AI Builders Digest — 2026年6月11日
📌 X / TWITTER
Andrej Karpathy — AI 研究者，前 Tesla AI 总监
Claude Fable 5 是一次超重量级发布。它与 Mythos 共享同一底层模型，但增加了安全防护。基准测试全面领先，是 SOTA。但 Karpathy 强调"定性地"看，这也是一个值得大版本号跳跃的质变——堪比从 GPT-3.5 到 GPT-4 的跨越。他说这不仅是一个更好的模型，而是一种全新体验，就像第一次使用 Claude Code 时的那种震撼感。 → https://x.com/karpathy/status/2064409694761054332
Andrej Karpathy — AI researcher, former Tesla AI Director — calls Claude Fable 5 a super exciting release. It’s the same underlying model as Mythos but with added safeguards. Benchmarks are SOTA across the board, but qualitatively, Karpathy emphasizes this is a major-version-bump-deserving step change forward — on the order of going from GPT-3.5 to GPT-4. It’s not just a better model; it’s a fundamentally new experience, like the first time using Claude Code. → https://x.com/karpathy/status/2064409694761054332
Swyx — AI 工程师，Latent Space 播客主理人
Fable 发布后，Swyx 分享了几个关键观察：NVIDIA 基础设施让 Anthropic 从签约到发布 Mythos 级模型仅用了 34 天。他还指出目前用 Claude Code + Fable 做代码审查几乎是免费的"alpha"——让 Fable 审查你的代码，准备好对你上线的每一行代码感到后怕。 → https://x.com/swyx/status/2064492823781789969 → https://x.com/swyx/status/2064421542503797186 → https://x.com/swyx/status/2064415059003248694
Swyx — AI engineer, Latent Space podcast host — notes that Anthropic went from deal signing to shipping a Mythos-class model in just 34 days on the NVIDIA stack. He also flags a massive alpha opportunity: using Claude Code with Fable for code review while it’s not pay-per-use — “be prepared to be in abject horror that you shipped anything to prod without a Fable Check first.” → https://x.com/swyx/status/2064492823781789969 → https://x.com/swyx/status/2064421542503797186 → https://x.com/swyx/status/2064415059003248694
Boris Cherny — Anthropic Claude Code 团队成员
作为构建 Claude Code 的人，Cherny 说 Fable 5 是自 Opus 4.5 以来最大的能力跃升。Opus 4.5 出现后他卸载了 IDE 因为所有编码都在终端完成。Fable 更进一步：从"帮你写代码的 agent"变成了"像同事一样跟你一起思考的 agent"。他强调了自验证循环（self-verification loops）的重要性——让模型长时间运行并自我校正，是实现高质量结果的关键。 → https://x.com/bcherny/status/2064469227499929651 → https://x.com/bcherny/status/2064431111154053187 → https://x.com/bcherny/status/2064426115255730578
Boris Cherny — Anthropic Claude Code team — says Fable 5 is the biggest step up since Opus 4.5. After 4.5 he uninstalled his IDE because all coding moved to the terminal. With Fable, Claude has stepped up from being a coding agent to a thinking partner that reasons alongside you like a colleague. He emphasizes self-verification loops as the key ingredient enabling models to run longer and deliver results closer to your intent. → https://x.com/bcherny/status/2064469227499929651 → https://x.com/bcherny/status/2064431111154053187 → https://x.com/bcherny/status/2064426115255730578
Alex Albert — Anthropic 研究员
经历过每一次 Anthropic 模型发布的内部人士。Albert 列出真正里程碑式的发布只有四次：Claude Opus 3、Sonnet 3.5、Opus 4.5，以及现在的 Fable 5。他说 Fable 让他不再把模型当成"工具"，而是一个可以放心委派模糊任务的协作者。他给出四条建议：给模型更大胆的任务、默认用高努力模式、利用自验证循环、让模型自己纠偏。 → https://x.com/alexalbert__/status/2064467657483829441 → https://x.com/alexalbert__/status/2064394410004304003
Alex Albert — Anthropic researcher who’s been through every model launch — names only four transformative moments: Claude Opus 3, Sonnet 3.5, Opus 4.5, and now Fable 5. With Fable, the model stopped feeling like a tool he directs and became a collaborator he can delegate ambiguous tasks to. His tips: give it bigger tasks, default to high effort mode, use self-verification loops, and let it self-correct. → https://x.com/alexalbert__/status/2064467657483829441 → https://x.com/alexalbert__/status/2064394410004304003
Aaron Levie — Box CEO
Levie 连发三条关于 AI 应用层建设的深度思考。核心理念：如果 AI 进步放缓的担忧存在，那么 Fable 5 就是最直接的回应——全维度能力大跳跃，将在几乎所有知识工作领域带来 agent 能力的重大提升。他特别推荐了一篇关键文章，阐述"不可训练之角"理论：AI 应用的价值不在于训练更好的模型，而在于做那些"不光彩"的工作——整理公司的私有现实让模型可以行动、给模型提供执行工具、与客户协作确保输出符合情境。 → https://x.com/levie/status/2064569513023328268 → https://x.com/levie/status/2064396746953023647 → https://x.com/levie/status/2064379199629181139
Aaron Levie — Box CEO — pushes back on AI slowdown narratives with Fable 5 as Exhibit A. The capability jump will deliver major agent improvements across nearly all knowledge work categories. He highlights a must-read post on the “untrainable corner”: AI application value is won not by training better models but by doing the unglamorous work — arranging a company’s private reality so a model can act on it, handing it tools, and working with customers so outputs fit context. → https://x.com/levie/status/2064569513023328268 → https://x.com/levie/status/2064396746953023647 → https://x.com/levie/status/2064379199629181139
Garry Tan — Y Combinator CEO
Tan 称 Fable 5 是"我见过的最大模型能量"。但他也展示了一个真实吐槽：Fable 5 在帮他修复 GStack 时也遇到了令人沮丧的卡壳时刻——提醒人们即使最强模型也仍需人类引导。另外，他分享了对旧金山政治的犀利评论。 → https://x.com/garrytan/status/2064573857911152710 → https://x.com/garrytan/status/2064559225859416186
Garry Tan — Y Combinator CEO — calls Fable 5 “the biggest model energy I’ve ever seen,” but also shares a real moment of frustration when Fable hit a wall while fixing GStack — a reminder that even the strongest models still need human steering. → https://x.com/garrytan/status/2064573857911152710 → https://x.com/garrytan/status/2064559225859416186
Guillermo Rauch — Vercel CEO
Vercel CLI 新增 AI Gateway 功能：可以创建 API key、用 –budget 设置消费上限、用 –refresh-period 控制配额周期。Rauch 将其比喻为"AI token 的虚拟信用卡"——让团队安全地管理 AI API 使用，防止账单爆炸。 → https://x.com/rauchg/status/2064551967461114111
Guillermo Rauch — Vercel CEO — announces Vercel CLI now supports AI Gateway API keys with budget caps and refresh periods: “Think of it as virtual credit cards for AI tokens.” → https://x.com/rauchg/status/2064551967461114111
Zara Zhang — 独立建造者
非技术背景的建造者视角：编码 agent 的障碍从来不是界面——聊天是最简单的 UI。真正的障碍是非技术人员不知道自己可以要什么。空白聊天框假设你已经知道可能性，而大多数人不知道。Zhang 分享了她如何在完全没有编码背景的情况下积累 30K GitHub 星标，并将在本周五举办线上分享，讲述她的"vibe coding"方法论。 → https://x.com/zarazhangrui/status/2064587398529606082 → https://x.com/zarazhangrui/status/2064486120386379950
Zara Zhang — independent builder with 30K GitHub stars despite no coding background — argues the barrier for non-technical people using coding agents was never the interface (chatting is the easiest UI ever) but knowing what to ask for. The blank chat box assumes you already know what’s possible. She’ll share her vibe coding methodology in a virtual talk this Friday. → https://x.com/zarazhangrui/status/2064587398529606082 → https://x.com/zarazhangrui/status/2064486120386379950
Nikunj Kothari — FPV Ventures 合伙人
Kothari 展示了 Fable 的一次性建站能力：听完 Invest Like the Best 播客谈论 S 曲线后，他将 transcript 丢给 Fable，一个 prompt 就生成了 escurves.com——一个梳理过去 200 年所有 S 曲线、拐点和泡沫历史的网站。工作流：Claude app 研究模式 → 梳理结构 → Claude Code prompt → 一次构建。 → https://x.com/nikunj/status/2064508462034501997 → https://x.com/nikunj/status/2064506504888373758
Nikunj Kothari — FPV Ventures partner — demonstrates Fable’s one-shot website building: after listening to a podcast on S-curves, he dumped the transcript to Fable and got escurves.com — a site cataloging every S-curve, inflection point, and bubble commentary from the last 200 years — all in one prompt. → https://x.com/nikunj/status/2064508462034501997 → https://x.com/nikunj/status/2064506504888373758
Dan Shipper — Every CEO
Shipper 和他团队的 Fable 5 评测（VIBE CHECK）出来了。作为每周都在深度使用前沿模型的团队，Every 对 Fable 的评价极具参考价值。他们还发布了一期 YouTube 视频讨论。同时，Amjad Masad（Replit CEO）宣布在 Replit 上使用 Claude Mythos 可享 75 折。 → https://x.com/danshipper/status/2064395458777108707 → https://x.com/danshipper/status/2064395167658860859
Dan Shipper — Every CEO — publishes Every’s “Vibe Check” review of Fable 5. As a team that uses frontier models intensely every week, their take carries weight. → https://x.com/danshipper/status/2064395458777108707 → https://x.com/danshipper/status/2064395167658860859
Josh Woodward — Google Labs / Gemini VP 软件的需求即将爆表。AI 降低了制造软件的成本，却没有降低对软件的需求——恰恰相反。 → https://x.com/joshwoodward/status/2064509357216428171
Josh Woodward — Google Labs / Gemini VP — “The demand for software is going to be off the charts.” AI lowers the cost of making software but doesn’t lower demand — it multiplies it. → https://x.com/joshwoodward/status/2064509357216428171
Peter Yang — AI 教育者 分享他用 Claude Fable 构建 F-Zero 风格赛车的 prompt。Fable 的长上下文和浏览器能力让他感到惊艳，但也指出 browser use 功能在某些场景下反而拖慢了速度。 → https://x.com/petergyang/status/2064550073594446059 → https://x.com/petergyang/status/2064577126385459265
Peter Yang — AI educator — shares his prompt for building an F-Zero style racing game with Fable, while noting browser use can sometimes slow things down. → https://x.com/petergyang/status/2064550073594446059
其他建造者动态：
- Thibault Sottiaux (OpenAI Codex/ChatGPT)：开玩笑说要收 1% 版税，并做了 Codex /goal 使用习惯调研
- Thariq (Anthropic Claude Code)：在东京 Code with Claude 活动现场
- Aditya Agarwal (South Park Commons)：从投资人视角看待 AI 里程碑
- Claude AI (官方)：宣布 Fable 5 全平台可用，Mythos 5 仅限 Glasswing 合作伙伴
🎙️ 播客
AI & I by Every — We Automated Everything With AI and Tripled Our Headcount → https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL
核心观点： Every 的 COO Brandon 采访 CEO Dan Shipper，讨论他的长文《After Automation》。核心悖论：Every 是你能找到的最"AI 原生"的公司——Slack 里 agent 和人类数量相当，人人都在用 Claude Code 和 Codex。然而自动化越深入，人类工作量不仅没减少，反而暴增。团队从 4 人扩张到 30 人，还在继续招人。
Dan 的解释框架：AI 使"昨日专家的能力"变得廉价且普及。任何人都能用 prompt 写代码、写文章、做设计，大量"接近正确但不够对"的输出充斥着工作流。这产生了两种新需求：① 需要更多真正的专家来将这些半成品打磨到真正可用的状态；② 专家自身利用 AI 的更高基线，去创造之前完全不可能做出的东西。
关键隐喻——阿喀琉斯与乌龟：你让 AI 做一件事，它让你惊叹，让你觉得自己的工作要没了。然后它停下来看着你问：“下一步做什么？“Dan 的观点是：再强的模型也会回头问你"什么重要”——因为什么重要在不断变化，且这种变化本身就在被 AI 加速。只要模型回到你这里来问方向，人类就始终不可替代。
Dan 的最终建议：“如果你跟着模型走——每次新模型出来就去学着用它——你不会有事的。而且你可能会发现，你可以做比以往更多、更好、更有成就感的工作。”
Podcast: AI & I by Every hosts Brandon (COO) interviewing Dan Shipper (CEO) on his essay “After Automation.” The paradox: Every is as AI-native as it gets, yet more automation has meant more human work — the team grew from 4 to 30. Dan’s framework: AI makes yesterday’s expert competence cheap and ubiquitous, flooding workflows with outputs that are close-but-not-quite-right. This creates demand for experts to (a) shepherd near-misses into production-quality work, and (b) use the raised baseline to build things previously impossible. The Achilles metaphor: AI amazes you, then stops and asks “what’s next?” Dan’s core thesis — ride the models, and you’ll be fine. → https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL
通过 Follow Builders 生成: https://github.com/zarazhangrui/follow-builders