Chat is all you need

matt palmer 2026-08-19 14.2 KB 4.2M 1.5K 𝕏 文章列表

本文探讨了以聊天作为主要交互界面的AI代理（如Grok Bot）如何重塑软件开发、个人软件和工作流自动化。作者认为，真正的力量来自底层系统而非界面，聊天界面可以替代传统UI，通过自然语言编写提示词来控制代理。文章分享了个人健身应用从传统软件转变为聊天驱动的案例，以及如何将聊天作为编码的“外循环”来保护上下文，并讨论了工作流自动化如何通过聊天简化。

AI代理聊天界面个人软件工作流自动化上下文保护自然语言编程软件定义效率工具

# Chat is all you need **作者**: matt palmer **日期**: 2024-06-16T00:45:40.000Z **来源**: [https://x.com/mattyp/status/2089758434921160877](https://x.com/mattyp/status/2089758434921160877) --- ![Image](/doc/4680/image/img_e1a092ed945b.jpg) Grok Bot has a single output: a message This might seem backwards. Power-user tools are built on complexity and breadth, right? Chat can’t possibly have my workflow. It’s a special workflow. I’m a developer. <video controls preload="metadata" poster="/doc/4680/image/img_8f7268e8f544.jpg" src="/videos/0145.mp4"></video> "Look with your special workflow" But what if "power" comes from the underlying system rather than the interface? Instead of code, we now write prompts. We swapped IDEs for TUIs because the terminal was the real tool - chat is just the interface. We don't need a user interface - at least, not in the traditional sense. We need a computer. And rather than using it ourselves, we can give it to an agent. That's Grok Bot: an agent with a computer. I've been using Grok Bot in 4 different ways, and they're chat, all the way down. ## Personal software AI lowered the cost of code and unlocked personalized software - some even say 'disposable' software. > **David Hoang@davidhoang**: [原文链接](https://x.com/davidhoang/status/1802140453292372272) > > I believe there’s going to be four main categories of software moving forward: > - Commercial software > - Boutique software > - Personal software > - Disposable software > Disposable is a sustainable way of executing a use case without the need to maintain it. > https://proofofconcept.pub/p/the-four-types-of-software-in-the… > > ![Illustration](/doc/4680/image/img_75539d2fcc44.jpg) > ![The concept of a disposable object is interesting and can be a great design (if done right). There are disposable cameras, contact lenses, and many others. In fact, the sourdough clam chowder bowl and tostada shell are two of the best examples of great disposable design. What about software that destroys itself when a certain job is done? I see in the future there are certain types of software created by people (or an AI) for a certain purpose and will dispose of it.](/doc/4680/image/img_ccf70fba9bce.jpg) > ![Image](/doc/4680/image/img_f19c79d2bcd4.jpg) But what is software? Software is logic, persistent storage, and an interface. Software takes inputs, does stuff with it, then returns outputs. Historically, we've described logic as code. Increasingly, we describe logic through English. Persistent storage used to be hard (databases) but now, persistent storage can really be… anything: A CSV, a repository, a JSON file. Interfaces, however, are notoriously difficult. They're designed around rigid workflows - they need to be opinionated, but malleable. All software comes with a tradeoff between simplicity and scope. What I build starts simple, then quickly becomes difficult to manage. Software exists to make life easier, yet most personal software turns into an unpaid second job.¹ We need something better. From software to chat My most recent bout of personal software was a strength training app. > **matt palmer@mattyp**: [原文链接](https://x.com/mattyp/status/2070129458305257746) > > I built a progressive overload and periodized strength training platform > Here's a demo - heavily inspired by @drmikeisraetel, @JeffNippard, & @strongerbysci > Under the hood, it uses @convex, which is pretty neat because I get realtime sync out of the box and a super simple > > ![Embedded video](/doc/4680/image/img_e23a71d58fe1.jpg) I spent weeks refining the UI, adding a database², building out formulae, and writing tests. I built a full preview deploy system and some nice feedback loops for iterating on my app. There was always a new edge case, a new thing to fix, a new problem. I would write down notes and fire off cloud agents. Slowly, it got better! Still, it breaks. One day I wondered - what if this was just a chat? Talking with Grok Bot, I extracted all of the core logic from my app. I rewrote it as a set of skills and MCP servers.³ I asked Grok to come up with a format for logging my workouts. I decided to use Git as the storage layer. Now, Grok sends me my workout each day. I log each workout with a 30 second voice note. If I want to change an exercise, I ask. If a progression looks off, I ask. If I want to add a feature, I ask. ![Image](/doc/4680/image/img_e56194a7f94a.jpg) The app is more flexible, simpler, and faster. I can add or delete features with a single message. There is no CI, no preview deploy, and no UI to break. I did this while walking to the gym. I update and refine the app as I go. I eventually wanted to build chat into my fitness app. Now my fitness app is chat. ## Code "Ok forget this guy - he's writing code with a chatbot." For the six of you that made it this far, hear me out. I'm not replacing the coding harness, I'm calling the coding harness. Some of the best engineers at [@SpaceXAI](https://x.com/@SpaceXAI) are shipping hundreds of PRs, starting in [@Bot](https://x.com/@Bot). Last week, I sat down with [@poteto](https://x.com/@poteto). She described Grok Bot as the perfect "outer loop:" a place to gather & shape context before handing it off. "Context is precious," she said. ![Image](/doc/4680/image/img_ffd02c0c5c65.jpg) Agents work on next token prediction: agent inputs directly influence outputs. Chat, files, what agents search for, what agents read. What agents do access is just as important as what they don't access. Lauren and others use Grok Bot to explore ideas, source information, gather context, then stage prompts to fire off Cloud Agents. That work protects the context of inner loop agents. Rather than mucking up my Cloud Agent with a bunch of exploratory research, I can send them only what's necessary - avoiding context poisoning.⁴ Code is abstracted yet another layer up - from Cloud Agents to "Bots" or agent managers. Luckily, Cursor makes this handoff seamless - agents started by Grok Bot are waiting for me inside my Cursor app. You, the human, chat with the outer loop agent. The outer loop agent kicks off inner loop agents. Coding tools are the inner loop, the outer loop is chat. ## Workflow automation If software is just logic, persistent storage, and an interface, then a workflow is a form of software. We create workflows, chaining together multiple tools, to achieve outcomes. Those workflows require a human. That's a problem, because humans make mistakes and most UIs are very bad. Accounting, shopping, and banking are painstaking operations. I may be dramatizing, but dealing with these websites introduces a real cognitive load. Have you ever booked the wrong flight? Ordered something to the wrong location? Thought you paid for something but didn't? I have. ![Image](/doc/4680/image/img_40a84426be10.jpg) More than personal ops - we all work in similar tools every day. The promise of AI is not to render us useless⁵, but to eliminate things that distract us from real work and creativity. Workflows and GUIs were meant to augment our craft, but too often they replace it. For many, this is the first time we're free to think, express, and dictate our preferences and have someone else implement them. The same way personal software can collapse into chat, consumer and business workflows can collapse, too. > **Krista Letz@kristaletz**: [原文链接](https://x.com/kristaletz/status/2089103618121314689) > > It's been really fun leveraging @bot for GTM! Have spent a few weeks iterating on some of my workflows, so I wanted to share how I use Grok Bot at SpaceXAI. The crazy thing is that we don't need MCPs, hosted markdown, or tools. We just need AI and a computer, with the same software we use. That's what Grok is: AI with a computer. Saved search I buy and sell lots of things online. If you've ever browsed a marketplace, you've likely saved and reused searches: It's a godsend. But saved search is really just pre-applied logic - it's a workflow. After I save a search, I start my very particular search process: - Is the listing legit? - Is the condition good? - Is the price fair? - Is it nearby? This is my "software," simple filtering logic I use to find what I'd like - logic saved search can't handle, but that Grok Bot can. My Grok Bot is now looking across Craigslist and Facebook Marketplace, checking each day for new espresso machines that match my criteria. ![Image](/doc/4680/image/img_22ef7449fd17.jpg) It can draft a message and send one for me. It can advise me on what to buy and what not to buy. It can make sure I don't make dumb mistakes (I often do when I'm tired). It's a much better interface than a web browser. Consumer and enterprise UIs are collapsing into chat. ## Information retrieval Much of knowledge work is search and retrieval. It's asking questions and getting answers. Tools like Notion and Slack revolutionized the way we communicate, but finding the right information is still terribly difficult - especially across multiple sources. Learning is limited by feedback loops. If it takes 2 hours to answer a question, I will have fewer "cycles" than someone who gets an answer in 2 minutes. This can be painful, because many of us were first introduced to plugins and MCP servers through flaky, brittle interfaces. But they've gotten much better - in large part thanks to the Grok Bot team. I've found Grok Bot plugins to be excellent - surprisingly excellent. I've hooked up our codebase, Slack, Notion, my email, my calendar, Figma… Multiple accounts, even. ![Image](/doc/4680/image/img_970e9000dedc.jpg) And Grok is scary good at finding things. This was the first thing I did when I joined Cursor. Onboarding at Cursor I onboarded at Cursor in about a week. I asked [@leerob](https://x.com/@leerob) a few questions on my first day. Then, I got access to Grok Bot. My onboarding turned into dozens of questions to Grok. Questions about the product, questions about IT, questions about where I should park, or when lunch was. Questions about what launches typically look like, who owns what, where things live. Every single time I thought "Grok won't be able to answer this." Every time, it really did. Sometimes, my years of professional 'experience' kicked in "who owns this thing? I should ping them on Slack." "NO! 🙅‍♂️" I would tell myself (with that emoji in my head) "I will ask Grok! I will not send my busy teammates more messages!" ![Image](/doc/4680/image/img_1a2ca06fe3ec.jpg) It turns out, at a company of Cursor's (and now SpaceX's) size, most people have asked your question and a surprising number of things have been documented. But code is more than software, it's a source of truth. Grok Bot has been able to look through our codebase and tell me where things are in the app.⁶ Imagine the wasted hours if I had to ping a person about each question - I probably wouldn't be writing this right now! And now... I'm using Grok Bot to onboard to [@SpaceXAI](https://x.com/@SpaceXAI). ## The future is chat Last year, we watched the IDE collapse to chat. Now, we're seeing other interfaces do the same. More than automating common tasks, Grok Bot has me reconsidering what I build. If I can emulate a workflow with an agent and a SaaS, do I really need to vibe code a solution? Or is an agent and a computer good enough? SaaS is thriving, but the consumers are now computer-use agents. Chat is replacing a surprising number of my other workflows. This is a bit scary: if not for the tools, what is a craftsman? But what if the real tool is English: the ability to articulate what we perceive and desire. And the real craft is having novel ideas about what we build. Chat is the most familiar human interface. We needed skeuomorphic design on the first iPhone because humans understood tactile interfaces. We need chat interfaces with AI, because humans know chat. And it turns out I like bothering my agents with voice notes the same way I like bothering my friends. Footnotes ¹ Unless you work in developer relations :) ² Shout out [@convex](https://x.com/@convex) & [@PlanetScale](https://x.com/@PlanetScale) ³ Packaged as [Agent Plugins](https://agent-plugins.org/) ⁴ Shout out [@dbreunig](https://x.com/@dbreunig) ⁵ The ramblings of effective altruists may lead us to this conclusion ⁶ [this tweet](https://x.com/mattyp/status/2086632898229322118?s=20), for example ## 相关链接 - [Elon Musk reposted](https://x.com/elonmusk) - [matt palmer](https://x.com/mattyp) - [@mattyp](https://x.com/mattyp) - [97K](https://x.com/mattyp/status/2089758434921160877/analytics) - [Jun 16, 2024](https://x.com/davidhoang/status/1802140453292372272) - [84K](https://x.com/davidhoang/status/1802140453292372272/analytics) - [Jun 25](https://x.com/mattyp/status/2070129458305257746) - [@drmikeisraetel](https://x.com/drmikeisraetel) - [@JeffNippard](https://x.com/JeffNippard) - [@strongerbysci](https://x.com/strongerbysci) - [@convex](https://x.com/convex) - [12K](https://x.com/mattyp/status/2070129458305257746/analytics) - [@SpaceXAI](https://x.com/@SpaceXAI) - [@Bot](https://x.com/@Bot) - [@poteto](https://x.com/@poteto) - [Aug 17](https://x.com/kristaletz/status/2089103618121314689) - [@bot](https://x.com/bot) - [1.3M](https://x.com/kristaletz/status/2089103618121314689/analytics) - [@leerob](https://x.com/@leerob) - [@SpaceXAI](https://x.com/@SpaceXAI) - [@convex](https://x.com/@convex) - [@PlanetScale](https://x.com/@PlanetScale) - [Agent Plugins](https://agent-plugins.org/) - [@dbreunig](https://x.com/@dbreunig) - [this tweet](https://x.com/mattyp/status/2086632898229322118?s=20) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [12:56 AM · Aug 19, 2026](https://x.com/mattyp/status/2089758434921160877) - [91.4K Views](https://x.com/mattyp/status/2089758434921160877/analytics) - [View quotes](https://x.com/mattyp/status/2089758434921160877/quotes) --- *导出时间: 2026/8/19 12:54:17* --- ## 中文翻译 # 聊天即一切 **作者**: matt palmer **日期**: 2024-06-16T00:45:40.000Z **来源**: [https://x.com/mattyp/status/2089758434921160877](https://x.com/mattyp/status/2089758434921160877) --- ![Image](/doc/4680/image/img_e1a092ed945b.jpg) Grok Bot只有一个输出：一条消息。这看起来可能有些倒退。高级用户工具不是建立在复杂性和广度之上的吗？聊天怎么可能承载我的工作流程？我的工作流程很特殊。我可是个开发者。 <video controls preload="metadata" poster="/doc/4680/image/img_8f7268e8f544.jpg" src="/videos/0145.mp4"></video> “看看你那特殊的工作流程吧” 但如果“强大”来自于底层系统，而非界面本身呢？我们不再编写代码，而是编写提示词。我们用TUI取代了IDE，因为终端才是真正的工具——聊天只是界面。我们不需要用户界面——至少不是传统意义上的界面。我们需要一台计算机。而且，与其我们自己使用它，不如把它交给一个智能体。这就是Grok Bot：一个拥有计算机的智能体。我一直在以四种不同的方式使用Grok Bot，而它们归根结底都是聊天。 ## 个人软件 AI降低了代码的成本，解锁了个性化软件——有些人甚至称之为“一次性”软件。 > **David Hoang@davidhoang**: [原文链接](https://x.com/davidhoang/status/1802140453292372272) > > 我相信未来软件将主要有四大类别： > - 商业软件 > - 精品软件 > - 个人软件 > - 一次性软件 > 一次性软件是一种可持续的执行用例的方式，无需维护。 > https://proofofconcept.pub/p/the-four-types-of-software-in-the… > > ![Illustration](/doc/4680/image/img_75539d2fcc44.jpg) > ![The concept of a disposable object is interesting and can be a great design (if done right). There are disposable cameras, contact lenses, and many others. In fact, the sourdough clam chowder bowl and tostada shell are two of the best examples of great disposable design. What about software that destroys itself when a certain job is done? I see in the future there are certain types of software created by people (or an AI) for a certain purpose and will dispose of it.](/doc/4680/image/img_ccf70fba9bce.jpg) > ![Image](/doc/4680/image/img_f19c79d2bcd4.jpg) 但软件是什么？软件是逻辑、持久化存储和界面的结合。软件接收输入，进行处理，然后返回输出。历史上，我们用代码来描述逻辑。如今，我们越来越多地用英语来描述逻辑。持久化存储曾经很困难（数据库），但现在，持久化存储实际上可以是……任何东西：一个CSV文件、一个代码仓库、一个JSON文件。然而，界面是出了名的难做。它们围绕固定的工作流程设计——它们需要有明确的倾向性，但又要有可塑性。所有软件都在简单性和功能范围之间进行权衡。我构建的东西一开始很简单，但很快就变得难以管理。软件的存在是为了让生活更轻松，然而大多数个人软件却变成了一份无薪的第二职业。¹ 我们需要更好的东西。从软件到聊天我最近一次做个人软件是一个力量训练应用。 > **matt palmer@mattyp**: [原文链接](https://x.com/mattyp/status/2070129458305257746) > > 我构建了一个渐进超负荷和周期化力量训练平台 > 这是一个演示——深受@drmikeisraetel、@JeffNippard和@strongerbysci的启发 > 底层使用了@convex，这非常棒，因为我开箱即用地获得了实时同步，而且超级简单 > > ![Embedded video](/doc/4680/image/img_e23a71d58fe1.jpg) 我花了数周时间打磨UI、添加数据库²、构建公式、编写测试。我构建了一个完整的预览部署系统，并为迭代应用建立了良好的反馈循环。总有新的边缘情况、新的问题要修复、新的难题要解决。我会记下笔记，然后启动云端智能体。慢慢地，它变好了！但它仍然会出问题。有一天我突发奇想——如果这只是一个聊天呢？通过与Grok Bot对话，我提取了应用中的所有核心逻辑。我将其重写为一组技能和MCP服务器。³ 我让Grok为我的训练日志设计一个格式。我决定使用Git作为存储层。现在，Grok每天给我发送我的训练计划。我通过30秒的语音备忘录记录每次训练。如果我想更换一个动作，我就问它。如果某个进展看起来不对劲，我就问它。如果我想添加一个功能，我就问它。 ![Image](/doc/4680/image/img_e56194a7f94a.jpg) 这个应用更加灵活、更简单、也更快。我可以用一条消息添加或删除功能。没有CI，没有预览部署，也没有会崩溃的UI。我是在走去健身房的路上完成这一切的。我边走边更新和完善这个应用。我最终想在我的健身应用中集成聊天功能。现在我的健身应用就是聊天。 ## 代码 “好吧，忘了这家伙吧——他在用聊天机器人写代码。”对于坚持读到这里的六位读者，请听我说完。我不是在取代编码工具链，我是在调用编码工具链。 [@SpaceXAI](https://x.com/@SpaceXAI) 一些最优秀的工程师正在从 [@Bot](https://x.com/@Bot) 开始，提交成百上千个PR。上周，我和 [@poteto](https://x.com/@poteto) 坐下来聊了聊。她将Grok Bot描述为完美的“外循环”：一个在交接之前收集和整理上下文的地方。“上下文是宝贵的，”她说。 ![Image](/doc/4680/image/img_ffd02c0c5c65.jpg) 智能体基于下一个词元预测工作：智能体的输入直接影响输出。聊天、文件、智能体搜索的内容、智能体阅读的内容。智能体能够访问什么，与它们不能访问什么同样重要。 Lauren和其他人使用Grok Bot来探索想法、获取信息、收集上下文，然后准备好提示词以触发云端智能体。这项工作保护了内循环智能体的上下文。与其用一堆探索性研究来搞乱我的云端智能体，我可以只发送必要的信息——避免上下文污染。⁴ 代码又被抽象了一层——从云端智能体到“机器人”或智能体管理器。幸运的是，Cursor让这种交接变得无缝——由Grok Bot启动的智能体已经在我的Cursor应用中等待我了。你，作为人类，与外循环智能体聊天。外循环智能体启动内循环智能体。编码工具是内循环，外循环是聊天。 ## 工作流自动化如果软件只是逻辑、持久化存储和界面的结合，那么工作流就是一种软件形式。我们创建工作流，将多个工具串联起来，以实现目标。这些工作流需要人来参与。这是个问题，因为人会犯错，而且大多数UI都非常糟糕。会计、购物和银行业务都是需要一丝不苟的操作。我可能有些夸张，但处理这些网站确实带来了实实在在的认知负担。你有没有订错过航班？把东西寄到了错误的地址？以为付了款但实际上没有？我都遇到过。 ![Image](/doc/4680/image/img_40a84426be10.jpg) 不仅仅是个人事务——我们每天都在类似的工作工具中工作。 AI的承诺不是让我们变得无用⁵，而是消除那些让我们无法专注于真正工作和创造力的干扰。工作流和GUI本应增强我们的技艺，但很多时候它们反而取代了它。对许多人来说，这是我们第一次可以自由地思考、表达、陈述自己的偏好，并让别人来实施。就像个人软件可以坍缩为聊天一样，消费级和商业工作流也可以坍缩。 > **Krista Letz@kristaletz**: [原文链接](https://x.com/kristaletz/status/2089103618121314689) > > 利用@bot进行GTM（市场进入）真的非常有趣！我花了几周时间迭代我的一些工作流，所以我想分享一下我在SpaceXAI如何使用Grok Bot。疯狂的是，我们不需要MCP、托管Markdown或工具。我们只需要AI和一台计算机，以及我们正在使用的软件。这就是Grok的本质：拥有计算机的AI。保存的搜索我在网上买卖很多东西。如果你浏览过市场平台，你可能保存并重复使用过搜索：这简直是天赐之物。但保存的搜索实际上只是预先应用的逻辑——它是一个工作流。保存搜索后，我会开始我非常特定的搜索流程： - 这个列表是真实的吗？ - 物品状况好吗？ - 价格合理吗？ - 离得近吗？这是我的“软件”，是我用来寻找心仪之物的简单过滤逻辑——保存的搜索无法处理这些逻辑，但Grok Bot可以。我的Grok Bot现在正在浏览Craigslist和Facebook Marketplace，每天检查是否有符合我标准的新意式浓缩咖啡机。 ![Image](/doc/4680/image/img_22ef7449fd17.jpg) 它可以为我起草消息并发送。它可以建议我买什么、不买什么。它可以确保我不会犯愚蠢的错误（我累的时候经常犯）。这是一个比网页浏览器好得多的界面。消费级和企业级UI正在坍缩为聊天。 ## 信息检索知识工作的很大一部分是搜索和检索。就是提出问题并获得答案。像Notion和Slack这样的工具彻底改变了我们的沟通方式，但找到正确的信息仍然非常困难——尤其是在跨多个来源的情况下。学习受到反馈循环的限制。如果回答一个问题需要2小时，我的“循环”次数就会比2分钟就能得到答案的人少。这可能很痛苦，因为许多人最初是通过不稳定、脆弱的界面接触到插件和MCP服务器的。但它们已经变得好多了——很大程度上要归功于Grok Bot团队。我发现Grok Bot插件非常出色——出乎意料地出色。我已经连接了我们的代码库、Slack、Notion、我的电子邮件、我的日历、Figma……甚至多个账户。 ![Image](/doc/4680/image/img_970e9000dedc.jpg) 而且Grok在查找东西方面好得惊人。这是我加入Cursor时做的第一件事。在Cursor的入职体验我大约在一周内完成了在Cursor的入职。第一天我问了 [@leerob](https://x.com/@leerob) 几个问题。然后，我获得了Grok Bot的访问权限。我的入职变成了向Grok提出的几十个问题。关于产品的问题、关于IT的问题、关于我应该在哪里停车或午餐时间的问题。关于发布通常是什么样子的、谁负责什么、东西放在哪里的问题。每次我都想“Grok肯定回答不了这个问题”。但每一次，它真的都回答了。有时候，我多年的职业“经验”会冒出来：“这东西谁负责？我应该在Slack上联系他们。” “不！🙅‍♂️”我会告诉自己（脑子里带着那个表情符号） “我要问Grok！我不会再给我忙碌的队友们发消息了！” ![Image](/doc/4680/image/img_1a2ca06fe3ec.jpg) 事实证明，在Cursor（现在是SpaceX）这样规模的公司里，大多数人都问过你的问题，而且有数量惊人的事情已经被记录在案。但代码不仅仅是软件，它还是事实的来源。 Grok Bot能够查看我们的代码库，并告诉我应用中的各个部分在哪里。⁶ 想象一下，如果每个问题都要去打扰一个人，那会浪费多少时间——我可能现在都不会写这篇文章了！而现在……我正在使用Grok Bot来入职 [@SpaceXAI](https://x.com/@SpaceXAI)。 ## 未来是聊天去年，我们目睹了IDE坍缩为聊天。现在，我们看到其他界面也在经历同样的过程。 Grok Bot不仅仅是自动化常见任务，它还让我重新思考我要构建什么。如果我能用一个智能体和一个SaaS来模拟工作流，我真的需要“vibe coding”一个解决方案吗？还是一个智能体和一台计算机就足够了？SaaS正在蓬勃发展，但现在的消费者变成了计算机使用智能体。聊天正在取代我大量其他工作流。这有点令人恐惧：如果不是为了工具，那工匠是什么？但如果真正的工具是英语：是表达我们所感知和渴望之事物的能力呢？而真正的技艺是对于我们要构建的东西拥有新颖的想法。聊天是最熟悉的人类界面。我们在第一代iPhone上需要拟物化设计，因为人类理解触觉界面。我们需要与AI的聊天界面，因为人类懂得聊天。事实证明，我喜欢用语音备忘录打扰我的智能体，就像我喜欢打扰我的朋友一样。脚注 ¹ 除非你在开发者关系部门工作 :) ² 感谢 [@convex](https://x.com/@convex) 和 [@PlanetScale](https://x.com/@PlanetScale) ³ 打包为 [Agent Plugins](https://agent-plugins.org/) ⁴ 感谢 [@dbreunig](https://x.com/@dbreunig) ⁵ 有效利他主义者的胡言乱语可能会让我们得出这个结论 ⁶ 例如 [这条推文](https://x.com/mattyp/status/2086632898229322118?s=20) ## 相关链接 - [Elon Musk reposted](https://x.com/elonmusk) - [matt palmer](https://x.com/mattyp) - [@mattyp](https://x.com/mattyp) - [97K](https://x.com/mattyp/status/2089758434921160877/analytics) - [Jun 16, 2024](https://x.com/davidhoang/status/1802140453292372272) - [84K](https://x.com/davidhoang/status/1802140453292372272/analytics) - [Jun 25](https://x.com/mattyp/status/2070129458305257746) - [@drmikeisraetel](https://x.com/drmikeisraetel) - [@JeffNippard](https://x.com/JeffNippard) - [@strongerbysci](https://x.com/strongerbysci) - [@convex](https://x.com/convex) - [12K](https://x.com/mattyp/status/2070129458305257746/analytics) - [@SpaceXAI](https://x.com/@SpaceXAI) - [@Bot](https://x.com/@Bot) - [@poteto](https://x.com/@poteto) - [Aug 17](https://x.com/kristaletz/status/2089103618121314689) - [@bot](https://x.com/bot) - [1.3M](https://x.com/kristaletz/status/2089103618121314689/analytics) - [@leerob](https://x.com/@leerob) - [@SpaceXAI](https://x.com/@SpaceXAI) - [@convex](https://x.com/@convex) - [@PlanetScale](https://x.com/@PlanetScale) - [Agent Plugins](https://agent-plugins.org/) - [@dbreunig](https://x.com/@dbreunig) - [this tweet](https://x.com/mattyp/status/2086632898229322118?s=20) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [12:56 AM · Aug 19, 2026](https://x.com/mattyp/status/2089758434921160877) - [91.4K Views](https://x.com/mattyp/status/2089758434921160877/analytics) - [View quotes](https://x.com/mattyp/status/2089758434921160877/quotes) --- *导出时间: 2026/8/19 12:54:17*