🌐 双语
Archive

AI Builders
Digest

2026-08-29 15 builders · 30 tweets · 1 podcasts · 2 blogs

🔥 热点话题

OpenAI终止与Cursor的合作OpenAI Ends Partnership with Cursor

OpenAI Codex与ChatGPT团队成员Thibault Sottiaux宣布,因信任问题,OpenAI将于11月12日停止通过Cursor提供模型访问。用户仍可通过自己的OpenAI API密钥或OpenAI的IDE扩展继续使用。OpenAI强调将继续支持广泛的工具生态系统和开源计划。此消息引发广泛讨论,Replit CEO Amjad Masad立即回应,表示Replit可免费访问OpenAI模型,并愿资助企业从Cursor迁移。Anthropic的Claude Code负责人Thariq也表示将继续与Cursor合作。

OpenAI Codex and ChatGPT team member Thibault Sottiaux announced that OpenAI is ending model access through Cursor effective November 12 due to trust issues. Users can still use their own OpenAI API keys or OpenAI's IDE extensions. OpenAI remains committed to the broader ecosystem of tools and open-source initiatives. The news sparked discussion, with Replit CEO Amjad Masad offering free OpenAI model access on Replit and funding transitions for businesses. Anthropic's Claude Code lead Thariq affirmed continued partnership with Cursor.
OpenAI Codex and ChatGPT team member Thibault Sottiaux announced that OpenAI is ending model access through Cursor effective November 12 due to trust issues. Users can still use their own OpenAI API keys or OpenAI's IDE extensions. OpenAI remains committed to the broader ecosystem of tools and open-source initiatives. The news sparked discussion, with Replit CEO Amjad Masad offering free OpenAI model access on Replit and funding transitions for businesses. Anthropic's Claude Code lead Thariq affirmed continued partnership with Cursor.
查看原文 →查看原文 →查看原文 →

Anthropic修复Claude Code质量问题Anthropic Fixes Claude Code Quality Issues

Anthropic Engineering发布更新,解释过去一个月Claude Code、Claude Agent SDK和Claude Cowork响应质量下降的报告。问题源于三个独立变更:默认推理努力从high改为medium(已恢复)、会话空闲后清除旧思考块的bug(导致遗忘和重复)、以及降低冗长的系统提示指令。所有问题已在4月20日(v2.1.116)解决,API未受影响。Anthropic重置了所有订阅者的使用限制,并改进内部测试、代码审查和提示变更控制,以避免类似问题。

Anthropic Engineering detailed three separate changes that degraded Claude Code, Agent SDK, and Cowork responses for some users: switching default reasoning effort from high to medium (reverted), a bug that repeatedly cleared prior thinking after idle sessions (causing forgetfulness), and a system prompt addition limiting verbosity that hurt coding quality. All fixed by April 20 (v2.1.116); the API was unaffected. Anthropic reset usage limits for all subscribers and is tightening evals, dogfooding of public builds, and prompt change controls.
Anthropic Engineering detailed three separate changes that degraded Claude Code, Agent SDK, and Cowork responses for some users: switching default reasoning effort from high to medium (reverted), a bug that repeatedly cleared prior thinking after idle sessions (causing forgetfulness), and a system prompt addition limiting verbosity that hurt coding quality. All fixed by April 20 (v2.1.116); the API was unaffected. Anthropic reset usage limits for all subscribers and is tightening evals, dogfooding of public builds, and prompt change controls.
查看原文 →

AI产品构建的3个月半衰期AI Product Building Has a 3-Month Playbook Half-Life

Meta AI高级总监Madhu Guru指出,AI产品构建的玩法半衰期约3个月。传统产品团队发现玩法后会沿用5年,但AI实验室有意识地每几个月丢弃旧玩法,因为回报递减。AI构建者应专注定义元原则:如何持续学习市场当前和3个月后的需求,以及如何以极致紧迫感执行。团队需以发明而非榨取玩法为目标。Box CEO Aaron Levie呼应称,AI领域强烈信念的半衰期最多6个月,列举了从OSS落后到RAG已死等一系列已翻转或无共识的观点,强调保持思维灵活。

Meta AI Senior Director Madhu Guru argued that AI product playbooks have roughly a 3-month half-life. Traditional teams milk a playbook for years; AI labs deliberately discard them as returns diminish. Builders should define meta-principles for continuous market learning and extreme-urgency execution, optimizing teams for invention rather than playbook extraction. Box CEO Aaron Levie added that strongly held AI beliefs now last at most 6 months, listing cycling views from OSS lagging to RAG being dead, and urged remaining flexible amid constant change.
Meta AI Senior Director Madhu Guru argued that AI product playbooks have roughly a 3-month half-life. Traditional teams milk a playbook for years; AI labs deliberately discard them as returns diminish. Builders should define meta-principles for continuous market learning and extreme-urgency execution, optimizing teams for invention rather than playbook extraction. Box CEO Aaron Levie added that strongly held AI beliefs now last at most 6 months, listing cycling views from OSS lagging to RAG being dead, and urged remaining flexible amid constant change.
查看原文 →查看原文 →

🛠️ 开发者工具与技巧

Anthropic推出Managed Agents:解耦大脑与双手Anthropic: Scaling Managed Agents by Decoupling Brain from Hands

Anthropic Engineering介绍Managed Agents,一种托管服务,通过持久接口运行长期代理。核心是将“大脑”(Claude与harness)与“双手”(sandbox和工具)以及session日志解耦。早期将所有组件放在单一容器导致“宠物”问题:失败难恢复、调试受限、无法灵活连接客户VPC。解耦后,harness通过execute调用sandbox,失败可重建;session日志在外,支持从任意点恢复。安全上,凭证永不进入sandbox,通过代理或初始化注入。此架构使p50 TTFT下降约60%,p95下降超90%,并支持多大脑多双手扩展。接口面向未来模型,不绑定特定实现。

Anthropic Engineering introduced Managed Agents, a hosted service for long-horizon agents built around durable interfaces. The key insight is decoupling the “brain” (Claude + harness) from the “hands” (sandboxes/tools) and the session log. Coupling everything in one container created pets that were hard to recover, debug, or connect to customer VPCs. After decoupling, the harness calls sandboxes via a simple execute interface, sessions survive harness crashes via an external log, and credentials never enter the sandbox. The change cut p50 TTFT by ~60% and p95 by over 90%, while enabling many brains and many hands. The design deliberately avoids assumptions about future harnesses or models.
Anthropic Engineering introduced Managed Agents, a hosted service for long-horizon agents built around durable interfaces. The key insight is decoupling the “brain” (Claude + harness) from the “hands” (sandboxes/tools) and the session log. Coupling everything in one container created pets that were hard to recover, debug, or connect to customer VPCs. After decoupling, the harness calls sandboxes via a simple execute interface, sessions survive harness crashes via an external log, and credentials never enter the sandbox. The change cut p50 TTFT by ~60% and p95 by over 90%, while enabling many brains and many hands. The design deliberately avoids assumptions about future harnesses or models.
查看原文 →

Guillermo Rauch:Web正走向两个极端Guillermo Rauch: The Web Evolves Toward Two Extremes

Vercel CEO Guillermo Rauch观察到Web正分化为两个极端:一端是超精美的人类体验(delight、探索、品牌、娱乐),由Three.js、TypeGPU等支撑;另一端是面向agent的纯信号内容、数据和API,markdown与MCP是冰山一角。中间地带将被agent即时生成UI所取代。他同时指出MCP爆炸式增长,mcp-handler等包下载量激增,并推广eve——一个让开发者完全拥有runtime、模型、skills、工具和sandbox的Git repo式agent方案。

Vercel CEO Guillermo Rauch sees the web splitting into two extremes: ultra-fancy human experiences for delight and brand, enabled by tools like Three.js and TypeGPU, versus pure, lossless agent-centric content, data and APIs, with markdown and MCP as early signals. The middle will be subsumed by agents generating just-in-time UI. He also highlighted MCP’s explosive growth via packages like mcp-handler and promoted eve as a way to own the full intelligence stack in a Git repo.
Vercel CEO Guillermo Rauch sees the web splitting into two extremes: ultra-fancy human experiences for delight and brand, enabled by tools like Three.js and TypeGPU, versus pure, lossless agent-centric content, data and APIs, with markdown and MCP as early signals. The middle will be subsumed by agents generating just-in-time UI. He also highlighted MCP’s explosive growth via packages like mcp-handler and promoted eve as a way to own the full intelligence stack in a Git repo.
查看原文 →查看原文 →查看原文 →

Peter Yang:Grok Bot才是非技术用户的正确终点Peter Yang: Grok Bot Is the Right End State for Non-Technical Users

AI教程作者Peter Yang认为Claude Cowork和ChatGPT Work只是部分解决方案,Grok Bot才是面向非技术用户的正确最终形态。人们能直观理解“云中运行的计算机”,却很难分清ChatGPT Work与Codex的区别,或Claude Cowork的实际工作原理。他同时分享了一则时机恰当的广告。

AI educator Peter Yang’s hot take is that Claude Cowork and ChatGPT Work are only partial solutions; Grok Bot is the right end state for capable AI agents aimed at non-technical people. Users can easily grasp “a computer that runs in the cloud,” whereas most cannot explain the difference between ChatGPT Work and Codex or how Claude Cowork actually works.
AI educator Peter Yang’s hot take is that Claude Cowork and ChatGPT Work are only partial solutions; Grok Bot is the right end state for capable AI agents aimed at non-technical people. Users can easily grasp “a computer that runs in the cloud,” whereas most cannot explain the difference between ChatGPT Work and Codex or how Claude Cowork actually works.
查看原文 →

🌍 其他动态

Eon联合创始人:数据基础设施必须为AI时代重塑Eon Co-Founders: Rethinking Legacy Data Infrastructure for the AI Era

The Takeaway:在AI时代,企业最持久的护城河是自己的数据,而传统数据栈无法支撑agent和全民构建者带来的速度与风险。

Eon联合创始人Ofir Ehrlich和Gonen Stein(前CloudEndure,被AWS收购)指出,模型与算力几乎零切换成本,真正稀缺的是企业历史数据。Google以1000万美元买下破产的Spirit Airlines数据就是明证。他们构建的云数据基础能跨多云映射、分类、安全摄取结构化和非结构化数据,在保护与恢复的同时让AI模型直接查询和训练。

安全威胁已从人类勒索软件升级到“合法权限”的非人类agent,速度极快。非技术员工用Lovable等工具构建agent时,往往把敏感数据带出组织边界。旧ETL工具太 niche,无法处理agent生成的海量噪声数据。Eon的目标是让数据团队在不损害生产、安全与合规的前提下,持续获得可用数据。

“我们生活在非常有趣的时代”——每个人都能成为构建者,既是巨大机遇也是巨大风险。云迁移曾是抽象概念,AI则人人可感,董事会正强推采用,Forward Deployed Engineers和PLG因此爆发。

The Takeaway: In the AI era, a company’s most durable moat is its own data, yet legacy stacks cannot handle the velocity and risk introduced by agents and citizen builders.

Eon co-founders Ofir Ehrlich and Gonen Stein (ex-CloudEndure, acquired by AWS) argue that models and compute have near-zero switching costs; the scarce asset is proprietary historical data. Google’s $10M purchase of bankrupt Spirit Airlines’ data is only the beginning. Eon’s cloud data foundation maps, classifies and safely ingests structured and unstructured data across hyperscalers so customers can both protect it and feed it to LLMs.

Threats have shifted from human ransomware to non-human agents that already hold legitimate permissions—and the velocity is extreme. Non-technical employees building with tools like Lovable routinely expose sensitive data outside organizational boundaries. Old niche ETL tools cannot cope with the noise and volume agents generate. Eon’s approach lets data teams continuously access usable data without compromising production, security or compliance.

“We live in very interesting times”—everyone inside an organization can now become a builder, which is both powerful and dangerous. Unlike the abstract cloud transition, AI is viscerally understood by boards, driving rapid adoption via Forward Deployed Engineers and product-led growth.
The Takeaway: In the AI era, a company’s most durable moat is its own data, yet legacy stacks cannot handle the velocity and risk introduced by agents and citizen builders.

Eon co-founders Ofir Ehrlich and Gonen Stein (ex-CloudEndure, acquired by AWS) argue that models and compute have near-zero switching costs; the scarce asset is proprietary historical data. Google’s $10M purchase of bankrupt Spirit Airlines’ data is only the beginning. Eon’s cloud data foundation maps, classifies and safely ingests structured and unstructured data across hyperscalers so customers can both protect it and feed it to LLMs.

Threats have shifted from human ransomware to non-human agents that already hold legitimate permissions—and the velocity is extreme. Non-technical employees building with tools like Lovable routinely expose sensitive data outside organizational boundaries. Old niche ETL tools cannot cope with the noise and volume agents generate. Eon’s approach lets data teams continuously access usable data without compromising production, security or compliance.

“We live in very interesting times”—everyone inside an organization can now become a builder, which is both powerful and dangerous. Unlike the abstract cloud transition, AI is viscerally understood by boards, driving rapid adoption via Forward Deployed Engineers and product-led growth.
查看原文 →

Josh Woodward:Gemini NotebookLM的作者来信是亮点Josh Woodward: Author Letters in Gemini NotebookLM

Google VP Josh Woodward(负责Google Labs、Gemini App、AI Studio)分享,Gemini NotebookLM的Expert Intelligence功能中,作者来信是他最喜欢的部分之一,细节贴心。

Google VP Josh Woodward (Google Labs, Gemini App, AI Studio) highlighted that one of his favorite touches in Expert Intelligence inside Gemini NotebookLM is the letter from the author.
Google VP Josh Woodward (Google Labs, Gemini App, AI Studio) highlighted that one of his favorite touches in Expert Intelligence inside Gemini NotebookLM is the letter from the author.
查看原文 →

Zara Zhang:Slop与否取决于是否包含独特人类经验Zara Zhang: Slop Is About Unique Human Perspective, Not AI Origin

Builder Zara Zhang认为,判断内容是否为slop,与是否由AI生成关系不大,关键在于是否由具体、独特的人类经验或视角驱动。人类生成的slop同样大量存在。她同时询问在Grok Bot虚拟电脑上登录真实X账号执行任务是否有封号风险。

Builder Zara Zhang argued that whether content is slop has less to do with AI generation and more with whether it is informed by specific, unique human experiences or perspectives—there is plenty of human-generated slop. She also raised the practical question of account risk when signing into a real X account inside Grok Bot’s virtual computer.
Builder Zara Zhang argued that whether content is slop has less to do with AI generation and more with whether it is informed by specific, unique human experiences or perspectives—there is plenty of human-generated slop. She also raised the practical question of account risk when signing into a real X account inside Grok Bot’s virtual computer.
查看原文 →查看原文 →

Nikunj Kothari给创始人的建议:好BP甚至不需要提“AI”Nikunj Kothari: Best Pitches Don’t Need the Word “AI”

FPV Ventures合伙人Nikunj Kothari提醒创始人:最好的BP甚至不需要出现“AI”一词;AI也不能成为唯一的“why now”。提前处理这两点会更有说服力。

FPV Ventures partner Nikunj Kothari advised founders that the best pitches do not even need the word “AI,” and that AI cannot be the only “why now.” Getting ahead of both points strengthens a pitch.
FPV Ventures partner Nikunj Kothari advised founders that the best pitches do not even need the word “AI,” and that AI cannot be the only “why now.” Getting ahead of both points strengthens a pitch.
查看原文 →

Dan Shipper:AI没有坏主意,只有弱模型Dan Shipper: In AI There Are No Bad Ideas, Just Weak Models

Every CEO Dan Shipper写道,在AI领域没有坏主意,只有弱模型——随着指数曲线继续,每个想法都会有它的日子。他同时宣布公司已设立Head of Evals,相关工作将带来显著改变。

Every CEO Dan Shipper observed that in AI there are no bad ideas, only weak models—every idea has its day as the exponential continues. He also announced that Every now has a Head of Evals whose work is already proving transformative.
Every CEO Dan Shipper observed that in AI there are no bad ideas, only weak models—every idea has its day as the exponential continues. He also announced that Every now has a Head of Evals whose work is already proving transformative.
查看原文 →查看原文 →

Aditya Agarwal:H1-B问题可以同时成立的多重真相Aditya Agarwal: Multiple Truths on H-1B Can Be True at Once

SPC普通合伙人、前Dropbox CTO Aditya Agarwal(自身经历过F1到绿卡全流程)指出:H-1B确实存在滥用,应制止;同时美国对真正高技能人才的留用门槛过高,尤其是高技能印度人;任何国家法律都应优先本国公民。这些观点可以同时成立。

SPC General Partner and former Dropbox CTO Aditya Agarwal (who personally navigated F1 through green card) argued that several things can be true simultaneously: H-1B abuse exists and should be stopped; the U.S. makes it too hard for truly high-skilled talent to stay; high-skilled Indians face particular difficulty; and every country’s laws should prioritize its own citizens.
SPC General Partner and former Dropbox CTO Aditya Agarwal (who personally navigated F1 through green card) argued that several things can be true simultaneously: H-1B abuse exists and should be stopped; the U.S. makes it too hard for truly high-skilled talent to stay; high-skilled Indians face particular difficulty; and every country’s laws should prioritize its own citizens.
查看原文 →

Garry Tan:数据中心创造繁荣与好工作Garry Tan: Datacenters Create Prosperity and Good Jobs

Y Combinator总裁兼CEO Garry Tan强调数据中心带来繁荣和好工作,并批评旧金山部分政策与官员。

Y Combinator President & CEO Garry Tan highlighted that datacenters create prosperity and good jobs, while also criticizing certain San Francisco policies and officials.
Y Combinator President & CEO Garry Tan highlighted that datacenters create prosperity and good jobs, while also criticizing certain San Francisco policies and officials.
查看原文 →

Matt Turck:热轮中的VC尽职调查Matt Turck: VC Due Diligence in a Hot Round

FirstMark合伙人Matt Turck分享了热轮融资中VC尽职调查的现实图景。

FirstMark partner Matt Turck shared a snapshot of what VC due diligence looks like in a hot round.
FirstMark partner Matt Turck shared a snapshot of what VC due diligence looks like in a hot round.
查看原文 →

Amjad Masad:增长Agent仍是未被充分挖掘的潜力Amjad Masad: Growth Agents Remain Untapped Potential

Replit CEO Amjad Masad指出增长agent仍有巨大未被挖掘的潜力,并在OpenAI-Cursor消息后强调Replit作为独立多模型替代方案的优势。

Replit CEO Amjad Masad noted that growth agents still represent largely untapped potential, and positioned Replit as an independent multi-model alternative after the OpenAI-Cursor announcement.
Replit CEO Amjad Masad noted that growth agents still represent largely untapped potential, and positioned Replit as an independent multi-model alternative after the OpenAI-Cursor announcement.
查看原文 →