🌐 双语
Archive

AI Builders
Digest

2026-11-10 15 builders · 29 tweets · 1 podcasts · 0 blogs

🔥 热点话题

Handshake AI:前沿实验室真正采购的数据是什么样的What Frontier Labs Actually Buy: Handshake AI Data Deep Dive

The Takeaway:训练前沿模型的数据市场三年从10亿美元涨到100亿美元,未来三年有望到1000亿美元,核心驱动是专家亲手打造的复杂RL环境与评分标准,而不是简单标注。

Handshake AI联合创始人Garrett Lord与团队(Nishant、Sahil、Nick Magazine、Jonas等)深入拆解了当前数据工作的真实形态。数据已从大学生画框、做情感分类,彻底变成投资银行家、医生、律师全职「创作」环境:模拟Outlook、Bloomberg、Excel、实验室仪器,填充真实一致的邮件、数据室、财务模型,再写数百行评分标准,定义「什么是好」。知识工作大多不可完美验证,95%正确等于100%错误,因为修错比从零做更费时间。

生命科学基准Vials上,几乎所有前沿模型(包括最强者)准确率都低于50%,有的甚至不到20%。模型连基本的凝胶电泳条带、质粒图谱百分比、组织病理切片坏死程度都读不准。金融基准Atlas上同样惨淡(sub 20%),一个LBO模型里2600多个公式,模型在338(h)(10)税收处理的关键一行上出错,整份交付物就彻底不可信。安全基准Care Bench则覆盖儿童情绪依赖、网络欺凌、早期诱导等12类风险,家长与临床专家共同标注,揭示模型在「拒绝后仍继续提供替代方案」上的危险盲点。

Garrett总结:「三年前这是一个十亿美元的市场,现在是一百亿。三年后我看不出它怎么可能不是一千亿。」数据正在从「标注」变成「专家全职创作+审计+后训练闭环」,而企业也开始自己建私有基准,把「什么是好、什么是坏」写进可量化的数字。
The Takeaway: The market for high-quality frontier training data has grown from $1B to $10B in three years and is on track for $100B, driven by expert-authored complex RL environments and rubrics rather than simple labeling.

Handshake AI co-founder Garrett Lord and team (Nishant, Sahil, Nick Magazine, Jonas and others) laid out the real work. Data has moved far beyond college students drawing boxes. Experts now author full simulated worlds—Outlook, Bloomberg, Excel, lab instruments—populate them with consistent emails, data rooms and financial models, then write hundreds of rubric rows defining “what good looks like.” Most knowledge work is only semi-verifiable. As one banker put it, “If it’s 95% right, it’s 100% wrong,” because fixing a flawed output takes as long as starting from scratch.

On the life-sciences benchmark Vials, nearly every frontier model scores under 50% (some under 20%). Models still fail at reading gel bands, plasmid map percentages and histopathology necrosis grades. Atlas (finance) is equally harsh: a single wrong tax-treatment line in a 2,600-formula LBO model renders the entire deliverable unusable. Care Bench covers 12 child-safety risk categories; parent + clinician annotations exposed models that refuse a harmful request yet immediately offer “safe alternatives” that keep a malicious actor engaged.

Garrett’s closer: “Three years ago, this was a billion-dollar market. It’s now a $10 billion market. Three years from now, I don’t see how it’s not a hundred billion dollar market.” Data is becoming full-time expert authorship plus audit plus closed-loop post-training—and enterprises are starting to build their own private benchmarks that codify both “good” and “bad.”
查看原文 →

Claude Fable 5.2 发布:创意与长时任务的新基准Claude Fable 5.2 Launch: New Bar for Creativity and Long-Horizon Work

Anthropic正式推出Claude Fable 5.2,全付费计划可用。官方账号同步公布了两个物理学预印本结果:极小宇宙中的突变相变,以及模形式自举技术精确极限。

Every CEO Dan Shipper的Day-zero Vibe Check最能说明体感:团队用它三天内做出带agent的视频剪辑工具Everycut;把孩子玩具照片变成可玩赛车游戏;把研究采访直接剪成纪录片。有人直接停工等模型重新上线才肯继续项目。代价是慢、吃token,日常写作与短对话仍建议用更轻的模型。

Anthropic研究员Alex Albert把达芬奇500年前的笔记本PDF丢给Fable 5.2,模型直接把发明做成可交互3D。他推荐的万能提示词:「读完整项目,找出所有坏掉、不一致、低效或不再需要的东西,能修的先修,其余列给我审。」

Claude产品负责人Boris Cherny宣布:订阅越久,Fable重置次数越多(Pro一次,Max/Ultra/Team Premium最多三次)。Swyx则调侃有人用Fable 5.2当demo直接证明了费马大定理,并顺带预告AIE大会有更大彩蛋。
Anthropic launched Claude Fable 5.2 on all paid plans. The official account also dropped two physics preprints: an abrupt transition in very small universes and the exact limit of a 15-year-old modular-bootstrapping technique used to rule out quantum-gravity candidates.

Every CEO Dan Shipper’s day-zero Vibe Check captured the feeling best: the team shipped an agentic video editor (Everycut) in three days, turned a toddler’s toy photo into a playable racing game, and turned research interviews into a finished documentary. One teammate simply stopped working on certain projects until the model returned. Trade-off: it is slow and token-hungry; for everyday writing or quick coding loops people still reach for lighter models.

Anthropic researcher Alex Albert fed Fable 5.2 a PDF of a 500-year-old Leonardo notebook and watched it turn the inventions into interactive objects. His go-to prompt for every new model: “Read this entire project and find everything that’s broken, inconsistent, inefficient, or no longer needed. Fix what you’re sure about and list the rest for me to review.”

Claude product lead Boris Cherny added loyalty resets: longer subscribers get more Fable resets (one for Pro, up to three for Max/Ultra/Team Premium). Swyx noted someone used the model as a live demo to prove Fermat’s Last Theorem and teased a bigger AIE talk reveal.
查看原文 →查看原文 →查看原文 →查看原文 →查看原文 →

Box CEO Aaron Levie:知识工作的「Claude Code时刻」正在到来Box CEO Aaron Levie: The Claude Code Moment for Knowledge Work Is Coming

Box CEO Aaron Levie指出,编码与计算机使用能力的跃升,加上行业特定理解与通用推理,正在把自动化带进动画、影视、游戏、制造、建筑、产品设计、机器人等从未有过「Claude Code时刻」的领域。这些类别没有天然的需求上限,「软件吞噬世界」终于以更大的规模通过AI真正发生。
Box CEO Aaron Levie argues that leaps in coding and computer-use, combined with domain understanding and general reasoning, are about to bring autonomous collaborators into animation, film, gaming, manufacturing, architecture, product design and robotics—fields that have not yet had their Claude Code or Codex moment. None of these categories has finite demand. “Software is eating the world” is finally happening at a much larger scale because of AI.
查看原文 →

🛠️ 开发者工具与技巧

Vercel CEO Guillermo Rauch:MCP Events 与云沙箱是抓住新模型算力的关键Vercel CEO Guillermo Rauch: MCP Events and Cloud Sandboxes Are the Missing Pieces

Guillermo Rauch对MCP依然极度看好:它的价值已超出「Model」本身,有望成为文明级的系统互通协议。新增的MCP Events解决了跨会话回调问题——agent部署完服务后可以订阅deployment.succeeded,而不用轮询。Vercel自己的MCP已支持。他同时强调:新一代模型能并行处理远超单机的工作量,瓶颈只剩野心和并行算力。最简单的用法是装Vercel CLI,启用Plugin,让agent调用/offload-to-sandbox skill即可自动开销云沙箱。
Guillermo Rauch remains extremely bullish on MCP: its utility already exceeds the “Model” part and could become a civilization-scale open protocol for system communication. MCP Events finally adds cross-session callbacks—an agent can subscribe to deployment.succeeded instead of polling. Vercel’s own MCP already supports it. He also notes the new generation of models can orchestrate far more parallel work than any single machine; the only remaining bottlenecks are ambition and parallel compute. The simplest path: install the Vercel CLI, enable the Plugin, and tell the agent to use the /offload-to-sandbox skill.
查看原文 →查看原文 →

OpenAI Codex负责人Thibault Sottiaux:Dots现已可控制多台机器OpenAI Codex Lead Thibault Sottiaux: Dots Can Now Control Many Machines

Thibault Sottiaux宣布Dots除了自带的小云电脑,现在还能直接控制用户的多台机器,进一步扩展了agent的执行边界。
Thibault Sottiaux announced that Dots can now control many of the user’s own machines in addition to the little cloud computer provided by default, expanding the agent’s execution surface.
查看原文 →

OpenAI产品Nan Yu:ChatGPT全局听写值得打开OpenAI Product Nan Yu: Turn On Global Dictation in ChatGPT

Nan Yu提醒:在ChatGPT设置里打开全局听写,就能获得与App内听写同样快速准确的转写路径,把订阅里尚未被充分使用的价值释放出来。
Nan Yu’s PSA: go into ChatGPT settings and enable global dictation. It uses the same fast, accurate transcription path as the in-app dictation and unlocks unused value in the subscription.
查看原文 →

Replit CEO Amjad Masad:自愈软件与企业级Vibe CodingReplit CEO Amjad Masad: Self-Healing Software and Enterprise Vibe Coding

Amjad Masad回应「谁来维护vibe coded应用」的质疑:自愈软件正在出现。他同时为日本SMFL数字创新团队的新书《Enterprise Vibe Coding》站台(11月27日出版),并分享了2024年7月向Paul Graham展示早期vibe coding时的反应截图。
Amjad Masad answered the perennial “who maintains vibe-coded apps?” question with self-healing software. He also promoted the upcoming Japanese book Enterprise Vibe Coding (Nov 27) from SMFL’s digital innovation team and shared the screenshot of Paul Graham’s reaction when Amjad first showed him vibe coding in July 2024.
查看原文 →查看原文 →查看原文 →

Meta AI Madhu Guru:产品经理应用AI把每个环节做深Meta AI Madhu Guru: PMs Should Use AI to Go Deeper on Every Stage

Madhu Guru建议产品负责人先画出自己的产品生命周期(市场理解→设计→开发→GTM),然后问:用AI后,PM能否在每个阶段都走得更深——更广地理解用户与市场、快速从想法到可运行原型、自己探索代码库、自己跑分析、真正理解GTM漏斗。这不是要PM取代所有角色,而是把广度与深度结合起来,提高整个职能的门槛。
Madhu Guru’s exercise for product leaders: draw your product lifecycle (market understanding → design → development → GTM). Then ask how much deeper PMs can go on every stage with AI—deeper user and market understanding, idea-to-working-prototype loops, codebase exploration, self-serve analytics, real GTM funnel insight. The point is not to replace every role, but to raise the bar so PMs combine breadth with far greater depth.
查看原文 →

🌍 其他动态

YC总裁Garry Tan:新前沿模型发布日 = 所有项目阶跃YC President Garry Tan: New Frontier Model Day = Step-Function for Every Project

Garry Tan表示自己可能永远不会厌倦前沿模型发布日——每个新模型都让手头所有项目立刻获得一次阶跃提升。
Garry Tan: “Red letter day: New frontier model = every one of your projects gets a step function improvement. I’m not sure if I’ll ever get sick of frontier model release days.”
查看原文 →

FPV Ventures Nikunj Kothari:人们其实看得出来你在用AI复读FPV Ventures Nikunj Kothari: People Notice When You Are Just Feeding AI

Nikunj Kothari引用Niklas Gruhn的文章提醒:如果你把一切都喂给AI再原样吐出,对方其实看得出来——因为他们很可能也在用同一批agent形成自己的判断。
Nikunj Kothari, citing Niklas Gruhn’s piece: if you feed everything to AI and simply regurgitate it, people notice—because they are probably already forming opinions with the same agents.
查看原文 →

Peter Yang:把孩子的画变成卡丁车游戏Peter Yang: Turning an 8-Year-Old’s Drawings into Kart Racers

Peter Yang展示了把8岁孩子的手绘直接变成可玩卡丁车角色的小实验,并顺手升级了背景音乐,感叹「这就是未来的工作」。
Peter Yang turned his 8-year-old’s drawings into playable kart-racer characters, upgraded the music on the fly, and captioned it “This is the future of work.”
查看原文 →查看原文 →

Aditya Agarwal:Nuance AI 大更新Aditya Agarwal: Major Nuance AI Launch

SPC GP Aditya Agarwal转发并推荐Nuance AI团队的最新重磅更新,邀请大家立刻试用。
SPC GP Aditya Agarwal highlighted a major launch from the Nuance AI team and encouraged everyone to try it immediately.
查看原文 →

Alex Albert:机器人也会被Bitter Lesson击中Alex Albert: Robotics Will Fall to the Bitter Lesson Too

Alex Albert在引用相关讨论时断言:机器人领域最终也会被Bitter Lesson收割,通用方法与规模会再次胜过手工设计的专用系统。
Alex Albert, quoting a related thread: “btw robotics will fall to the bitter lesson too.”
查看原文 →