与 AI 共设计：OpenAI 产品设计内幕

我想有些人可能会以为我们领先了两年，但实际上我们与所有进展的节奏非常贴近。这是一种截然不同的工作方式。脚下的事物每时每刻都在变化，非常令人兴奋。这种“边走边摸索”的感觉真的很有趣——我们会去尝试，会去转动曲柄，不断迭代，持续前进。

I think some people might assume we're like 2 years ahead thinking like that, but we're running very closely with where all of these advancements are going. And so that's just like a very different way of working. Things are changing underneath your feet all day long. And it's very exciting. It's really fun to be like, I don't know, we're going to figure this out as we go. We're going to try it. We're going to turn the crank. We're going to keep iterating. We're going to keep going.

欢迎来到 Dive Club。我是 Reid，在这里设计师永不停歇地学习。今天的嘉宾是 OpenAI 的产品设计主管 Ian Silber。我们将深入探讨他们的工作方式、以 AI 为材料进行设计的体验、Ian 对设计师角色演变的看法，以及更多内容。但在开始之前，我必须知道：一个人究竟是如何成为 OpenAI 的产品设计主管的？

Welcome to Dive Club. My name is Reid, and this is where designers never stop learning. Today's episode is with OpenAI's head of product design, Ian Silber. So, we're going to go deep into all of the ways that they work, what it's like designing with AI as a material, Ian's thoughts on how the role of designer is evolving, and a lot more. But before we get into all of that, I had to know how the heck does somebody become the head of product design at OpenAI?

我在 Instagram 工作了大约 8 年，那是一段非常棒的时光。几乎所有我们做过的重大项目，我都能以某种方式参与其中。后来 Kevin 和 Mikey 创办了一家新公司，我就加入了他们。那是 Artifact，一个类似新闻的 AI 产品。非常有趣，团队也很棒。但与此同时，我一些同样离开 Instagram 的好朋友对我说：“嘿，我们正在做一件事，差不多是在开发一款游戏。我们也不完全清楚自己在做什么，但很希望你能来一起干。”我当时处于职业生涯的一个节点，心想：我应该去试试。这听起来有点疯狂，我甚至算不上硬核玩家，完全没有游戏背景。从这个角度看，我其实不太够格，但团队里既有非常亲密的朋友——能遇到这样的机会总是很幸运——也有我职业生涯中合作过的最有才华的人，包括设计师和非常出色的工程师。我们从 Instagram 汲取了很多经验，它本身就带有一些游戏化的机制和行为。所以我们非常侧重社交方面。那款游戏包罗万象，简单来说，我们做的本质上就是浏览器里的《我的世界》。但我们试图将《我的世界》和《Roblox》结合起来：Roblox 允许任何人创建游戏，拥有完整的市场，人们不断创作游戏，然后你可以来玩，创作者还能从中赚钱，形成了一个完整的产业。但真正去制作游戏时，体验并不那么友好。而我们喜欢《我的世界》的易用性，你可以一对一地搭建东西。所以我们想将两者结合：如果让任何人都能进入这个世界进行建造，同时让其他人参与其中会怎样？我们在游戏中加入了多种机制，但我特别喜欢的一点是，任何人都可以创建自己的版本，然后让别人来玩。这非常有趣。做游戏很不一样，有趣的是，你很难判断什么才是好主意——你总觉得“哦，那个可能会好玩”，所以很容易把范围铺得很广。

I had been at Instagram for like 8 years. It was amazing. I mean, most everything that the big stuff that we'd worked on, I was able to be involved in in some way. Kevin and Mikey were starting a new company. I went and joined them. It was Artifact. It was like this sort of news kind of AI thing. Ton of fun, great team, but I had some really good friends that had also left Instagram that were like, 'Hey, we're starting this thing. We're kind of building a game. We don't totally know what we're doing, but we'd love for you to come work with us on it, too.' I don't know. I was at the point in my career where I was like, I should just try it. Some crazy stuff. I'm not even like a hardcore gamer or I have zero background in gaming. In that way, I was kind of unqualified, but I mean, the team was like that they were assembling was both really close friends, which is always like fortunate if you ever get that opportunity. But also some of the most talented people that I've had a chance to work with in my career. Both from design, but also like really great engineers. I think we took a lot of lessons from Instagram, which is like a little bit of a has some game type mechanics or behavior. So, we leaned a lot into like the social side and and that sort of thing. And it was a game, but it was it was a little bit of everything. The TLDR of what we built was Minecraft in the browser, essentially. But we were kind of trying to combine Minecraft meets Roblox, where Roblox is like anybody can create these games, and there's this like full marketplace. People are constantly creating games, and then you know, you can come and play them, and and people can actually like make money off making these games, and it's this whole sort of industry. But when you really go to make the games, it's like pretty comp like it's not the most approachable experience. But we loved how approachable Minecraft was, where you're just like one-to-one building things. And so, we wanted to kind of combine that together and say, 'What if you could let anybody come into this world and build, but also let other people participate in that?' And we built a lot of different mechanics into the game, but one of the things that I really liked was the idea that anybody could kind of create their own sort of version of it, and and then let other people play that. It was ton of fun. Working on a game is very different. One thing that was interesting is like it's kind of hard to know like everything's a good idea in some ways. You feel like, 'Oh, that that would probably be fun.' And so, like the amount of scope you can easily get into is quite wide.

是啊。我们这才明白为什么那些 3A 大作要花 8 年才能发布——除非你做的是小型手游。很容易就能看出它有多庞大。那么，你是怎么来到 OpenAI 的呢？

Yeah. And we realized why like a these AAA studios take 8 years to to to release a game. Unless you're making, you know, a small mobile game. Like it's easy to it's easy to kind of see how sprawling it can really be. So then, what's the story? Like how did you get to OpenAI?

首先，当时 ChatGPT 或 GPT-4 刚发布，我们创业大约一年。我们的创始人离开了一段时间，回来后说：“伙计们，这改变了一切。我们需要思考这对所有事情意味着什么。”他没有给出具体方向，只是说：“这是真家伙。”他在大学学过 AI，有相关背景，他说：“是的，现在是一个真正不同的时间点。”我们当时用 AI 来制作游戏。回想起来很有趣，那时我们大概有 GitHub Copilot，可以做些自动补全，还有大量在 ChatGPT 和 VS Code 之间复制粘贴的工作。我当时做很多前端开发，那体验令人难以置信，但想想从那以后进步了多少，又觉得好笑——就像石器时代，而那只过了两年。石器时代，我基本上是在手写所有前端代码。真实情况是，我和另一位设计师都收到了当时 OpenAI 产品负责人的联系，我们过去都和他共事过。他刚加入 OpenAI，负责 ChatGPT，他说：“看，这东西发展得很好。我们有一个很棒的团队，但需要快速扩张，需要有人帮我们思考如何达到下一个规模。”他当时并不知道我们在一起工作，所以试图分别招聘我们俩。我们聊了之后，他发现我们还有一个工程团队，于是一来二去，我们八个人作为一个团队一起加入了 OpenAI。那段时间很有趣：周五我们还在讨论某个随机的游戏机制，下周一就已经深入 OpenAI 的“战壕”，琢磨下一步要发布什么以及整个公司如何运作。

So, first of all, this was when ChatGPT or GPT-4, I think it kind of come out when we're like a year into the the startup. Our founder went away and came back, and he's like, 'Guys, this is this is changing everything. We need to like think about what this means for everything.' It wasn't any specific direction, but he was just like, 'This is like the real deal.' He studied AI in in college and has a background in all that stuff, and he was like, 'Yeah, this is like a true kind of different point in time now.' We were using it to make the game. And this was like so funny, but way back in the day, this was like I guess we had GitHub Copilot, so you could kind of do some auto-complete stuff. And a lot of copying and pasting between ChatGPT and and and back into VS Code or whatever. I was doing a lot of front end at the time. It was incredible, but it's funny to think how far the work has come since then. Like stone age, and it was only 2 years ago. Stone age. Like I was basically writing all this front end, you know, by hand. And basically what happened was I mean, the true story is that me and the designer were both sort of had been reached out to by the head of product at OpenAI at the time, who we had both worked with in the past. He had recently joined OpenAI. He was leading ChatGPT, and he was like, 'Look, this thing's going great. We have a great team, but we need to grow it really fast, and we need people that can help us figure out how to like get to this next scale.' He didn't actually realize that we were working together when he was kind of trying to sort of hire both of us. And so, we talked, and he also found out about the engineering team that we had, and one thing led to another, we all ended up joining as a team. There were eight of us. It was a pretty funny like time where we went from like talking about some random specific game mechanic on like a Friday, and then like that next Monday, you know, we're we're like deep in the OpenAI trenches trying to figure out what we're launching next and and how this whole place works.

快速插播一条消息，然后我们继续。如果你和我一样，最近经常用代码做原型，但问题是你必须在两种工具间选择：要么是没连接到实际代码库和设计系统的工具，要么是本地主机原型，分享起来超级麻烦。这就是为什么我喜欢 Descript 正在做的事情。只需一键，你和团队中的每个人都可以直接在代码库中制作原型，而无需打开 IDE。Descript 会提取你的设计语言，为你提供完美的探索沙盒，没有任何技术障碍。当你准备好后，右上角有一个方便的分享按钮，可以发送给团队中的任何人。这真的很重要，你现在就可以连接代码库并开始制作原型。只需访问 dive.club/descript。网址是 d e s s n。我万万没想到，过去几个月我的设计工作流程会发生如此巨大的变化。

Real quick message, and then we can jump back into it. If you're like me, then you're prototyping a lot in code lately. But the problem is you kind of have this choice between a tool that isn't hooked up to your actual code base and design system, or a local host prototype that's super annoying to share. That's why I love what Descript is doing. In one click, you and everyone on your team can prototype directly in your code base without ever opening an IDE. Descript extracts your design language and gives you the perfect sandbox to explore without any of the technical hurdles. And when you're ready, there's a nice little share button top right, and you can send it to anybody on your team. It's a pretty big deal, and you can connect your code base and start prototyping today. Just head to dive.club/descript. That's d e s s n. Never in a million years did I think my design workflow would change so drastically in the last few months.

其中很大一部分是 Paper 的新快照工具。它是一个 Chrome 扩展，可以让你从你的在线网站复制任何组件或元素，然后直接粘贴到 Paper 中作为可编辑的图层。所以，我经常从生产环境中抓取一些东西，让 Claude 立即在 Paper 的画布上以六种不同的方式呈现出来。然后当我准备将概念发回给 Claude 时，由于 Paper 的画布使用真实的 HTML 和 CSS，整个过程是无缝的。对我来说，这种工作流程很像设计的未来，而且你今天就可以开始这样做。只需访问 dive.club/paper 即可尝试。现在进入正题。

And a big part of it is Paper's new snapshot tool. It's a Chrome extension that lets you copy any component or element from your live website, and then paste it directly into Paper as editable layers. So, all the time I'll grab something from prod, have Claude immediately spin it up six different ways on Paper's canvas. And then when I'm ready to send a concept back to Claude, it's seamless because Paper's canvas uses real HTML and CSS. That workflow feels a lot like the future of design to me, and you can start doing this today. Just head to dive.club/paper to try it out. Now on to the episode.

在我们深入探讨当下的事情之前，我想聊聊你刚融入 OpenAI 的头几个月。与过去的经历相比，OpenAI 的“战壕”有哪些独特的特征？

Before we get all into the present day stuff, I want to talk a little bit about some of those first few months even as you're assimilating. Like what were some of the defining characteristics that made the OpenAI trenches unique compared to past experience?

从第一天起，这里就截然不同。我清楚地记得第一次全体大会。他们会在这些大会上展示：“这是我们今天所处的位置，这是我们看到的未来可能实现的事情。比如这些模型的发展方向。” 然后你真的会感受到进步的节奏以及随之而来的东西。那一刻我就想：“哇，这个地方不一样。”

From day one, it was just such a different place. I remember very distinctly my first all-hands. You know, they have these all-hands, and they kind of show you, 'Here's where we're at today, and here's what we're seeing is possible with the future. Like where these models are going.' And it really just sinks in the rate of progress and what is coming from that. Moment I was like, 'Wow, this place is different.'

我认为最大的不同是在一个研究主导的环境中工作。OpenAI 最初是一个研究实验室，这种 DNA 依然存在。这与我非常熟悉的 Instagram 或那些从消费产品起步的公司截然不同。OpenAI 始于一个研究实验室，试图弄清楚他们要构建什么。而 ChatGPT 是作为一个低调的研究预览发布的。从这个角度来看，整个公司的运作方式都不同。公司非常以使命为导向，真正思考如何向未来长远发展。

I think the biggest difference is working in a research-led environment. OpenAI started as a research lab, and that DNA is very much there. It's very different than something like Instagram, which I was very familiar with, or maybe a more kind of company that started as a consumer product. OpenAI started as a research lab, just trying to figure out what they're going to build. And ChatGPT launched as a low-key research preview. When you think about it through that lens, it's just a different way the entire company approaches it. The company is super mission-driven and really thinking about how we develop really far forward into the future.

我们能否更深入地探讨一下，当设计师与一个研究实验室紧密合作时，如何才能茁壮成长？这如何改变了日常或每周的设计实践？

Can we go a little deeper into what it looks like to thrive as a designer when you're working closely with a research lab? Like how does that shift the practice of design on a daily or weekly basis?

我们的大部分工作是弄清楚模型擅长什么，然后将其包装成一个人们能够理解并使用的产品。我们希望设计师非常接近模型端，深入思考、尝试，看看它在哪里出错、在哪里失败，理解我们可以用来调整行为的机制。但很大程度上，我认为归根结底就是要有极大的好奇心。在这里工作你不必是技术专家，但我认为你必须非常好奇，非常想了解人们会如何实际使用它？很多时候你基本上就是拿到这个新东西，然后我们必须想办法把它产品化，把它变成人们可以在电脑或手机上使用的东西，然后感叹：“哇，看看这东西能做什么。” 是的，这其实就是把模型本身当作产品来思考。

So much of our work is figuring out what the models are good at, and then trying to wrap that in a product that people can understand and can use. We like our designers to be really close to the model side, and really thinking deeply about and playing with and trying to see where it breaks, where it falls down, understanding the kind of mechanics we can use to tweak the behavior. But a lot of it is really just having a ton of curiosity, I think, is what it really comes down to. You don't have to be technical to work here, but I think you have to be really curious and really interested to see how might somebody actually use this? A lot of times you basically just get handed this new thing, and we have to figure out how we can productize that, how we can get that into a thing that people can now go to their computer or their phone and be like, 'Wow, look at this thing it can do.' Yeah, it's really just a lot of thinking about the model as the product.

我认为我们试图花更多时间的一件事是，很多时候我们会问：“我们能不能不用像素来做这件事？能不能用 token 来做？能不能用模型本身来做？” 并且总是试图推动更多这样的做法。哪些事情可以直接在对话中、在你使用体验的流程中发生，而不是去想我们需要什么样的新定制 UI？显然需要平衡，我认为有时候你确实需要 UI，有时候不需要。这只是一种新的工具或新的材料，我认为你必须考虑如何与之合作。

I think one thing we try to spend more time on is honestly like a lot of times it's like, 'What can we do this without pixels? Can we do this with tokens? Can we do this with the model itself?' And always trying to push to do more of that. What can happen directly in the conversation and in the flow of how you're using the experience versus trying to think about what new kind of bespoke UI we need for this? There's obviously a balance, and I think there's, you know, sometimes where you really want that, and sometimes where you don't. It's just a new tool or new kind of material that I think you have to think about working with.

我想对于大多数听众来说，他们可能不太了解在像素之外进行设计是什么样的。那么，我们能深入探讨到什么程度？也许有没有一些你们内部正在做的、更多发生在模型层面的设计决策或探索的例子？

I think for most people listening, they probably don't have as much of a grid for what it looks like to design outside of the pixels. So, like how deep can we go on that? Maybe is there an example of some of the design decisions or explorations that you all are doing internally that are happening more at the model level?

我给你举一个我们正在思考的例子，我们其实还在实验阶段。如果你考虑如何向新用户介绍 ChatGPT，会有很多产品都采用的传统引导流程。比如有导览、创建账户，也许还会问一些问题来了解你。这很有效，人们也习惯了，但你可以完全不同的方式来思考，对吧？你可以说，实际上我们有一个超级智能的模型，它可能能更好地理解这个人的目标是什么、他们想学什么、他们可能有什么问题、我们应该告诉他们哪些功能来帮助他们完成任何任务。我们真的在剥离很多传统做法，并试图说：实际上，让我们思考一下如何给模型提供上下文，告诉它这个人是全新的，可能需要一些指导，或者需要了解某个特定功能，或者甚至需要理解这个东西到底是什么，因为它与你过去用来获取信息或学习的传统工具截然不同。那么，我们如何让模型来完成工作，而不是写一堆静态的解释来说明这是什么？于是你就开始思考系统提示词和模型行为之类的东西，这是一种全新的、有趣的工作方式。有了这个，你可以快速原型设计：如果我们什么都不做，它通常会输出什么。那么，如果我们给它不同的上下文，或者我们尝试告诉它这个或那个，行为会如何改变？它是否变得更友好、更容易使用，或者帮助人们更好地理解你能用这个东西做什么？所以我认为这是一个例子，希望从事这方面工作的设计师能花更少的时间在 Figma 或任何你用来画像素的工具上，而花更多的时间真正思考你如何与这个东西互动，以及模型确实是核心产品这一事实。

I'll give you one example that we're thinking about right now which we're still just honestly experimenting with. If you think about how you introduce ChatGPT to a new user, there's this very traditional onboarding flow that lots of products do. There's kind of this tour and account creation and maybe trying to ask you some questions to learn about you. That's been effective and people are used to that but you know, you could think about that totally different, right? You could say, well actually we have this super intelligent model that could probably do a much better job trying to understand what this person's goals are, what they're trying to learn or what questions they may have, what functionality we might want to tell them about that will help them with whatever task they're trying to carry out. We're really stripping back a lot of maybe what you might traditionally do and trying to say, well actually let's think about how we should give this context to the model that this person is brand new and they might need some hand-holding or they might need to learn about a specific feature or honestly understand what this thing even is because it's so different than what maybe you've traditionally used before to get information or to learn or whatever. And so how can we let the model do the work versus trying to write a bunch of static explanation of what this thing is. And so there you start to get into thinking about the system prompt and the model behavior and that sort of thing which is just like a totally fun new way of working. With that you can quickly prototype: here's what it would typically output if we did nothing. Well, what if we gave it a different context or what if we tried telling it this or that and how does that actually change behavior? Does that make it more friendly or easier to use or help people understand better what you can do with this thing? And so I think that's one example where designers working on this are hopefully spending a lot less time in Figma or whatever tool you use to draw pixels and more time really thinking about how you interact with this thing and the fact that the model really is the core product.

有没有一套指导原则，帮助你思考何时采用界面级别的解决方案，何时将其外包给模型？因为我想象可能有各种不同定制界面的潜力。

Is there like a set of guiding principles that help you think about when to reach for interface level solutions and when to outsource it to the model? Because I would imagine there's probably a lot of potential for all kinds of different bespoke interfaces.

我知道你现在做了学习体验。你如何在保持聊天作为交互范式的简单性的同时，又充分发挥这种体验的潜力？

I know you did the learning experiences now. How do you maintain the simplicity of chat as an interaction paradigm while also reaching for the full potential of what this experience can be?

这个问题问得真好。我们没有原则，也许应该有。我觉得目前更多是靠直觉。你说得完全对，纯文本绝对不是终极形态，我们也看到了这一点。所以现在聊天能做各种事情。你可以让它做点什么，比如让它写东西，它不会只是说“好的”然后给你一堆文本，再问“你想让我改什么吗？”，你提要求，它再给新回复。我们当时审视了这个问题，进行了批判性思考。我们发现 ChatGPT 的首要用例之一是写作，这是一个很有趣的设计主导的练习，值得聊聊。我们看了人们使用 ChatGPT 的数据，很大比例的人用它来写作，辅助写作。而写作内部又有许多不同的用例，其中有些地方挺别扭的。比如经典的场景：有人把整段内容复制下来，粘贴成邮件，底部还带着“如果你想要更短请告诉我”之类的。这是一方面，还有交互循环的问题。如果你只想改一个词，就会很繁琐，因为你得说“把第二段改成这样那样”。所以我们想更偏向直接操作。现在，当你让它写作时——我们还在逐步推出，所以有些写作场景你会看到，有些还不会，但我们希望覆盖所有写作用例——你会得到一个容器，它把你写的内容放进去，你仍然可以用聊天说“哦，把它变长点，变短点”，它会以传统方式工作，但现在你可以选中文本的某一部分，删除它，要求特定修改，只针对那一处。这是一种更符合人体工程学的工作方式。这是一个例子，混合了模型行为以及与模型的协作：决定何时显示这个功能，何时不显示，以及如何在这个写作块中直接操作。我们考虑的下一件事是，如何确保我们构建的系统，每当添加这种与纯文本略有不同的东西时，能成为一个系统，让我们构建这些不同的积木块，最终让模型能够根据用户手头的任务，以定制化的方式组合和调用它们。

That's a really good question. We don't have principles. We probably should. I think it's more at this point a little bit intuition. You're totally right that just text is definitely not the end-all, and we're seeing that. So we have all sorts of things that you can do with chat now. You can ask it for something, and if you're asking to write something, instead of it just saying "sure" and then giving you a bunch of text, you know, "do you want me to change anything?" and you ask and you get a new turn. What we did there was we looked at that and we just thought critically. We saw that one of our number one use cases for ChatGPT, and this was a fun design-led exercise, a fun one to talk about. We looked at the data of how people use ChatGPT. A huge percentage of people use ChatGPT to write, to help them write. And within that, there are lots of different use cases for how you do that. There are a bunch of things that are just kind of funky with it. For example, there's the classic: you see somebody copy the entire thing and they paste that, and it's like an email to somebody, and at the bottom there's like "let me know if you want this shorter or something." That's one thing, but also just the interaction loop. If you want to just change one word, it was kind of tedious because you would then have to say "okay, change the second paragraph to say this or that." So we wanted to lean into more direct manipulation. With that, now when you ask it to write, and we're still rolling this out, so for some writing cases you'll see it, others you won't, but we want to get this out to all writing use cases. Now you get this container where it puts what you were writing into the container, and you can still use chat to just say "oh, actually make this longer, make it shorter" and it'll work the traditional way, but you can now select a certain part of text. You can delete it. You can ask for a specific change, and it's just targeting that one thing. So it's a way more ergonomic way of working. That's an example of where it's a mix of model behavior and working with the model to say when should it show this, when should it not, and then how do you manipulate stuff directly in this writing block. The next thing we think about too is how do we make sure we're building a system so that anytime we add one of these things that feels a little bit different than pure text, how can we make that a system so that we're building these different building blocks so that eventually the model, I think, will be able to compose and pull these together in bespoke ways specific for what that user's task is at hand.

你认为 OpenAI 最优秀的系统思考者有哪些特质？

What do you think are some of the traits of the best systems thinkers at OpenAI?

人们很容易只考虑自己正在做的那个用例。但如果你思考人们如何使用 ChatGPT，它是非常流动的。你可能在做一个事情，然后流到另一个事情；前一秒你在问即将到来的旅行该带什么，下一秒你在帮它写一封给老板的邮件，再下一秒你又在做研究。人们以非常不同的方式在这些上下文之间切换。所以我认为最优秀的系统思考者不仅考虑自己的功能，还考虑这个功能如何扩展系统。我们非常努力思考的一件事是，这些产品需要哪些底层原语来满足用户需求，这样你一旦构建了它，就能让其他所有事情变得更好。我认为你可以看到一些例子开始成形，比如技能（skills）就是这些产品开始拥抱的一个新原语。我们越能找到这些最深的抽象，它们封装了你想要做的事情，那么我们就可以将其用于模型可能需要的其他功能或任务。最终我们希望拥有一个系统，不仅人类能理解，而且模型也能推理并知道何时以及如何使用它。

It's easy to kind of think about the one use case that you're working on right now. If you think about how people use ChatGPT, it's very fluid. You might be working on one thing and then flowing into something else, or one moment you're asking a question about what to pack for a trip coming up, and then next you're helping it write an email to your boss, and then next thing you know, you're doing some research. People move between these contexts in very different ways. So I think the best systems thinkers are thinking not just about their feature, but how this feature extends the system. One thing we're really trying to think a lot about is what are the underlying sort of primitives that these products need in order to serve the user's needs, so that if you build this once, it's going to make every other thing that we're trying to do better. I think you can see some of these examples that are starting to crystallize around, for example, skills is like this new primitive that I think these products are starting to embrace. The more we can find these deepest abstractions that encapsulate what it is you're trying to do, then we can use that for other features or other tasks that the model might want to do. Eventually we want to have a system that not only humans can understand, but a model can reason about and know when and how to use this.

我想聊聊什么会被发布以及如何发布，因为你在很多方面处于研究驱动创新的熔炉之上，所有这些新能力不断涌现。这是一个非常通用的产品，被全世界几乎每个人使用。你可以朝几乎任何方向发展，所以有大量的实验在进行。跟我聊聊作为 OpenAI 的设计师，有效引导一个想法是什么样的？你如何从原始实验到真正推出像直接操作或新的学习体验这样的东西？

I kind of want to talk about what gets shipped and how, because you're in many ways sitting on top of this cauldron of research-led innovation, and all these new capabilities are popping up. It's a very generalizable product that is being used by literally everyone in the world. You can go in almost any direction, and so you have all of this experimentation that's happening. Talk to me a little bit about what it looks like to effectively steward an idea as a designer at OpenAI. How do you go from raw experiment to actually getting something like that direct manipulation or the new learning experiences out the door?

很多都是从原型或设计开始的。当然不只是设计师，任何人——工程师、设计师、产品经理、研究员——都会有这些好主意。但从设计师的角度来看，众所周知，构建一个能用的版本已经变得容易多了。最好的版本是这样的：设计师有了想法，然后用 Codex 或任何你想用的工具，构建出真正的版本，这些版本不只是可点击的原型，而是实际上有实时模型响应，并与模型一起工作。所以我认为这打开了大门。比如这些数学功能：那是一位设计师在思考，如果你问一个问题，它返回 LaTeX 格式的数学表达式，这看起来很过时。这不是你应该学习的方式。你应该能够与这些东西互动。所以那位设计师观察到了这一点，拼凑了一个版本，然后实际上用很多东西尝试了，他们说：“嘿，你拿到模型，告诉模型，如果你通常会这样回复，现在试试这样回复。”当我们分享这个时，它显然非常有价值，团队围绕它团结起来，最终帮助它完善，扩展用例，并发布了它。所以我认为有很多这样的例子。总的来说有很多自下而上的东西。这是我认为 OpenAI 不同的另一件事：大量的自下而上的工作。

So much of it starts with a prototype or design. It's not just a designer, of course; anybody has these great ideas—an engineer, a designer, a PM, a researcher. But from a designer's perspective, as everybody knows, it's become much easier to build a working version of something. The best versions of these are like a designer will have this idea, and now with Codex or whatever tool you want to use, you can build real versions of this that aren't just clickable prototypes but are actually like live model responses and are working with the model. So I think that opens the door. For example, these math ones: that was a designer who was just thinking about if you ask for something and it gives you back LaTeX, like math expressions and all that, it seems pretty archaic. That is not the way that you should be able to learn. You should be able to interact with these things. So that was a designer who just observed that, put together a version, but then actually just tried it with a bunch of things where they said, "Hey, you know, you get the model and you tell the model if you would typically respond with this, try respond with this now." I think it was just so clearly valuable when we shared that around that the team rallied around it and eventually helped harden it, helped expand the use cases, and shipped it. So I think there are a lot of pockets like that. There's a lot of bottoms-up stuff in general. That's another thing that I think is different at OpenAI: just tons of bottoms-up stuff.

我认为这可能是行业发展的方向，因为现在任何人都能实现一个想法。你只需启动 Codex 或任何工具，基本上就能交付。这是我们经常思考的问题：我们希望赋能每个人去拥有这些想法，然后问题就变成了如何编辑并确保所有内容作为一个系统协调一致，并且我们交付的是正确的东西。

And I think this is probably the way some of the industry is going, because anybody can build an idea now. You can just fire up Codex or whatever tool and ship it basically. That's something we think a lot about: we want to empower everybody to have these ideas, and then it becomes about how you edit and make sure it all fits together as a system and we're shipping the right things.

回顾你刚加入时，你提到你还在从 ChatGPT 复制粘贴，而现在一切感觉就像从指尖射出闪电。这如何改变了团队协作的方式、想法呈现的方式？你看到设计实践因这项新技术发生了怎样的转变？

Looking at when you joined, you mentioned you were copying and pasting from ChatGPT, and now everything feels like shooting lightning bolts out of your fingertips. How has that changed the way teams collaborate, the way ideas get presented? How have you seen the practice of design shift as a result of this new technology?

当我加入时，那是两年半前，有些设计师在使用 Origami。有些设计师技术很强，真的在用 API 工作。比如，他们会在 playground 里尝试不同的东西。我觉得这对更熟悉传统设计工具的设计师来说不那么容易上手。快进到今天，显然你有了——这真的发生在过去几个月，想想都觉得疯狂。先是有了 Cursor，我觉得那很酷，但即使那样，对团队的一些部分来说还是有点难以触及。而现在有了 Codex，以及这些产品未来的发展方向，你有了能真正为你做实际工作的工具。这更多是关于把你的想法拿出来，放进去，并恰当地表达出来。整个行业都在努力弄清楚我们需要什么样的工具来让它超级高效。例如，如何把它连接到设计系统，这样它就不会只是吐出随机的 UI，而是感觉适合我们系统的东西。那么，你如何在原型设计、生产工作、Figma 和所有这些之间建立联系？但我看到越来越多的设计师在拥抱这一点。看到这个真的很酷：突然间，不再是 Figma 原型或静态的东西，甚至不是视频录制，而是非常互动。我可以去点击、摆弄它。在很多情况下，我可以开始提问，或者看到实际的模型行为会是什么。而且这不仅仅是关于设计，这是很酷的一点。例如，我们使用一个内部智能体，它就像一个数据科学家。任何在某个项目上的设计师都可以提问：“人们实际上是怎么用这个的？给我数据，看看它是怎么被使用的，用例是什么？”无论你的问题是什么。现在你可以对你所做的决策有更充分的了解。当你在规模化工作时，这非常有价值。

So when I joined, two and a half years ago, some designers were using Origami. Some designers were very technical and were really working with the API. For example, they'd be in the playground trying different stuff. I think it wasn't as accessible to designers more familiar with traditional design tools. Fast forward to today, obviously you now have—and this really happened in the last few months, which is crazy to think. First you had Cursor, and I think that was cool, but even that was a little bit inaccessible to some parts of the team. And now with Codex and the future where these products are going, you have tools that can really go and do real work for you. It's more about getting your idea out and into it and expressing it properly. There's a whole effort going around to figure out what tooling we need to make this super effective. For example, how to hook it up to a design system so that it's not just spitting out random UI, but something that feels like it would fit within our system. So how do you make the connection between prototyping, production work, Figma, and all of this? But I think we're seeing more and more of our designers embracing this. It's been really cool to see: all of a sudden, instead of a Figma prototype or static thing or even a recording of a video, it's very interactive. I can go and tap around and play with it. And in many cases, I can start asking questions or see what the actual model behavior would be. And it's not just about design, which is one cool thing. For example, we have an internal agent we use which is like a data scientist. Any designer working on something can ask questions: 'How do people actually use this? Give me the data on how this is used, what are the use cases?' Whatever your question might be. Now you can be way more informed with the kind of decisions you're making. And when you're working at scale, that's really valuable.

我喜欢这一点，因为它说明了我们对其他人的依赖在减轻，从而能够探索并把东西发布到世界上。你描述的一切都让我想起，我多少次被困在，比如，Metalab 之类的地方，努力回忆怎么做 SQL 查询。我就是想要这个答案，你知道吗？一旦这些东西继续发展，每个人都拥有它们，那将会很疯狂。我觉得每个人都能做更多的事情。另一个有趣的事情是理解什么时候该用什么工具包。我认为我们这个行业还在进化，对吧？在某些情况下，你肯定应该编写一个实时原型，或者使用工具来构建一个实时原型。在其他情况下，纸上草图也很有价值，是过程中正确的一部分。我觉得这方面的讨论很有趣。很容易偏向一边或另一边，但有时你想要白板，有时想要纸上草图，有时想要线框图，有时想要 Figma，有时想要这些实时原型。我认为最好的设计师会理解并凭直觉知道什么时候该用哪一个。因为有时你想要非常广泛，尝试一千个或一百个想法。我不认为我们已经有很好的工具来帮助我们做那部分，那种广泛的探索。我认为这是一个我感兴趣看到很大发展的领域。因为你不想只拿一个原型。很容易沉迷于那个原型，然后想，‘我们怎么改进这个？’而你的想法可能完全偏离了方向。我认为我们需要弄清楚这种深度和广度的问题，不要过度偏向一边。作为设计师，你需要非常灵活，理解什么时候该去哪里。特别是对于一个没有丰富前端背景的设计师，突然之间你能做出这个东西，感觉棒极了，你会想，‘这是我一生中做过的最伟大的东西，因为它功能完整’，但它可能完全错误。

I love that because it speaks to this lightening of reliance that we have on a bunch of other people to be able to explore and put something out into the world. Everything you're describing is like, man, the amount of times I've been stuck in, I don't know, Metalab or something trying to remember how to do a SQL query. It's like I just want this answer, you know? It's going to be wild to think once these things continue to develop and everybody has these. I think everyone will be able to do so much more. And the other thing that is interesting to think about is just understanding when to turn to what toolkit. I think we're still kind of evolving as an industry, right? In some cases you should definitely be coding a live prototype or using a tool to build a live prototype. In other cases a paper sketch is also valuable and a right part of the process. And I think the discourse has been interesting. It's easy to swing one way or the other, but sometimes you want a whiteboard, sometimes a paper sketch, sometimes a wireframe, sometimes Figma, and sometimes these live prototypes. I think the best designers will understand and intuit when to go to which one. Because sometimes you want to go super wide, you want to try a thousand ideas or a hundred ideas. And I don't think we have great tools yet to help us do that part, the kind of wide exploration. I think that's one area I'm interested in seeing evolve quite a bit. Because you don't want to just take one prototype. It's pretty easy to get obsessed with that and just be like, 'How do we refine this?' And your idea might be way off. I think there's this depth and breadth thing that we need to figure out and not swing too far one way. As a designer, you just need to be really flexible understanding when to go where. Especially as a designer who doesn't have a rich front end background, and all of a sudden you can make this thing and it feels amazing and you're like, 'This is the greatest thing I've ever made in my life because it's fully functional,' but it might be completely wrong.

100%。而且，你知道，我最初是设计师兼工程师的混合体。老实说，暴露年龄了，我觉得大多数设计师都会编码——当时我们在写前端代码。我那时非常喜欢 37signals 那帮人，那种你非常接近构建实时体验的东西。我认为很多人更接近代码，把它当作一种材料。你有点像很多设计师兼工程师的混合体。当然，也有像图标设计师和视觉设计师这样的专家，但传统的软件设计真的是扮演两种角色。然后随着行业的发展，人们开始专业化，你有了这种新的产品设计，它肯定更偏向纯设计，然后我们会有专业工程师来构建它。我认为现在正在改变，它正在摆回来，两者可以更多地重叠。所以有一个大问题：什么样的技能组合，你在哪里、什么时候工作，以及如何工作？这正在发生相当大的变化。

100%. And, you know, I started as a hybrid designer-engineer. Honestly, dating myself, I think most designers coded—we were writing front end code at the time. I was really into the 37signals crew and that kind of stuff where you're just really close to building this live experience. I think a lot of people were much closer to code and were using it as a material. You were kind of a lot of designer-engineer hybrids. Of course you had specialists like iconographers and visual designers, but the traditional kind of software design was really about playing both roles. And then as the industry grew, people started to specialize, and you had this new kind of product design, which was definitely more on the pure design side, and then we're going to have specialist engineers to build that. I think that's now changing, it's kind of swinging back where both can overlap way more. So there's a big question around what skill set, where and when do you work, and how do you work? It's just shifting quite a bit.

那我们就深入探讨一下，因为作为领导层的一员，你必须思考这走向何方。我们未来想如何工作？如果你进行外推，这看起来会是什么样子？那么当你思考哪些技能变得更有价值时，你提到了好奇心。

Let's pull on that a little bit then, because as somebody in a leadership position, you're having to think about where this is going. How do we want to work in the future? If you do extrapolate, what does this kind of look like? So when you think about the skills that become more valuable, you mentioned curiosity.

你最近在设计师身上寻找的其他特质还有什么特别在意的吗？

Is there anything else that's top of mind for you that you're looking for in designers today?

是的，从某些方面来说，我认为优秀设计师的本质并没有改变。我们团队做了很多这样的工作：深入思考我们要解决什么问题、为谁构建产品，尝试各种想法然后收敛到一个方案上。当然还有工艺、品味以及对交互模式的理解。这些都没有改变，而且同样重要。我认为关键在于理解如何与快速变化、不断演进的事物合作。它不是一成不变的，每天都在变化，可以变形并做很多不同的事情。然后你有了这些新工具，所以你可以用非常不同的方式表达你的想法。回想我在 Instagram（当时属于 Facebook）的时光，吸引我加入那家公司的是他们对设计工具的深度投入。我加入时用的是 Quartz Composer。他们在上面构建了各种自定义功能，Mike Matas 和其他几个人主导了这项工作。他们构建了 Origami，那是 Brandon Walkin 做的。看到设计师从 Sketch 转向使用 Origami 后，他们的想法能更好地传达出来，这真是太酷了。他们不仅要思考外观，还要思考它是如何工作的、感觉如何、如何在状态之间切换。当时我们正从 Web 转向移动端，这非常重要，因为很多都关乎你如何与这个东西交互。那个工具在那个时间点做得非常好。我看到的是，拥抱它的人能够提升他们的工作、工艺和作为设计师的影响力。我现在看到同样的转变正在发生，借助这些新工具。我认为 Codex 或 Cursor 或任何东西都能真正为设计师解锁一种新的自我表达方式。

Yeah, in some ways I think nothing has changed as far as what a great designer is. Our team does a lot of this: thinking really deeply about what problem we are trying to solve, who we are building this for, and trying a bunch of different ideas and converging on one. And then of course craft, taste, and understanding of interaction patterns. None of that has changed and is just as important. I think it's just all this understanding of working with something that is changing very quickly, that is going to evolve. It's not set in stone; it's changing every day and can shapeshift and do many different things. And then you have these new tools, so you can express your ideas very differently. Going back to my time at Instagram, which was part of Facebook, one thing that drew me to that company was their deep investment in design tools. When I joined, it was Quartz Composer. They had all this custom stuff built on top of it that Mike Matas and a few others led the charge on. They built Origami, that was Brandon Walkin. It was so cool to see that when a designer went from Sketch to using Origami, you could totally see their ideas come across much better. They had to think through not just how it looks but how it works, how it feels, and how you move between these states. At the time, we were shifting from web to mobile, and that was really important because so much of it was about how you interact with this thing. That tool did a really good job for that point in time. What I saw is the people that embraced it were able to up-level their work, craft, and impact as a designer. I see that same shift happening now with these new tools. I think Codex or Cursor or anything can really unlock a new way of expressing yourself for a designer.

是的，因为这不仅仅是关于视觉，还关乎感觉，但在你的案例中，内容本身也是设计，这真的很有趣。我从未在这样一个地方工作过，那里内容本身和它所在的圆角半径一样都是设计的一部分。

Yeah, because it's not even just about the visuals, it's also about how it feels, but in your case it's also like the content itself is design, which is really interesting. I've never even worked in a place where that is as much of the design as the corner radius that it sits in.

是的，这很有趣，因为我们很多人可能都在用户生成内容就是人们所见的地方工作过。所以 Instagram 有点像是一个外壳，对吧？当你设计时，你是在设计外壳，然后你不知道里面会放什么内容。你无法控制。但这介于两者之间，因为我们确实对放入的内容有一些控制，但又不是完全控制。你不知道用户会具体如何使用、接近或提问。但你可以稍微调整模型的行为和响应方式。所以它介于无控制和有控制之间。这完全不同。

Yeah, and it's so interesting because a lot of us have probably worked in places where user-generated content is kind of what people see. So Instagram is kind of a shell, right? When you're designing, you're sort of designing the shell and then you don't know what content's going to go in there. You don't have control over that. But this is like something in between because we do have some control actually over what goes in, but not really. You don't know exactly the way the user is going to use this or approach or ask. But you can play a little bit with how the model should behave and how it should respond. So it's somewhere in between this kind of no control and some control. It's just totally different.

过去一年，Dive Club 让我非常清楚地认识到，设计实践正在发生变化。旧的反馈流程在当今世界已经不太适用了。这就是为什么我很高兴地宣布 Inflight 正式进入公开测试阶段。它是我一直想要的反馈工具，专为以 AI 速度运转的世界而构建。我可以分享我的原型，在视频演示中提供上下文，Inflight 让我轻松获得继续前进所需的确切反馈，无论是投票选择方向，还是获得发布新想法的许可。所有这些都可以通过一个链接实现，我可以把它放到 Slack 中，甚至与高级用户分享以测试新原型。我每天都在使用 Inflight，它彻底改变了我分享工作的方式。所以我非常期待你试用这个产品。如果你想讨论它，只需发邮件到 rid@inflight.co。

One thing that Dive Club has made abundantly clear to me over the last year is that the practice of design is changing. And the old process of getting feedback just doesn't quite cut it in today's world. That's why I'm excited to announce that Inflight is officially in open beta. It's the feedback tool that I've always wanted and it's built for a world that moves at the speed of AI. So I can share my prototypes, give context in video walkthroughs, and Inflight makes it easy to get the exact feedback that I need to move forward, whether it's voting on directions or maybe even getting the green light to ship a new idea. And all of this is available in a single link that I can drop into Slack or maybe even share with power users to test out a new prototype. I use Inflight every day and it's totally transformed the way that I share work. So I'm excited for you to try the product. And if you ever want to jam about it, just email me at rid@inflight.co.

你之前跟我提到，头两年有点像拼命坚持。你看到了惊人的规模，那种大多数人无法体验到的规模。现在你似乎到了另一阶段，作为设计领导者，你正在尝试哪些更有意识的转变？

So you mentioned to me that the first two years were kind of holding on for dear life. You're seeing absurd scale, like the level of scale that most people don't get to experience. Now that you're kind of on the other end of this, like what are some of the more intentional shifts that you're trying to make as a design leader?

我仍然在拼命坚持。但我试图不让所有事情都一股脑地涌向设计团队，而是想办法在我们想要做的事情上更加有意识。我认为有一些经典的东西。例如，我们之前从来没有一个设计系统团队，因为我们进展太快了，现在我们正在建立它。酷的是，我们真的试图从第一性原理来思考，如何原型化想法以及如何与模型配合。所以我们有一个完整的系统叫做动态用户界面库，它允许我们设计模型可以解释的东西。这是一种新的思考方式。所以有这一整套东西。还有就是要弄清楚作为设计师需要哪些系统和工具，以便能够快速上手，在这种新的工作方式中做出出色的工作。当然还有我们的流程是什么？老实说，它每天都在变化。有时我们在 Figma 的世界里，有时我们在原型化东西，通过 Slack 快速发送东西、给出反馈，行动非常迅速。但我认为我们正在试图找到一个稍微更有意识的方法，同时保持诚实，你必须愿意卷起袖子，快速尝试。

I'm still holding on for dear life. But I'm trying to not just let it all come at us as a design team and try to figure out how we can be more intentional about some of the things that we want to work on. I think there is the classic stuff. For example, we didn't really ever have a design systems team because we were moving so fast, and we're establishing that now. What's cool about that is we're really trying to think of it from first principles, how you prototype ideas and how it works with the model. So we have a whole system called the dynamic user interface library, which allows us to design things that the model can then interpret. That's a new way of thinking about this. So there's that whole thing. There's figuring out what are the systems and tools that you need as a designer so that you can hit the ground running and do great work in this new way of working. And then of course there is what is our process? Honestly, it changes day-to-day. Sometimes we're off in Figma land and sometimes we're prototyping things and firing off things over Slack and giving feedback and moving really quickly. But I think we're trying to figure out a slightly more intentional approach while keeping ourselves honest that you have to be willing to just roll up your sleeves and try things really fast.

我们能稍微澄清一下吗？让我们像墙上的苍蝇一样，观察 OpenAI 设计师一周的工作。它是怎么运作的？

Can we add a little bit of clarity there? Let us be a fly on the wall for a week as a designer at OpenAI. How does it work?

我们有很多传统的仪式，我们一直在摸索。很多设计师会有一个想法然后去探索它。例如，我们有一个频道叫做 PD whip，就是设计师的工作进展。你只需把东西扔进去。它必须是原型或视频之类的。我们努力让它成为人们可以轻松回应的东西。这是一个我们用了很久的地方，一个开放的地方，只是分享东西，抛出想法，即兴发挥。

We have a lot of traditional rituals that we're always figuring out. A lot of designers will just have an idea and explore it. We have a channel, for example, that we call PD whip, which is designers' work in progress. You just throw stuff in there. It has to be a prototype or video or something. We try to make it something that people can easily react to. It's a place we've had for a really long time, an open place to just share stuff and get ideas out and riff.

我们确实会做评审，我认为让一群人一起碰撞想法通常很有价值，这本质上就是在构建想法。我们还在摸索是否要做设计评审，如果做的话具体怎么做，而日常工作中，我觉得这取决于你在做什么。我们有一个比较传统的结构，你会和一位 PM 以及一位工程师合作，老实说，很多工作就是构建和迭代。我们努力拥抱这样一个过程：我们得慢慢摸索。我们绝对不会在确认方案可行之前，就花时间去打磨一个完美的解决方案。所以我认为，先做出一个能在产品里试玩的早期版本是一个很好的里程碑，然后从那里开始迭代，花大量时间讨论：有没有其他方式来做这件事？我们有没有考虑过它如何与这个结合，以及如何把所有点连起来。当然，现在当我们发布东西时，我们必须考虑：这是否要推送给所有用户？它能否很好地扩展系统？如果不能，有没有其他方式或事情可以做，来构建系统，让所有部分都 cohesive 地结合在一起？老实说，这是我们正在努力做得更多的事情。

We do do crits and I think that they are generally valuable to just get a bunch of people to kind of riff on ideas and that's really about building ideas. We're still figuring out do we do design reviews and if we do exactly how and then day-to-day I mean I guess it depends on what you're working on. We do have a traditional kind of structure where you have a PM that you're working with and an engineer and honestly a lot of this is like building and iterating. We try to embrace that process of like we're going to have to feel this out. We definitely don't spend our time trying to craft the exact perfect solution until we know that it even works and so I think trying to get to an early version that we can play with in the product is a really good milestone and then iterating from there and spending a bunch of time talking about like is there another way we could do this? Have we looked at how this fits with this and connecting all the dots. Of course now when we ship stuff we have to be thoughtful about is this something we want to ship to all of our users and will this extend the system nicely and if not are there other ways or other things that we can do to build out the system so that it all comes together cohesively which is something we're trying to do more of to be totally honest.

你已经多次谈到系统思维的重要性。所以这显然是 OpenAI 设计工作的一个核心原则。我们讨论了好的系统思维是什么样子，以及它如何发挥作用。那么，有没有一些你试图避免的陷阱？比如，好的系统思维的反面是什么样子？

You've talked about the importance of systems thinking multiple times now. So it's obviously like a core tenant of what it looks like to design at OpenAI. We talked about what it looks like when it's good and when it's working. Are there pitfalls that you're trying to avoid where maybe it's like what's the opposite of good systems thinking look like?

你知道吗？我认为我们试图平衡的是：如何推出那些实验性、早期且会不断变化的东西？这是研究实验室的本质，但另一方面，如何以让真正重要的事情感觉 cohesive 的方式来做？所以我认为更多是理解这种平衡：好吧，这是新的，我们不完全确定，而这是我们认为需要真正固化并做对的事情。那么，不好的系统思维是什么样子？老实说，我认为就是如果你太狭隘，只想着如何尽快推出这个东西，而没有完全理解我们还有一堆其他类似的事情在同时进行。如果我们能把所有事情整合起来——这其实是我的工作。这可能就是为什么我经常思考这个问题：我的工作就是试图连接所有这些事情，然后说，如果我们在这里少做一点呢？如果我们实际上少做这些事情，但用这种方式来做，也许就能把一切整合起来。

You know? I think what we're trying to balance I would say is how do we put out things that are experimental and early and going to change? That's like the research lab nature but then how do you do that in a way that for the things that truly matter feel cohesive. So I think it's more about just understanding the balance of like okay this is new we don't totally know versus like this is something that we feel like we really need to harden and we're going to get right. I guess what is what is not good systems thinking look like? I you know honestly I think it's just if you're a little bit too blinders on and just trying to say well how do we get this thing out as quickly as possible and not totally understanding that we have a bunch of other things going on that are actually pretty similar and if we all came together and that's honestly my job. That's why I think probably why I think about this a lot is like my job is to kind of try to connect all this stuff and say well what if we did less here actually? What if there's a way that we actually did fewer of these things but we did it this way maybe that pulls it all together.

当我思考 ChatGPT 的演变时，我们正在努力解决的问题是：如何收紧我们正在实验的许多新东西，把它们变成一种让你明白何时该用哪个工具，或者如何清晰地展示你能用这个东西做什么——随着我们推出新功能或新进展，我们理解它们可能适合这个系统，但同时也要认识到，老实说我们不知道，因为明天可能就会出现某种东西，完全改变你与这个东西互动的方式。所以你必须非常灵活。

As I think about the evolution of ChatGPT that's something that we're trying to figure out is just like how do we tighten up a lot of the new things that we're kind of experimenting with and turn it into something that feels like you understand when to go to which tool or how to expose what you should what you could do with this thing in a way that is clear as we have new kind of features or new advancements like we understand that that could maybe fit into this system all while recognizing that honestly we don't know because like there could be something tomorrow that comes out that completely changes the way you might want to interact with this thing. So you have to be super flexible.

那么，我们能从动态界面库的角度来谈谈吗？你们在那里遇到的设计挑战或机遇是什么？你们如何设计这个系统，让它能在所有平台上原生渲染？你们如何让它变得可交互？你们如何确保它真正为体验增加价值，而不是仅仅重复以前的做法，而是尝试思考 AI 或 AGI 版本的任何你可能与之互动或工作的东西。然后，我认为未来的大事是这些东西如何整合在一起，以及模型最终如何理解它，这样它也许就能开始帮助我们组合这些组件，而不需要设计师手工制作所有东西。我们还没有完全做到，但我认为这很快就会成为可能。

Can we talk about that through the lens of the dynamic interface library then? What are some of the design challenges or opportunities that you all are wrestling with there? How do you design this system so that it can render everywhere natively? How do you figure out how to make it interactive? How do you make sure that it's truly adding value to the experience and not just like doing what was done before but trying to think of like the AI kind of AGI pill version of whatever you might be trying to kind of interact with or work with. And then I think the big thing looking forward is how do these things come together and like how does the model eventually kind of have an understanding of it so that maybe it can start to help us compose these with without having to have a designer hand craft all of these. We're not quite there yet but I do think that that will be possible very soon.

所以当你思考这些时，你是在考虑这些独立的组件，它们堆叠成一个系统，设计师可以用，工程师可以用，模型也可以用，那么这一切如何整合在一起？

So when you think about these you're thinking about these like individual components that kind of stack up to a system that a designer can use an engineer can use and the model can use and how does that all kind of come together?

这引发了一些有趣的问题：在一个你几乎放开对界面所有控制的世界里，设计师未来的角色到底是什么？

It begs some interesting questions about what the future role of a designer even is in that world where you're kind of loosening all grip on what the interface can be.

我认为这是一个重要的问题。我确实认为，随着这些工具让人们能够创造更多东西，你会需要一位编辑。一个很好的类比是：在相机出现之前，如果你想得到一个人的静态图像，你必须坐下来画。你必须能画肖像，不是每个人都能做到。如果你想画，那很耗时且昂贵。现在，突然之间，任何人都可以拍照。但仍然有优秀的摄影师，那些精通技艺、有品味的人，也有那些只为个人生活拍照的人，技艺在那里并不重要，重要的是记忆和瞬间等等。类似地，现在任何人都可以为任何东西编写软件。但我认为，仅仅因为任何人都能做到，并不意味着它就会是好的、正确的构建方向或正确的直觉。我认为，作为设计师或任何职能，你的工作将越来越多地是帮助编辑、指导和策展。我不认为设计师的工作会很快消失。我认为，如果说有什么不同的话，那就是我们将能够做更多，但核心技能仍然同样重要，因为我认为人们能够区分好的软件和坏的软件，对吧？我们都知道那是什么感觉，它不仅仅是关于软件是否能用，而是关于它如何工作、感觉如何，以及它是否为我解决了正确的问题。

I mean I think it's an important question. I do think that as all these tools let people make more things like you know you're going to need an editor. I think a good example like a good analogy of course is like before the camera if you wanted a still of somebody you had to sit down and paint that thing. You had to be able to paint a portrait and not everybody could do that. If you wanted to it was you know time consuming and expensive. Now all of a sudden anybody can take a photo. There are still excellent photographers and people that like master the craft and people that have taste and then there's you know people that use photography for their personal life and it doesn't matter like the craft doesn't really matter there. It's about the memories and and moments and all of that. There's a similar way you could think about that with the fact that now anybody can write software for anything. I think that just because anybody can do it doesn't necessarily mean it's going to be good or the right thing to build or the right instinct. I think we will as designers or or any function I think like your job's going to be more and more about kind of helping edit and direct and curate. I don't think like the the job of a designer is going away anytime soon. I think that if anything I think we'll be able to do more but I think that the core skills will still be just as important cuz I think people can recognize good software from bad software right? And we all know what that feels like and it's not just about does it work but like how it works and how it feels and is it solving the right problems for me.

你谈到过这种张力：一方面是为当前能力或模型最新产出而设计，另一方面是超级 AGI 的部分。那么你如何处理这种张力？现在你脑子里有哪些更 futuristic 的想法在打转？

You've talked about this tension between designing for like present capabilities or maybe the most recent thing that the models have produced versus you know the super AGI pill part of it. So like how do you deal with that tension and what are some of the more futuristic things that are rattling around your brain right now?

我们经常思考的一件事是能力差距。

One thing we think a lot about is the capability gap.

能力差距指的是，模型现在实际上已经能够做很多事情了，对吧？如果你看看 Codex 以及它能做什么，有时它会消耗大量 token，花费很长时间。再看看人们今天使用的 ChatGPT-4，对吧？实际上，模型实际能力与用户使用之间存在差距，而这差距是最近才出现的。我认为作为设计师，我们必须深入思考：如何以一种让人们理解其能力的方式去呈现它，以及如何提供工具，让他们能在 ChatGPT 中完成真正有意义的工作。这是我们反复思考的问题。我们如何塑造产品，提供工具，让你能真正利用模型的所有能力——如果它们愿意花时间为你工作的话。而我们还没有做到这一点，对吧？比如，在 ChatGPT 内部你能做的功能是有限的，不像 Codex，Codex 可以操作你的电脑，花大量时间，运用各种技能等等。再次提醒自己，仅仅三个月就发生了这么多变化。接下来的三个月会是什么样？再接下来呢？我也提醒自己，我们做这件事才三年左右。想想计算机刚出现时，它们才存在三四年或五年，再看看现在，每一年都有进步，能做的事情越来越多。

The capability gap is saying that the models actually are now at a point where they can do a lot, right? If you look at Codex and what it can do, sometimes it's spending a lot of tokens, it's taking a long time. And then what people use, let's say ChatGPT-4 today, right? There's actually now a gap between what the models are actually capable of, and that actually just happened very recently. I think as designers we have to think a lot about how do you actually expose that in a way that people understand what they can do with it, and how do you give them the tools so that they can go out and do real meaningful work right in ChatGPT. And so that's something we think about a ton. How do we shape the product, give you the tools so you can really take advantage of everything that these models are able to do if they really sit and spend the time to work for you. And we just haven't done that yet, right? Like there's a limited amount of functionality that you can do inside of ChatGPT, unlike if you think about Codex, Codex can go and use your computer and spend a bunch of time and use the different skills and all of that. And again, you always have to remind yourself so much happened just in three months. What's the next three months going to look like? What's the next? I also try to remind myself that we're working on this thing three years or whatever it is into its existence. When computers first came out and they were three or four years into their existence, or five, you know, and then where they are now, and just each year how do you have that advancement and then more and more things were able to be done.

说起来有趣，顺便提一下，有个很好玩的播客叫 30 for 30，是个体育播客。嗯，好的。所以 30 for 30 是那个节目，但他们也有播客。你听过关于 Madden 的那一期吗？

It's funny, side note, but there's this really fun podcast 30 for 30, which is like the sports podcast. Yeah, okay cool. So 30 for 30 is like the show but they also have a podcast. Have you heard the one about Madden?

没有。

No.

好的，有一集关于 Madden 橄榄球游戏的节目非常棒。顺便说一句，这游戏是我童年的支柱，所以我听得津津有味。我等不及看你怎么把话题拉回来。我记得故事是这样的：他们在 NES 上做了这个游戏，或者可能是在 NES 之前，我不记得了。他们想让 Madden 做代言人，Madden 非常感兴趣，但他看了看说：“场上只有八个人？”或者五个？橄榄球是几个人？11 个？我不知道。他们说：“比特数不够，技术上不可能把 22 个球员放在场上，我们总共只能放 10 个。”然后他说：“等你们能做到的时候再来找我。”哇。然后一两年后，芯片进步了，他们突然就能做到了。那是 Madden 最基础的版本，终于有了 11 个球员，然后他说：“好，我代言，我想参与，我喜欢。”这个播客很有趣，因为你能听到他在录音棚里录制音效等等。但这是个很棒的故事，也让人思考技术是如何不断进步的。我们和这些模型打交道时都明白，它们并不完美。你会遇到问题，比如为什么它在某些情况下会失败，或者它很擅长这个但在别的事情上很差。你只能相信一年后，甚至更短，六个月后，你外推一下，看看它会走向何方。这有点难以想象，但我经常思考，20 或 30 年前设计软件或构建软件是什么样子，那时你不断遇到各种限制，处理各种细节。我们讨论上下文管理和上下文窗口之类的东西，我不知道，5 年或 10 年后这些会不会显得过时？甚至提示工程这个概念，我前几天还在想。我们如此强调如何精确格式化和结构化，Twitter 上到处都是那些彩色编码的复制粘贴图形，但现在我甚至不再想这些了。是的，一切变化太快。所以我们的工作就是试着回到原点：你必须处理当下的现实，因为如果你走得太超前，你就无法把 11 个球员放在场上。所以你无法那样设计游戏。所以你必须找到平衡，既要真正理解它现在能做什么，又要能推动它，说我们实际上想做到这个。那么我们如何确保我们努力为那个目标设定愿景？

Okay, there's this really great episode about Madden football the game. This is a pillar of my childhood by the way, so I'm on the edge of my seat. I cannot wait to see how you loop this back in. I just remember the story from that where they'd made this game on NES I think, or maybe it was even before NES, I don't remember. They wanted Madden to be the spokesperson, and Madden was super into it, but he looked at it and he was like, 'There's only eight people on the field,' or whatever, five people on the field. Like football has how many is it? 11? I don't know. They're like, 'They're not enough bits to do that. It's technically not possible to put 22 players on the field. We can only put 10 total.' And he was like, 'Come back to me when that's possible.' Wow. You know, and then like a year or two later, the chips advanced, and then all of a sudden they were able to make it. And that was like the very basic version of Madden where they finally had the 11 players, and then he was like, 'Cool, I'll endorse it. I want to be part of this, I love it.' And it's a really fun podcast because you hear him in the booth recording his sound bites and stuff like that. But it was a cool story of just also thinking about how much technology advances on a regular basis. And I mean we all know this as you're working with these models. They're not perfect. You run into these issues, like why it falls down in a certain case, or it's really good at this but actually really bad at some other thing. And you just have to have faith that a year from now, or even less, six months, and you extrapolate out where that's going. It's kind of hard to imagine, but I think a lot about that and what it must have been like to design software or build software 20, 30 years ago when there were all these limitations you were constantly running into and you're working with these details. I mean, we talk about context management and context window and that sort of thing, and I don't know, is that going to feel archaic in 5, 10 years? And so even the concept of prompt engineering, I was thinking about the other day. Like we put so much emphasis on exactly how to format and structure, and you had all those copy and paste graphics that would be all over Twitter with the different color coding, and now I'm like, I don't even think about that anymore. Yeah, it's all changing so quickly. So it's our job to kind of try to think about going back to, well, you do have to deal with the realities of what it is now, because if you go too far ahead, you can't put 11 people on the field. So you can't design the game that way. So trying to bring it back. So you have to kind of find the balance, and that's where really understanding what it's capable of now and then also being able to push it and say, well we actually do want to be able to do this. So how can we make sure we're trying to help set some vision for that?

作为领导者，你是否有意识地做些什么来帮助设计团队推动那个未来，或者让他们在探索中更具生成性、更敢于冒险？

Is there anything that you are doing intentionally as a leader to help the design org push on that future or be more generative or more risk-taking in what they're exploring?

我认为这是个好问题，可能我们需要更明确什么时候该做。我认为大问题就是找到合适的时机。有些情况下你想做，有些情况下你不想，而我想做的是找到更多时间，让你能在两者之间切换。这其实是我们团队正在努力解决的问题。比如，我应该花多少时间做那种工作，多少时间做执行？我认为你必须非常灵活。

I think it's a good question and probably something we need to be more clear about when to do it. I think that's the big question is finding when the right time is to do that. I think in some cases you want to do it, in other cases you don't, and I think what I'd like to do is find more time where you can kind of flip-flop between doing both. I think that's something we're kind of honestly trying to figure out as a team. Like how much time should I be spending doing that kind of work versus more execution, and I think you have to be able to be very fluid.

我们聊了很多。还有什么我们没谈到，但你认为能描绘出在 OpenAI 做设计的样子、你们的文化以及你们运作方式的？

We covered a lot of ground. What have we not talked about yet that you think paints a picture or shines an accurate light on what it looks like to design at OpenAI and the culture that you have and just the way that you all operate?

现实是节奏非常快。我们快速解决问题，快速更新自己的想法，必须随着技术和团队一起快速进化。所以有些人可能以为我们领先两年，但实际上我们紧跟着所有进步的步伐。这是一种非常不同的工作方式。你知道，脚下的事情整天都在变化，非常激动人心。说“我不知道，我们边走边解决，我们试试，我们转动曲柄，我们不断迭代，我们继续前进”真的很有趣。

The reality is that it is very fast-paced. We are figuring things out quickly. We kind of try to update our own thinking very quickly, and we have to evolve quickly with the technology and as a team. And so I think some people might assume we're like 2 years ahead thinking like that, but we're running very closely with where all of these advancements are going. And so that's just like a very different way of working. You know, things are changing underneath your feet all day long, and it's very exciting. It's really fun to be like, I don't know, we're going to figure this out as we go. We're going to try it. We're going to turn the crank. We're going to keep iterating. We're going to keep going.

我想追问一下你之前说的。你谈到在很多方面，成为一名优秀的设计师并没有改变。你是在解决问题。

I kind of want to push on something that you said earlier. You talked about how in many ways, being a great designer hasn't changed. You're solving problems.

但你谈到的很多事情，感觉更像是被动应对技术，而非用户行为。我想这可能会给设计实践带来一些麻烦，比如你在 Instagram 那样的地方可能会遇到的情况。

But yet a lot of the things that you are talking about do feel almost more reactive to the technology rather than user behaviors. And I got to imagine that that throws a little bit of a wrench into the design practice that maybe you would have at Instagram or something like that.

当然。我想说的是，我认为优秀的设计师可以两者兼顾。优秀的设计师既能说“哇，我们有这项新技术。好的，酷。我们怎么把它包装起来？”但另一方面，其他设计师或者我们设计师在其他时候会想，实际上它需要做到这个。它需要在这方面非常出色，或者这是它可能的工作方式。现在这是一种融合，老实说，你只需要再多考虑一件事。就像你不仅要考虑问题、用户问题等等，你还要考虑技术做不到但应该能做到的事情。所以你必须能够兼顾所有这些。所以我认为这基本上只是扩展了流程的一部分，我认为对设计师来说最好的流程是思考：我们如何推动它前进？如果它能做到 X、Y、Z，我们如何让它更有用？或者这里有一种理想的工作方式，然后我们可以与工程或研究部门合作，让它实现。

Sure. So what I should say, I think to make it clear, I think that great designers can do both. Great designers can both say, oh wow, we have this new technology. Okay, cool. How do we package that up? But then I think other designers or other times our designers are thinking about, well actually it needs to do this. It needs to be really good at this or here's how this could work. And it's just a blend now and you just have to think about one more thing, honestly. It's like you have to think not just about the problems, the user problems and all that, you have to think about what the technology can't do that it should be able to do. So you have to be able to do all that. So I think it's basically just extending one part of the process and I think the best process for designers is thinking like, well how do we push this forward? How do we make this be more useful if it could just do X, Y, and Z or here's kind of an ideal way this might work and then we can like work with engineering or research or whatever to get it there.

好的，那么对于正在收听、受到对话启发并想加入团队的人来说，你今年肯定会持续招聘。你在设计候选人身上寻找的主要信号是什么？你会怎么做来判断这些信号是否存在，以及他们是否是在这种环境中茁壮成长的人？

Okay, so for somebody listening who's inspired by the conversation, they want to join the team. You'll be hiring throughout this year, I'm sure. What are some of the main signals that you would be hunting for in a design candidate? And what would you be doing to figure out if they're present and if they're the type of person that would thrive in this environment?

我的意思是，我认为我们总是在寻找几种候选人类型。我总是对那些充满活力的后起之秀感到非常兴奋。我觉得这很有趣。我过去常说，当我两年前刚加入时，你不需要有 AI 背景就能来这里工作。你必须对技术充满好奇。这大体上仍然正确，但我认为现在已经有了足够的时间，我们可以找到那些已经开始实验和摆弄这些东西的人，他们理解它的优点、缺点、需要发展的方向，以及我们如何推动这些工具、技术和我们正在构建的产品。我认为有足够的东西可以玩。我一直尊重那些深入做某个副项目或对某个想法充满热情的人。你需要基本功。你需要擅长做一般的产品设计，但我认为另一部分就是真正的好奇心和对想法的兴趣，理想情况下你确实花时间玩过它，理解它，并且希望你对我们可以推动的方向有很多想法。我们处于一个幸运的位置，你不仅仅是在被动应对技术，而是在帮助塑造它的发展方向，并帮助扩展我们可以放入产品中的能力，让用户能够用它做更多事情。所以，那些对这种工作方式感到兴奋的人。

I mean, I think there's just a few candidate types that we always look for. I'm always very excited about the kind of more up-and-coming people that just have tons of energy. I think that it's funny. I used to say like when I first joined two and a half years ago, you don't need a background in AI to come work here. You have to be curious about the technology. And that's still generally true, but I think that there's enough now there's been enough time where finding people that have started to experiment and play with this stuff and understand what it's good at, what it's bad at, where it needs to go and how we need to push on these tools and the technology and the products that we're all building. I think there's like enough there to play with. I've always respected people that will go deep on some side project or get really passionate about some idea. You need the fundamentals. You need to be great at just kind of like doing general product design, but I think that that other piece just like being truly curious and interested in the idea and ideally you've really spent time playing with it and understand and hopefully you have lots of ideas for where we can push things. We're at this fortunate place where you're not just reacting to the technology, but you're hopefully helping shape where this is going and helping to expand the capabilities that we can put into the product into people's hands so that they can do more with it. And so people that are excited about that kind of way of working.

好的，Ian，感谢你今天来和我们分享这些。你们对设计的状态产生了巨大影响，甚至互动方式也是如此，就像创造了这些我们都在其基础上构建的范式，所以听到一些幕后故事以及你们今天如何运作，非常有趣。

Well, Ian, I appreciate you coming on and sharing this with us today. You all have had a massive impact on state of design and even interacting like just creating these paradigms that we're all building on top of and so it's been fun to hear a little bit of the behind the scenes and how you all operate today.

是的，这非常有趣。谢谢你邀请我。

Yeah, this is super fun. Thanks for having me.