🌐 双语
Archive

AI Builders
Digest

2026-11-20 14 builders · 33 tweets · 1 podcasts · 0 blogs

🔥 热点话题

超级智能离不开物理世界:Arda CEO Bob McGrewSuperintelligence Needs the Physical World: Arda CEO Bob McGrew

The Takeaway:没有物理世界的变革,就不可能真正实现超级智能。数字自动化研究者只是一次性的S曲线加速,而非递归自我改进。

Bob McGrew是Arda的CEO兼联合创始人,曾在OpenAI担任首席研究官,早年是Palantir的关键早期高管。他指出,如果只看数字化生活,过去十年几乎没有变化。即便是Waymo,第三次乘坐后人们就只抱怨交通。研究者不是瓶颈,真正的限制是GPU数量、实验吞吐和物理设施。自动化研究可能带来像推理能力跃迁那样的巨大一次性提升,但不会触发失控的自我改进循环。

“I don't think you are actually gonna be able to get to superintelligence without changes in the physical world.”

当假设生成变得廉价,稀缺的就是湿实验室、制造线和执行能力。Arda正用类似Cursor的方式为制造工程师构建平台,让自动化产线能像重新培训工人一样在七天内重新配置,重点服务高混、快速迭代的领域:国防、消费电子和数据中心供应链。他们采用Palantir式的前线部署工程师模式,从Anduril的单个执行器虚拟构建起步,已扩展到无人机、导弹和Fury协作战斗机。最终愿景是机器制造机器,甚至走向自我复制的星际探针,让智能真正走出数据中心。
The Takeaway: You cannot reach superintelligence without changing the physical world. Automating researchers is a one-time S-curve, not recursive self-improvement.

Bob McGrew, CEO and co-founder of Arda, previously served as Chief Research Officer at OpenAI and was an early key executive at Palantir. He argues that digital life has barely changed in a decade. Even a Waymo ride loses its magic by the third trip and becomes just another traffic complaint. Researchers were never the real bottleneck at OpenAI—GPU count, networking, and experiment throughput were. Automating research can deliver a large one-time gain similar to the jump from GPT-4o to o3 with reasoning, but it does not create a runaway loop.

“I don't think you are actually gonna be able to get to superintelligence without changes in the physical world.”

When hypotheses become cheap, the scarce resources are wet labs, manufacturing capacity, and physical action. Arda is building a Cursor-like platform for manufacturing engineers so that automated lines can be reconfigured in seven days—the same speed as retraining human workers. The focus is high-mix, fast-changing domains: defense, consumer electronics, and data-center supply chains. Using a Palantir-style forward-deployed engineering model, they started with a single actuator virtual build at Anduril and have expanded across drones, missiles, and the Fury collaborative combat aircraft. The longer arc is factories that build factories and ultimately self-replicating systems that can expand intelligence beyond the data center.
查看原文 →

ChatGPT 推出全新移动端体验ChatGPT Rolls Out New Mobile Experience

OpenAI Codex & ChatGPT 团队成员 Thibault Sottiaux 宣布,ChatGPT 正在推出全新移动端体验,目前先向 Pro 用户开放。新设计让常用功能更容易触达,用户可以把 Codex、Projects、Images 等固定到标签栏,自己的“dot”也有独立标签。部分用户会默认看到新界面,也可在设置中随时开关。他调侃这是“Le ChatGPT 2.0”,并表示团队“crazy enough”要在 12 亿用户的现有体验上原地改进。
OpenAI Codex & ChatGPT team member Thibault Sottiaux announced a new mobile experience for ChatGPT, starting with Pro users. The redesign makes favorite features easier to reach; users can pin Codex, Projects, Images and more to the tab bar, and their personal “dot” gets its own tab. Some users will see the new design by default and can toggle it in Settings. He called it “Le ChatGPT 2.0” and noted the team is “crazy enough to think they can succeed at improving the experience used by 1.2B users in place.”
查看原文 →查看原文 →查看原文 →

Gemini 4 Neon 与 Mac 端电脑任务能力Gemini 4 Neon and Computer-Use on Mac

Google VP Josh Woodward 宣布 Gemini 应用在 Mac 上现已支持在电脑上执行任务,这是用户呼声最高的功能之一。Peter Yang 对比了 Gemini 4 Neon 与 Claude Opus 5.5:Neon 在写作和个性上已持平甚至更好,速度远超 Carbon,日常 99% 任务够用;Opus 仍在 3D 和设计上略胜。他特别指出 Google 拥有超过 10 万员工可做内部测试和统计显著的 A/B,是巨大优势,且 100 美元/月的 Gemini Ultra 订阅(含 YouTube Premium Family 和 20TB 存储)性价比极高。
Google VP Josh Woodward announced that the Gemini app on Mac can now perform tasks on the user’s computer—one of the most requested features. Peter Yang compared Gemini 4 Neon to Claude Opus 5.5: Neon matches or beats Opus on writing and personality while being far faster than Carbon, making it the default for 99% of everyday work; Opus still leads on 3D and design. Yang highlighted Google’s structural advantage of 100,000+ employees who can run statistically significant internal A/B tests, and called the $100/month Gemini Ultra plan (including YouTube Premium Family and 20 TB storage) a steal.
查看原文 →查看原文 →查看原文 →

工程师的价值被彻底重新定价Engineering Value Is Being Completely Repriced

Vercel CEO Guillermo Rauch 给出了清晰的数学:结果导向、有创造力的工程师现在是 1000x;纠结于 tabs-vs-spaces、只见树木不见森林的工程师变成了 -10x;普通知识工作者变成了 2x。这与他们在创业公司和大型企业中观察到的 token 消耗分布一致——工程师和软件工厂将指数级消耗更多 token。Anthropic 的成功再次证明:编码是 AI 的头号杀手级应用。
Vercel CEO Guillermo Rauch put hard numbers on the shift: an outcome-oriented, creative engineer is now a 1000x engineer; a tabs-vs-spaces, miss-the-forest engineer is now a -10x engineer; the average knowledge worker is a 2x engineer. This matches the token-consumption distribution Vercel sees across startups and enterprises—engineers and software factories will consume tokens exponentially. Anthropic’s success underlines the same point: coding is the #1 killer app of AI.
查看原文 →

💰 创业成功案例

Replit:用一句话构建个人代理,并与代理签合同Replit: Build a Personal Agent with One Prompt and Sign a Contract with It

Replit CEO Amjad Masad 展示了用一句话即可构建类似 Muse 的个人代理,并强调“与代理签合同”的范式——他预期很多人会跟进。这延续了 Replit 把代理能力直接交到用户手中的路线。
Replit CEO Amjad Masad showed how to build a Muse-like personal agent with a single prompt and highlighted the paradigm of having a formal contract with your agent—something he expects many will copy. It continues Replit’s push to put agent capabilities directly into users’ hands.
查看原文 →查看原文 →

Anthropic 的成功与 SPC 的长期复利Anthropic’s Success and SPC’s Long-Term Compounding

SPC 普通合伙人 Aditya Agarwal 祝贺 Anthropic,并指出两位关键建设者都曾在 SPC 深度停留:Tom Brown 2016 年加入时毫无 ML 背景,靠与其他成员一起啃论文最终联合创立 Anthropic;Thariq 关掉游戏公司后加入,因同伴推荐尝试 Claude Code 而参与构建了今天几乎人人使用的工具。把有野心的人放在一起,多年后会产生非凡结果。
SPC General Partner Aditya Agarwal congratulated Anthropic and noted that two key builders spent formative time at SPC. Tom Brown arrived in 2016 with no ML background, worked through papers with other members, and later co-founded Anthropic. Thariq joined after winding down a gaming company; a fellow member asked if he had tried Claude Code, and he soon helped build a tool most people now use daily. Putting ambitious people next to each other produces extraordinary outcomes years later.
查看原文 →

Vercel Board:代理的“无头 Slack”与公司创世纪Vercel Board: Headless Slack for Agents and the Genesis of a Company

Guillermo Rauch 将 Vercel Board 描述为代理的“无头 Slack”和 Redux 式状态管理器——没有沉重的 UI,只是基础设施。安全团队可以只看漏洞计数和 SEV 分级,内部工厂甚至用 TUI 或纯 Slack 代理运行。他进一步提出:一家公司的创世纪,就是那个帮助你建立公司的代理。
Guillermo Rauch described Vercel Board as “headless Slack” for agents and the Redux state manager of your agent fleet—no heavyweight UI, pure infrastructure. Security teams focus on vulnerability counts and SEV classifications; internal factories run as TUIs or entirely via Slack agents. He went further: a company’s genesis is the agent that helps you build the company.
查看原文 →查看原文 →

🛠️ 开发者工具与技巧

从单任务到完整岗位:RoleCUA 与代理数据From Single Tasks to Whole Jobs: RoleCUA and Agent Data

Aditya Agarwal 指出代理正从执行单个任务升级到承担完整岗位,数据必须跟上。Fig 的 RoleCUA 映射专家角色数月真实工作中的每一个动作,并与任务和目标绑定,为训练更强大的代理提供结构化基础。
Aditya Agarwal noted that agents are moving from single tasks to whole jobs, and the data must follow. Fig’s RoleCUA maps expert roles across months of real work, tying every action to the task and objective it serves—providing structured ground truth for more capable agents.
查看原文 →

软件既是艺术也是实用,AI 既是工业革命也是文艺复兴Software Is Both Art and Utility; AI Is Both Industrial Revolution and Renaissance

前 Anthropic Claude Code 成员 Thariq 在 AIE Code 上强调:工厂解决的是实用性,而我们还需要更好的艺术。他同时用“an age of miracles”来形容当前时刻。
Former Anthropic Claude Code member Thariq argued at AIE Code that software is both art and utility, and AI is both an Industrial Revolution and a Renaissance. Factories handle utility; we still need to make better art. He also simply called the moment “an age of miracles.”
查看原文 →查看原文 →

OpenClaw 嵌入式 harness 修复落地OpenClaw Embedded Harness Fix Lands

Peter Steinberger 表示已在 OpenClaw 中修复了嵌入式 harness 的问题(针对少数使用插件而非 Codex 的用户),大多数 harness 尚未更新。
Peter Steinberger reported landing a fix in OpenClaw for the embedded harness (for the minority of users relying on the plugin instead of Codex). Most other harnesses have not yet been updated.
查看原文 →

🌍 其他动态

社交媒体没有准确性代价,只有夸张激励Social Media Has Zero Cost for Inaccuracy and Maximum Incentive for Hyperbole

Swyx 指出一个残酷现实:你可以发“达到 SOTA 的概率已降为零”并收获大量点赞,三个月后对方真的达到 SOTA,你再发道歉贴又能收获更多点赞,对方甚至会打印出来感谢你给的动力。算法奖励的就是这种循环,几乎没有后果。
Swyx observed that you can claim “odds of reaching SOTA have dropped to zero,” collect likes, then three months later post an apology when they do reach SOTA and collect even more likes—while the subject of the post prints it out and thanks you for motivation. There are no consequences for inaccuracy and every incentive for hyperbole; the algorithm is tuned for exactly this.
查看原文 →

保护老年人免受网络攻击与诈骗的四层方案Four Layers to Protect the Elderly from Network Attacks and Scams

Nikunj Kothari 提出针对老年人(尤其是父母和祖父母)的防护思路:浏览器层(安全浏览器拦截恶意网站)、网络层(类似 Pi-hole 或手机 VPN)、支付层(可疑交易拦截)、手机层(只允许亲友来电 + 克隆声音检测)。最终可打包成带保险的产品,由子女购买并接收异常警报。
Nikunj Kothari outlined four layers to protect the elderly: a secure browser that only allows safe sites, a network-level device or VPN that blocks known-bad destinations, payment-layer flagging of scam-like transactions, and phone-level controls that restrict calls to known contacts while detecting cloned voices. The whole stack could be packaged with insurance, bought by adult children who receive alerts when something looks off.
查看原文 →

谁在阻止当前的基础设施攻击?Who Is Working on Stopping the Current Infrastructure Attacks?

Every CEO Dan Shipper 直接发问:谁在处理眼下正在发生的基础设施攻击?AI 领域的人能提供什么帮助?
Every CEO Dan Shipper asked bluntly: who is working on stopping the infrastructure attacks happening right now, and how can people in AI help?
查看原文 →

Bostrom 警告与 YC 的认同Bostrom’s Warning and YC’s Agreement

Y Combinator 总裁兼 CEO Garry Tan 转引并强烈认同 Nick Bostrom 的观点:“if nobody builds it, everyone dies”。
Y Combinator President & CEO Garry Tan quoted and strongly endorsed Nick Bostrom’s line: “if nobody builds it, everyone dies.”
查看原文 →