从零设计 Perplexity 的 AI 浏览器 Comet

想象一下，你以设计师身份加入 Perplexity，入职第一天设计副总裁就把全新 AI 浏览器 Comet 的钥匙交给你。

Imagine you join Perplexity as a designer and the very first day the VP of design hands you the keys to their all-new AI browser Comet.

我的第一次会议就是，嘿，我刚开始做这个浏览器的早期阶段。你想不想直接接手这个项目？

My first meeting was like, hey, so I started very early stages of this browser thing. Do you just want the keys to like own this?

你会如何处理设计系统和产品策略？关键决策如何获得支持？

How would you approach the design system, product strategy? What about getting buy-in on key decisions?

我的第一反应是，好吧，我要翻遍 Figma 文件，找到这个设计系统。找到那些我可以用来快速构思的组件。然后很快发现，根本没有设计系统。

My first instinct was all right, I'm going to go through Figma files and find this design system. I'm going to find the pieces that I could use to just start riffing on ideas. And quickly found out there is no design system.

如何打造一种能让人产生共鸣的体验？

How do you craft an experience that makes people feel something?

AI 和工具很有意思，因为现在还太早期。我们正在摸索模式。

AI and tooling is interesting because it's so early. We're figuring out the patterns.

欢迎来到 Dive Club。我是 Rid，这里是设计师永不止步的学习之地。本周的嘉宾是 Escha Vera，她将分享设计 Perplexity 全新 Comet 浏览器的历程。我最喜欢的部分之一，是她如何将各种不同的 AI 工具整合在一起，为用户体验做出超越像素的贡献——比如添加音乐、生成邀请卡等等。那么，让我们直接深入她 Perplexity 之旅的起点。

Welcome to Dive Club. My name is Rid and this is where designers never stop learning. This week's episode is with Escha Vera and she's going to share the journey of designing the all-new Comet browser from Perplexity. And one of my favorite parts is learning how she's stitching together all of these different AI tools to contribute to the user experience beyond just the pixels. Things like adding music or generating invite cards and a lot more. So, let's start off by diving right into the beginning of her Perplexity journey.

我离开 Descript 时，感觉自己的职业生涯可能结束了。我在一个产品上花了太长时间，离开时觉得自己做了件蠢事，而且不知为何感觉自己找不到工作了。所以我花了大约一年时间，和 CEO、其他公司、初创公司交谈，试图弄清楚我想做什么，是否还想留在 AI 领域，或者是不是该远离它，回到常规工作中。但我没找到什么感觉对的事。要么是我关于伦理的问题得到了糟糕的答案，要么是他们将 AI 整合到工具中的方式让我不舒服，要么是他们根本没有答案，这让我不安。我很早就从 Descript 认识了 Henry。我联系他，问他是否知道有哪些公司在招人，或者有适合我的机会。我也是这样接触到 Daylight 的。但他只是给我和 Henry 发了一封介绍邮件，说“嘿，创始设计师对创始设计师，你们应该认识一下。”那是在我加入 Perplexity 的一年前。所以我们只是偶尔通信。当时 Henry 专注于组建品牌团队，我说：“嘿，等你关注产品的时候，联系我。”大约一年后，在 Daylight 项目上折腾了一段时间，做了 OS 和 Motif 之后，我重新联系了 Henry。他已经准备好并正在寻找产品设计师。我得到了所有我关心的关于 AI、如何使用 AI 等细节问题的满意答案——这些细节是我作为设计师，愿意把自己的名字放在公司推出的产品背后所看重的。我问 Henry：“嘿，你知道我会做什么吗？有没有什么研究需要我先做？或者我可以在正式入职前开始准备什么？”他说：“嗯，可能是一些桌面端的东西。”我说：“好吧，我不知道那是什么意思。这等于没说。我一直在做各种东西，所以无所谓。我想我只能盲人摸象了。”结果我入职那天，正是他们宣布 Comet 的日子。

I left Descript and I kind of felt like maybe my career is over. I'd spent so long on one product that I kind of left feeling like I did myself a disservice by doing that and I felt unhirable for whatever reason. So I spent like a year just talking to CEOs and other companies, startups, trying to figure out what I wanted to do, if I even wanted to stay in the AI space or if I kind of felt like it was time to distance myself and just go back to the regular grind. And I didn't really find anything that felt good. Either I got bad answers to questions I had about ethics or just like the way that they were approaching the integration of AI into their tooling or whatever it was or they didn't have answers and that made me uncomfortable. So I know Henry early on from Descript. I reached out to him and asked him if there was any companies that he knew that was kind of looking for someone or someone that might be a good fit for me. And that's how I got connected with Daylight as well. But he just sent me an intro email with Henry of like, "Hey, founding designer to founding designer, y'all should know each other." And that was like a year before I joined. So, we just had some casual correspondence. At the time, Henry was focused on building up the brand team and I said, "Hey, whenever you're focused on product, hit me up. Let me know." About a year later, after kind of spinning the wheels on Daylight stuff, working on the OS and working on Motif, I reached back out to Henry. He was ready and looking for product designers. I got all the right answers to the questions that I wanted surrounding like AI, how it's used, just like all of these gritty details that I care about as a designer who's like putting my name behind the products that these companies are pushing out. And I asked Henry, hey, like do you know what I'll be working on? I'm like, "Is there any research I should be doing? Like, is there anything I can get started on like ahead of my actual start date?" And he was like, "Yeah, maybe some desktop stuff." And I was like, "Okay, well, I don't know what that means. That doesn't really mean anything. I've been doing stuff forever, so it's whatever. I guess I'm going in blind." And the day I joined was the day they announced Comet.

哇。

Wow.

所以我直到醒来开始工作，在 Twitter 上看到公告才知道 Comet。我当时想：“哦，这真有意思。”我和 Henry 的第一次会议，他说：“嘿，我刚开始做这个浏览器的早期阶段，但你想不想直接接手这个项目？”那真是太酷了。我的意思是，能接手这样一个产品，我感到非常荣幸。

So, I didn't even know about Comet until I woke up, started working, saw the announcement on Twitter, and I was like, "Oh, that's really interesting." My first meeting with Henry. He was like, "Hey, so I started like very early stages of this browser thing, but do you just want the keys to like own this?" So that was very cool. It was very I mean just so honored to be handed a product like that.

也许我们可以继续聊聊这段经历。你用了“接手”这个词，这可不是小事。第一个月到底怎么样？那是一个很大的适应过程。作为一个新设计师、新团队，你从哪里开始？而且，你现在接手了一个在 Twitter 上备受炒作的大型产品。

Maybe we could just continue on with the journey a little bit. You use the phrase being handed the keys. It's a pretty big deal. What the heck was that first month like? That's a big time ramp up process. Like where do you even start as a new designer, new team, and oh yeah, you've now inherited this massive product that has this whole hype cycle attached to it on Twitter.

从外部看，Perplexity 是一个非常精致的产品。我的意思是，它有一些小毛病、一些 bug，有些 UX 可以改进，但整体感觉不像是有 20 个设计师在设计，而像是一个小团队。有很多连续性和一致性。所以我的第一反应是，好吧，我要翻遍 Figma 文件，找到这个设计系统。找到那些我可以用来快速构思的组件。然后很快发现，至少在 Figma 里没有设计系统。也许这很常见，但我从没在任何这样运作的地方工作过。但代码就是法律，设计系统就在代码里。这里很多设计师实际上直接贡献前端代码。所以这很有趣，意识到 Figma 文件是有的，但每个人都是根据需要自己制作组件来完成设计。你知道，每个人都有品味和常识，他们可以从网上已有的东西中提取，但基本上就是截图，然后基于网上的东西制作组件。这对 Perplexity 核心来说可能没问题，因为它是一个基于 Web 的产品。Comet 则涉及 C++ 代码和 Web 的混合，有两个不同的团队。Web 团队有他们需要的东西，有构建有趣 UI 所需的基础设施，但 C++ 团队什么都没有——他们需要线框图、像素完美的设计、原型。所以我开始自己制作一个设计系统，因为我自己需要一些东西来快速迭代 C++ 和浏览器 Chrome 部分。团队看到后说：“嘿，这太棒了。我们完全可以把它用在 Perplexity 上。”所以这从 Comet 设计系统扩展成了 Perplexity 设计系统，两者之间有了连续性。它们实际上是同一个系统。

From the outside looking at Perplexity, it's a very polished product. I mean, it has some quirks, has some bugs, there's some UX things that could be better, but overall it feels not like 20 designers are designing it. It feels like a small team. There's a lot of just continuity and consistency. So my first instinct was all right, I'm going to go through Figma files and find this design system. I'm going to find the pieces that I could use to just start riffing on ideas and quickly found out there is no design system at least in Figma. Maybe this is common place, but I've never worked anywhere that operated like this. But code is law that the design system is in the code. A lot of the designers here are actually contributing directly to front end. So that was interesting just kind of realizing okay there are Figma files but everyone's kind of just making their own components as they need to fulfill a design you know everyone has taste and just like common sense and they can pull from what is on the web but it's really like screenshotting and then like making the component based on what's on the web and that's probably fine for Perplexity core because it's a web-based product. Comet has overlap between C++ code and the web and it's two different teams and the web team kind of has what they need and they have the sort of infrastructure they need to be able to build out fun UIs but the C++ team didn't have anything they needed mock-ups they needed pixel perfection they needed prototypes and so I started by just making a design system like I made it for myself because I needed something to quickly iterate on the C++ and like the core or browser Chrome pieces and the team kind of saw that and was like, "Hey, this is sick. Like, we could totally use this for Perplexity." So, that kind of exploded the scope of that from a Comet design system to a Perplexity design system that kind of has continuity between the two. They're literally the same system.

这是 Figma 吗？还是代码？还是两者都有？

Is this Figma? Is this code? Is it both?

两者都有。我的大部分工作是在 Figma 中完成的，主要与 Champori 合作，他是这里前端、设计和品味的全明星之一。所以我和他一起在前端方面做了很多工作，实现了对基于 Web 的设计系统的更改。

It's both. So it was majority of like my work was in Figma working primarily with Champori who's kind of one of the all-stars here on front end and just design and taste in general. So I worked a lot with him on the front end side implementing changes to the web-based design system.

你在设计系统方向上走了多远？尤其是考虑到你们要覆盖的界面面积很大，肯定有快速交付真实屏幕的压力。所以我感兴趣的是那个交叉点，即设计系统在多大程度上促进了产出和节奏，而不是过度架构、陷入细节泥潭。

How far did you go into the design system direction? Especially as I'm sure there's probably some pressure to ship real screens quickly given the amount of surface area that you have to take on. So, I'm interested in that intersection point, like where's the sweet spot where the design system is contributing toward the output and tempo of what you're building versus overarchitecting something and getting kind of lost in the sauce a bit.

Figma 设计系统中的大部分内容只是组件的核心原子：按钮组件内使用的文本组件、包含图标组件的列表项组件。我们只构建这些核心组件的原子，这样就不必把更大的部分也组件化。那些部分可以保持灵活，由设计师自行决定如何实现。我们只是根据界面区域建立内边距和间距的原则与逻辑。所以界面区域逻辑已经内置于 Figma 的自动布局中。整个目的是设置组件，让任何设计师都能轻松使用，而无需每次都重新制作。这样就有了一个事实标准，不必在设计文件中翻找，同时提供如何组合设计元素的指导。就我个人而言，我把内边距直接放在组件里，这样你在布局时就不必太担心文本和下方按钮之间的距离。内边距已经免费内置在组件中，并且是 8 的倍数，有时微单元是 4 或 2，但通常是 8、16、24、32、40。我们甚至不需要考虑这些，它已经内置于系统中。显然，在编码和工程端也要做同样的工作。确保令牌命名一致，尽力做到一一对应，这样即使新工程师进来，查看 Figma 设计，检查足够的信息就能知道使用了哪些组件、哪些颜色变量，然后使用设计系统在工程端进行组装。他们甚至不需要担心内边距，这些都已经免费提供。还有另一个部分是原型代码。我个人仍在尝试找出最佳方法。

Most of what is in the design system in Figma is just the core atoms of components: a text component that's used inside of a button component, list item components which have an icon component. Just building out the atoms of those core components so that we don't actually have to componentize the bigger pieces. Those can be left alone and be flexible, up to the discretion of the designer to make it work. Just establishing principles and the logic for padding and spacing depending on the surface area. So there's surface area logic that's baked into auto layout within Figma. The whole point was setting up components that any designer can easily use without having to remake it every time. Having a source of truth for that instead of having to dig around design files to try and find something, and providing the guidance for how to put a design piece together. For me personally, I put padding into the components themselves so that when you're laying out the design, you don't have to worry too much about the distance between a piece of text and a button below it. You're going to get that padding for free baked into that component. And it's going to be divisibles of eight, sometimes four or two for a micro unit, but 8, 16, 24, 32, 40. We don't really have to even think about that. It's just baked into the system. And obviously doing that work on the coding and the engineering end as well. Making sure that tokens are named the same, really doing our best to make it a one-to-one so that even a new engineer could come in, look at a design in Figma, inspect it enough to see what components are being used, what color variables are being used, and then use the design system on their end to put it together. And they don't even have to worry about padding. They get all that for free. There's also another piece of that which is prototyping code. I'm personally still trying to figure that out, like what's the best way to do that.

我试过很多原型工具，都还行。大多数针对移动端交互，对于过渡或状态都还可以，但不太适合桌面环境下的复杂用户流程。所以我发现目前最好用的是 Claude 和 Cursor。你能详细说说这部分流程吗？

I've tried so many prototyping tools. They're all okay. Most of them are geared towards mobile interactions. Pretty much all of them are okay for transitions or states, but aren't great for complex user flows in a desktop environment. So, what I found kind of just is working the best right now is Claude and Cursor. Can you go a little bit deeper on that part of the process for you?

到目前为止，流程是在 Figma 中组合关键屏幕，使用 Claude 或 Cursor 描述我要原型化的东西，利用这些截图来构建初始前端。虽然不完美，但要么足够接近，要么我很容易就能进去修复 HTML、CSS 等，使其正确。这对于需要文本输入、处理带提前输入的建议列表的流程特别有帮助。另外，我还想出了一些巧妙的方法，让工程师更容易接近目标。例如，Comet 中带有旋转行星的引导登录屏幕。那可能应该是一个 3D 对象，最终也会是，但迫于时间和资源，我最终使用了 Perplexity Labs 来制作。我基本上只是拿了一个我们已有的品牌行星纹理，制作了一个无缝版本，然后用 Perplexity Labs 描述：嘿，我想要一个在白色背景屏幕中央旋转的球体，添加一个按钮来切换纹理，添加一些控制来调整光照，再加一个导出 360 度视频的按钮。就这样。我导出视频并裁剪，这实际上就被投入生产了。我只用了大约两分钟。我想这有点 hacky，但确实有效。我一直在学习这些工具的边界和限制。在工具内部摸索边缘在哪里，并尽力推动它。但一旦我理解了它，或者掌握了如何与工具沟通以获得我想要的东西，那么我就会想：好吧，我还能在这个工作流程中加入哪些其他工具，以超越这个单一工具的限制？

So far, the process is putting together key screens in Figma, using Claude or Cursor to describe the thing I'm trying to prototype using those screenshots to actually build the initial front end. It's not perfect, but it's either close enough or it's easy enough for me to go in and fix the HTML, the CSS or something to actually get that right. And that's been really helpful for flows that require text input, dealing with a suggestion list with type ahead. Also, just figuring out quirky ways of getting things closer for the engineers. For instance, the landing screen for onboarding within Comet with the sort of rotating planet. That should probably be a 3D object and it will be eventually, but pressed for time and resources, I ended up using Perplexity Labs to make that. I basically just took a branded texture for the planets that we had, made a version of it that was seamless, used Perplexity Labs to describe like, hey, I want a spinning orb in the middle of the screen on a white background, add a button to be able to swap out the texture, add some controls so I could play with lighting, and add a button to export a 360-degree video. And that was that. I just exported that video and cropped it. And that's literally what ended up being put into production. And I did that in like two minutes. It's kind of hacky, I guess, but it works. And I'm constantly learning what the boundaries and what the limits are for these tools. Like working within the tool to figure out where the edges are and doing my best to push that. But once I feel like I've kind of grokked that or built the skill into how to communicate with the tool to get what I want, then it's like, okay, well, what other tools can I put into this workflow to actually do even more than what I'm limited to within this one thing?

这就是我反复提到的关键词：工作流程。很久以来第一次，感觉我们有很多不同的工具可以拿来，拼接到这个工作流程中，创造出以前……我想到了你做的那些行星。

That's the key word that I keep coming back to is workflow. For the first time in a long time, it does feel like there are many different tools that we can pull from and stitch together into this workflow to create an artifact that previously... I think about those planets that you made.

它们很美，对吧？几年前作为设计师，这绝对超出了我可能交付的范围。我根本不可能不去摆弄那种东西。但它真的很特别，对吧？你发布那周，我在 Twitter 上看到它被称赞的次数，我都数不清了。而且听到那只是你说“如果……”，然后随便拼凑出来的，真是太酷了。

They're beautiful, right? That would have clearly sat outside of my box of potential deliverables as a designer a couple years ago. There's just no way I wouldn't be messing with that type of artifact. And yet, it was really special, right? The amount of times that I saw that celebrated on Twitter in the week that you released it, I lost track. And it's so cool to hear that that was you just saying, you know, what if, just hacking something together.

我不能完全把制作那个纹理的功劳归于自己，对吧？我有点像是拿我们现有的东西，为了那个语境进行编辑和操作。但这也是 Perplexity 的美好之处之一。我可能还会说很多次。品牌团队非常棒。整个就是设计团队。品牌团队和产品团队之间的隔阂只有齐腰高。两个团队之间有大量的协作和沟通。尽管我只是一个产品设计师，但我在品牌方面贡献了很多。有机会以那种方式与品牌设计师合作，而不是仅仅交接“嘿，我们需要一些营销材料”，而是在品牌 Slack 上问：“我在做这个东西。有人有素材吗？有人有想法吗？有人有空和我一起讨论吗？”我的意思是，那是梦想。

I can't take full credit for making that texture, right? I kind of took what we were working with and edited and manipulated it in order to work for that context. But it's one of the beautiful things about Perplexity. Again, I'm probably going to say this a lot. The brand team is so good. It's all just the design team. The separation between the brand team and the product team is a waist-high wall. There's so much collaboration and communication between the two teams. I'm contributing a lot on the brand side even though I'm just a product designer. But having the opportunity to work with brand designers in that way instead of it just being a handoff, "Hey, we need some marketing materials," but pinging the brand Slack and being like, "I'm working on this thing. Does anyone have any assets? Does anyone have any ideas? Does anyone have bandwidth to jam with me on this?" I mean, that's the dream.

我采访完 Fee 后立刻想，天哪，不知道和 Fee 一起工作是什么感觉，因为我确信那样我会看起来像个更好的设计师。

I finished interviewing Fee and was immediately like, man, I wonder what it'd be like to have a Fee to work with because I'm sure that I would look like a better designer as a result.

亲身体验太棒了。不可思议。FE 简直是天赐之物。

It's amazing firsthand. It's incredible. FE is a godsend.

我们来聊聊引导流程。这是我想深入探讨的事情之一，因为它非常有创意，艺术指导等等。显然，品牌团队也发挥了作用。但跟我说说那个过程吧。你从哪里开始思考什么是引导用户的最佳方式？我想唤起什么样的感觉？你是如何最终确定现在这个方案的？

Let's talk about the onboarding a little bit. That was one of the things that I wanted to dig into because it's so creative, art direction, everything. Obviously, the brand team played a role, too. But talk to me about that process. Where did you even start thinking about what is the best way to onboard users? What types of feelings do I want to evoke? How did you land where you landed?

一开始和任何典型的引导设计流程没什么不同，就是弄清楚我们需要从用户那里获取哪些关键信息，以便让他们进入浏览器后的首次体验尽可能好。比如从其他浏览器导入配置文件。从历史记录开始会让你的回答好得多，因为它有所有上下文。初始设置，比如你想给这个配置文件起什么名字？选一个主题色。将浏览器设为默认，这有点难办，要求很高。实际上这是一个非常轻量的操作，但人们不这么认为，但它对留存率和尝试新浏览器的整体体验质量有很大影响。所以只是把这些做好，确保我们没有太多屏幕。你添加的屏幕越多，流失就越多。虽然浏览器可能有点不同，如果你特意下载了浏览器，可能多几个步骤也没问题，但我真的想让它尽可能简洁。一旦我们有了这些，就觉得，好吧，这有点无聊。我可以在样式上玩点花样，但它感觉还是产品引导。我真的很想让它感觉特别，就像对我来说是开箱时刻。

It was no different than any other typical onboarding design process at the start, just figuring out what's the key information that we need to get from the user in order to make their first experience when they actually get into the browser as good as possible. So that's things like importing profiles from other browsers. Starting with history is going to make your answers so much better because it has all of that context. The initial setup like what do you want to name this profile? Pick a theme color. Setting the browser as default, which is kind of a tall order and a high ask. The reality is it's a really lightweight action, but it's not perceived as one, but it makes a huge difference in just the retention and the overall quality of your experience trying a new browser. And so just getting those in place, making sure that we don't have too many screens. The more screens you add, the more drop off you have. Although it might be a little bit different with a browser, like if you've gone out of your way to download the browser, you could probably get away with a few extra steps, but I really wanted to keep it as minimal as possible. Once we had that, it's like, okay, well, this is kind of boring. I can play with the styling a little bit, but it feels like product onboarding. And I really wanted it to feel like something special, like it felt like an unboxing moment for me.

是啊。就是加入音乐这个想法。如果我不去实现它，它就不会发生。带有动画和标志展示的片头序列。我想让它感觉像你启动 Nintendo 64 游戏时那样，就像老式电子游戏，启动时在标题画面之前出现制作人员名单。我想我可能潜意识里受到了《原神》的启发。我其实从来没玩过《原神》，但在我做 Comet 和引导流程时，我女朋友经常坐在沙发上玩《原神》，我能听到声音和配乐。或者有人事后指出，《原神》的标题画面有点让人联想到 Comet 的引导，现在被指出来我也看到了，但那只是潜意识的东西。我想我主要是受到电子游戏和那种体验的启发，有点想在产品设计中看到更多那样的东西，而引导流程是实现这一点的完美方式，无论是通过图像、动画还是音乐和声音。

Yeah. Just the idea to put music in like that. That wouldn't have happened if I didn't just make it happen. The intro sequence with the animation and the logo reveal. I wanted it to feel like when you start up a Nintendo 64 game, just like an old school video game when you start up and you get the production credits before the title screen. I think I was low-key subconsciously inspired by Genshin Impact. I've never actually played Genshin Impact, but while I was working on Comet and onboarding, my girlfriend was often on the couch just playing Genshin Impact, and I could just hear the sounds and the soundtrack. Or someone pointed this out after the fact, but the title screen for Genshin Impact is kind of evocative of the Comet onboarding, and I see it now that it was pointed out, but that was just a subconscious thing. I think I was mostly largely just inspired by video games and that experience and kind of wanting to see more of that in product design, and onboarding was the perfect way to do that, whether that's with imagery or animations or music and sound.

是的，音乐对我来说很重要。我设计很久了，从未在流程中添加过音乐作为交付物，但你感觉到它确实彻底改变了我的初始体验。我记得最近唯一有类似功能的产品是 iOS 上 Bandcamp 的登录界面。在登录界面，他们有一段非常可爱的、类似《超人总动员》风格的主题音乐，而且已经存在好几年了。没人谈论它。我从未见过其他类似的东西。但这只是一个非常棒的时刻。它把原本非常标准无聊的登录界面变成了有点愉悦感的东西。它有独特的身份，我很想看到更多这样的东西。而且对我来说，我对音乐充满热情。我认识很多制作人。我在音乐圈很久了。Comet 的主题曲基本上是我去找一个朋友说：“嘿，做点怀旧的东西。”那就是 Comet 片头主题的提示。所以那首曲子的标题是“在睡衣派对上比别人都早醒来，Wii 还开着”。

Yeah, the music was a big deal for me. I've been designing for a long time and I've never added music to a flow as a deliverable, but you felt it like it definitely transformed the entire initial experience for me. I think one of the only products that in recent memory that has something like that is Bandcamp's login screen on iOS. When you're on the login screen, they have this really cute Incredibles-esque sounding theme music and it's been there for years. No one ever talks about it. I've never seen anything else like it. But it was just this really neat moment. It turned what was otherwise a very standard boring login screen into something that had a little bit of delight. It had a distinct identity and I would just love to see more of that. And I think for me, I'm passionate about music. I know so many producers. I've been in the music scene for so long. The Comet theme was basically me going to one of my friends and being like, "Hey, make something nostalgic." That was the prompt for the Comet intro theme. So, the title of that track is "Waking Up Before Everyone Else at the Sleepover, and the Wii Is Still On."

我喜欢这个。

I love that.

而且它非常契合。制作人的名字是 Nesso。我必须提一下他。他太被低估了。他非常……在节目笔记里。他能驾驭任何风格。能把他拉进 Perplexity 的一些事情里真是太棒了。漫画游戏的所有音效都是 Nesso 做的。直播的一些曲目要么是 Nesso 的原创，要么是和我合作的其他制作人的，但这就是 Sunno 发挥作用的地方。

And it fits that so well. The producer's name is Nesso. I have to give him a shout out. He's highly slept on. He's so... in the show notes. You can nail any genre. And it's been awesome being able to pull him into Perplexity stuff a little bit. All of the sound effects for the comic game, all done by Nesso. Some of the tracks for the live stream are either originals from Nesso or just from some of the other producers I work with, but that's where Sunno comes in.

所以就像……它刚出来时我试过，然后有点搁置了，再也没碰过。后来我有点好奇，又捡起来看看有什么新东西。它太不可思议了。没有足够多的人在谈论它有多好。它对于某些没人谈论的具体事情很有用。比如混音人声或使用 Splice 的人声，这些通常不是你真正想用的，但却是你手头有的。

So like is something that like... I tried initially when it came out and then I kind of put it off and never touched it again. And then I kind of was curious and just picked it back up seeing what else is new. And it's incredible. Not enough people are talking about how good it is. It's good for specific things that no one is talking about. Remixing vocals or taking splice vocals, which are often not really what you want to use, but it's what's available to you.

用 Suno 作为输入来重新混音那些人声或改变歌词，或者把人声旋律变成乐器。如果你喜欢某个人声的节奏或韵律，把它变成萨克斯风非常容易。这和任何其他工具一样：你用得越多，就越能掌握如何与它沟通，从而得到你想象中的东西。我们现在正在制作一张 EP，完全是我用他的样本生成音乐，发给他，然后他再根据这个灵感从头创作。如果没有尝试这种工作流程，我们永远不会做出这样的作品。你好像有音乐背景，但很多听众可能没有。所以这只是一个机制，让你可以通过音频表达自己，否则你根本没有权利或能力做到。我觉得这又回到了我们之前讨论的：设计师能带来的可能性正在扩展，真正拥有体验意味着什么？你带来了行星的视频，带来了这段音乐，所有这些都与你在前 20 秒内用 Comet 唤起的感受紧密相连，但它不在 Figma 里，不在像素里，但你设计了它，太酷了。

Using that as an input through Suno to actually remix those vocals or change the lyrics, or like turning a vocal melody into an instrument. If you have a flow or a cadence in some vocal that you like, it's so easy to like turn that into a saxophone. It's the same with like any other tool. The more you work with it, the more skill you build in how to communicate with it in order to get what you're envisioning. And we are kind of like working on an EP right now that's like entirely the result of me generating music using samples from him, sending it to him, and then he's just making something from scratch inspired by that. Something that like we would have otherwise never made if we weren't experimenting with this type of workflow. And it sounds like you have a music background, but a lot of people listening probably don't. So it's just this mechanism that you can use to express yourself through audio that you otherwise would have no right or ability to do. You know, I think it just goes back to what we were talking about earlier where we're seeing this expansion of the box of what a designer can potentially bring to the table and what it means to truly own the experience, right? Like you brought in the video of the planets. You're bringing in this music here and like all of that is so tied to the feeling that you were able to evoke with Comet in the first like 20 seconds and yet it's not in Figma. It's not in pixels, you know, but you designed it and so cool.

太酷了。接下来能聊聊邀请函吗？我知道那是你从头到尾负责的。我在 Twitter 上看到很多。有什么故事吗？

So cool. Can we talk about the invites next? I know that was something that you owned end to end. I've seen a bunch of them on Twitter. What's the story there?

我们决定一开始只通过邀请制发布，以便进行压力测试，确保一切正常，不会过载。有趣的是，这些反馈显然来自早期用户，而不是普通大众或 Comet 的目标受众。但我们还是决定做邀请制。我们需要一个邀请系统。第一个想法是二维码，那是最简单的，但有点无聊。每个人都做过二维码。虽然有一些巧妙的设计，但感觉还是不够特别。我真的很想让它变得特别，让每个人都兴奋。所以我做了一些尝试。几周前，我们推出了 Perplexify Me，那只是一个实验，一个内部“如果这样会怎样”的想法，后来变成了一个功能：上传头像，然后转换成 Perplexity 品牌风格。

We decided to do an invite only rollout initially just to stress test everything to make sure nothing broke. We didn't get overloaded. It's interesting because that feedback is obviously coming from early adopters and not the lay person or really who the target audience is for Comet. But yeah, so we decided we're going to do invites. We needed an invite system. The first thing was like first pass QR code. That's the easiest thing, but it's kind of boring. Everyone has done QR codes. There's been some clever things, but it still just doesn't feel special. And I really wanted to make it feel special. I wanted everyone to be excited. And so I played around a bit. And a couple weeks prior, we had launched Perplexify Me, which was just an experiment, like an in-house what if idea that turned into a thing, which was just the ability to like upload a profile picture and have it be converted into like a style that is within Perplexity's brand.

我试过自己的头像。很酷。我非常喜欢输出结果。

I did it on me. It was cool. I like the output a lot.

我觉得可以在邀请函上做点类似的事情。所以我基本上拿了一些现有的品牌资产，有的是 fee 或其他品牌团队成员做的，有的是创意大使用 MidJourney 分享的风格代码。我花了一天时间浏览已有的素材、风格代码，在 MidJourney 里尝试各种东西，试图弄清楚如果我们做生成式邀请函，它会是什么样子。试了几种方案，有些因为不可扩展而放弃了，因为可扩展性也很重要。但最终我用那些作为 MidJourney 的输入，得到了一种我觉得既符合品牌又具有可扩展性的风格。我把这些图片导入 Photoshop，编辑构图、图层，调整混合模式，以达到我想要的某种效果，这些效果在 MidJourney 里无法完全实现，或者我在生成迭代时没想到。然后我把这些图片重新输入 MidJourney，得到了一部分我认为可以作为训练数据的好素材。接着我用 Civot AI 来制作模型。我尝试了大约十几个模型，选了三款我觉得不错的，然后用 FAL AI 来设置生成流程。在 FAL 上，你可以直接使用 Civot 生成的模型代码。我最终用了三个模型，你可以混合多个模型。我选了三个模型，调整它们之间的权重。然后开始尝试提示词，基本上是从 MidJourney 里用的提示词改编而来，但做了一些修改。我再次学习 FAL 的工作原理以及如何与它沟通，找到了一个有效的提示词。我还调整了很多其他参数，比如模型迭代次数来生成图像。有很多参数可以调整。我花了一天时间调整这些参数，观察输出结果，并确保它长期可扩展。然后我回到提示词，找出可以变量化的部分，比如颜色，创建了一些可替换的颜色变量。例如，如果提到蓝色，在另一个版本的提示词中可能变成蓝绿色或其他颜色。还有添加或删除元素，比如黑洞或运动，有些图像运动感更强，有些更液态化，有些更云状。我构建了一个包含可替换或变量化部分的提示词。然后在工程端设置好，运行时随机化这些变量。我做了压力测试，可以生成多达 10,000 张图片，每张都独一无二，但差异程度保持在可辨识的范围内，一眼就能认出是 Comet 邀请函。

I thought there was an opportunity to do something there with the invites. So, I basically just kind of took some brand assets that we currently have either made by fee or other people on the brand team or like the creative ambassadors who are using MidJourney and sharing style codes. So, I kind of just spent like a day going through artifacts that we already had, going through style codes, going in MidJourney, and just throwing stuff at the wall, trying to like figure out what could an invite even be if we made it generative. Tried a couple things that either didn't end up being scalable because that was also important. But I eventually used that as an input within MidJourney and eventually got a style of something that I thought was good within brand and was scalable. Took those, brought them into Photoshop, edited them, the composition, the layering, played with blend modes to get like a certain effect that I wanted that I couldn't quite get out of MidJourney or like didn't come to mind when I was in the generative iteration part. Used those, reinputed those into MidJourney, got just a subset of what I thought would have been good fodder for training data. Used Civot AI to actually make a model. So I played around there, made a dozen or so models, picked like three that I thought were good, and then used FAL AI to actually set up the flow for the generations. So with FAL, you can just take a model code that you generated with Civot. I ended up using three. You can like blend multiple models together. So I picked out three models, played around with the weights of like how I wanted those models to be weighted against each other. Started playing with prompts, kind of like pulled from what I was using in MidJourney, but editing it a little bit. Just again just learning how FAL works and how to talk with it and found a prompt that worked there. Played a lot with other parameters around how many times it'll run through the model in order to generate the image. There's like a slew of parameters that you can play with. Spent like a day playing with those, just seeing what outputs I got and making sure that it was scalable over time. And then I went back to the prompt and just found pieces that could either be variableized, like color, and then just created some color variables that could be swapped out. Like if blue is mentioned, maybe it's turquoise or whatever else in another version of that prompt. Additions in subtractions, including a black hole or motion, like some have more motion than others. Some are kind of more liquidified, some are more cloudy. Just building a prompt that has pieces that can be swapped or variableized. And then just set that up on the engineering side so that it runs through those at runtime and just randomizes them. I stress tested this and you can get like up to 10,000 generations and they are all unique but stay within a level of variance that feels distinguishable and like identifiable as a Comet invite.

你对这个工作流程之前有多少了解？你以前用过这些工具吗？做过大规模生成吗？你刚才列举了一堆我都不熟悉的工具。你是怎么搞明白这一切的？

How much of this workflow did you have any level of familiarity with before? Like had you used these tools before? Had you done large scale generation before? I mean, you just even listed off a bunch of tools that I'm not even familiar with. So, how'd you figure all this out?

我过去几年一直在自己做一些实验。在 Twitter 上，我算是 AI 的倡导者，但我努力做一个好榜样，展示如何合乎道德地、创造性地使用它：你可以用这个、那个和那个，把它们组合起来，在工作流程中以不同顺序使用等等。所以我对训练自己的模型有一定了解，虽然不是专门针对 Civot。我以前用过其他工具，比如 Stable Diffusion、Flux。有很多选项我试过一两次。我只是通过信息流知道它们的名字。在做这个之前，我从未用过 FAL 或 Civit，但最终它们成了有效的工具。

I mean, I've been doing just experimentation on my own for a couple of years now. Like on Twitter, I'm kind of an advocate for AI, but I try to be a good example of like how to use it ethically or like just creatively like, hey, you can use this and this and this and put them together and use them in different order within your workflow or whatever. So, I had some familiarity with training my own models, not specifically with Civot. I had used other tools in the past. I'd used stable diffusion, flux. There's a number of options that I had either tried once or a couple of times. Just had been familiar with by name as those had come up in my feed or whatever. I had never used FAL or Civit prior to doing this and it ended up being the one that worked.

再次强调，这种工作流程的理念，以及找到方法和工具来解锁新能力，感觉像是设计师的全新疆域。听到这个幕后故事，真的很酷。

Again, this idea of workflows and just figuring out ways and tools to unlock new capabilities feels kind of like this whole new frontier for being a designer. So, really, really cool to hear this backstory.

这绝对是整个事情中最令人兴奋的部分之一。不仅仅是指 Perplexity，而是设计师使用 AI 这件事本身。这就是让我兴奋的地方。不要把输出看作最终结果，而是看作中间产物。要么把它放进另一个工具再拿回来，要么就是不断反馈循环，反复迭代或混搭输出。我分享过的任何东西，都是这样反复循环几十次的结果。

It's definitely one of the more exciting pieces about the whole thing in general. Like not even just perplexity, but just in general designers using AI. That's what's getting me excited. Looking at the outputs not as outputs, but as throughputs. Either putting it into another tool and then putting it back in or just constant feedback loop of reiterating or remixing an output. Like anything I've ever shared is the result of constant just regurgitating that flow like dozens of times.

我分享过的任何东西都有成百上千个源头才最终得到那个结果。

Anything I've ever shared has hundreds of parents that led to that result.

我想稍微拉远一点，更广泛地谈谈 Comet 背后的产品策略，以及设计在其中扮演的角色。似乎有很多关于 Tutonomy 的内容。可能要做成千上万个微决策。但有没有其他你纠结的核心策略决策，或者项目中某个转折点，改变了最终走向，我们可以深入探讨，从而了解你作为设计师如何应对模糊性，并逐步解决设计过程中一些更复杂的部分。

I want to zoom out for a second and talk a little bit more broadly about the product strategy behind Comet, the role that design played in that. It seems like there was quite a bit about Tutonomy. Probably making thousands of micro decisions. But were there other core strategy decisions that you were wrestling with or maybe even like an inflection point in the project that kind of changed where you ended up landing that we could drill into to understand a little bit about how you as a designer wrestle with ambiguity and kind of work your way through some of the more complex parts of the design process.

最让我纠结的问题就是：我们到底为谁做这个产品？用户是谁？每个人都需要浏览器，每个人都在用浏览器，但 Comet 具体是为谁做的？为什么要用 Comet？Comet 面向的是普通用户，至少最初是这样。这有点像一种策略，网撒得越广越好。Comet 是为你父母做的，是为那些买了新电脑就毫不犹豫装 Google Chrome 的人做的，因为他们从没试过别的，或者他们不是从 Arc 或 DIA 转过来的，而是从 Edge 或 Chrome 过来的。这影响了早期很多关于 Comet 外观的决策。一部分原因是它基于 Chromium。现在大多数浏览器都是 Chromium 内核，你要么用 Chromium，要么用 Firefox，基本上不可能自己从头做一个。苹果的 Safari 是个例外。所以，我们用了 Chromium，这样 Chrome 和 Chromium 的很多东西我们都能免费获得。然后就是决策：我们要改什么？我认为整体上还有很大的改进空间，但我们在创新和熟悉之间要走到哪一步？产品早期的很多版本都更偏向熟悉感。所以我们会收到反馈说：“嘿，我本来以为 Comet 会是什么疯狂的东西，或者我从没见过的东西，但安装之后它看起来就跟 Chrome 差不多，不管好坏。”有些人会说：“呃，为什么它没什么不同？”而另一些人会说：“嘿，这太好了，我不需要学新东西，所有东西都在我预期的地方。”这有助于把 AI 用例和助手推得更靠前，作为真正需要专注或尝试的东西。即使只是弄清楚我会用助手做什么，也有一定的学习曲线。助手到底能做什么？所以，通过让 Chrome 和浏览器体验更偏向熟悉，用户就有更多精力去关注那些东西，而不是像 Arc 那样。我还在用 Arc，很难放弃侧边栏标签。那彻底改变了我的生活。Arc 开创了这一点，但 Arc 也做了很多其他事情，把用户吓跑了，这影响了他们扩大市场份额的能力。有些东西感觉有点随意，有些东西我觉得很出色。但当用户只是想尝试一个新浏览器时，这增加了认知负担。我认为我们最初试图避免这一点，当时我谈到只通过邀请制发布，这意味着我们的初始用户很可能是重度用户或技术尝鲜者。大多数反馈是“你需要垂直标签”……嗯，这里存在矛盾。这会很有趣。随着我们开放邀请，我们已经开始看到这一点。随着时间的推移，我们引入了更多普通用户，并且正在为正式发布做准备。那种反馈开始减少了，只是因为有些人从未用过垂直标签，所以他们不理解其价值，也就不会要求。但有趣的是，现在如果你看看任何其他值得考虑的浏览器，它们都提供垂直标签。讽刺的是，我觉得 Chrome 是唯一不提供的。

The one that's top of mind at least was just establishing like who are we building this for? Like who is the user? Everyone needs a browser. Everyone uses browsers, but who is Comet for specifically? Why would you use Comet? Comet is for the lay person, at least initially. It's kind of a strategy thing. It's as wide a net as possible. Comet is for your parents. It's for the person who gets a new laptop and they just install Google Chrome without thinking about it because they've never really tried anything else or they're not someone who's coming from Arc or DIA. They might be coming from Edge or Chrome. And so that influenced a lot of the early decisions on just what does Comet look like? Part of it was it's Chromium. It's a Chromium build as most browsers are these days. You're either Chromium or you're Firefox. It's like basically impossible to make your own at this point. Apple is an outlier with Safari. So, it's Chromium. So, we got a lot of Chrome and Chromium stuff for free. And then it was decisions on like, okay, well, what do we change? I think there's a ton of room for improvement across the board, but like how far do we want to go in making it novel versus familiar? And a lot of these early revs of the early days of the product are leaning far more on the familiar side. So, we'll get feedback that like, hey, I was expecting Comet to be, you know, something crazy or something that I've never seen before, but I installed it and it just kind of looks like Chrome for better or for worse. Some people are like, uh, why isn't it different? And other people are like, hey, this is great. I don't have to learn something new. Everything is where I expect it to be. That helps kind of push the AI use cases and the assistant a little bit more as being the one thing to really focus in on or try. There's a bit of a learning curve in even figuring out what I would use the assistant for. What can the assistant even do? So, by making the Chrome and the browser experience a little bit more on the familiar end, we have more bandwidth from the user to be able to focus in on those things instead of doing what Arc did. I mean, I use Arc still. It's hard to move away from sidebar tabs. Like, that changed my life fundamentally. Arc started that, but Arc also did a bunch of other stuff that kind of scared users away. It kind of hurt their ability to grow their market share. Some things felt a little bit arbitrary. Some things again I think are brilliant. But it adds cognitive load to the user when they're just trying to try a new browser. That I think we were trying to avoid initially when I was talking about the decision to make it a rollout with invite only and that meaning that our initial users are probably going to be power users or people who are tech adopters. Majority of feedback is you need vertical tabs like... yeah there's a dissonance there. It'll be interesting. And we're already starting to see as we open up invites. So like we're getting more lay people in as time has gone on and we're ramping up for general launch. That feedback is starting to dwindle a little bit just because there's people who have never used vertical tabs and so they don't understand the value of that. So they're not really asking for it. But it is interesting now if you look at any other browser that you should probably consider using offers vertical tabs at this point. I think Chrome is the only one that doesn't, ironically.

嘿，很快说一下，让我告诉你全新的 Dive Talent Network。我亲手汇集了超过一百位我认识的最有才华的设计师和开发者，这样我就可以把他们推荐给我最喜欢的公司。所以，如果你正在听这个，并且对新的机会持开放态度，这个人才网络是匿名的，压力非常小，而且是一种简单的方式来看看有什么机会，而不必在社交媒体上发帖。所以，如果你有兴趣加入，或者你正在寻找下一个员工，请访问 dive.club/talent。

Hey, really quickly, let me tell you about the all-new Dive Talent Network. I've hand assembled over a hundred of the most talented designers and builders that I know so I can recommend them to my favorite companies. So, if you're listening to this and you're open to new opportunities, the talent network is anonymous and super low pressure and it's just an easy way to see what's out there without having to post on social media. So, if you're interested in joining or maybe you're looking for your next hire, head to dive.club/talent.

也许我们可以深入聊聊 AI 这部分。你在 Chrome 方面做的很多决策，让你至少能在那里追求一些新颖的元素。那么，我们能谈谈一些设计方向吗？你探索了什么，最终又是如何确定发布版本的？

Maybe we could drill into just the AI piece then. So, a lot of the decisions that you are making as far as the Chrome go, they're affording you the ability to pursue at least an element of novelty there. So, could we talk a little bit about some of the design directions? What were you exploring and how did you arrive at what ultimately shipped?

工程那边有一个智能体团队，他们来自经验非常丰富的地方，所以很多决策或者技术实现方式其实不由我决定。我的职责是让它变得可用，帮助找到用例，然后提供反馈来改进智能体，使其能够完成那些任务。我的意思是，存在一个表面区域问题：它放在哪里？它看起来像什么？我如何与之交互？我把它当作一个隐藏的扩展来处理。它只是一个侧边栏。我们也探索过让它成为一个浮动窗口的可能性。我们仍然会探索一些这样的选项，所以你可以把它从侧边栏中分离出来，但侧边栏是最简单、最快速的做法。然后还有更深层次的问题：它是基于标签页还是基于窗口？如果我在不同标签页之间切换，助手是跟着我，还是刷新，每个标签页都有一个新助手？如果我们做基于窗口的，那么我们就得确保清除对话历史是直观的、合理的，并且是你会去做的事情，因为我发现人们不会清除对话历史。

There's an agent's team on the engineering side and they're coming from places where they have so much experience and that already that like a lot of those decisions or the way the technology works was not really up to me. It was up to me to make it usable and to like help find the use cases to then provide feedback to improve the agent to be able to do those things. I mean there's the surface area question on like where does this live? What does it look like? How do I engage with it and kind of approached it as a hidden extension. It's just a sidebar. There were explorations of that being a floating window potentially. We'll still explore doing some of that stuff even still like as an option. So you can kind of undock it from the sidebar, but the sidebar was just the easiest, quickest thing to do. And then there were deeper questions of like is it tab-based or is it window-based? If I'm going between tabs, does the assistant follow me or does it refresh and is it kind of like a new assistant per tab? If we did the window-based stuff, it's like, okay, well, now we got to make sure that clearing out the thread is intuitive and makes sense and is something that you would do because what I found is that people wouldn't clear the thread.

Perplexity 在线程中的工作方式就是上下文构建的地方。如果你开始问与该线程完全无关的问题，它就会因为喂入了那些内容而开始混淆。所以在 Perplexity 中开启新线程很重要。我们有机会改变它的工作方式，让这一点不那么重要，但这是我们不得不解决的问题。

The way perplexity works in a thread, that's where the context builds. And so if you start asking questions that are totally unrelated to that thread, it'll start getting confused because it'll feed in that stuff. And so like starting a new thread is important in perplexity. There's opportunities to like change that the way that it works so that that's less important. But that was something that we had to tackle.

想要让比较 Windows 笔记本和 MacBook 变得容易。比如我应该能去 Windows 页面和 apple.com 上查看那些笔记本的页面，然后让 Comet 告诉我哪个性价比最高，我到底该买哪个。一种实现方式是做成基于窗口的，这样我可以打开 Apple 页面问问题，切换到 Windows 页面问同样的问题，它知道我之前是在 Apple 页面上问的。另一种方式是做成基于标签页的，那么为了做到这一点，我们实际上需要能够提及标签页，或者让智能体足够聪明，能查看你打开的标签页，这样你甚至不需要在问题中明确或啰嗦地描述你想要什么。

Wanting to make it easy to like compare between a Windows laptop and a MacBook. Like I should be able to go to Windows and go to apple.com go to the pages for those laptops and have comet tell me like what's the best value like which one should I actually buy and so one way of doing that is by making it window-based so I could open up Apple ask the question switch over to Windows ask the same question it has the context that I was asking this from Apple and then that happens the other way of doing that where it's tab based it's like okay well in order to do that then we actually got to be able to mention tabs or we got to make the agent smart enough to look at your open tabs so that you don't even have to like be explicit or verbose in your question or whatever you're trying to get out of it.

结合分类器、推理，以及知道去检查你的历史记录、所有打开的标签页，判断哪些是相关的，然后给出正确的结果。这有点像试错，一开始直觉上觉得应该是基于窗口的，试了之后发现它有缺陷，有时会让人困惑，最终转向了基于标签页的方案，这并非我的第一直觉。那不是我原本以为我想要的，直到我们试了之后才意识到还有其他方式，无论是从智能体的功能上，还是引入像提及标签页这样的机制，都能让这些东西工作起来。

The combination of like classifiers and in inference and also just knowing to like check your history, check the all tabs that are open decide what's relevant and give you the right thing. So that was a bit of just trial and error, like having a gut feeling that it should be window-based, trying it out, realizing that it has pitfalls. It's kind of confusing sometimes and ended up going towards tab-based, which was not my first instinct. Like that's not what I thought I wanted until we tried it and then realized that there's other ways whether that's functionally how the agent works or introducing mechanisms like mentioning tabs to be able to make those things work.

我把浏览器看作是迈向操作系统或智能体操作系统的第一步。如果 Comet 还能查看你电脑上的内容，或者操作你电脑上安装的其他原生应用，而不仅仅是标签页中打开的网页应用，那就太酷了。这有点像进入了 MCP 的领域，我可以把 Slack、Notion 和 Comet 都用它们的原生应用整合在一起，实现不同的工作流。

I see the browser as the first step towards an OS or an agentic OS. Like it would be so cool if Comet could also look at what's on your computer or act on other first-party apps that are installed on your computer, not just web apps that you have open in a tab. Kind of gets into the MCP realm where I could sling Slack and Notion and Comet together all using their native apps to enable different workflows.

如果是基于标签页的，那就意味着你在每个标签页上执行一次查询。该查询存在于该标签页的上下文中。所以我们也在探索，比如如果我们有一个命令中心，或者一个专门聚焦于智能体的界面，它不必存在于标签页内，那会怎样？它可以是一个独立的东西，或者我可以安排查询、一次运行多个查询并管理它们、将它们串联起来，同时仍然能够提及标签页或与浏览器内的内容交互，但不必被锁定在“我必须打开一个网页标签页才能使用助手”这种模式里。

If it's tab based, that means you're doing a query per on that tab. That query lives within the context of that tab. So also exploring like okay well what if we had like a command center what if we had like a surface area that's just focused on the agent that doesn't have to live within a tab. It can be its own thing or I could schedule queries or run many queries at once and be able to manage those, string them together, still being able to appmention tabs or do anything engage with stuff within the browser itself, but not being locked down to like, okay, I've got to have a tab open on a web page in order to use the assistant.

我想回到从基于窗口到基于标签页的转变，以及它最终如何违背了最初的直觉。在设计过程中，还有其他地方你最终主张的东西与你最初的想法有所不同吗？

I want to go back to the switch between like window-based to tab-based and how that ended up going against the initial instinct. Were there any other places in the design process where you ultimately advocated for something that was a little bit different than what you originally thought?

那绝对是最大的之一，我私下里还挺懊恼的。你知道，就是那种我仔细思考过，得出了这个观点，并且能够用理由支撑它，我很有信心觉得这就是该尝试的方向，结果却发现它不是正确答案。我当时就想，“天哪，”

That's certainly like one of the biggest ones that I lowkey feel bad about. You know, it's like the kind of thing where like I thought through this thing, I came out with this opinion and I was able to back that up with why and I felt confident that like this is the thing to try and then it ended up not being the right answer. And I was like, "Oh my god,"

我大概 30 天前也经历过非常类似的事，我把团队带进了一个两周的兔子洞，事后看来完全是错的，而我当时完全以为是对的。所以我理解。这是作为设计师最困难的部分之一。

I'm like 30 days from having a very similar thing where I sent us down like a two week rabbit hole that was just in retrospect wrong and I totally thought it was right. So I get it. It's like one of the hardest parts about being a designer.

你刚才提到的是新标签页。我们到底应不应该有一个新标签页？

Something you just touched on or mentioned is just the new tab page. Should we even have a new tab page?

对。

Yeah.

比如，当你点击新标签页时，实际显示一个界面，与 Arc 那种聚光灯模式相比，有什么好处？Arc 没有新标签页，只有搜索。

Like what are the benefits of actually having a surface area show up when you hit new tab versus like Arc which is like a spotlight model. There is no new tab page. It's just search.

我们对这两种方案都做过探索。现在也还在做，仍然对能把它推到多远保持开放态度，北极星在哪里？但最终决定先从一个新标签页开始，给我们一个品牌化的界面。有一些品牌化的很酷的小部件，比如“试试助手”小部件，它会随机执行一个智能体式查询，是一种快速一键式的方式，让你体验神奇时刻，也许能种下一颗种子，让你觉得“哦，这个浏览器能做我没想到的事情”；或者 Comet 游戏，有点像 Comet 的离线版。但有一个入口点，比如最近访问的网站，这很有价值，也是进入 Perplexity 本身以及你的整个 Perplexity 账户和库的入口。所以基于所有这些原因，从一个新标签页开始是合理的，但从长远来看可能并不合理。可能有一个更好的解决方案。所以不要因为我们现在有一个新标签页而感到太糟糕，即使我们并不真正想要它。

Have done explorations on both of those. Still doing that like still like open to how far can we push this and and what is the north star here? But ultimately decided let's start with a new tab page that gives us a branded surface area. Have some branded really cool widgets like a try assistant widget which just like does a random agentic query just to like a quick one-click way to like get you to a magic moment to maybe plant a seed of like oh this browser can do something I didn't think was even possible or the comet game which is kind of like the offline equivalent for comet but having an entry point for that recent sites like that's pretty valuable an entry point into like perplexity proper and your whole perplexity account and your library. So like it just made sense to start with a new tab page for all of those reasons, but probably doesn't make sense in the long run. Like there's probably a better solution. And so just like not trying to feel too bad that we have a new tab page right now and even though we don't really want one.

我脑子里有一个决策谱系：一方面是一些低层次的视觉组件细节，听起来你基本上是自己做主。你感到有自主权，直接做决定，发布出去，我们之后再解决。另一端是像智能体用窗口还是标签页这样真正重大的战略决策。中间则是一片混乱的灰色地带。我希望你能帮我稍微理清这片灰色地带。具体来说，作为一个新团队的成员，你还在摸索如何与这些人合作。你如何看待适当的协作程度与直接做决定之间的平衡？你什么时候策略性地带人一起参与，让关键利益相关者加入进来，这在你的实践中是怎样的？

There's kind of this spectrum of decisions that I have in my brain where on one hand you have some of the lower level visual component details that it sounds like you're kind of just owning. You feel the autonomy, just make that call, ship it, and we'll figure it out. And then on the other end, there's these really meaty strategic decisions like windows versus tabs for the agent. And then there's this messy middle ground in between. And I'd kind of like you to help me make sense of that messy middle ground a little bit. Specifically, as someone who's on a new team, you know, like you're still kind of even figuring out how to work with these people a little bit. How do you think about the right level of collaboration versus just making the call? When do you strategically bring people along for the ride, loop in key stakeholders, and what does that look like in your practice?

一开始是持续的。我不断地打扰每个人，但这只是新团队入职过程的一部分，就像试图弄清楚事情在哪里。我们是否有针对这个特定问题的现有模式或解决方案？我们考虑过这个吗？只是试图了解团队的想法，他们对这件事的答案是什么？以及他们为什么那样回答？通过学习，随着时间的推移，我就不必再打扰他们了。我可以自信地知道他们会说什么，或者知道他们的方法会是什么。所以对我来说，随着我在团队中逐渐成熟，对可能出现的那些问题的答案有了更好的把握，这种情况变得越来越少。

Initially, it was constant. I was constantly bugging everyone, but it was like a it was just part of the onboarding process for a new team, like just trying to figure out where things are. If we have a pre-existing pattern or a solution for this particular problem, have we thought about this? Just trying to get a sense of like how the team thinks, what are their answers to this thing? And just like why did they answer it that way? And learning so that over time I don't have to bug them. I can be confident that I know what they would say or I know what their approach would be. And so it's kind of gotten less and less, at least for me personally, as I've matured within the team, and have a better sense of answers to those questions that might come up.

这也有点反常，因为我是 Comet 上唯一的设计师。Gunner 负责 iOS 和移动端，他是移动端之王。Champuri 负责大部分智能体和核心的 Perplexity 回答内容。有人负责企业级产品，有人负责空间相关的东西。所以每个人都有自己的专长，覆盖的领域也相当广泛。我刚开始的时候，Comet 上只有我一个人。现在我们又雇了两位设计师来帮我，这很棒。所以现在有移动端的人帮我做 Comet，包括 iOS 和 Android。还有一个人刚开始帮我做桌面端的东西。这就像是重新打开了大门。我一直在找他们，说：“嘿，告诉我我做错了什么，帮我重新审视所有这些决定。”同时我也在努力给他们提供所有背景信息，比如我们是怎么走到这一步的，为什么做那些决定，以及我们还想在哪些方面改进。现在我们有了更多的同步会议和头脑风暴，这很有趣，也很酷，非常令人兴奋。这也有助于缓解我的焦虑，那种必须做出决定或做出正确决定的压力。我是个设计师，我有决策瘫痪的问题。

It's also maybe a bit of an anomaly because I was the only one on Comet. There's Gunner, who's on iOS and mobile, and he's the king of mobile. There's Champuri, who owns most of the agent and core Perplexity answer stuff. People are tackling enterprise, people are more around spaces. So everyone has their specialty and a surface area that's quite broad. I started and I was the only one on Comet at this point. We've hired two extra designers to help me. So that's been great. So we have someone on mobile helping me with Comet, so iOS and Android. Someone just started to also help me with desktop stuff. It's kind of like opening the doors again. I'm pinging them all the time: 'Hey, tell me everything I did wrong. Help me second-guess all of these decisions.' While also trying to give them all the context of how we got to where we are and why we made those decisions and where we still want to improve. And now we're having a lot more syncs, a lot more jam sessions, which is fun and very cool to see and very exciting. And it's helping ease some of my anxiety, feeling like I have to make a decision or the right decision. I am a designer. I suffer from decision paralysis.

我每天都穿同样的衣服，因为我不想在现实生活中处理这些事，毕竟这是我的工作。

I wear the same thing every day because I don't want to deal with that in my real life because it's my job.

在这么大的领域里做出这么多决定，真是了不起。你的背景确实很棒。你是 Descript 的第一位设计师，对吧？没错。所以你曾是 Descript 的第一位设计师。你提到了 Motif 和 Daylight。所以在日常工具方面，这可不是你第一次尝试了。那么，在设计这些其他工具的经验中，有没有什么经验教训或原则，你在做 Comet 时会借鉴？

It's an impressive feat to own that many decisions across that large of a surface area. You do have a pretty awesome background. You know, you were the first designer at Descript, right? Is that correct? Yeah. So you had the first designer at Descript. You mentioned like Motif and Daylight. So it's not your first rodeo in terms of everyday tools. So are there lessons learned or maybe principles that you find yourself pulling from based off of your experience designing these other tools while working on Comet?

我整个职业生涯都在做设计系统。我是在一家 UX 公司工作时接触到设计系统概念的。我的主要客户是微软，那是在 2008、2009 年左右。所以我帮助建立了 Metro 的设计语言，后来变成了 Windows 8、Windows Phone、Xbox、Surface 平板。我和那些团队合作，建立了一个可以跨平台的设计系统。在微软，最重要的是无障碍设计。所以这在我职业生涯早期就根深蒂固了，我深知它的重要性，以及设计系统或系统设计的重要性。当时，设计系统还是 PDF 文件，不是可用的工具，而是参考文档。从那以后，我去过的每个地方，我基本上都是独行侠。在 Descript，直到 2022 年我都是唯一的设计师。在那之前我在 OpenTable，负责餐厅产品，我也是唯一的设计师。团队其他人都做网站、品牌或 iOS 应用。每个人都想做 iOS 消费者应用，那是 iOS 7 刚推出时的热门。但我想，“不，我要做赚钱的东西，我要解决难题。”结果我还是成了独行侠。我是那个产品唯一的设计师，然后把它带到了 Detour。Descript 之前也是一样。后来我们把资产卖给了 Bose，然后把它变成了 Descript。把它剥离出来成为一个独立工具，因为 Descript 是我们为 Detour 做的内部工具。所以我独自一人太久了，甚至没有其他设计师可以一起讨论。我和 CEO 合作的方式就像和设计师合作一样。所以我锻炼出了很强的能力，能够做出这些决定，或者为自己建立一个可扩展的框架，无论产品是什么，领域是什么。这对我来说非常宝贵。我之前提到过，在 Descript 待了那么久，我觉得自己没成长，对自己不利。有一段时间，我觉得在同一个地方待太久，停滞不前。只是在产品和行业领域内成长，但作为设计师没有成长。但之后，加入 Perplexity，加入设计团队，我才意识到那段时间我一直在积累技能和力量。正是这些让我在 Comet 的头几个月里能够独立快速上手。

I've been doing design systems my entire career. I kind of got introduced to the concept of design systems when I was working at a UX firm. My main client was Microsoft. This was in like 2008, 2009. So I helped establish the design language for Metro, which turned into Windows 8 or Windows Phone or the Xbox, the Surface tablet. I was able to work with those teams to establish a design system that can live across all platforms. At Microsoft, the huge thing was accessibility. And so that was just ingrained in me early on in my career of how important that is and how important design systems are, or systems design is. And at the time, design systems were PDFs. They were not things that you could use. They were things that you would reference. And everywhere I've been since then, I've mostly been a lone wolf. I was the only designer at Descript until like 2022. I was at OpenTable before that and I was working on the restaurant product. I was the only designer on the restaurant product. The whole rest of the team was on the website or brand or the iOS app. Everyone wanted to work on the iOS consumer app. It was the shiny thing when iOS 7 was first introduced. But I was like, 'Oh no, I want to work on the thing that makes the company money. I want to work on the hard problem.' But I ended up being a lone wolf. I was the only designer on that product and then kind of took it to Detour. Pre-Descript was the same thing. Then we sold assets to Bose and kind of turned that into Descript. Spun that off as its own tool because Descript was an internal tool that we made for Detour. So I've just been alone for so long that I didn't actually even have designers to riff on. I was working with a CEO in the way that I would work with designers. And so I just gained a lot of muscle in being able to make those decisions or set myself up with a framework that can scale no matter what the product is or what the surface area is. And that's proven really valuable for me. I mentioned at the beginning of this that when I stayed at Descript for so long, I felt like I did myself a disservice that I didn't grow. There was a period of time where I felt stagnant being in the same place for so long. Just kind of growing within an industry area within a product but not growing as a designer. But after that, post that, joining Perplexity, joining a design team, realizing that I was just building skill and muscle that whole time. That's kind of what enabled me to hit the ground running with Comet on my own for the first couple of months.

你有没有什么框架或经验法则，关于如何将 AI 集成到工具中？因为你也一直在思考这个问题。我是说，Descript 在 AI 还没火的时候就已经做了很多。所以我很好奇你在这方面有什么想法。

Do you have any frameworks or rules of thumb or something like that as it relates to integrating AI into a tool? Because you've been thinking about that a lot, too. I mean, Descript is doing a ton with AI way before it was cool. So, I'm curious if anything comes to mind on that front.

AI 和工具的结合很有趣，因为现在还太早。我们正在摸索模式。我们做的第一件事就是聊天，比如支持式的聊天界面和提示词。我认为这两者都有问题。它们在可用性上不太理想。这是一种工程方法，是一种无头 UI。它非常强大，但需要很多前期学习才能知道如何使用以及能做什么。这会导致新用户和普通用户出现“闪烁光标”问题。就像，“好吧，我有个空白输入框。我到底该拿它怎么办？我甚至能用它做什么？”所以在 Descript，提示词和聊天被当作二等公民。我们用的是自定义指令，而不是“提示你想要什么”。就像，“你真的知道自己想要什么吗？好，那你就去做。否则，这里有个界面。这里有滑块、下拉菜单和描述，帮你选择你想要的东西。”但把所有功能都放进去，比如生成片段，这是 Descript 的一个功能：你可以拿一个播客，生成社交媒体片段。你可以用提示词来做，直接告诉它去做。但我们最终做的是构建一个界面。所以这个功能就是“片段生成片段制作片段”。当你打开它时，它会显示，“好的，你想要这些片段多长？选个滑块。你想要多少个片段？有没有特定主题要关注？格式是什么？这是给 Instagram、TikTok 还是 Twitter 用的？”这些问题我们可以通过一个通用用户界面提前得到答案，实际上在后端它会开始构建那个提示词，但用户不需要知道提示词是什么。要么是秘方我们不想让你知道，要么就是多余的信息。如果你不想学怎么做提示工程师，你根本不需要知道这些。它就在后台默默进行。如果你愿意，也可以自定义。

AI and tooling is interesting because it's so early. We're figuring out the patterns out the gate. The first thing that we did was chat, like support-style chat interfaces and prompts. Both I think are problematic. They're less than ideal for being usable. They're an engineering approach. It's a headless UI. It's so powerful, but it requires so much upfront to learn how to use it and what's possible. And it leads to a blinking cursor problem for new users and lay people. It's just like, 'Well, I got this blank input. What the hell do I do with this? What can I even do with this?' And so with Descript, prompting and chat was treated like a second-class citizen. It was custom instructions instead of 'prompt what you want.' It was like, 'Do you really know what you want? Cool. Go for it. Otherwise, here's an interface. Here's sliders and drop-downs and a description to help you choose what you want.' But just putting all of these features, like generate clips for instance, as an example of a feature within Descript: you can take a podcast and you can generate clips for social media. You could do that with a prompt. You could just tell it to do that. But what we ended up doing was building an interface. So the feature is 'clip generate clips make clips.' When you open that up, it shows you just like, 'Okay, how long do you want these clips to be? Choose a slider. How many clips do you want? Is there a certain topic that you want to focus on? What's the format? Is this for Instagram or TikTok or Twitter?' There are questions that we can get answers to upfront just with a general user interface that actually on the back end starts to build that prompt, but the user doesn't need to know what the prompt is. Either it's secret sauce and we don't want you to know, or it's just overkill. It's information you don't really need to know if you're not interested in learning how to be a prompt engineer. It just happens in the background under the hood. And if you want, you can customize it.

你可以点击自定义指令按钮，并在上面添加自己的提示词。但总的来说，Dcript 的做法是将 AI 产品化——要么通过功能实现而不称之为 AI，要么如果是 AI，也不以提示词或聊天为主导，而是通过界面获取我们所需的信息，从而真正给出好的结果。这种方法对于 Comet 的产品设计有多大的适用性？尤其是你们已经为一个真正智能体式的世界奠定了基础，这个世界还要应对无限更大的可能性范围，因为你不仅仅是在编辑东西，而是在浏览器中做任何事情。那么，你能谈谈在设计这个更智能体式的世界时，未来界面面临的一些挑战或背后的思考吗？

You can hit the custom instructions button and add your own prompt on top of that. But that in general was the approach with Dcript was productizing the AI whether it's by feature and not calling it AI or if it is AI not leading with a prompt or chat but leading with getting information that we need through an interface to actually give you a good result. How applicable is that approach to designing the product going to be for Comet? Especially as you all have laid the foundation for a truly agentic world that also accounts for an infinitely larger spectrum of possibilities because you're not just editing something, you're doing literally anything in a browser. So can you talk a little bit about some of the challenges or thinking behind what future interfaces look like as you're designing this more agentic world?

第一个问题就是发现和教育，让用户了解 Comet 能做什么，或者 Comet 能为你个人、独特地做什么。一种方法是建立一个查询库。比如我们有一组虚拟查询，或者每个人都会用到的基本查询，我们知道这些查询是有效的。想法是，这些查询能播下创意的种子——你在库中浏览查询，看到它属于某个类别，比如邮件，然后利用这个机制来播下 Comet 能做什么的种子。还有另一层，就是个性化查询。实际上，如果你导入浏览器数据，或者你使用过 Comet 并有历史记录、书签，我们就能为你生成个性化查询，这些查询是专门为你定制的，作为另一种帮助播下创意种子的方式。所有这些都会有效。你可以直接点击“尝试并评论”按钮，它就会生效。或者，如果你觉得某个操作你会反复使用，你可以点击保存，将其保存为快捷方式。但我经常发现它作为一个跳板很有用，比如：哇，我不知道 Comet 还能做这个。我不想要这个具体的查询，但让我编辑一下。比如，我保存它，然后编辑那个快捷方式，现在我就有了自己的东西。未来的版本可能允许分享这些查询。把你的 Comet 查询配置文件设为公开，这样别人就能看到并学习。这是可发现性的一部分。然后是可用性，即在产品中实时捕捉你当下的需求。我们可以通过零建议来实现，当你点击全能框或助手输入框时，我们知道你在哪个网站。我们已经对主要域名上哪些操作有意义有了很好的了解。比如，最终这一切都将由 AI 推断和生成，但一开始，相当一部分是硬编码的——我们做了工作，找出哪些是热门域名，以及在这些域名上哪些是热门操作，然后把这些放到零建议中。所以当你在 Notion 上点击输入框时，你会看到一系列针对 Notion 的建议，这些建议可能与你当时想做的事情相关，也可能不相关，但随着时间的推移，它会更好地推断你的意图。根据你有一个空白文档还是已填满的文档、一个包含电子表格的文档，以及你之前刚刚执行的操作，我们可以用 AI 做很多事情来推断下一步可能是什么。你在做什么工作流程，从而建议一个 Comet 可以执行的操作？然后是后续操作。如果你有一个复杂的查询，或者你想做更复杂的事情，设计智能体的方式不仅仅是提交查询、获得答案，而是实际提交查询，看到浏览器在运行，看到它在思考什么，它如何学习一个网站，它如何学会使用那个网站，然后如果它卡住了或需要后续操作，就暂停智能体，提示你需要信用卡信息，比如我们需要你的批准才能发送这封邮件，而不是自动发送。就像你给它盖上最终的印章。或者只是建议后续操作，比如：嘿，你做了这件事。还有另一个查询。它类似于零建议，但实际上是一个后续建议。

One is just the problem of discovery and education on what Comet can do or what Comet can do for you like personally and uniquely for you. And so one approach to that is just a query library. Like we've got a set of dummy queries or just basic queries that everyone has that we know will work. The idea is maybe those plant seeds of ideas like you read this query in the library. You're browsing it. You see that it falls under a category of like email and just using that as a mechanism to plant seeds of what Comet can do. Then there's another layer of that which is personalized queries. Actually, if you import your browser data or you have used Comet and you have a history, you have bookmarks, we're able to generate personalized queries for you that are specific and unique to you as another way of just helping plant the seeds of ideas. Like all of these are going to work. You can just click the try and comment button and it'll work. Or you could hit save and save it as a shortcut if it's something that you like see yourself doing repeatedly. But I've often just found it useful as a springboard of like, oh wow, I didn't know Comet could do that. I don't want this query specifically, but let me edit it. Like, let me save it and then edit that shortcut and now I've got my own thing. Future version to that could be being able to share those queries. Making your Comet query profile public so people could like see that and learn from. That's one piece on like the discoverability. Then there's usability in terms of actually being in the product in real time capturing you at that moment. So we can do that with like zero suggest when you click into the omni box or the assistant input. We know what site you're on. We already know like we have a good idea of what makes sense to do on a major domain for instance like eventually all of this will just be AI and inferred and generated but to start like a good fair amount of it is just hardcoded like we did the work to figure out what are popular domains what are popular things to do on those domains put those in zero suggest so when you're on Notion and you click on the input you've got a series of like suggestions that are specific to Notion that may or may not be relevant to what you're trying to do at that moment but will get better over time at inferring that. Depending on whether you have an empty document or a document that's filled, a document with spreadsheets in it, depending on the actions that you just took prior to that, there's a lot that we could do with AI to infer what might be the next step. What workflow are you doing to suggest an action that Comet can do? Then there's like follow-ups. If you have a complex query or something that you want to do that's more complex, designing the agent in a way that's not just submit query, get answer, but actually submit query, see the browser working, see what it's thinking, how it's learning on a website, how it's learning how to use that website, and then if it gets stuck or it needs a follow-up or something, pausing the agent, prompting that you need your credit card info, like we need your approval to send this email before it just sends automatically. Like put your final stamp on it or just suggesting follow-ups like, hey, you did this thing. There's another query. It's similar to a zero suggest, but it's actually like a follow-up suggest.

嗯，我们涵盖了很多内容。我很感谢你深入探讨了你的想法，你不仅在设计 AI 体验，而且还在真正地实验和摆弄所有这些工具。那么，在我放你走之前，你认为设计师在 AI 及其在我们实践中所扮演的角色方面，还有什么应该更多思考的吗？

Well, we've covered a lot of ground. I appreciate you getting into the weeds of what you're thinking about as someone who's not only designing AI experiences, but you really experimenting and tinkering with all of these tools. So maybe before I let you go, is there anything else that you think designers should be thinking about more when it comes to AI and the role that it plays in our practice?

大多数人认为，哦，你只是输入了一些东西。你输入了你想要的，然后点击执行，就得到了你想要的结果。要么是老虎机机制，要么你只是从菜单中下单。当然，AI 可以那样，但它不必如此。它本质上不是那样的。AI 本质上并没有那么局限。它实际上是如此开放。我做的很多事情都是在 Figma、Photoshop 之间来回切换，至少在图像生成或艺术创作方面是这样。比如，在传统工具和 Midjourney 之间来回切换。用我自己的艺术作品作为输入，用我自己的艺术作为模型的素材，再用那个模型生成的图像作为输入。把那个输出拿过来，继续迭代和混搭。把那个输出作为另一个输入，再结合另外两个输入。真的不只是停留在输入文字、点击执行，而是做很多你作为艺术家或创意人员通常会做的事情，然后利用它来获得变化。可能成功，可能失败，可能你花时间迭代却一无所获，但有时你会得到一个意外的惊喜。有时你会得到一些东西，比如：嘿，这真是一个很酷的想法。我没想到可以这样做。让我现在回去，做点什么，我可能会用不同的方式做。但要把 AI 作为创意的跳板，作为沟通工具，而不是用来绕过或避免工作的工具。重点不是你能做得多省事，而是因为 AI 作为你工作流程中的工具，你能做多少更多的事情。无论是在开始、中间还是结束，它都非常开放和灵活，而且随着 MCP 等概念的引入，它变得越来越开放和灵活，我觉得这非常令人兴奋，并试图教给别人或向别人展示光明——就像当年大家都讨厌 Photoshop 一样。它当时非常有争议。30 年前，Photoshop 上了新闻，因为把一匹马从山脚移到山顶的树旁，以制作一个合适的书籍封面，这件事引起了巨大争议。当时有专门的新闻报道，因为这个想法太激进了："哦，你在操纵这张照片。"

Most people think, oh, you just typed something. You just typed what you wanted and hit go and you got what you wanted. Either slot machine mechanic or you just ordered from a menu. And sure, AI can be that, but it doesn't have to be. It's not inherently that. AI is not inherently that limiting. It's actually so open-ended. A lot of what I do is back and forth between Figma, Photoshop, at least like in image generation or art stuff. Like it's a back and forth between traditional tools and Midjourney as an example. Using my own art as inputs, using my own art as fodder for a model, using images from that model as inputs. Taking that the output of that and just continuing to iterate on that and remix that. Take that output, put it in as another input along with two other inputs. Really not just sticking to type words, hit go, but doing a bunch of what you would normally do as an artist or a creative and then using that to get variance on something. Maybe it's a hit, maybe it's a miss, maybe it's, you know, you might spend time iterating and not get anything, but sometimes you're going to get a happy accident. Sometimes you're going to get something that's like, hey, that's actually a really cool idea. I didn't think of doing that. Let me go back now and like do something and I might do something in a different way. But like using AI as a springboard for ideas, using AI as a communicative tool and less as a tool to get around or avoid doing work. It's less it's not about how little can you do, it's about how much more can you do because of AI as a tool within your workflow. whether it's at the beginning, the middle, the end, it's so open-ended and flexible and becoming increasingly even more with the introduction of like MCP as a concept that I just find really exciting and trying to like teach people or show people the light that like everyone hated Photoshop too. It was very controversial. Like Photoshop was in the news 30 years ago because of how controversial it was to move a horse from a bottom of a hill up to the top of a hill next to a tree to make a good format for a book cover. There was new segments on that because it was just so radical of an idea like, "Oh, you're manipulating this photo."

这太不对了。虽然我觉得对 AI 的批评有些完全合理，也有一些我们讨论得不够或没有正面回应的合理担忧。但我也认为，随着技术越来越好、模型越来越小、越来越快，很多两三年前反对 AI 的论点要么不再相关，要么相关性降低了。我真心相信，在合乎道德地使用 AI 来做好事、解决好问题方面，存在着巨大的力量和机会。我希望人们在决定对 AI 的看法时能更谨慎一些，多做研究。我鼓励每个人都去做自己的研究，保持足够的好奇心。不要只听你在 Twitter 上看到的，不要陷入共识。如果你在使用 AI，就要注意它本身以及你使用方式的影响。牢记伦理和道德。我认为所有这些对于这个工具的长期发展和未来都很重要。会有不良行为者，我们需要追究他们的责任。问题不在于技术，也不在于所有人。这是一个容易被滥用的花园。如果实施 AI 的公司没有足够警惕去防止这些事情，那么闸门就会打开。它不会消失，我们必须与之共存。我只是尽力引导方向。我最终决定留在 AI 领域，而不是在设计职业生涯中退缩，因为我感到有道德义务去推动它朝着让我感到舒适的方向发展。我无法接受退出舞台，看着它走向我不认同的道路，而我知道自己本可以产生影响。

That's so wrong. While I feel like there are totally valid arguments against AI, there are valid concerns that I don't think we talk enough about or address well enough directly head-on. But I also think that a lot of those arguments are continuously getting more and more outdated as the technology gets better, as these models get smaller, as they get quicker, as you know, a lot of stuff that the arguments against AI 2 years ago, 3 years ago are either no longer relevant or are just less relevant. And I truly believe that there is great power and opportunity in using AI ethically for good purpose and to solve good problems. I hope people are a little bit more mindful of whether they're deciding how they feel about AI, doing more research on it. Like I encourage everyone to do your own research. Be curious enough. Don't just listen to what you see on Twitter. Don't fall into consensus. If you are using AI, just be mindful about the impact it has, just the technology itself or how you're using it. Keep ethics and morals in mind. I think all of that is important just for the longevity and the future of this tool. Like there will be bad actors and we need to hold bad actors accountable. It's not the technology that's the problem and it's not everybody that's the problem. It's a garden that is ripe for abuse. If companies that are implementing AI aren't mindful enough to be able to prevent those things, the floodgates are open. It's not going away. We have to live with it. And I'm just trying to do my best to guide the ship. I ultimately decided to stay in AI and not back away from it in terms of my design career in the industry because I felt a moral obligation to be part of moving that forward in a direction that made me feel comfortable. I couldn't ultimately couldn't live with exiting the scene and watching it go down a path that I wasn't comfortable with, that I knew that I could have had an impact on.

我觉得这非常鼓舞人心，而且我认为你很好地阐述了这种区别，以及我的一些挫败感——看到围绕 AI 的叙事，以及把工具当作老虎机来使用（这在很多方面几乎取代了对艺术的需求），与像艺术家一样使用工具、挖掘创造力并提升天花板之间的对比，而你正是后者的典范。所以，非常感谢你今天来参加节目，给我们一些幕后故事，分享一些工作流程和你正在尝试的东西。我可以真诚地说，我是你作品的忠实粉丝，无论是 Comet 还是 Dscript 等等。所以，谢谢你的时间。

I find it very inspiring and for what it's worth I think you did a really good job of articulating that difference and some of my own frustrations in seeing even the narrative around AI and what it looks like to use the tool as a slot machine which kind of almost replaces the need for art in a lot of ways versus using the tools as an artist and tapping into that creativity and raising the ceiling and you're exemplifying that so well. So, I really appreciate you coming on today and giving us a little behind the scenes and sharing some of the workflows and what you're experimenting with and I can genuinely say that I am a big fan of your work both with comet and dscript and everything. So, thanks for the time.

谢谢。非常感谢。很感激。

Thank you. Thank you so much. Appreciate it.

在你走之前，我想花一分钟介绍一下我最喜欢的产品，因为我经常被问到我的技术栈。Framer 是我建网站的工具，Genway 是我做研究的方式，Granola 是我在评审时记笔记的工具，Jitter 是我为设计做动画的工具，Lovable 是我用代码实现想法的工具，Mobin 是我寻找设计灵感的地方，Paper 是我像创意人一样设计的工具，而 Raycast 是我每一步的快捷方式。我精心挑选了这些公司，这样我就能全职做这些节目了。所以，支持节目的首要方式就是去看看它们。你可以在 dive.comclub/partners 找到完整列表。

Before I let you go, I want to take just one minute to run you through my favorite products because I'm constantly asked what's in my stack. Framer is how I build websites. Genway is how I do research. Granola is how I take notes during crit. Jitter is how I animate my designs. Lovable is how I build my ideas in code. Mobin is how I find design inspiration. Paper is how I design like a creative. and Raycast is my shortcut every step of the way. Now, I've hand selected these companies so that I can do these episodes full-time. So, by far the number one way to support the show is to check them out. You can find the full list at dive.comclub/partners.