June 15, 2026

Box CEO Aaron Levie made two big-picture arguments. Endorsing Microsoft CEO Satya Nadella's post on AI "learning loops," he argues the companies that win will be those that get their unique IP and institutional knowledge into an architecture that captures every future AI gain. In his words, "so much of the power and value will accrue to wherever can best leverage any AI system against their information," which is why he expects the applied-AI layer to keep gaining value.

In a separate, more contrarian take, Levie argues the real winner of a model being "pulled back" is open-weights models: now that the once-theoretical risk of a model becoming unavailable has a precedent, every country has more incentive to build sovereign AI, eroding US leadership over time. His warning to US policymakers is that regulating at the model layer rather than the applied layer pushes other nations toward open weights that mostly aren't American today.

Box CEO Aaron Levie 抛出了两个宏观判断。在转发 Microsoft CEO Satya Nadella 关于 AI "learning loop" 的文章时,他认为未来的赢家,会是那些把自己独有的 IP 和组织知识沉淀进一套架构、从而能吃下 AI 每一次进步红利的公司。用他的话说,"绝大部分的力量和价值,会流向那些最能把任意 AI 系统作用于自身信息之上的人",这也是他看好应用层 (applied-AI layer) 会持续增值的原因。

在另一个更具争议的观点里,Levie 认为某个模型被"撤回"之后,真正的赢家是 open-weights 模型:既然"模型可能变得不可用"这个过去只停留在理论上的风险如今有了先例,那么每个国家都更有动力去发展主权 AI (sovereign AI),长期看会削弱美国的领导地位。他给美国政策制定者的警告是:在模型层而非应用层做监管,会把其他国家推向 open weights,而这些模型如今大多并非来自美国。

Swyx (Cognition, Latent Space) shared early hands-on notes on "ultracode," Anthropic's token-heavy parallel-agent mode: it's "scarily good at burning tokens," but you have to set up your repo to parallelize properly to exploit the subagent fan-out. His bigger point is to think of subagents as "subroutines but intelligent" — once you notice how much knowledge work is just "yakshaves after yakshaves that require some judgment and intelligence," you see dynamic agent workflows aren't only for coding.

He also amplified Satya Nadella's framing of "loops as IP": "the real opportunity is not in picking the best model but instead in building a learning loop on top of models where human capital and token capital compound. You can offload a task, or even a job, but you can never offload your learning."

Swyx (Cognition、Latent Space) 分享了他对 "ultracode" 的早期实操心得 (这是 Anthropic 那个极其消耗 token 的并行 agent 模式):它"烧起 token 来好得吓人",但你得把自己的 repo 配置好、让它能真正并行,才能吃到 subagent fan-out 的红利。他更大的观点是:把 subagent 理解成"有智能的子程序 (subroutines but intelligent)"。一旦你意识到大量知识工作其实就是"一个接一个、需要判断力和智能的杂活",你就会明白动态 agent workflow 并不只属于写代码。

他还转发并力挺了 Satya Nadella 关于"把 loop 当作 IP"的论述:"真正的机会不在于挑选最好的模型,而在于在模型之上构建一套 learning loop,让人力资本和 token 资本一起复利。你可以外包一个任务、甚至一份工作,但你永远无法外包自己的学习。"

Replit CEO Amjad Masad endorsed the same Satya Nadella post, calling it "the most inspiring positive-sum vision for AI in the enterprise" — a vote for a frontier ecosystem where every organization owns the learning loop that encodes its institutional knowledge, rather than a winner-take-all race for the single best model.

Replit CEO Amjad Masad 也力挺了同一篇 Satya Nadella 的文章,称其为"企业 AI 领域最鼓舞人心的正和 (positive-sum) 愿景":他支持的是一个前沿生态,让每个组织都拥有那套承载自身组织知识的 learning loop,而不是一场只为争夺"单一最强模型"的赢家通吃式竞赛。

Y Combinator CEO Garry Tan offered two takes. On the open-versus-closed debate, he frames open source as "the escape hatch for businesses to be able to continue to control their own destiny long term."

And on careers, he predicts the next generation of world-changers will be those most adept at making "long-running multi-stage multi-team agent tasks work extremely well, and at high volume" across both their personal and work lives.

Y Combinator CEO Garry Tan 给出了两个判断。关于开源与闭源之争,他把开源比作"让企业能长期掌控自身命运的逃生舱"。

关于职业,他预测下一代改变世界的年轻人,几乎一定是那些最擅长把"长周期、多阶段、多团队的 agent 任务大规模、高质量地跑通"的人,而且会把这种能力贯穿进自己生活和工作的方方面面。

OpenAI's Thibault Sottiaux (Codex & ChatGPT) announced that Codex can now see and set its own /goal — a generalization of meta-prompting where the agent derives its own task from your intent. His framing principle: "Everything we build, we build also as a tool for the agent."

OpenAI 的 Thibault Sottiaux (负责 Codex 和 ChatGPT) 宣布,Codex 现在能看到并自行设定它自己的 /goal,这是 meta-prompting 的一种泛化:让 agent 根据你的意图自行推导出该做的任务。他奉行的原则是:"我们造的每样东西,同时也是给 agent 用的工具。"

Vercel CEO Guillermo Rauch flagged a milestone: an open, community-driven ecosystem has passed 700,000 "skills," all organic.

Vercel CEO Guillermo Rauch 提到了一个里程碑:这个完全由社区自发驱动的开放生态,"skill" 数量已经突破 70 万个。

Builder Zara Zhang shared a sharp lesson on building AI "skills": don't start by writing one. "You don't make a good skill by writing a skill. You make it by doing the thing, fixing it 20 times, then telling the AI to bottle up everything you just did." In her words, "you make a skill by ending with one, not starting with one."

Builder Zara Zhang 分享了一条关于打造 AI "skill" 的犀利心得:别一上来就去写。"做出好 skill 的方式,不是去写一个 skill,而是真的去把那件事做一遍、反复修正 20 次,然后再让 AI 把你刚做的一切打包封装起来。"用她的话说,"好 skill 是干完之后倒推出来的,不是一开始凭空写出来的。"

The Takeaway: Computing just flipped from retrieval to generation, and Nvidia's job is to manufacture intelligence the way power plants manufacture electricity — a shift Jensen Huang says we're only ~$1 trillion into a future $20-trillion-a-year industry.

Speaking to an audience of global investors and manufacturers, Nvidia founder and CEO Jensen Huang reframed AI not as a chatbot novelty but as an industrial revolution on the scale of electricity and the internet. His central metaphor: 300 years ago, Siemens built the dynamo, which turned motion — wind, steam, water — into electrons. Nvidia builds the modern dynamo: electrons go in, and numbers come out. Those numbers, recombined, become language, proteins, physics, robotics. "These tokens are intelligence. That's it. That's what we do. It's not that hard."

The deeper shift is from a 60-year-old retrieval paradigm — where data centers store files you fetch later — to real-time generation, where content is produced fresh for every person. "Every single pixel that you see, every single sound that you hear in the future, every video you see in the future will be originally generated, not retrieved," which is why the world needs vastly more "generators."

For investors, Huang offers a five-layer cake: energy at the base, then chips and networking, then infrastructure (land, power, data centers), then models like OpenAI and Anthropic, then applications — where $100B of venture capital flowed last year alone.

His most contrarian beat is on jobs. He recounts how a famous computer scientist predicted computer vision would wipe out radiology a decade ago; instead, demand and the number of radiologists rose, because their purpose (diagnosing disease alongside doctors) outlived the task of reading scans. "You may or may not lose a job to an AI. But you will absolutely lose a job to someone who uses AI." His advice: ignore the Terminator talk and engage.

核心要点: 计算刚刚从"检索"翻转为"生成",而 Nvidia 要做的,就是像发电厂制造电力一样去制造智能。Jensen Huang 说,这个未来每年规模高达 20 万亿美元的产业,我们才刚投入大约 1 万亿美元。

面对一群来自全球的投资人和制造业者,Nvidia 创始人兼 CEO Jensen Huang 把 AI 重新定义为一场堪比电力和互联网的工业革命,而非聊天机器人式的新鲜玩意。他的核心比喻是:300 年前 Siemens 造出了 dynamo (发电机),把风、蒸汽、水这些"运动"转化为电子;而 Nvidia 造的是现代版 dynamo,电子进去,数字出来。这些数字重新组合,就成了语言、蛋白质、物理、机器人。"这些 token 就是智能。就这么回事。这就是我们干的活。其实没那么难。"

更深层的转变,是从一个已有 60 年历史的"检索"范式 (数据中心存好文件、你之后再取),转向实时生成,为每个人现场生产内容。"未来你看到的每一个像素、听到的每一段声音、看到的每一段视频,都将是原生生成的,而不是被检索出来的",正因如此,世界需要多得多的"生成器"。

面向投资人,Huang 给出了一个"五层蛋糕":底层是能源,往上是芯片与网络,再往上是基础设施 (土地、电力、数据中心),然后是模型层 (OpenAI、Anthropic),最顶层是应用,光去年一年就有 1000 亿美元 VC 涌入。

他最反共识的一段是谈工作。他讲到十多年前一位著名计算机科学家预言 computer vision 会让放射科消失;结果恰恰相反,需求和放射科医生的数量都涨了,因为他们的目的 (和医生一起诊断疾病) 比"读片"这个任务活得更久。"你也许会、也许不会被一个 AI 抢走工作。但你一定会被一个会用 AI 的人抢走工作。"他的建议是:别理会那些 Terminator 式的言论,去拥抱它。