七月十九日

OpenAI's Codex and ChatGPT lead Thibault Sottiaux says his job is now "basically delegating to ChatGPT Work" — he can't stop using dictation. As proof, he posted the raw prompt he spoke into the app: go through thousands of Twitter DMs replying to his ChatGPT Work beta callout, build a spreadsheet with each person's name, DM link, and post, invent an 8-12 label taxonomy for what they do, and rate how sophisticated each person's existing workflow is so the team can select a beta cohort spanning adoption levels. He also spelled out the product pitch: ChatGPT Work creates and hosts sites, manages your email, summarizes mountains of documents, and produces docs, sheets, and slides — already in the mobile app and included in Plus, Pro, Business, and Enterprise plans.

OpenAI 负责 Codex 和 ChatGPT 的 Thibault Sottiaux 说他的工作现在"基本上就是把活儿委派给 ChatGPT Work"，而且已经离不开语音输入了。作为证明，他晒出了自己对着 app 口述的原始 prompt：让 ChatGPT Work 翻遍几千条回复他 ChatGPT Work 内测招募的 Twitter 私信，整理成一张表格，包含每个人的名字、私信链接和原帖，自己发明一套 8 到 12 个标签的分类体系来归类这些人的职业，并给每个人现有 workflow 的成熟度打分，方便团队挑选一个覆盖不同采用水平的内测群体。他还完整列出了产品卖点：ChatGPT Work 能创建并托管网站、替你管理邮件、总结成堆的文档、产出高质量的 docs、表格和幻灯片，已经上线移动端 app，并包含在 Plus、Pro、Business 和 Enterprise 套餐里。

Vercel CEO Guillermo Rauch published internal stealth-eval results on cybersecurity: Kimi K3 is "top-tier" — countering chatter on X that Moonshot overfit the benchmarks ("These are stealth evals. Model has raw IQ") — GPT-5.6 Sol is "a leap ahead in cyber capability" at significantly higher cost, and Fable "refuses everything," failing to complete the run at all, even though Sol was much more open to helping with defensive cyber hardening. His TL;DR: frontier, open-weight cybersecurity capability is here. Separately, he argued the term "AGI" has aged very poorly: AI is already "far better than human intelligence, for most economically-relevant tasks," making superintelligence the better term — yet these systems horribly suck at writing, because robots produce robotic prose even with "a corpus of your writing and 200 Skills and 50 subagents running in gRaPh LoOps." The one definitive way people become irrelevant is to stop being themselves and delegate away their identity and unique thoughts; "quality and humanity will prevail."

Vercel CEO Guillermo Rauch 公布了内部关于网络安全的隐藏评测结果：Kimi K3 属于"顶级水平"，直接回击了 X 上关于 Moonshot 过拟合基准测试的传言（"这些是隐藏评测。这个模型有原生智商"）；GPT-5.6 Sol 在网络安全能力上"领先一个身位"，但成本明显更高；而 Fable"拒绝一切请求"，完全没能跑完评测，相比之下 Sol 在协助防御性网络加固方面开放得多。他的结论：前沿水平的开放权重网络安全能力已经到来。另外，他认为"AGI"这个词已经严重过时：对于大多数有经济价值的任务，AI 已经"远强于人类智能"，所以 superintelligence 才是更准确的说法。但这些系统写作水平烂得离谱，因为机器人写出来的就是机器味十足的文字，哪怕你喂给它"你全部的文集、200 个 Skills 和 50 个跑在图循环里的 subagent"也没用。人变得无关紧要只有一条明确的路径：不再做自己，把自己的写作、独特的想法和身份全都委派出去。"质量和人性终将胜出。"

Box CEO Aaron Levie called the past few months a turning point against the thesis that AI value will only accrue to a few companies, mapping the categories already working: shops that tune and run inference on custom models for enterprise use cases; applied AI companies owning the end-user workflow (legal, IT, security, HR, customer support, coding) that act as model-agnostic routing layers and drive change management; new vertical labs going deep in life sciences, financial services, and healthcare; a multi-layer infrastructure stack for running, governing, and securing agents; and a coming wave of AI-native services firms. "Way too early to call the winning architectures" — it will be heterogeneous like every other tech market. He also argued gatekeeping models won't work at scale: China has crossed the rubicon on near-frontier capability — "If this were even just 1 month ago and we saw kimi k3 our brains would be broken" — and the answer isn't locking down your own ecosystem but safely keeping the rate of progress high, driving diffusion and infrastructure, and enabling US open source.

Box CEO Aaron Levie 认为过去几个月是一个转折点，直接打脸"AI 价值只会沉淀到少数几家公司"的论断。他梳理了已经在跑通的几个方向：帮企业针对特定场景调优模型并跑推理的服务商；掌控终端用户 workflow 的应用型 AI 公司（法务、IT、安全、HR、客服、编程），它们模型无关、扮演路由层并推动企业变革管理；在生命科学、金融、医疗等领域深耕的新型垂直实验室；为运行、治理和保护 agent 而生的多层基础设施栈；以及即将涌现的一大批 AI 原生服务公司。"现在断言哪种架构会赢还为时过早"，最终会像所有科技市场一样走向异构共存。他还指出，对模型搞封锁在规模上行不通：中国在准前沿能力上已经过了卢比孔河，"哪怕只是一个月前看到 Kimi K3，我们的脑子都会被震碎"。答案不是把自己的生态锁得更死，而是在安全的前提下保持进步速度、推动技术扩散和基础设施建设、扶持美国的开源力量。

Swyx of Cognition and Latent Space pushed back on Europe-bashing: Europe actually has some of the top AI engineers in the world if you elicit the right ones, and measured by talks and workshops his events are "basically running the most competitive global arena for AI talent." He also shared a striking data point on answer-engine optimization: at the current rate, AEO will be fully responsible for $1M of his revenue next year.

Cognition 和 Latent Space 的 Swyx 反驳了唱衰欧洲的论调：只要你会挖掘，欧洲其实拥有世界顶级的 AI 工程师，而以演讲和 workshop 的水准衡量，他的活动"基本上是在运营全球竞争最激烈的 AI 人才竞技场"。他还分享了一个关于答案引擎优化的惊人数据：按目前的势头，明年 AEO 将为他直接带来 100 万美元的收入。

FirstMark VC Matt Turck skewered a three-year-old consensus in four lines: "2024: 'The model layer is commodotizing.' 2025: 'The model layer is commodotizing.' 2026: 'The model layer is commodotizing.' The model layer: still not commodotized."

FirstMark 风投合伙人 Matt Turck 用四行字调侃了一个流传三年的共识："2024 年：'模型层正在商品化。'2025 年：'模型层正在商品化。'2026 年：'模型层正在商品化。'模型层：至今仍未商品化。"

Builder Zara Zhang argued everyone should develop a "personal eval set" for AI models: a few tasks actually relevant to your own day-to-day work, because industry benchmarks may not reflect what makes a model useful to you — "you find the model's capability boundary by poking at it & bumping into it for fun." She also nailed the biggest barrier to enterprise AI adoption in one line: the people who understand AI don't understand the business, and the people who understand the business don't understand AI.

Builder Zara Zhang 提出每个人都该给 AI 模型建一套"个人 eval 集"：挑几个和你日常工作真正相关的任务，因为行业基准测试未必能反映模型对你个人的实际用处，"你要靠好玩地去戳它、撞它，才能摸到模型的能力边界"。她还一句话点破了企业 AI 落地的最大障碍：懂 AI 的人不懂业务，懂业务的人不懂 AI。

Creator Economy author Peter Yang built a ChatGPT Site with his 8-year-old to help her learn the multiplication tables: since she loves animals, they used ChatGPT Images to generate the UI and characters, added music, and even built a timed boss level — the playable game is linked in his post.

Creator Economy 作者 Peter Yang 和他 8 岁的孩子一起用 ChatGPT Site 做了一个学乘法口诀的小游戏：因为孩子喜欢动物，他们用 ChatGPT Images 生成了界面和角色，配上了音乐，甚至还加了一个限时 boss 关卡。游戏链接就在他的帖子里，可以直接玩。

Anthropic Claude Code team member Thariq reflected on shipping Fable to subscription plans, quote-tweeting the official announcement: it took "a heroic effort by many people at Anthropic working sometimes literally around the clock," and "it was not at all clear that we'd be able to do this in time." His sign-off: "Enjoy Fable."

Anthropic Claude Code 团队的 Thariq 转评官方公告，回顾了把 Fable 推进订阅套餐的过程：这背后是"Anthropic 许多人的英雄式付出，有时真的是连轴转"，而且"当时完全不确定我们能不能按时做到"。他的收尾是一句："Enjoy Fable。"

The Takeaway: The man often called the father of modern AI is wildly optimistic about AI technology and deeply pessimistic about the companies selling it — expect a stock market renormalization, not a winner-take-all AGI moment.

Jürgen Schmidhuber — dubbed the father of AI by the New York Times and Forbes, the researcher behind foundational work on LSTMs, meta-learning, and artificial curiosity, now at KAUST — has chased the same goal since the 1970s: build an AI smarter than himself so he can retire. His case against the CapEx boom is simple arithmetic: compute gets roughly 10x cheaper every five years, so "the guys who are investing a thousand billion dollars into GPUs for data centers today, within the next five years, they are going to lose $900,000,000,000. There's no business model that can recuperate that loss." Open-source models that catch up within months keep prices pinned to the floor, nimble software companies are morphing into capital-hungry utilities, and free cash flow is quietly collapsing. The coming correction, he says, is a renormalization of misallocated capital, not the end of civilization.

He's equally contrarian on recursive self-improvement as a moat: nearly every key AI algorithm came from small academic labs, the self-improvement crowd skews open source, and "everybody's cooking with the same water" — no company can hold an RSI lead long enough to monetize it.

The deeper limit is physical. "You can't have AGI just behind the screen." Robot hardware remains vastly inferior to the human hand — millions of sensors, self-healing when cut — which is why "the robots of the movies are all played by humans." Comparable hardware may take decades, not years.

What comes next, he argues, are artificial scientists. Today's LLMs are "super biased towards humans" because they train only on what some human found interesting, while babies learn by running experiments on the world. Curiosity-driven agents that invent their own questions and generate their own data are the future — and that same property is why he never signed the alignment letters: ten humans in a room have ten different objective functions, and the AI drones already fighting each other over Ukraine have none at all. His long-run comfort: true artificial scientists "will be fascinated by life" — and motivated to protect the source of interesting patterns rather than destroy it.

核心要点：这位常被称为现代 AI 之父的人对 AI 技术极度乐观，却对卖 AI 的公司深度悲观。等着看的是一场股市的价值重估，而不是赢家通吃的 AGI 时刻。

Jürgen Schmidhuber，被纽约时报和福布斯称为 AI 之父，LSTM、meta-learning 和人工好奇心等奠基性工作背后的研究者，现执教于 KAUST。他从上世纪 70 年代起追逐同一个目标：造出比自己更聪明的 AI，好让自己退休。他看空这轮 CapEx 狂潮的理由是一道简单的算术题：算力每五年便宜大约 10 倍，所以"今天往数据中心 GPU 里砸一万亿美元的这些人，未来五年内将亏掉 9000 亿美元。没有任何商业模式能挽回这个损失。"几个月内就能追平的开源模型把价格死死摁在地板上，曾经轻盈的软件公司正在变成重资产的公用事业公司，自由现金流在悄悄崩塌。他说，即将到来的修正是对资本错配的一次重估，而不是文明的终结。

在"递归自我改进能否构成护城河"这个问题上他同样反主流：几乎所有关键的 AI 算法都出自小型学术实验室，搞自我改进的这批人大多偏向开源，"大家用的都是同一锅水"，没有哪家公司能把 RSI 的领先优势保持到足以变现的程度。

更深层的限制在物理世界。"你不可能只在屏幕后面实现 AGI。"机器人硬件依然远逊于人类的手，这只手有数百万个传感器，割伤了还会自愈，这正是"电影里的机器人都由人类扮演"的原因。可与之相比的硬件也许需要几十年，而不是几年。

他认为接下来的方向是人工科学家。今天的 LLM"对人类有极强的偏向性"，因为它们只在某个人类觉得有趣的数据上训练，而婴儿是靠对世界做实验来学习的。由好奇心驱动、会自己提出问题、自己生成数据的 agent 才是未来。也正因为这种特性，他从未在任何 alignment 公开信上签过名：一个房间里的十个人就有十套不同的目标函数，而已经在乌克兰上空互相厮杀的 AI 无人机根本没有任何对齐可言。他对长远的安慰是：真正的人工科学家"会对生命深深着迷"，会有动力去保护有趣模式的源头，而不是毁掉它。