
Sources: https://x.com/claudeai/status/2069468701548531895, https://x.com/claudeai/status/2069468699766005847, https://x.com/claudeai/status/2069468698071494976
PODCASTS
No Priors: Biohub: The Future of Biology is Open-Source with Co-Founders Mark Zuckerberg, Priscilla Chan, and Head of Science Alex Rives
The Takeaway: Biohub is betting that biology becomes an engineering discipline when frontier AI is tightly coupled with frontier wet-lab science and open-source tooling.
Mark Zuckerberg, Priscilla Chan, and Biohub head of science Alex Rives described Biohub as a long-term tool-building institution for biology, not a traditional disease-cure lab. Their core diagnosis is that science moves slowly because labs work in silos, data is hard to share, and useful tools often vanish when a postdoc leaves. Biohub's answer is shared infrastructure: open tools, large biological datasets, and AI models that can learn from protein, cell, and system-level data. The most specific insight is that biology lacks the internet-scale corpus that trained LLMs, so the lab must create new measurement methods and datasets rather than merely train on what already exists. As Zuckerberg put it, the theory is not that Biohub alone cures diseases, but that it accelerates the whole field.
核心判断：Biohub 的赌注是，当 frontier AI 与 frontier wet-lab science、open-source tooling 紧密结合时，biology 会从发现型科学转向工程型科学。
Mark Zuckerberg、Priscilla Chan 和 Biohub head of science Alex Rives 把 Biohub 描述成一个长期建设生物学工具的平台，而不是传统意义上的治病实验室。他们对科学低效的诊断很具体：实验室彼此割裂，数据难以共享，很多有用工具随着博士后离开就消失。Biohub 的解法是共享基础设施：开放工具、大规模生物数据集，以及能从 protein、cell、system-level 数据中学习的 AI model。最关键的洞察是：生物学没有像互联网文本那样的 LLM 训练语料，所以 Biohub 必须发明新的测量方法和数据集，而不是只训练已有数据。Zuckerberg 的意思很清楚：他们不是要独自治愈疾病，而是要加速整个科学共同体。
Source: https://www.youtube.com/@NoPriorsPodcast
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders