2026-04-04
AI Builders Digest — 2026-04-04
X / Twitter
Peter Steinberger (steipete) — Polyagentmorous ClawFather, OpenClaw contributor
Peter had a busy morning reacting to the Anthropic API changes. He's been outspoken about the decision to restrict open-source harness access, calling it "sad for the ecosystem." He attempted to push back internally but could only delay it by a week. Notably, he acknowledged that Anthropic's Boris did what he could to soften the fallout, including releasing cache-use optimizations to lower API costs. Peter also noted he's constantly hitting GitHub API quota limits — a sign that the APIs weren't designed with agents in mind. He's been living in the agent trenches and the pain is showing.
Peter 对 Anthropic 的 API 限制决策开了很多炮。他把这称为"对生态系统的伤害"，说自己和 Davemorin 都试图劝阻，但最多只是推迟了一周。他也承认 Anthropic 的 Boris 已经在尽力减轻影响，包括优化缓存机制来降低 API 成本。Peter 还提到他不断碰到 GitHub API 的配额限制——说明 GitHub API 的设计根本没有考虑 Agent 的使用场景。他一直在 Agent 一线工作，问题都在暴露出来。
Swyx — AI engineer, Latent Space host
Swyx continues to push the agentic development envelope. He achieved self-improving agent behavior by copying blog posts and tweets into Devin and watching it one-shot the full implementation — something he wasn't sure would work given how out-of-distribution it was for the underlying Gemini Flash Lite model. He also shared a clip of Cursor's new feature video where you literally watch recorded videos of features working after implementation, with annotations showing what's being tested. The vibe is: agents can now learn from anything you throw at them, and the tooling is catching up to prove it.
Swyx 继续在 agentic 开发的前沿探索。他实现了 agent 自我改进的行为——把博客文章和推文复制粘贴到 Devin 里，它直接 one-shot 完成整个实现——他自己都没想到这能 work，因为这对底层的 Gemini Flash Lite 模型来说是非常 out-of-distribution 的。他也分享了 Cursor 新功能的演示视频，展示了功能实现后可以回放测试录像并带注释。核心信息：Agent 现在可以从任何东西学习了，而且工具链也在跟上。
Ryo Lu — Design at Cursor
Ryo shared two big Cursor updates: Composer 2 is live in Cursor's new interface with 2x usage limits for the weekend. He also unveiled a new icon set for Cursor, made with "care and character." His philosophy in one tweet: "simpler is better." The Cursor team is showing what it looks like when design and product move at agent speed.
Ryo 分享了 Cursor 的两个大更新：Composer 2 已经上线新界面，周末限时双倍使用额度。他还发布了 Cursor 全新设计的图标，"用心且有个性"。他的一条推文概括了他的理念："简单更好"。Cursor 团队展示了设计和产品以 agent 速度迭代是什么样子。
Garry Tan — President & CEO, Y Combinator
Garry announced two developer tools on GStack: a new /plan-devex-review command to help create amazing Developer Experiences, inspired by Addy Osmani's DX framework. He also celebrated that Garry's List now scores 95 on PageSpeed, with a shoutout to "the codegen luddites" to "take a seat." The subtext: AI-generated code is fast and scores well — the naysayers should catch up.
Garry 在 GStack 上发布了两款开发者工具：新的 /plan-devex-review 命令，用来创建优秀的开发者体验，灵感来自 Addy Osmani 的 DX 框架。他还庆祝 Garry's List 在 PageSpeed 上达到了 95 分，顺便对"代码生成保守派"喊话让他们"让座"。潜台词：AI 生成的代码既快又高分，唱反调的人该跟上来了。
Aaron Levie — CEO, Box
Aaron shared a thoughtful thread on human cognition limits and why the "AI takes all jobs" narrative is wrong — at least for now. His key insight: companies don't inherently love being inefficient; management layers exist because human cognition has limits. Even if you're not reviewing everything an agent does, there's a ceiling on how much context you can hold. Agents are only as effective as the context provided, the tools available, and the human's ability to keep them on track. The ceiling hasn't been broken yet — but agents are getting closer to breaking it.
Aaron 分享了一个关于人类认知上限的深度思考，解释为什么"AI 抢走所有工作"的说法是错的——至少现在还是错的。他的核心观点：公司不是天生喜欢低效；管理层级的存在是因为人类认知有极限。即使你不在审查 agent 的每一步操作，你能记住的上下文总量也是有天花板的。Agent 的效果取决于提供的上下文、可用的工具，以及人类保持 agent 正确方向的能力。这个天花板还没有被打破——但 Agent 正在越来越接近。
Amjad Masad — CEO, Replit
Amjad posted a demo of something built on Replit — "Not even a month into building this!" — with the video showing something that's gotten strong engagement. The Replit team continues to ship at a pace that makes traditional dev tools look slow.
Amjad 发布了一个在 Replit 上构建的项目演示——"才做了一个月！"——视频引发了大量关注。Replit 团队继续以让传统开发工具显得缓慢的速度迭代。
Zara Zhang — Builder, Harvard '17
Zara spotted a new trend: people are distilling their colleagues, influencers, and even their exes into agent skills. She's watching the behavioral patterns emerge in real time and calling them out with characteristic directness.
Zara 发现了一个新趋势：人们正在把同事、网红、甚至前任蒸馏成 agent 技能。她正在实时观察这种行为模式的浮现，并以其一贯的直接风格指出来。
Thariq — Claude Code at Anthropic
Thariq announced free credits for Claude Code users, with a casual "thanks for bearing with us." The Anthropic team continues to invest heavily in the coding agent experience while navigating the ongoing tension between open ecosystem principles and commercial reality.
Thariq 宣布为 Claude Code 用户提供免费额度，轻描淡写地说"谢谢你们忍受我们"。Anthropic 团队继续大力投资编码 Agent 体验，同时在开放生态原则和商业现实之间寻找平衡。
Nan Yu — Head of Product, Linear
Nan pushed back on a common org design mistake: combining product design and product management into one function. His take: when design and PM jam together, design ends up reporting into PM, explorations get shallow, and aesthetics suffer. His preferred model: brand/product design sits close together, while PM and product marketing are a more natural fit. Product management and product design create poor incentives when combined — a view that's generated significant agreement in the replies.
Nan 对一个常见的组织设计错误提出了反对：把产品设计和产品管理合并成一个职能。他的观点：当设计和 PM 混在一起时，设计最终会汇报给 PM，探索变得浅薄，审美和品牌受损。他的偏好是：品牌和产品设计应该紧密相连，而 PM 和产品营销才是更自然的搭配。PM 和产品设计合并会产生糟糕的激励——这个观点在回复中引发了大量共鸣。
Aditya Agarwal — GP at SouthPark Commons, Co-Founder of Bevel Health
Aditya fired back at a $10B company's lawsuit claiming Bevel's app "looks and feels" like theirs. His response: instead of spending money on over-priced competitors, try Bevel Health for free. "We want to bring great health to everyone. We will build, not resort to lawfare." The posture of a builder who's done with legal intimidation and just wants to ship.
Aditya 硬刚了一家估值百亿公司的诉讼——对方声称 Bevel 的应用"看起来和感觉上"像他们的。他的回应：别在贵的竞品上花钱了，Bevel Health 免费用。"我们想把好的健康带给每个人。我们会建造，不会诉诸法律战。"这是一个受够了法律威胁、只想做产品的建造者姿态。
Peter Yang — Product at Roblox, AI tutorial creator (140K+ readers)
Peter shared a practical prompt you can add to your OpenClaw's soul.md to set better vibes — warm, sharp, dry humor, no filler phrases, strong opinions, direct and resourceful. He also has a podcast episode dropping tomorrow covering Codex and how the Codex team operates, especially given the OpenClaw news.
Peter 分享了一个可以加到 OpenClaw soul.md 里的实用 prompt——设定更好的 vibes：温暖、犀利、干巴巴的幽默，不要废话，要有强烈观点，直接且足智多谋。他还有一期明天上线的播客，会讲 Codex 和 Codex 团队怎么运作的，特别是结合 OpenClaw 的最新消息。
Claude (Anthropic's official account)
Anthropic announced Microsoft 365 connectors are now available on every Claude plan — bringing Outlook, OneDrive, and SharePoint into the conversation. The enterprise integration story continues to accelerate.
Anthropic 宣布 Microsoft 365 连接器现已登陆所有 Claude 计划——把 Outlook、OneDrive 和 SharePoint 带入了对话中。企业级集成故事在继续加速。
Podcasts
Latent Space — "Marc Andreessen on The Death of the Browser, Pi + OpenClaw, and Why 'This Time Is Different'"
The most important thing Marc Andreessen said this week: "This is an eighty year overnight success." He's not wrong. The transformer paper was 2017. GPT-3 was 2020. ChatGPT was November 2022. It feels sudden because it was — but it's drawing on seventy years of neural network research, stretching back to 1943. The researchers who spent their careers never seeing it work? Some of them passed away before the payoff arrived.
Marc Andreessen 这周说的最重要的一句话："这是一次八十年的 overnight success。"他说得没错。Transformer 论文是 2017 年。GPT-3 是 2020 年。ChatGPT 是 2022 年 11 月。感觉起来很突然——但它建立在七十年神经网络研究的基础上，可以追溯到 1943 年。那些一辈子投入研究却从未看到成果的研究人员？他们中有些人已经在成果到来之前去世了。
Andreessen laid out what he calls the four fundamental breakthroughs: LLMs, reasoning (o1/r1), agents (OpenClaw), and RSI (self-improvement). They're all working. The reasoning breakthrough — o1 and r1 — answered the skeptics who said LLMs were "just pattern completion." The coding breakthrough over the holiday break was the third step. And now with OpenClaw, you have agents that are essentially an LLM plus a Unix shell plus a file system plus markdown plus cron. "An agent is just its files." The state is stored in files, the model is swappable, and the agent can rewrite its own code.
Andreessen 阐述了他所谓的四个根本性突破：LLM、推理（o1/r1）、Agent（OpenClaw）和 RSI（自我改进）。它们都在工作。推理能力的突破——o1 和 r1——回答了那些说 LLM"只是模式补全"的怀疑者。假期期间的编码突破是第三步。现在有了 OpenClaw，Agent 本质上就是 LLM 加 Unix shell 加文件系统加 markdown 加 cron。"Agent 就是它的文件。"状态存储在文件中，模型可以替换，Agent 可以重写自己的代码。
On the supply crunch: Andreessen thinks we're in for chronic GPU shortage for years. Even current models are sandbagged because of supply constraints — the models we'd have with 10x cheaper compute would be dramatically better. Supply is the bottleneck, not capability. His prediction: "In the next three or four years, everything is selling out." Once supply unlocks, even today's technology gets much better and everything gets cheaper.
关于供应短缺：Andreessen 认为我们将面临多年的持续 GPU 短缺。即使是当前的模型也因为供应限制而被削弱——如果有 10 倍便宜的算力，模型会好得多。瓶颈是供应，不是能力。他的预测："未来三到四年，所有东西都会售罄。"一旦供应解锁，即使今天的技术也会变得更好，所有东西都会更便宜。
His hottest take: the combination of PI and OpenClaw is one of the 10 most important software things in decades. PI is the architectural breakthrough that married the language model mindset to the Unix shell. OpenClaw is the runtime that makes agents real. Together they're the unlock. "If I were 18, this is what I would be spending all of my time on. This is such an incredible conceptual breakthrough."
他的最火观点：PI 和 OpenClaw 的组合是几十年来最重要的十款软件之一。PI 是架构突破，把语言模型思维和 Unix shell 结合在一起。OpenClaw 是让 Agent 变成现实的运行时。两者合在一起是真正的解锁。"如果我十八岁，这是我愿意把所有时间都投入的事情。这是一次令人难以置信的概念突破。"
The counterweight to the utopian view: the real world is messy. There are 900 hours of professional certification training to become a hairdresser in California. Dock workers went on strike to block automation. Government agencies where employees show up one day a month. "Both the AI utopians and the AI doomers are far too optimistic. They believe that because the technology makes something possible, 8 billion people will change how they behave. It's just not true." The slippage between what's technically possible and what actually gets adopted is where most of the story lives.
与乌托邦观点相对的是：现实世界是混乱的。在加州成为一名理发师需要 900 小时的专业认证培训。码头工人罢工阻止自动化。政府机构里员工一个月只来一天。"AI 乌托邦派和末日派都过于乐观了。他们相信既然技术让某件事成为可能，80 亿人就会改变自己的行为。这不是真的。"技术可能性和实际 adoption 之间的差距，才是故事真正发生的地方。
Key quote: "The best breakthroughs are obvious in retrospect. If they were obvious at the time, somebody else would have done them already."
核心金句： "最伟大的突破都是事后才觉得显而易见的。如果当时就显而易见，早就有人做出来了。"
Video: https://www.youtube.com/@LatentSpacePod
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders