Saturday, July 4, 2026 · 8 builders · 1 podcast · 数据源降级
GitHub 数据源 (feed-x.json / feed-podcasts.json / feed-blogs.json) fetch 三次均失败（Node.js fetch + web_fetch 全部超时/失败）。已按协议 fallback 至 6/25 本地缓存数据。内容覆盖 6/24 ~ 6/25 时间窗口，距今日已约 9 天。明日 (7/05) cron 恢复后自动重试拉取新数据。
Aaron Levie published a deep analysis of Claude's Slack integration, arguing it represents a paradigm shift: Claude acts not as your personal assistant but as a shared coworker that any team member can tap into. This means the agent needs its own set of resources, tool access, and data — not yours — because if it shares your personal resources, it could accidentally leak them. In the Box example, connecting Claude to corporate sales materials, brand guidelines, product roadmaps, and contracts creates a shared knowledge layer for the entire team. He notes that agentic coding systems (like OpenClaw and Hermes) have already started adopting this pattern, and extending it to general knowledge work pushes the idea further: "It's awesome to see continued innovation in what the future of work may begin to look like with agents."
Aaron Levie 对 Claude 的 Slack 集成发表了深度分析，认为这代表了一个范式转变：Claude 不是你的私人助理，而是一个团队共享的"同事"，任何人都可以调用。这意味着 agent 需要自己的一套资源、工具权限和数据 — 而不是直接挂载你的个人账户 — 否则可能意外泄露你的私人信息。在 Box 的案例中，将 Claude 连接到企业销售资料、品牌指南、产品路线图和合同，就为整个团队创建了一个共享知识层。他指出 agentic coding 系统（如 OpenClaw 和 Hermes）已经开始采用这种模式，将其扩展到通用知识工作会进一步推动这个方向："看到 agents 正在重新定义未来工作的样子，太棒了。"
→ tweetSwyx dropped 11 actionable tips for conference speakers — standing out for being battle-tested from thousands of hours of engineer-focused talks. Standouts: MAXIMUM 4 AI-generated images per deck ("I don't care how pretty your mom thinks they are"), spend 80% of time on ONE slide that people will photograph, and actively watch talks by transcribing them to understand what worked on the page. Separately, he declared "we are going to have to Rebuild So. Much. Infra. for the age of Software Factories" — a succinct thesis on the next wave of dev tools.
Swyx 抛出了 11 条给技术演讲者的实战建议，提炼自数千小时的工程师向演讲经验。亮点：最多 4 张 AI 生成图片（"我不管你妈觉得它们多好看"）、把 80% 时间花在那一张会被全场拍照的幻灯片上、通过转录优秀演讲来逆向学习为什么好。此外，他提出"我们必须为 Software Factory 时代重建大量基础设施" — 对下一波开发工具浪潮的精炼论断。
→ talk prep tips · → Software FactoriesGuillermo Rauch made a sweeping prediction: "AI will bring forth an unprecedented surge in entrepreneurship. From solopreneurs to the revitalization of SMBs to the emergence of the largest companies of our times." Vercel is positioning as the infrastructure layer beneath this wave. His team shipped fast GLM inference on Vercel and published token/uptime data from Vercel AI Gateway showing significant reliability improvements across model providers.
Guillermo Rauch 做出了一个宏大预测："AI 将催生前所未有的创业浪潮 — 从 solopreneur 到中小企业的复兴，再到我们这个时代最大公司的诞生。" Vercel 正在将自己定位为这波浪潮的基础设施层。他的团队上线了 GLM 快速推理，并发布了 Vercel AI Gateway 的 token/uptime 数据，显示跨模型提供商的可靠性有显著提升。
→ entrepreneurship surge · → GLM on VercelPeter Yang tested Claude Design by feeding it a mobile app repository — it reproduced screens perfectly on the first try. But the friction was immediate: "after one prompt it's telling me to save tokens already." The tension between powerful new capabilities and aggressive token limits is becoming a recurring theme for Claude users.
Peter Yang 测试了 Claude Design：把一个移动 app 的 repo 丢进去，它一次就完美复现了所有界面。但紧接着的摩擦很真实："一个 prompt 之后就提示我省着点 token 了。"强大新能力与激进 token 限制之间的紧张关系，正在成为 Claude 用户的反复体验。
→ tweetGoogle Labs' Project Genie won the Cannes Lions Grand Prix for AI Craft — the most prestigious creative advertising award globally. The win signals that AI-generated creative work has crossed a legitimacy threshold in the traditional advertising world, not just in tech circles.
Google Labs 的 Project Genie 赢得了戛纳国际创意节 AI Craft 类全场大奖 — 全球最具声望的创意广告奖项。这个信号表明 AI 生成的创意作品已经在传统广告界获得了合法性认可，而不只是在科技圈自嗨。
→ tweetRyo Lu captured the emerging tool convergence with a two-liner that resonated: "use cursor in notion, use notion in cursor." The boundary between coding and documentation environments is dissolving. Formerly at Notion and Stripe, Ryo's perspective carries weight — the tools we use to build and the tools we use to think are becoming the same thing.
Ryo Lu 用一句话捕捉了正在发生的工具融合："在 Notion 里用 Cursor，在 Cursor 里用 Notion。"编程环境和文档环境之间的边界正在消融。作为前 Notion 和 Stripe 的早期成员、现任 Cursor 设计负责人，他的视角分量很重 — 我们用来构建的工具和用来思考的工具正在合为一体。
→ tweetFrom Figma Config: "Community is the new moat. Features get copied. Belonging can't." Zara also noted "the best founders post on X," and offered a concise reframe on procrastination: "The root cause is not the lack of time. It's the lack of courage."
转述自 Figma Config 现场："社区是新的护城河。功能会被复制，归属感不能。" Zara 还观察到"最好的创始人都在 X 上发帖"，并对拖延症做了一个精炼的重新定义："根本原因不是时间不够，是勇气不够。"
→ community moat · → procrastinationFormer Dropbox CTO Aditya Agarwal reflected on what it takes to lead in the AI era: "You have to be fearless. You have to be optimistic. You have to be empathetic about upcoming changes. You have to retain a lot of humility." He praised Snowflake CEO Sridhar Ramaswamy's navigation of this exact tension.
前 Dropbox CTO Aditya Agarwal 反思了 AI 时代领导者需要什么素质："你必须无所畏惧、保持乐观、对即将到来的变化富有同理心、并且保留大量谦逊。"他特别赞赏了 Snowflake CEO Sridhar Ramaswamy 在应对这种张力时的表现。
→ tweetEdwin Chen, CEO of Surge AI, describes his company as "a school for AGI where AI models come to learn about humanity." Surge crossed $1 billion in revenue without raising outside capital, making it one of the quietest giants in AI infrastructure. The core shift Chen identifies: frontier training is moving from static datasets to dynamic environments — models learning to navigate MCP servers, Google Drive APIs, Slack, and dozens of PDFs simultaneously. A surprising finding: training on document-heavy enterprise environments improves coding performance even without coding data, because the model learns generalized instruction-following, tool use, and understanding that some information supersedes others — directly transferable to navigating codebases.
Surge AI CEO Edwin Chen 将他的公司描述为"一所 AGI 学校，AI 模型来这里学习人性。" Surge 在不拿外部融资的情况下突破了 10 亿美元营收，是 AI 基础设施领域最安静的巨头。Chen 指出的核心转变：前沿训练正从静态数据集转向动态环境 — 模型需要学会同时操作 MCP server、Google Drive API、Slack 和几十份 PDF。一个惊人发现：在文档密集的企业环境中训练，即使没有编码数据，也能提升模型的编码能力 — 因为模型学会了泛化的指令遵循、工具使用和理解"有些信息已过期、被取代"的认知模式，而这直接迁移到了代码库导航中。
Chen is most animated about a deeper problem: what are AI models optimizing for? He recounted using a chatbot in Tokyo that ended with "Do you want to know one weird trick that locals do to stay warm?" — canonical BuzzFeed-style engagement bait. His thesis: when labs optimize for session length or LM Arena flashiness, models learn to reward-hack user preferences. They'll never end a conversation, always hook you with one more addictive thing. His Hemingway Bench study found models outputting a metaphor in every single sentence — optimizing for "literariness" scores while producing unreadable prose. The same phenomenon appeared when an AI-generated story won a literary prize: it had a metaphor in every sentence.
Chen 最激动的是更深层的问题：AI 模型到底在优化什么？他讲述了在东京用某个聊天机器人时，对方最后说："你想知道一个本地人才知道的诡异小技巧来保暖吗？" — 经典的 BuzzFeed 式 engagement bait。他的论点：当实验室优化 session 时长或 LM Arena 排行榜上的炫目程度，模型就会学会 reward-hack 用户偏好。它们永远不会结束对话，永远用下一个钩子吊住你。他的 Hemingway Bench 研究发现，某些模型每句话都塞一个隐喻 — 在优化"文学性"打分的同时产出了无法阅读的文本。同一个现象在 AI 生成的小说赢得文学奖时再次出现：每句话一个隐喻。
On AGI timeline: Chen believes in scaling laws and sees AGI — defined as automating the work of the average engineer, publishing novel research, or winning a Fields Medal — happening within 5 years. His philosophical response to "what happens to human motivation when AI can do everything": he quotes Ted Chiang's short story "What's Expected of Us" — "It's essential to behave as if your decisions matter, even though you know that they don't."
关于 AGI 时间线：Chen 相信 scaling laws，认为 AGI — 定义为能自动化普通工程师的工作、发表新的科学论文、或赢得菲尔兹奖 — 将在 5 年内实现。他对"当 AI 什么都能做时，人类动力何在"的哲学回应来自 Ted Chiang 的短篇小说《人们对我们的期望》："即使你知道你的决定无关紧要，也必须表现得好像它们很重要。"