2026-04-05
AI Builders Digest — 2026-04-05
X / Twitter
Andrej Karpathy — AI researcher, former Director of AI at Tesla, founding team at OpenAI
Karpathy has been thinking out loud about two big themes this week. First: personal AI wikis as the right model for personalization. He highlighted Farzapedia — a personal knowledge base built on files you own, in universal formats, inspectable and interoperable — as superior to the black-box "the AI gets smarter the more you use it" approach. His argument: explicit memory in files means your data stays yours, works across any AI tool, and can even be fine-tuned into model weights. The "File over app" philosophy in action. He followed up by applying this to government accountability — arguing AI breaks the historical bottleneck where intelligence (not access) was what prevented citizens from processing the flood of public data: omnibus bills, lobbying disclosures, federal budgets, voting records. AI can finally make government legible to the governed.
Karpathy 这周主要在思考两个主题。第一：个人 AI wiki 是个性化更好的方式。他推荐了 Farzapedia——一个基于你自己拥有的文件构建的个人知识库，通用格式、可检查、可互操作——比"AI 用得越多越聪明"的黑盒方式好得多。他的逻辑：把记忆存在文件里意味着数据永远属于你、能被任何 AI 工具使用，甚至可以微调到模型权重里。"File over app"理念的实践。他随后把这个思路扩展到政府问责——认为 AI 打破了历史瓶颈：以前普通公民缺乏的不是信息获取渠道，而是处理海量公共数据（综合法案、游说披露、联邦预算、投票记录）的 intelligence。AI 终于可以让政府变得对人民透明了。
Aaron Levie — CEO, Box
Levie dropped two substantive threads. The first on why the context layer is irreplaceable in enterprise AI: every company has different data, different access levels, different workflows — and no model can know all of it. Even within a banking team, different bankers see different documents. This isn't solvable at the model layer; it's a context engineering problem. His second thread traced how agent architecture has matured. A year ago, the best practice for document-heavy agents was embedding-based chunk retrieval — necessary because context windows were small. Today, with bigger windows and better reasoning, agents can work more like humans: search broadly, read deeply, know when to look again. Box rebuilt their agent stack around this shift, and Levie sees this as the unlock for qualitatively new use cases.
Levie 发了两个深度 thread。第一个讲为什么 context layer 在企业 AI 里不可替代：每个公司有不同的数据、不同的访问权限、不同的 workflow——没有任何模型能知道所有这些。即使在同一个银行团队里，不同的客户经理能看到的文件也不同。这不是模型层能解决的问题，而是 context 工程的问题。第二个 thread 讲 agent 架构的演进。一年前，处理大量文档的 agent 最佳实践是基于 embedding 的分块检索——必须这么做，因为 context window 太小了。现在，随着 context window 变大、推理能力变强，agent 可以更像人一样工作了：广泛搜索、深入阅读、知道什么时候该再找一遍。Box 围绕这个转变重建了他们的 agent 技术栈，Levie 认为这是解锁质的飞跃的关键。
Peter Steinberger — Polyagentmorous ClawFather, OpenClaw contributor
Steinberger kept it short and punchy: "There's a big wave coming" — 1,043 likes and climbing. No thread, no context needed. Sometimes the vibe is the message.
Steinberger 保持了简短有力的风格："There's a big wave coming"——1043 个赞还在涨。没有 thread，没有解释。有时候气场本身就是信息。
Thariq — Claude Code at Anthropic
Thariq with the rare cooking-related agent tweet that somehow landed. A "POV: you're cooking" clip got 722 likes — apparently agents in the kitchen is a cultural moment. Meanwhile the serious work continues: he's one of the people living in the Claude Code trenches, shipping and iterating.
Thariq 发了一条少见的和烹饪相关的 agent 推文，莫名其妙就火了。"POV: you're cooking"的片段获得了 722 个赞——看来 agent 进厨房成了一个文化事件。同时正经的工作也在继续：他是 Claude Code 一线的建设者之一，持续迭代交付。
Garry Tan — President & CEO, Y Combinator
Garry announced 14 security bug fixes landed for GStack, half from community PRs — a good sign that the open-source flywheel is spinning. He's also pushing toward an "L8 software factory," with adaptive, smarter code review. The subtext: YC portfolio companies are writing a lot of code, and the AI-assisted review layer is becoming a genuine product.
Garry 宣布 GStack 上 14 个安全 bug 修复已上线，其中一半来自社区的 PR——这是开源飞轮转起来的好迹象。他还在推进"L8 软件工厂"的目标，实现适应性、更智能的代码审查。潜台词：YC 投资组合里的公司正在写大量代码，AI 辅助的代码审查层正在成为一个真正的产品。
Zara Zhang — Builder, Harvard '17
Zara shared a prompting tip from Peter Steinberger: always ask the model "Do you have any questions?" — a simple metacognitive check that can unlock better outputs. Small habits, real gains.
Zara 分享了 Peter Steinberger 的一个 prompting 小技巧：每次都问模型"你有什么问题吗？"——一个简单的元认知检查，可以解锁更好的输出。小的习惯，大的收获。
Peter Yang — Product at Roblox, AI tutorial creator (140K+ readers)
Peter asked a question at 3am Shanghai time: MacBook Pro or Mac Studio for vibe coding and running local models? 30 replies later, the Mac Studio consensus seems to be winning for pure local model throughput, but the MacBook Pro crowd argues mobility matters. The eternal builder dilemma, now with LLM flavor.
Peter 在上海凌晨 3 点发了一个问题：vibe coding + 本地跑模型，MacBook Pro 还是 Mac Studio？30 个回复后，Mac Studio 似乎在纯本地模型吞吐量上占了上风，但 MacBook Pro 阵营认为移动性不可忽视。建造者的永恒困境，现在有了 LLM 版本。
Nan Yu — Head of Product, Linear
Nan quote-tweeted a few things this week but nothing substantive enough to highlight. Quiet week on main.
Nan 这周转推了几条内容，但没有实质性的值得单独讲。主场比较安静的一周。
Nikunj Kothari — Partner, FPV Ventures
Nikunn posted a photo of flowers with a reminder to stop and smell them. Sometimes the builders need to breathe too.
Nikunj 发了一张花的照片，提醒大家别忘了闻一闻花。有时候建造者也需要喘口气。
Podcasts
No Priors — "AI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus"
Liam Fedus spent years training some of the most powerful AI systems in the world — including as VP of Post-Training at OpenAI, where he was part of the team that shipped GPT-4 and launched ChatGPT. Now he's building Periodic Labs, which he describes as "an AI foundation lab for atoms." The pitch: most AI excitement has been about bits — writing, coding, reasoning. The next wave is connecting these systems to the physical world: materials science, chemistry, drug discovery, semiconductor design. And unlike language, atoms don't publish their data on the internet.
Liam Fedus 曾训练过世界上一些最强大的 AI 系统——包括在 OpenAI 担任 Post-Training VP 期间，参与了 GPT-4 和 ChatGPT 的发布。现在他在建立 Periodic Labs，他把这描述为"原子的 AI 基础实验室"。核心观点：大部分 AI 的热度都在 bits 层面——写作、编程、推理。下一波是将这些系统连接到物理世界：材料科学、化学、药物发现、半导体设计。但和语言不同，原子不会把自己的数据发布到网上。
The conversation opens with a question that sounds simple but isn't: why are so many physicists running AI companies? Fedus's answer: physics teaches you to think from first principles, to be careful, and to work with systems where the feedback loop is reality itself. He traces his own path from dark matter research, to Google Brain during the Cambrian era of the transformer, to OpenAI for the productization push, and now to Periodic. "Science ultimately isn't sitting in a room thinking really hard," he says. "You have to conduct experiments. You have to learn from them. You have to interface with reality."
对话从一个听起来简单但实际不简单的问题开始：为什么这么多物理学家在运营 AI 公司？Fedus 的回答：物理学教会你从第一性原理思考，教你谨慎，教你和那些反馈回路本身就是现实的系统打交道。他追溯了自己的路径：从暗物质研究，到 Google Brain 经历 transformer 的"寒武纪大爆发"时代，到 OpenAI 做产品化，再到 Periodic。"科学最终不是在房间里拼命思考，"他说，"你必须做实验。你必须从中学习。你必须和现实对接。"
On the data problem: language models got easy data from the internet. Atoms don't have that. Periodic uses a hybrid — simulation data, experimental data, literature — but the key insight is that the system needs to be in a closed loop. Run an experiment, get data, find patterns, design the next experiment. The AI doesn't just store data; it actively drives the scientific process. Fedus is candid about the limits: "It's not like coding where you can check if the tests pass in milliseconds. The experiment takes time, the physical infrastructure has lead times." But the combination of better AI reasoning, better robotics, and better data loops is finally making the physical world tractable.
关于数据问题：语言模型从互联网上获得了容易获取的数据。原子没有这个条件。Periodic 使用混合方式——模拟数据、实验数据、文献数据——但关键洞见是系统需要处于一个闭环中。做实验、获取数据、发现规律、设计下一个实验。AI 不只是存储数据；它积极推动科学过程。Fedus 对局限性很坦诚："这不像编程，你可以在几毫秒内检查测试是否通过。实验需要时间，物理基础设施有前置周期。"但更好的 AI 推理、更好的机器人、更好的数据循环的组合，终于让物理世界变得可触及了。
The most quotable moment comes when the host invokes The Diamond Age — a Neal Stephenson novel about matter compilers that pipe resources into homes. Fedus's response: "You're going from systems that aren't just writing essays, not just writing software, but to literally generating matter." He believes Periodic's work could accelerate physical development by an order of magnitude or two, and that the combination of atoms and bits will eventually feel as natural as software does today. "If our physical world could keep up at some fraction to our digital world, I think life will just feel dramatically different."
最值得引用的时刻是主持人提到了《钻石时代》——Neal Stephenson 的小说，讲的是物质编译器把资源输送到每家每户。Fedus 的回应："你正在从那些不只是写文章、不只是写软件的系统，走向真正在生成物质的系统。"他相信 Periodic 的工作可能将物理世界的发展速度加速一两个数量级，bits 和 atoms 的结合最终会像今天的软件一样自然。"如果我们的物理世界能以数字世界几分之一的速度跟上，我认为生活会感觉完全不同。"
One unexpected thread: Fedus predicts recursive self-improvement in AI software will happen relatively soon — within the domain of software engineering, the feedback loops are tight enough. But connecting AI to the physical world requires a fundamentally different kind of patience, and that's the gap Periodic is built to bridge.
一个意外的 thread：Fedus 预测 AI 软件的递归自我改进会相对较快发生——在软件工程领域，反馈循环足够紧密。但把 AI 和物理世界连接起来需要一种根本不同的耐心，而这正是 Periodic 存在要弥合的鸿沟。
Watch here: https://www.youtube.com/watch?v=Iu4gEnZFQz8
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders