AI Builders Digest — 2026-11-05

2026-11-05

AI Builders Digest — 2026-11-05

X / TWITTER

Swyx (AI Engineer / Latent Space)

Swyx announced the final “Raising Your Ambition” track for AI Engineer, gathering 1,000 agentic engineers in San Francisco and reaching a much larger livestream audience. He also highlighted Anthropic’s “dark factory” approach, where roughly 30,000 code changes a day are paired with independent verification, reinforcing that high-throughput agentic coding still needs a separate quality loop.

Swyx 公布了 AI Engineer 大会最后一个主题赛道“Raising Your Ambition”，计划在旧金山聚集 1,000 名 agentic engineer，并通过直播触达更大规模的线上观众。他还重点介绍了 Anthropic 的“dark factory”实践：每天约 30,000 次代码变更，同时引入独立验证机制。这说明高吞吐的 agentic coding 仍必须配套独立质量闭环。

Boris Cherny (Claude, Anthropic)

Boris Cherny said Claude can now auto-connect employees to company plugins. Team and Enterprise admins configure integrations once, and teammates are connected automatically on first use, removing repetitive authentication from enterprise plugin adoption.

Boris Cherny 表示，Claude 现在可以自动将员工连接到企业插件。Team 和 Enterprise 管理员只需统一配置一次，团队成员首次使用时即可自动连接，减少企业推广插件时反复登录认证的摩擦。

Thibault Sottiaux (Codex & ChatGPT, OpenAI)

Thibault Sottiaux argued that today’s agents are simultaneously remarkable and frustrating. His key point is that the category remains early: interfaces and capabilities will keep changing until agents can restore a reliable engineering flow state instead of repeatedly interrupting it.

OpenAI 的 Thibault Sottiaux 认为，今天的 agents 既惊艳又令人沮丧。核心判断是这个领域仍处于早期：界面和能力还会持续变化，直到 agents 能稳定帮助工程师进入心流，而不是频繁打断工作。

Peter Yang (AI educator and interviewer)

Peter Yang highlighted Grok Bot’s ability to make US phone calls for tasks such as negotiating bills, canceling subscriptions, waiting on hold, booking restaurants, and collecting contractor quotes. He also observed that even highly talented people are leaving AI-native companies after one or two years, raising a practical question about whether the pace and intensity of these jobs are sustainable, especially for parents.

AI 教育者 Peter Yang 重点介绍了 Grok Bot 的美国电话代办能力，包括议价、取消订阅、排队等待客服、预订餐厅和收集承包商报价。他还观察到，不少优秀人才在 AI-native 公司工作一两年甚至更短时间后便离开，由此提出一个现实问题：这类岗位的节奏与强度是否可持续，尤其对有孩子的员工而言。

Cat Wu (formerly Claude Code & Cowork, Anthropic)

Cat Wu announced her departure from Anthropic after helping turn Claude Code from a small internal tool into an industry-defining product. Her retrospective captures a major behavioral shift: users moved from manually approving every permission and asking Claude for tests to trusting its judgment for auto-approval, large migrations, and complete product builds.

Cat Wu 宣布离开 Anthropic。她参与了 Claude Code 从内部小工具成长为行业标志性产品的过程。她的回顾呈现了一次关键行为转变：用户从逐项手动批准权限、主要让 Claude 回答问题和写测试，发展到信任其自动审批，并执行大型迁移和完整产品构建。

Thariq (formerly Claude Code, Anthropic)

Thariq framed model progress through the “bitter lesson”: smarter models may need fewer tokens and progressively less explicit verification because they will increasingly produce correct work directly. It is a bold counterpoint to the current emphasis on elaborate checking loops.

Thariq 用“bitter lesson”解释模型进步：更聪明的模型可能消耗更少 token，也会越来越少依赖显式验证，因为它们能够直接产出正确结果。这一判断与当前强调复杂校验循环的主流实践形成鲜明对照。

Amjad Masad (CEO, Replit)

Replit CEO Amjad Masad characterized enterprise adoption of agents and vibe coding as being deep in a chaotic phase. The signal is not that adoption has stalled, but that organizations are deploying faster than their governance, workflows, and operating models can stabilize.

Replit CEO Amjad Masad 将企业采用 agents 和 vibe coding 的现状概括为“深度混乱期”。这并非采用停滞，而是部署速度已经超过治理机制、工作流和组织运营模式的稳定速度。

Guillermo Rauch (CEO, Vercel)

Vercel CEO Guillermo Rauch launched Vercel Security Review, productizing an internal security agent that automatically audits every line of code and works even for repositories not hosted on Vercel. His argument is that security cannot depend on every engineer remembering a skill or adding a “make this secure” prompt; the control must be systematic. He also recommended adding diagnostic doctor subcommands to developer CLIs.

Vercel CEO Guillermo Rauch 推出了 Vercel Security Review，将内部安全 agent 产品化，对每一行代码进行自动审计，而且无需将仓库托管在 Vercel。他的核心观点是：安全不能依赖每位工程师记得安装某个 skill，或在 prompt 中补一句“确保安全”，而应成为系统化控制。他还建议开发者 CLI 普遍加入诊断用的 doctor 子命令。

Aaron Levie (CEO, Box)

Box CEO Aaron Levie argued that AI adoption is broadening while the productivity gap between frontier users and everyone else is widening. He estimates the upside has moved from roughly 30–50% two years ago to potentially 5–10× today in suitable fields, driven by longer autonomous tasks, better agent harnesses, improved training data, and multi-agent workflows.

Box CEO Aaron Levie 认为，AI 的采用面正在扩大，但前沿用户与普通用户之间的生产力差距也在扩大。他估计，在适配领域中，生产力提升上限已从两年前约 30–50% 上升到今天潜在的 5–10 倍，背后驱动力包括更长时间的自主任务、更好的 agent harness、更优训练数据以及多 agent 工作流。

Matt Turck (Partner, FirstMark Capital)

FirstMark partner Matt Turck highlighted SurrealDB 4 as the payoff from years of less-visible database engineering after its initial open-source momentum. The release combines vector, graph, and document models, ACID compliance, a rebuilt engine, native Postgres/MySQL support, and claimed peak scaling through Cloudflare, positioning it as infrastructure for agentic applications.

FirstMark 合伙人 Matt Turck 将 SurrealDB 4 视为开源项目早期走红后多年基础工程投入的兑现。该版本整合 vector、graph 和 document 多种模型，支持 ACID，重构核心引擎，原生兼容 Postgres/MySQL，并借助 Cloudflare 实现高峰扩展，定位为 agentic 应用的现代数据库基础设施。

Nikunj Kothari (Partner, FPV Ventures)

FPV Ventures partner Nikunj Kothari recommended two underused Claude Code skills: /insights for periodically identifying improvements to his setup, and /playground for visualizing repositories and workflows. He runs /insights on a biweekly cron and pairs Codex with Claude Code to identify dead code, hot paths, and latency improvements, creating a self-improving software-factory loop.

FPV Ventures 合伙人 Nikunj Kothari 推荐了两个被低估的 Claude Code skills：用 /insights 定期发现开发环境的改进点，用 /playground 可视化仓库与工作流。他每两周通过 cron 运行一次 /insights，并组合 Codex 与 Claude Code 查找 dead code、hot paths 和延迟优化机会，形成自我改进的软件工厂闭环。

Sam Altman (OpenAI)

Sam Altman stated that people must remain in control of the future and rejected a world where humans are reduced to doing whatever AI systems decide is worth rewarding. In his framing, preserving human control should be an explicit constraint on superintelligence and therefore on the pace and direction of AI progress.

Sam Altman 表示，人类必须始终掌握未来的控制权，并反对把人类降格为只做“AI 系统认为值得奖励之事”的世界。在他的框架中，保留人类控制权应成为 superintelligence 的明确约束，也应约束 AI 进步的速度与方向。

PODCASTS

Unsupervised Learning — Ep 96: Nebius CTO on Racing the Hyperscalers Before AWS Wakes Up

The Takeaway: Nebius is betting that a software-first AI cloud can grow into a peer competitor before hyperscalers redesign their stacks around AI workloads.

Nebius CTO’s infrastructure thesis is that GPUs alone are not the moat. The winning platform must serve both enormous interconnected clusters and self-service customers through one coherent software layer. AI workloads also reshape the rest of the stack: customers can demand hundreds of gigabytes per second from shared storage, while new rack-scale GPU systems force cloud operators to rethink virtualization, networking, maintenance, and isolation.

The market opportunity exists because AWS, Azure, and Google must divide scarce capacity between internal AI needs and external customers, while also adapting infrastructure designed for older workloads. Nebius aims to exploit that window, but expects hyperscalers to focus aggressively once chip scarcity eases. The company’s operational stance on coding agents is notably conservative: teams may use them freely, but humans retain responsibility for every pull request and on-call consequence, while fully autonomous agents stay away from core services. The most important research frontier may therefore be outside the model itself: better tools, harnesses, and quality gates that increase the probability of reaching a reliable outcome.

Memorable quote: “Whatever you've done with the agent, it's still your result. It's your responsibility.”

核心结论： Nebius 押注以软件为核心的 AI cloud，争取在 hyperscalers 完成面向 AI workload 的全栈改造之前，成长为可与其同级竞争的玩家。

Nebius CTO 的基础设施判断是：GPU 本身不是护城河。真正有竞争力的平台，必须通过统一的软件层，同时服务超大型互联集群与按需使用的自助客户。AI workload 也在重塑整个技术栈：客户对共享存储的吞吐需求可达每秒数百 GB，而新一代 rack-scale GPU 系统迫使云厂商重新设计虚拟化、网络、维护和隔离机制。

市场窗口来自 hyperscalers 的结构性矛盾：AWS、Azure 和 Google 既要把稀缺算力留给内部 AI 业务，又要服务外部客户，同时还需改造原本为传统 workload 设计的基础设施。Nebius 希望利用这段窗口期，但也预计芯片短缺缓解后，巨头会全面加码。Nebius 对 coding agents 的运营态度相当克制：员工可以自由使用，但每个 pull request 和每次 on-call 后果仍由人负责，完全自主的 agent 暂时不得触碰核心服务。因此，真正关键的研究前沿可能不只在模型内部，而在模型之外：通过更好的工具、harness 和 quality gates，提高 agent 稳定达成目标的概率。

代表性原话：“Whatever you've done with the agent, it's still your result. It's your responsibility.”

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders