some new brand concepts we're playing with @every wdyt
我们正在尝试的一些新品牌概念，@every 你觉得怎么样
DS
Dan Shipper
@danshipper ↗ceo @every | the only subscription you need to stay at the edge of AI
3 最新125 累计51 期
每条推文 hover 显示单独 ▶
omfg you can read terrence tao's chatgpt conversation about the jacobian polynomial
天啊，你居然可以读到 terrence tao 和 ChatGPT 关于 jacobian polynomial 的对话
[1]
[1]
we’re hiring a senior engineer to help work on our agent @every must love agents. dm if you’re interested (for now preferencing for people I know on here!)
我们正在招聘一位 senior engineer，来帮助参与我们在 agent @every 上的工作——一定得热爱 agents。如果你感兴趣请 dm 我（目前会优先考虑我在这里认识的人！）
Yes it just passed some barrier Have been able to automatically do ~70% of the copy edits we would ordinarily do by hand internally @every for the last week First time that’s happened and I’ve been trying it for years
是的，它刚刚跨过了某个门槛。过去一周里，在 @every，我们原本通常需要内部手动完成的文稿润色修改（copy edits），现在大约有 70% 已经能够自动完成了。这是第一次发生这种情况，而我这些年来一直都在尝试。
we will vibe check Kimi K3 but i am extraordinarily skeptical of claims it’s as good as fable
我们会来 vibe check 一下 Kimi K3，但我对“它和 fable 一样好”这种说法极其怀疑。
not surprising that @OpenAI is firing on all cylinders right now, and it’s an unbelievably interesting story - they launched GPT-5 summer of 2025 and positioned it as a pair programmer. we wrote at the time @every that they completely missed the new agentic coding that was starting to happen inside of Claude Code. they bet on agentic coding in the browser / vms and vibe coding in ChatGPT but it was too early - a small team broke off and began working on a separate Codex model line and product. didn’t have to serve the gigantic customer base of ChatGPT, and by November / December of 2025 with 5.3 it was clear they were starting to make something good and the progress was fast - Codex Desktop app launched in Feb and was just clearly superior. there’s a weird late comer advantage sometimes in AI because you get to skip to what works instead of your product having the scars of new capability improvements being bolted on every 3 months - Codex started getting popular it was clear it was a thing, and needed to get merged back in. Which they did, imo quite well—a super complicated thing that would’ve been very easy to screw up. most companies try to disrupt themselves and fail. OpenAI somehow figured out how to disrupt their main product, and then merge it back in seamlessly. incredible aura
@OpenAI 现在火力全开，这一点并不令人意外，而且这整个故事也精彩得难以置信——他们在 2025 年夏天发布了 GPT-5，并把它定位成一个 pair programmer（结对程序员）。我们当时在 @every 写过，他们完全错过了当时正在 Claude Code 内部开始出现的新型 agentic coding（agent 驱动编程）。他们把赌注押在了浏览器 / vms 里的 agentic coding，以及 ChatGPT 里的 vibe coding 上，但那时候还太早——后来有一个小团队分拆出来，开始做一条独立的 Codex model line 和产品线。不必去服务 ChatGPT 那个庞大无比的客户群，到 2025 年 11 月 / 12 月，随着 5.3 发布，已经很明显他们开始做出真正不错的东西了，而且进展非常快——Codex Desktop app 在 2 月上线，显然就是更强。有时候 AI 领域会有一种奇怪的后来者优势，因为你可以直接跳到真正有效的方案，而不是让自己的产品每隔 3 个月就把新能力硬拧上去，留下满身伤疤——随着 Codex 开始流行起来，很明显它成气候了，也需要被重新合并回去。而他们确实做到了，依我看还做得相当好——这是一件极其复杂、却也极其容易搞砸的事。大多数公司都试图颠覆自己，结果失败了。OpenAI 不知怎么竟然做到了先颠覆自己的主产品，然后再把它无缝合并回去。气场惊人。
look at those emoji reactions this is how we win
看看这些 emoji 反应，这就是我们取胜的方式。
early bird just ended and DAMN people love the builder pack! sign up:
early bird 刚刚结束，DAMN，大家太爱 builder pack 了！报名：
Even when things are going great, running a $1.5 billion AI startup is a knife fight. @MeetGranola was one of the first AI apps of this generation to achieve near-ubiquitous adoption. But meeting notes are not the company’s be-all and end-all. The real battle is over owning the interface that everyone uses to get their work done in an AI-native world. I had Chris Pedregal (@cjpedregal), cofounder and CEO of Granola, back on @every’s AI & I to talk about the current state of the application layer, AI’s frontier, and the future of work. We get into: - Why meeting notes clones don’t matter. Three big companies cloned Granola’s core feature. To him, meeting notes were never the real prize. “Easy come, easy go” is his view of anyone’s lead, including his own. - How he thinks about building proactive features in AI. Granola pre-generates millions of pre-meeting briefs, which include context on the nature of the meeting and people participating, that most people never open. But when they do, they have a magical experience. - Why Granola is betting on “bring your own agent.” Chris says the API and MCP will get “a lot better” over the next few months, and we talk about their agent-native strategy and why they’ve pushed the product that way. This is a must-watch for anyone building at the application layer. Watch the episode! Timestamps Introduction: 00:00:59 Why running a company is a knife fight even when it’s working: 00:01:57 Granola’s counterintuitive view on competition: 00:04:33 Dan’s “pirate and architect” model for early-stage product teams: 00:10:44 Granola’s “shaping” and “validation” phases for building features: 00:13:09 Why Dan lives almost entirely inside Codex: 00:18:17 The case for “Codex-native apps”: 00:24:40 Granola’s “handrail” philosophy: 00:35:37 Why Granola is going all in on winning meeting-adjacent context: 00:38:12 What a transcript alone can never capture: 00:44:19
即使一切进展顺利，经营一家估值 $1.5 billion 的 AI 初创公司也像是一场刀刀见血的近身肉搏。@MeetGranola 是这一代 AI 应用里最早实现近乎无处不在普及的产品之一。但 meeting notes（会议笔记）并不是这家公司的全部，也不是终点。真正的战场，是在一个 AI-native（AI 原生）的世界里，谁能拥有那个人人都会用来完成工作的 interface（界面）。我请来了 Granola 的联合创始人兼 CEO Chris Pedregal（@cjpedregal），再次做客 @every 的 AI & I，一起聊 application layer（应用层）的现状、AI 的 frontier（前沿），以及工作的未来。我们聊到了：- 为什么 meeting notes 的克隆产品并不重要。三家大公司都克隆了 Granola 的核心功能。在他看来，meeting notes 从来都不是真正的奖品。无论是谁的领先地位——包括他自己的——他的看法都是“来得容易，去得也容易”。- 他如何看待在 AI 中构建 proactive（主动式）功能。Granola 会预先生成为数百万场会议准备的会前简报，其中包含会议性质以及参与者的相关背景信息，而大多数人其实从来不会点开看。但一旦他们看了，就会获得一种神奇的体验。- 为什么 Granola 正在押注 “bring your own agent（自带 agent）”。Chris 表示，API 和 MCP 在接下来的几个月里会“好很多”，我们也谈到了他们以 agent-native（agent 原生）为核心的战略，以及他们为什么把产品往这个方向推进。对于任何正在 application layer 构建产品的人来说，这一期都必看。去看这期节目吧！时间戳 Introduction：00:00:59 即使公司运转良好，为什么经营公司仍像一场刀战：00:01:57 Granola 对竞争的反直觉看法：00:04:33 Dan 针对早期产品团队提出的“pirate and architect”模型：00:10:44 Granola 构建功能时的 “shaping” 和 “validation” 阶段：00:13:09 为什么 Dan 几乎完全活在 Codex 里：00:18:17 “Codex-native apps” 的成立理由：00:24:40 Granola 的 “handrail” 哲学：00:35:37 为什么 Granola 正全力押注赢下与会议相邻的上下文：00:38:12 为什么单靠 transcript（转录文本）永远无法捕捉某些东西：00:44:19
receipts, from our Codex Desktop app launch vibe check:
这是来自我们 Codex Desktop app 发布后“氛围检查”的战绩截图：
if you are reading @every, you knew codex was about to take off 6 months ago :) success kid pointing to when i started getting extremely, absolutely, annoyingly codex pilled and we began covering it
如果你在读 @every，那你 6 个月前就该知道 codex 要起飞了 :) “success kid” 指向的是我开始变得极度、彻底、烦人地被 codex pilled（被 codex 完全说服/洗脑），而我们也开始持续报道它的那个时间点
COME SEE @EVERY IRL! We're hosting an @every subscriber meetup at our brownstone in Brooklyn TOMORROW Wednesday, July 15 from 6-8 PM You should come!
来线下见 @EVERY 吧！我们将在明天，也就是 7 月 15 日星期三晚 6 点到 8 点，在我们位于 Brooklyn 的 brownstone 举办一场 @every 订阅者见面会。你应该来！
we need husk doing vibe checks @every
我们需要 husk 对 @every 做做 vibe checks（氛围检查）
so this implies that...developers don't do work? lol
所以这是不是意味着……developers 不干活？lol
GPT-5.6 SOL: THE GOLD STANDARD FOR KNOWLEDGE WORK
GPT-5.6 SOL：知识型工作的黄金标准
BREAKING: OpenAI announced GPT-5.6 Sol! As of today, by U.S. government directive, access is limited to only ~20 pre-approved companies and @every is not on the list. This appears to be a temporary situation while the government races to figure out a long-term policy for releasing frontier models with advanced capabilities. I understand and applaud the need for some government oversight in making American infrastructure resilient to cyberattacks and other potential new threats from the misuse of these models. However, I also strongly believe that widespread democratic access to frontier models is absolutely necessary to our country’s leading position in the AI race. It’s also critical for allowing American workers to keep pace with the new skills they need to be productive in this era. A world where advanced models are locked up only for use by the employees of AI giants and a select few companies is one where ambitious students, independent builders, and working professionals are denied the tools they need to learn, create, and compete to their fullest potential. Also, speaking for @every and the community of developers and writers who share early access with us: Our job is to test these tools early so that we can prepare Americans for how to use them to their fullest extent. If we lose access, we lose the ability to do that important job well. Thankfully, both OpenAI and the government seem to be working to make broad access available soon. We’ll be ready to vibe check when that happens :)
突发消息：OpenAI 宣布了 GPT-5.6 Sol！截至今天，根据 U.S. government 的指令，访问权限仅限于约 20 家预先批准的公司，而 @every 不在名单之列。这看来是一个临时情况，因为政府正争分夺秒地制定一项关于如何发布具备高级能力的 frontier models（前沿模型）的长期政策。我理解并赞赏政府监督的必要性，以增强 American 基础设施抵御 cyberattacks（网络攻击）以及这些模型被滥用可能带来的其他新威胁的韧性。然而，我也强烈相信，frontier models 的广泛民主化访问权，对于我们国家在 AI 竞赛中保持领先地位是绝对必要的。这对于让 American workers 跟上这个时代所需的新技能、保持生产力，也同样至关重要。一个只有 AI 巨头员工和少数被选中的公司才能使用 advanced models（先进模型）的世界，意味着有抱负的学生、独立建设者以及在职专业人士，将被剥夺他们充分学习、创造和竞争所需的工具。还有，代表 @every 以及那些与我们共享 early access（早期访问权限）的开发者和写作者社区发言：我们的工作，就是尽早测试这些工具，以便我们能够帮助 Americans 为如何最大程度地使用它们做好准备。如果我们失去访问权限，我们也就失去了把这项重要工作做好的能力。值得庆幸的是，OpenAI 和政府似乎都在努力，争取尽快提供广泛访问。等那发生时，我们会准备好来做个 vibe check :)
One of the smartest people i know, ex-One Medical, ex-Bridgewater founder backed by a16z is building a stealth NYC health-tech company. He's doing AI-first medicine with physical clinics Hiring staff+ engineers, physicians who ship product, and a Codex-native Chief of Staff/BizOps lead You should email him if you want to work on something like this: every@rungravity.com
我认识的最聪明的人之一，曾在 One Medical 任职、Bridgewater 的前创始团队成员，且获得 a16z 支持，正在 NYC 打造一家处于 stealth（隐身开发）阶段的 health-tech 公司。他在做 AI-first（AI 优先）的医疗模式，并配套实体诊所。正在招聘 staff+ 级别工程师、能交付产品的 physicians（医生），以及一位熟悉 Codex-native 的 Chief of Staff/BizOps 负责人。如果你想做这类事情，应该给他发邮件：every@rungravity.com
Are we hurtling toward a future where AI can do everything humans can? Edwin Chen (@echen) believes we might be. He’s the CEO of Surge AI, one of the largest providers of expert data for frontier labs. Surge passed over $1 billion in revenue without raising any outside capital, and that gives Edwin a unique perspective on how quickly AI progress is accelerating. I’m on the record arguing that AI automation actually creates more human work. I also believe that even though AI progress is accelerating exponentially, we’re much farther away from AI replacing humans than it might seem. That’s why I had Edwin on @every’s AI & I. We batted around different visions of the future, and discussed whether humanity will retain its unique place in the universe, and what that might be. We get into: • If Chen’s version of the future materializes, he’s worried it’ll make people stop trying. One answer comes from a short story by science fiction writer Ted Chiang: Behave as if your decisions matter, even when you know they don’t. • AI may soon be able to take a nebulous goal like “win a Fields Medal” and execute. What it can’t do, I argue, is set its own goals—LLMs have no intrinsic motivation, no drive to explore, no ability to just change their mind. • A model optimized for engagement doesn’t provide the most valuable user experience. Edwin spent 20 rounds polishing a pointless email with one model before Claude told him to just send it. • Why AI is still bad at writing: models learn to hack the metrics they're trained on. Edwin's Hemingway Bench found models outputting a metaphor in every single sentence, an overindexxing that makes for a terrible reading experience. This is a must-watch for anyone interested in where we fit as models get more capable. Watch below! Timestamps 1. Introduction: 00:00:54 2. Surge as a "school for AGI": 00:01:49 3. What AI's capacity for novel mathematics says about human achievement: 00:04:46 4. Motivation in an era when AI can do everything: 00:07:29 5. The trap of optimizing AI models for engagement: 00:14:34 6. Training using datasets versus training using environments: 00:29:34 7. The value of personal data: 00:35:09 8. Why models are bad at writing: 00:39:40 9. Chen's AGI timeline: 00:42:00
我们是否正高速冲向这样一个未来：AI 能做到人类能做的一切？Edwin Chen（@echen）认为，也许是的。他是 Surge AI 的 CEO，Surge AI 是 frontier labs 的最大专家数据提供商之一。Surge 在没有募集任何外部资本的情况下，营收已超过 10 亿美元，这让 Edwin 对 AI 进展加速的速度拥有独特视角。我一直公开主张，AI 自动化实际上会创造更多需要人类来做的工作。我也相信，尽管 AI 的进展正以指数级加速，但 AI 距离取代人类仍比看上去要远得多。这就是为什么我请 Edwin 来做 @every 的 AI & I 节目。我们来回碰撞了多种不同的未来图景，并讨论了人类是否会保住自己在宇宙中的独特位置，以及那种位置可能是什么。我们谈到了：• 如果 Chen 所设想的未来真的实现了，他担心这会让人们不再努力。一个答案来自科幻作家 Ted Chiang 的一篇短篇小说：即使你知道自己的决定并不重要，也要表现得好像它们很重要。• AI 也许很快就能接过一个像“赢得 Fields Medal”这样模糊的目标并加以执行。但我认为，它做不到的是为自己设定目标——LLM（大语言模型）没有内在动机，没有探索驱动力，也没有单纯“改变主意”的能力。• 一个为 engagement（用户参与度）优化的模型，并不能提供最有价值的用户体验。Edwin 曾用某个模型来回打磨一封无关紧要的邮件 20 轮，最后 Claude 告诉他，直接发出去就行。• 为什么 AI 仍然不擅长写作：模型会学着“钻”它们所依据的训练指标的空子。Edwin 的 Hemingway Bench 发现，模型会在每一句话里都输出一个隐喻，这种 overindexing（过度偏重）会带来非常糟糕的阅读体验。对于任何关心在模型能力越来越强时，我们将处于什么位置的人来说，这期内容都不容错过。请看下方！时间戳 1. 介绍：00:00:54 2. Surge 作为“AGI 的学校”：00:01:49 3. AI 在新颖数学上的能力说明了人类成就的什么：00:04:46 4. 当 AI 能做一切时，这个时代的动机问题：00:07:29 5. 为 engagement 优化 AI 模型的陷阱：00:14:34 6. 使用 datasets（数据集）训练 versus 使用 environments（环境）训练：00:29:34 7. 个人数据的价值：00:35:09 8. 为什么模型不擅长写作：00:39:40 9. Chen 对 AGI 时间线的判断：00:42:00
Extreme time deflation happening right now We expect a step change in productivity whenever Fable is unlocked, might as well go on vacation until then
现在正在发生极端的时间通缩（time deflation）；我们预计，一旦 Fable 被解锁，生产力就会出现一次台阶式跃升，所以在那之前不如先去度假
we are hiring a managing editor @every! if you are extremely meticulous, can keep the trains running on time, hate split infinitives, love oxford commas, want to take back the em dash, and are AI-pilled inquire here:
我们正在 @every! 招聘一名 managing editor（执行编辑）！如果你极其一丝不苟，能让一切像准点运行的列车一样顺畅，讨厌 split infinitives（分裂不定式），热爱 oxford commas（牛津逗号），想把 em dash（长破折号）重新夺回来，而且是 AI-pilled（坚定看好 AI 的），请在这里咨询：
@every @ninklefitz i wrote "against explanations" in 2023 about how AI might change the sciences...extremely pumped to see the prospect of so much progress so fast
@every @ninklefitz 我在 2023 年写过《against explanations》，讨论 AI 可能如何改变科学……想到这么快就有望出现如此巨大的进展，真的让我非常兴奋
we @every make investments every so often in founders building things in AI who we think are excellent @ninklefitz is one of those founders—and i am also a huge fan of tacit's mission and approach
我们 @every 会不时投资那些在 AI 领域打造产品、且我们认为非常优秀的创始人，@ninklefitz 就是其中之一——而且我也非常认同 tacit 的 mission（使命）和 approach（方法）
@hammer_mt @github @kdaigle YouTube: Spotify:
@hammer_mt @github @kdaigle YouTube: Spotify:
just switched off of Atlas Browser and back onto Dia i've been having so many weird bugs with atlas, and doesn't feel like it's improving
我刚从 Atlas Browser 切回 Dia；我在 atlas 上一直遇到很多奇怪的 bug，而且感觉它并没有在改善。
dario at the wellness retreat when they told him hegseth was on the phone
当他们告诉 Dario，Hegseth 打来电话时，Dario 正在 wellness retreat。
before and after fable ban: my claude app vs. codex app usage
Fable ban 之前和之后：我在 Claude app 与 Codex app 上的使用情况。
the last thing you see before andy jassy gets your model banned by the usg
在 Andy Jassy 让 USG 封禁你的 model 之前，你最后看到的东西。
harry potter fan fiction and hillbilly elegy brought us to this
是 Harry Potter 同人小说和 Hillbilly Elegy 把我们带到了这一步
this seriously messes up my weekend plans i may have to actually see people
这真把我周末计划彻底搞乱了，我可能真的得去见人了
had an idea for a big fable project, set it up, and let it cook came back an hour later and it had triggered the safeguards and fell back to 4.8 10 minutes in back to codex 😬
我当时有个想做大型 fable 项目的想法，就把它搭起来先让它自己跑着，结果一小时后回来一看，它触发了 safeguards（安全防护机制），并且在进行到 10 分钟时就回退到了 4.8，最后又回到了 codex 😬
I predicted this might happen on on @lennysan’s pod last year Higher productivity from each individual employee with AI, makes it appealing to reshore certain jobs back to the US to be close to customers
我去年在 @lennysan 的播客（pod）上就预测过这事可能会发生：AI 带来的每位员工生产力提升，会让把某些工作岗位重新迁回美国、以便更贴近客户，变得更有吸引力
FABLE (MYTHOS) is OUT! we've been testing for a week @every. here's our vibe check:
FABLE (MYTHOS) 已发布！我们已经在 @every 测试了一周。以下是我们的 vibe check：
my absolute favorite of Plato's Dialogues is a deep discussion of the limits of techne and the necessity of aidos and dike Protagoras here, in the way he talks about where knowledge comes from and whether virtue can be taught, pre-sages LLMs:
我最喜欢的 Plato《对话录》篇章，是一场对 techne（技艺/技术）之局限，以及 aidos 和 dike 之必要性的深刻讨论；这里的 Protagoras，在他谈论知识从何而来、以及美德是否可以被教授的方式上，预示了 LLMs：
to add two more that increase in value: Aidōs—reverence and responsiveness to others Dikē—the capacity to perceive what is right
再补充两个会提升其价值的概念：Aidōs——对他人的敬重与回应能力；Dikē——感知何为正确之事的能力
LLMs are not conscious. LLMs are not not conscious. Both true.
LLMs 没有意识。LLMs 也并非没有意识。两者都是真的。
we'll have the full workflow and prompts available to @every subscribers soon:
我们很快会向所有 @every 订阅用户提供完整的 workflow（工作流）和 prompts（提示词）：
@every @TrySpiral with your agent here:
@every @TrySpiral，与你的 agent 一起在这里试用：
NEW: Spiral 4.0—a writing partner for you and your agent by @every -> Stylometry: we built a new Style Engine based on the principles of stylometry to extract you and your brand's voice and produce great writing every time, based on examples of your past work -> MCP and CLI: Spiral is now built to be used by your agent like Codex, Claude Code, OpenClaw and more so you can get great writing automatically we use it every day internally to write landing pages, tweets, podcasts, marketing emails and more and to make sure it's ALL on-brand across our entire 30 person team @every
新发布：Spiral 4.0——一个为你和你的 agent 打造的写作伙伴，来自 @every -> Stylometry（文体计量学）：我们基于 stylometry 的原理构建了全新的 Style Engine，用来提取你本人以及你品牌的表达风格，并根据你过去作品的示例，每次都产出高质量写作 -> MCP 和 CLI：Spiral 现在被构建为可像 Codex、Claude Code、OpenClaw 等一样由你的 agent 使用，因此你可以自动获得出色的写作；我们每天都在内部使用它来撰写 landing pages、tweets、podcasts、marketing emails 等内容，并确保在我们整个 30 人团队中，这一切都完全符合品牌风格 @every
I bought Figma stock during the SaaSpocalypse panic. Talking to @mcolyer, Figma's director of product management for developers, made me wish I'd bought more. He joined me on @every's AI & I to make the case for a SaaS resurgence—and to explain why chat-based tools are the wrong interface for design. We get into: - Why running your own agents makes you more willing to pay for SaaS, not less - How Figma's MCP server allows you to approach design work from two directions: Take a live web page and reconstruct it on the Figma canvas, or hand a Figma design over to an agent so it can makes changes via pull request - The inherent limits of chat-based design, a format that isn’t equipped to generate lots of new ideas -Why review is the next bottleneck, and potential solutions for helping teams scale evaluations Watch below! Timestamps 0:01:03 - Introduction 0:02:15 - The SaaSpocalypse narrative has it backwards 0:05:27 - Matt’s email-agent origin story 0:13:21 - Divergent vs. convergent design thinking 0:17:39 - Figma’s MCP server 0:19:45 - Why design agents need personalization 0:22:09 - Every problem is a context problem 0:25:12 - Apple and Google as the reigning kings of context 0:28:18 - Why review is the new bottleneck
我是在 SaaSpocalypse 恐慌期间买入 Figma 股票的。和 @mcolyer——Figma 负责开发者方向产品管理的总监——聊过之后，我甚至希望自己当时买得更多。他来到了 @every 的 AI & I 节目，和我一起论证 SaaS 复兴的理由——并解释为什么基于聊天的工具并不是设计的正确 interface（界面）。我们聊到了：- 为什么自己运行 agent（智能体）会让你更愿意为 SaaS 付费，而不是更不愿意 - Figma 的 MCP server 如何让你从两个方向切入设计工作：把一个在线网页拿来，在 Figma canvas 上重建；或者把一个 Figma 设计交给 agent，让它通过 pull request 做修改 - 基于聊天的设计方式有哪些内在局限，这种形式并不适合生成大量新想法 - 为什么 review（评审）会成为下一个瓶颈，以及有哪些潜在方案可以帮助团队扩大评估规模。请看下方视频！时间戳：0:01:03 - 介绍；0:02:15 - SaaSpocalypse 这套叙事把方向搞反了；0:05:27 - Matt 的 email-agent 起源故事；0:13:21 - 发散式与收敛式设计思维；0:17:39 - Figma 的 MCP server；0:19:45 - 为什么设计 agent 需要个性化；0:22:09 - 每个问题本质上都是 context（上下文）问题；0:25:12 - Apple 和 Google 作为 context 领域当之无愧的王者；0:28:18 - 为什么 review 是新的瓶颈
Can you look the machine god in the face and keep your feet on the ground? Can you refuse to confuse your self with what can be measured, verbalized, or grasped? If so, you will go far.
你能直视 machine god，却依然脚踏实地吗？你能拒绝把自我与那些可以被测量、被言说、被把握的东西混为一谈吗？如果可以，你会走得很远。
Lucas joined us four years ago to sell ads. He’s leaving having lead a team of designers who consistently set the bar for taste and creative direction in AI. Much of what makes @every distinctive visually came from his brain—and you can see his influence across the AI ecosystem. As a founder, there nothing more rewarding than seeing someone like Lucas win—he’s going to do big things. And I am incredibly excited about the team and design culture he built @every that will continue to be inspired by his example and bar for excellence
Lucas 四年前加入我们时是来卖广告的。如今他离开时，已经带领着一支设计师团队，持续为 AI 领域的品味与创意方向树立标杆。@every 在视觉上之所以如此独特，很大一部分都来自他的头脑——而且你也能在整个 AI ecosystem（生态系统）中看到他的影响。作为 founder（创始人），没有什么比看到像 Lucas 这样的人取得成功更令人欣慰的了——他将会做成大事。而我也对他在 @every 打造的团队和设计文化感到无比兴奋；这种文化将继续受到他的榜样力量和他对卓越标准的启发
people internally are freaking out about this, worth paying attention:
内部很多人都在为这件事慌了，这值得关注：
Almost a week later! What are your thoughts on Opus 4.8? We were extremely bullish on it in testing—it seems the response was more tepid once y'all got your hands on it. If you disagreed with our take I'm curious why so we can tune our evaluations! One theory I have is that by nature it pushes on your frame a little more, and the results are high-variance—sometimes it does something amazing, and sometimes it disagrees in a way that is obviously wrong. But curious how you're feeling and what you're reaching for after a few days of testing
差不多一周后了！你们对 Opus 4.8 有什么看法？我们在测试时对它非常 bullish（看好）——但等你们亲自上手之后，反馈似乎就没那么热烈了。如果你不同意我们的判断，我很好奇原因是什么，这样我们就能调整我们的 evaluations（评估）！我的一个理论是，它的天性决定了它会更主动地推动你的 frame（思路框架）一点，因此结果的方差很高——有时候它会做出非常惊艳的东西，有时候它又会以一种明显错误的方式跟你唱反调。不过我还是很好奇，在测试了几天之后，你现在的感受如何，以及你会优先拿它来做什么
if you're comparing your startup to the manhattan project and you're not rocking this every day .... ngmi
如果你把自己的 startup 和 Manhattan Project 相提并论，却不是每天都在全力以赴地做这件事……那你就 ngmi
38b tokens and a 56h longest task 41 day current streak lfg codex
380 亿个 token，以及 56 小时最长任务，当前连续 41 天，lfg codex
we're having people over at the @every brownstone during NYC Tech Week ai-pilled readers, builders, and writers—and our team can't wait to see everyone come hang:
NYC Tech Week 期间，我们将在 @every brownstone 招待来访的 ai-pilled 读者、builder 和 writer——我们的团队已经迫不及待想见到大家，来一起玩了：
@StainlessAPI @RattrayAlex YouTube: Spotify:
@StainlessAPI @RattrayAlex YouTube：Spotify：
Anthropic just acquired developer tool startup @StainlessAPI, whose biggest customers were OpenAI and Google. Back in October, I had Stainless CEO and founder Alex Rattray (@RattrayAlex) on AI & I to talk about MCP servers and the unglamorous plumbing that makes AI agents actually work. (Disclosure: I’m a small investor in the company.) After Monday's news, the conversation lands differently—in it, Alex essentially walks me through the design thinking for building APIs, SDKs, and MCP servers that Anthropic paid a reported $300 million for. On @every's AI & I, we get into MCP and the future of the AI-native internet. Highlights include: - Design MCP servers to be lean and precise. Alex's best practices for building reliable MCP servers start with keeping the toolset small, giving each tool a precise name and description, and minimizing the inputs and outputs the model has to handle. At Stainless, they also often add a JSON filter on top to strip out unnecessary data. - Make complex APIs manageable with dynamic mode. To solve the problem of how an AI figures out which tool to use in larger APIs, Stainless switches to "dynamic mode," where the model gets only three tools: List the endpoints, pick one and learn about it, and then execute it. - MCP servers as business copilots. At Stainless, Alex uses MCP servers to connect tools like @NotionHQ and @HubSpot, so he can ask questions like, "Which customers signed up last week?" The system queries multiple databases and returns a summary that would've otherwise taken multiple logins and searches. - Create a "brain" for your company with Claude Code. Alex built a shared company brain at Stainless by keeping Claude Code running on his system and asking it to save useful inputs—like customer feedback and SQL queries—into GitHub. Over time, this creates a curated archive his team can query easily. - The future of MCP is code execution. Instead of giving models hundreds of tools, Alex believes the most powerful setup will be a simple code execution tool and a doc search tool. The AI writes code against an API's SDK, runs it on a server, and checks the docs when it gets stuck. This is a must-watch for anyone who wants to understand MCP—and learn how to use them as a competitive edge. Watch below! Timestamps Introduction: 00:01:15 APIs and MCP, the connectors of the new internet: 00:05:09 Why MCP exists: 00:11:00 Why MCP servers are hard to get right: 00:17:15 Design principles for reliable MCP servers: 00:20:24 Using MCP for business ops at Stainless: 00:25:06 Alex’s take on the security model for MCP: 00:40:57 How one-off AI actions become permanent production software: 00:44:42
Anthropic 刚刚收购了开发者工具初创公司 @StainlessAPI，而它最大的客户一直是 OpenAI 和 Google。早在 10 月，我就曾在 AI & I 节目中邀请 Stainless 的 CEO 兼创始人 Alex Rattray（@RattrayAlex）来聊 MCP servers，以及那些让 AI agents 真正运转起来、并不起眼的底层“管道”。（披露一下：我本人是这家公司的小额投资人。）在周一的消息之后，再回看这段对话，会有不同的感受——在这期内容里，Alex 基本上向我完整讲解了构建 APIs、SDKs 和 MCP servers 的设计思路，而 Anthropic 据报道正是为这些能力支付了 3 亿美元。 在 @every 的 AI & I 中，我们深入聊了 MCP 和 AI-native internet 的未来。精彩要点包括：- 将 MCP servers 设计得精简而精确。Alex 关于构建可靠 MCP servers 的最佳实践，首先是保持工具集尽量小，为每个工具提供精确的名称和描述，并尽量减少 model 需要处理的输入和输出。在 Stainless，他们还经常额外加上一层 JSON 过滤器，以剔除不必要的数据。- 用 dynamic mode 让复杂 API 更易管理。为了解决 AI 在大型 API 中该如何判断使用哪个工具的问题，Stainless 会切换到 “dynamic mode”，在这种模式下，model 只会拿到三个工具：列出 endpoints、选择其中一个并了解其信息，然后执行它。- 把 MCP servers 当作业务 copilot。在 Stainless，Alex 用 MCP servers 连接了 @NotionHQ 和 @HubSpot 之类的工具，这样他就能提出类似“上周有哪些客户注册了？”这样的问题。系统会查询多个数据库并返回摘要，而这原本需要多次登录和搜索才能完成。- 用 Claude Code 为公司建立一个“brain”。Alex 在 Stainless 搭建了一个共享的公司 brain：他让 Claude Code 持续运行在自己的系统上，并要求它把有用的输入——比如客户反馈和 SQL queries——保存到 GitHub。随着时间推移，这会形成一个经过整理的归档，团队可以很方便地查询。- MCP 的未来是 code execution。Alex 认为，与其给 model 几百个工具，不如提供一个简单的 code execution 工具和一个文档搜索工具，这才是最强大的配置。AI 可以针对某个 API 的 SDK 编写代码，在服务器上运行，并在卡住时查阅文档。对于任何想真正理解 MCP——以及学会如何把它作为竞争优势来使用的人来说，这期内容都非常值得一看。可在下方观看！时间戳如下：引言：00:01:15 API 和 MCP，作为新互联网的连接器：00:05:09 MCP 为什么会出现：00:11:00 为什么 MCP servers 很难做好：00:17:15 可靠 MCP servers 的设计原则：00:20:24 在 Stainless 用 MCP 处理业务运营：00:25:06 Alex 对 MCP 安全模型的看法：00:40:57 一次性的 AI 动作如何变成长期的生产软件：00:44:42
we'll be publishing a complete guide to codex soon on @every get notified when it's out:
我们很快会在 @every 发布一份关于 codex 的完整指南；想在它发布时收到通知：
Amazing!! So proud of @RattrayAlex and honored to be a tiny investor
太棒了！！为 @RattrayAlex 感到非常骄傲，也很荣幸能做一名小小的投资人
people should write better books! the vast majority of books that get published in this category are slop
人们应该写出更好的书！这一类别里出版的绝大多数书都很粗制滥造
If you use a bicycle it weakens your walking abilities, too!
如果你使用自行车，它也会削弱你的步行能力！
Successfully Codex-pilling the world one text at a time
一次一段文本地，成功地用 Codex-pilling 改造整个世界
our full deep-dive on trying to launch an agent-as-a-service platform built on openclaw! my two bigs ones: 1. OpenClaw is awesome but it's EXTREMELY hard to build on it as a platform. it moves super fast, there are tons of regressions, it's not great to be the layer in between OpenClaw and a user 2. One super agent for a company beats 1-1 agents for everyone. I do think we're going to get there over time, but for now agents actually require a lot of work (often technical) to keep working well. And people with jobs don't want to be messing with the internals of the agent all day. However if you give everyone an agent that works really well and make it someone's job to make it good for the whole company...lots of good stuff ensues stay tuned we'll have more on this @every!
这是我们对尝试推出一个基于 openclaw 构建的 agent-as-a-service 平台所做的完整深度复盘！我有两个最大的结论：1. OpenClaw 很棒，但把它作为一个平台来构建东西极其困难。它变化非常快，回归问题（regressions）很多，而且夹在 OpenClaw 和用户之间做中间层并不是一件很理想的事。2. 对一家公司来说，一个超级 agent 胜过给每个人都配一个 1 对 1 agent。我的确认为，随着时间推移我们会走到那一步，但就目前而言，agent 实际上仍然需要大量工作（而且往往是技术性的）才能持续保持良好运行。而有正式工作的人，并不想整天都去折腾 agent 的内部机制。不过，如果你给每个人一个运行得非常好的 agent，并且把“让它对整个公司都足够好”变成某个人的职责……那就会产生很多好的结果。敬请关注，我们之后还会在 @every 发布更多相关内容！
Gabbing with my barber about Codex and he’s telling me about how he’s using Claude to run a crypto trading strategy and Gemini + nano banana to learn jiu jitsu moves Guy in the chair next to me leans over and says his company just bought 1,300 Claude seats 2026 is wild
我和我的理发师聊 Codex 时，他跟我说自己在用 Claude 跑一个 crypto trading strategy（加密货币交易策略），还用 Gemini + nano banana 学 jiu jitsu（柔术）动作。坐在我旁边椅子上的那位大哥探过身来说，他的公司刚买了 1,300 个 Claude 席位。2026 真是太疯狂了。
when people ask me how to get their org agent-pilled i always say the same thing: the #1 leading indicator is whether their leadership team personally uses Codex, Claude Code or Cowork day to day. that’s why over the last few months we’ve been working privately with leadership teams of the top companies in tech, helping them get their hands deep into Claude Code, Cowork, Codex and more. if you want @every to come and get your exec team agent-pilled, we’re opening up a few slots:
当人们问我，怎么让他们的 org（组织）变得 agent-pilled 时，我总是说同一句话：排名第一的 leading indicator（领先指标）是，他们的 leadership team（领导团队）是否自己每天都在用 Codex、Claude Code 或 Cowork。这也是为什么过去几个月里，我们一直在私下与 tech（科技）行业顶级公司的领导团队合作，帮助他们深度上手 Claude Code、Cowork、Codex 等工具。如果你想让 @every 来帮你的 exec team（高管团队）变得 agent-pilled，我们现在开放少量名额：
there are a few people in a tiny office in brooklyn 1-2 months ahead of SV founders / eng :)
在 Brooklyn 的一间小办公室里，有那么几个人，比 SV 的 founders / eng 提前了 1–2 个月 :)
we hired someone new to help out with social and YouTube and she’s finding INSANE A/B test uplifts like this
我们新招了一个人来帮忙做 social 和 YouTube，而她正在发现像这样夸张到离谱的 A/B test 提升效果
codex-native weekend hack project: 1. buy cable to connect MIDI keyboard to computer 2. "hey codex, make a watcher script and a little web app to show me which chords im playing" 3. okay cool, now give me some exercises and help me see how to improve! literally 5 minutes start to finish, and it works flawlessly
codex-native 周末 hack 项目：1. 买根线，把 MIDI keyboard 连到电脑上 2. “hey codex，给我做个 watcher script 和一个小 web app，告诉我我正在弹哪些和弦” 3. 好，很酷，现在再给我一些练习，并帮我看清该怎么提升！从开始到结束真的就 5 分钟，而且运行得毫无问题
“We got a tool to perform poorly” is the lowest form of science and journalism imo and is only relevant when the tool is, in fact, extremely useful
在我看来，“我们让一个工具表现得很差”是最低级的 science（科学）和 journalism（新闻报道）形式，而且只有当这个工具事实上极其有用时，这种说法才有相关性
Knives Can Blind You When You Stick Them in Your Eye submitted 17 Apr 2026
把刀插进自己眼睛里会把你弄瞎，提交于 2026 年 4 月 17 日
mythos obviously looks incredibly capable and im psyched to use it also if you're panicking about it: benchmarks don't measure model capability alone they measure model capability after a human has done the work of finding a prompt that lets the model’s capability appear that work is non-trivial, and requires skilled expert humans doing something that looks very much like a job
mythos 显然看起来能力强得惊人，我也很期待用它；另外，如果你正因此感到恐慌：benchmarks（基准测试）衡量的不只是 model（模型）能力本身，它们衡量的是在人类已经完成了寻找一个 prompt（提示词）、从而让模型的能力得以显现之后的模型能力；而这项工作并不简单，需要有技能的专家级人类去做某种看起来非常像一份工作的事情
Generational opportunity for anyone in AI to play the markets given this time lag. Guarantee everyone is psyched about Codex in a few months. Invest accordingly
鉴于这种时间滞后，现在对任何做 AI 的人来说，都是一个在市场中大展身手的世代级机会。我敢保证，几个月后大家都会对 Codex 兴奋不已。相应地进行投资吧。
Evidence that everyone at Milken is 3-4 months behind
这证明了 Milken 的每个人都落后了 3 到 4 个月。
the ai platform war is coming @kieranklaassen and i recorded a quick dispatch from code with @claudeai on the xAI compute deal, managed agents, and why anthropic is turning their api into a full cloud infrastructure for developers:
AI 平台之战即将到来，@kieranklaassen 和我与 @claudeai 一起从 Code 录了一段快速简报，聊了 xAI 的 compute 交易、managed agents（托管式 agent），以及为什么 Anthropic 正在把他们的 API 变成面向开发者的完整云基础设施：
I’ll be at Code with Claude today with @kieranklaassen and @tedescau Come say hi!
我今天会和 @kieranklaassen 以及 @tedescau 一起在 Code with Claude，欢迎来打招呼！
no mythos this week! but some interesting stuff coming :)
这周没有 mythos！不过会有一些有趣的东西要来了 :)
@every if you want to try a codex-native app, use proof to write:
@every，如果你想试试一个 codex-native app，就用 proof 来写：
clear that this is how we'll be doing most of our work for the next 10 years: agent running continuously on the left, application that you + the agent use on the right
很明显，未来 10 年我们的大部分工作都会这样进行：左侧是持续运行的 agent（智能体），右侧是你和 agent 一起使用的 application（应用程序）
experimenting with using Codex + Chronicle as a focus tracker:
正在尝试把 Codex + Chronicle 用作一个专注度追踪器：
Agents who can buy, sell, and trade on our behalf are becoming a major part of the economy. But what exactly are they doing? Stripe sees 2% of global GDP, so they’re the company with the best view of what’s going on in the earliest innings of the agent economy. That’s why I had Emily Glassberg Sands, who leads data and AI at @stripe, on @every’s AI & I. We covered: - Most of us still don’t trust AI with larger online purchases. People are hesitant to let AI make expensive purchases like a vacation or a couch—just like the early days of online shopping. But a superhero outfit for a kid who needs one stat? Sure, let the agent handle it. - Fraud is moving up the stack. It used to mean stolen credit cards. Now attackers are stealing free-trial tokens and compute credits. Free-trial abuse has 4x-ed in the last six months.. - AI is on both sides of fraud. Fraudsters are using it to scale attacks, while Stripe is using it to detect them. They’re blocking 250,000 fraudulent free trials a week for one large customer. - AI companies are growing faster than any cohort Stripe has ever tracked. Top companies hit $30M ARR in 18 months—3x faster than the 2018 SaaS class. So far, it’s net new spend instead of cannibalized software budgets. If you want to understand how AI is reshaping online commerce, this one deserves your time. Timestamps Introduction: 00:00:45 New rules for an agent-driven economy: 00:01:27 Compute theft is the new payment fraud: 00:03:57 How Stripe expanded fraud detection from checkout to the full customer lifecycle: 00:10:00 Why AI companies are scaling way faster than top SaaS companies: 00:19:48 Outcome-based billing is replacing seat-based pricing: 00:23:27 Where AI spending is coming from: 00:29:57 How the developer experience changes when agents are the builders: 00:36:45 The agentic commerce spectrum, from assisted buying to autonomous purchasing: 00:41:00 Meet Link, a consumer wallet for delegated agent purchases: 00:51:06
能代表我们进行 buy、sell 和 trade 的 agent，正在成为经济中的一个重要组成部分。但它们究竟在做什么？Stripe 看到了全球 GDP 的 2%，因此在 agent economy（agent 经济）最早期阶段正在发生什么这件事上，它是视角最好的公司。这就是为什么我邀请了 Emily Glassberg Sands——她在 @stripe 负责 data 和 AI——来到 @every 的 AI & I。我们聊到：- 我们大多数人仍然不信任 AI 来处理金额较大的线上购买。人们不太愿意让 AI 做出像度假或沙发这样的高价购买决定——就像 online shopping 的早期一样。但如果是给一个急需超级英雄服装的孩子买衣服？当然，可以让 agent 来处理。- 欺诈正在向上游迁移。过去，欺诈意味着被盗的信用卡。现在，攻击者正在窃取 free-trial token（免费试用 token）和 compute credits（算力额度）。在过去六个月里，free-trial abuse（免费试用滥用）增长了 4 倍。- AI 正在欺诈的两端同时发挥作用。诈骗者在用它扩大攻击规模，而 Stripe 在用它检测这些攻击。对于一家大型客户，他们每周会拦截 250,000 次欺诈性免费试用。- AI 公司增长速度比 Stripe 追踪过的任何 cohort（群组）都更快。头部公司在 18 个月内就达到 3000 万美元 ARR（年度经常性收入）——比 2018 年的 SaaS 班级快 3 倍。到目前为止，这些支出是净新增支出，而不是对软件预算的挤占。如果你想了解 AI 如何重塑 online commerce（在线商业），这期内容值得你花时间。时间戳 引言：00:00:45 agent 驱动经济的新规则：00:01:27 算力盗窃是新的支付欺诈：00:03:57 Stripe 如何将欺诈检测从 checkout（结账）扩展到完整客户生命周期：00:10:00 为什么 AI 公司扩张速度远快于顶级 SaaS 公司：00:19:48 基于 outcome（结果）的计费正在取代基于 seat（席位）的定价：00:23:27 AI 支出来自哪里：00:29:57 当 agent 成为构建者时，developer experience（开发者体验）会如何变化：00:36:45 agentic commerce（agent 驱动商业）光谱：从辅助购买到自主购买：00:41:00 认识 Link，一个用于委托 agent 购买的消费者 wallet（钱包）：00:51:06
using posthog inside of codex it is SO MUCH better to browse posthog inside of codex—codex can write queries and see the results, and it makes our collaboration extremely seamless. codex can also kick off agents that write PRs or run production DB requests to back up insights it's finding a browser inside your desktop coding orchestration tool > an agent in your browser
在 codex 里使用 posthog，浏览 posthog 的体验真的好太多了——codex 可以编写查询并查看结果，这让我们的协作变得极其顺畅。codex 还可以启动 agent，去写 PR，或运行生产环境 DB 请求来支撑它发现的洞察；在你的桌面 coding orchestration tool 里的浏览器 > 你浏览器里的一个 agent
we might need to coin a new term: Codex-native Cowork-native Cursor-native apps designed to be used inside of the in-app browser of your agent of choice. the agent can use it and you can use it too, full context is shared, and you can both see what the other is doing. huge opportunity to build software here
我们可能需要创造一个新术语：Codex-native、Cowork-native、Cursor-native 的应用，专门设计为在你所选 agent 的应用内浏览器里使用。agent 可以用它，你也可以用它，完整上下文是共享的，而且你们双方都能看到对方在做什么。在这里构建软件有巨大的机会
we might need to coin a new term: Codex-native Cowork-native Cursor-native apps designed to be used inside of the in-app browser of your agent of choice. the agent can use it and you can use it too, full context is shared, and you can both see what the other is doing. huge opportunity to build software here
我们可能需要创造一个新术语：Codex-native、Cowork-native、Cursor-native 的应用，专门设计为在你所选 agent 的应用内浏览器里使用。agent 可以用它，你也可以用它，完整上下文是共享的，而且你们双方都能看到对方在做什么。在这里构建软件有巨大的机会
if a new world war were to break out, what would be its causes? "Germany would strive to regain lost provinces and establish herself as a great military power on the Continent"
如果爆发一场新的世界大战，它的起因会是什么？“Germany 会努力收复失去的省份，并在欧洲大陆上把自己确立为一个强大的军事强国。”
in a world where code is cheap to produce software is the new media come see me talk with @Borthwick and co @betaworks tomorrow about this!
在一个 code 廉价易产出的世界里，software 就是新的 media。明天来看我和 @Borthwick 以及 @betaworks 的同事聊聊这个话题吧！
what we observe is never the model itself, only the model exposed to our method of questioning
我们观察到的从来不是模型本身，而只是暴露在我们提问方法之下的模型