🌐 双语
Archive

AI Builders
Digest

2026-11-04 18 builders · 29 tweets · 1 podcasts · 0 blogs

🔥 热点话题

Anthropic Claude Code 核心成员离职与 AI 编码范式转变Claude Code Core Team Departure and the Shift in AI Coding

Anthropic Claude Code 与 Cowork 团队成员 Cat Wu 宣布已于上周五离职。两年前 Claude Code 还只是内部小工具时,她让它构建 RL 环境,自己从未手写一行代码却成功运行,从此加入团队。从最初用户手动批准每个权限、主要用来写测试,到如今默认信任 Claude 自动批准、完成大规模迁移和产品构建,Claude Code 已成为行业定义性产品。她感谢团队与用户的反馈塑造了产品。

Former Claude Code 成员 Thariq 引用 bitter lesson 指出:最聪明的模型将需要越来越少的 token,不再需要验证工作或检查 bug,它们会直接正确。OpenAI Codex 与 ChatGPT 团队的 Thibault Sottiaux 则表示,agents 今天已令人难以置信,却仍如此令人沮丧,改变会一次次发生,我们还非常早期,期待重新找回 flow state。
Anthropic Claude Code and Cowork team member Cat Wu announced her last day was Friday. Two years ago when Claude Code was a small internal tool, she asked it to build an RL env, steered it for tens of hours without writing a single line of code herself, got hooked, and joined the team. From early days of manual permission approvals and using Claude mainly for questions and tests, the default is now trusting Claude's judgment to auto-approve while it handles large migrations and builds new products. Helping turn Claude Code into an industry-defining product was the journey of a lifetime.

Former Claude Code member Thariq noted this is the bitter lesson at work: the smartest models will need fewer and fewer tokens, they won't need to verify their work or check for bugs, they will just be right. OpenAI Codex & ChatGPT's Thibault Sottiaux added that agents today are incredible yet still so frustrating. Things will change again and again. We're still very early and can't wait to crack the flow state. Long journey ahead.
查看原文 →查看原文 →查看原文 →

Sam Altman:人类必须掌控未来,不能让超级智能决定奖励什么Sam Altman: Humans Must Stay in Control of the Future

OpenAI CEO Sam Altman 明确表态:人们应始终掌控未来。OpenAI 反对将人类角色降级为“AI 系统认为值得奖励的事情”。人们不应把未来的控制权交给超级智能,这一要求最终应约束 AI 进步。
OpenAI CEO Sam Altman stated that people should always be in control of the future. OpenAI is against the notion that it would be acceptable for people to be relegated to a role of doing things that “AI systems think makes sense to reward in humans”. People should not cede control of the future to superintelligence, and that requirement should ultimately constrain AI progress.
查看原文 →

Aaron Levie:AI 产出差距达到历史最高,前沿用户可获 5-10 倍生产力提升Aaron Levie: AI Output Divergence at All-Time High, 5-10X Gains at the Frontier

Box CEO Aaron Levie 指出,尽管 AI 采用面最广,产出差距却可能达到历史最高。几年前即使在编码这类能力对齐领域,顶尖用户也只能获得 30-50% 生产力提升,还要花大量时间清理模型错误。如今模型能处理更长任务、干预更少,agentic harness 支持多 agent 并行,生产力上限已从 30-50% 跃升到 5-10 倍。随着 recursive self-improvement,这一差距只会继续拉大。
Box CEO Aaron Levie observed that the divergence in AI output may be the highest it’s ever been even as adoption is the broadest. A few years ago the best productivity gains in a capability-aligned field like coding were maybe 30-50% if you were very good at it, with significant time spent cleaning up mistakes and limited context windows. Now models handle far longer tasks with limited human interaction, harnesses support agentic multi-agent work, and the upper limits on productivity gains have gone from 30-50% to likely 5-10X depending on the field. This distance is only getting wider with the rate of progress from RSI.
查看原文 →

企业 Agent 与 Vibe Coding 进入混乱采用期Enterprise Agents and Vibe Coding Enter Chaotic Adoption Phase

Replit CEO Amjad Masad 表示,我们正深陷企业采用 agents 与 vibe coding 的混乱阶段。
Replit CEO Amjad Masad noted we are deep in the chaotic phase of enterprise adoption of agents and vibe coding.
查看原文 →

💰 创业成功案例

Nebius CTO:在 AWS 真正觉醒前抢跑 AI 云赛道Nebius CTO: Racing the Hyperscalers Before AWS Fully Wakes Up

The Takeaway:真正的护城河是软件平台能力,而非单纯 GPU 供应;独立 AI 云仍有窗口,因为 hyperscalers 受内部需求和芯片短缺制约。

Nebius CTO(前 Yandex 核心团队)Danila 分享了从 Yandex 分离后的 AI 云路径。他们拒绝“Neo Cloud”标签,自称 AI Cloud,目标直接对标 AWS 与 Google Cloud。核心信念是软件是护城河:同一平台既能服务数万 GPU 互联集群,也能服务半小时单 GPU 的自助用户。硬件团队来自 Yandex 大规模运维经验,软件则完全重写,软件工程团队规模远超硬件。

新一代 NVIDIA 芯片上线过程混乱但常态化:原型不断迭代、实验室样品与最终设计差异大、虚拟化/存储/网络适配、有时甚至是全球首个大规模集群并回馈补丁给 NVIDIA。客户差异主要不在规模或 workload,而在成熟度:新实验室只在乎 GPU/存储/性能/可用性,甚至不在乎地点;企业客户则要求合规、安全审计与源码审查。

Inference 优化远未解决,行业“几乎什么都没优化好”。Disaggregated prefill/decode 仍处原型与手工阶段。瓶颈不是电力、冷却或 HBM,而是执行能力——项目经理在多重短缺中导航的能力本身就是护城河。Hyperscalers 行动慢,是因为芯片不够用、必须优先内部研究与产品,且旧基础设施改造成本高。他个人最大误判是低估单一客户规模(如今 30MW 有时不够一个客户,共享存储吞吐可达 700GB/s)。未来关注点转向 agent 结果优化与 harness,而非单纯更大模型。
The Takeaway: The real moat is software platform capability, not just GPU supply; independent AI clouds still have a window because hyperscalers are constrained by internal demand and chip shortages.

Nebius CTO Danila (from the original Yandex core team) described the path after the clean separation from Yandex. They reject the “Neo Cloud” label and call themselves the AI Cloud, with the next competitive horizon being AWS and Google Cloud. The founding bet is that software is the moat: the same platform must serve both tens-of-thousands-GPU interconnected clusters and a self-serve customer running one GPU for half an hour.

Hardware expertise came from Yandex-scale operations; the software stack was rewritten from scratch and the software org grew far larger than the hardware team. Bringing new NVIDIA generations live is a continuous iteration mess—prototype changes, lab samples that differ from earlier designs, virtualization/storage/networking adaptation, and occasionally being among the first to stand up large clusters and feed patches back to NVIDIA.

Customer differences are less about scale or workload type and more about maturity: new labs care mainly about GPUs, storage, performance and availability (location often irrelevant); enterprises demand compliance, security posture and source-code reviews. Inference optimization is far from solved—“as an industry we didn’t optimize shit.” Disaggregated prefill/decode remains prototype and white-glove. The single bottleneck is ability to execute amid shortages of everything; experienced project managers who can hold the full context become the moat. Hyperscalers moved slower because chip allocations are insufficient for both internal needs and external market, and retrofitting infrastructure built for other workloads is a massive investment not yet justified by the size of the business. His biggest personal underestimate was client scale (30 MW sometimes insufficient for one customer; shared storage routinely hitting 700 GB/s). Looking ahead he cares more about outcome optimization and agent harnesses than larger models.
查看原文 →

SurrealDB 4:面向 Agentic 时代的现代多模型数据库SurrealDB 4: The Modern Database for the Agentic Era

FirstMark VC Matt Turck 指出 SurrealDB 4 是重大时刻。构建数据库是漫长苦活:开源初期热度之后是数年不那么光鲜的打磨。现在它已是真正的世界级产品——多模型(vector、graph、document)、ACID、重建引擎、原生 Postgres/MySQL 支持,并借助 Cloudflare 实现峰值 100 亿+ 行/秒的扩展性。这是 agentic 时代的现代数据库。
FirstMark VC Matt Turck called SurrealDB 4 a major moment. Building a database is an incredible grind—after the initial open-source pop come years of unglamorous hard work. SurrealDB 4 is now that product: multi-model (vector, graph, document), ACID compliant, with a rebuilt engine, native Postgres/MySQL support and massive scalability via Cloudflare handling 10B+ peak rows/sec. The modern database for the agentic era.
查看原文 →

🛠️ 开发者工具与技巧

Claude 企业插件自动连接与 Vercel 安全 Agent 全面开放Claude Auto-Connects Company Plugins; Vercel Ships Security Agent

Anthropic 的 Boris Cherny 宣布:团队与企业管理员可预先配置公司插件,成员首次使用时自动连接,无需逐个登录。

Vercel CEO Guillermo Rauch 推出内部安全 agent(代号 @hacker),现已对所有仓库开放(不限于 Vercel 托管)。每一行代码自动审计,能抓住其他 agent 漏掉的问题。该能力由安全团队全权开发,相当于把安全团队下载成软件。同时推荐 CLI 加入 doctor 子命令,Vercel CI 可加速任意仓库,Security Review 提供保护。
Anthropic’s Boris Cherny announced that Team and Enterprise admins can set up company plugins so teammates are connected automatically the first time they use them—no more signing in one by one.

Vercel CEO Guillermo Rauch revealed the internal security agent (lovingly called @hacker) is now available for every repo, whether hosted on Vercel or not. Every line of code is audited automatically and catches issues other agents miss. It was built entirely by the security team—the closest thing today to downloading a security team as software. He also recommends shipping doctor subcommands in CLIs; Vercel CI accelerates any repo and Security Review protects them.
查看原文 →查看原文 →查看原文 →查看原文 →

Claude Code 实用技能与多 Agent 日常工作流Claude Code Power Skills and Multi-Agent Daily Workflows

FPV Ventures partner Nikunj Kothari 分享两个仍被低估的 Claude Code 默认技能:/insights 与 /playground。他每两周用 cron 跑 /insights,结合 X 书签形成自我改进循环;/playground 用于可视化架构、模拟学习循环、与 Codex 协作找死代码与热点路径。他同时高频使用 ChatGPT Codex、Claude 桌面、Bot 与 Muse,各自在模型与体验上推进前沿。
FPV Ventures partner Nikunj Kothari highlighted two still-underused default Claude Code skills: /insights and /playground. He runs /insights on a bi-weekly cron paired with X bookmarks for a self-improving loop on setup upgrades; /playground visualizes architectures, simulates learning loops, and pairs with Codex to find dead code and hot paths. He regularly reaches for ChatGPT Codex, Claude desktop, Bot and Muse apps—each pushing the frontier on models plus customer experience.
查看原文 →查看原文 →

Grok 电话功能:五类高价值真实场景Grok Phone Feature: Five High-Value Real-World Use Cases

Peter Yang 推荐 Grok Bot 新电话功能(目前限美国号码),并给出五个实用场景:协商账单、取消订阅、代为排队等待、预订餐厅、批量获取承包商报价。他录制了真实通话演示视频。
Peter Yang highlighted Grok Bot’s new phone-call feature (US numbers only for now) with five practical use cases: negotiate any bill, cancel subscriptions, wait on hold, make reservations, and collect contractor quotes. He shared a video of a real call demonstration.
查看原文 →

AI Engineer 大会最终主题:Raising your AmbitionAI Engineer Final Track Theme: Raising your Ambition

Swyx 宣布 AIE 年底最终 track 主题为 Raising your Ambition,嘉宾包括 jxnlco、trq212、Jarred Sumner、Sourcegraph 的 sqs、Armin Ronacher、dexhorthy、StrongDM 的 BuiltByJustin,以及 echelon 自克隆整个 Adobe 套件后的首次重要公开演讲。线下 1000 名顶尖 agentic 工程师,线上直播覆盖 200 万 MAU。
Swyx announced the final AIE track theme “Raising your Ambition” featuring jxnlco, trq212, Jarred Sumner, Sourcegraph’s sqs, Armin Ronacher, dexhorthy, StrongDM’s BuiltByJustin, and a special keynote from echelon in his first major appearance since cloning the entire Adobe suite. 1000 top agentic engineers in person in SF, livestreamed to 2M MAU.
查看原文 →

🌍 其他动态

AI 原公司人才流动加速与工作可持续性讨论Accelerating Talent Churn at AI-Native Companies

Peter Yang 观察到,过去离职未满 2-3 年会被视为跳槽,如今顶尖人才在 AI 原公司工作 1-2 年甚至更短就离开已越来越常见。原因多样,但他开始思考这些岗位对有孩子的人是否可持续。
Peter Yang noted that leaving a company before 2-3 years used to look like job-hopping; it now feels increasingly normal for very talented people to leave AI-native companies after 1-2 years or less. People leave for many reasons, but he wonders how sustainable these jobs are, especially for parents.
查看原文 →

Google Labs Dreambeans 登陆印度Google Labs Rolls Out Dreambeans in India

Google Labs 宣布 Dreambeans 已向印度用户开放,并征集下一站建议。
Google Labs announced Dreambeans is now available to users in India and asked where it should roll out next.
查看原文 →