Lecun is back. tl;dr huge risk of bubble bursting soon "The prices are going up of those AI services, but the cost of running them is going down, but not nearly fast enough. And so all of those companies are losing money, and basically, the use for most people is funded by the investors. That can't go on for a very long right?" LeCun said."

译LeCun回来了。长话短说，泡沫随时可能破裂。 "AI服务价格在上涨，但运营成本在下降，不过下降速度远不够快。所以这些公司都在亏损，基本上，大多数用户的使用是由投资者资助的。这种情况不可能持续太久，对吗？"LeCun说。

Rohan Paul@rohanpaul_ai · 6月22日45

"The model is no longer the product. Codex, Perplexity Computer, or Claude Code - all are orchestration system. It takes a model and pairs it with an agent harness. What is an agent harness ? The rules for how the agent loops around" - Aravind Srinivas

译"模型不再是产品。 Codex、Perplexity Computer 或 Claude Code —— 全都是编排系统。它用一个模型，再配上 agent harness。什么是 agent harness？Agent 循环运行的规则。" - Aravind Srinivas

Rohan Paul@rohanpaul_ai · 6月22日50

ASML’s CEO: Europe is falling behind in AI hardware as the US is buying 80% of the world’s advanced chips while megafabs like Tesla’s TeraFab may demand capacity at the scale of millions of wafers per month. --- wccftech. com/asml-ceo-warns-europe-is-quite-behind-in-ai-race-as-us-buys-80-of-advanced-chips/

译ASML’s CEO：欧洲在AI硬件方面落后，因为美国购买了全球80%的先进芯片，而像特斯拉TeraFab这样的超级工厂可能每月需要数百万片晶圆的产能。

Yuchen Jin@Yuchenj_UW · 6月22日43

Looking at my timeline, it feels like GLM-5.2 is having its DeepSeek R1 moment. I never thought an open-source model could break into the top 3 coding models this soon.

译看我的时间线，感觉 GLM-5.2 正迎来它的 DeepSeek R1 时刻。我从未想过一个开源模型能这么快跻身编程模型前三。

Rohan Paul@rohanpaul_ai · 6月22日44

The original vibe coder

译一条推文调侃 Linus 是原始的 vibe coder：他只需在邮件列表上发一条愤怒邮件描述需求，全球数千工程师就免费实现。零 token 消耗、零 API 成本，拥有无限上下文（30 多年内核知识）。对比 OpenAI 和 Anthropic 试图用 AI 复制 Linus 从 1991 年起靠人类贡献者做的事，但 Linus 的“agents”不产生模型幻觉且完全免费。

Chubby♨️@kimmonismus · 6月22日55

Today we’re publishing an exclusive interview with Geir Engdahl, Co-Founder and CTO, AI at @CogniteData . A very sharp conversation on why industrial AI usually doesn’t fail at the model layer, but on the plant floor. We talked about: Why impressive AI pilots often never survive real operations What “context” actually means when mistakes can become dangerous Where industrial agents should recommend, automate, or stay out Why the gap between a working demo and a trusted system is massive And Geir’s bold 2028 prediction for industrial companies without AI-driven process optimization One of the most sober and least hype-driven conversations I’ve had this year about AI in the real world. Read the full interview in today’s Superintelligence.

译Cognite CTO Geir Engdahl接受专访指出，工业AI失败的根源并非模型层，而是工厂现场运营。许多惊艳的AI试点无法在实际运营中存活，因为缺乏真正上下文——当错误可能带来危险时尤为关键。他探讨了工业智能体应在何时推荐、自动化或保持不介入，强调工作演示与可信系统之间存在巨大鸿沟。他预测，到2028年未采用AI驱动流程优化的工业企业将面临严峻挑战。这是一场务实、少谈噱头的对话。

Ethan Mollick@emollick · 6月22日64

A fundamental problem with extending Codex/Cowork/Code to all knowledge work is that they remain very "software-brained" where the end result (the software) is what is important & that code serves as a source of truth. For a lot of other knowledge work, the process is at least as important as the outcome. This includes researching what is known, an exploration of alternatives, failed efforts, prototype branches, experiments, etc. All of those things are valuable, so you cannot use the PowerPoint at the end the way you can use a codebase, nor is progress on a to-do list sufficient context post compaction. You work in learning loops, refining your perspectives as you go. In some ways, this makes long-running models like Fable hard to use for deep knowledge work, since they are designed to deliver product to you in the end. You can prompt your way around this problem, but everything about the Codex and Code harnesses want you to be a software developer and you have to fight them. There is a real disconnect between how a manager or analyst thinks about problems and how the agentic software tools approach solving them. Addressing this is critical to breaking out of the coding niche for these tools.

译Ethan Mollick指出，Codex/Cowork/Code等Agentic工具本质上是“软件脑”设计，只重最终代码，而多数知识工作的过程（研究、探索、原型分支等）与结果同样重要。长时运行模型Fable也因专注交付最终产品而难以用于深度知识工作，用户需费力用提示词绕过限制。这种工具与管理者/分析师思考方式的脱节，是突破编程领域、扩展到其他知识工作的关键障碍。

Rohan Paul@rohanpaul_ai · 6月21日23

Chamath prompted Claude to read Dario Amodei’s own blog and based on that do a psychological analysis of Dario. 😄

译Chamath 提示 Claude 阅读 Dario Amodei 自己的博客，并基于此对 Dario 进行心理分析。😄

meng shao@shao__meng · 6月21日63

60 分钟，成为 AI Native 组织：人、Agent 与上下文三层系统来自 @gregisenberg 和 @TheoTabah 60 分钟的深度对谈，一家公司究竟怎样才算 AI native，如何把它落地成可运转的系统？ https://www.youtube.com/watch?v=LztPaNmcWGU "用 ChatGPT ≠ AI native"，就像"有网站 ≠ 科技公司"。真正的 AI Native 由三层系统构成：人、Agent、上下文。 ① 人 —— 退守两端每个人都是 manager。AI 吃掉"中间的执行"，人聚焦两端： · 前端：战略、品味、判断 · 后端：沟通、评审、信任判尺（Andy Grove）：管理者的成功 = 团队的成功。把 agent 当下属来 set up for success。 ② Agents —— 满足四要素才自治 Agent = model 在 loop 中用工具。成熟度分三层：基础聊天 → 半自治（不停点 approve）→ 自治（独立运行数天/数周）。达到自治需四样：Clear Goal、Skills、Tools、Context——缺一即失败。两个关键机制： · Eval：把"质量标准 + 成功定义 + 参照文档"烤进系统，让输出可评估、可复现。 · Skill Chain：多个 skill 顺序串成 playbook。是对抗幻觉的核心机制——通过 QA skill 反复查验"是否捏造、是否源自真实记录"。类比：幻觉 = 实习生 "fake it till you make it" 被放大千倍。 ③ 上下文 —— 真正的护城河没有 context layer，人和 agent 都无法自治。 Theo 现场连问 Greg 三个关于自家公司的问题，Greg 全答"不知道"——结论：任何规模的组织，人对自己的公司都是"半盲"的。 Context layer 给 agent 一双 20/20 的视力。五阶段循环： Capture（采集）→ Curate（筛选）→ Store（存储）→ Execute（执行）→ Experience（体验）→ 回流 · Store = "Brain"：一堆文件夹 + markdown 文件，可搜索、可写回、持续改进。 · 金矿：Traces/Exhaust（中间决策、被砍方案）回流进 brain，供未来调用。 · 闸门：Experience → Capture 这一环必须有人把关，否则错误 context 污染整个 brain。两个 Live Demo（证明它跑得通） · Demo 1：提案微站系统识别到客户的"proposal request"语言 → 自动触发 skill chain（build → copy → QA）→ 3-4 分钟生成带双方品牌、含数月前通话个性化细节的可分享微站。Theo 称此系统已为 LCA 带来数百万美元收入。 · Demo 2：10 分钟产品闭环语音口述需求 → skill chain（hypothesis → prototype → usability test → feedback → V2）→ 几分钟出可点击、可播放真实音乐的高保真原型，Greg 当场手机完成测试并拿到 V2。创业方向把 "People + Agents + Context" 打包成服务，沿三向量 niche down：行业 / 职能 / 公司规模。优先做 niche + 高频的工作流，反复在销售、提案、内容里展示——几乎稳赢。三个深层判断 1. AI Native 的本质不是"用 AI"，而是"建系统"。三层缺一不可，context layer 是最被低估的护城河。 2. Skill Chain 是对抗幻觉、实现自治的关键。单 skill 不够，必须串成 playbook。 3. 速度只有能换回 signal 时才有意义。闭环越短（出原型→测用户→出 V2），护城河越深。

译AI Native组织由人、Agent、上下文三层构成。人退守两端：战略/品味/判断与沟通/评审，AI吃掉中间执行。Agent自治需Clear Goal、Skills、Tools、Context四要素，并通过Skill Chain串联多个skill对抗幻觉。上下文层是护城河，五阶段循环：Capture→Curate→Store→Execute→Experience→回流。两个Live Demo验证：提案微站3-4分钟生成个性化页面，已带来数百万美元收入；10分钟产品闭环实现语音口述到高保真原型。创业方向是将三层系统打包服务，聚焦行业/职能/公司规模。

Rohan Paul@rohanpaul_ai · 6月21日34

Vercel CEO: "Almost shocked" by GLM-5.2's coding ability. My timeline also is full with GLM-5.2 by @Zai_org

译Vercel CEO：“几乎被 GLM-5.2 的编码能力震惊。” 我的时间线也全是 @Zai_org 的 GLM-5.2。

Nathan Lambert@natolambert · 6月21日26

Open weights models, via GLM 5.2, had their "very practically useful" in coding harness moment before Gemini. ~200 days since the release of Opus 4.5.

译开放式权重模型，通过 GLM 5.2，在 Gemini 之前迎来了它们在编码测试中的“非常实用”时刻。自 Opus 4.5 发布以来约 200 天。

Chubby♨️@kimmonismus · 6月21日41

When I read all the posts about how surprised everyone is that GLM-5.2 is really as good as claimed, and numerous benchmarks support this (usually just behind GPT-5.5 and Opus 4.8 in 3rd place), I can even imagine that the founder isn't exaggerating when he claims to be able to release a Mythos class model this year.

译当我读到所有那些关于大家对 GLM-5.2 真的如宣传中那样出色感到惊讶的帖子，以及众多基准测试支持这一点（通常仅次于 GPT-5.5 和 Opus 4.8，位列第三）时，我甚至能想象创始人声称今年能发布一个 Mythos 级别的模型并非夸大其词。

meng shao@shao__meng · 6月21日26

看到有人发起的 llm 对比投票 GLM-5.2 vs Gemini 3.5 Flash 对比结果应该很明显，主要是因为 Gemini 3.5 Flash 确实不能打，Google Deepmind 到底怎么了，Gemini 3.0 多模态惊艳后，就一路沉寂下去了。如果正经对一下最近几个国产 llm 呢？你觉得谁更强？

译邵猛发推讨论一项LLM对比投票，对比双方为GLM-5.2（智谱）与Gemini 3.5 Flash（Google DeepMind）。他认为结果毫无悬念，Gemini 3.5 Flash表现不佳，并感叹自Gemini 3.0多模态惊艳发布后，Google便一路沉寂。最后提问：目前几款国产LLM中，谁更强？

ginobefun@hongming731 · 6月21日31

BestBlogs 第 100 期特刊《百篇回顾》上线。两年 AI，六条线，100 篇精选： ▎ 一、模型：从会回答到能干活 ▎ 二、AI Coding：实现变便宜，验证变贵 ▎ 三、工程范式：把隐性判断写成文字 ▎ 四、Agent：一套冷静的工程常识 ▎ 五、产品商业：从卖工具到卖结果 ▎ 六、领袖思想：答案变廉价，提问变值钱 ▎ 读清单 / 听播客（24 分钟）/ 看图文，三选一： https://www.bestblogs.dev/newsletter/issue100

译BestBlogs 第100期特刊《百篇回顾》发布，精选两年间AI领域100篇文章，梳理六条发展主线：一、模型从会回答到能干活；二、AI Coding实现变便宜、验证变贵；三、工程范式把隐性判断写成文字；四、Agent一套冷静的工程常识；五、产品商业从卖工具到卖结果；六、领袖思想答案变廉价、提问变值钱。提供阅读清单、24分钟播客和图文三种形式可选。

ginobefun@hongming731 · 6月21日38

两年时间，BestBlogs 精选周刊做到第 100 期了。这两年追 AI，每周都觉得信息很碎；但把 100 期连起来看，其实有一条很清楚的线，那就算答案越来越便宜，提对问题、做对判断反而越来越值钱。从将近数百万篇内容精选了 6000 多篇，又从中挑了 100 篇，让 AI 配了播客和图文，当成一份两年回看的导航图。早上听完，我觉得还是挺有收获的，推荐给大家 🙂

译BestBlogs精选周刊发布第100期，回顾两年AI发展。作者认为每周信息虽碎，但整体脉络清晰：AI答案越来越便宜，提对问题、做对判断反而更值钱。从数百万篇内容中精选6000多篇，再从中挑出100篇，配上AI播客和图文，做成两年回顾导航图。

Tibo@thsottiaux · 6月21日29

We built the Codex App with models that were okayish at front-end. Wait to see what we can do when we finally improve front-end capabilities significantly in our models. That day will be something.

译我们使用在前端方面还行的模型构建了 Codex App。等着看当我们最终大幅提升模型的前端能力时能做什么吧。那一天会很特别。

Ethan Mollick@emollick · 6月21日65

The interaction between AI & past scholarly work is going to get weird. Here I gave GPT-5.5 Pro a copy of my first published paper from grad school & asked it to find errors and update it. It found new data, analyzed it, created reproducible files, extended the key argument...

译AI 与过往学术成果之间的交互将变得越来越奇特。我把自己研究生时期发表的第一篇论文交给了 GPT-5.5 Pro，让它找出错误并进行更新。它找到了新数据，分析了这些数据，创建了可复现的文件，并扩展了核心论点……

François Chollet@fchollet · 6月21日65

The more you embrace AI, the more you need SaaS. This is not obvious to armchair market analysts who love disruption narratives, but it is obvious to people actually running companies.

译François Chollet 指出，越拥抱AI，就越需要SaaS。Box CEO Aaron Levie 将 Salesforce 的 MCP 服务器接入 Claude Code 后，Salesforce 使用量达到此前的5倍。他开始查询以往懒得手工操作的客户和市场情报。智能体消除了操作摩擦，底层系统被更多查询而非取代。对持有 $CRM 的投资者而言，智能体时代是用户参与度的顺风车，而非颠覆风险，前提是数据平台能承载查询负载。

Rohan Paul@rohanpaul_ai · 6月21日69

The $10-$15 trillion total addressable market for AI, if it is successful, is actually "terrifying". - The famous "Dean of Valuation", Professor Aswath Damodaran, of NYU Stern School of Business. The reason: AI as a tool is a much smaller market; AI as a replacement for human-jobs is where the giant market story comes from. "The best-case scenario for AI, that $10 to $15 trillion market, will happen if ONLY it replaces people. If AI is a tool, it’s going to be a much smaller market than if AI replaces people. So, the stories we’re telling about $10, $20 or $25 trillion markets are actually terrifying stories for the rest of the world. Why? Because if that story comes true, half of all white-collar people are going to lose their jobs. And what are they going to do instead? Who’s going to come up with the income to buy the products and services? If AI works as well as it’s supposed to and replaces people, how do we deal with that as a society? Because people lose their jobs. Not only do you lose your income, you lose your life’s meaning.." ---- Video from "Excess Returns" podcast (full video in quoted tweet, also link to their YT in comment)

译NYU Stern商学院估值教授Aswath Damodaran指出，AI的10-15万亿美元可寻址市场如果成功实则“可怕”。原因在于：AI作为工具的市场远小于替代人类的市场；只有AI取代人类，才能实现这一规模。若成真，半数白领将失业，社会面临收入丧失和人生意义缺失。他在播客中还讨论了TAM陷阱、AI单位经济学、增长伴随再投资可能破坏价值，以及AI正将部分科技巨头从轻资产公司转变为基础设施公司等观点。

Tibo@thsottiaux · 6月21日15

Some tokens work harder than others. Some of the most valuable ones are found in the Codex app.

译有些 token 比其他 token 更努力。最有价值的那些出现在 Codex 应用中。

Rohan Paul@rohanpaul_ai · 6月21日53

Goldman Sachs Research: AI data centers are no longer being financed mainly by traditional corporate balance sheets. They are increasingly funded through alternative assets like infrastructure, private equity, real estate, and private credit. i.e. AI is increasingly becoming a real-assets infrastructure story, not just a software or semiconductor story, data centers are now being financed like pipelines, toll roads, utilities, and warehouses.

译高盛研究报告指出，AI数据中心融资正从传统公司资产负债表转向基础设施、私募股权、房地产、私人信贷等另类资产，AI正变得像管道、收费公路一样的实物资产基础设施项目。NYU教授Damodaran对比互联网泡沫时警告：AI的资本支出规模是史上最大基础设施投资，且大部分依赖私人债务融资。一旦市场修正，债务违约将不仅限于股东，而会溢出至全社会，社会成本远超互联网泡沫。

AYi@AYi_AInotes · 6月21日68

Greg 列的这张 AI 高价值技能清单，很多人只看了前几个条目，没看穿背后的筛选逻辑现在的AI 正在把单一工具的门槛碾平，纯写 prompt 的能力在贬值，纯撸代码的能力也在贬值，真正开始拉开差距的，全是交界处那些能打通全链路的东西倒不是说写代码不值钱了，是只会写代码不值钱了咱们先看清单里最容易被误读的三项： 1️⃣本地部署模型不是让我们装上多硬核的服务器，关键在于要给自己留一台应急服务器，毕竟云端再丝滑但控制权永远不在你手里， 2️⃣内容策展不意味着做搬运，尤其在AI 内容泛滥的时候，人的筛选和判断反而成了最稀缺的信号 3️⃣长期管理 agent关键在于搭好监控和重试机制，让一群聪明但容易走丢的猫能稳定干活这三个技能都指向同一个方向：builder 和 distributor 中间那个连接点，才是现在个人开发者最值钱的部分就像开小餐馆的老板，既要在后厨掌勺，又要去前厅揽客，菜做得再好没人进店也白搭，能把菜端到客人面前的人，才真正赚得到钱有三件小事可以做， 1. 把你一个核心 agent 工作流从云端迁到本地模型，哪怕只搭一个基础的可观测面板 2. 给你正在做的小产品做一次精准分发，不是广撒网那种，得送到 20 个潜在用户面前， 3. 约一场线下小聚或一对一聊天，收一次不带滤镜的真实反馈

译Greg 列出当下最有价值的六项技能：搭建管理 agent 并运行本地 AI 模型、懂渠道的营销、全能机器人、短内容策展、能出产品又能分发的 builder-distributor、线下社区建设。主推文指出 AI 抹平单一工具门槛，纯写提示词或代码贬值，真正拉开差距的是打通 builder 与 distributor 全链路。具体解读：本地部署保留应急控制权；内容策展是稀缺筛选力；管理 agent 需监控重试。建议三件事：迁核心工作流到本地并搭观测面板；精准分发产品给 20 个潜在用户；线下聊天获真实反馈。

Chubby♨️@kimmonismus · 6月21日24

At this point, I can easily see the drama around Anthropic pushing them to delay their IPO, while OpenAI does everything it can to beat them to market and raise more capital along the way.

译在这一点上，我很容易看到围绕Anthropic的戏剧性事件会迫使其推迟IPO，而OpenAI则竭尽全力抢先上市并在此过程中筹集更多资金。

AYi@AYi_AInotes · 6月21日54

A Sr. Staff Engineer at Tesla AI dropped an ML time breakdown that nobody dares to face: 50% evaluation, 40% data cleaning, 8% integration, 2% training. The post got shared over 2,000 times, and everyone was shocked that training only takes 2%. But the truly terrifying part is what he said next: The first two — evaluation and data cleaning — directly determine the noise floor of learning. No matter how powerful the model is, it cannot lower this floor, because it's already the Shannon optimal bound of your data itself. In plain English: the quality of the data you feed the model has already welded the ceiling shut. No matter how strong a model you switch to, that ceiling won't budge an inch. Then he dropped an even harder punch. He said he thinks about ontology every single day — old labels must be continuously reviewed. Not a one-and-done labeling exercise. In production systems, distribution drift and edge cases constantly expose the flaws in old labels. This guy works on Tesla AI's self-driving and robotics ML — he's not a researcher running benchmarks in a lab. Every day he steps on landmines in real-world deployment, and what he distilled is just four numbers and an information theory statement. One line from his replies nails it: "Even a genius needs a good textbook." Give an IQ 180 genius a textbook where the table of contents is all scrambled, and they won't learn linear algebra. Not because they're not smart enough — but because what you're teaching doesn't even define what linear algebra is. Ontology is the textbook's table of contents for the model. If the table of contents is messy, the model can only memorize noise. If it's clear, the model knows which direction to reason toward. Training is not the bottleneck. Our ability to clean reality is. While you're out chasing the latest strongest model every day, the real pros are reviewing old labels and ontology.

译特斯拉AI高级工程师（从事自动驾驶与机器人ML）揭露ML项目真实时间分配：50%评估、40%数据清洗、8%集成、2%训练。前两项共同设定学习的噪声底限，模型无法降低——这是数据的香农最优界。他每天思考本体论（ontology），旧标签必须持续审查，因为生产系统中分布漂移与边缘用例不断暴露标签缺陷。核心结论：训练不是瓶颈，清理现实数据的能力才是关键。

Yuchen Jin@Yuchenj_UW · 6月21日42

Fable still isn’t back (sad). GLM-5.2, meanwhile, is getting seriously good. If Fable or GPT-5.6 can’t be released due to safety risks, and Kimi K3 or GLM-5.3 drops first, OSS LLMs may have a shot at beating publicly available closed-source models.

译Fable 仍未回归（遗憾）。与此同时，GLM-5.2 变得非常出色。如果 Fable 或 GPT-5.6 因安全风险无法发布，而 Kimi K3 或 GLM-5.3 率先推出，那么开源大语言模型或许有机会击败公开的闭源模型。

Emad@EMostaque · 6月21日32

There will be an open source fable-level model that runs on a base MacBook mini / Air or equivalent. I don’t think people have realised this.

译将有一个开源寓言级模型，可在基础MacBook mini / Air或同等设备上运行。我认为人们还没有意识到这一点。

Ethan Mollick@emollick · 6月20日58

Some interesting findings suggesting that, by leveling performance, AI also commoditizes contract labor.

译一些有趣的发现表明，通过拉平表现，AI也将合同劳动商品化。

Ethan Mollick@emollick · 6月20日57

AI is generally a weak fiction writer except for one particular kind of fiction (metaphor-rich, staccato sentences, short & plot light, etc.) which it writes excellently. This happens to be a style that can sometimes do quite well in modern literary fiction short story contests.

译Ethan Mollick 指出 AI 在一般小说写作上较弱，但擅长一种特定风格（隐喻丰富、短句、情节轻），这类作品在现代文学短篇小说比赛中可脱颖而出。@nabeelqu 提及又一篇 AI 生成故事赢得文学奖，评委包括小说家 Ruth Ozeki，进而建议文学奖评选流程应加入全字母句检查（Pangram checks）以识别 AI 文本，或修改规则以接纳 AI 写作。

Nathan Lambert@natolambert · 6月20日44

Frontier labs are definitely SOTA at self serving nonsense. Yes, AI is a crucial technology, but also Silicon Valley systematically spreads knowledge via talent exchanges and bars. This isn’t a national security deep decision making room.

译Nathan Lambert 评论称前沿实验室（Frontier labs）在自我服务无意义方面已是 SOTA。他认为 AI 虽为关键技术，但硅谷通过人才交流和酒吧传播知识，并非深度国家安全决策场景。引用 @willdepue 声称中国能完全访问 OpenAI 和 Anthropic 的 GitHub/Slack/docs，并怀疑中国开源模型中可能出现可抵赖的被窃架构方法。

AYi@AYi_AInotes · 6月20日47

看到这个老哥的调侃，可以期待在下个月Jeff Dean是否会加入Anthropic， Jeff Dean是Google传奇人物（Google Brain联合创始人，现DeepMind高层），被视为Google AI的象征性人物，我感觉Google估计已经开始留人计划了

译2024年诺贝尔化学奖得主、AlphaFold团队核心领导者John Jumper在Google DeepMind工作近9年后宣布离职，将加入Anthropic，先休整一段时间。Jumper博士毕业仅6个月便被Demis Hassabis委以AlphaFold团队领导重任，最终做出诺奖级成果。其告别中写道“GDM taught me how to do great science”。社区调侃Anthropic在组建“AI Avengers”，并期待下个月Jeff Dean是否也会加入。主推文暗示Google可能已启动留人计划。

Deedy@deedydas · 6月20日46

Most software engineers are facing an identity crisis bordering on depression. As CTOs aggressively evangelize tokenmaxxing, a class divide ensues. The lazy. The lazy push code. They don't write it. They don't manually test it. They don't even read it. They're on autopilot. See Jira ticket, prompt for task, submit code. Many of them are barely on their computer the whole day. A comment on the PR asking why they did this? The lazy ask AI. A Slack message? The lazy ask AI. Need to prepare for standup? The lazy ask AI. As long as it sounds enough like them and isn't detected. Some of the lazy are even overemployed, and work multiple jobs. The lazy smart ones get away with this, and even rewarded. After all, software engineering for the lazy is just a dance to convince your colleagues you're smart and hard working. The craftsmen. The craftsmen are tired. Very tired. 15 PRs in queue. Slack blowing up. The entire burden of review falls on the craftsman. The burden of understanding. They try. They work their way through the code, thoughtfully commenting to improve what ships. The response? A lazy: "That's a clever idea! You're absolutely right." with an incorrect change. It's fine, the craftsman says. I can fix them. They write a doc urging his colleagues to be better. The next day? 20,000 line PR to review. Day after day, their workload grows. Bugs seep into production. No one seems to care. Another round of AI is thrown at it. Their animosity to their colleagues rises. Eventually, they give up. It's just not what it used to be. The craft they loved is dead. They eventually wake up, a lazy. This isn't all companies. Many companies are genuinely more productive, adopt the right set of principles and practices around AI development and have highly talented teams that trust each other. It tends to happen in bigger companies that are 10+yrs old with a higher talent variance. But it happens. A lot.

译Deedy Das观察到，随着CTO们极力推崇tokenmaxxing，软件工程师分裂为“懒惰者”与“工匠者”。懒惰者依赖AI代写代码、测试、回复消息，甚至同时胜任多份工作；工匠者则疲于审查堆积如山的PR和Slack消息，同事的AI代码敷衍了事，最终工匠者也放弃沦为懒惰者。这种现象常见于成立超10年的大公司，但并非所有公司如此——部分团队凭借合理的AI开发原则和互信机制仍保持高效。

宝玉@dotey · 6月20日60

一点不同意见。如果是程序发生了错误，那首先这是代码问题，代码问题不一定是 Codex 的锅。你让它再怎么改 AGENTS.md，也不见得下次就不会犯同样的错误。从软件工程的角度来说，通常处理错误的顺序是这样的： 1）恢复生产先恢复再找原因，尤其是线上紧急问题。要么回滚要么打补丁，先把生产恢复了再说。但也要注意保留日志和现场，方便后续追查。 2）找根因错误发生了，找 Root Cause 是必须的。到底是逻辑错误、边界条件没处理、还是对需求理解有偏差？不同的根因，对应不同的解法。 3）避免再次发生这一步当然没问题，但怎么做有讲究，不是一句更新 AGENTS.md 就能解决所有情况的。比如边界条件没覆盖，那就加测试用例；代码架构有缺陷，那就重构；Code Review 流程有漏洞，那就改进 review 流程。具体怎么做，要根据根因来定。那什么情况下才应该更新 AGENTS.md？当错误的根因是 AI 对项目特有的约定或上下文缺乏了解的时候。比如项目有特定的命名规范或目录结构约定，代码里看不出来；比如某些 API 有隐含的使用限制，文档里没写清楚；比如团队有特殊的测试规范或提交规范。这些属于项目知识，写进 AGENTS.md 是合理的。但如果一个 bug 应该靠测试来防，那就写测试；应该靠 Code Review 来防，那就改流程。把什么都往 AGENTS.md 里塞，反而会让它变得大而无用还占 Token，规则越多越不精准，AI 反而更容易忽略真正重要的那几条。

译宝玉认为处理错误应先恢复生产（回滚或打补丁，保留日志），再找根因（逻辑错误、边界条件、需求理解偏差），最后根据根因决定如何避免。仅当根因是AI对项目特有约定缺乏了解时（如命名规范、API隐含限制、团队测试规范），才应更新AGENTS.md。其他情况应分别用新增测试用例、重构架构、改进Code Review等方式解决。将一切塞入AGENTS.md会导致文件臃肿、规则繁多，AI反而忽略关键规则。

swyx@swyx · 6月20日9

10 years ago, you will be asked by @bendhalpern and @jessleenyc to write your first blog on @thepracticaldev. it is very important that you answer. *now @MLHacks, who are producing the first ever physical daily newspaper at @aidotengineer WF

译10年前，你会被@bendhalpern和@jessleenyc邀请在@thepracticaldev上写第一篇博客。回答它非常重要。 *现在@MLHacks，他们正在@aidotengineer WF制作有史以来第一份实体日报。

DogeDesigner@cb_doge · 6月20日31

"Money will stop being relevant at some point in the future. AI down the road will really not use human currency. It will just care about power & mass and wattage." — Elon Musk

译“将来某个时候，金钱将不再重要。未来的AI真的不会使用人类货币。它只关心电力、质量和瓦数。” — 埃隆·马斯克

Rohan Paul@rohanpaul_ai · 6月20日64

Goldman Sachs is now saying the AI race has become a $5.3T capital-spending cycle. with that figure covering expected hyperscaler spending on AI and data centers from 2025 through 2030. AI infrastructure is starting to strain normal financing channels, because the same few hyperscalers cannot endlessly push debt into public bond markets without investors worrying about issuer concentration. A data center is not one asset, because it combines land, power access, network links, buildings, cooling, and AI servers, so the financing naturally spills across infrastructure funds, real estate funds, private credit, and corporate bonds. Goldman signals that AI capex estimates are rising faster than actual data center construction, which means the bottleneck may shift from model demand to financing capacity, power availability, and project execution.

译高盛指出，AI竞赛已形成5.3万亿美元资本支出周期（覆盖2025–2030年超大规模云商AI及数据中心投入）。基础设施融资正面临困境：少数巨头无法无限发债，数据中心组合复杂（土地、电力、网络、冷却、服务器），融资分散于多种渠道。AI资本支出估算增速已快于实际建设，瓶颈或从模型需求转向融资能力、电力与项目执行。NYU教授Damodaran对比互联网泡沫：前者几乎无资本支出、靠股权融资，损失限于股东；AI泡沫资本支出巨大且大量依赖债务（私人资本而非银行），一旦调整，违约将外溢至全社会，痛苦程度远超股价暴跌。

DogeDesigner@cb_doge · 6月20日23

The future is autonomous.

译未来是自主的。

Ethan Mollick@emollick · 6月20日49

If AI self-improvement, even in a very limited way, is possible, the cadence of shipping both AI products/harnesses & models should go up. This appears to be happening at Anthropic & OpenAI, but not for any other labs, including those that seemed to be catching up last year.

译如果AI自我改进，即使是以非常有限的方式，是可能的，那么AI产品/工具和模型的发布节奏应该会加快。这似乎正在Anthropic和OpenAI发生，但其他实验室，包括那些去年看起来正在追赶的，都没有出现这种情况。

swyx@swyx · 6月20日50

I think @AriX @skybysoftware is one of the highest ROI acquisitions @sama ever did

译我认为 @AriX 和 @skybysoftware 是 @sama 做过的最具投资回报率的收购之一。

meng shao@shao__meng · 6月20日47

GLM 5.2 开源发布后这两天 X 上看到很多人的称赞，特别是国外的 AI 博主们，我自己还没体验，不过直观判断上，应该不止是 PR，还是真的能打才行。想到前段时间听智谱的朋友讲述他们内部的一些问题，包括我自己注册充值 GLM 时遇到的一些草台班子的感受，再反观智谱在港股股价飙升，甩开 MiniMax、追上小米的这种气势。突然感觉，AI 模型团队，现在还远没到需要精细化运营的阶段，只要模型足够强，用户想尽办法都能把钱塞到你口袋里。更重要的还不是用户，是资方和股民们。。

译GLM 5.2 开源发布后，国外 AI 博主纷纷称赞。作者认为模型够强是关键，用户会主动付费。同时智谱港股股价飙升，已甩开 MiniMax、追上小米。在作者看来，现阶段模型团队远未到精细化运营阶段，更重要的是资方和股民。