七月三日

Box CEO Aaron Levie argues that deploying AI in the enterprise beyond simple chatbots will take real work to align AI systems with the underlying business processes they touch. Most enterprise workflows weren't designed for agents to "just drop into" — they run on fragmented data, legacy systems agents can't connect to, and institutional rather than documented knowledge. To run agents reliably at scale, companies have to clean up data, modernize IT, build evals, and redesign where humans stay in the loop. This, he says, is why applied AI companies are expanding forward-deployed engineer (FDE) efforts and launching "deploycos," making the FDE "one of the most critical jobs in tech going forward."

Box CEO Aaron Levie 认为，要让 AI 在企业里真正落地（而不只是当个聊天机器人），必须下大功夫把 AI 系统与它所涉及的底层业务流程对齐。企业里的大多数 workflow 从一开始就不是为 agent「即插即用」设计的，它们跑在割裂的数据、agent 连不上的老旧系统，以及只存在于人脑、没有文档化的机构知识之上。想让 agent 大规模可靠运行，公司得先把数据清理干净、把 IT 系统现代化、建立 evals，并重新设计人类在哪个环节继续参与。他说，这正是为什么众多应用型 AI 公司都在扩张 forward-deployed engineer（FDE）团队、成立「deployco」，也让 FDE 成为「未来科技行业最关键的岗位之一」。

Vercel CEO Guillermo Rauch introduced AI Gateway Rules, pitching the AI Gateway as "a CDN for AI models" that can dynamically re-route or deny model traffic without a redeployment. The motivation was concrete: when Fable was suddenly retired, teams worried about production workloads depending on it, and models get retired often as GPU capacity stays contested — so you can now rewrite a model "route" on the fly, e.g. remap anthropic/claude-fable-5 to anthropic/claude-opus-5. Separately, he highlighted new private connectivity between Vercel services, registered as bindings in vercel.json in a single line of code and readable across Node, Python, and Docker.

Vercel CEO Guillermo Rauch 推出了 AI Gateway Rules，把 AI Gateway 定位成「AI 模型的 CDN」，可以在不重新部署的前提下动态改路由或拦截模型流量。动机很具体：当 Fable 被突然下线时，很多团队担心依赖它的生产负载会受影响，而由于 GPU 产能一直紧张，模型被下线其实相当频繁。所以现在你可以即时重写模型「路由」，比如把 anthropic/claude-fable-5 映射到 anthropic/claude-opus-5。此外他还介绍了 Vercel 服务之间新的私有互联能力：只需在 vercel.json 里用一行代码注册 binding，就能在 Node、Python、Docker 里直接读取。

Anthropic's Cat Wu (Claude Code + Cowork) says Claude Tag is driving productivity across Anthropic's entire org — eng, product, data, sales, and marketing — with their internal version landing 65% of product PRs. She's shared the CEO/CTO playbook for rolling it out and why security was designed in from day one. To help other teams try it, Anthropic is granting $25k in credits for Claude Enterprise orgs and $2.5k for Claude Team orgs to use Claude Tag through September 1.

Anthropic 的 Cat Wu（负责 Claude Code + Cowork）表示，Claude Tag 正在整个 Anthropic 公司层面提升生产力，覆盖工程、产品、数据、销售和市场团队，而他们的内部版本已经能合入 65% 的产品 PR。她还分享了推广落地的 CEO/CTO playbook，以及为什么安全从第一天起就被设计进去。为了让其他团队试用，Anthropic 向 Claude Enterprise 组织发放 2.5 万美元额度、向 Claude Team 组织发放 2500 美元额度，可在 9 月 1 日前用于体验 Claude Tag。

Anthropic's Boris Cherny (Claude Code) announced that Artifacts in Claude Code are expanding to Pro and Max plans, calling them "life changing" in his own workflow.

Anthropic 的 Boris Cherny（负责 Claude Code）宣布，Claude Code 里的 Artifacts 功能将扩展到 Pro 和 Max 订阅计划，并说这个功能在他自己的工作流里「改变了游戏规则」。

Anthropic's Thariq addressed the many questions about Fable's availability on subscription plans: while Fable will come off subscriptions after July 7, the team aims to restore it as a standard part of subscriptions "as soon as capacity allows."

Anthropic 的 Thariq 回应了大家关于 Fable 是否会保留在订阅计划里的疑问：虽然 Fable 会在 7 月 7 日之后从订阅中移除，但团队的目标是「一旦产能允许」，就把它重新作为订阅的标准组成部分恢复回来。

OpenAI's Thibault Sottiaux (Codex & ChatGPT) teased a coming model, GPT-5.6 Sol Ultra, urging people to "stash your hardest prompts somewhere" ahead of its release.

OpenAI 的 Thibault Sottiaux（负责 Codex 和 ChatGPT）预告了一款即将发布的模型 GPT-5.6 Sol Ultra，并提醒大家在它发布前「先把你手里最难的 prompt 存起来」。

FirstMark Capital VC Matt Turck released a deep conversation with NVIDIA's Bryan Catanzaro on the Nemotron models and what it takes to build a top AI lab. The through-line: why a chip company puts hundreds of researchers on building AI models and then gives them away for free. They dig into whether open source is catching the frontier, whether the US is falling behind China, training a 550B model in 4 bits, the hybrid Mamba-Transformer architecture, and the contrarian case that "open AI is safer than closed."

FirstMark Capital 的 VC Matt Turck 发布了一期与 NVIDIA 的 Bryan Catanzaro 的深度对话，聊 Nemotron 系列模型，以及打造一个顶尖 AI 实验室究竟需要什么。核心线索是：一家芯片公司为什么要投入几百名研究员去做 AI 模型，然后再免费送出去。两人深入探讨了开源是否正在追上前沿、美国是否正被中国甩开、如何用 4 bit 训练一个 550B 的模型、混合 Mamba-Transformer 架构，以及一个反直觉的观点：「open AI 比 closed 更安全」。

AI educator Peter Yang shared how he's getting the most out of Fable before July 7: prep context with cheaper models first, plan with Fable but execute with another model, and use lower effort (like Medium) while babysitting what it does. He also shared a hands-on kids project — his 8-year-old drew a dragon, they used Codex image generation to create more poses via voice feedback, then printed sticker sheets — a simple template for a summer activity with kids.

AI 教育者 Peter Yang 分享了他在 7 月 7 日之前如何把 Fable 用到极致：先用更便宜的模型准备好上下文，用 Fable 来做规划、但让另一个模型来执行，并用较低的 effort 档位（比如 Medium）同时盯着它在做什么。他还分享了一个亲子小项目：8 岁的女儿画了一条龙，他们把画上传给 Codex，用图像生成配合语音反馈做出更多姿势，再打印成贴纸，给暑假在家的孩子提供了一个简单可复制的玩法模板。

Builder Zara Zhang offered a few sharp takes. "The root of AI slop isn't bad style. It's no substance," she wrote. She relayed a recent college grad who fed lecture decks to an AI and had it teach the material instead of attending class — often finding the AI taught better than the professors. And her practical tip for multi-agent work: "talk in groups, not DMs."

Builder Zara Zhang 抛出了几个犀利观点。她写道：「AI slop 的根源不是风格差，而是没有实质内容。」她还转述了一位应届毕业生的做法：把课程讲义喂给 AI、让 AI 来讲这门课，而不是去上课，结果常常发现 AI 讲得比教授还好。她给多 agent 协作的实用建议是：「在群里聊，别用私聊。」

Every CEO Dan Shipper flagged a growing UX gap with long-running agents: with Fable, "it can go off for hours at a time and then comes back with a 2 paragraph explanation of what it did." His takeaway: "we need better ways for AI to tell us stories" about the work it actually did.

Every CEO Dan Shipper 点出了长时间运行的 agent 带来的一个越来越明显的体验缺口：用 Fable 时，「它可以一口气跑好几个小时，然后回来只给你两段话解释它做了什么。」他的结论是：「我们需要更好的方式，让 AI 把它到底做了哪些工作『讲成故事』告诉我们。」

Anthropic's Claude account announced "Built with Claude: Life Sciences," a global virtual hackathon with the Gladstone Institutes running July 7–13, with a $100k prize pool for researching and building with Claude Science and Claude Code. It also noted that Claude Fable 5 is now available in Claude Tag, alongside a conversation with Boris Cherny and Cat Wu on the path from Claude Code to Claude Tag.

Anthropic 的 Claude 官方账号宣布了「Built with Claude: Life Sciences」，这是一场与 Gladstone Institutes 合办的全球线上黑客松，时间为 7 月 7 日至 13 日，设有 10 万美元额度奖池，鼓励大家用 Claude Science 和 Claude Code 做研究与开发。同时它还提到，Claude Fable 5 现已在 Claude Tag 中上线，并发布了一段 Boris Cherny 与 Cat Wu 的对谈，讲述从 Claude Code 到 Claude Tag 的演进之路。

The Takeaway: Nuclear's real bottleneck was never physics — it's the willingness to build, iterate, and verticalize like a hardware startup instead of a modeling-and-simulation shop.

Isaiah Taylor is the founder and CEO of Valar Atomics, which just became the first private company since the discovery of fission to make nuclear power, going from Delaware incorporation to its first split atom in two years and four months. His central claim is that most of the nuclear industry is quietly "a modeling and simulation industry," producing exquisitely detailed "paper reactors" while almost nothing gets built. Valar's bet is the opposite: nuclear is a hardware execution problem.

Taylor's framing is that nuclear "has never had its Ford moment or its Tesla moment" — reactors were built as one-off civil infrastructure, never mass-manufactured. He borrows a metric from video games he calls "tick rate": the time between splitting one atom and the next. The goal is to drive that from months toward minutes, because nuclear's economics are dominated by how fast and cheaply you can produce plants, not by fuel, which is cheap.

The safety philosophy is contrarian. Rather than obsess over lowering the odds of an accident, Valar engineers for consequence: its regulatory safety case assumes everything in the plant has failed. To prove it, they scram the reactor, cut all electrical power and pumps, and let passive water jackets shed decay heat over two days with no moving parts.

The most striking thread is verticalization. Quoted $5M and a 2.5-year wait for a reactor protection system, five engineers locked themselves in a conference room and built one in six weeks for roughly $400k. They even invented their own grade of concrete — a rebar-free "modular citadel" whose seams follow a sine wave so radiation can't slip straight through — after a three-week "rock hunt" collecting samples across the US. "We just run toward gunfire on the most complicated things every time," Taylor says. His underlying bet: cheaper energy induces its own demand, forever.

关键要点：核能真正的瓶颈从来都不是物理，而是有没有意愿像硬件创业公司那样去真造、去迭代、去垂直整合，而不是当一家只做建模与仿真的公司。

Isaiah Taylor 是 Valar Atomics 的创始人兼 CEO。这家公司刚刚成为自核裂变被发现以来第一家真正发出核电的私营公司，从在 Delaware 注册到第一次裂变，只用了两年零四个月。他的核心论点是：核能行业的绝大部分其实悄悄地是「一个建模与仿真行业」，产出极其精细的「纸上反应堆」，却几乎什么都没真正建出来。Valar 的赌注恰恰相反：核能是一个硬件执行问题。

Taylor 的说法是，核能「从未迎来自己的 Ford 时刻或 Tesla 时刻」，反应堆一直被当作一次性的土木工程来建，从未被大规模制造过。他借用了一个来自电子游戏的指标，叫「tick rate」：从裂变第一个原子到下一个原子之间的时间。目标是把它从以月计压缩到以分钟计，因为核能的经济性主要取决于你能多快、多便宜地造出电站，而不是燃料，燃料本身很便宜。

它的安全理念也很反直觉。Valar 不是一味纠结于降低事故发生的概率，而是围绕「后果」来做工程：它提交给监管方的安全论证，直接假设电站里的一切都已失效。为了证明这一点，他们会紧急停堆、切断全部电力和水泵，让被动式水套在没有任何运动部件的情况下，用两天时间把衰变热带走。

最令人印象深刻的一条线索是垂直整合。一套反应堆保护系统被报价 500 万美元、要等两年半，于是五名工程师把自己关进会议室，用六周、约 40 万美元就做出了一套。他们甚至发明了自己的一种混凝土：一个无钢筋的「模块化城堡（modular citadel）」，接缝按正弦波走线，让辐射无法直穿而过，而这背后是一场为期三周、跑遍全美采集样本的「找石头行动」。Taylor 说：「我们每次都是径直冲向最复杂问题的枪林弹雨。」他底层的赌注是：更便宜的能源会永远不断地催生出对自身的需求。