七月二日

Peter Yang, creator of practical AI tutorials, ran his Fable 5 "vibe check" and called it a genuine step function above any other model: "It's still really f***ing good... This is a step function above any other model. Hope GPT 5.6 can match." He also shipped a new video tutorial walking through five use cases worth trying Fable 5 on — finding Fable-worthy work, getting life and business advice, making projects ship-ready, planning your next big thing, and refactoring a codebase — with real output shown for each.

Peter Yang，实用 AI 教程的作者，做了一次 Fable 5 的"体验测评"，认为它相比其他任何模型都是一次真正的阶跃式提升："它还是真的太强了……这是一个 step function 级别的飞跃，希望 GPT 5.6 能追上。"他还发布了一个新视频教程，演示了值得用 Fable 5 尝试的五个场景——找到适合 Fable 的高价值工作、获取生活和商业建议、把项目打磨到可交付状态、规划下一件大事，以及重构代码库——每个场景都展示了真实的输出。

Google Labs announced it is retiring MusicFX and MusicFX DJ on July 31, 2026, folding everything it learned from those early real-time AI music experiments into Google Flow Music, its longer-term tool for creating, sharing, and remixing original music.

Google Labs 宣布将于 2026 年 7 月 31 日下线 MusicFX 和 MusicFX DJ，把从这些早期实时 AI 音乐实验中学到的一切都并入 Google Flow Music——它面向长期的原创音乐创作、分享与二次混音工具。

Replit CEO Amjad Masad says that now that building is easy, Replit's focus has shifted to getting entrepreneurs to market — helping them reach their first customer and first dollar. To that end, you can now sell your Replit apps on Whop, which he calls one of the best places on the internet to monetize your creations.

Replit CEO Amjad Masad 表示，如今"构建"已经变得很容易，Replit 的重心转向了帮助创业者走向市场——触达第一个客户、赚到第一块钱。为此，你现在可以在 Whop 上出售你的 Replit 应用，他称 Whop 是互联网上把作品变现的最佳去处之一。

Vercel CEO Guillermo Rauch showed WordPress running on Vercel Fluid with Active CPU from a single Dockerfile.vercel, backed by MySQL on PlanetScale, with 30-second deploys (including the cloud docker build) from one command. He also announced a dry-run step for agentic deployments — mirroring how agents already run node --check, tsc --noEmit, or next build to verify their work before pushing — to minimize the cost and risk of agent-driven deploys.

Vercel CEO Guillermo Rauch 演示了用一个 Dockerfile.vercel 让 WordPress 跑在 Vercel Fluid 的 Active CPU 上，数据库用 PlanetScale 上的 MySQL，一条命令即可在 30 秒内完成部署（包括在云端跑 docker build）。他还发布了面向 agent 部署的 dry-run（预演）步骤——就像 agent 在推送前已经会跑 node --check、tsc --noEmit 或 next build 来自查一样——用来把 agent 驱动的部署的成本和风险降到最低。

Box CEO Aaron Levie argues Devin's new "agentic mapreduce" is a concrete answer to why the world will need 100X more AI inference: swarms of agents that fan out over huge amounts of data (here, an entire code repo) to do work humans never could. As he quotes it, Devin "maps relevant signals across the repo, fans out focused agents over bounded shards, reduces their findings into one report, then verifies serious vulnerabilities in isolated sandboxes before marking them confirmed." Levie sees the same pattern coming to pharma, banking, and any industry drowning in unstructured data — and notes these token-hungry workloads only pencil out when you can mix frontier and lower-cost models, which he calls a major value proposition for the applied-AI layer.

Box CEO Aaron Levie 认为，Devin 新推出的"agentic mapreduce"（智能体版 MapReduce）正是"世界为何需要 100 倍 AI 推理量"的一个具体答案：一大群 agent 分散铺开去处理海量数据（这里是整个代码仓库），完成人类根本做不到的工作。用他引用的话说，Devin"在整个仓库中映射出相关信号，把聚焦的 agent 分派到各个有界分片上，再把它们的发现归约成一份报告，最后在隔离的沙盒中验证严重漏洞、确认无误后才标记为已确认"。Levie 认为同样的模式会进入医药、银行以及任何被非结构化数据淹没的行业——他还指出，这类极度消耗 token 的工作负载，只有在能够混合使用前沿模型和低成本模型时才算得过账，这正是应用层 AI 的一大价值主张。

FirstMark VC Matt Turck dug into Lime's surprising IPO: how a scooter company carrying $1B in debt that had literally flagged "substantial doubt" about surviving the year pulled off a successful public offering. His read: impressive financial engineering paid off the toxic loans and converted the rest to equity, Uber's 22% stake and direct rider referrals provide a backstop, and Lime has quietly been free-cash-flow positive for three straight years with revenue up nearly 30% YoY — the last survivor of the micromobility bloodbath.

FirstMark 的 VC Matt Turck 深挖了 Lime 出人意料的 IPO：一家背着 10 亿美元债务、甚至公开表示"对能否撑过今年存在重大疑问"的滑板车公司，是怎么成功上市的。他的解读是：漂亮的财务工程还清了有毒贷款、把其余债务转为股权；Uber 持有 22% 股份并直接为其导流乘客，是强力的靠山和合作方；而且 Lime 已经悄然连续三年实现自由现金流为正、营收同比增长近 30%——它是这场微出行大洗牌中最后的幸存者。

Builder Zara Zhang shared a counterintuitive rule for agent skills: "You don't START with a skill. You END with a skill. Building the skill is the last step of your workflow, not the first." In other words, get the workflow working first and only codify it into a reusable skill at the end. She also posted a quick tip that you can swap Codex's model over to GLM.

Builder Zara Zhang 分享了一条关于 agent skill（技能）的反直觉规则："你不是从一个 skill 开始，而是以一个 skill 收尾。构建 skill 是你工作流的最后一步，而不是第一步。"换句话说，先把工作流真正跑通，最后才把它固化成一个可复用的 skill。她还发了一个小贴士：你可以把 Codex 的模型切换成 GLM。

FPV Ventures partner Nikunj Kothari says the talent vortex pulling people into OpenAI and Anthropic is only intensifying — in the past two months, four of his personal friends left very established roles to join the labs, drawn by the dual pull of building one of the most consequential companies and getting pre-IPO liquidity. His contrarian aside: most VCs won't actually make life-changing money, and some have deflected to the labs knowing exactly that.

FPV Ventures 合伙人 Nikunj Kothari 表示，把人才吸进 OpenAI 和 Anthropic 的"漩涡"只会越来越强——过去两个月里，他有四位私交好友离开了非常稳固的岗位加入这些实验室，动力来自双重吸引：既能参与打造最具影响力的公司之一，又能拿到 IPO 前的流动性。他有一个逆向观点：大多数 VC 其实赚不到能改变人生的钱，而有些人正是看清了这一点才转投实验室。

Anthropic's Claude account confirmed Fable 5 is back: all paid plans with usage included can access it through July 7, using it for up to 50% of your weekly usage limit before switching to another model (or continuing via usage credits), with a blog post detailing the release. The team also reminded users that if a request is mistakenly flagged in Claude Code, running /feedback (or using the thumbs buttons on the web and in Cowork) helps tune the classifiers and reduce false positives over time.

Anthropic 的 Claude 账号确认 Fable 5 回归了：所有含用量额度的付费套餐都可以在 7 月 7 日前使用它，最多可占用你每周用量上限的 50%，之后需切换到其他模型（也可以继续通过用量额度使用），并配有一篇博客详细介绍此次发布。团队还提醒用户，如果在 Claude Code 中某个请求被误判拦截，运行 /feedback（或在网页端和 Cowork 里用点赞/点踩按钮）可以帮助调优分类器，并随时间减少误报。

The Takeaway: In the age of AI you can build anything, so the harder and more valuable skill is knowing what you should actually build and maintain versus what you should just buy.

Natalia, head of consulting at Every, manages an AI agent "employee" named Claudie that now handles real operational work every day — running dashboards, managing the CRM, keeping its own LinkedIn and Twitter, and looping to self-evaluate against feedback. Yet she just hired a human ops person and bought a real CRM (Attio) plus Asana, retiring the CRM she had vibe-coded on top of Google Sheets. Her lesson: AI is exceptional at executing a standard operating procedure, but real software is a compilation of thousands of logical rules a model won't one-shot — and someone still has to maintain data quality and supply taste and direction.

The other throughline is Codex, which she calls the single greatest tool she's adopted. As a non-engineer she loves that the terminal and browser live inside the chat, letting her offload file-system and architecture decisions. Her favorite "loop": she gave Codex an overnight goal to populate her CRM from months of emails and call notes, went to sleep, and woke up to what would have been weeks of work already done. She's also built a 13-hour Codex app to coordinate her 81-year-old father's care across nurses and family, and makes illustrated "zines" to learn topics like anatomy and Aristotle on the go.

Her framing for the shift: "Knowledge work now is turning into something like gardening, where when you're gardening, you're creating the conditions for the growth to happen, but you're not like making the plant with your hands." Her advice to executives: start from the systems and OKRs you already have, hand AI that same shared context, and define one task really well before chasing an "AI-first" rebuild.

要点：在 AI 时代你几乎什么都能做出来，所以更难、也更有价值的能力，是判断哪些东西你真正应该自己构建并长期维护，哪些应该直接买现成的。

Natalia 是 Every 的咨询业务负责人，她管理着一个名叫 Claudie 的 AI agent"员工"，如今它每天都在处理真实的运营工作——运行各种看板、管理 CRM、维护自己的 LinkedIn 和 Twitter，还会循环运行、对照反馈做自我评估。但她还是招了一位真人运营，并买了真正的 CRM（Attio）加上 Asana，把之前在 Google Sheets 上"氛围编程"拼出来的 CRM 退役了。她的教训是：AI 极其擅长执行标准作业流程（SOP），但真正的软件是成千上万条逻辑规则的编译产物，模型无法一次性搞定——而且总得有人来维护数据质量、提供品味和方向。

另一条主线是 Codex，她称之为自己用过的最强工具。作为一个非工程师，她很喜欢终端和浏览器都内置在聊天界面里，让她可以把文件系统和架构层面的决策交出去。她最爱的一个"loop"：她给 Codex 设定了一个通宵目标——根据几个月的邮件和通话记录把 CRM 填好，然后去睡觉，醒来时发现原本要花几周的工作已经完成了。她还用 Codex 花了 13 小时做了一个应用，用来在护士和家人之间协调她 81 岁父亲的照护，并且会做插画版的"知识小册子（zine）"，随时随地学习解剖学、亚里士多德等主题。

她这样描述这种转变："如今的知识工作正在变得像园艺——你在做园艺时，是在创造让其生长的条件，而不是用双手把植物本身造出来。"她给高管们的建议是：从你已有的系统和 OKR 出发，把同样的共享上下文交给 AI，先把一件事真正定义清楚、做到位，再去追求所谓"AI 优先"的整体重构。