Grok Bot + Grok 4.6 + Cursor Origin - is Claude Code dead?
频道: How I AI
视频: https://www.youtube.com/watch?v=8ONFvAtboZ4
原文语言: en
统计: 共 22 轮 · Claire Vo 19
[0:00] Claire Vo
I don't know if you know this, but the whole open AI versus Anthropic thing is old news. Everybody I know is secretly becoming a Grok boy. Yep, ever since Cursor was acquired by SpaceX for, I believe it was 60 billion American dollars, everybody's been really excited about the new products released from the Cursor team. And in fact, I'm hearing more and more positive things about the Grok models for coding and the Grok models in general. In today's episode of How I AI, we're going to go through all the releases from the Cursor, xAI, SpaceX, whatever, the Elon Cinematic Universe of code builders and model builders. We're going to go through all their products, and I'm going to tell you what I think of Grokbot, Origin, the new GitHub competitor from Cursor, as well as the Grok model. Let's get to it. This episode is brought to you by bolt.new, the AI app builder for people who have ideas and want to ship them. Most AI tools spit out code that looks great in a demo and falls apart the second you try to do anything real with it. Or they lock you into their own platform with no real way out. Bolt is different. [music] You describe what you want to build, a startup MVP, a landing page, an internal tool, a side project, and Bolt generates production-ready code in minutes. Connect Stripe [music] or any other MCP, hook up your domain, and deploy it live. Founders are using Bolt to build businesses doing real revenue.
不知道你们注意到没有，OpenAI 和 Anthropic 之间的那点恩怨，已经是老黄历了。我认识的人现在一个个都在偷偷变成 Grok 党。真的，自从 Cursor 被 SpaceX 收购——我记得是 600 亿美金——大家就一直在期待 Cursor 团队发的新产品。而且我确实听到越来越多好话，说 Grok 模型写代码不错，Grok 模型整体也不错。今天这期 How I AI，我们就把 Cursor、xAI、SpaceX，管它叫什么吧，整个「马斯克宇宙」里做代码工具和做模型的这帮人最近放出来的东西全过一遍。我会挨个讲他们的产品，然后告诉你我怎么看 GrokBot、Origin（Cursor 做的 GitHub 竞品），还有 Grok 模型。咱们开始。本期节目由 bolt.new 赞助，这是一款面向「有想法、想赶紧做出来」的人的 AI 应用搭建工具。大多数 AI 工具吐出来的代码，演示的时候光鲜亮丽，你一旦真拿它干点正事就散架了；要么就是把你锁死在他们自己的平台上，根本没有退出的路。Bolt 不一样。[music] 你只要描述你想做什么——一个创业公司的 MVP、一个落地页、一个内部工具、一个业余项目——Bolt 几分钟就能生成可以直接上生产的代码。接上 Stripe [music] 或者任何一个 MCP（让 AI 调用外部工具和数据的那套标准接口），绑好你的域名，就能部署上线。已经有创业者用 Bolt 做出了真正在赚钱的生意。
[1:32] Claire Vo
Product managers are shipping prototypes their teams actually use.
产品经理们用它做出来的原型，团队是真的在用。
[1:36]
[music]
[music]
[1:36] Claire Vo
Designers and marketers are launching campaigns without waiting in line.
设计师和市场同学不用再排队等排期，自己就能把 campaign 发出去。
[1:40]
[music]
[music]
[1:41] Claire Vo
Anyone can build, engineering can ship, everyone wins. You just need an idea and a weekend. Check it out at bolt.new/howiai. I'm going to start with the most accessible product on our list today, Grokbot. In case you missed it, the Cursor/xAI team released GrokBot, which is a chat-style agent in a desktop and mobile app that you can use to do knowledge work for you. It's very similar to an open claw, but it's simpler, hosted, and easier to use. If you look at GrokBot just from a UI perspective, and I'm just pulling up the marketing site right now, we'll look at my GrokBot in a minute. It is definitely pulling from that iMessage terminal streamline chat experience. And it's doing something I really love. Not everybody loves this, but I love a multi-agent experience. I want each of my agents to have a job. I want them to have a name. I don't want one agent to rule them all. And the team has really leaned into this ethos with GrokBot. Now, I've been testing GrokBot. I had a couple days early access, and then I've been testing it earnest over about the past week, and I think there are some pros to GrokBot, and I think there are some cons to GrokBot, but I do suspect that GrokBot is going to catch the attention of Cursor customers who want an all-purpose agent for their employees. So, let's go into my GrokBot and show you a little bit about how it works. Now, as I said, this has this very like iMessage-style experience. It's a left bubble and right bubble, and you just chitty-chat with your GrokBots. I have set up a couple of GrokBots in my own account. There is Proddy McProd, my product manager GrokBot that's hooked up to you Chat PRD.
人人都能动手做，工程团队负责发布上线，皆大欢喜。你需要的只是一个想法和一个周末。去 bolt.new/howiai 看看吧。我先从今天名单里最容易上手的产品讲起——GrokBot。如果你没跟上进度：Cursor/xAI 团队发布了 GrokBot，这是一个聊天式的 agent（能替你干活的智能体），有桌面端也有手机 app，你可以让它帮你处理各种知识工作。它和 OpenClaw 很像，但更简单、是托管在云端的、也更好上手。单从界面看 GrokBot——我现在打开的是它的官网页面，一会儿再看我自己那几个 GrokBot——它的灵感明显来自 iMessage 那种终端式的、极简的聊天体验。而且它做了一件我特别喜欢的事。不是所有人都喜欢这个，但我喜欢多 agent 的玩法。我希望我的每一个 agent 都有自己的岗位，都有自己的名字。我不想要「一个 agent 统治一切」。GrokBot 团队在这个理念上贯彻得非常彻底。我已经测 GrokBot 有一阵了，先拿到了几天的抢先体验，然后正儿八经地用了大概一周。我觉得 GrokBot 有它的优点，也有它的缺点，但我确实预感它会打动那批 Cursor 客户——那些想给员工配一个万能 agent 的公司。那我们进我的 GrokBot，给你看看它到底怎么用。刚才说了，它就是很典型的 iMessage 式体验：左边一个气泡，右边一个气泡，你就跟你的这些 GrokBot 唠嗑。我在自己的账号里建了好几个 GrokBot。一个叫 Proddy McProd，是我的产品经理 GrokBot，接了 ChatPRD。
[3:32] Claire Vo
I have one of the out-of-the-box GrokBots, commitment tracker, that just keeps me honest with things I've told people that I will do for them, basically emails I need to reply to. I have a money maker bot that chases invoices and payments and sales deals. I have a case study buddy cuz I'm doing a bunch of case studies. And I have a GrokBot that monitors all my data for Chat Perdition tells me trends that I should pay attention to. So, I think the magic of GrokBot, if I had to tell you one thing that is the magic of GrokBot, it is plugins. So, I've always thought that Cursor was the best MCP client. I still to this day, when I want to connect to different MCPs in my stack, would still default generally to Cursor. I thought the connector experience was really nice and I thought the harness did a really nice job of talking to all these systems. They have brought that great plugin MCP experience into GrokBot. But, there's one killer feature, which is for any one of these connectors, whether it's Gmail or Slack or an MCP, you can connect multiple accounts to GrokBot. So, if you have two Gmails, you can connect two Gmails. If you're in seven different Slacks, you can connect the different Slacks. This is something that Codex has not figured out. Claude has not figured out. No one seems to figure out that I have approximately a dozen email addresses and Slack accounts and different things that I need to log in, even if it's the same service. This experience of setting up multiple accounts per connector, huge huge huge benefit. You could like package it up and send it off. It would have product market fit with me.
我还开了一个官方自带的 GrokBot：承诺追踪器（commitment tracker），专门盯着我答应过别人要做的事，别让我食言，说白了就是那些我该回而没回的邮件。我有一个「搞钱 bot」，负责追发票、追回款、追销售单子。我还有一个「案例研究搭子」，因为我最近在做一堆案例研究。另外还有一个 GrokBot 盯着 ChatPRD 的全部数据，告诉我哪些趋势值得我关注。所以，如果只让我说一件事——GrokBot 的魔力到底在哪儿——我会说是 plugin（插件）。我一直觉得 Cursor 是最好的 MCP 客户端。直到今天，我要连自己技术栈里那些 MCP 的时候，默认还是会先想到 Cursor。它的 connector（连接器，也就是把 agent 接到外部服务上的那层）体验做得非常好，它的 harness（承载模型跑起来的那套外壳程序）跟这些系统对话也处理得很漂亮。他们把这套优秀的 plugin/MCP 体验原封不动带进了 GrokBot。但这里面还有一个杀手级功能：不管是 Gmail、Slack 还是某个 MCP，任何一个 connector，你都可以往 GrokBot 上接多个账号。你有两个 Gmail，就接两个 Gmail。你同时待在七个不同的 Slack 里，就把七个都接上。这件事 Codex 没搞明白，Claude 也没搞明白。好像没人意识到，我手上有差不多十几个邮箱、Slack 账号和各种要登录的东西，哪怕它们是同一个服务。「每个 connector 可以配多个账号」这个体验，是天大天大天大的好处。你哪怕把这一个功能单独打包发出来，对我来说也算找到产品市场契合了。
[5:17] Claire Vo
As you can see here, I have already four email addresses connected to Gmail. I have my Chat Perdition um and a couple other business ones that I'm working on. Having all four of these connected and being able to traverse all those accounts in GrokBot is a killer feature. Just thank you to the Cursor {slash} XAI team for doing this. If you did this and this alone, I would be very happy. But, they did not do that and that alone. GrokBot ships with tons and tons of out-of-the-box uh plugins. These are generally the ones that were available in Cursor as well. And so, I've installed your basic productivity ones. I've installed GitHub and I've installed the ChatGPT plugin. And really, I've been using Grokbot as a way to just traverse and use my data. So, again, it's like a little bit of a fancy MCP client, but it does come with something special, which is every Grokbot comes with a computer. And so, you can actually go into virtual machine, and this is Prod McProd's computer. It's not very exciting, but it can use Chrome, it can use terminal, and it has some files. And so, again, this is like a little bit of a open claw light, where it has virtual machine. That virtual machine can use the web, and it has connectors and MCPs.
你看这儿，我已经往 Gmail 上接了四个邮箱地址。有 ChatPRD 的，还有另外几个我在做的业务的。四个全接上，而且能在 GrokBot 里横跨这几个账号来回查，这就是杀手级功能。真心谢谢 Cursor/xAI 团队做了这件事。哪怕你们只做了这一件，我也已经很满足了。但他们不止做了这一件。GrokBot 出厂就带了海量的 plugin，基本就是原来 Cursor 里能用的那一批。所以我装了几个常规的效率类插件，装了 GitHub，还装了 ChatGPT 插件。说白了，我一直是把 GrokBot 当成一个「穿梭并使用我自己数据」的工具在用。所以它有点像一个花哨版的 MCP 客户端，但它确实带了点特别的东西：每一个 GrokBot 都自带一台电脑。你可以真的点进那台虚拟机，这就是 Proddy McProd 的电脑。看着没什么意思，但它能用 Chrome、能用终端，里面还有一些文件。所以你看，这又有点像一个轻量版的 OpenClaw：有虚拟机，虚拟机能上网，还带着 connector 和 MCP。
[6:41] Claire Vo
Now, what do I think of the experience using Grokbot? Well, I will say setting up a Grokbot super simple. You can simply create a bot. It will start up that machine. It will start up a bot, and then you can just tell it what to do. So, this one just got stood up. What's the first thing I want you on? And I say, I want you to be a family manager bot that keeps track of my personal email and personal calendar. Can't spell calendar. And again, it's going to go ahead and say, "Great, I'm going to find all your connectors. I'm going to set it up, and I'm going to learn myself." Really, really simple, streamlined onboarding experience, and it's going to go ahead and give me pretty good responses back and do work on my behalf. It's pretty chat-based, nothing fancy, but it's it's genius is in its simplicity. Now, what I don't like about Grokbot? Well, I don't like that it's so simple and I don't like that I can't hack it. If you look at my open claws, they are chaotic and technical and impossible to use and high maintenance. And I love them. I love them. Um Grok Bot, I have a little bit less control over. I manage them a little bit less tightly and so I love them less. You know, they're like my children.
那我用下来对 GrokBot 的感受是什么？我得说，建一个 GrokBot 简单到不行。你直接点「创建一个 bot」，它就把那台机器启动起来，把 bot 拉起来，然后你只要告诉它干什么就行。比如这个刚建好，它问我「第一件事想让我做什么？」我说：我想让你当一个家庭管家 bot，帮我盯着我的私人邮箱和私人日历。（calendar 我都拼不对。）然后它就会说：「好的，我这就去找你所有的 connector，把它们配上，然后自己学起来。」上手流程真的特别简单顺畅，之后它给我的回复质量不错，也确实能替我干活。整体就是聊天式的，没什么花哨的东西，但它的高明之处恰恰就在于简单。那我不喜欢 GrokBot 什么呢？我不喜欢它这么简单，我不喜欢我没法折腾它。你看我那几个 OpenClaw，一个个混乱、技术味重、难用得要命、还得天天伺候。可我爱它们，是真爱。而 GrokBot，我对它的掌控少了一些，我打理它们打理得没那么勤，所以我也就爱得少一些。你懂的，它们就像我的孩子一样。
[8:11] Claire Vo
The child that the joy is in the challenge and I will tell you Grok Bot has not challenged me because it works simply and it works out of out of the box. The other thing I don't like about Grok Bot is you don't get to pick what model you're working with. You really actually don't get to have that sort of like soul.md control over what Grok Bot does, how it's configured, etc. And you know me, I just like to craft as if out of clay my agents. So, I like that experience. I don't think everybody else does. So, if you're looking for that like highly tunable, highly hackable, highly transparent setup for your agent, Grok Bot is not that. Open Claw or Hermes or any of those is is a little bit more like that. And then of course, it doesn't run on my machine. I don't have control over it. It's a third-party system. Some good, some bad. That being said, it works and the connector and kind of like UX affordances are quite nice. I would say just related to that, the last thing that I don't love about Grok Bot is whatever model it's using and maybe it's Grok, maybe it's Sounds a lot like Claude Slop. Is I just no vibes. Just no vibes. The vibes are bad. It's like I don't want to hang out with this person. Doesn't really make me cheerful.
养孩子的乐趣就在那份挑战里，而我可以告诉你，GrokBot 没给我带来任何挑战——因为它很简单就能跑起来，开箱即用。我不喜欢 GrokBot 的另一点是：你没法选它用哪个模型。你其实也拿不到那种 soul.md 级别的控制权，去决定 GrokBot 做什么、怎么配置，等等。你们了解我，我就喜欢像捏泥巴一样亲手把我的 agent 一点点捏出来。我喜欢那种体验，但我不觉得所有人都喜欢。所以如果你想要的是那种高度可调、高度可折腾、高度透明的 agent 配置，GrokBot 不是。OpenClaw 或者 Hermes 那一类会更接近一点。当然还有一点，它不跑在我自己的机器上，我控制不了它，它是个第三方系统。有好处也有坏处。话说回来，它是能用的，而且 connector 和整体的交互设计（UX affordance）都相当不错。跟这个相关的，我最后一个不太喜欢 GrokBot 的地方是：不管它底下用的是什么模型——可能是 Grok，也可能是……听起来非常像 Claude 那股「AI 味」——总之就是没气质，就是没气质，气质很差。那种感觉就像「我不想跟这个人待在一块儿」，一点都逗不乐我。
[9:36] Claire Vo
I have tuned my open claws to be exactly who I want to talk to, when I want to talk to them. This I'm getting like too much of that like classic slop, not this, not that slop, terrible names for products like a scope knife. Um just stuff that I would never say, and so I think the personality tuning and the voice tuning across all these harnesses are really important. But they're even more important if you are going to do this multi-agent strategy where you're going to have people name their agents and have them do work for you. A couple use cases that I think you could use Grokbot for, again, I've made I like to make my agents have roles. So, I made a PM one, I made a data analyst one, I made a deal desk one, I made a finance one, I'm making a family one. I think just thinking that way, what kind of teammate do I want, and then giving that teammate the tools it needs, and getting it the instructions it needs is a great way to design your agents. Now, I would demo more, but I actually think this is all about workflow. So, the product itself is super simple. I think the use cases are what make it interesting, and then the connector experience being top-tier, I think it's going to help the Cursor XAI SpaceX team get their claws in the enterprise use case. So, I'm really curious, drop in the comments, tell me what your use cases are. Um I will continue to share ones, and I will let you know if I decide I'm going to migrate any of my open claws over to Grokbot. If you want me to move over, SpaceX team, just make Grokbot way harder to manage, and then apparently I will fall in love with it and never use anything else. I'm going to keep an eye on this, but TLDR, super simple, great connectors, use your agents as employees, and if you like a bad time, stay with stay with open claw. This episode is brought to you by Jira by Atlassian. The teamwork graph in Jira delivers 44% more accurate agent results
我那几个 OpenClaw，我已经调教成了我想聊天的时候正好想聊的那种人。而这个，我收到的全是那种经典的 AI 腔套话——「不是这个，而是那个」的腔调，还有给产品起的一堆烂名字，比如什么「scope knife（范围刀）」。反正都是些我这辈子都不会说的话。所以我觉得，在所有这些 harness 上，人格调教和语气调教都特别重要。而如果你要走多 agent 这条路——让用户给自己的 agent 起名字、让它们替自己干活——那这件事就更加重要。说几个我觉得可以拿 GrokBot 来做的场景。前面说了，我喜欢给我的 agent 定角色。我做了一个产品经理的，一个数据分析师的，一个 deal desk（销售报价审批）的，一个财务的，现在在做一个管家庭的。我觉得，先想清楚「我想要一个什么样的队友」，然后把这个队友需要的工具给它、把它需要的指令交代清楚——这是设计 agent 非常好的思路。我本来还能再多演示一些，但我其实觉得这东西的关键全在工作流上。产品本身极其简单，真正让它有意思的是使用场景；再加上顶级的 connector 体验，我觉得这些会帮 Cursor/xAI/SpaceX 这帮人把爪子伸进企业级市场。所以我特别好奇，评论区告诉我你们的使用场景是什么。我也会继续分享我自己的，如果我决定把哪个 OpenClaw 迁到 GrokBot 上，也会告诉你们。SpaceX 团队，你们要是想让我搬过去，就把 GrokBot 做得难管理一万倍，那看样子我就会爱上它、再也不用别的了。我会持续盯着这个产品。一句话总结：极其简单、connector 一流、把你的 agent 当员工用；如果你就喜欢给自己找罪受，那就继续留在 OpenClaw 吧。本期节目由 Atlassian 旗下的 Jira 赞助。Jira 的 Teamwork Graph（团队协作知识图谱）能让 agent 的结果准确率提升 44%，
[11:32] Claire Vo
with 48% less token usage. That's a huge difference when working with AI coding agents like Claude, Cursor, Codex, or Copilot. The hardest part of shipping with AI isn't the code, it's [music] the context. What's the right ticket? What did the specs say? What got decided in Slack? The Teamwork Graph pulls all of that from across your entire stack from Jira and Confluence to GitHub and feeds it directly to your agents before they write a single line. You assign the work, the agent gets [music] everything it needs, and a PR surfaces when it's ready. No digging, no context switching, no broken flow. Same team, smarter agents. Try it free at jira.dev.
同时把 token 消耗降低 48%。在跟 Claude、Cursor、Codex 或者 Copilot 这类 AI 编码 agent 打交道时，这个差距是巨大的。用 AI 交付东西，最难的部分从来不是代码，而是 [music] 上下文。该看哪张 ticket？spec 里到底是怎么写的？Slack 上最后拍板的是什么？Teamwork Graph 会把这些信息从你整个技术栈里捞出来——从 Jira、Confluence 一直到 GitHub——在 agent 写下第一行代码之前就直接喂给它。你派活，agent 就拿到 [music] 它需要的一切，活干完了自动冒出一个 PR。不用再翻箱倒柜，不用切上下文，心流不会被打断。还是原来那个团队，但 agent 变聪明了。上 jira.dev 免费试用。
[12:15]
[music]
[music]
[12:15] Claire Vo
That's j i r a .d e v. Okay, the next release from the Cursor team is Origin, the GitHub replacement announced at Cursor's conference this year and finally released in early access today, which will be a couple days ago when we get this episode live. So, Origin is an early access beta and they've called it kind of code base inside the app. So, if you're in the Cursor web app or you're in the desktop app, you're really talking about code base, which is of course the right primitive for something that is code hosting. Now, the whole value proposition that Cursor has put in front of us in terms of why they should build a GitHub replacement is they are basically building a agent native GitHub replacement, which means that it's got all the get primitives, it's got code, it's got diffs, it's got pull requests, it's got all that same get primitives stuff, but it's built in a UI way, it's built in a UX way for agents to collaborate with and in particular, of course, the Cursor cloud agents and the cursor desktop app. And and and the cursor CLI. So, it's it's Git. We're all stuck with Git. We've agreed on Git. Now, have we agreed on GitHub? And it seems like the tides are turning. But, Cursor's bet is that we are going to want a agent-native GitHub, and that GitHub is not going to get there fast enough. So, what does that look like tactically? Well, Origin has free posts.
就是 jira.dev。好，Cursor 团队的下一个新东西是 Origin——他们今年大会上官宣的 GitHub 替代品，今天终于开放了 early access（早期体验）；等这期节目上线的时候，大概就是几天前的事了。Origin 现在还处在 early access 的 beta 阶段，他们把它叫做「应用里的 code base（代码库）」。所以不管你是在 Cursor 网页版还是桌面端，你打交道的其实就是 code base 这个概念——对一个代码托管产品来说，这当然是最合适的基本单元。Cursor 摆在我们面前的整套价值主张，也就是他们为什么要做一个 GitHub 替代品，核心在于：他们要做的是一个 agent 原生（专为 AI 智能体协作而设计）的 GitHub。意思是 Git 那套基本能力它全都有——代码、diff（改动对比）、pull request，Git 该有的一样不缺——但它的 UI 和 UX 是按「让 agent 参与协作」的思路做的，尤其是要配合 Cursor 的云端 agent、Cursor 桌面端，还有 Cursor CLI（命令行工具）。所以说到底还是 Git。我们都被 Git 绑住了，大家在 Git 这件事上已经达成共识了。但我们在 GitHub 上达成共识了吗？看起来风向正在变。而 Cursor 赌的是：我们迟早会想要一个 agent 原生的 GitHub，而 GitHub 自己走到那一步不会够快。那具体落到实处是什么样？首先，Origin 里可以建仓库（repo），而且是免费的。
[13:53] Claire Vo
You can, of course, import GitHub repos. I'll tell you a little bit about my experience there. It has pull requests, which you can see. Um agents can reply to comments. They can be assigned as reviewers. All those sorts of things. And there's a small set of extensions, um CI/CD extensions, for example, like build preview branches in Vercel for those who want it. Okay, so what does that actually look like? Well, you go into the Cursor web app, and you click code base, and then you basically import your GitHub repos. Very smart. Now, it was not smart to release this the day that GitHub had a major outage, or maybe it was genius to release this the day that GitHub had a major outage. So, I had some trouble with this GitHub import, but all you do is click sync from GitHub, you authorize your GitHub account, pick which repositories you want to sync in, and then they are here. As you can see, you have kind of the same primitives. You have code. So, this is your code view. You have pull requests. So, you have pull requests, and then you have some settings. What I would say is with the GitHub integration, it's really it seems like it's just a wrapper on the GitHub API. So, yes, in theory, this has been kind of redesigned. You can see here, like, there's some nice things about how bug bot feedback has been shown, or how Cursor feedback has been shown. It has some like smart little suggestions at the top that are are nice. But at the end of the day, this is this is just like GitHub with maybe less features, late and maybe less integrated. Much more integrated with cursor, so I appreciate that and I can see the vision. You can see the vision, I promise.
当然，你可以把 GitHub 上的仓库导进来——我待会儿讲讲我自己的体验。它有 pull request，你能看到 PR；agent 可以回复评论，可以被指派成 reviewer（代码审查者），诸如此类。另外还有一小撮扩展，比如 CI/CD 方面的扩展——想用的话可以在 Vercel 上构建预览分支之类。好，那实际用起来长什么样？你进 Cursor 网页版，点 code base，然后把 GitHub 仓库导进来。挺聪明的。不过挑 GitHub 大面积宕机那天发布这东西就不太聪明了——当然，也可能反而是神来之笔。总之我导入 GitHub 的时候确实碰到了点麻烦。步骤本身很简单：点「从 GitHub 同步」，授权你的 GitHub 账号，选好要同步哪些仓库，然后它们就都在这儿了。你看，基本单元是一样的：有代码，这是代码视图；有 pull request；再往下是一些设置。我想说的是，就 GitHub 集成这块，它看着真就是在 GitHub API 外面包了一层。是，理论上界面重新设计过——你看这里，Bug Bot 的反馈怎么呈现、Cursor 的反馈怎么呈现，确实做得挺好；顶部还有几条聪明的小建议，也不错。但说到底，这就是个功能可能更少、上线更晚、集成可能更弱的 GitHub。当然它跟 Cursor 的整合要深得多，这点我认可，我也看得到那个愿景。你也能看到那个愿景，真的。
[15:42] Claire Vo
But right now, if I'm using a GitHub hosted repo and I'm not ready to go like all in on the cursor ecosystem, there's just not enough here in this early launch to get me to pull over. And so, yes, I can like at cursor here and ask it to fix the Vercel preview branch issue. Please, why is it failing? Cuz this is failing. I can do that. Nope, we got an error. So, maybe I can't do that. Um GitHub's having an outage right now, so maybe this is a GitHub issue, maybe this cursor issue. But again, like today's today origin, I'm seeing that they're setting the groundwork for getting you to get your repos into cursor. They're setting the groundwork for Bug Bot and Cursor to be first-class citizens in your repo. I'm sure there's some affordances on like the CLI side that make it easier for cloud agents to open PRs and it's just going to be just like a nice buttoned-up experience. But again, if you have invested deeply in GitHub, which everybody has, and you have your automations, and you have your actions, and you have your code owners, and you have all this stuff, we're going to have to see more from this experience, more from why AI native, why agent native matters for us all to move over to this new product.
但眼下，如果我的仓库托管在 GitHub 上，我又还没准备好把身家全押进 Cursor 这套生态，那这次早期发布给的东西还不足以让我搬家。是，我可以在这儿 @cursor，让它去修 Vercel 预览分支的问题——「拜托，这个为什么挂了？」因为它确实挂了。我可以这么干。……不行，报错了。看来我干不了。呃，GitHub 现在正在宕机，所以这可能是 GitHub 的问题，也可能是 Cursor 的问题。但还是那句话，今天这个 Origin，我看到的是他们在铺路：先让你把仓库搬进 Cursor，先让 Bug Bot 和 Cursor 在你的仓库里成为一等公民。我相信在 CLI 那一侧也有一些便利设计，能让云端 agent 更容易开 PR，整个体验会很妥帖、很顺手。但话说回来，如果你已经在 GitHub 上投入很深——所有人都投入很深——你有自动化、有 actions、有 code owners（代码归属人），有一整套家当，那我们还得看到这个产品拿出更多东西，看到更多「为什么 AI 原生、为什么 agent 原生对我们真的重要」的理由，才可能集体搬过去。
[17:16] Claire Vo
Now, literally came out today. They're not even calling it beta. They're calling it early beta. So, I have faith. I'm excited to see what this looks like, but again, just the current GitHub integrated sync, it's like a little slower. I get the redesign, but it's not giving me life. And I'm just too embedded in the GitHub ecosystem. I need to see something wow from this in order for me to click unlock. You know, and I and I've spent an hour with it, 2 hours with it, not that much. I've messed around with opening PRs inside Cursor. Again, like it's redesigned. It's synced to GitHub though, so it kind of feels like I'm being slow migrated over to Cursor and that they just want me to get like that they just want to get their claws in me and then really wow me with this experience, but out of the box, I'm not yet feeling the vision, although I'm excited to see how this plays out. It's worth experimenting with. I think it's worth keeping your eye on. It's interesting. No one is happy with GitHub right now. The stability is like terrible and they really haven't yet given us that like AI native completely reimagined get experience that we're hoping for. That being said, it is super early on Cursor Origin. I think we are mostly in the very, very, very early stages of a very, very, very long migration and it will be interesting to see if Cursor becomes the new centralized source of truth for code or if it continues to just operate in the writing code agents space.
这东西字面意义上就是今天才出来的。他们连 beta 都不叫，叫 early beta（早期测试版）。所以我是有信心的，我也很期待它最后能长成什么样。但眼下这个跟 GitHub 同步的集成，就是慢了点。重新设计我 get 到了，可它没让我惊艳。而我在 GitHub 生态里陷得太深了，得看到点「哇」的东西，我才会点下那个「解锁」。我也就跟它待了一两个小时，不算多。我在 Cursor 里试着开了几个 PR，还是那句话，界面是重做过的，但它是跟 GitHub 同步的，所以感觉像是我正被一点点迁移到 Cursor 上去——像是他们想先把爪子勾住我，然后再用体验把我镇住。但开箱这一刻，我还没感受到那个愿景，尽管我很想看看后面怎么演。它值得试一试，我觉得也值得持续关注，挺有意思的。现在没人对 GitHub 满意，稳定性差得一塌糊涂，而且他们确实还没给出我们期待的那种 AI 原生、彻底重新想象过的 Git 体验。话虽如此，Cursor Origin 真的还非常早期。我觉得我们现在处在一场非常非常非常漫长的迁移的非常非常非常早期阶段，接下来就看 Cursor 会不会成为代码新的中心化「事实来源」，还是继续只在「写代码的 agent」这个赛道里玩。
[18:52] Claire Vo
All right, so those are our two products. I would say I like Rock bot. I have not yet seen the light on Origin. The last thing we're going to talk about is Grok 4.6. This is the model that has everybody I know texting me saying, "I think I kind of like the Grok models." Now, I think we're all trying this because Grok has defaulted. I've noticed some of their models to Grok. Um not sure if I love it, but that means we give it a try. And you know, we're going to put benchmarks aside. Everybody picks the benchmarks that they like. I want to talk about the How I AI vibe review. For anybody who is new, whenever a new model comes out, I do a couple things. I test PRDs, I test prototypes, I test design, wireframes, technical changes, and if I like talking to the model. I run all these evals blind, so I run them against models, and then I actually go through and grade all of these designs live myself. I've made some changes to the How I AI bench. In particular, this design one where I let the model decide how it wants to redesign the page. And I've made a eval for a very complicated claims adjudication wireframe um that has a lot of complex interactions. So, those were two new things that we added to the How I AI vibe bench. And then we ran Grok 4.6 through these models. I graded them.
好，这就是我们要讲的两个产品。总结一下：GrokBot 我喜欢；Origin 我还没被点亮。最后要聊的是 Grok 4.6。就是这个模型，让我认识的一堆人给我发消息说：「我好像有点喜欢 Grok 这几个模型了。」大家都在试它，我猜是因为 Cursor 把一部分地方默认切成了 Grok——我注意到他们有些模型默认就是 Grok 了。呃，不确定我喜不喜欢这个安排，但至少它逼着我们去试一试。另外，benchmark（跑分）我们先放一边，谁都会挑对自己有利的那套跑分。我想讲的是 How I AI 的「体感测评」。第一次看的朋友说明一下：每次有新模型出来，我都会做这么几件事——测 PRD（产品需求文档）、测原型、测设计和线框图、测技术改动，再看看跟这个模型聊天舒不舒服。所有 eval（评测）我都是盲测跑的：先在各个模型上跑一遍，然后我自己一个个把这些设计当场打分。我对 How I AI 这套评测也做了些调整，特别是加了一道「设计」题：让模型自己决定要怎么重新设计这个页面。我还加了一个非常复杂的理赔审核线框图 eval，里面有很多复杂交互。这就是我们给 How I AI 体感基准新加的两项。然后我们把 Grok 4.6 也拉进来跑了一遍，我打了分。
[20:24] Claire Vo
AI judges them. I judge them. And then we put them together in a presentation that I show with you before I even know what the scores are. So, let get to the How I AI clear-weighted index of Grok 4.6 compared to other models. All right, so here are the results. Now, remember, this is my benchmark, so I get to decide how we grade things, and so it is 70% my taste, 30% our LLMs a judge taste. And interestingly enough, Grok 4.6 is right up there with my my favorite, 5.6 Soul, beating out both Sonnet 5 and Opus 5 on the clear index. I am not surprised by this. I do love GPT 56 whole. It's my default. But I am surprised to see 46 up here. I didn't hate it. Now, if you look at the overall recommendation by task, you'll see that 5.6 is my favorite direct writer on PRDs. I just like the clean way of writing. It's very matter-of-the-fact. It's very comprehensive, a little bit technical, but not overly complex. I also like a GPT 5.6 prototype. And so, um I rated those quite highly. Opus 5 got graded by the LLM as a judge as the best implementer for a technical, like bug triage problem. And then, absolutely zero surprise, I continue to like Sonnet 5 for an open Claude agent chit-chat experience. This always wins. I know Sonnet 5 and 4 very pithy AI back-and-forth open Claude model. I just have not found anything better. Now, if we look at my opinions about prototypes, it's really interesting to see that when following art direction, I like 5.6, but when given broad decisions to make its own design decisions, I like Grok 4.6. Now, I suspect that this is because I can just spot GPT and Claude slop a mile away. A mile away.
AI 打一遍分，我打一遍分，然后把结果做成一份演示，而且是在我自己都还不知道分数的情况下当场跟大家一起看。那我们就直接看 How I AI 的「Claire 加权指数」里，Grok 4.6 跟其他模型比是什么水平。好，结果在这儿。记住，这是我的基准，所以怎么打分我说了算：70% 是我的口味，30% 是我们「LLM 当评委」的口味。有意思的是，Grok 4.6 直接冲到了跟我的心头好 GPT-5.6 齐平的位置，在 Claire 指数上把 Sonnet 5 和 Opus 5 都压下去了。前面那点我不意外，我确实很爱 GPT-5.6，它是我的默认选择。但看到 4.6 站到这个位置我是意外的——我一点都不讨厌它。再看按任务给的总体推荐：写 PRD 我最喜欢 GPT-5.6 那种直给的写法，干净、就事论事、覆盖全面，带点技术味但不至于过度复杂。原型我也喜欢 GPT-5.6，所以那几项我打分很高。技术类问题——比如 bug 分诊——LLM 评委把 Opus 5 评成了最佳「实现者」。然后毫无悬念，OpenClaw 那种 agent 闲聊体验，我依然最喜欢 Sonnet 5。这一项永远是它赢。Sonnet 5 和 4 就是那种言简意赅、你来我往的 OpenClaw 型模型，我到现在也没找到更好的。再看我对原型的评价，很有意思：如果给了明确的美术方向，我更喜欢 GPT-5.6；但如果放手让模型自己做设计决策，我更喜欢 Grok 4.6。我猜这是因为 GPT 和 Claude 那股「AI 味」我隔着一里地就能闻出来。真的，一里地。
[22:41] Claire Vo
GPT 56 loves forest green. Claude loves a brown, tan, orange combo. So, I think I was just enjoying this breath of fresh air that was Grok 4.6. That was not those two things. But in terms of executing complex UIs, I still really like 56 whole. And when following art direction, I still like 5-6 Soul. Okay, so let's see what it actually shipped. So, first question I was trying to answer in this benchmark is can it follow explicit design instruction? And GPT 5-6 Soul and Grok both followed my instructions well and did a relatively good non-slop job of following instructions. GPT 5-6 just continues to crush at this doc routing app that I have it make where it has to give a very complex but easy to parse at a glance UI for managing I guess like container ships. Crushes, it always does better than the competition, but I think that Grok did a nice job with my technical incident triage app as well as this editorial page. Now, when it's not given any design, we saw that every model did pretty good. We saw a couple good designs out of 5-6. We saw some clean designs out of Claude Sonnet 5s, and my favorite one was actually this design from 4-6, which is Claude Sonnet adjacent. We see an orange, we see a brown, but was actually quite cute and a good um interactive kind of ordering system for a coffee shop. So, I was really pleased with that one.
GPT-5.6 特别爱用森林绿，Claude 特别爱那套棕色、驼色加橙色的组合。所以我想，我只是在享受 Grok 4.6 带来的那口新鲜空气——它不是那两种味儿。但要说执行复杂 UI，我还是更喜欢 GPT-5.6；要说跟着美术方向走，我也还是更喜欢 GPT-5.6。好，那我们看看它到底交出了什么东西。这套基准我想回答的第一个问题是：它能不能听懂并执行明确的设计指令？GPT-5.6 和 Grok 都把我的指令执行得挺好，而且做得相对不「AI 味」。GPT-5.6 在我让它做的那个「单据路由」应用上一如既往地碾压——那个界面要非常复杂，但又得一眼就看明白，用来管理集装箱货轮之类的东西。它一直比对手做得好。不过我觉得 Grok 在我的「技术事故分诊」应用上做得不错，这个编辑类页面也做得不错。而完全不给设计方向的时候，我们看到每个模型都挺不错：GPT-5.6 出了几个好设计，Claude Sonnet 5 出了几个很干净的设计，而我最喜欢的其实是 4.6 的这个——风格挺「Claude Sonnet」的，有橙色、有棕色，但真的挺可爱，是个交互做得不错的咖啡店点单系统。这一个我很满意。
[24:26] Claire Vo
Dense information architecture, I of course loved 5-6. I think it's the best at complex UI and designing things that are not overwhelming, but are technically complete. And then in terms of just like taking a generic design and making it little better, I saw some good results from Opus and Sonnet. But when you aggregated all this up, all the ones I scored, good and bad, the two that bubbled up to the top were again 5-6 soul as well as 4-6. Now, what's really funny about this? Again, this is 70% my taste, 30% the LLM is a judge. If you take my taste out of it, the LLM hates Grok and it loves Claude Opus, it loves Sonnet, and it does not like GPT 5-6 soul. Again, this is not weighted by an Anthropic model. I actually use GPT 5.5 to grade all these things because it's actually the harshest judge. Um so, while the models still like Claude, I, Claire, like GPT 5-6 and I like Grok Grok 4-6 for design. Now, this isn't my experience talking to the model, although you all know I love to do that. Um this is pure outputs. So, all I will say is my conclusion here is the Grok model is a competitor. None of it was terrible. I liked a lot of it compared to the models that you would use out of the box. And combined with Cursors, Harness, its investment in new ways of thinking about things like Git, and fun products like Grokbot, I would say it's time for all of us to maybe become some level of a Grok boy. I am so excited about all this new stuff. It's really fun to look at. Again, true transparency, I'm still spending a lot of my time in Codex and spending a lot of time with the 5-6 models. But, given this experience, I'll think about spinning up Cursor a little bit more for coding and see if I can continue to like and get value out of Grok Grok Grok.
信息密度高的界面架构，我当然还是最爱 GPT-5.6。我觉得它做复杂 UI 最强，能把东西做到既不让人眼花，又在技术上完整。至于「拿一个普通设计稍微改好一点」，Opus 和 Sonnet 都给了不错的结果。但把所有东西汇总起来——我打过分的、好的坏的全算上——冒到最上面的还是那两个：GPT-5.6 和 Grok 4.6。有意思的地方来了：还是那句话，70% 是我的口味，30% 是 LLM 评委。如果把我的口味剔掉，纯看 LLM 评委：它讨厌 Grok，爱 Claude Opus，爱 Sonnet，而且不喜欢 GPT-5.6。注意，这个权重并不是由 Anthropic 的模型给出来的——我其实是用 GPT-5.5 给这些东西打的分，因为它才是最狠的评委。所以呢，模型们还是偏爱 Claude，而我，Claire，喜欢 GPT-5.6，设计这块我也喜欢 Grok 4.6。当然，这次讲的不是我跟模型对话的体验——虽然你们都知道我很爱聊这个——这次纯看产出。所以我的结论就一句话：Grok 这个模型是有竞争力的。没有哪一项是烂的，跟你平时开箱即用的那些模型比，它很多地方我都挺喜欢。再加上 Cursor 那套 harness（跑模型的外壳/框架），加上他们在 Git 这类事情上重新思考、真金白银地投入，再加上 GrokBot 这种好玩的产品，我想说：也许是时候，我们都该多少变成个「Grok 粉」了。这一堆新东西真的让我很兴奋，看着就好玩。再次坦白，我现在大部分时间还是泡在 Codex 里、用 5.6 系列的模型。但有了这次体验，我会考虑多开一开 Cursor 来写代码，看看能不能继续喜欢 Grok、能不能从 Grok 身上持续榨出价值。Grok、Grok、Grok。
[26:32] Claire Vo
Thanks for joining this mini episode of How AI. I want to hear from you whether you love Grokbot, if you think Cursors going to replace GitHub, and whether or not you've switched to Grok for any of your coding tasks. See you soon. Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiai.com. See you next time.
感谢收看这期 How I AI 的迷你特辑。我想听听你们的想法：你喜不喜欢 GrokBot？你觉得 Cursor 会取代 GitHub 吗？还有，你有没有在某些编码任务上换用了 Grok？回头见。非常感谢收看。如果你喜欢这个节目，请在 YouTube 上点赞订阅，或者更好——给我们留言说说你的想法。这档播客在 Apple Podcasts、Spotify 或者你常用的播客 App 上都能找到，也欢迎给我们打分写评价，能帮更多人发现这个节目。所有往期内容和节目介绍都在 howiai.com。下次见。