生成式 AI 设计：从 0 到 1——对话 Luma AI 设计主管

欢迎来到 Dive Club。我是 Rid，这里是设计师永不止步的地方。本周的嘉宾是 Tuhin Kumar，他是 Luma AI 的设计主管，Luma 是当今生成式 AI 领域的领军公司之一。我们将深入探讨设计一个真正优秀的 AI 产品所涉及的模式、交互以及所有思考。这一集我最喜欢的部分之一是我们稍微拉远视角，Tuhin 谈到了他在 Apple、Airbnb 和 Meta 的经历，以及将生成式 AI 作为一种新材料来工作如何改变了他对设计过程的看法，甚至对设计实践的未来意味着什么。那么，让我们先了解一下 Tuhin，听听他来到 Luma 的历程。

Welcome to Dive Club. My name is Rid and this is where designers never stop learning. This week's episode is with Tuhin Kumar who's the head of design at Luma AI which is one of the companies really leading the charge on generative AI today. So we're going to get ultra specific on the types of patterns and interactions and all of the thinking that goes into designing a truly great AI product. And one of my favorite parts of the episode is when we kind of zoom out a little bit and Tuhin talks about his time at Apple, Airbnb and Meta and all of the ways that working with generative AI as a new material really changes how he thinks about the design process and what it means for the future of the practice even. So let's start by getting to know Tuhin a little bit and hearing about the journey that led him to Luma.

我学的是工程专业，在工程学到一半的时候开始写博客，然后想要一个更漂亮的博客。所以我自学了 HTML 和 CSS。后来发现这还不够。当时是 WordPress 时代。所以我又自学了 PHP。整个过程结束后，我发现写博客本身还行，但我在印度做的另一件事是什么？当时还没有正式的名称。iPhone 刚出来，Twitter 成了我了解世界的窗口。我关注了当时 Twitter 上所有的 10 位设计师，那正是这个领域，这让我开始理解整个设计领域。所以大学大部分时间我都在自学设计、做自由职业，然后来到了一家叫 Pulse 的小公司。那是一个新闻阅读应用。我是他们的第一个设计师。我每天都是 Pulse 的用户，所以当我发现你在做这个时，我觉得太棒了，我一直想知道是谁做的。

I studied engineering and somewhere mid-engineering I started blogging and then I wanted a prettier blog. So I taught myself HTML and CSS. Then I realized oh that's not enough. Back then this was WordPress era. And so then it's like oh actually I have to teach myself PHP. I taught myself that. And then by the end of the whole process I was like the blogging part is fine. What was this other thing I did in India? There was no real name for it back then. iPhone had just kind of come out and that was when Twitter honestly became my really peak into the world. I would follow all the 10 designers at that point who were on Twitter really it was that field and that kind of led me to like understand oh this whole field. So most of my college I basically just started like picking up design doing freelance and then kind of got here for a really small company called Pulse. It was a newsreading app. I joined them as a first designer. I was a daily user of Pulse. So as soon as I found out that you were working on that I was like this is amazing I've always wondered who did it.

是的，Pulse 让我觉得生活很有趣。他们找到我是因为我写了一个 Hacker News 主题修改器 Georgifi，那里的网页工程师至今还在用，我发誓我已经 10 年没更新了，我都不知道它怎么还在插件商店里，但它确实在 Chrome 商店里。我就是这样进入这个领域的。从那里到 Facebook，再到另一家初创公司、Airbnb，然后 Apple 一直是我想去的地方，很大程度上是因为 Steve Jobs 在 2000 年代末引领了设计的价值，我觉得至少需要理解它的运作方式。显然，我加入时连 Jony 都已经离开了。但那是一次非常有趣的经历。之后我又像钟摆一样在大公司和小公司之间来回切换。所以当我又感到渴望时，我想加入一个更小的团队。AI 正在发生，而且很明显它正在以惊人的速度发展，你需要一个真正促进这种速度的环境。这就是我来 Luma 的原因。

Yeah, the Pulse I feel like life is really funny. The folks found me because I had written this Hacker News theme modifier Georgifi and really the web engineer there uses it to this date I have not updated it I kid you not in 10 years I don't even know how it's in the plug-in store but it is in the Chrome store and so that's kind of how I got in here from there Facebook another startup Airbnb and then Apple was always just like I got to work there for the most part in ways like product design digital design do owe it to Steve Jobs to usher in like the value of design in the late 2000s that was like I need to at least understand what makes this tick. Obviously by the time I joined even Jony had left. But that was like a really interesting experience and from there I was like back to I keep doing this like pendulum of like big company small company big company small company and so that was kind of when I was itching again of like okay I want to join something a little smaller. AI was happening and it was pretty clear this is moving at breakneck speed and you needed to be in an environment that really promoted that. And so that's kind of what led me to Luma.

快速插播一条消息，然后我们继续。我全力支持 Paper 作为下一个伟大的设计工具。这话出自一个教过比任何人都多的人使用 Figma 的人之口。但我准备好了。是时候用一款将创意置于核心、而非系统和流程的工具了。我们设计和创建软件的方式变化比以往任何时候都快。而我们正在使用的工具是为昨天的任务打造的。相反，Paper 将设计师放在首位，并构建了我一直想要的各种功能。所以，如果你想成为第一批使用下一代设计工具的人，请访问 dive.club/paper 加入列表。早期访问将于 5 月启动。网址是 dive.club/paper。你知道我是 Jitter 的超级粉丝，但他们最新的版本无疑是最好的。现在，Jitter 拥有无限画布用于动画设计。使用 30 秒内我就知道这应该是动态设计的方式。它让迭代变得容易 10 倍。与团队成员协作更有意义。你可以通过单个文件中的多个画板跨格式缩放内容。Jitter 太棒了。这个产品现在好得令人难以置信。你一定要试试。只需前往 dive.club/jitter 开始使用。网址是 J I T E R。好了，现在回到节目。

Real quick message and then we can jump back into it. I'm all in on Paper as the next great design tool. And this is coming from someone who's taught Figma to more people than just about anyone. But I'm ready. It's time for a tool that puts creativity at the heart of everything, not systems and processes. I mean, the way that we design and create software is changing faster than ever. And the tools we're using were made for yesterday's tasks. Instead, Paper is putting designers first and building all kinds of features that I've always wanted. So, if you want to be one of the very first people to use the next generation of design tooling, then head to dive.club/paper to get on that list. Early access launches in May. So, that's dive.club/paper. Look, you know how big of a Jitter fan I am, but their latest release is hands down the best one yet. Now, Jitter has an infinite canvas for animation design. And I mean, within 30 seconds of using it, I knew that this was always the way that it should be for motion design. It makes iteration 10 times easier. Collaborating with team members makes way more sense. And you can scale content across formats using multiple artboards in a single file. Jitter is crazy. Like, this product is so incredibly good now. You got to try it out. Just head to dive.club/jitter to get started. That's J I T E R. Okay, now on to the episode.

你能给我们介绍一下你加入时产品的状态，以及最初几个月或项目在设计方面是如何取得进展的吗？

Can you give us a rundown of the state of the product when you joined and what those first few months or initiatives looked like where you were getting momentum on the design front?

Luma 的旅程很有趣。我第一次见到创始人是在 2 月。他联系我只是想多了解如何招聘设计师。当时他们还在做很多 3D 和 NeRF 相关的工作。他们试图弄清楚应该找什么样的设计师。最初的对话就是这样开始的。随着我更多地了解他们想做什么以及遇到的问题，我觉得这其实很令人兴奋，让我们多聊聊，看看是否有合作的可能。那是 2 月的事。当时我们还在讨论为什么是 3D，虽然视频模型刚刚开始出现，但大家都不确定它是否会成为主流。团队仍然非常专注，整个团队都有 3D 背景，所以他们非常关注 3D 方面，我就没有继续深入。我想，好吧，他们似乎对 3D 很感兴趣，但那不是我的专长。然后我们在 6 月左右再次联系，这时视频方面的想法已经开始萌芽，但公司仍然处于生成式 3D 领域。我当时还在 Apple。

The Luma journey was interesting. I first met the founder in Feb. He had reached out to like just learn a little bit more about how to hire designers. Back then they were still doing a lot of like 3D and NeRF. They were like trying to figure out okay what kind of designers should we go after. And so honestly the original conversations started there and then as I learned a little bit more about what they were trying to do what kind of problems they were running into I was like actually this is kind of exciting let's talk more let's try and see if there is something here. That was in Feb and then this was when I think we were talking and it was a little bit of why 3D though I think video models had just started coming but it was one of those like yeah we don't know if it's actually going to be a thing anywhere. And so the team was still very much like focused, the entire team was 3D background and so they were very focused on like the 3D aspects and so I let that go. It was like okay they seem very interested in 3D that's not particularly my thing and so then we caught up again in like June and by this point I think seeds of the idea of the video side of things had started happening but the company was still in the generative 3D land of things. I was still at Apple.

整个对话更多地围绕将生成式与 3D 结合起来意味着什么？你将设计描述为知道什么是可取的。那么，你能稍微展开一下，谈谈这如何塑造了你设计界面的方式吗？

The entire conversation was a lot more around what would it mean to bring generative and 3D together? You described design as knowing what is desirable. So, can you unpack that a little bit and talk about how that shapes the way that you've designed the interface?

我认为这是那种老生常谈的设计与艺术之争。但实际上，真正设计精良的东西确实具有艺术品质。我们能深入探讨一下从 0 到 1 的设计过程吗？设计师必须把脑海中的想法具体化，有时这些想法受到其他 10 个想法的启发，但最终需要有人将其解决成有意义的东西。既然你在这种 AI 原生产品中工作，你是否看到设计的价值主张以其他方式转变？你现在有机会做更多事情。做不做由你决定，但前提是别人会愿意去做。

I think this is one of those perennial quintessential design versus art and blah blah blah. But really, the things that are really well designed do have an artlike quality. Can we dive into the 0 to 1 design process? The person designing has to actually take the idea that's in their head, in some cases influenced by 10 other ideas people have built upon it, but then someone has to resolve it into something that makes sense. Are there other ways that you've seen the value proposition of design shift now that you're working in this type of AI native product or you now have the opportunity to do more. Whether you do it or not is up to you. The caveat being someone else will be willing to do it.

公司之前发布了一个叫 Genie 的文本转 3D 模型。这展示了我们可能成为一家生成式公司的理念：可以吸引合适的研究人员，获得合适的算力来训练这些模型，等等。所以这次整个对话更多地围绕“将生成式与 3D 结合意味着什么”展开。在整个对话中，我基本上一直在试探：3D 作为纯粹的 3D 与更广泛的创意之间有多大的弹性空间。一旦我发现确实有足够的空间，而且没有人说“不不不，我们为什么要做别的？3D 才是正道”，我就觉得这是个很好的时机。另一个让我觉得和这些人交谈非常有意思的原因是，我认为这是当时唯一还在谈论产品的研发公司。我聊过当时大多数做类似事情的公司，它们基本上都是纯研究实验室，而我作为一个职业生涯都在做产品的人，这有点学习上的挑战，因为真正的研究实验室与研发产品公司之间的激励机制、时间尺度等都略有不同。所以这两点契合了，这就是我加入 UMA 的原因。但我加入那天，公司严格来说还在做生成式 3D。我记得第一周我们开了一次团队会议，团队大概有 10 到 15 个人，我记不清了。CEO Ahmed 说，我们应该讨论下一步做什么。那时才提到视频模型的想法，而且和行业人士、投资者交流后，发现这个方向正在升温。那时我才意识到，哦，我加入了一家视频模型公司。经典的“入职一周后转型”，你懂的。对我来说，这其实是个幸运的意外。我觉得这太棒了，因为我一直想拓宽领域，为什么只局限于一种模态？从某种意义上说，AI 领域真正的突破有两个：一是语言现在成为输入，二是任何模态都可以是任何模态。这是两个最根本的解锁。在那个世界里，试图把自己限制在一种模态上，对我来说总是有点短视。所以这对我来说简直是再好不过的消息了。

The company had released a text-to-3D model called Genie. And so that had shown the idea of like okay this can be a generative company. We can attract the right kind of researchers. We can get the right kind of compute to train these models and all of that stuff. And so this time the entire conversation was a lot more around like what would it mean to bring like generative and 3D together. And throughout the whole conversation, I would basically do this thing where I would try and understand how much flex there was in 3D as like only 3D versus creativity at large. And once I saw there was like enough give there honestly and nobody was like no no no why would we ever do anything else? 3D is where it's at. I was like okay this feels like a very opportune time. The other thing that really made it very interesting to talk to these folks was I think it was the only research company that still was talking about product. I had talked to like most companies uh that were doing stuff back then and they were pretty much like full-on research labs and like trying to be a product person who has all their career done product. It is a little bit of a learning challenge because the incentives the like time scales all of those are just a little bit different when you're in a proper research lab versus when you're in a research and product company. And so those two stars align and that's kind of what led me to UMA. But the day I joined the company was still technically doing generative 3D. I think the first week we had this like group team huddle. I think the team was what like 10 15 folks I don't even remember. Ahmed the CEO he's like so we should talk about what we're going to do next. And that's when like the idea of the video model and that that's something that you know in talking with folks in the industry investors that that's something that's been picking up steam. And that's really was when like, oh, I guess I've joined a video model company. Uh, the classic pivot a week after you joined, you know. Yeah. I I think like in my case, it turned out to be a happy accident. I was like that's amazing cuz I had always wanted to broaden the sphere of like why one modality in some ways the real like unlock in the whole like AI stuff there are two. One language is now an input and then two any modality can be any modality. Like those are the two most like foundational unlocks. And so in that world like trying to limit yourself to one modality always felt a little like shortsighted to me. And so for me this was like this could not be more amazing news.

好的。那我们能深入聊聊从零到一的设计过程吗？我想了解，在公司你入职一周后有点转型的情况下，你是怎么弄清楚这产品可能是什么的。而且你在录音前跟我说过一句很有意思的话：从零到一是你打上自己印记的机会。那么你的印记是什么样的？

Okay. So can we dive into the zero to one design process? I want to learn about how you even figured out what this could be after you know the company kind of pivots a little bit a week into joining. And you said something interesting to me before we started recording. You said zero to one is your opportunity to put your signature on it. So what does your signature look like?

我认为这就像那些老生常谈的设计与艺术之争，等等。但真正优秀的设计确实具有艺术般的品质。是的，理论上我理解艺术不是设计，设计也不是艺术，但它们确实有一种特殊的品质，不是理性的设计方法。所以我说打上印记，并不是指仅仅为了留下痕迹而做点什么。我的意思是打上印记，让它拥有某种本质或灵魂，能在更本能的层面与一些人对话。如果你坐在设计评审会上，有人问你为什么这么做，你诚实的回答会是“因为我觉得这样更好”，真的没有道理可讲。这就是我所说的打上印记。作为设计师，这听起来可能有点玄，但我认为设计师的基本职责之一是决定什么东西应该存在于这个世界上，这是一个巨大的机会，但我也认为这是一种道德责任。尤其是在不同公司工作过之后，我对商业、商业激励、用户激励等有了更全面的理解，但我还是回到同一个想法：是的，这些都没问题。但整个设计职业之所以存在，是因为我们需要一个为用户而战的职能。现在人们说设计师应该谈论业务指标、利益相关者对齐等等。是的，但整个职能的诞生是因为我们需要一个只为用户而战的职能。所以我觉得很有趣，现在我们看待设计时，认为它应该做这个、平衡那个。是的，但作为职能，我们也有道德责任：我们选择将某些东西带入世界。在那一刻，你对事物的形态有最大的影响力。所以我说印记，指的是比仅仅因为觉得酷而做 fancy 事情更深层的含义。

I think this is like one of those perennial quintessential like design versus art and blah blah blah. But really like the things that are really well designed do have an artlike quality. Like yes, I understand theoretically art is not design and design is not art, but they do have that like special quality that is not a rational approach to design. And so when I say putting your signature, I don't mean like just doing something cuz you want to leave your mark. I mean putting a signature so that it can have somewhat of an essence or soul that speaks to some people on a more visceral level. There's no you couldn't like honestly rationalize it if you sat in a design critique and someone asked you why did you do it this way and honestly the answer will be cuz I just thought it was better like really there is no rhythm or rhyme for that kind of reasoning. So that's what I mean by putting your mark. As designers like this is going to sound like too heady but I think as designers one of the fundamental things we are charged with is what should come into existence in the world and that is you know a great opportunity but that is I also think that is a moral responsibility especially having worked at like different companies uh I feel like I have a fuller picture of what you know business means business incentives means what user incentives means and all of that stuff but I come back to the same idea. Yeah, all of that's fine. But the whole profession of design came into being because we wanted a function to fight for the user. Now it's like, well, designers should talk business metrics and this and stakeholder alignment. It's like, yeah, but the whole function really, it was brought into existence because we needed a function to fight for just the user. And so it's like really interesting to me that we now look at this as like well design should do this and balance this and balance that. It's like yes, but we also as a function have this moral responsibility of like we are choosing to bring something into the world. At that point, you have the most influence on what the shape of something can be. And so when I say signature, I mean it at that like deeper level than you know just doing something fancy cuz you thought it was the coolest thing ever.

我在产品中感受到了这一点，不管怎么说。有好几次，甚至在我打开灵感面板时，那个小背景动画让我进入一种近乎平静的状态，我就想，他们本不必这么做，但他们做了，我很欣赏。所以那个产品里确实有很多这样的细节，很明显，我也会转达给团队。

I felt that in the product for what it's worth. Like there were multiple times where I think even when I hit the brainstorm panel and I felt like this little background animation and it kind of took me to this like almost peaceful state where I was like okay they didn't have to do that they didn't have to do that and they did and I appreciated it. So uh there's definitely a lot in that product where it's clear and I'll pass it on to the team as well.

这是我学到的另一件事。我认为你可以通过和志同道合的人在一起，让生活变得非常轻松。有些环境，你知道，人们经常问——我想我们在面试工程或设计候选人时也被问过——比如“你是怎么说服 XYZ 投资这个的？”老实说，答案是我们不需要说服。我们面试了每个人。很多核心团队成员是我带来的，关键是我们在哲学上一致。我们共享一个世界观：我们应该如何制造东西？什么样的东西应该存在？在其中投入多少关爱？设计师很擅长为我们想要的任何东西辩护，有点靠忽悠。但当大家真正一致时，就很神奇。你不需要证明什么。只是说，好吧，我们有时间限制，怎么最高效地完成？这时，有经验、在其他地方做过类似事情的团队成员会让你的生活变得极其轻松。

So this is the other thing that I've learned. I think you can make your life very easy by just surrounding yourself with like-minded people. There are environments, you know, often time people will ask, I think we've been asked when like we interview candidates for engineering or design where, oh, how did you like convince XYZ to invest in this? And honestly, I think the answer is we didn't have to. We interviewed every person. A lot of the core team I brought them on and like the whole point was there was this philosophical alignment. We shared a worldview of how should we make things? What kind of things should come into existence? What kind of love and care is put into that? Designers are really good at justifying anything we want. But a little bit of like, you know, bullshitting our way through it. But really, it's magic when everybody is like aligned. You don't have to justify things. It's a little bit of like okay but we have this kind of time constraint. How do we do it in like the most efficient way? And that's where like you know having experienced team members who've done this before in other places makes your life ridiculously easy.

我采访过很多真正注重工艺的公司的设计师，以前我总想问他们：“你们是怎么做到的？到底怎么做的？”但现在我已经不再问这个问题了，因为答案非常清楚——就是每个人都比普通公司的人更在乎，就这么简单。我觉得这是我在 Facebook 的 Paper 团队学到的。Paper 团队，可惜已经没了。我问我的朋友 Brandon：“你怎么能 justify 这么多工作，比如这个纸卷动画之类的？”他的回答是：“因为我们面试了每一个加入 Paper 团队的工程师，如果他们不认同，就不能加入团队。”就这么简单。我当时想：“好吧，我明白了。”所以现在，在多个地方看到之后，是的，你要么懂，要么不懂。

I've interviewed so many designers at these really craft-oriented companies, and I used to try to get at, 'Oh, how are you doing it? How are you doing?' And I've actually stopped asking the questions because it's super clear. It's just like, well, everyone just cares a lot more than your average company, and it's as simple as that. I think I learned this at Facebook from the Paper team. Paper, RIP. But I would ask my friend Brandon, 'How are you able to justify so much work and like this paper curl animation and all of this stuff?' And his answer was, 'Because we hired and interviewed every single engineer who joins the Paper team, and if they're not aligned, they don't get to join the team.' It's as simple as that. I was like, 'Okay, I see.' And so yeah, now having seen it at multiple places, yeah, you either get it or you don't get it.

好的，我想从稍微不同的角度来谈谈这个想法。现在我们讨论的是以我们认为正确的方式构建产品。还有一个概念叫“渴望性”，你第一次和我聊天时提到过。你说你把设计描述为“知道什么是令人渴望的”。那么你能稍微展开一下，谈谈这如何塑造了你设计界面的方式吗？

All right, I want to talk about this idea but from a slightly different angle, where right now we're talking about building products the way that we believe products should be built. And there's also this concept of desirability, which you mentioned to me the first time we chatted. You said you described design as knowing what is desirable. So can you unpack that a little bit and talk about how that shapes the way that you've designed the interface?

我相信的是我们试图探索的想法的本质，我们可以真正放大它。当然，有工作要做。你知道，使用它的人——很多 AI 创作者、视频制作人、平面设计师——最终，他们当然是在寻找一个素材。他们可以把它带到他们的工具里，无论他们当前的生产流程是怎样的。这当然是知道工作是什么。但除了典型的“给个提示，说点什么，你就能得到一个视频，然后走人”之外，我们还能带来什么？从商业角度来看，这对我来说并不具有粘性。一旦更好的模型出现，你就会切换。我们在 LLM 领域和其他领域都看到了这种情况。所以这始终是我们——你可以说相信，或者说我们想得足够深，知道产品必须给你别的东西。这是第一个想法。第二部分是因为语言作为输入。另一条线索是，你必须小心不要拉得太紧，尤其是在像 Dream Machine 这样的专业环境中，但存在另一个实体的概念。你可以试图抗拒它，我认为 Chad Gibb 做得很好，他坚持说“不，这是一个工具”。但你是在和这个东西对话。在深层的潜意识层面，你永远无法把它看作一个工具。你必须把它看作比工具更多一点的东西。然后由你决定你想在多大程度上将其人格化等等。所以对我们来说，这是第二部分：如何将其置于创意工作流程中？如果你仔细想想，创造力实际上是相当孤立的。这是一个非常孤独的活动。无论你和多少人一起即兴创作和头脑风暴，当关键时刻到来时，设计者必须真正把脑海中的想法——有时受到其他 10 个想法的启发——但最终必须有人将其解决成有意义的东西。在这个过程中，没有人陪着你。这真的是孤立的。我们的想法是，如果我们能设计整个体验，让它带有一点点创意伙伴的感觉，不让人觉得这是一个只等待命令的仆人，而是它有自己的想法。如果你是一个新手，它可以帮助你的想法变得更好，因为我们看到很多其他情况：很多对设计、摄影、电影不太了解的人进来，只是从各处捡起提示词然后拼凑起来。你会看到像“Octane, 8K renderer, glitchy noise”这样的东西。我想，这些术语放在一起根本没有任何意义。所以我们希望，如果你使用这样的东西，a）它应该感觉像是一种伙伴关系。不仅仅是你必须与机器合作并告诉机器你想要什么。它应该感觉像是一种伙伴关系。b）随着你越来越多地使用它，特别是如果你不是这个领域的资深人士或没有正规的创意艺术背景，你可以学习。所以“Brainstorm”很大程度上就是从那个角度出发的：你可以带着最简单的想法来。我设计整个语言架构时，我们最喜欢的测试流程的例子是“红色厨房”。你是什么意思？红色乡村厨房？红色现代意大利厨房？红色农舍厨房？什么样的红色厨房？一方面，我们展示这些变体，但 Brainstorm 让你在这些方面更深入，这样下次你来的时候，你就学会了词汇，学会了其他方式，学会了红色如何与赤陶色联系，赤陶色如何与砖色联系，然后整个厨房设计中的材质性游戏。这就是我们想要注入的那种东西，让它感觉与众不同。所以这些是我们试图在对话作为基础之上构建的东西，然后是创意伙伴的想法。当时，我认为“智能体”甚至还不是一个术语。回想起来，很明显我们只是触及了那个想法的表面。但当我们开始时，想法非常明确：不是试图构建一个 AI 艺术家，而是想构建一个创意伙伴。

The thing that I believed was the essence of the idea that we were trying to explore, and we could really amplify that. Of course, there's a job to be done. You know, people who are using it — a lot of AI creators, video production people, and people in graphic design — end of the day, of course, they are looking for an asset. They can take it to their tool and whatever their current production pipeline flows are. That is knowing, of course, that's the job to be done. But what can we bring to the table where it's not just your typical 'Here's a prompt, say something, you'll get a video, off you go'? That doesn't feel, to me, even from a business perspective, sticky. The day a better model comes, you will switch. And we've seen this play out in LLM land and different lands. So that was always something we — you can say believe in, or you can say we thought deep enough about it that we knew the product had to give you something else. So that was one idea. The second part of this was because of language being an input. One of the other things that is a thread that you've got to be a little careful in how much you pull on it, especially in a professional setting like Dream Machine, but there is this idea of another entity that is there. You can try to resist it, and I think Chad Gibb does a good job of really resisting — 'No, it's a tool.' But you're talking to this thing. You cannot, at a deep, deep subconscious level, ever think of this as a tool. You'd have to think of it a little more than a tool. And then it's up to you how much you want to personify this and all of that stuff. So for us, that was the second part: how do you contextualize this in the creative workflow? If you think about it, creativity is actually pretty isolating. It's a very solitary activity. No matter how many people you're jamming and brainstorming with, when the rubber meets the road, the person designing has to actually take the idea that's in their head — in some cases influenced by 10 other ideas people have built upon it — but then someone has to resolve it into something that makes sense. And in that, there is no one with you. It is truly isolating. And our thinking was that if we could design the whole experience where it could have just a dash of creative partnership, where it didn't feel like this was a servant just waiting for your commands, but that it had ideas of its own. It could help your ideas get better if you were someone who was a novice, because that was a lot of the other things we were seeing: a lot of people who didn't really know a lot about design, photography, film coming in and just picking up prompts from here and there and putting them together. You would see things like 'Octane, 8K renderer, glitchy noise.' I'm like, those terms don't even mean anything together. And so our hope was that if you use something like this, a) it should feel like this is a partnership. It's not just you having to work with the machine and tell the machine what you want. It should feel like a partnership. And b) as you use it more and more, especially if you're someone who hasn't been deep into this field or has a proper creative art background, you can learn. So 'Brainstorm' was very much from that lens: you can come with the simplest idea. Our favorite example of testing these flows when I was designing the entire language architecture was 'red kitchen.' What do you mean, red rustic kitchen? Red modern Italian kitchen? Red farmhouse kitchen? What kind of red kitchen? One aspect is we show those variations, but Brainstorm lets you kind of go deeper on these, so that next time you come, you learn the vocabulary, you learn other ways, you learn how red can be connected to terracotta, how that can be connected to brick, and then their entire play of materiality in kitchen design. That's the kind of stuff we wanted to imbue in terms of what makes it feel different. So those were the things we were trying to build on top of conversation as a foundational thing, and then this idea of a creative partner. Back then, I don't think 'agent' was even a term. Looking back, it's pretty clear we were just scratching the surface of that idea. But when we started, the idea was very much: instead of trying to build an AI artist, we wanted to build a creative partner.

设计 AI 产品时，有没有一些挑战是你直到真正加入像 Luma 这样的公司才意识到的？

Are there any challenges that come with designing an AI product that you might not even be aware of until you actually join a company like Luma?

很多。我认为时间尺度非常不同。它们既不同又有点不可预测。不可预测的原因是研究是不可预测的。我经常这么说。我认为这是设计第一次遇到一个比过去 10-15 年的设计更不按时间表来的职能。这真的不是因为有人故意难以合作。而是因为他们实际上在发明以前不可能的东西。即使是一篇别人撰写并发表的论文，团队试图采用那些方法，那些论文也是从研究的角度设计和撰写的。它们不是从产品化的角度设计和撰写的。所以当你开始产品化时，你会意识到，“哦，这个方法，我不知道，太贵了，太慢了，实际上无法大规模收敛，只适用于精心挑选的数据集，而不是真正的大规模。”

Lots of them. I think the time scales are rapidly different. They're both different and a little bit unpredictable. The reason they're unpredictable is because research is unpredictable. I say this all the time. I think this is the first time design has met a function that is even more timeline-averse than design has been historically for the last 10-15 years. And it really is not because someone is trying to be difficult to work with. It's because they're actually inventing stuff that wasn't possible before. Even when it's like a paper that someone else has written and published, and the team is trying to just adopt those methodologies, those papers were designed and written from a research POV. They were not designed and written from a productization point of view. And so when you start productizing it, you realize, 'Oh, this method is, I don't know, too costly, too slow, doesn't actually converge at scale, is good for handpicked datasets, but not really at scale.'

然后你必须解决每一个问题。而每一个问题的解决方案都是未知的。这不是工程，至少不是过去 15 年我们看到的常规工程——比如，哦，这个不行，我们知道另外五种方法可以解决这个性能问题。不，这真的是未知的。可能他们一个周末就能搞定，但也可能得花 3 个月才能解决。这让事情变得非常不确定。但加上当前 AI 领域的竞赛，意味着一旦你有了成果，尽快推向市场对公司和对用户都是最有利的。

And you then have to solve each of those things. And each of those things, the solve is unknown. This is not engineering, at least like normal engineering that we've seen for the last 15 years, wherein, oh hey, this thing doesn't work. Oh yeah, we know five other ways we can solve this problem of performance that we're hitting. No, this is really it's unknown. It could be they can figure it out in a weekend, but it could also be it'll take them 3 months to figure it out. And that makes it very uncertain. But coupled with the race that is AI field right now means the moment you do have something ready. It is in the company's best interest and your users's best interest to get that out in the market as soon as possible.

我认为从 0 到 1 的阶段非常擅长建立一些大的边界，比如创意合作、一个核心的构思等等。但一旦你有了这些，新能力就像不断涌现，你要弄清楚它们如何融入产品，然后尽快推出。因为如果你不这么做，别人就会做——现在人们很容易切换到他们喜欢的任何工具。所以从公司的可持续性以及用户将其用于关键任务项目的角度来看，越早推出，对他们越好，他们就能越早尝试这些功能，模型也会变得更好。

I think the 0 to 1 phase is really good at trying to establish some I would say big borders of like okay creative partnership one big thing ideation as a core thing okay one big border to establish but once you have those like these new capabilities are a little bit like as they come figure out how they fit in the product and get them out the door because if you don't someone else will right now like it's very easy for people to switch to whatever tool that is of their preference and so from a company's sustainability perspective from users like using this on a project that is mission critical to them. The sooner it gets out, the better it is for them, the more they can try these things and better the model gets.

所以，你可能第一个月领先一点。但因为你的反馈循环，下一次更新时，你就比别人快两个月。这个差距只会越来越大，复利效应是真实的。在 AI 领域，这一点更加明显。那么如何平衡呢？棘手之处在于，如果你的产品最终只是核心模型的一个包装，无论是你的模型还是别人的，那没问题——你可以不断发布新功能，每个模型、每个能力都变成一个按钮。但当你进入智能体式工作流时，你会有对话，有各种不同的路径，这就变得非常棘手了，因为你必须设计产品。我认为我们正在学习的是，你必须把产品设计得极其模块化和有弹性，能够随着新能力的出现而热插拔。这对我们来说是一个很好的教训，因为过去我们试图做成一个庞大的单体结构，很明显现代化才是我们应该采用的方式。但正是这些因素让事情变得特别困难。最后，你无法预知边缘场景。我的意思是，我们从 2007 年的电容触控开始，很快，大概三四年内，我们就有了工具来真正感受电容触控解锁了什么。一开始大家还在做快速按钮。直到我们有了滑动解锁，才意识到手势是这个媒介的全新能力。从那时起，许多设计师开始真正体验这个媒介。他们设计时仿佛在控制手势，控制触摸点，在 Origami 或任何你使用的工具中，你可以控制所有这些。这在研究领域是不存在的。你可以用 Comfy，但不一样。你必须建立一些关于技术发展弧线的直觉。你需要在脑海中看到：曲线是什么？它们走向哪里？落在哪里？在很多方面，你不得不猜测：大概两个月后我们可能会有这个能力，六个月后我们肯定会有这个根本性的阶跃变化能力，然后你从那里倒推。比如，如果这个能力大约在六个月后落地，那么我们现在不应该做出哪些根本性决策，因为六个月后它们会完全错误。不是说不能改变，但有些事情太基础了——如果这个能力在三到六个月后落地，你为什么还要做这个决定？所以你必须向前思考，我认为这不是我们过去需要具备的特质。说实话，我们过去看到问题，知道手头的工具，然后解决它们。即使是技术和触控设备，怎么做也很清楚。现在你还需要假设一点技术弧线，所以你必须更接近这个媒介。你必须更了解正在发生的研究，比如我们离实时视频模型还有多远？你至少应该有一个理论上的答案。基于此，你日常的许多决策可能会改变，你甚至可以在某个时候引导研究朝那些方向发展。这就是不同之处。

And so, you know, you're probably first like 1 month soon. But because of your feedback cycle, next time you do an update, you are 2 months sooner than everyone else. And that gap only increases like compounding is real. And in AI field, it is like even more so real. So how we balance that with like you know the tricky part is if you don't do that in a way where end of the day your product is just trying to be a wrapper around the core model whether yours or someone else's if you're doing that that's fine you can just keep shipping things each model each capability becomes a button but when you get that in like an agentic workflow you have conversation you have all these like different pathways that's where it gets like really tricky cuz you have to design the product this is something that I think honestly we are now learning that you have to design the product in a way where these are extremely modular and resilient and you can just hot swap them as new capabilities come and so that's like a good lesson for us I would say cuz I think we tried to be one big monolithic thing in the past it is pretty clear modernization should have been the way we had approached this but those are the things I would say that makes it like particularly difficult and lastly you don't know the kind of edge scenarios in that like what I mean is we started with like in 2007 with capacitive touch And soon within I would say 3 4 years we had tools where you could actually feel what capacitive touch unlocked. And in the beginning everyone was still doing really fast buttons. And it wasn't until we got the slide to unlock that we realized oh gestures are a whole new capability of this medium. From that point onwards many designers were like actually experiencing the medium. They were designing these as if like they were controlling the gestures. They were controlling the touch points and in origami or whatever tool you were using, you could have control over all of these things. That doesn't exist in research land. You can kind of use comfy, but it's not the same. And you have to build a little bit of like intuition of where the arc is going. Like you kind of in your head be able to see like okay, what are the curves and where are they going? Where are they landing? And in many ways you kind of have to like guess a little bit of like I think in like two months we will likely have this in six months we will definitely have this like fundamental step change capability and then you kind of like work a little bit backwards from that. It's like okay if that is roughly landing in 6 months what fundamental decisions we shouldn't be making right now because in 6 months they will be completely the wrong decisions. Not to say you can't ever change, but then there are some things that are so foundational that you know it's like okay if this is landing in like 3 to 6 months why would you make this decision and so you do have to like think a little bit forward which I think is not a trait that I think we've ever had to honestly we we see the problem. We know the tools at our hand and we solve them. Even with like tech and touch devices it was pretty clear how to do that. Now you kind of have to also assume like the arc of technology a little bit and so you do have to be a little closer to like the medium. You do have to be a little aware of like what is the research happening like you know something like how far are we from real-time video models you you at least should have some theoretical answer to that question and based on that a lot of the decisions you are making dayto-day might shift you can nudge research at some point towards those directionalities so that's how it's different

你能分享一个例子吗？比如哪种问题你可以做出战略决策，认为这不值得解决？

Is there an example that you can share of a type of problem where you can make the strategic decision like this is not worth solving?

我可能更多能说的是我们通过艰难方式学到教训的例子。对于智能体相关的工作，我们当时还在用 3.5 模型，我开始写基本的系统提示词之类的东西。这个过程的第一步是，语言不应该成为我们投资的方向，并让大家达成共识。然后问题是，你究竟如何解决这个问题？所以我们几个人聚在一起，我写了一个粗略的 Notion 文档，列出了思考这个问题的广泛方式。非常宽泛地说，我觉得用“架构”这个词来形容我写的内容可能都有点尴尬，但大致就是这些阶段：智能体应该如何思考，什么时候应该重新思考和评估所有东西。然后开始工作：理论上这一切都很好，证明它吧。于是开始写第一个系统提示词，安装所有工具，花时间在终端里，因为这是最快的方式。整个过程我们都在用 3.5 模型，所以不得不创建大量特殊规则：如果发生这个，就做这个；在这种情况下，就这样做；这是你分解问题的方式。结果呢，现在有了 4 和 40 模型，我们意识到一件事：每次你用大语言模型处理这类问题时，你实际上是在给自己的智能设限，因为你写的那些逻辑路径和规则本身就是限制。

I can probably tell more of the ones where I think we learned this the hard way. So for the agentic stuff we were still dealing with 3.5 when I had started like writing the basic system prompts and stuff and the way that process worked was the first step was like no language should be something we invest in and kind of getting everyone aligned on that idea to like okay but like how are you even approaching this problem you know so like a few of us got together I just wrote a rough notion doc of these are the broad like ways to think about the problem this is very broadly speaking like I think architecture is probably even a cringe word to describe what I wrote in that document, but roughly this is these are the like stages. This is how like the agent should think. This is when it should like rethink and re-evaluate all of that stuff. From there began the work of okay, this is all great in theory land. Prove it. And so like writing the first system prompt to like just seem like you know getting all the tools installed and spending time in terminal because it was the fastest way to get this all done. And so that whole process we were dealing with 3.5 and so we had to create so many of these like special case rules like if this happens do this in this case do this this is how you break it down. Lo and behold now like 4 and 40 like I think one of the things we are realizing is every time you approach in the LLM side of things a problem like that you're putting a ceiling on the intelligence yourself because the way you wrote those logic pathways and those rules are the limits.

无论模型变得多聪明，它也只能做那些事，因为你明确告诉它不要做别的。事后看来，我本该用上自己刚才那套关于“知道发展轨迹”的智慧。我本该对自己说：“不，不，不。等到我们发布的时候，我们就会拥有它。”这正是一个典型例子，说明方法学如何随时间快速改进。但这个问题很棘手。我认为 AI 领域始终存在一种拉扯：你可以再等两个月，更好的方法就会出现，但你也确实等了两个月。我觉得这得具体情况具体分析，每家公司、每个功能，你都得做出判断：什么时候值得把现有的东西发出去，什么时候其实还不够好，得等更好的模型出来再发——因为总会有更好的东西。我认为这些领域都远未触及天花板，视频模型也一样。没错，三个月后我们可能会有更好的视频模型，六个月后还会有更好的。但显然我们不能一直等到理想状态出现，因为它永远不会来。不过在核心架构方面，知道这一点会有点帮助，你可以围绕它来规划工作。否则你只会不断重写。比如我们大概一个月前重写了智能体架构，因为结构化响应之类的功能出来了，我们觉得这比我们之前不得不写的所有逻辑都好太多了。对吧？所以我们就重写了。这没关系。

No matter how intelligent the model gets, that is all it can ever do because you're literally telling it not to do anything else. And I think in hindsight, we should have had this wisdom that I was spewing a second ago of knowing the arc and everything. I should have taken that myself and been like, 'No, no, no. By the time we launch, we would have it.' That's a perfect case of where clearly the methodologies got better very quickly over time. This one is tricky though. I think there is a constant push and pull in AI of like you can wait for two more months and a better method will come, but you will also have waited for two months. And I think this is on a case-by-case basis, every company, every feature, you kind of have to make the call of when is it worth it to still ship the thing you have versus actually no, it's not good enough, hold on, wait for it to ship because you have a better model that we can't have, because there will always be a better thing. I think none of these are anywhere close to plateauing, even video models. Like yeah, in 3 months we will probably have a better video model, in 6 months even a better one than that. But obviously we're not going to just wait till the ideal happens because it's never going to happen. But in core architecture stuff, I think knowing that helps a little bit, and you can just plan your work around that. Otherwise you just end up rewriting. Like we rewrote the agent architecture I think a month before because structured responses and all of that stuff came out, and we were like, 'This is so much better than all the logic that we were having to do.' Right. And so we rewrote it. It's okay.

听你谈到有多少设计输出被捕捉到系统提示中，而且你多次用到“语言”这个词，甚至深入到终端，这很有意思。这个角色本身也随着媒介发生了巨大的变化。

It's interesting to hear you talk about how much of the design output is being captured in the system prompts, and you've used the word 'language' quite a few times now, and even getting into terminal, and just how much the role is even changing based off of the medium.

我认为可能有两部分。一部分是我个人对自己所做之事的看法，以及我想做更多的事。过去十年左右，我并不觉得自己是个 UI 设计师。我做得还行，但从未真正有那种感觉。Facebook 是第一次让我接触到“产品设计师”这个概念，那正是我们的职位头衔。其理念是你的工作不仅仅是创造设计或像素输出，你的设计实际上是要思考整个产品。然后在 Airbnb，我们被称为“体验设计师”，不是用户体验那种，而是真正思考客人最终会获得什么样的体验，以及任何能控制这一点的因素。这就是你作为设计师的角色。所以从那个广义的角度来看，是的，我会称自己为设计师。从技术 UI 设计的角度看，我觉得自己更接近产品架构师而非设计师。但老实说，有了 AI 之后，很多方面都变得更容易应用了。你能够表达“这是我脑子里的大致架构”，但现在你不仅能说出来，还能实际在 playground 里跑出一个粗略版本，展示给别人，然后那些比我聪明得多的人可以在此基础上构建。我认为这是设计含义的根本转变。以前有 UI 设计，我们有平面设计，我们看静态的东西。然后随着 UI 设计，交互变得很重要，比如原型设计和交互设计，要知道从状态 A 或视图 A 到视图 B，中间有 100 个微时刻把你从这里带到那里。这成了你必须考虑的事情。我认为在这个案例或隐喻中，我们只是走得更深了。你现在必须理解你可以通过什么材料来施加影响，因为它不需要是确定性的代码。现在它更像是概率性的语言，可以塑造“这可能会发生”或“那可能会发生”。这听起来很直观，但当你真正开始使用时，它非常不直观，尤其是要关闭你大脑中 10 到 15 年编程等非常确定性的思维方式。你开始设计，然后发现自己会停下来想：“等等，我为什么要这样做？那是老方法。”我会称之为思考这个问题的旧方式。让我们换个思路。所以我认为，知道存在哪些能力，以及有多少能力你可以仅通过写四行提示就能创造出来，比如 UI 中的概念药丸就是这样来的。

I think there are two parts to it probably. One is my personal view of what I think I do and what I want to do more of. I think for the last decade or so, I didn't feel like a UI designer. I was decent at it, but I never quite felt it. And I think Facebook was the first time where this idea of a product designer was exposed, like that was our role title. And the idea that your job is not to just create designs or pixels output, your design is to actually think about the whole thing. And then Airbnb, I think we were called experience designers, and not in like user experience. It really meant what is the end experience a guest is going to have, and whatever vector can control that, think about it. That's the role you have as a designer. So in that broad view, yeah, I would call myself a designer. In the tech view of UI design, I feel closer to a product architect than a designer. But a lot of this, I think, has become much easier to apply honestly with AI. It's so much easier to be able to express 'this is the rough architecture that's in my head' but I can't just talk about it now, I can actually go and execute a rough version of it in a playground and then show it to people, and then people who are way smarter than me can build on top of it. And so that, I think, is a fundamental shift in what design means. From before, UI design was there, we had graphic design, we would look at static things. Then with UI design, interaction kind of became a big thing, like prototyping and interactive design, and knowing between state A or view A and view B, what were the 100 in-between micro moments that happened that took you from here to there? That became a thing you had to think about. And I think in this case or metaphor, I think we're just going deeper. You now have to understand the material through which you can influence, because it doesn't need to be deterministic code. Now it is like probabilistic language which can shape 'this can happen' or 'then that can happen'. It sounds intuitive, but when you actually start working with it, it is very non-intuitive, especially trying to turn off your brain's 10-15 years of very deterministic ways of thinking in programming and stuff. You start designing and you find yourself catching yourself being like, 'Wait, why am I doing it this way? That's the old way.' I would call it the old way of thinking about this problem. Let's shift gears and think about it this way. So I think knowing what capabilities exist, how many of the capabilities you can literally give birth by just writing four lines of prompts, like concept pills in the UI came like that.

你能分享一下概念药丸的故事吗？

Can you share the story on the concept pills?

是的，我的意思是，它始于一个想法：哦，那会很不错，因为再说一次，人们只是把这些当作提示，但我们希望它感觉像一个构思伙伴。那么我们可以怎么考虑呢？当然，你总是可以问任何问题，所以在 Dream Machine 中，你当然可以得到视频、图片，但你也可以直接获得想法。你可以直接问“给我更多类似上面的想法”，它就会给你想法。在首次将其产品化的过程中，我们试图思考：当人们试图即兴发挥提示时，他们往往是新手，不是提示工程师。他们词汇量很有限，那么我们如何给他们提供选项，让他们理解“好吧，稍微更细致地谈论这个的方式是什么，仍然在同一个想法周边”。所以这不完全是头脑风暴，头脑风暴要广泛得多，而是说，不，在这个领域内，你还能怎么说？在这个领域里还有哪些东西可以让你快速探索，用潜在空间的方式来思考，即你当前所在潜在空间的更外围区域。我们就是这样开始的。起初，我们想，哦，让我们做一些项目匹配之类的事情，或者有一个硬编码列表。非常直觉地，我觉得那对我来说就是不对的。我想，我们有语言模型，为什么非得这样做？所以我就用 Claude 写了一些东西，我想是那时候，只是为了证明是否可行。事情很简单：对于给出的每个术语，找出正确的关键词集，然后对每个关键词，找出其他有创意的替代词。然后我不得不在那个系统提示上做了很多提示工程，因为否则有时它会开始写文章。然后就是，好吧，你如何清理响应的问题。但这就是最初想法的由来。一旦这个想法实现了，一些更像文章式的响应就引出了头脑风暴功能。我们想，“天哪，这东西实际上可以走得很深。”

Yeah, I mean it started off with the idea that, oh it would be nice because again going back to people were thinking of these just as prompts, but we wanted this to feel like an ideation partner. And so what are some ways we can think about it? Of course you can always ask anything, and so in Dream Machine, of course you can get a video, you can get an image, but you can also just get ideas. You can literally just ask 'give me more ideas like the one above' and it will give you ideas. In trying to first productize that, we were trying to think of okay, but when people are trying to riff on prompts, often they come, people who are novices are not prompt smiths. They would come with very simple vocabulary, and so how can we give them options that lets them understand okay, what is the slightly more nuanced way of talking about this, still in the same periphery of ideas. So it's not quite about brainstorm which is much more go broad, but this is like no, within this realm, how else could you have said it? What are other things that are in this realm that let you quickly explore in like the latent space way of thinking about it, the more peripheral regions of the latent space where you're right now. That's kind of how we started off with. At first it was oh, let's do some projects matching and this and that, or have a hardcoded list. Very intuitively, I feel like that just felt wrong to me. I was like, we have LMs, why the hell would you have to do this? And so I really just wrote a few things in Claude, I think at that time, just to prove out could it be done. And this was something as simple as: hey, for every term that's being given, figure out the right set of keywords, and for each of those keywords, figure out other creative alternates. And then I had to prompt-smith a lot on that system prompt, because otherwise sometimes it would start writing essays. Then it was like, okay, how do you sanitize the response problem. But that's kind of how the original idea came. Once that came, some of the more essay kind of responses is what led to brainstorm. We were like, 'Holy, actually this thing can go so deep.'

我们其实可以把另一篇不需要的随笔形式，重新包装成另一个功能。所以，如果你抽象地思考，你不太可能把头脑风暴看作一个功能。但当你在这个媒介中与这些东西打交道时，每个回应都是一种新方向的可能性。这就是概念构建的方式，也是头脑风暴的由来。它们都源于同一个出发点：只是和语言模型玩耍。

Let's actually take this other essay form that we didn't need here. Let's repackage it as another feature. So really, if you think in abstract, you wouldn't likely think of brainstorm as a feature. But when you're in the medium working with these things, each response is a possibility of a new direction you can take it in. And so that's how concept builds happen. That's how the brainstorm stuff happens. And they're all coming from the same place of just playing with language models.

一旦模型就绪，我们就回过头来优化：什么样的提示词更适合我们的模型？这样用户就不用过多考虑这些。他们只需带着头脑中最简单的想法来，不必费力进行提示词工程或这种文字炼金术。你可以自然地说话，做个人就好。只管来，表达出来，模型和中间的系统应该负责接收你的输入，并以最佳方式表达出来。

Once we had the models ready, then we hooked it back to optimizing: what kind of prompt works with our model better? So that the users don't have to think too much about this stuff. They can just come with the most simple version of the idea they have in their head and not have to work so hard with prompt smithing or this alchemy of words. You can just come and say things naturally. Just be human. Just come and express, and it should be the model and the systems in between that take your input and do what is the best way to express that.

云制品在那个时候是交接流程的一部分吗？还是你作为设计师只是确认了这是可行的？

Is the cloud artifact part of the handoff process at that point, or you just as the designer being like, okay, I have confirmed this is possible?

我认为当我们提出其中一些想法时，我们甚至还没有语言工程师。所以语言团队就是我和 iOS 上那个对此非常热情的人。他帮我设置了终端之类的东西，我们只是在这些想法上即兴发挥，真正去验证它。不，这是可行的。所以当时只有我在做提示词工程，因为真正特别的地方在哪里？你只是用 50 种不同的方式让它做事，看看哪些在广泛范围内有效。我认为评估的概念当时还不存在；这完全是凭感觉，因为那是早期。所以我们只是觉得，好吧，这对这些查询感觉不错。

I think when we came up with some of these, I don't think we even had a language engineer. So the language team was me and the person on iOS who was very passionate about it. He helped me set up with all the terminal and stuff, and we were just jamming on these ideas to really prove it out. No, this can be done. And so at that time it was just me prompt smithing, because really what's special there? You're just saying it to do things in 50 different ways and seeing what works across a broad set. I think the idea of eval was not a thing then; it was very much vibes because it was early days. So we were just like, okay, this feels good for these queries.

过了一段时间，我们开始有了一组 10 到 20 个提示词来测试这些。所以我有三四个 Claude 聊天记录，基本上都用完了上下文内存限制。实际上，每次我回去，Claude 都会说请开始新对话。那时你确实会得到一些提示词，但当时我不知道的是，这与 playground 模型的行为非常不同，因为 Anthropic 或 OpenAI 在公开版 ChatGPT 中做了很多清理。然后你把它拿到 playground 里。你决定：是用 OpenAI 的模型还是 Anthropic 的模型？然后你在 playground 里工作，因为它又略有不同——不是天壤之别，只是几处不同。所以你就用它。现在，我们显然有了适当的评估和整个团队；这是一个更规范的过程。但那时，是的，系统提示词最终成了有效的东西。然后，一旦我们有更多人帮忙，就去运行评估，迭代这个。现在，老实说，我甚至认不出那段代码。我知道它的大致形状，但当我看到时，我会想，这是干什么的？我都不认识了。

After a while we started having a set of 10 or 20 prompts that we wanted to test these with. So that was like I have three or four Claude chats where I've basically run out of context memory limits. Literally every time I now go back, Claude just says please start a new thing. At that point you do get to some prompt, but back then, something I didn't know was that is very different from how the playground models behave, because there's a lot of sanitization Anthropic or OpenAI is doing in the public version of ChatGPT. So then you take that in the playground. You decide: are you using OpenAI's model or Anthropic's models? And then you work in the playground because it's slightly different again—not day and night, but just a few things are different. So you use that. At this point, obviously we have proper evals and a whole team; it's a much more proper process. But back then, yeah, the system prompt ended up becoming the thing that works. Then someone, once we had more folks who could help us, just go and run eval, iterate on this. At this point, honestly, I can't even recognize the code. I know the rough shape of it, but when I look, I'm like, what is this doing here? I don't even recognize it anymore.

你谈到了与研究合作，以及语言系统提示词元素。既然你在这种 AI 原生产品中工作，你还看到设计价值主张有哪些其他转变方式？

You talked a little about collaborating with research, and about the language system prompt element. Are there other ways you've seen the value proposition of design shift now that you're working in this type of AI-native product?

对我来说是双重的，至少我看到的那些，尤其是如果你在一家拥有应用研究或基础模型研究团队的公司工作。一是老生常谈：有了一个能力，我们该用它做什么？这是典型的产品设计流程。同样：嘿，我们现在可以做到这个，弄清楚它如何在产品中运作。我认为对于 AI，我们谈了一些应用方面的事情。有很多应用方面的事情你实际上可以做——如果你使用 ComfyUI，你可以在视觉模型方面做。所以我们团队现在有另一个人正在研究如何插值嵌入以获得更好的结果。而他们是一名设计师，不是研究员甚至应用研究员。所以你可以对应用方面做出贡献。然后我认为新的东西，尤其是当你与基础模型团队合作时，是你常常可以塑造和指导研究应该研究什么。他们并不是凭空想出这些想法，对吧？一旦产品推出，它就是为产品服务的。总有一些研究非常看重的方向，因为它们使训练过程更高效，或者你可以用正确的数据获得更多，而不是总是数据饥渴。但还有很多东西，比如人们应该如何与相机控制交互。设计师可能能够提出相当好的原型。不要过于纠结于一个，因为我们不知道是否有直接的研究方法。但如果你有两三个这样的想法，去找你的研究员谈谈。真的展示给他们看。很可能你展示的三四个中至少有一个会非常兴奋，因为那是他们梦寐以求的东西。他们只是没能想出正确的做法或最优雅的形式。但一旦他们看到，他们也能将其与他们知道的 50 种设计模式联系起来。同样：哦，实际上那很容易。我有另一种研究技术，然后有这篇新的研究论文。让我试试看。这就是你如何触及棘手部分：在研究中，你不会得到你要求的确切结果。这就是事物的本质。它可能只有 60% 或 70%。所以还有另一个设计阶段。所以把那个设计看作是你与研究团队合作的一个想法的粗略草图。一旦能力实现，那时你可能就在做那个想法的生产版本了。

For me it's twofold, at least the ones I've seen happen, especially if you're working at a company that has either applied research or a foundation model research team. One is the good old: there's a capability, what should we do with it? That's your typical product design process. It's the same way: hey, we can now do this, figure out how it should work in the product. I think with AI, we talked about some of the applied stuff. There's a lot of applied stuff you can actually do—if you use ComfyUI, you can do it on the visual model side. So someone else on our team right now is playing with how to interpolate the embeddings to get better results. And they're a designer, not a researcher or even applied researcher. So there's that layer of contribution to the applied side you can do. And then the thing that I think is new, especially if you're working with foundation model folks, is often you can shape and guide what research should even research. They're not coming with these ideas in a vacuum, right? Once a product is out, it is in the service of product. There are always some arcs that are very dear to research because they make the process of training much more efficient, or you can get a lot more with the right data instead of always being data hungry. But then there's a lot of stuff like how should people work with camera control. Designers probably are able to come up with pretty good prototypes of what that could be. Don't get too hung up on one because we don't know if there's a direct research methodology for it. But if you have two or three of these, go talk to your researchers. Really show them. Chances are at least one of the three or four you show it to will get very excited because that is something they have dreamt about. They just couldn't quite figure out the right way of doing it or the most elegant form of it. But once they see it, they're also able to connect it with the 50 design patterns they know that will apply there. It's the same: oh, actually that's very easy. I have this other research technique and then there's this new research paper. Let me try something. That's kind of how you get to the tricky part: in research, you wouldn't get exactly what you asked for. That's just the nature of things. It would be like 60 or 70 percent. And so there is another phase of design that happens. So think of that design as a rough sketch of an idea that you collaborate with the research team on. Once the capability happens, that's when you're probably doing the production version of that idea.

但如果你带着那种“第一次讨论就该定下来”的心态去，那设计确实能扮演一个新角色，只要你有本事。在应用端，比如模型行为，这些都很重要，但你得学些新技能。设计师或者有敏锐审美眼光的人能带来很多价值。你对创意行业的工作流程有深刻理解，这对研究人员很有帮助，因为它真正影响研究的方向。所以大的方向是：帮助研究团队决定该把精力投到哪里，基于应用端需要什么；帮忙做系统工程；再加上一点应用，把已有的东西用新的有趣方式连接起来。在语言方面，概念药丸头脑风暴是简单的例子，但你可以根据技能水平做得更复杂。还有就是把能力推向市场，快速反应，思考问题和系统，构建模块化系统以便快速插拔，因为速度是关键。这些和以前不太一样。以前我记得有些项目我会花几周时间思考正确的设计并迭代。现在仍然有项目可以那样做，但大部分时间，这个过程是分散的：在研究开始前做一点，研究进行中做一点，但一旦事情确定下来，就非常快节奏了。

But if you go in with that mentality that what you're first talking about is not like a one-and-done kind of approach, then yeah, I think that's a new kind of role that design gets to play if you have the chops. On the applied side, in model behavior, all of those are very important, but you probably have to pick up some new skills. There's so much role for designers or just people with a very keen aesthetic eye that you can bring to the table. You have deep knowledge about workflows in creative industries, and that's very helpful for research folks because it truly impacts the direction of research. So those are the big buckets: helping see where research teams should channel their energy based on what would be nice to have on the applied side, helping with systems engineering, plus a dash of applied to connect existing things in slightly new interesting ways. On the language side, concept pill brainstorms are simple examples, but you can be as complex as your skill sets allow. And then there's the good old capability to get it to market, being fast on your feet, thinking through problems and systems, and having a modular system for quick plug-and-play because velocity ends up being key. Those are new and slightly different than before. Previously, I remember projects where I would spend weeks just thinking about the right design and iterating. There are still projects where you get to do that, but for the most part, it's spread out: you get to do that a little before research happens, a little as research is happening, but once things converge, it's very much go-go-go.

我相信现在有人在听，心想：天哪，我听了这么多期，感觉每个来的设计师都在给我加更多要学的技能，我的责任清单越来越长。所以反过来，根据你在 Luma 的经验，你觉得在 AI 世界里，传统上我们和设计关联的技能有没有变得不那么有价值？

I'm sure there's somebody listening right now that is like, man, I've listened to a bunch of episodes and it feels like every designer that comes on is just tacking on more skills that I have to learn and my bucket of responsibilities just getting bigger and bigger. So maybe the flip side of that based off of your experience at Luma, do you think that any skills that we would traditionally associate with design become less valuable in an AI world?

我从来没这么想过，但这是个很好的问题，因为你说得对。每个人都在加码。好像设计师还得懂销售和商业。好吧，那我干脆自己开公司算了。我觉得有些技能可能没那么重要了，但不会降到零。这是个现实问题。如果我现在是刚毕业的设计师或正在学设计，真的很难决定该投资哪些技能。以前的设计师被框在了一个观念里——部分是我们自己，部分也是行业——就是能做出好的视觉作品。即使考虑体验，最终也还是好的视觉作品。这就是科技设计师的工作：正确的字体、正确的 UI、完美清晰的东西、正确的间距、节奏、和谐等等。但这些都是视觉属性。这些不会消失。我不太想说它们变得不那么重要，因为我不想让人以为“太好了，视觉设计不重要了”。不，我跟你讲，这仍然是基础，是让你进门的东西。因为如果连基本的品味和执行能力——做出好的、美的、高效的、有效的、令人向往的东西——都没有，那就不行。这些是基础技能。现在把这些看作放大器。它们不定义你，但如果你有这些技能，你会走得更远。把它看作投资。未来 5-10 年的事情还很模糊，但我们可以看出一些趋势。把它看作未来投资。如果你今天没有这个技能，就做不成了？不是的。你的基础技能仍然是经过时间考验的，能让你进门。你会找到理想的工作，和合适的人一起，但现在你有机会做更多。做不做由你。但问题是，别人会愿意做。所以真正的情况是：一旦你到了那个位置，从宏观上看，我们每个人都在设计自己的生活，在这里，我们每个人都在设计自己想要的职业。就像人们说商业技能对设计师很重要一样。一个很棒的设计师有商业技能，和一个没有的，你会选哪个？如果其他条件相同，你会选有商业技能的。但如果你要找的是一个很棒的设计师——典型的优秀视觉 UI、正确的 UX——和一个设计一般但商业很厉害的人，那取决于公司。在非常早期的阶段，公司可能更喜欢后者，但一旦他们有两三个设计师，他们会很快意识到产品对了但用起来不舒服，然后就需要前者。所以最终，团队技能必须平衡。这不是“你不做这个就完了”。把它看作投资，在所有你能投资的地方，哪些能给你最大回报？这确实像一种新材料。就像触摸屏出现时，能快速原型的人走得更远。这是事实。回头看，这不意味着不能原型的人就什么都做不了。不，他们仍然有价值。原型能力仍然很难获得，但如果你还能做其他事，那就更有价值。然后你也看到有人把原型做到极致，如果设计也做得好，就成了设计工程师。很神奇，我们能有这么多选择，可以获取不同的技能组合，而真正基础的东西很少。

I don't think I've ever thought about it that way, but that is a very good question because you are right. Everyone just adds more. It's like they should be aware of sales and business. It's like, okay, I will just be a one-man company at this point. I think some of those skills might be a little less important. Maybe I don't think they go to zero though. I think this is a real issue. If I were a new grad designer or studying design right now, it is actually hard to understand how I would think of what skills to invest in. Designers before this were so boxed into the idea—and some of it is us, but some of it is also the way the industry boxed us in—to the idea of being able to create something that is a good visual artifact. Even if experience is considered, at the end of the day it's a good visual artifact. That was the job of a designer in tech especially: the right typography, the right UI, the perfect crisp looking things, the right spacing, rhythm, harmony, all of that. But very much visual attributes. Those don't go away. I'm hesitant to say they become less important because what I don't want people to take away is like, oh great, visual design doesn't matter. No, I kid you not, that is still baseline, the thing that will get you in the door. Because if that basic taste and execution capability to make something good, beautiful, efficient, effective, and desirable—going back to that word—that's a foundational skill. Think of these at this point as amplifiers. They don't define you, but if you have these skill sets, you're going to go much further. Think of it as investing. We're starting to see the arc of things in fuzzy, not fully definite ways for the next 5-10 years. Think of it as future investment. And if you don't have the skill today, you just can't make it? No. Your baseline skills are still the time-tested things that will get you in the door. You will land the perfect job you've been looking for with the right set of people, but you now have the opportunity to do more. Whether you do it or not is up to you. The caveat being someone else will be willing to do it. So that's the real thing: once you're in that situation, at the grand level we each design our lives, and in this case we each design the exact career we want to have. It's the same way when people would say business skills are super important for designers to understand. Would you have a really great designer who has that skill set versus one who doesn't? If all things equal, you will prefer the one. But if you're looking for someone who is a great designer—typical great visual UI, right UX—versus someone who is okay at that but also really good at business, it depends on the company. In super early stages, a company might prefer the latter, but the moment they have two or three designers, they will quickly realize they got the right product but it doesn't feel good to use, and they'll need the former. So at the end of the day, the team skills have to be balanced. This is not a "if you don't do this, you're done." Think of it as an investment, and of all the places you could invest, what are the ones that would give you the most return? This does feel like it is a new material. The same way when touch came, people who could prototype went further. That's just a fact. Looking back, we see that doesn't mean people who didn't prototype couldn't do anything. No, they were still valuable. It's still a very hard skill set to acquire, but if you could do the other stuff, it was more valuable. And then you also see people who pushed so much on the prototyping side that if you did that well with enough design, you became a design engineer. It's magical that we get to have so many choices in the different strands of skills we can acquire, and there are very few things that are like baseline you should know this.

那么，你对什么感到兴奋？因为如果你不兴奋，你就不会享受其中。所以，这不是必要条件，但却是加分项。而且，就像原型设计被证明是职业发展的良好放大器一样，这感觉也会是一个坚实的放大器，让你能为公司做出超出当前项目的贡献。你们现在正在招聘。所以，如果有人正在收听，受到启发，想要加入，你目前最看重的放大器是什么？

From there it's like what are you excited about this? Because if you're not excited you're not going to enjoy it. So, it's not a requirement, but it is a nice to have. And it does feel like something just the way prototyping proved to be a pretty good amplifier for careers. It does feel like it would be a pretty solid amplifier and like something that you can contribute to the company beyond just the immediate project that you're on. You're hiring right now. So, if someone's listening, they're inspired, they want to join, what are the amplifiers that you care the most about right now?

我认为成长心态是最重要的前提。如果没有成长心态，你甚至不会问自己“我现在还能学什么”这样的问题。所以，成长心态和好奇心是最根本的东西。我们是一个小团队，所以我会说，擅长视觉、擅长产品设计这些基本技能很重要。我们还没有到可以大量指导别人的阶段，所以我们确实在寻找处于职业生涯中期、能够自给自足的人。但如果你渴望更多，如果 AI 让你兴奋，如果你好奇这里还能发生什么，我们如何超越提示框和获取输出，去探索另外 100 种可能，而且这些想法就在你脑海中沸腾，那么这里就是你的归属。

I think growth mindset is like one of those before everything. If you don't have the growth mindset, you wouldn't even be asking this question of like what else can I learn right now. So fundamentally that and curiosity are like the most foundational things. We're a small team, so I will say like you know being good at visuals, being good at product design, like the good old basic stuff is important. We're not yet at the point you know where we can like really mentor people a lot, so we definitely are looking for people who are in the middle stages of their career and you know are self-sufficient. But if you're hungry for more, if AI is exciting to you, if you're curious of like what else can happen here, how do we move beyond just like you know prompt boxes and getting outputs to like what are 100 other things, and these ideas are just bubbling in your head, yeah this is the place to come and be.

我想稍微拉远一点，或许再多给一些背景，谈谈你如何看待 Luma 的组织架构。因为如果我现在让一个随机听众说出世界上最具设计驱动力的公司，我保证苹果和 Airbnb 都会在名单上，而这正是你过去的经历。那么，你从那些公司汲取了什么？作为一名设计师在那里是什么感觉？这些又如何塑造了你为 Luma 设计组织的方式？

I want to zoom out for a second and maybe even just give a little bit more context about how you're thinking about the Luma org. Because if I were to ask a random listener right now, name the top design-driven companies in the world, I guarantee both Apple and Airbnb would be on that list, which is your past experience. So, what are you drawing from at those companies and what it was like being a designer there? That's then shaping how you design the org for Luma.

老实说，我认为这是当你身处苹果或 Airbnb 时才能拥有的奢侈，因为它们是成熟的产品。它们有清晰的产品市场契合度，是成熟的产品，所以作为产品团队、设计团队，你确实可以深入钻研一个问题。在这两家公司，项目通常要花好几个月。我在这些公司做任何项目的平均时间都是 8 到 12 个月。我几乎所有的项目都那么长。而且，别误会，它们都是非常深入的事情，比如整个窗口管理功能，或者 Airbnb 整个设计系统的重新设计。但它们确实跨越了那样的时间跨度。所以我从中汲取的是对变得更好或比昨天更好的不懈追求。这些都是很好的北极星。但有时你也必须意识到，即使 Airbnb 也不像苹果，苹果显然也不像 Airbnb。它们是完全不同的业务，存在于完全不同的领域。所以它们有非常不同的业务动态和市场动态，影响着事情的运作方式。同样的情况很可能也适用于 Luma，也适用于任何正在收听并在其他公司担任类似角色的人。所以不要把这些视为原型，认为设计就应该这样。从中提取更通用的部分。

I would honestly call it a luxury that you get to have when you are Apple or Airbnb because they are established products. They have a clear product-market fit. They are mature products and so you actually can afford as a product team, as a design team, to really go deep on one problem. And in cases of both the companies, honestly, spend months. I think my average time for any project at either of these companies was 8 to 12 months. Almost everything I worked at these companies was that long. And they were like, don't get me wrong, they were very deep things. They were like, you know, a whole window management feature or a whole redesign of the entire design system of Airbnb. But like they did span that kind of time horizon. So the thing that I draw from is the relentless pursuit of trying to be good or better than where you were yesterday. These are like good north stars to have. But at some point you also have to realize even Airbnb is not like Apple and Apple obviously is not like Airbnb. They are completely different businesses. They exist in completely different sectors. And so they have very different business dynamics and market dynamics that affect how things work and how things don't work. And so the same is likely going to be true of Luma and the same is probably true for whoever is listening and is in a similar role at a different company. So try to not think of these as like this is the archetype. This is exactly how design should be. Take the parts that are more universal from there.

所以，我觉得乔尼·艾维留下的那副给每个新员工的卡片中，有一张让我印象深刻：信任专家。在苹果尤其如此，因为你可能会遇到一个在皮克斯工作过、一生都在做角色设计的人。所以你在给他们关于拟我表情之类的反馈之前，至少应该思考一秒钟。你知道，就像这样：老兄，这人一辈子都在做这个，他们很可能是世界上这方面的顶尖专家，不只是在国家层面，而是全世界。这两家公司的另一个共同点是极其尊重个人贡献者。我尽可能地将这一点带到 Luma，不仅在设计团队，也在更大的产品和工程方面。我们是一个小团队，尤其是在 AI 领域，小团队实际上可能比层级繁多的大团队拥有更多的阿尔法。你越是赋能工程师、个人贡献者或设计师，进展就越快，想法就越多。历史上，你不会把 Facebook 和 Airbnb、苹果相提并论，但这实际上是我从 Facebook 学到的核心东西之一。任何人都可以提出任何想法。在一夜黑客马拉松中创造出来的产品功能，现在被数十亿人使用，数量惊人。比如视频通话就是某人在一夜黑客马拉松中创造的。我当时想，“哦，如果能直接视频通话某人会很酷”，然后它就成了一个产品功能。这太疯狂了。所以当你给人们那种自主权时，美好的事情就会发生。不同的公司运作方式不同。在 Facebook，是“去试试，我们看看”。在苹果，有点像是“去试试，但然后我们会真正成为整个体验的编辑，然后把它推向世界”。两者都可行。基本原理是一样的：你重视做工作的人，给他们做正确工作的自主权。最后，这些公司拥有而我们正在努力增加但尚未完全达到的，就是时间和空间。好事多磨。这听起来像是废话，但好事确实需要时间。你需要学会在时间限制下能深入多少，而不是试图编织一个完美的想法网络。因为如果没有足够的时间，一切都会崩溃。好的一面是，我们是一个小团队，你拥有全部自主权。所以一旦这个版本发布，没有人会阻止你提出更好的版本。这就是我认为与苹果和 Airbnb 不同的地方。但最根本的是，这两家公司共享的是设计作为一种非常乐观的世界观。把设计师看作乐观主义者其实很有趣，因为我们是世界上最有评判力的人之一。我们喜欢批评一切。你去一家餐厅，会想“为什么这样？”但在所有尖刻背后，我认为是一种不懈的希望：它可以变得更好。

So I think like one of the things that Jony Ive left these like deck of cards that every new hire would get. One of them that has stuck with me is trust the expert. Especially true in case of Apple because you would have probably someone who has worked at Pixar and done characters all their life. So you should at least have one second of thought before you give them feedback on Memojis or something. You know, it's like one of those, it's like dude this is all this person has done. They are probably the leading expert on this in the world. Like not even in the country, like actually in the world. The other thing both these places had is huge huge respect for individual contributors. That is something that as much as I can I do try to bring over to Luma both on the design team but also on the larger product and engineering side of things. Which is we're a small team and especially with AI it's possible that small teams actually have way more of an alpha than much larger teams with layers of management. The more you can empower engineers and individual contributors or designers, the faster you can go, the more ideas come to the table. Like historically, you wouldn't think of Facebook in the same breath as Airbnb and Apple, but this one actually is one of the core things I picked up from Facebook. Anyone could just come up with any idea. Like the number of product features that were created in an overnight hackathon that are now used by billions is shocking. Like literally video chat was a thing someone created in an overnight hackathon. I was like, "Oh, it would be cool if you could just video call someone" and then it became a product feature. It's insane. And so when you give people that kind of agency, beautiful things happen. Different companies work differently. At Facebook it was like go try, we'll see. At Apple, it's a little bit of like go try, but then we will, you know, really be the editor of the entire experience and then bring that out into the world. Both can work. The fundamentals are still the same. You know, you value the people who do the work. You give them the agency to do the right kind of work. And lastly, the thing that these companies have that I think we're trying to increase more of but are not fully there is just the time and space. Good things take time. This is like, you know, one of those duh kind of statements, but good things do take time. You need to then learn how much you can go deep on something given the time constraints you have versus trying to craft this masterful web of ideas. But because you don't have the right set of time, it all comes crashing. And the good part is it's a small team and you have all the agency. So once this version is out, nobody will stop you from coming up with the better version of this idea. And so that's I would say something that is actually different from the Apples and the Airbnbs of the world. But really in its most fundamental form, the thing that both companies share is design as a very almost optimistic view of the world. It's really funny to think of designers as optimistic because we are one of the most judgy people in the world. We like critique everything. You go to a restaurant, you're like why is this like this? But behind all the snark and whatever, I think is this relentless hope of like it can be better.

对我来说，这在苹果和爱彼迎这样的地方极其真实。我的希望是，在这一切中，我们能够将那种乐观的设计和产品理念带到 Luma 以及这个 AI 领域——目前这里似乎所有人都认为速度是唯一重要的东西。但我的希望是，如果我们做对了，速度当然有其价值，但还有很多其他东西能让产品有粘性，让人们真正爱上我们的产品，而不仅仅是爱上我们的模型。

And to me that is extremely true at places like Apple and Airbnb. And my hope is in all of this we're able to bring a little bit of that idea of the optimistic aspect of design and product at large at Luma and in this AI field where right now everyone seems to think velocity is the only thing. But my hope is if we do this right, it's sure velocity has its value, but it's a lot of other things that make the product sticky, make people actually love our product, not just love our models.

你几次提到成长型思维，所以也许我可以把问题抛回给你。回顾你在这些公司的经历，有没有一个特定的转折点或某个领域，让你觉得自己作为设计师真正成长了？

You've talked about a growth mindset a few times now, so maybe I'll even flip the question back at you. When you reflect on your time at these companies, was there a specific inflection point or a particular area where you feel like you really grew as a designer from those experiences?

加入爱彼迎后，有些事情变了。我不知道是什么，真的不知道。在 Pulse 的时候，我太新了。我对整个国家都很陌生。我记得和联合创始人一起去租房子，因为我没有任何信用评分。我不知道那是什么东西。所以我的一个联合创始人不得不陪我去，为我做担保。房东叫 Bryant，我记得他的名字，很酷的家伙，一个老头，他说：“哦，是的，我看得出来你刚落地。你完全摸不着头脑。”所以在 Pulse 的时候，我试图做事，但老实说，Pulse 更多的是理解设计意味着什么。Facebook 是一个阶梯式的变化，让我看到了设计中的“砂砾”是什么样子。所有人都在那里：Tim Vanam、Wilson、Mike Mattis、Supa。老实说，名字多到能说上几个小时，但所有人都在那里。我最初在 Twitter 上关注的那 10 个人，全都在那里。这让我既看到了他们真正擅长的东西，也看到了归根结底他们也是人。一旦你开始和他们一起参加评审，你就能看到他们哪些方面是世界一流的，哪些是他们的弱点。而且每个人的情况都不同，包括我。所以这让一切变得正常多了，几乎让我有了锚点和根基，意识到这不是一个你成为明星设计师的幻想。不，每个人都有长处和短处。一个人不可能同时做最好的品牌设计、最好的交互设计和最好的产品设计。这根本不可能。所以从那里开始，知道什么让你兴奋，你想深入哪个领域？那很大程度上是 Facebook 的经历。对我来说，Claraara，我加入的下一个初创公司，是一次成长的经历。我又独立了。我觉得自己积累了足够的技能，能够以更广阔的视角看待事物，并且能够做从品牌设计到产品的所有事情，而且在几乎没人知道 AI 是什么的时候进入了 AI 领域。这对我来说非常有趣。一点一点地，我能够独立完成这些，这让我对自己的技能组合变得非常自信。现在收到来自同样那 10 个人的评论和私信，其中一些人称赞我的工作，赞扬，对我们已经完成或正在尝试的事情说好话。那是一个巨大的信心提升，但也让你意识到，当你做你深信不疑的事情时，它们就是好东西。

Something changed when I joined Airbnb. I don't know what, I really don't know what. Like at Pulse, I was too new. I was new to the whole country. I remember going to rent a house with my co-founder because I didn't have any credit score. I didn't know that was a thing. So one of my co-founders had to come with me and really vouch for me. And Bryant, I remember the landlord's name, really cool guy, an old dude, and he's like, "Oh yeah, I can tell you just landed. You're so lost." So at Pulse, I was trying to do the thing, but honestly Pulse was very much about understanding what design meant. Facebook was a step change in seeing what the grits look like in design. Everyone was there: Tim Vanam, Wilson, Mike Mattis, Supa. Honestly, there are more than enough people to name for hours, but everyone was there. All those 10 people I started following on Twitter, they were all there. That exposed me to both the things they were really good at, but also that at the end of the day, they were people. Once you started sitting in crits with them, you could see the things they were excellent and world-class at, and the things that were their weaknesses. And they were different for everyone, including me. So that made it much more normal to be like, okay, it almost anchored and grounded me a little bit to realize it's not this imaginary thing where you become a star designer. It's like no, everyone has strengths and weaknesses. One person can't be doing the best branding and the best interaction design and the best product design. It's just not possible. So from there, knowing what excites you, what do you want to go deeper into? That was a lot of Facebook. For me, Claraara, the next startup I joined, was a coming-of-age experience. I was back on my own. I felt like I had picked enough skills where I could see things in a broader perspective, and I was able to do everything from brand design to product, and be in AI at a point where nobody even knew what AI was. That was very interesting for me. Bit by bit, the fact that I could do that on my own made me become very confident in my own set of skills. Now getting comments and DMs from the same 10 people, some of them complimenting the work, praising, talking nicely about the things we had done or were trying to do. That was a huge confidence boost, but also made you realize that when you do things you deeply believe in, they are good things.

然后老实说，当我开始在爱彼迎工作时，我觉得自己准备过度了。我不知道脑子里发生了什么变化，但我记得即使在面试环节中，我也没有一丝怀疑自己会得不到这份工作。我根本没有想过“哦，也许我得不到”。我就是觉得，你可以称之为过度自信，可能有一点。我心想“哦，这个来自 Facebook 的炙手可热的设计师”，可能有一部分是那样。但我从未觉得自己不能胜任这份工作。对我来说，在爱彼迎重要的是确保我能真正对公司产生影响。所以第一年的项目有点实验性。第二年，我深入研究了既在工作上困难又在组织上具有挑战性的事情。然后第三年，我认为在我的职业生涯中第一次有了一位真正看到我的经理，他给了我 wink。我觉得自己真的势不可挡。我突然开始领导与不同组织的总经理的会议，致力于某个设计系统，确保他们的整个产品线能够采用这个设计系统。同时还要重新设计设计系统，进行品牌评审等等。所以我认为那可能就是转折点：有一个真正相信你的人，有一个能为你提供保护伞的人，尤其是在有很多其他议程的大公司里。那是改变人生的。

Then honestly when I started at Airbnb, I felt overprepared. I don't know what changed in my mind, but I remember even through the interview loops I had zero shred of doubt I wouldn't get the job. It just was not a thought of "oh maybe I won't get it." I was like no, you can call it overconfidence, possibly a little bit of that. I was like "oh this hot shot designer from Facebook," probably some of it. But I never felt that I couldn't do the job. For me, what was important at Airbnb was to make sure I was able to actually have impact on the company. So the first year's project was a little bit very experimental. The second year I just went deep into something that was both work-wise hard but also organizationally challenging. And then in my third year, I think for the first time in my career I had a manager who I would say truly saw me, who just gave me winks. I think I was just like I am truly unstoppable. I was suddenly leading organizational meetings with the GMs of a different org to work on some design system and make sure their entire product line could adopt the design system. Meanwhile also redesigning the design system, having brand reviews and all of this stuff. So I think maybe that's the inflection point: having someone who actually believes in you, having someone who is able to provide the air cover for you, especially in larger companies where there's a lot of other agendas going around. That was life-changing.

你做的这些事情，你有时从未意识到自己正在变得更好。我从不认为自己擅长视觉设计，但在爱彼迎职业生涯的后期，我却在指导别人做视觉设计。我仍然不认为自己是一个好的视觉设计师，但我想这些事情总是相对的。

Some of these things you work on, you kind of never realize you're getting better at it. I never thought of myself as good at visual design, but towards the end of my career at Airbnb, I was mentoring people on visual design. I still don't think of myself as a good visual designer, but I guess these things are always relative.

我仔细审视过你的用户界面，我要说你是一个好的视觉设计师。

I've taken a fine comb through your UI, I'm going to say you're a good visual designer.

我不知道。我觉得我看别人的作品时会想，他们怎么能看到这个？这看起来太漂亮了。我不认为自己达到了那个水平。我能让东西看起来漂亮，但没到那个程度。但没关系。我擅长的事情，我知道自己擅长。所以没关系。

I don't know that. I feel like I look at some other people's work and I'm like, how could they even see this? This looks so pretty. I don't think I'm there. I can make things look pretty, but not that level of pretty. But that's okay. The things I'm good at, I know I'm good at. So it's okay.

这也是我要强调的。我对你所说的一些事情进行串联，我的理解是，认识到你不可能在每个技能上都出色这一点非常重要。即使我们在 Twitter 上崇拜的人，他们可能也不是每个技能都出色。每个人实际上都是人。我很喜欢你的转折点是关于人和内在信念，而不是……是的。

And I think that's what I'm going to underline, too. My interpretation of just connecting the through lines of some of the things that you've been saying is how important it is to recognize the fact that you can't be great at every skill. And even the people that we look up to on Twitter, they're probably not great at every skill. Everyone actually is human. So much of I love that your inflection point was about people and this inner belief rather than... Yeah.

然后我学到了 XYZ。我个人职业生涯中抵制最久的事情就是做经理。我就是不想当经理。我最热衷的事情是确保个人贡献者拥有我职业生涯中很多时候没有得到的空间。所以我认为，因为我做了很长时间的个人贡献者，我对那部分工作和职业道路有一种非常不同的同理心。

And then I learned XYZ. The thing that I resisted for the longest time in my career personally was being a manager. I just didn't want to be one. The thing that I'm most passionate about is being able to make sure the individual contributors have the space that many times I didn't get to have in my career. And so I think because I was an IC for so long, I have a very different kind of empathy for that body of work and that career track.

我会不惜一切代价保护他们的创作精神、创作能量，并将其引导出来。你仍然需要给予反馈等等，但创作能量需要在某些方面得到保护和维护。人们认为设计团队非常排外，但你应该邀请每个人参与对话。然而，想法确实很脆弱，你不能随便在上面戳洞。你还必须对想法有足够深入的思考，才能在一个相近的频率波长上进行讨论。如果你只是肤浅地思考想法，那么期望每个人都把你带入讨论是不公平的，因为你没有足够深入地思考。无论你是设计师还是工程师，这都不重要。重要的是你是否足够深入地思考过，因为当你真正剖析一个想法时，这是尊重的表现。你已经做了功课，弄清楚了，思考过了，并且带来了一个原创的观点。那么谁在乎你是设计师、工程师还是其他人？我不在乎。你只需要思考过。这是保护创作能量的另一种方式，因为不保护它很容易。我见过设计师精疲力竭，有些人从设计转行，再也不做设计师了——他们成了工程师。它需要被保护。苹果在这方面做得很好，他们把设计师都留在象牙塔里。Airbnb 也很擅长保护设计团队的创作能量。我相信其他世界级的设计团队也把这一点作为让人们蓬勃发展的核心：保护那个能给你所有好想法的东西。这就是你如何想出好东西。你必须像保护任何其他资源一样保护它。

I'll do whatever it takes to preserve their creative spirit, their creative energy and channel it. You still need to give feedback and all that, but creative energy needs to be preserved and protected in some ways. People think of design teams as very exclusive, but you should invite everybody for a conversation. However, ideas are truly fragile and you can't just go around poking holes in them. You also have to have thought about the idea deeply enough to be able to have a discussion at a similar frequency wavelength. If you're just shallowly thinking of ideas, it's unfair to expect everyone to bring you into the discussion because you haven't thought deeply enough. Whether you're a designer or an engineer, that doesn't matter. What matters is whether you've thought deeply enough, because it's a sign of respect when you're truly dissecting an idea. You have done your homework, figured it out, thought about it, and have an original point of view to bring to the conversation. Then who cares if you're a designer, engineer, or someone else? I don't care. You just need to have thought about it. That's another way of protecting the creative energy, because it's very easy to not have that protected. I've seen designers get burnt out, some switch from design to never being a designer again—they become engineers. It needs to be protected. Apple is really good at that; they keep all designers in an ivory castle. Airbnb is also pretty good at preserving the creative energy of the design team. I'm sure other world-class design teams also make it a core aspect of how they allow people to flourish: by preserving the thing that gives you all your good ideas. That's how you come up with good stuff. You have to preserve it like any other resource.

我们覆盖了很多内容。在你离开之前，我想再做一个通用的总结，确保我们涵盖了你所有想谈的话题。也许我们可以对听众说几句，他们被你所说的一切所启发，一直在关注一些 AI 的东西，但他们不在一个前瞻性的公司，这些东西还没有真正渗透到他们的日常工作中。你有什么建议或经验想分享给那个人吗？

We've covered a ton of ground. Before I let you go, I want to give one more generic catch-all to make sure we're hitting everything you want to talk about. Maybe we could address the person listening who is inspired by everything you're saying, has been keeping tabs on some AI stuff, but isn't at a forward-looking company and none of it is really seeping into their day-to-day yet. Any advice or learnings you want to share with that person?

如果你在推特圈子里，你会听到品味最重要，高能动性最重要。我认为能动性确实很重要。当你没有能动性时，你会变得悲观和尖刻。但当你能动性不足时，你也会开始缺乏好奇心。所以我会说，不要追求高能动性。那不是你可以直接追求的东西。你可以培养它，但也许从保持好奇心开始。如果你在一个 AI 不是前沿的地方，不是每个地方都需要是前沿。确实有一阵风潮，每个人都跳进 AI，因为那是该做的事。也许你很幸运不在那样的公司。你的公司可能更有意图。所以第一个要回答的问题是：你被它吸引吗？这里有什么让你感兴趣吗？因为如果不是，你只是在追逐另一个 meme 周期。就像那些进入加密货币领域的人，因为他们认为这会让他们致富，而不是因为他们被共享所有权的想法所吸引。对我来说，那是加密货币最令人兴奋的部分。对于 AI，首先，它让你好奇吗？有没有任何部分让你兴奋、有点害怕，但也感兴趣？其次，如果你看不到它，你可以在日常生活中找到小方法使用它。如果你是设计师，使用像 Cursor、Lovable、v0 这样的工具来快速制作小型原型非常容易。即使你在苹果工作，NDA 满天飞，你也能找到方法做到这一点。你已经领先很多人了，因为你不仅仅在发推文，你实际上在体验它，理解什么有效、什么有限制。这对于文本到语言的原型设计、编码、文本到图像、文本到视频都是如此。如果你是一个有创造力的人，看看文本到图像或文本到视频能否为你的创作实践带来新方法。这可能是你一直想要的东西的开始吗？我一直想做的一件事是时装系列，所以对我来说，这些工具是在通勤或观看东西时快速触及那个实践种子的方式。这是另一种做法。最后，你总是可以换工作—— wink wink，我们在招聘。但在此之前，你可以做很多副项目。我最近看到一个项目，是一个人和一个 AI 一起做播客聊天。我觉得“这太酷了”。这个领域非常新，才两年左右。回到 2007 年，你最多在 2009 年。大多数有趣的公司那时甚至还没开始。所以还有很多事情可以做。追随你自然好奇的东西，看看有没有让你兴奋的部分。你不需要学习所有东西——智能体、多模态模型以及你推特信息流中的其他 50 件事。选择一两个真正让你兴奋的，深入下去，看看它会变成什么形式。

If you're in the Twitter sphere, you'll hear that taste is all that matters, high agency is all that matters. I think agency is really important. When you don't have agency, you become pessimistic and snarky. But when you don't have enough agency, you also start lacking curiosity. So I would say don't go after high agency. That's not something you go after. You can cultivate it, but maybe start by just being curious. If you're at a place where AI isn't at the forefront, not every place needs to be. There was definitely a trend of everyone jumping into AI because it was the thing to do. Maybe you're lucky you're not at a company like that. Your company is probably more intentional. So the first question to answer is: are you drawn by it? Is anything here interesting to you? Because if it's not, you're just chasing another meme cycle. It's like people who got into crypto because they thought it would make them rich, not because they were fundamentally allured by the idea of shared ownership. For me, that was the most exciting part about crypto. For AI, first, is it intriguing to you? Is there any part that excites you, mildly scares you, but also interests you? Second, if it's not something you can see, there are small ways you can use it in your day-to-day life. If you're a designer, using things like Cursor, Lovable, v0 for quick tiny prototypes is very easy. Even if you work at Apple with NDAs galore, you can probably find ways to do this. You're already ahead of many people because you're not just tweeting about it; you're actually experiencing it and understanding what works and what the limitations are. This can be true for text-to-language prototyping, coding, text-to-image, text-to-video. If you're a creative person, see if text-to-image or text-to-video can bring a new approach to your creative practice. Could this be the beginning of something you've always wanted? One thing I've always wanted to do is a fashion line, so for me, these tools are a quick way to tap into that seed of practice while commuting or watching something. That's another way to do this. Lastly, you can always change your job—wink wink, we are hiring. But before that, there are so many side projects you can do. One I saw recently was a podcast where a person and an AI just talk. I thought, "That's so cool." The field is super fresh, barely two years old. Going back to 2007, you're at best at 2009. Most interesting companies didn't even start at that point. So there's so much left to do. Follow things you're naturally curious about and see if there are parts that excite you. You don't have to learn about everything—agents, multimodels, and all the other 50 things in your Twitter feed. Pick the one or two that actually excite you and go deeper on them, and see what form it takes.

也许它真的会带来你的下一份工作，或者你可能会说服你所在的公司，因为通过构建这些，你向他们展示了这个领域能带来的价值。在我看来，这两件事都是高自主性的人会做的，而不是只说“哦，我公司不让我做”。问题是，谁在阻止你？这就是我的建议。

Maybe it'll actually lead to your next job, or maybe you will actually convince the company you're at, because by building these, you show them the value of what this domain can bring. Both of these in my head are things that high agency people would do, instead of just saying, 'Oh, my company doesn't let me do it.' It's like, so who's stopping you? That's kind of my advice.

我喜欢这个。太棒了。伙计，今天真的非常开心。感谢你来做客，分享你的想法和你正在做的事情。我一直很享受关注你的旅程和目前的产品。所以，感谢你的时间。

I love it. I love it. Well dude, this has been really really fun. Thanks for coming on today and sharing a little bit about what you think and what you're working on. I'm really enjoying just following along with your journey and the products so far. So, appreciate your time.

谢谢你，也谢谢你邀请我。

Thank you and thank you for having me.

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工具栈是什么。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，支持这个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.