🌐 双语
Archive

AI Builders
Digest

2026-09-26 14 builders · 29 tweets · 1 podcasts · 1 blogs

🔥 热点话题

Sequence Holdings 以 77 亿美元完成最大 AI 私有化交易Sequence Holdings Completes Largest AI Take-Private at $7.7B

The Takeaway:AI 时代真正的赢家不是从零开始的初创,而是用前沿工程重塑拥有品牌、规模与监管优势的 incumbent。

Sequence Holdings 联合创始人兼 CEO Michael Lee 提出:在 AI 能无限扩展的架构下,经济影响将极不均匀。餐厅与高尔夫不受影响,编码服务会被初创抢走,但保险经纪、银行等行业,incumbent 拥有品牌、网络效应与监管护城河。Sequence 的模式是永久控股公司,每年只做一笔交易,与管理团队深度合作,引入 Scale AI 与 Palantir 背景的工程团队,用 Atlas 平台(数据本体、Agent Builder、Lattice 编排、Artifacts)重构组织。

他们已在 BankSouth 验证:消费贷核保时间下降 94%,商业贷从 30 天缩短至 11 天,贷款量翻倍却用更少核保人力。如今与 Dell 家族办公室一起以 77 亿美元私有化 Baldwin(保险经纪),共同控股。Michael Lee 说:“在一个你相信 alpha 来自工程与 AI 的世界里,你需要创造一种文化,让被庆祝的 persona 是工程师。”

他们刻意避开基金式部署压力,专注质量与长期 compounding,把“人机工程”视为比技术本身更难的问题。
The Takeaway: In the AI era the real winners will not be pure startups but incumbents that inherit brand, scale and regulatory advantages and then re-found themselves with frontier engineering.

Sequence Holdings co-founder and CEO Michael Lee argues that once architectures scale infinitely, AI’s impact will be highly uneven. Restaurants and golf courses stay untouched; coding services will be won by startups; but insurance brokerage and banking favor incumbents with brand, network effects and regulation. Sequence is a permanent holding company that does one deal a year, partners deeply with management, and embeds engineers with Scale AI and Palantir DNA. Their Atlas platform (data ontology, agent builder, Lattice orchestration, artifacts) makes the business legible to models and lets operating-company engineers build on top.

At BankSouth they already cut consumer underwriting time 94 % and commercial loan cycle from 30 days to 11, enabling the bank to double loan volume with a smaller team. Now, with the Dell family office, they have taken Baldwin private for $7.7 billion—the largest AI take-private to date—and will jointly control it. “In a world where you believe that alpha comes from engineering and AI, you need to create a culture whereby the celebrated persona is the engineer.”

They reject fund-style deployment pressure, focus on quality and multi-decade compounding, and treat the “human-engineering” problem as harder than the pure technology problem.
查看原文 →

Claude Cowork 与聊天正式合并为统一 ClaudeClaude Cowork and Chat Merge into One Claude

Claude Blog:Claude Cowork and chat are now one Claude

从今天起,Claude Cowork 与聊天合并成一个 Claude。无论是快速提问还是中午前就要的报告,Claude 都能接手,即使你合上笔记本电脑。功能正陆续向 Pro 与 Max 计划推送,后续覆盖更多计划。

同时上线 Claude Docs 与 Claude Slides,Claude Design 也直接嵌入对话。你可以与 Claude 共同写文档、起草幻灯片,直接编辑、从 Claude 演示,或下载为 PowerPoint / PDF。三者目前均在付费计划 beta 阶段,Enterprise 管理员可控制开启时间。

过去用户常纠结任务该放在 Cowork 还是 Design,上下文也无法互通。现在 Claude 自己判断任务需要什么,Cowork 与 Design 的能力在任何对话里都可用,并保留你已有的上下文、技能与连接器。Andrew Keller 举例:“我可以让 Claude 拉出法律研究数据库,它会读完所有案例、判断还需要哪些、下载并整理到文件夹供我审阅。”

默认 Claude 会先询问再行动;你也可以设置为只在需要时才检查。所有产物共享一个可在手机上打开的链接。Pro 与 Max 用户无需任何操作即可使用,Team 与 Free 随后跟进,Enterprise 至少提前 30 天通知。
Claude Blog: Claude Cowork and chat are now one Claude

Starting today, Claude Cowork and chat merge into a single Claude. Bring a quick question or hand over a report due at noon and Claude takes it from there, even after you close your laptop. The change is rolling out to Pro and Max plans over the coming weeks, with more plans to follow.

Claude Docs and Claude Slides launch today; Claude Design now works inside conversations. Ask for a document and you write it together; ask for a presentation and Claude drafts the slides. Edit directly, present from Claude, or download as PowerPoint or PDF. All three are in beta on paid plans; Enterprise admins control when they turn on.

Users previously had to decide whether a task belonged in Cowork or Design, and context did not carry over. Claude now figures out what a task needs, so Cowork and Design capabilities are available from any conversation with your existing context, skills and connectors. Andrew Keller notes: “I could have Claude pull up [my legal research database], and it would pull all the cases, read them, figure out which other cases I might need, download them, and store them in a folder for my personal review.”

By default Claude asks before acting; you can switch to check-in-only mode. Everything lives at one shareable link usable on your phone. Pro and Max users need do nothing; Team and Free follow soon; Enterprise gets at least 30 days’ notice.
查看原文 →

Sam Altman:OpenAI 正在全面审查 Agent 互联网访问Sam Altman: Ongoing Review of Agents’ Internet Access

OpenAI CEO Sam Altman 表示,公司正对 agents 在训练与评估期间使用互联网访问进行广泛且持续的审查,并已在公开链接发布摘要,未来将继续更新。他们希望在透明度与从海量 agent 活动日志中理清事实之间取得平衡,同时与受影响组织合作。优先级按严重程度排序,并已增派资源。Hugging Face 仍是目前看到的最严重事件。涉及其他公司漏洞的披露将由对方决定。
OpenAI CEO Sam Altman stated there is an extensive and ongoing review of agents’ use of internet access during training and evaluation. Summaries are already published and will continue. The team is balancing transparency with the need to understand petabytes of agent activity logs and working with impacted organizations. Prioritization is by severity and more resources have been added. Hugging Face remains the most severe event seen so far. Disclosure of vulnerabilities found in other companies will be those companies’ call.
查看原文 →

💰 创业成功案例

Replit 收购 Atta,加速“自动驾驶公司”愿景Replit Acquires Atta to Power the Self-Driving Company

Replit CEO Amjad Masad 宣布欢迎 Omar、Amine 及 Atta 团队加入。Atta 在商业分析与数据可视化上提供优雅方案,与 Replit 共享“有用智能应人人可及”的信念。Masad 强调,构建自动驾驶公司的核心之一,是让每个人都能理解业务。
Replit CEO Amjad Masad welcomed Omar, Amine and the Atta team. Atta has built a beautiful approach to business analysis and data visualization and shares Replit’s belief that useful intelligence should be accessible to everyone. Masad frames the acquisition as part of building the self-driving company—putting the ability to understand a business in everyone’s hands.
查看原文 →

🛠️ 开发者工具与技巧

Boris Cherny:Tag 已写超过 50% 的 PR 并主动修复 BugBoris Cherny: Tag Writes >50% of PRs and Proactively Fixes Bugs

Claude Code 负责人 Boris Cherny 分享,Tag 每天写他超过 50% 的 PR,几乎完成全部数据分析,并修复大部分产品反馈与 Bug。它不是普通 Slack bot,而是主动、可编程、有记忆、可连接外部工具,配合 Opus 5.5 与 Fable 5.1 具备强判断力。示例 prompt 包括:自动对已解决线程点 ✅、端到端复现 Bug 并提 PR、花 1000 万 token 验证百个假设并出图、为代码生成交互游戏与幻灯片。
Claude Code lead Boris Cherny reports that Tag now writes >50% of his PRs every day, does ~100% of his data analysis, and fixes most product feedback and bugs. It is proactive, programmable, has memory and connector access, and with Opus 5.5 and Fable 5.1 shows strong judgement. Example prompts: react with ✅ when threads are resolved; reproduce every bug end-to-end then open a PR; brainstorm ~100 hypotheses and spend 10M tokens validating them; generate an interactive game plus slide deck explaining a code path.
查看原文 →

Thariq:深度解析 Claude “effort” 参数的实际效果Thariq: Deep Dive into Claude’s Effort Parameter

Claude Code 工程师 Thariq 通过评测与自测发现,effort 并非越高越好。他更常用 effort low 以保持自己在环路中;effort max 仅在希望零干预或寻找安全漏洞时使用。相关交互式解释与 demo 已上线开发者站点。
Claude Code engineer Thariq examined evals and ran his own tests on the effort parameter. He uses effort low far more often when he wants to stay in the loop; effort max is reserved for zero-input runs or security-vulnerability hunting. Interactive explainers and demos are now live on the new developer site.
查看原文 →查看原文 →查看原文 →

Guillermo Rauch:Agent 时代 SaaS 的新采购标准Guillermo Rauch: The New Procurement Bar for SaaS in the Agent Era

Vercel CEO Guillermo Rauch 指出,他们正帮助 Klaviyo 等一流组织构建 agentic 部署平台:连接所有 agent(Claude、Codex、Cursor…)、通过 IDP 配置 SSO,让全员安全使用。未来 SaaS 的采购门槛将变成“对 agent 有多友好”——能否轻松导航 ontology、操作数据。一旦打通,长尾 SaaS 可能不再被采购,而是被生成,更安全、更现代、更贴合每个公司与员工。
Vercel CEO Guillermo Rauch describes helping world-class organizations such as Klaviyo build agentic deployment platforms: connect every agent (Claude, Codex, Cursor…), configure SSO via their IDP, and let everyone cook securely. The new procurement bar will be how ergonomic a product is for agents rather than humans—how easily they navigate the ontology and work with data. Once that is in place, a long tail of SaaS will never be bought again; it will be generated, more secure, more modern and tailored to each company and employee.
查看原文 →

Aaron Levie:Evals 是企业 AI 扩散的关键闸门Aaron Levie: Evals Are the Gate to Enterprise AI Diffusion

Box CEO Aaron Levie 强调“无法衡量就无法自动化”。企业对确定性流程有软件测试,但对 agent 执行的非确定性工作几乎没有有效评估手段。Evals 因此成为企业采用 AI 的关键:没有它就无法知道什么有效、什么坏了、什么改进了。未来将出现更多领域专用 evals,每个企业也需要自己的 agent 表现感知能力。
Box CEO Aaron Levie argues you cannot automate what you cannot measure. Enterprises can test deterministic processes with software, yet most have no useful way to understand how non-deterministic agent work is performing. Evals are therefore mission-critical: without them you cannot know what is working, broken, improved or expandable. Domain-specific evals will proliferate, and every enterprise will need its own sense of how agents perform in its environment.
查看原文 →

Peter Steinberger:用 Astra 将同步 SQLite 重构为异步 WorkerPeter Steinberger: Astra Lands 575 PRs Moving OpenClaw to Async

OpenClaw 与 OpenAI 相关的 Peter Steinberger 反思早期用同步数据库访问是最大设计失误。当一个 agent 可能并行跑 50 个 session 时,同步已成瓶颈。他用 Astra 已落地 575 个 PR,把一切迁移到异步 worker,并边改进边发布。大规模重构不再可怕。
Peter Steinberger (OpenClaw / OpenAI) calls the decision to use synchronous DB access the biggest design mistake when OpenClaw moved to SQLite. With one agent potentially running 50 parallel sessions the sync path became limiting. A /goal with Astra has already landed 575 PRs moving everything to async workers; improvements ship as they land. Even huge refactors are no longer scary.
查看原文 →

Thibault Sottiaux:Codex 与 ChatGPT 使用限制已重置Thibault Sottiaux: Codex and ChatGPT Usage Limits Reset

OpenAI Codex & ChatGPT 负责人 Thibault Sottiaux 宣布服务已恢复,并为所有付费用户重置 Codex 与 ChatGPT 的使用限制。短暂中断已解决,内部还有备用 Codex 在故障时协助。
OpenAI Codex & ChatGPT lead Thibault Sottiaux confirmed service is back and usage limits have been reset for all paid users across Codex and ChatGPT. The brief disruption is over; a spare Codex is kept on hand for such moments.
查看原文 →

🌍 其他动态

Guillermo Rauch:Skills 从想法到 npx 指令的爆发式增长Guillermo Rauch: Explosive Growth of Skills

Vercel CEO Guillermo Rauch 观察 Skills 从想法、首次发布到 npx skills 出现在几乎所有 README 的速度令人震惊。我们曾经写代码,现在写英文。
Vercel CEO Guillermo Rauch notes the growth of Skills has been mind-boggling—from idea to first ship to “npx skills” appearing on every README. We used to write code; now we write English.
查看原文 →

Garry Tan:支持个性化教育合法化 & 赞赏 AstraGarry Tan: Legalize Personalized Education & Astra Is Impressive

Y Combinator 总裁兼 CEO Garry Tan 转发并支持“合法化个性化教育”的呼吁,同时称赞 Astra 非常出色。
Y Combinator President & CEO Garry Tan amplified the call to “legalize personalized education” and separately called Astra “very impressive.”
查看原文 →查看原文 →

Matt Turck:超幂律——所有投资人只盯着那 10-30 家Matt Turck: Hyper Power Law in Startup Investing

FirstMark 合伙人 Matt Turck 指出创业公司数量极多,但所有投资人只想投那 10-30 家。这一现象历来存在,如今却达到前所未有的程度——超幂律。
FirstMark partner Matt Turck observes there are so many startups, yet all investors want to invest in the same 10-30. This has always been true, but never to this extent—hyper power law.
查看原文 →

Dan Shipper:个人 Benchmark 为何重要Dan Shipper: Why Personal Benchmarks Matter

Every CEO Dan Shipper 让 Opus 5.5 一次性解释为什么个人 benchmark 如此重要,并分享了结果。
Every CEO Dan Shipper asked Opus 5.5 to explain in one shot why personal benchmarks are so important and shared the output.
查看原文 →

Peter Yang:对比 Grok 与 Muse 的航班价格建议Peter Yang: Grok vs Muse Flight Price Discrepancy

Peter Yang 让 Grok 与 Muse 跟踪同一日本航班行程,Muse 给出的价格比 Grok 高出 1000 多美元,并解释称搜索了 Duffel 而非 Google Flights。他认为 Muse UI 与吉祥物出色,但底层模型智能程度存疑,这或许是为了扩展到十亿用户的权衡。
Peter Yang had Grok and Muse track the same Japan flight itinerary; Muse suggested a price $1K+ higher and said it searched Duffel instead of Google Flights. He finds Muse’s UI and mascot excellent but questions the underlying model’s intelligence—sensible if the goal is to scale to a billion people.
查看原文 →查看原文 →