2026-05-01
AI Builders Digest — May 1, 2026
X / TWITTER
Andrej Karpathy (Former Tesla AI Director, OpenAI Founding Team)
Karpathy's Sequoia Ascent fireside chat is making the rounds. His core argument: LLMs are about far more than speeding up what already existed. Three new paradigms he flags — (1) menugen: apps fully engulfed by LLMs with zero classical code needed; (2) .md skills over .sh scripts: instead of writing a bash script to install software, just describe it in English and hand it to an LLM interpreter; (3) LLM knowledge bases: something fundamentally impossible with classical code because it requires computation over unstructured data from arbitrary sources. He also dives into the "jaggedness" of LLMs — why a single model can refactor 100k lines of code coherently yet hallucinate about car washes. The short version: you're either in the RL data distribution ("on rails") or you're off-roading in the jungle. Finally, the agent-native economy: decomposition of products into sensors, actuators, and logic across 1.0/2.0/3.0 paradigms, and the emerging skill set of agentic engineering.
Karpathy 还在推他那句最近反复引用的金句。核心论点：LLM 的价值远不止"加速已有工作"。他点名了三个新方向——(1) menugen：整个应用可以被 LLM 吞掉，不需要一行传统代码；(2) .md 技能取代 .sh 脚本：安装软件不需要写 bash 脚本，直接用自然语言描述扔给 LLM；(3) LLM 知识库：这是传统代码根本做不到的事，因为涉及对任意来源、非结构化数据的计算。关于 LLM 的"参差不齐"问题（同一个模型能重构 10 万行代码却会在洗车这种事上产生幻觉），他的解释是：要么在 RL 训练数据的轨道上运行，要么在丛林里用砍刀开路。最后提到了 Agent 原生经济——把产品拆解成传感器、执行器、逻辑，散落在 1.0/2.0/3.0 的计算范式中。
Sam Altman (CEO, OpenAI)
OpenAI shipped a big Codex upgrade — now handles non-coding computer work. Altman frames it as the next frontier: computer use that isn't just code generation. Also dropped "artificial goblin intelligence achieved" which predictably broke the internet (6,200+ likes).
OpenAI 给 Codex 做了大升级——现在不止能写代码，还能处理通用计算机任务。Altman 把这个定位成下一个前沿方向。另外那条"AGI 已实现"的玩笑推文不出意外地炸了，6200 多个赞。Sam 最近的风格越来越放松了。
https://x.com/sama/status/2049946120441520624
https://x.com/sama/status/2050021650641695108
Aaron Levie (CEO, Box)
Mandatory reading for anyone building software in the agent era. Levie's thesis: as agents become the biggest users of software, all software goes headless. The UI era ends; APIs win. His three-part framework for what comes next: (1) Seats still matter for people, but each seat comes with API usage on behalf of that user — agents need to work with your data seamlessly or you're DOA; (2) Agent seats make sense for stateful agents with their own workspace and permissions, but pricing them like human seats is nearly impossible — one company might use 1 agent, another 1,000; (3) Consumption pricing becomes dominant for headless use beyond seat allotments. The kicker: agents operating on data will dramatically exceed what humans do with their tools today. Every platform that goes headless will need this model.
如果你在 Agent 时代做软件，这篇必读。Levie 的核心判断：随着 Agent 成为软件的最大用户，所有软件都会变成无头（headless）模式。UI 时代结束，API 为王。他的三层框架：(1) 人类用户仍然按席位订阅，但每个席位都要附带 API 配额让 Agent 代为操作——如果 Agent 不能顺畅访问你的数据，你离死不远了；(2) Agent 席位适用于有独立工作空间和权限的有状态 Agent，但按人类席位定价几乎不可能——有的公司可能只用一个 Agent，有的用 1000 个；(3) 超出席位配额的用量走消耗式计费。最后他点出了关键：Agent 对数据的操作量会远超人类。每一个走向无头模式的平台都需要这套模型。
Ryo Lu (Design, Cursor)
Two things from Ryo this week. First: a demo of adding the Cursor SDK to ryOS — now he can edit his OS by just chatting. The future of operating systems as conversational interfaces, right there. Second, a sharp thread on opinionated vs. general-purpose design: "no opinion is opinionated." Every decision about what concepts exist, what's visible, what composes — that's already a worldview. His argument: great systems don't multiply products, they reduce concepts to their essence. Flexibility isn't the absence of opinion; it's the result of a strong opinion held in cohesion.
Ryo 这周发了两条值得关注的内容。第一，演示了把 Cursor SDK 集成到 ryOS，现在可以直接聊天改操作系统——这大概就是未来 OS 的样子。第二条是关于"有观点 vs. 通用设计"的深度思考：所谓"没有立场"本身就是一种强烈立场——只要决定哪些概念存在、什么可见、什么可组合，你就已经在表达世界观了。他的核心观点：伟大的系统不是堆产品，而是把概念简化到本质。灵活性不是没有立场，而是立场坚定后的自然结果。
https://x.com/ryolu_/status/2049872551955013713
https://x.com/ryolu_/status/2049866003287576978
Peter Steinberger (Polyagentmorous ClawFather, OpenClaw + OpenAI)
Steinberger has been in the security ecosystem trenches — worked with NVIDIA, OpenAI, Microsoft, GitHub, Tencent Hunyuan, Convex, Atlassian, and Blacksmith to "get secure the claw." Also doing evangelism work: if you tried OpenClaw in group chats and got mixed results, try again — "it IS SO GOOD NOW." His case for Codex harness over GPT for agent work is getting traction.
Steinberger 这几个月一直在安全生态里摸爬滚打——和 NVIDIA、OpenAI、Microsoft、GitHub、Tencent Hunyuan、Convex、Atlassian、Blacksmith 合作，给 OpenClaw 做安全加固。另外他还在热情推广：如果之前在群聊里用 OpenClaw 效果一般，再试一次——"现在真的很好用了"。他推荐用 Codex harness 替代 GPT 做 Agent 工作的观点也在获得认同。
https://x.com/steipete/status/2049976855617314991
https://x.com/steipete/status/2049988836160074022
Guillermo Rauch (CEO, Vercel)
Vercel CEO asked v0 what it would look like if Vercel shipped GitHub. Two prompts. The result is predictably unsettling in a good way.
Vercel CEO 用 v0 问了两个 prompt：如果 Vercel 发布 GitHub 会长什么样。结果嘛……有点让人不安，但那种好的不安。
Cat Wu (Claude Code, Anthropic)
Claude Security is now in public beta, built directly into Claude Code on the web. Point it at a repo, get validated vulnerability findings, and fix them in the same interface where you're writing code.
Claude Security 现在已进入公开测试阶段，直接集成在网页版 Claude Code 里。对着代码库一指，能拿到经过验证的漏洞报告，还能在写代码的同一界面里直接修复。
Amjad Masad (CEO, Replit)
Replit's meta-dogfooding continues. Amjad's frame: they treat Replit as customer zero — and expect "insane ROI." The example he shares is exactly that. Also dropped "Prompt → LLC" as the new startup elevator pitch template.
Amjad 的逻辑：把 Replit 当作零号客户来用，期望获得"疯狂的 ROI"。他分享的案例印证了这一点。另外那条"Prompt → LLC"正在成为新的电梯演讲模板。
https://x.com/amasad/status/2049934937688854993
https://x.com/amasad/status/2049921597499445677
Nikunj Kothari (Partner, FPV Ventures)
All these MCPs and CLIs are making one thing clear: "big" models are going to orchestrate our lives. His progression: first terminal, then computer use, soon the entire OS. Message: be in the path of these models or you'll cease to exist.
所有这些 MCP 和 CLI 正在说明一件事：大模型将成为我们生活的指挥者。路径是：先是终端，然后是 computer use，接下来是整个操作系统。他的建议：要么进入这些模型的路径，要么消失。
Garry Tan (President & CEO, Y Combinator)
YC President on GBrain — an MIT-licensed open source knowledge wiki that pairs well with OpenClaw/Hermes-style personal AI. Three weeks in, he's confident GBrain is its own category, not competing with general-purpose retrieval tools.
YC 掌舵人在推 GBrain——一个 MIT 许可的开源知识 wiki，特别适合搭配 OpenClaw/Hermes 这类个人 AI 使用。三个星期下来，他的判断是 GBrain 自成一家，不是去和通用检索工具竞争。
https://x.com/garrytan/status/2050096324100682097
https://x.com/garrytan/status/2050095919157350644
Amanda Askell (Philosopher & Ethicist, Anthropic)
A rare personal thread from Askell — she's been seeing internet fiction written about her that's "completely made up" and asserted with confidence. Her reaction: she should be "the millionth item on people's list" of things to write fiction about, somewhere below paper cups and the right way to caulk a bathtub. Classic Askell dry humor, while noting her actual work is "top tier in terms of interestingness."
Askell 发了一条罕见的个人推文——她发现自己成了互联网同人小说的主角，内容完全虚构却写得信心十足。她的反应：自己应该只是人们写同人的百万分之一候选，甚至排在纸杯和浴室防霉之前。典型的 Askell 式冷幽默。同时她也认真地说，她做的实际工作"有趣程度是顶级的"。
Dan Shipper (CEO, Every)
Shipper's experimenting with Codex + Chronicle as a focus tracker — logging where attention goes while working. Also sharing what looks like a notable Every product launch around AI-native company tools.
Dan Shipper 在用 Codex + Chronicle 做注意力追踪——记录工作时注意力花在哪里。同时似乎有 Every 新的 AI 原生公司工具产品即将发布。
https://x.com/danshipper/status/2050010481751167187
https://x.com/danshipper/status/2049913064561258986
Aditya Agarwal (GP, South Park Commons; Ex-Facebook, Dropbox)
An episode plug for his South Park Commons conversation with Nikesh Arora, CEO of Palo Alto Networks. Core argument: AI is both the threat and the only viable solution in cybersecurity. Incumbents got it wrong when ChatGPT dropped — security was never built into AI from the start. The only way out of the current chaos is more AI.
Aditya 在推广他和 Palo Alto Networks CEO Nikesh Arora 的对话播客。核心论点：AI 是网络安全领域里的威胁，也是唯一可行的解药。传统安全厂商在 ChatGPT 出现时判断失误——安全从未被内置进 AI。走出当前混乱的唯一办法是用更多 AI。
Peter Yang (Product, Roblox)
Two practical questions from Yang: best workflow for self-updating OpenClaw (and it breaking half the time — a real problem statement), and whether anyone's cracked great YouTube thumbnail generation with GPT Image 2 yet.
Peter Yang 提了两个实际问题：如何让 OpenClaw 自我更新不频繁崩溃（这是真实痛点），以及有没有人已经搞定用 GPT Image 2 生成油管缩略图。
https://x.com/petergyang/status/2050026309615792509
https://x.com/petergyang/status/2050022980688966081
Zara Zhang (Builder)
Using ChatGPT Image 2 to make slides — her verdict: "SO fun."
Zara 在用 ChatGPT Image 2 做 PPT，评价是"太好玩了"。
PODCASTS
Training Data — "Demis Hassabis on Building DeepMind, AlphaFold, and the Final Stretch to AGI"
DeepMind 创始人 Demis Hassabis 近日登上 Training Data 播客，分享了他从游戏神童到 AGI 追梦者的完整路径。以下是本次对话的核心提炼。
The Takeaway: Hassabis 的经历证明了一条反直觉的创业原则——提前 5 年押注正确方向是天才，押注 50 年是疯子。他在 15 岁时就把 AGI 定为人生目标，然后用游戏、神经科学、学术研究绕了一大圈才回到原点。这种"先绕远路再回到核心"的策略，恰恰是 DeepMind 最终成功的隐秘原因。
最有力的一个细节：他 17 岁做的 Theme Park 卖了 1000 万份，成千上万的玩家沉浸在那个 AI 驱动的模拟世界里——他意识到，这不只是娱乐，而是人们第一次真正愿意和 AI 共存。这种体验让他确信：AI 值得他用一辈子来追求。
关于 AGI 时间线，他维持了一贯的判断：2030。"我们从 2010 年开始，认为这需要 20 年。现在看来，整个领域基本上是按计划推进的。"
AlphaFold 被问作最自豪的时刻，但 Hassabis 更兴奋的是下一步：Isomorphic Labs 正在解决药物发现的后半程——已知蛋白质结构，现在要自动设计能精准结合的化合物。"把 99% 的探索放在硅片里做，只把验证那一步留到实验室。如果能做到，药物发现时间可以从 10 年压缩到数周，甚至数天。届时所有疾病都在射程内。"
关于 consciousness，他承认自己无法给出一个完整定义，但指出了必要条件：self-awareness、self-other distinction、时间上的连续性。"我们会从行为角度接近这个问题，但关于体验层面，永远会有一道鸿沟——因为我们永远不会和人工系统有相同的基质。"
关于物理世界和信息的关系，他提出了一个独特的观点：信息可能比物质和能量更基础。"如果你从信息处理的角度看宇宙，AI 的意义就比我们以为的更深远——它本质上是在组织和理解信息。"
最后被问到 AGI 之后读什么书，他推荐了 David Deutsch 的《The Fabric of Reality》——那是他希望在 AGI 实现后能回答的问题。
DeepMind 创始人 Demis Hassabis 近日登上 Training Data 播客，分享了他从游戏神童到 AGI 追梦者的完整路径。以下是本次对话的核心提炼。
核心结论： Hassabis 的经历证明了一条反直觉的创业原则——提前 5 年押注正确方向是天才，押注 50 年是疯子。他在 15 岁时就把 AGI 定为人生目标，然后用游戏、神经科学、学术研究绕了一大圈才回到原点。这种"先绕远路再回到核心"的策略，恰恰是 DeepMind 最终成功的隐秘原因。
最有力量的一个细节：他 17 岁做的 Theme Park 卖了 1000 万份，成千上万的玩家沉浸在那个 AI 驱动的模拟世界里——他意识到，人们第一次真正愿意和 AI 共存，而这种体验让他确信：AI 值得他用一辈子来追求。
关于 AGI 时间线，他维持了一贯的判断：2030。"我们从 2010 年开始，认为这需要 20 年。现在看来，整个领域基本上是按计划推进的。"
AlphaFold 被问作最自豪的时刻，但 Hassabis 更兴奋的是下一步：Isomorphic Labs 正在解决药物发现的后半程——已知蛋白质结构，现在要自动设计能精准结合的化合物。"把 99% 的探索放在硅片里做，只把验证那一步留到实验室。如果能做到，药物发现时间可以从 10 年压缩到数周，甚至数天。届时所有疾病都在射程内。"
关于 consciousness，他承认自己无法给出完整定义，但指出了必要条件：self-awareness、self-other distinction、时间上的连续性。"我们会从行为角度接近这个问题，但关于体验层面，永远会有一道鸿沟——因为我们永远不会和人工系统有相同的基质。"
关于信息与宇宙的关系，他提出了一个独特的观点：信息可能比物质和能量更基础。"如果你从信息处理的角度看宇宙，AI 的意义就比我们以为的更深远。"
最后被问到 AGI 之后读什么书，他推荐了 David Deutsch 的《The Fabric of Reality》——那是他希望在 AGI 实现后能回答的问题。
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders