June 18, 2026

Box CEO Aaron Levie laid out a detailed playbook for what the "Applied AI" layer looks like at scale, pushing back on the early critique that vertical AI apps are just a thin layer on top of an LLM. He argues the moat comes from four things: building tuned interfaces and bespoke tools that bridge raw intelligence and real workflows, acting as a model router that balances frontier models with cheaper ones based on deep task-specific evals, driving real change management through forward-deployed engineers, and running domain-specific go-to-market that speaks each industry's language. His key caveat is that even if the bitter lesson eventually means model intelligence alone solves everything, enterprises need help changing today, so most of the applied work matters no matter how good the models get.

Box CEO Aaron Levie 详细拆解了规模化的 "Applied AI"（应用层）到底长什么样，并反驳了早期那种 "垂直 AI 应用不过是套在 LLM 外面的一层薄壳" 的说法。他认为护城河来自四件事：构建经过调校的界面和定制工具，把底层智能和真实工作流连接起来；做一个 model router，基于针对具体任务的深度 eval，在前沿模型和更便宜的模型之间做平衡；通过 forward-deployed engineer 来真正推动企业的变革管理；以及做贴合行业语言的、领域专属的 GTM。他最关键的提醒是：即便 bitter lesson 最终意味着光靠模型智能就能解决一切，企业现在就需要帮助去做转型，所以无论模型变得多强，这些应用层的工作大多依然重要。

Vercel CEO Guillermo Rauch argued the AI SDK is more relevant than ever given the intense model competition, pointing out that GLM 5.2, an open model, just surpassed Opus 4.8 in Vercel's Next.js Evals. His bigger point: the world needs a practical, opinionated way to build and deploy agents, the same way React needed Next.js to become an actual web application framework.

Vercel CEO Guillermo Rauch 认为，在如此激烈的模型竞争格局下，AI SDK 比以往任何时候都更重要，并指出开源模型 GLM 5.2 刚刚在 Vercel 的 Next.js Evals 上超过了 Opus 4.8。他更大的观点是：这个世界需要一种实用且有明确主张的方式来构建和部署 agent，就像当年 React 需要 Next.js 才真正成为一个能做实际 Web 应用的框架一样。

YC President and CEO Garry Tan put a contrarian number on the temporary Fable 5 ban, estimating it cost roughly $12M per working hour in lost developer productivity — modeling 5M frontier AI-coding daily actives, a $90/hr fully-loaded cost, about 17.8% of work routed to Fable, and Fable being ~15% more productive. Separately he argued that AI now gives technical founders access to business thinking and business founders access to technical thinking, with the net result being more startups that actually work.

YC 总裁兼 CEO Garry Tan 给那次短暂的 Fable 5 封禁算了一笔很反直觉的账：他估计这每个工作小时大约造成了 1200 万美元的开发者生产力损失——按 500 万前沿 AI 编码日活、每小时 90 美元的全负担成本、约 17.8% 的工作被分流到 Fable、以及 Fable 平均高出约 15% 的生产力来测算。另外他还提出，AI 如今让技术型创始人也能接触到商业思维、让商业型创始人也能接触到技术思维，最终结果是会出现更多真正跑得通的创业公司。

Replit CEO Amjad Masad announced a "Design with Claude, Ship with Replit" integration, wiring Claude's design output directly into Replit's build-and-deploy flow so a design can move straight into a shippable app.

Replit CEO Amjad Masad 宣布了一个 "Design with Claude, Ship with Replit" 的集成，把 Claude 的设计产出直接接入 Replit 的构建和部署流程，让一份设计可以一路走到能上线的应用。

OpenAI CEO Sam Altman announced that Noam — someone he says he's wanted to work with since the very beginning of OpenAI — is finally joining, quipping "only took 10 years" and predicting it will be worth the wait. In a follow-up he joked the team offers "no explanation as to why Noams are so good at AI; we attribute their success, as all else, to divine benevolence."

OpenAI CEO Sam Altman 宣布 Noam 终于要加入了——他说这是自 OpenAI 创立之初就最想合作的人之一，并调侃 "只花了 10 年"，相信这份等待是值得的。在后续的一条帖子里他开玩笑说，团队 "无法解释为什么这些 Noam 在 AI 上这么强；我们和对待其他一切一样，把他们的成功归功于神的恩典。"

OpenAI's Thibault Sottiaux, who works on Codex and ChatGPT, reminded developers that the Codex App, CLI, and SDK all work with any open-source model, not just OpenAI's own. The note lands as model-neutral tooling becomes a bigger competitive theme across the industry.

负责 Codex 和 ChatGPT 的 OpenAI 的 Thibault Sottiaux 提醒开发者，Codex 的 App、CLI 和 SDK 都可以搭配任意开源模型使用，而不仅仅是 OpenAI 自家的模型。这条提醒恰逢 "模型中立的工具链" 正成为全行业越来越重要的竞争主题。

Anthropic's Claude account announced that Claude Design is now in beta on all paid plans across web and desktop, and that Claude Design and Claude Code now work together in both directions: hand a design off to build, or start in Claude Code and sync design projects from your terminal. Work can be exported to PDF and PowerPoint, and the redesigned editor adds direct on-canvas controls to drag, resize, and align elements.

Anthropic 旗下的 Claude 账号宣布，Claude Design 现已在 Web 和桌面端面向所有付费方案开放 beta，并且 Claude Design 和 Claude Code 现在可以双向协作：既可以把设计交付给开发，也可以在 Claude Code 里起步、再从终端同步设计项目。成果可以导出为 PDF 和 PowerPoint，重新设计的编辑器还新增了在画布上直接拖拽、缩放和对齐元素的控件。

Linear's head of product Nan Yu pushed on a common confusion in the discourse around "taste," arguing it isn't just taste in aesthetics, the same way "design" isn't just visual design. His point: half the online arguments on the subject are people talking past each other because they collapse those broader meanings into the narrow ones.

Linear 的产品负责人 Nan Yu 针对当下关于 "taste（品味）" 的讨论里一个常见的混淆发难：他认为 taste 不只是审美上的品味，就像 "design（设计）" 不只是视觉设计一样。他的观点是：网上关于这个话题的争论有一半其实是各说各话，因为大家把这些更宽泛的含义压缩成了狭义的理解。

Builder Zara Zhang offered two pointed takes. First, don't use AI for writing until you've developed your own taste and voice, because if you can't recognize slop, you'll happily ship it. Second, on vibe-coded personal apps: building the thing takes a day, but finding out whether you'll actually use it takes a week, and most products are built for an idealized user who doesn't exist, so you should build for the lazy, forgetful real human instead.

Builder Zara Zhang 抛出了两个很尖锐的观点。第一，在你形成自己的品味和声音之前，不要用 AI 来写作，因为如果你连 slop（劣质内容）都认不出来，你就会心安理得地把它发出去。第二，关于 vibe coding 出来的个人 app：把东西做出来只要一天，但要搞清楚自己到底会不会真的用它，得花上一周；大多数产品都是为一个根本不存在的理想用户而造的，所以你应该为那个又懒又健忘的真实人类去设计。

FPV Ventures partner Nikunj Kothari went after tranched funding rounds, arguing they hurt the next employee who joins by inflating the 409A valuation above what the lead preferred investor actually paid. He offered a simple tell: if a company raises a large dollar amount at under 10% dilution of the published valuation, it was almost certainly a tranched round — and he's now seeing these creep from seed and Series A up into Series B.

FPV Ventures 合伙人 Nikunj Kothari 把矛头对准了分期交割（tranched）的融资轮，认为它会坑到下一个加入的员工，因为它把 409A 估值抬到了高于领投优先股投资人实际支付价格的水平。他给了一个简单的判别方法：如果一家公司以低于其公布估值 10% 的稀释比例融到了一大笔钱，那几乎可以肯定这是一轮分期交割——而且他现在发现这种做法正从种子轮和 A 轮蔓延到 B 轮。

Every CEO Dan Shipper announced an investment in Tacit, a founder he's a fan of, and pointed back to his 2023 essay "Against Explanations" on how AI might reshape the sciences, saying he's extremely pumped at the prospect of so much scientific progress happening so fast.

Every CEO Dan Shipper 宣布投资了 Tacit（他很欣赏的一位创始人），并重新提起了他 2023 年写的文章 "Against Explanations"，讲 AI 可能如何重塑各门科学，表示自己对这么多科学进展能如此之快地发生感到无比兴奋。

The Takeaway: the flood of AI-generated code isn't slop drowning developers — it's the early slope of a curve where the scarce, defensible skill becomes giving agents enough of your personal context that they can "complete your thought."

Kyle Daigle, COO of GitHub and now also Microsoft's chief marketing officer for developers, has a front-row seat to the agent economy: 17 million pull requests were created by agents in a single month, and GitHub is on track for roughly 14 billion commits this year. His contrarian read is that this output isn't worthless slop. "We're all just actually getting to the point where we're no longer in the super early adoption... no matter where you're building or what tools you're using, all of that code ends up on GitHub."

On the $200-to-$2,000 subscription fear everyone has, his answer isn't only cheaper tokens — it's automatic model routing tied to task intent, so the system quietly downshifts to a small model for a find-and-replace instead of burning frontier tokens. He's also bullish on "frontier tuning" enterprise models on a company's own documents and chats, an idea he admits first sounded like "a magic parlor trick" but that he now sees as where the alpha hides: "sometimes that's where the alpha is. It's like where it feels like this is too simple to work."

Maybe the most human insight: Daigle runs a personal agent he named Baxter that reads everything he writes and says, then sends him a daily comms report critiquing how he communicates. He leans on a thesis from GitHub's old chatops days — people take critical feedback better from a robot than from another human. His real AI loop is less about shipping software and more about a recursive self-improvement loop for himself.

On open-source maintainers drowning in agent PRs, his stance is restraint: give maintainers more controls and tooling rather than imposing one standard from the top, and only cement a system if one organically emerges from the community.

核心要点：AI 生成的代码洪流并不是淹没开发者的 "slop（垃圾）"，而是一条曲线的早期爬坡段——在这条曲线上，真正稀缺、真正有护城河的能力，变成了给 agent 足够多关于你的上下文，让它能 "把你的想法补全"。

Kyle Daigle 是 GitHub 的 COO，如今同时还担任微软面向开发者的首席营销官，他对这场 agent 经济有着第一线的观察：单月就有 1700 万个 pull request 是由 agent 创建的，GitHub 今年有望达到约 140 亿次 commit。他有一个反直觉的判断：这些产出并不是没有价值的垃圾。"我们其实都正走到这样一个节点——不再处于非常早期的尝鲜阶段……无论你在哪里构建、用什么工具，所有这些代码最终都会落到 GitHub 上。"

对于人人都担心的 "200 美元订阅变成 2000 美元" 问题，他的答案不只是更便宜的 token，而是与任务意图绑定的自动 model routing：系统会悄悄地为一次查找替换降档到小模型，而不是去烧前沿模型的 token。他还看好对企业模型做 "frontier tuning"，用公司自己的文档和聊天记录来微调。他坦言这个想法一开始听起来像 "魔术表演里的小把戏"，但现在他认为 alpha 恰恰藏在这里："有时候 alpha 就在那种感觉上简单到不可能奏效的地方。"

也许最具人情味的洞见是：Daigle 自己跑着一个他起名叫 Baxter 的私人 agent，让它读他写的、说的一切，然后每天给他发一份点评他沟通方式的 "comms report"。他援引了 GitHub 早年 chatops 时代的一个论点——比起被另一个人批评，人们更容易接受来自机器人的批评意见。他真正的 AI 循环，与其说是关于交付软件，不如说是一个针对他自己的递归式自我提升循环。

至于被 agent 提交的 PR 淹没的开源维护者，他的立场是克制：给维护者更多的控制权和工具，而不是自上而下强加一个统一标准，只有当社区自发地涌现出某种做法时，才会把它固化下来。