AI 建造者日报 — 2026年06月16日
📌 X/TWITTER
【swyx】AI 开发者与布道者，Latent Space 播客联合主持人
swyx 分享了 Anthropic UltraCode 的使用体验——这是一项尚未被广泛采用的强大工具。“subroutines but intelligent”（智能化的子程序）是其核心理念：当你意识到大量知识工作本质上是一连串需要判断力的"yak shave"时，动态工作流（dynamic workflows）的价值就不仅限于编码任务了。他还引用了微软 CEO Satya Nadella 关于"学习循环作为知识产权"的论述：真正的机会不在于挑选最好的模型，而在于在模型之上构建学习循环，让人力资本和 token 资本持续复利增长。“你可以外包任务，甚至外包工作，但你永远无法外包你的学习。”
[原文链接] https://x.com/swyx/status/2066415484149633329 [原文链接] https://x.com/swyx/status/2066235625695850526
swyx shared hands-on impressions of Anthropic UltraCode, noting it’s still under the radar but “scarily good at burning tokens.” The core idea: “subroutines but intelligent.” He also amplified Satya Nadella’s thesis on learning loops as IP — the real opportunity is building a compound learning system on top of models where human and token capital grow together. “You can offload a task, or even a job, but you can never offload your learning.”
【Thomas Sottiaux】OpenAI Codex & ChatGPT 团队成员
Codex 现在可以查看并自行设定 /goal。团队将开发的一切也都构建为 agent 可用的工具。这是元提示（meta prompting）的泛化版本——让 agent 基于你的意图自行定义任务。
[原文链接] https://x.com/thsottiaux/status/2066270561081454989
Codex can now see and set its own /goal. Everything the team builds is also built as a tool for the agent. This is a generalization of meta prompting: the agent sets its own task based on your intent.
【Nan Yu】Linear 产品负责人
“反方观点：现在每个人都在结对编程，只是搭档是个机器人。“一句犀利的观察点出了 AI 编码工具已经悄然成为开发者标配的现实。
[原文链接] https://x.com/thenanyu/status/2066190061419282602
“Counterpoint: everyone pair programs now, with a robot.” A sharp observation that AI coding assistants have quietly become the default programming partner.
【Guillermo Rauch】Vercel CEO
Vercel AI 技能市场突破 70 万个技能，全部有机、社区驱动。Rauch 称之为"开放的 AI 生态系统”——一个不由单一公司控制、由社区共建的 AI 工具生态正在成形。
[原文链接] https://x.com/rauchg/status/2066299732277031042
The Vercel AI Marketplace has passed 700,000 skills — entirely organic and community-driven. Rauch calls it “the open AI ecosystem,” highlighting a community-powered alternative to centralized AI platforms.
【Aaron Levie】Box CEO
Levie 发表了两条重量级观点。其一：企业的竞争优势将取决于能否将其独特的 IP、机构知识和数据转化为 AI 可用的架构。“真正的机会不在于挑选最好的模型，而在于构建让人力资本和 token 资本复利增长的学习循环。“其二：开源权重模型将成为最大赢家。当模型可以被随时撤回的先例已经确立，各国将有更强的动力发展主权 AI，而开源权重模型——目前多数并非来自美国——将成为各国的首选方案。Levie 呼吁美国重新审视在模型层而非应用层监管 AI 的策略。
[原文链接] https://x.com/levie/status/2066237607244427761 [原文链接] https://x.com/levie/status/2066167615618466060
Levie dropped two heavyweight takes. First: competitive advantage goes to companies that can format their unique IP and institutional knowledge for AI systems — “the real opportunity is not in picking the best model but in building a learning loop where human and token capital compound.” Second: open weights models will be the biggest winner. Now that the precedent exists for models being pulled back, countries have even more incentive to develop sovereign AI — and open weights, mostly not from the US, becomes the default choice. He urges the US to reconsider regulating AI at the model layer versus the applied layer.
【Garry Tan】Y Combinator CEO
“开源是企业长期掌控自身命运的安全阀。“Tan 还预测：下一代改变世界的年轻人，将是那些最擅长让长时间运行、多阶段、多团队的 agent 任务大规模高效运转的人——贯穿他们个人和工作生活的每一个方面。
[原文链接] https://x.com/garrytan/status/2066307697574862905 [原文链接] https://x.com/garrytan/status/2066269412391637050
“Open source is the escape hatch for businesses to control their own destiny long term.” Tan also predicts the next generation of world-changers will be those most adept at making long-running, multi-stage, multi-team agent tasks work at high volume across every part of their lives.
【Zhang Rui】建造者，Follow Builders 作者
关于如何做出好 skill 的两条精辟见解：“好的 skill 不是写出来的，是做出来的——反复做 20 次，然后让 AI 把你刚做的一切打包。“以及"skill 是结束时的产物，不是开始时的起点。“这是在 AI agent 工具链实战中凝练出的经验之谈。
[原文链接] https://x.com/zarazhangrui/status/2066388749244854771 [原文链接] https://x.com/zarazhangrui/status/2066394505037926426
Two nuggets on building good AI skills: “You don’t make a good skill by writing a skill — you make it by doing the thing, fixing it 20 times, then telling the AI to bottle up everything you just did.” And: “You make a skill by ending with one, not starting with one.” Battle-tested wisdom from the agent tooling trenches.
🎙️ 播客
Training Data: LIVE — Jensen Huang on Building the Dynamo of the Intelligence Age
一句话要点： Jensen Huang 将 AI 革命定义为人类历史上第三次"茧壳化"基础设施浪潮——继电网和互联网之后，智能生成层将包裹整个地球，而 NVIDIA 在其中扮演"智能发电机"的角色。
Jensen Huang 在这次演讲中给出了理解 AI 产业的全景框架。他将计算机的历史划分为两个时代：过去 60 年是"检索式计算”——人类预录制内容、存储为文件、需要时取出；而从 ChatGPT 出现至今，我们进入了"生成式计算”——每一次交互都实时生成全新的内容。推理（reasoning）是生成能力的自然延伸，而 Agent 则是推理能力的应用：“AI 从理解信息，进化到能够执行工作。”
NVIDIA 的核心定位被他描述为一个跨越 300 年的对称结构：300 年前西门子发明了发电机（原子的运动 → 电子）；今天 NVIDIA 制造的是智能发电机（电子 → 数字/token）。一个 DGX 机柜重两吨、价值 400 万美元、包含 150 万个零件——“一台超级昂贵的机器，但我们像造手机一样批量生产它们。”
他提出的"五层蛋糕"投资框架尤为值得关注：底层是能源（几十年来最大规模的电力基础设施投资机会），第二层是芯片与网络，第三层是数据中心基础设施，第四层是模型层（不只是语言模型，还包括蛋白质、基因组、物理世界等一切有结构可学习的数据），第五层是应用层（去年单年就有 1000 亿美元 VC 资金涌入）。
关于 AI 与就业的讨论，他直指核心：区分"任务"和"目的”。放射科医生的案例证明——AI 自动化的是阅片任务，但让放射科医生更高效，结果是更多患者入院、更多扫描、需要更多放射科医生。“你不是被 AI 取代，而是被使用 AI 的人取代。“他呼吁各国积极投资 AI，而非因恐惧而回避。“AI 是我整个职业生涯中消除技术鸿沟的最大力量。”
[YouTube] https://www.youtube.com/watch?v=2UpQbeAZuqA
Takeaway: Jensen Huang frames the AI revolution as the third “cocooning” infrastructure wave in human history — after the electrical grid and the internet, an intelligence-generation layer will envelop the planet, with NVIDIA as its dynamo.
Huang lays out a panoramic framework for understanding AI’s industrial structure. The 60-year era of “retrieval computing” (store files, retrieve later) has given way to “generative computing” — every interaction produces original output in real time. Reasoning emerged from generation, and agents emerged from reasoning: “AI went from understanding to doing work.”
He describes NVIDIA through a 300-year symmetry: Siemens invented the dynamo (motion in, electrons out); NVIDIA builds the intelligence dynamo (electrons in, tokens out). A single DGX rack weighs two tons, costs $4 million, and contains 1.5 million parts — “the most expensive piece of equipment in the world, but we crank them out like phones.”
His “five-layer cake” investment framework: energy (the biggest grid investment opportunity in generations), chips & networking, data center infrastructure, models (not just language — proteins, genes, physics, anything with learnable structure), and applications ($100B in VC last year alone).
On jobs: distinguish task from purpose. Radiologists — AI automated scan reading, making them more productive, leading to more patients, more scans, and more radiologists hired. “You won’t lose your job to AI. You’ll lose your job to someone who uses AI.” He urges countries to engage rather than retreat: “AI is the greatest force for eliminating the technology divide in my entire career.”
通过 Follow Builders 生成: https://github.com/zarazhangrui/follow-builders