从 Dropbox 到 ChatGPT：OpenAI 产品革命内幕

你曾是 Dropbox 的产品负责人，后来去了 Instacart。现在你是史上最具影响力产品的产品经理。

You were a product leader at Dropbox, then Instacart. Now you're the PM of the most consequential product in history.

我当时不知道在这里会做什么。那是个研究实验室。第一个任务就像修百叶窗之类的。

I didn't know what I would do here. It was a research lab. The first task was like fix the blinds or something like that.

当有人给你一艘火箭飞船时，别问坐哪个座位。我们着手构建一个超级助手。它本该是一个黑客松的代码库。

When someone offers you a rocket ship, don't ask which seat. We set out to build a super assistant. It was supposed to be a hackathon codebase.

它之前叫什么？

What was it called before?

它本来要叫“与 GBD3.5 聊天”，因为我们真的不认为它会成为成功的产品。

It was going to be chat with GBD3.5 because we really didn't think it was going to be a successful product.

然后 Sam Altman 就说：“嘿，让我发条推文。”

And then Sam Alman's just like, "Hey, let me tweet about it."

这是 AI 的一个模式。你只有在发布之后才知道该打磨什么。我的梦想就在那里。我们每天都在发布。等到人们听到这个的时候，他们就能用上 GPT-5 了。

This is a pattern with AI. You won't know what to polish until after you ship. My dream is there. We ship daily. By the time people hear this, they're going to have their hands on GPT5.

大约全球 10% 的人口每周都在使用。规模带来了责任。它感觉更鲜活一点，更有人性。这个模型有品味。

About 10% of the world population uses every week. With scale comes responsibility. It just feels a little more alive, a bit more human. This model has taste.

你们的 CPO Kevin Weil 说，让我问问你关于“是否最大化加速”这个原则。

Kevin Wheel, your CPO, said to ask you about this principle of is it maximally accelerated?

我真的想直接跳到重点。为什么我们现在不能做这个？我一直觉得我在这里的部分职责就是设定节奏和静息心率。

I just really want to jump to the punch line. Why can't we do this now? I always felt like part of my role here to just set the pace and the resting heartbeat.

大家总是想知道，聊天是这一切的未来吗？

Everyone's always wondering, is chat the future of all of this stuff?

聊天是当时最简单的发布方式。我对它如此火爆感到困惑。我更困惑的是有多少人抄袭了。

Chat was the simplest way to ship at the time. I'm baffled by how much it took off. I'm even more baffled by how many people have copied.

ChatGPT 现在给我的 newsletter 带来的流量比 Twitter 还多。

Chat GPT is now driving more traffic to my newsletter than Twitter.

这是那种极具留存力的能力。我对我们在搜索方面做的事情感到非常兴奋。

That is the type of capability that has been incredibly retentive. I've been really excited about what we've been doing in search.

你能让我们一窥长期的发展方向吗？

Can you give us a peek into where this goes long term?

ChatGPT 感觉有点像 MS-DOS。我们还没有构建 Windows，一旦构建了，一切就会变得显而易见。

Chatbt feels a little bit like MS DOS. We haven't built Windows yet and it will be obvious once we do.

今天的嘉宾是 Nick Turley。Nick 是 OpenAI 的 ChatGPT 负责人。他三年前加入公司，当时公司主要还是研究实验室。他帮助提出了 ChatGPT 的想法，并将其从零发展到超过 7 亿周活跃用户、数十亿美元营收，可以说是人类历史上最成功、最具影响力的消费软件产品。Nick 非常了不起。他一直很低调。这是他第一次接受大型播客采访，你们有福了。我们谈了很多事情，包括刚刚发布的 GPT-5。非常感谢 Kevin Weil、Claire Vo、George O'Brien、Joanne Jen 和 Peter Ding 为这次对话推荐话题。如果你喜欢这个播客，别忘了在你最喜欢的播客应用或 YouTube 上订阅和关注。如果你成为我 newsletter 的年度订阅者，你将免费获得一年一大堆很棒的产品，包括 Lovable、Replit、Bolt、Nadant、Linear、Superhuman、Descript、Whisper Flow、Gamma、Perplexity、Warp、Granola、Magic Patterns、Raycast、Chapar D 和 Mobin。请访问 lenny'snewsletter.com 并点击 bundle。就这样，我请出 Nick Turley。本期节目由 Orcus 赞助，Orcus 是开源 Conductor 背后的公司，该编排平台为现代企业应用和智能体式工作流提供动力。传统自动化工具跟不上节奏。孤立的低代码平台、过时的流程管理和断开的 API 工具在今天事件驱动、AI 驱动的智能体式环境中显得不足。Orcus 改变了这一点。借助 Orcus Conductor，你获得一个智能体式编排层，可以在企业规模下实时无缝连接人类、AI 智能体、API、微服务和数据管道。可视化和代码优先开发、内置合规性、可观测性和坚如磐石的可靠性确保工作流随着你的需求动态演进。这不仅仅是自动化任务。它是在编排自主智能体和复杂工作流，以更快地交付更智能的结果。无论是现代化遗留系统还是扩展下一代 AI 驱动的应用，Orcus 都能加速你从想法到生产的旅程。了解更多并开始构建，请访问 orcus.io/lenny。那就是 oke kes.io/lenny。本期节目由 Vanta 赞助，我非常高兴邀请到 Vanta 的 CEO 兼联合创始人 Christina Casiopo 参加这个非常简短的对话。

Today my guest is Nick Turley. Nick is head of chatbt at OpenAI. He joined the company 3 years ago when it was still primarily a research lab. He helped come up with the idea of chat GPT and took it from zero to over 700 million weekly active users, billions in revenue, and arguably the most successful and impactful consumer software product in human history. Nick is incredible. He's been very much under the radar. This is the first major podcast interview that he has ever done and you are in for a treat. We talk about all the things including the just launched GPT5. A huge thank you to Kevin Wheel, Claire Vo, George O'Brien, Joanne Jen, and Peter Ding for suggesting topics for this conversation. If you enjoy this podcast, don't forget to subscribe and follow it in your favorite podcasting app or YouTube. And if you become an annual subscriber of my newsletter, you get a year free of a bunch of incredible products including Lovable, Replet, Bolt, Nadant, Linear, Superhum, Dcript, Whisper Flow, Gamma, Perplexity, Warp, Granola, Magic Patterns, Raycast, Chapar D, and Mobin. Check it out at lenny'snewsletter.com and click bundle. With that, I bring you Nick Turley. This episode is brought to you by Orcus, the company behind Open-source Conductor, the orchestration platform powering modern enterprise apps and agentic workflows. Legacy automation tools can't keep pace. Siloed low code platforms, outdated process management, and disconnected API tooling fall short in today's event-driven AI powered agentic landscape. Orcus changes this. With Orcus Conductor, you gain an agentic orchestration layer that seamlessly connects humans, AI agents, APIs, microservices, and data pipelines in real time at enterprise scale. Visual and codeforce development, built-in compliance, observability, and rock solid reliability ensure workflows evolve dynamically with your needs. It's not just about automating tasks. It's orchestrating autonomous agents and complex workflows to deliver smarter outcomes faster. Whether modernizing legacy systems or scaling nextgen AIdriven apps, Orcus accelerates your journey from idea to production. Learn more and start building at orcus.io/lenny. That's oke kes.io/lenny. This episode is brought to you by Vanta and I am very excited to have Christina Casiopo, CEO and co-founder of Vanta, joining me for this very short conversation.

很高兴来到这里。我是这个播客和 newsletter 的忠实粉丝。Vanta 是这个节目的长期赞助商，但对于一些新听众来说，Vanta 是做什么的，面向谁？

Great to be here. Big fan of the podcast and the newsletter. Vanta is a longtime sponsor of the show, but for some of our newer listeners, what does Vanta do and who is it for?

当然。所以，我们在 2018 年创办了 Vanta，专注于创始人，帮助他们开始构建安全计划，并通过 SOC 2 或 ISO 27001 等合规认证来获得所有艰苦安全工作的认可。如今，我们帮助超过 9,000 家公司，包括一些初创公司家喻户晓的名字，如 Atlassian、Ramp 和 LangChain，启动和扩展他们的安全计划，最终通过自动化合规、集中 GRC 和加速安全审查来建立信任。

Sure. So, we started Vanta in 2018 focused on founders, helping them start to build out their security programs and get credit for all of that hard security work with compliance certifications like SOCK 2 or ISO 2701. Today, we currently help over 9,000 companies, including some startup household names like Atlassian, Ramp, and Lang Chain, start and scale their security programs, and ultimately build trust by automating compliance, centralizing GRC, and accelerating security reviews.

太棒了。我从经验中知道这些事情需要大量时间和资源，没有人愿意花时间做这个。

That is awesome. I know from experience that these things take a lot of time and a lot of resources, and nobody wants to spend time doing this.

这非常符合我们的经验，无论是在公司成立之前还是在某种程度上成立期间。但我们的想法是，通过自动化、AI 和软件，我们帮助客户以高效的方式与潜在客户和客户建立信任。你知道我们的玩笑，我们创办这家合规公司是为了让你不必这样做。

That is very much our experience, but before the company and to some extent during it. But the idea is with automation, with AI, with software, we are helping customers build trust with prospects and customers in an efficient way. And you know our joke, we started this compliance company so you don't have to.

我们感谢你这样做。而且你为听众提供了特别折扣。他们可以在 vanta.com/lenny 获得 Vanta 的 1000 美元折扣。那就是 venta.com/lenny，立减 1000 美元。谢谢你，Christina。

We appreciate you for doing that. And you have a special discount for listeners. They can get $1,000 off Vanta at vanta.com/lenny. That's venta.com/lenny for $1,000 off. Thanks for that, Christina.

谢谢。

Thank you.

Nick，非常感谢你参加我的节目，欢迎来到播客。

Nick, thank you so much for joining me and welcome to the podcast.

谢谢你邀请我，Lenny。

Thanks for having me, Lenny.

我已经有无数个问题想问你了，然后你们决定在我们录制这一周发布 GPT-5。所以，现在我至少有无数个问题要问你。我希望你有很多时间。首先，恭喜发布。它明天就要来了，就在我们录制后的第二天。恭喜。你感觉怎么样？我猜这是巨大的工作和压力。你还好吗？

I already had a billion questions I wanted to ask you and then you guys decided to launch DPT5 the week that we're recording this. So, now I have at least two billion questions for you. I hope you have I hope you have a lot of time. First of all, just congrats on the launch. It's coming tomorrow, the day after we're recording this. Just uh congrats. How you feeling? I imagine this is an ungodly amount of work and stress. How you doing?

这是忙碌的一周，但你知道，我们已经为此工作了一段时间。所以，把它发布出来感觉真的很好。

It's a busy week, but you know, we we've been working on this for a while. So, it also feels really good to get it out.

所以，等到人们听到这个的时候，他们就能用上 GPT-5 和最新的 ChatGPT。最简单的理解方式是什么，它解锁了什么，人们能用它做什么。给我们一个大概的介绍。

So, by the time people hear this, they're going to have their hands on GPT5 and the newest Chat GPT. What's the simplest way to just understand what this is, what it unlocks, what people can do with it. Give us kind of the the pitch.

我对 GPT-5 感到非常兴奋。我认为对大多数人来说，它会感觉像是一个真正的阶跃变化。

I'm so excited about GPD5. It uh I think for most people is going to feel like a real step change.

如果你是普通的 ChatGPT 用户，我们这周有 7 亿用户，你可能已经用 GPT-4 一段时间了，甚至不会去想驱动产品的模型是什么。而 GPT-5 给人的感觉是完全不同的。我会讲很多细节，但归根结底，体验感很好。至少我们这么觉得，也希望用户有同感。越来越多的人注意到的是这一点。他们不看学术基准，也不看评估，而是直接试用模型，感受它的表现。仅就这一点，我就非常兴奋。我已经用了一段时间了。但它也是我们发布过的最聪明、最有用、最快的前沿模型。就纯智能而言，一种衡量方式是学术基准。在许多标准基准上，无论是数学、推理还是原始智能，这个模型都是最先进的。我尤其对它在编程上的表现感到兴奋，无论是 SWE-bench 这样的常见基准，还是实际的前端编码，都非常出色。我觉得 GPT-5 在这方面实现了真正的阶跃式改进。但无论你怎么衡量智能，它都非常了不起，我想人们会感受到这种升级，尤其是那些还没用过 03 的人。

If you're the average ChatGPT user and we have, you know, 700 million of them this week, you've probably been on GPT-4 for a while. You probably don't even think about the model that powers the product. And GPT-5 is just categorically different. I'll talk about a lot of the specifics, but at the end of the day, the vibes are good. At least we feel that way; we hope users feel the same. Increasingly, that's what most people notice. They don't look at academic benchmarks or evaluations. They try the model and see what it feels like. On that dimension alone, I'm so excited. I've been using it for a while. But it is also the smartest, most useful, and fastest frontier model we've ever launched. On pure smarts, one way to look at that is academic benchmarks. On many standard ones, whether it's math, reasoning, or raw intelligence, this model is state-of-the-art. I'm especially excited about its coding performance, whether it's SWE-bench, a common benchmark, or actual front-end coding, which is really good. That's an area where I feel there's a true step change in GPT-5. But no matter how you measure the smarts, it's quite remarkable, and I think people will feel the upgrade, especially if they weren't using 03 already.

除了智能，第二点是它真的非常有用。编程是实用性的一个方面，无论你是问编程问题，还是在 vibe coding 一个应用。但它也是一个很好的写作者。我以写作为生，对内对外都是。我刚刚写了一篇大博客文章，周一发布了。这个东西是个不可思议的编辑。相比一些旧模型，它有品味，我觉得这非常令人兴奋。对我来说，这在我的日常工作中非常有用。还有其他领域，比如它在健康方面是最先进的，需要时很有用。但同样，你无法真正用用例或数据表达的是模型的“感觉”。它感觉更生动一点，更人性化一点，这种感受只有试过才能体会。所以我对这个很满意。还有，如前所述，它更快。它像 03 一样会思考，但你不需要手动告诉它。它会动态决定何时需要思考，不需要时就立即响应。这感觉比用 03 快得多。也许最令人兴奋的是我们让它免费可用。这是我们在 OpenAI 能独特做到的，因为很多有订阅模式的公司会把它放在付费计划后面。对我们来说，如果我们能扩展规模，我们就会做。这感觉太棒了。我们之前对 GPT-4 也这么做了。所以每个人明天都能试用 GPT-5，希望如此。

The second thing beyond smarts is that it's just really useful. Coding is one axis of utility, whether you have coding questions or you're vibe coding an app. But it's also a really good writer. I write for a living, internally and externally. I just wrote a big blog post that we published Monday. This thing is such an incredible editor. Compared to some older models, it's got taste, which I think is really exciting. To me, that's something truly useful in my day-to-day. There are other areas too, like it's state-of-the-art on health, which is useful when you need it. But again, the thing you can't really express in use cases or data is the vibe of the model. It just feels a little more alive, a bit more human, in a way that's hard to articulate until you try it. So I feel good about that. And as mentioned, it's faster. It thinks just like 03 did, but you don't have to manually tell it to do that. It'll dynamically decide to think when it needs to, and when it doesn't, it responds instantly. That ends up feeling quite a bit faster than using 03 did. And maybe the most exciting thing is that we're making it available for free. That's something we can uniquely do at OpenAI because many companies with a subscription model would gate it behind a paid plan. For us, if we can scale it, we will. That feels awesome. We did that with GPT-4 as well. So everyone will be able to try GPT-5 tomorrow, hopefully.

做这样一个东西需要多长时间？我不知道有没有简单的答案，但你们做 GPT-5 做了多久？

How long does something like this take? I don't know if there's a simple answer to this, but just how long have you guys been working on GPT-5?

我们已经做了一段时间了。你可以把 GPT-5 看作是多种努力的结晶。我们有推理技术，也有更经典的后训练方法，所以很难说它从什么时候开始。但它确实是我们一段时间以来开发的多种技术的终点。

We've been working on it for a while. You can view GPT-5 as a culmination of a bunch of different efforts. We had the reasoning tech, we had more classic post-training methodologies, so it's really hard to put a beginning on it. But it really is the end point of a bunch of different techniques we've been developing for a while.

你能透露一下 ChatGPT 未来的愿景吗？GPT 整体上，从表面看，很长时间以来都是同样的想法，只是大脑更聪明了。我很好奇长期来看它会走向何方。

Can you give us a peek into the vision for where ChatGPT is going? GPT in general is going, like if you look at the surface, it's been kind of the same idea with a much smarter brain for a long time. I'm curious where this goes long term.

稍微回顾一下，把 ChatGPT 看作一个无处不在的产品。全球大约 10% 的人口每周都在使用它。我们现在有大约 500 万企业客户。它本身已经是一个成熟的品类。但真正开始时，我们目标是构建一个超级助手。当时我们就是这么说的。事实上，我们使用的代码库叫 SA server。它本来是一个黑客松的代码库，但事情的发展有点不同。所以从某些方面看，这仍然是我们的愿景。我不多谈它的原因是我觉得“助手”这个词在我们试图创造的思维模型上有点局限。你会想到一个非常拟人化的东西，可能很实用。但坦白说，对大多数人来说，拥有一个助手并不特别有共鸣，除非你在硅谷，是经理之类的。所以这个说法不完美，但我们真正设想的是一个能帮你完成任何任务的实体，无论是在家、工作还是学校，任何情境。它是一个知道你目标的实体。不像现在的 ChatGPT，你不需要详细描述你的问题，因为它已经理解你的总体目标，并了解你的生活背景。这是我们非常兴奋的一点。给它更多生活输入的另一方面是给它更多的行动空间。我们非常兴奋地让它随着时间的推移，能像一个聪明、有同理心的人用电脑为你做事一样。一旦你给它这样的工具，你能为人们解决的问题类型将和今天在聊天机器人里做的非常不同。这是更多的输出。我经常想，如果我是一个通用智能，我成了 Lenny 的实习生会怎样？我可能不会特别有效，尽管我有刚才提到的两个属性。这是因为与这项技术建立关系也非常重要。这可能是让我兴奋的第三点：构建一个能随着时间真正了解你的产品。你看到我们今年早些时候推出了改进的记忆功能。那只是开始。所以它真的感觉像你的 AI。

To back up a bit, think of ChatGPT as this ubiquitous product. About 10% of the world population uses it every week. We have like five million business customers now. It's an established category in its own right. But really, when we started, we set out to build a super assistant. That's how we talked about it at the time. In fact, the codebase we use is called SA server. It was supposed to be a hackathon codebase, but things turned out a little differently. So in some ways, that is still the vision. The reason I don't talk about it more is because I think 'assistant' is a bit limiting in terms of the mental model we're trying to create. You think of this very personified human thing, maybe utilitarian. And frankly, having an assistant is not particularly relatable to most people unless they're in Silicon Valley and they're a manager or something. So it's imperfect, but really what we envision is an entity that can help you with any task, whether at home, work, or school, in any context. It's an entity that knows what you're trying to achieve. Unlike ChatGPT today, you don't have to describe your problem in minute detail because it already understands your overarching goals and has context on your life. That's one thing we're really excited about. The inverse of giving it more inputs on your life is giving it more action space. We're really excited to allow it, over time, to do what a smart empathetic human with a computer could do for you. The limit of the types of problems you can solve for people once you give it access to tools like that is very different than what you might do in a chatbot today. That's more outputs. I often think, okay, I'm a general intelligence, what would happen if I became Lenny's intern? I wouldn't be particularly effective despite having both of those attributes. It's because building a relationship with this technology is also incredibly important. That's maybe the third piece I'm excited about: building a product that can truly get to know you over time. You saw us launch some of those things with improved memory earlier this year. That's just the beginning. So it really feels like your AI.

所以，我不确定超级助手是否还是最贴切的类比，但我觉得人们就把它当作他们的 AI。我认为我们可以把这样一个 AI 放进每个人的口袋，帮他们解决实际问题。无论是保持健康、创业，还是对任何事情有个第二意见。日常生活中有太多不同的问题可以帮到人们，这正是激励我的地方。

So, I don't know if Super Assistant is still the right exact analogy, but I think people just think of it as their AI. And I think we can put one in everyone's pocket and help them solve real problems. Whether or not that's becoming healthy, whether or not that's starting a business, whether or not that's just having a second opinion on anything. There's so many different problems that you can help with people in their daily life. And that's what motivates me.

所以，我在这里读到的弦外之音是，愿景是让它成为人们的助手，而不是取代人类。这感觉像是拼图中非常重要的一块。也许就谈谈这个吧。

So, an interesting kind of between the lines that I'm reading here is the vision is for it to be an assistant for people, not to replace people. It feels like a really important piece of the puzzle. Maybe just talk about that.

AI 确实让人害怕。我理解，你知道，几十年的电影已经给人们植入了一种特定的心智模型。即使只看今天的技术，我想每个人都会有那么一刻，AI 做了一件对他们来说非常私人的事情，然后你会想，嘿，AI 永远做不到这个。对我来说，那是一些奇怪的音乐理论问题，我当时想，哇，这东西对音乐的理解竟然比我还深，而那是我热爱的东西。所以这自然让人害怕。我认为对我们来说，长久以来真正重要的事情是构建一些让你觉得有帮助、但由你掌控的东西。随着这些东西变得智能体式，这一点就更加重要了，对吧？那种掌控感。这可以是小事，比如我们构建了一种在智能体模式下观察 AI 正在做什么的方式。并不是说你真的会一直盯着它看，但它给了你一个心智模型，让你感到掌控，就像你在 Waymo 里那样，用过 Waymo 的人都知道，你会看到一个屏幕，可以看到其他车辆。你并不是真的会一直看，但它让你感觉你知道这东西是如何运作的、正在发生什么。或者我们总是跟你确认事情。这有点烦人，但它让你坐在驾驶座上，这很重要。因此，我们总是把技术以及我们构建的技术视为放大你能力的东西，而不是取代它。随着牌局越来越强大，这一点变得很重要。

AI is really scary to people. And I understand, you know, there's decades of movies on AI that have a certain mental model kind of baked in. And even if you just look at the technology today, once everyone I think has this moment where the AI does something that was really deeply personal to them and you're like kind of thought, hey, the AI can never do that. For me, it was like weird music theory things where I was like, wow, this thing actually understands music better than I do and that's like something I'm passionate about. So it's naturally scary. And I think the thing that's been really important to us for a long time is to build something that feels like it's helpful to you, but you're in the driver's seat. And that's even more important as the stuff becomes agentic, right? Like the feeling of being in control. And that can be small things like, you know, we built this way of sort of watching what the AI is doing when it's in agent mode. And it's not that you actually are going to watch it the whole time, but it gives you a mental model and makes you feel in control in the same way that when you're in a Waymo, you get that screen for those of you who have tried Waymo. You know, you can see the other cars. It's not like you're going to actually watch, but it gives you the sense that you know how this thing works and what's happening. Or we always check with you to confirm things. It's a little bit annoying, but it puts you in the driver's seat, which is important. And for that reason, we always view technology and the technology that we build as something that amplifies what you're capable of rather than replacing it. And that becomes important as the deck gets more powerful.

好的。你提到了 ChatGPT 的起步。我在另一个采访中读到过。你加入 OpenAI 时，ChatGPT 基本上只是一个内部实验项目，用来测试 GPT-3.5。然后山姆·奥特曼就说：“嘿，让我发个推文，看看人们是否觉得这有趣。”等等等等。我认为它是历史上最成功的消费产品，无论是增长速度、用户数还是收入，都简直离谱。你能让我们一窥那个早期阶段吗？在它成为人人痴迷的东西之前。

Okay. So you mentioned the beginnings of ChatGPT. I was reading in a different interview. So you joined OpenAI. ChatGPT was kind of just this internal experimental project that was basically a way to test GPT-3.5. And then Sam Altman's just like, "Hey, let me tweet about it. Maybe see if people find this interesting." yada yada yada. It's the most successful consumer product in history, I think, both in growth rate and users and revenue and just absurd. Can you give us a glimpse into that early period before it became something everyone's obsessed with?

是的。所以我们决定要做一些面向消费者的东西，我想大概是在 GPT-4 完成训练的时候，实际上主要出于几个原因。我们当时已经有一个产品，就是开发者产品。那其实是我最初来帮忙做的事情，而且它对使命来说一直非常棒。事实上，它已经成长起来，现在是一个开放平台，大概有 400 万开发者吧。但当时它还处于早期阶段，我们遇到了一些限制，因为有两个问题。第一，你不能很快地迭代，因为每次你改变模型，就会破坏每个人的应用。所以很难尝试新东西。第二，很难学习，因为我们得到的反馈是最终用户到开发者再到我们，所以非常间接。我们渴望快速向 AGI 迈进，感觉我们需要与消费者建立更直接的关系。所以我们试图找出从哪里开始，按照 OpenAI 的经典风格，尤其是在当时，我们组织了一个黑客松，让爱好者们一起捣鼓 GPT-4，看看我们能创造出什么很棒的东西，也许能发布给用户。每个人的想法都是某种超级助手。比如更具体的想法，我们有一个会议机器人，可以拨入会议，愿景是它最终能帮你主持会议。我们还有一个编码工具，现在回头看，可能超前了。挑战在于，当我们测试这些更定制化的想法时，每次人们都想用它来做其他事情，因为它是一项非常非常通用的强大技术。所以经过几个月的原型制作，我们召集了同样的志愿者团队，这真的是一个志愿者团队，对吧？我们有来自超级计算团队的人，他之前开发过 iOS 应用。我们有研究团队的人，他这辈子写过一些后端代码。他们都是这个初始 ChatGPT 团队的一部分，我们决定发布一些开放性的东西，因为我们只是想要一个真实的使用案例分布。我认为这是 AI 的一个模式，你真的必须发布才能理解什么是可能的、人们想要什么，而不是先验地推理。所以 ChatGPT 最终就这样诞生了，因为我们只是想尽快获得学习，我们在假期前发布了它，想着我们会回来获取数据，然后慢慢缩减。显然，那部分结果完全不同，因为人们真的很喜欢这个产品。所以我记得当时经历了一些过程，比如“哦，天哪，仪表盘坏了。等等，人们喜欢它。我确定这只是病毒式传播，热度会消退的”，然后变成“哦，哇，人们在留存，但我不明白为什么”。最后我们逐渐进入了产品开发模式，但这有点偶然。

Yeah. So we had decided that we wanted to do something consumer-facing, I think right around the time that GPT-4 finished training, and it was actually mainly for a couple reasons. We already had a product out there which was our developer product. That's actually what I came in to help with initially, and that has been amazing for the mission. In fact, it's grown up and now it's the open platform with, I don't know, 4 million developers I think. But at the time it was early stage, and we were running into some constraints with it because there were two problems. One, you couldn't iterate very quickly because every time you would change the model you would break everyone's app. So it was really hard to try things. And then the other thing was that it was really hard to learn because the feedback we would get was like the feedback from the end user to the developer to us. So it was very disintermediated, and we were excited to make fast progress toward AGI, and it just felt like we needed a more direct relationship with consumers. So we were trying to figure out where to start, and in classic OpenAI fashion, especially back then, we put together a hackathon of enthusiasts just hacking on GPT-4 to kind of see what awesome stuff we could create and maybe ship to users. And everyone's idea was some flavor of a super assistant. Like they were more specific ideas, like we had a meeting bot that would call into meetings, and the vision was maybe it will help you run the meeting over time. And we had a coding tool which, full circle now, probably ahead of its time. And the challenge was that when we tested those things, every time we tested these more bespoke ideas, people wanted to use it for all this other stuff because it's just a very, very generically powerful technology. So after a couple months of prototyping, we took that same kind of crew of volunteers, and it was truly a volunteer group, right? We had someone from the supercomputing team who had built an iOS app before. We had someone on the research team who had written some backend code in their life. They were all part of this initial ChatGPT team, and we decided to ship something open-ended because we just wanted a real use case distribution. And this is a pattern with AI, I think, where you really have to ship to understand what is even possible and what people want, rather than being able to reason about that a priori. So ChatGPT came together at the end because we just wanted the learnings as soon as we could, and we shipped it right before the holiday thinking we would sort of come back and get the data and then wind it down. And obviously that part turned out super differently because people really liked the product as is. So I remember sort of going through the motions of like, oh man, dashboard's broken. Oh wait, people are liking it. I'm sure it's just going viral and stuff is going to die down to like, oh wow, people are retaining, but I don't understand why. And then eventually we kind of fell into product development mode, but it was a little bit by accident.

哇。我不知道 ChatGPT 是从一个黑客松项目中诞生的。绝对是最成功的黑客松项目。我喜欢在谈论我们的黑客松时讲这个故事，因为我真的希望人们觉得他们可以发布他们的想法，过去确实如此，我们也会继续让这成为现实。

Wow. I did not know that ChatGPT emerged out of a hackathon project. Definitely the most successful hackathon project. I like to tell this story when we talk about our hackathons because I really do want people to feel like they can ship their idea, and it's certainly been true in the past and we'll continue to make it true.

也许你不想分享这些，但我很好奇那个团队是谁。

Maybe you don't want to share these things but I wonder who that team was.

团队大部分成员还在。

The team's largely still around.

实际上，参与 GPT-5 研发的一些研究人员，他们一直都是 ChatGPT 团队的成员。工程师还在，设计师也还在。我想我也还在。所以，团队还在继续运营，但显然我们已经大幅扩张，而且不得不这样做，因为随着规模扩大，责任也随之而来。我们很快就要达到 10 亿用户，你必须开始以与这种规模相匹配的方式行事。

Some of the researchers working on GPT-5, actually, you know, they were always part of the ChatGPT team. Engineers are still around, designers are still around. I'm still here, I guess. So, you got the team still running things, but obviously we've grown up tremendously, and we've had to, because with scale comes responsibility. We're going to hit a billion users soon, and you kind of have to begin acting in a way that is appropriate to that scale.

好的。那我想在这里花点时间。我不确定这是否 100% 准确，但我相信 ChatGPT 是历史上增长最快、最成功的消费产品。也是对人们生活影响最大的。感觉它现在已经成为社会的一部分了。我妻子会和它聊天。我每个问题都会去问它。语音模式。我妻子会说：“让我问问 ChatGPT。”它现在已经成为我们生活的一部分了。而且我觉得这还只是开始。很多人甚至都不知道到底发生了什么。作为领导者，你有没有停下来反思过，觉得“天哪”？

Okay. So, let me spend a little time there. I don't know if this is 100% true, but I believe it is that ChatGPT is the fastest growing, most successful consumer product in history. Also the most impactful on people's lives. It feels like it's just part of the ether of society now. My wife talks to it. Every question I have, I go to it. Voice mode. My wife's just like, "Let me check with ChatGPT." It's just such a part of our life now. And I think it's still early. So many people don't even know what the hell is going on. As someone leading this, do you ever just take a moment to reflect and think about just like holy cow?

能运营这样的产品，真的让人感到谦卑，我经常要掐自己一下。我也需要偶尔退后一步，好好思考，这在事情发展如此之快的时候真的很难。我喜欢在公司设定快节奏，但为了有自信地这样做，我每周至少需要一天完全断开联系，只思考该做什么，处理一周的事情等等。另外，我从未做过如此依赖经验的产品，如果你不停下来观察和倾听人们的行为，你会错过很多，无论是实用性还是风险方面。通常，在产品发布时，你知道它会做什么。你不知道人们是否会喜欢——那总是经验性的——但你知道它能做什么。而 AI 因为很多是涌现出来的，你真的需要在发布后停下来倾听，然后根据人们尝试做的事情和尚未奏效的事情进行迭代。仅就这一点而言，我认为停下来观察正在发生的事情非常重要。

It's quite humbling to get to run a product like that, and I have to pinch myself very frequently. I also have to sometimes sit back and just think, which is really hard when things are moving so quickly. I love setting a fast pace at the company, but in order to do that with confidence, I need at least one day every week that I'm entirely unplugged, just thinking about what to do and processing the week, etc. The other thing is I've never worked on a product that is so empirical in its nature, where if you don't stop and watch and listen to what people are doing, you're going to miss so much, both on the utility and on the risks. Normally, by the time you ship a product, you know what it's going to do. You don't know if people are going to like it—that's always empirical—but you know what it can do. With AI, because so much of it is emergent, you really need to stop and listen after you launch something, and then iterate on the things people are trying to do and on the things that aren't quite working yet. For that reason alone, I think it's very important to take a break and just watch what's going on.

好的。所以你每周休息一天。不是休息——这么说不对。你是花一天时间思考，深度工作。

Okay. So you take a day off every week. Not off—that's not the right way to put it. You take a day of thinking time, deep work.

我需要这样。是的。我需要在周六或类似时间彻底断开连接。

I need it. Yeah. And I need to hard unplug on a Saturday or something like that.

在周六，比如接下来……

On a Saturday like the next...

但否则是不可能的。这已经是三年的马拉松了。

But it's just not possible otherwise. This has been a giant marathon for three years now.

就像冲刺马拉松。

Like a sprint marathon.

冲刺马拉松，没错，或者间歇训练之类的。我不知道该如何准确描述 OpenAI 的发布节奏，但你必须以一种可持续的方式设定自己，即使这不是 AI，没有我刚才提到的那些有趣特性。我认为你也需要这样做，但尤其是在 AI 领域，去观察是很重要的。

Sprint marathon, that's right, or interval training or something. I don't know how to exactly describe the OpenAI launch cadence, but you got to set yourself up in a way that is sustainable, even if this wasn't AI and it didn't have the interesting attributes that I just mentioned. I think you would need to do that, but especially with AI, it's important to go watch.

沿着这个思路，我和你在 OpenAI 的很多同事聊过。Joanne 特别提到，紧迫感和节奏是你工作方式的重要组成部分，你觉得即使在你是历史上增长最快的产品、疯狂增长的时候，不断在团队中创造紧迫感也非常重要。谈谈你对团队中节奏和紧迫感重要性的理念。

So along those lines, I talked to a bunch of people that work with you at OpenAI. Joanne specifically said that urgency and pace are a big part of how you operate, that you find it really important to create urgency within the team constantly, even when you are the fastest growing product in history, growing like crazy. Talk about your philosophy on the importance of pace and urgency on teams.

她这么说真是太好了。我在 ChatGPT 上花了很多时间做两件事。当我们决定做的时候，我们已经原型设计了很久，我当时就说：“我们 10 天后就发布这个东西。”我们做到了。所以那可能是一个特定时刻的事情，我只是真的想确保我们去学到一些东西。但从那以后，我花了很多时间思考 ChatGPT 最初为什么会成功。我认为其中有一些因素是我们做了很多其他公司有 LLM 技术却从未发布的事情。我只是觉得在所有我们可以优化的东西中，尽可能快地学习是极其重要的。所以我开始围绕这一点召集大家，这采取了不同的形式。有一段时间，当我们还是那个规模时，我主持每日发布同步会议，所有需要做决定的人都参加，我们只是讨论该做什么，从昨天开始调整方向等等。显然在某个时刻这无法扩展，但我一直觉得我在这里的部分角色是思考产品方向，同时也为我们的团队设定节奏和静息心跳。再说一次，这在任何地方都很重要，但尤其重要的是，当找出人们喜欢什么和什么有价值的唯一方法是将它带到外部世界时。因此，我认为这已经成为 OpenAI 的超能力，我很高兴 Joanne 认为我在这方面起了作用，但这真的需要大家共同努力。

Well, it's nice of her to say that. I spent a lot of time on two things with ChatGPT. When we decided to do it, we had been prototyping for so long, and I was just like, "In 10 days we're going to ship this thing," and we did. So that was maybe a moment in time thing where I just really wanted to make sure that we go learn something. But ever since then, I've spent so much time thinking about why ChatGPT became successful in the first place. I think there was some element of doing things where there were many other companies that had technology in the LLM space that just never got shipped. I just felt like of all the things we could optimize for, learning as fast as possible is incredibly important. So I started rallying people around that, and that took different forms. For a while, when we were of that size, I ran this daily release sync, and it had everyone who was required to make a decision in it, and we would just talk about what to do and pivot from yesterday, etc. Obviously at some point that doesn't scale, but I always felt like part of my role here was to think about the direction of the product, but also to set the pace and the resting heartbeat for our teams. Again, this is important anywhere, but it's especially important when the only way to find out what people like and what's valuable is to bring it into the external world. For that reason, I think it's become a superpower of OpenAI, and I'm glad that Joanne thinks I had some part in that, but it really has taken a village.

我喜欢这个说法，团队的静息心率。这是一个完美的比喻，把节奏等同于静息心率。

I love this phrase, the resting heart rate of your team. That's such a perfect metaphor for just the pace being equivalent to your resting heart rate.

我实际上是在 Instacart 学到的，当时我到了那里，因为疫情，有一段时间是全员参与。我想我们解散了所有团队，所以有一个全公司的站会。

I actually learned that at Instacart when I showed up there, because we were in the pandemic and it was kind of all hands on deck for a while. There was this company-wide standup, I think, because we disbanded all teams.

好的，顺着这个思路，我问了你的 CPO 凯文·威尔该问你什么，他说让我问你关于“这是否已最大程度加速？”这个原则。谈谈这个吧。

Okay, so along these same lines, I asked Kevin Weil, your CPO, what to ask you, and he said to ask you about this principle of 'Is this maximally accelerated?' Talk about that.

有趣的是，我们现在好像有个 Slack 表情符号专门用来表达这个，因为我以前常这么说。现在我会试着换个说法。有时候我就是想直接跳到重点，比如“好，为什么我们现在不能做这个？”或者“为什么不能明天做？”我觉得这是和团队一起排除大量障碍的好方法，尤其是如果你来自大公司。我们开始从大型科技公司招人后，我觉得他们习惯了“我们一周后再看这个”或者“我们下个季度再回来看看能不能排上计划”。我把它当作一种思维练习，就像有人问：“如果这是最重要的事情，你想真正最大程度地加速它，你会怎么做？”这并不意味着你就真的那么做，但它确实是一个很好的强制功能，能帮你理解什么是关键路径，什么是可以稍后处理的事情。我一直觉得执行力极其重要。这些想法到处都是。每个人都在谈论“嘿，个人 AI”，你可能也看到了相关新闻。我真的认为执行力是这个领域最重要的事情之一，而这是一个工具。所以这变成了一个梗，还挺好笑的。就像一个小小的粉色 Slack 表情符号，人们把它放在任何他们想推动的问题上。

It's funny, we have a Slack emoji apparently for this now, because I used to say that. Now I try to paraphrase. Sometimes I just really want to jump to the punch line of like, okay, why can't we do this now, or why can't we do it tomorrow? And I think that it's a good way to cut through a huge number of blockers with the team, and just instill, especially if you come from a larger company. At some point we started hiring people from larger tech companies. I think they're used to, let's check in on this in a week, or let's circle back next quarter to see if it can go on the plan. And I just kind of as a thought exercise, I was like, people asking like, okay, if this was the most important thing and you wanted to truly maximally accelerate it, what would you do? That doesn't mean that you go do that, but it's really a good forcing function for understanding what's critical path versus what can happen later. And I've just always felt like execution is incredibly important. These ideas are everywhere. Everyone's talking about, hey, personal AI, you might have seen news on that. And I really think that execution is one of the most important things in the space, and this is a tool. So it's funny that that became a meme. It's like a little pink Slack emoji that people just put on whatever they're trying to force the question.

我正想问那个表情符号是什么。所以它是一个粉色的。里面是不是有“Max”之类的？

I was going to ask what the emoji was. So it's a little pink. Is there something in there like 'Max'?

那是一个 Comic Sans 字体的表情符号，写着“这是 Maximalist 吗？”所以那里的文化是，当有人在做什么事情时，问题、推动力就是：这是否已最大程度加速？有没有办法能更快？有没有什么我们可以解决的阻碍？

It's a comic sans emoji that says, 'Is this Maximalist?' And so the culture there is when someone is working on something, the question, the push, is: is this maximally accelerated? Is there a way we can do this faster? Is there anything we can unblock?

是的。而且你知道，我们很少用这个，对吧？因为它需要适合具体情境。有些事情你不想尽可能快地加速，因为你想要流程，我们在这方面非常谨慎，流程是一种工具。而我们拥有大量流程的领域之一就是安全，因为，a，风险已经很高了，尤其是这些模型，你知道，GPT-5 在很多方面都是前沿，但 b，如果你相信指数增长，我相信，而且大多数从事这项工作的人也相信，你必须为真正需要流程的时候练习。当然，当然。这就是为什么我认为将产品开发速度（必须非常快）与前沿模型等事情分开非常重要，对于前沿模型，确实需要严格的流程，比如红队测试、系统卡、外部输入，然后才能放心地发布，确保经过了适当的安全保障。所以再说一次，这是一个微妙的概念，但我发现它在我们需要时非常非常有用。而对于所有产品开发来说，如果你不快速行动，你就死定了。所以推出东西很重要。

Yeah. And you know, we use that sparingly, right? Because it needs to be appropriate to the context. There are some things where you don't want to accelerate as quickly as possible, because you kind of want process, and we're very deliberate on that, where process is a tool. And one of the areas where we have an immense amount of process is safety, because, a, the stakes are already really high, especially with these models, you know, GPT-5 which is the frontier in so many different ways, but b, you kind of, if you believe in the exponential, which I do, and most people who work on this stuff do, you have to practice this for a time where you really, really need the process. For sure, for sure. And that's why I think it's been really important to separate out the product development velocity, which has to be super high, from, okay, for things like frontier models, there actually needs to be a rigorous process where you red team, you work on the system card, you get external input, and then you put things out with confidence that it's gone through the right safeguards. So again, it's a nuanced concept, but I found it very, very useful when we need it. And for everything product development, you're dead on arrival if you don't move fast. So it's important to get stuff out.

我们得把这个作为梗开源，这样其他团队就能借鉴这种方法。

We got to open source as memes so that other teams can build on this approach.

当然。

Absolutely.

有趣的是，ChatGPT 不仅是史上增长最快、最成功的消费产品，留存率也高得惊人。有人分享过这些数据：一个月留存率大约 90%，六个月留存率大约 80%。首先，这些数字准确吗？你能分享一下吗？

So interestingly with ChatGPT, and it's not a surprise, but not only is it the fastest growing, most successful consumer product ever, retention is also incredibly high. People have shared these stats that one-month retention is something like 90%, six-month retention is something like 80%. First of all, are these numbers accurate? Can you share that?

显然，我能分享的内容有限。但我们的留存率数字确实非常令人兴奋，这也是我们真正关注的。我们完全不关心你在产品上花了多少时间。事实上，我们的动机只是解决你的问题，如果你真的喜欢这个产品，你会订阅。但我们没有动机让你长时间留在产品里。不过，如果长期来看，比如三个月周期等，你还在使用这个东西，我们显然非常非常高兴。对我来说，这从一开始就是房间里的大象。就像，嘿，这可能是一个很酷的产品，但这真的是你会回头使用的东西吗？看到强劲的留存率数字，以及随着时间推移留存率的提升，即使我们的用户群体从早期采用者变成了更普通的人，这都非常不可思议。

I'm obviously limited on what exactly I can share. But it is true that our retention numbers are really exciting, and that is actually the thing we look at. We don't care at all how much time you spend in the product. In fact, our incentive is just to solve your problem, and if you really like the product, you'll subscribe. But there's no incentive to keep you in the product for long. But we are obviously really, really happy if, over the long run, you know, 3-month period etc., you're still using this thing. And for me, this was always the elephant in the room early on. It's like, hey, this may be a really cool product, but is this really the type of thing that you come back to? And it's been incredible to not just see strong retention numbers, but to see an improvement in retention over time, even as our cohorts become less of an early adopter and more the average person.

是的，所以这一点我觉得人们并不真正理解这有多罕见。当一个产品的用户群来试用，然后留存率随时间下降，然后又回升，人们几个月后回来使用更多，这被称为微笑曲线，这极其罕见。

Yeah, so that note is something that I don't think people truly understand how rare this is. When a product's cohort of users comes, tries it out, and then retention over time goes down, and then it comes back up, people come back to it a few months later and use it more, and that's called a smiling curve or smile curve, and that's extremely rare.

是的，是的，是的。我知道有很多微笑，不只是团队里。我觉得我必须承认，其中一部分不是产品的原因。我认为人们实际上正在以一种非常有趣的方式适应这项技术。这也是为什么产品需要不断进化。把任务委托给 AI 这个想法对大多数人来说并不自然。你不会在生活中想着“我能委托什么”，硅谷的某些圈子会这样做，因为他们处于自我优化模式，试图委托所有能委托的事情。但我认为对世界上大多数人来说，这其实很不自然，你真的需要学习，“好吧，我真正的目标是什么，另一个智能能帮我什么？”我认为这需要时间，一旦人们有足够的时间使用产品，他们就会明白。但当然，我们在产品上也做了很多事情，无论是让核心模型更好，还是搜索和个性化等新功能，或者只是我们开始做的标准增长工作。这些东西当然很重要。

Yeah, yeah, yeah. I know there's some smiling going on, not just on the team. And I feel I have to acknowledge that some of it is not the product. I think people are actually just getting used to this technology in a really interesting way. And this is why the product needs to evolve too. This idea of delegating to an AI is not natural to most people. It's not like you're going through life and figuring out what can I delegate, like certain spheres of Silicon Valley do that, because they're in a self-optimization mode and they're trying to delegate everything they can. But I think for most people in the world, it's actually quite unnatural, and you really have to learn, okay, what are my goals actually, and what could another intelligence help me with? And I think that just takes time, and people do figure it out once they've had enough time with the product. But then of course there's been tons of things that we've done in the product too, whether it's making the core models better, whether it's new capabilities like search and personalization, and all that kind of stuff, or just standard growth work too, which we're starting to do. That stuff matters, of course.

所以你可能已经在回答这个问题了，但让我直接问一下。

So you might be answering this question already, but let me just ask it directly.

人们可能会看着这个说，好吧，他们正在这个神级智能之上构建某种层。当然，它会增长得非常快，留存率也会非常惊人。你们到底在模型之上做了什么，让它增长这么快、留存这么多？有没有什么特别有效、显著推动指标的事情可以分享？

People may look at this and be like, okay, they're building this kind of layer on top of this godlike intelligence. Of course, it will grow incredibly fast and retention will be incredible. What the heck are you guys actually doing that sits on top of the model that makes it grow so fast and retain so much? Is there something that has worked incredibly well that has moved metrics significantly that you can share?

我的意思是，我们学到的一件事，我一会儿再回答那个问题，但我们在 ChatGPT 上学到的一件事是，模型和产品之间真的没有区别。模型就是产品。因此，你需要像产品一样迭代它。我的意思是，你显然通常从发布一个非常开放的东西开始。至少如果你是 OpenAI，那是一种套路。但然后你真的要看看人们在试图做什么。好吧，他们试图写作，试图编码，试图获得建议，试图获得推荐，你需要系统地改进这些用例。这和产品开发工作非常相似。显然，方法论有点不同，但发现过程是一样的。你得和人交谈，你得做数据分析，你得尝试东西并获取反馈。所以这是我们一直在非常自觉地做的一部分工作，即在人们关心的用例上改进模型。还有所谓的“感觉”这种东西，我相信你知道，这也是我对 GPT-5 感到兴奋的一点，因为它的感觉真的很好。所以这也是，我们有一个模型行为团队，他们真正关注的是这个模型的个性是什么，它如何说话和交流。所以有那种工作。我想说这大概是我们看到的留存率提升的三分之一左右。然后我认为另外三分之一是我所说的产品研究能力。它们肯定是研究驱动的，有研究成分，但它们实际上是新的产品功能或能力。搜索就是一个例子，如果你记得在过去的日子里，也就是大概 20 个月前，你和 ChatGPT 说话，它会说“根据我的知识截止日期”或“我无法回答，因为那发生得太近了”之类的话。而这是一种非常有留存力的能力，原因很充分。它只是让你能用产品做更多事情。个性化，比如高级记忆的概念，让东西能随着时间真正了解你，是另一个这样的能力例子。我认为那是另一个很大的部分。然后第三部分是你会在任何产品中做的事情，那些东西也存在。比如不用登录是一个巨大的成功，因为它消除了大量摩擦，我想我们一开始就有这个直觉，但我们从未做到，因为我们没有足够的 GPU 或其他限制去真正做那件事。所以也有传统的产品工作。所以我经常把它想成大致三分之一、三分之一、三分之一，但实际上，我们仍在学习，我们计划大幅发展产品，这就是为什么我确信会有新的杠杆。

I mean, one thing we've learned, I'll answer that question in a minute, but the one thing we've learned with ChatGPT is that there really is no distinction between the model and the product. The model is the product. And therefore, you need to iterate on it like a product. By that I mean, you obviously typically start by shipping something very open-ended. At least if you're OpenAI, that's kind of a playbook. But then you really have to look at what are people trying to do. Okay, they're trying to write, they're trying to code, they're trying to get advice, they're trying to get recommendations, and you need to systematically improve on those use cases. And that is pretty similar to product development work. Obviously, the methodology is a bit different, but the discovery is the same. You got to talk to people, you got to do data science, and you got to try stuff and get feedback. So that's like one chunk of work that we've been very consciously doing, improving the model on the use cases people care about. And there's also such a thing as vibes, as I'm sure you know, and that's one of the things that I'm excited about in GPT-5 is that the vibes are really good. So that too, we have a model behavior team and they really focus on what is the personality of this model and how does it speak and talk. So there's that kind of work. I would say that's maybe a third of the retention improvements that we see, roughly. And then I think another third is what I would call product research capabilities. They're research-driven for sure, they have a research component, but they're really new product features or capabilities. And search is one example of that, where if you remember in the olden days, aka maybe 20 months ago or something, you would talk to ChatGPT and be like, "As of my knowledge cutoff" or "I can't answer that because that happened too recently" or something like that. And that is a type of capability that has been incredibly retentive, for good reason. It just allows you to do more with the product. Personalization, like this idea of advanced memory where things can really get to know you over time, is another example of a capability like that. I think that's another good chunk. And then the third stuff is the stuff you would do in any product, and those things exist too. Like not having to log in was a huge hit because it removed a ton of friction, and I think we had this intuition from the beginning, but we never got to it because we didn't have enough GPU or other constraints to really go do that. So there's the traditional product work too. So I often think about it sort of as roughly a third, a third, a third, but really, we're still learning and we're planning to evolve the product a ton, which is why I'm sure there's going to be new levers.

你提到了一个我想快速回顾的事情。你说从黑客马拉松到 Sam 发推说 ChatGPT 上线，大概用了 10 天。

You mentioned something that I want to come back to real quick. You said that it was something like 10 days from hackathon to Sam tweeting about ChatGPT being live.

你知道，黑客马拉松发生得更早，我们原型设计了很长时间，但在某个时刻，我们基本上对尝试构建更定制化的东西失去了耐心，而且这主要是因为每当我们测试时，人们总是想做所有其他事情。所以从我们决定要发布到真正发布，用了 10 天。而我们测试了很长时间的研究，是我们所说的“指令遵循”的演变，这个想法是，这些模型不是仅仅完成句子，而是能真正遵循你的指令。所以如果你说“总结这个”，它真的会这么做。研究从那里演变成一种聊天格式，我们可以进行多轮对话。所以那项研究花了远不止 10 天，并且在后台慢慢酝酿，但这个产品的产品化非常非常快。而且很多东西没有包含进去。我记得我们没有历史记录，这当然是用户给我们的第一个反馈。模型有很多缺点，但能够像我刚说的那样迭代模型，真是太酷了。在 ChatGPT 之前，把模型当作产品是不存在的，因为我们更像发布硬件那样发布它，比如发布 GPT-3，然后我们开始研究 GPT-4，这些是巨大的、高投入的研发项目，需要很长时间，规格就是那样，然后你得再等一年。而 ChatGPT 真正打破了这一点，因为我们能像软件一样对它进行迭代改进。而我的梦想是，如果我们能像软件领域那样每天甚至每小时发布一次，那就太棒了，因为你可以修复问题等等。但当然，在保持个性完整、不使其他能力倒退的同时做到这一点，有各种各样的挑战。所以这是一个开放的领域。

You know, the hackathon happened much earlier and we were prototyping for a long time, but at some point we basically ran out of patience on trying to build something more bespoke, and again that was mostly because people always wanted to do all this other stuff whenever we tested it. So it was 10 days from when we decided we were going to ship to when we shipped. And the research we'd been testing for a long time was kind of an evolution of what we'd called instruction following, which was the idea that instead of just completing the sentence, these models could actually follow your instructions. So if you said summarize this, it would actually do so. And the research had evolved from that into a chat format where we could do it multi-turn. So that research took way longer than 10 days and was kind of baking in the background, but the productization of this thing was very, very fast. And lots of things didn't make it in. Like I remember we didn't have history, which of course was the first user feedback we got. The model had a bunch of shortcomings, and it was so cool to be able to iterate on the model like the thing I just talked about. Treating the model as a product was not a thing before ChatGPT because we would ship it more like hardware, where there'd be a release like GPT-3 and then we would start working on GPT-4, and these were giant, big-spend R&D projects that would take a really long time, and the spec was whatever the spec was, and then you'd have to wait another year. And ChatGPT really broke that down because we were able to make iterative improvements to it just like software. And really my dream is that it would be amazing if we could just ship daily or even hourly like in software land, because you could just fix stuff etc. But there are of course all kinds of challenges in how you do that while keeping the personality intact and not regressing other capabilities. So it's an open field to get there.

这是一个很好的例子，是不是最大程度加速了？好吧，我们要发布 ChatGPT。好吧，10 天。天哪。我们一直在谈论 ChatGPT。显然，它是一种聊天界面。每个人都在想，聊天是不是这一切的未来。有趣的是，Kevin Weil 在播客中提出了一个非常深刻的观点，我一直铭记在心，他说聊天实际上是在超级智能之上构建的天才界面，因为这是我们与各种智力水平的人类互动的方式。它从较低端的人扩展到超级聪明的人。所以作为一种扩展这个频谱的方式，它非常有价值。也许谈谈这个，聊天是不是 ChatGPT 的长期界面？我猜它叫 ChatGPT。

This is such a good example of is it maximally accelerated? Okay, we're going to ship ChatGPT. Okay, 10 days. Holy moly. We've been talking about ChatGPT. Clearly, it's a kind of chat interface. Everyone's always wondering is chat the future of all of this stuff. Interestingly, Kevin Weil made this really profound point that has always stuck with me when he was on the podcast that chat is actually a genius interface for building on a superintelligence because it's how we interact with humans of all variety of intelligence. It scales from someone at the lower end to a super super smart person. And so it's really valuable as a way to kind of scale this spectrum. Maybe just talk about that and just is chat the long-term interface for ChatGPT? I guess it's called ChatGPT.

我觉得我们应该在某个时候去掉“chat”或去掉“GPT”，因为它太拗口了。我们被这个名字困住了。但你知道，无论我们怎么处理，产品都会发展。

I feel like we should either drop the chat or drop the GPT at some point because it is a mouthful. We're stuck with the name. But you know, no matter what we do with that, the product will evolve.

我同意，自然语言确实有某种深刻之处。它确实是与人交流最自然的形式，因此，用自然语言与软件交流感觉很重要。但我觉得这和聊天不同。聊天在当时是最简单的发布方式。我对它作为一个概念能如此流行感到困惑。更让我困惑的是，有那么多人复制了这种范式，而不是尝试与 AI 交互的其他方式。我仍然希望这种情况会发生。所以，我认为自然语言会一直存在，但必须是一轮一轮的聊天交互这个想法，我觉得真的很局限。这也是我不太喜欢“超级助手”这个类比的原因之一，尽管我们过去经常用，因为如果你那样想，就会觉得像是在和一个人说话。但是，你知道，GPT-5 在制作优秀的前端应用方面非常出色。所以，我看不出有什么理由不让 AI 以某种方式渲染自己的 UI。你显然希望让它可预测且体验良好。但对我来说，把终极界面想象成聊天机器人感觉很局限。这几乎有点反乌托邦，我不想通过某个界面的代理来使用我所有的软件。我喜欢在 Figma 里工作，我喜欢在 Google Docs 里工作。对我来说，这些都是很棒的产品，它们不是聊天机器人。所以，我的观点是：自然语言，赞成；聊天，反对。总的来说，我只是希望看到更多关于人们如何与 AI 交互的消费端创新。可能性太多了，你必须去尝试。这就是聊天能流行的原因，我们只是做了，人们喜欢。所以，我希望在那里看到更多，我们也会尽自己的一份力。

I think that I agree that there's something profound about natural language. It really is the most natural form of communicating to humans, and therefore it feels important that you should be communicating with your software in natural language. I think that's different from chat though. I think chat was the simplest way to put something to you know, to ship at the time. I'm baffled by how much it took off as a concept. I'm even more baffled by how many people have copied the paradigm rather than trying out a different way of interacting with AI. I'm still hoping that will happen. So, I think natural language is here to stay, but this idea that it has to be a turn-by-turn chat interaction, I think, is really limiting. And this is one of the reasons I don't love the super assistant analogy, even though we used to always use it, because if you think that way, then you kind of feel like you're talking to a person. But, you know, and GPT-5 is amazing at making great front-end applications. So, I don't see a reason why you wouldn't have AIs that can render their own UI in some way. And you obviously want to make that predictable and feel good. But it feels limiting to me to think of the end-all-be-all interface as a chatbot. It actually kind of feels dystopian almost, where I don't want to use all my software through the proxy of some interface. I love being in Figma. I love being in Google Docs. Those are all great products to me and they're not chatbots. So, yes on natural language, but no on chat is where I would describe my point of view. And I'm just hoping in general that we see more consumer innovation on how people interact with AI. There are so many possibilities and you just got to try stuff. That's why chat stuck, you know, we just did it and people liked it. So, I'm hoping that we see more there and we'll try to do our part.

你提到你们有点被 ChatGPT 这个名字困住了。也许这就是答案的一部分，但我很好奇，你们早期有没有做出一些偶然的决定，这些决定一直延续下来，并且基本上改变了历史？

You mentioned that you kind of got stuck with this name ChatGPT. Maybe this is part of the answer, but I'm curious, are there any accidental decisions you guys made early on that have stuck and have essentially become history changing?

太多了，有趣的是，你根本没有时间思考，然后它们就变得超级重要。你知道，名字就是其中之一，我们前一天晚上从“与 GPT-3.5 聊天”改成了“ChatGPT”。稍微好一点，但还是真的很糟糕。

There are so many, and it's funny because you have no time to think about them and then they end up being super consequential. You know, the day was one, you know, we went from Chat with GPT-3.5 to ChatGPT the night before. Slightly better but still really bad.

之前叫什么？

What was it called before?

本来打算叫“与 GPT-3.5 聊天”。因为我们真的没想到它会成为一个成功的产品。我们实际上试图尽可能保持极客风格，因为它本质上就是如此。这是一个研究演示，而不是产品。所以我们不觉得那名字不好。但是，你知道，在最初发布时，免费是一个大问题。我觉得我们没有意识到这一点，因为 GPT-3.5 模型在此之前已经在我们的 API 中至少六个月了。我认为任何人都可以构建类似的东西。可能在模型方面没有那么好，但我认为它会火起来。所以，免费并提供漂亮的 UI 非常重要，这是你现在认为理所当然的。这就是为什么我认为分发和界面在 2025 年仍然持续重要。

It was going to be Chat with GPT-3.5. Because we really didn't think it was going to be a successful product. We were trying to actually be as nerdy as we could about it because that's really what it was. It was a research demo, not a product. So we didn't think that was bad. But, you know, in the original release, making it free was a big deal. I don't think we appreciated that because the GPT-3.5 model was in our API for at least 6 months prior to that. I think anyone could have built something like this. Might not have been quite as good on the modeling side, but I think it would have taken off. So making it free and putting a nice UI on it was very consequential in the way that you take for granted now. And this is why I think that distribution and the interface are continuously important even in 2025.

付费业务，现在是一个巨大的业务，无论是在消费端还是企业端，它的诞生最初只是为了拒绝需求。我们并没有头脑风暴“哦，AI 最好的变现模式是什么？”而是“什么样的变现模式，或者什么样的机制能让我们拒绝那些不如真正想用的人那么认真的人？”订阅恰好具有这种特性，然后它就发展成了一个大业务。

The paid business, which now is a giant business, both in the consumer space and in the enterprise space, the birth of that was just to turn away demand originally. It was not like we brainstormed, 'Oh, what is the best monetization model for AI?' It was really, 'What monetization model has, or what mechanism would allow us to turn away people who are less serious than the people who are really trying to use it?' And subscriptions just happened to have that property, and it grew into a large business.

是的，我认为在功能完善之前就发布这些非常奇特的能力是另一件事。这感觉像是一个战术决定，但它成为了一种剧本，因为我们会学到很多。还记得我们发布代码解释器的时候吗？发布后我们学到了很多。现在它在 ChatGPT 中被称为数据分析之类的，因为我们确实得到了真实世界的用例，然后我们可以优化。所以我认为随着时间的推移，有很多决定被证明是相当重要的，但我们不得不非常非常快地做出这些决定。

Yeah, I think shipping really funky capabilities before they were polished is another thing. That feels like a tactical decision, but it became a playbook because we would learn so much. Remember when we shipped Code Interpreter? We learned so much after we shipped it. Now it's known as Data Analysis in ChatGPT or something like that, just because we actually got real-world use cases back that we could then optimize. So I think there have been a lot of decisions over time that proved pretty consequential, but we made them very, very quickly as we have to.

每月 20 美元感觉是其中重要的一部分。感觉现在大家都这么做。

The $20 a month feels like an important part of this. Feels like everybody's just doing that now.

哦，那个，实际上，我记得我当时有点恐慌，因为我们真的需要推出订阅，因为那时我们每次都要把产品下线。就像，我不知道你还记不记得，我们有一个“失败鲸”。上面有一首 AI 生成的小诗。他们说，“我们必须把这个发布出去。”我记得我给一个我非常尊敬的人打电话，他在定价方面非常厉害。我说，“我该怎么办？”我们聊了很多，但我没有时间采纳大部分反馈。所以我做的是在 Discord 上发了一个 Google 表单，里面有，我想，定价时应该问的四个问题。

Oh, that one actually, I remember I had this kind of panic attack because we really needed to launch subscriptions because at the time we were taking the product down every time. It was like, I don't know if you remember, we had this like fail whale. There's like a little AI-generated poem on it. They were like, 'We had to get this out.' And I remember calling up someone I greatly respect who is incredible at pricing. And I was like, 'What should I do?' And we talked a bunch, and I just ran out of time to incorporate most of that feedback. So what I did do is ship a Google form to Discord with, I think, the four questions you're supposed to ask on how to price something.

是的。没错。

Yeah. Exactly.

是的。它确实有那四个问题，我清楚地记得，我得到了一个价格。这就是我们如何得到 20 美元的。但第二天早上，有一篇新闻文章说，“你不敢相信 ChatGPT 团队为产品定价问的四个天才问题。”真是，要是你知道就好了。所以，在极度公开的环境中构建，人们会从你所做的事情中解读出比实际更多的意图。但我们最终定了 20 美元。当时我们在争论稍微高一点的价格。我经常想知道会发生什么，因为很多其他公司最终都复制了 20 美元的价格点。所以我想，我们这样定价是不是抹掉了一大块市值？但最终，我不在乎，因为这些东西越容易获得越好。而且我认为这个价格点在西方国家，对很多人来说，相对于他们获得的价值是合理的。更重要的是，我们能够半定期地把东西降到免费层。我们总是尽可能这样做，包括 GPT-5。

Yeah. It literally had those four questions, and I remember distinctly, I got a price back. And that's kind of how we got to $20. But the next morning there was a press article like, 'You won't believe the four genius questions the ChatGPT team asked to price their product.' It was like, if only you knew. So there's something about building in this extreme public where people interpret so much more intentionality into what you're doing than might have actually existed at the time. But we got with the 20. We were debating something slightly higher at the time. I often wonder what would have happened because so many other companies ended up copying the $20 price point. So I'm like, did we erase a bunch of market cap by pricing it this way? But ultimately, I don't care because the more accessible we can make this stuff, the better. And I think this is the price point that in Western countries has been reasonable to a lot of people in terms of the value that they get back. And more importantly, we're able to push things down to the free tier semi-regularly. And we always do that when we can, including with GPT-5.

所以，这个调查，正式名称是 Van Westendorp 调查，就是你们最终为 ChatGPT 定价的方式。

So the survey, just to give it the official name, the Van Westendorp survey, is how you guys ended up pricing ChatGPT.

它是谷歌搜索结果的第一名。

It was the top Google result.

这发生在 chat 拥有实时信息之前，否则它也许能自己定价，但呃，是 Discord 加 Google 论坛加一篇关于那方法的博客文章让我们走到了那一步。

This was before chat had real time information otherwise it could have maybe priced itself but uh it was discord plus google forum plus a blog post on that methodology that um got us there.

这太不可思议了。多么有趣的故事。这就是 Superhuman 的 Rahulvore 在他的第一轮文章中推广的那项调查。

That is incredible. What a fun story. This is the survey that Rahulvore at superhuman popularized in his first round article.

是的。是的。是的。没错。没错。呃，是的，千万别把我当作定价专家。我觉得你你你在这方面有更合适的人选。

Yeah. Yeah. Yeah. That's right. That's right. Uh yeah, definitely don't bring me on here as a pricing expert. I think you you you have got better people for that.

无论对错，它现在都是世界上增长最快、疯狂创收的业务。所以，呃，我不会感觉太糟。

Whether it was right or wrong, it is now the fastest growing insane revenue generating business in the world. So, uh I wouldn't feel too bad.

不，结果很好。是的，

No, it worked out. Yeah,

结果很好。呃，顺便说一句，我使用的是每月 200 美元的档位，所以显然还有空间。

it worked out. Uh and by the way, I'm on the 200 a month tier, so there's clearly room.

谢谢。谢谢。你知道，那个故事也很有趣，因为你知道，最初 Plus 计划的目的就是为了先推出正常运行时间，然后推出我们无法扩展到所有人的功能，而在某个时候，Plus 档位的人太多了，它就失去了那个特性。嗯，所以我们想出 200 美元档位的主要原因是，我们有太多令人难以置信的研究，这些研究实际上非常非常强大。嗯，比如你知道的 03 Pro，或者你知道的明天的 GPD5 Pro。嗯，只是有一个载体把它提供给真正关心的人，这很令人兴奋，尽管它有点违反了 SaaS 页面应该有的标准样子。嗯，看到 10 倍的跳跃有点刺眼。所以嗯，感谢你成为那个订阅者，也感谢所有观看的人订阅任何档位。嗯，这很棒。

Thank you. Thank you. You know that the story of that one is is interesting too because you know originally the purpose of the plus plan was to be able to ship first uptime and then be able to ship capabilities that we couldn't scale to everyone and at some point we got so many people in the plus tier that it just lost that property. Um so the re the main reason we came up with the $200 tier is just we had so much incredible research that's actually really really powerful. um like you know 03 Pro or to you know tomorrow GPD5 Pro. Um and just having a vehicle of shipping that to people who really really care is exciting even though it kind of violates the standard way a SAS page should look. Um it's like a little jarring to see the see the 10x jump. So um thank you for being a subscriber on that and thank you everyone else who's watching you subscribe to any tier. Um it's it's great.

我只是想在这个池塘里抛一根钓鱼线，还有没有其他类似的故事？你分享了 chat with GPT 3.5 作为最初名字的不可思议的故事，以及你是如何想出定价的。还有别的吗？

I'm just going to throw a fishing line into this pond of are there any other stories like this? You shared this incredible story of chat with GPT 3.5 being the original name, how you came up with pricing. Is there anything else?

企业版也很有趣，因为我们在企业中看到了如此多令人难以置信的采用，同时尝试建立开发者业务、消费者业务和企业业务，这客观上有点疯狂。但你知道，那里的故事是，在第一个月或第二个月，我就很清楚，大部分使用都是工作性质的使用，实际上比今天多得多，今天产品上有这么多消费者，你知道，它有点超越了流行文化，但在当时，你知道，写作、编码、分析之类的东西，我们很快就自然进入了 90% 的财富 500 强公司，这种方式我可能之前在 Dropbox 见过，那是我两份工作之前，我们有过类似的故事，从那以后有了更多 PLG 公司，但我们做企业版的真正原因，我记得我们在争论是做企业版还是推出 iOS 应用，因为团队就是这么小，他们这样做的原因是，我们开始在公司里被禁止，因为他们都觉得，你知道，无论对错，隐私和部署的故事等等都不存在，所以我只是觉得，天哪，我们必须做点什么，否则我们会错过一个世代性的机会来构建一个工作产品，你知道，我们真的把 AGI 定义为，你知道，在经济上有价值的工作上超越大多数人类，或者我可能说错了，但你知道，我认为嗯，我认为我们就是这样说的。嗯，所以我觉得我们必须出现在那里。当时这是一个相当快速的决定，但它已经成长为一个巨大的业务。我们刚刚达到 500 万企业订阅者，我想一两个月前是 300 万。所以这有点像这个衍生品有了自己的生命，我对此非常非常兴奋。嗯嗯，原因很明显。

Enterprise interesting one too because we've been seen so much um incredible adoption in the enterprise and it's sort of objectively crazy to try to take on building a developer business and a consumer business and a develop and and an enterprise business and an and all at once. But you know the story there is in in like month one or or two I it was like very clear that most of the usage was like kind of worky usage actually much more than today where you've got so many like kind of consumers uh on the product and you know it's kind of sort of transcended into pop culture but at the time it was like you know writing coding analysis that kind of stuff and uh we were pretty quickly in you know organically in like 90% of Fortune 500 companies in a way that I had seen maybe at Dropbox back when I you know that two jobs ago where we kind of had a similar story and since then there's been more PLG companies but the real reason we did enterprise I remember we were debating should we do enterprise or should we launch an iOS app because that's how small the team was and the reason they did yeah did is we were starting to get banned in companies because they all you know felt you know rightfully or wrongfully that you know the the privacy and deployment story etc wasn't there so I was just like man we have to do something we're going to miss out on a generational opportunity to build a a a a work product and you know we've literally really define AGI as, you know, outperforming most humans at economically valuable work or I probably butchered that, but you know, I think um I think that's the way we put it. And um um so it I feel like we had to be present there. And it was a fairly, you know, quick decision at the time, but it's grown into an immense uh business. We just hit 5 million um business subscribers, up from three, I think, u a month or two ago. So it is kind of this spin-off that's taking a life of its own that I'm really really excited about. um um for for obvious reason

这要处理很多事情，呃，平台基本上是 API，消费者产品，历史上增长最快、最成功的产品，还有 B2B 方面，这显然是一个巨大的业务，呃，你有没有什么启发式方法来做这些权衡，同时做所有这些，保持理智并取得成功？

That is a lot to be handling uh the platform essentially the API the consumer product the fastest growing most successful product in history and also the B2B side which is uh clearly a massive business uh do you have any kind of heristics for how to make these trade-offs do all this at once and stay sane and be successful

呃，这是个好问题，首先，我不再负责开发者业务了，我们找到了一个更有能力的人来做那件事，嗯，他很棒。所以我仍然负责各种形式的 chat，但你知道，幸运的是，我不必做那个权衡。OpenAI 在做，我也可以深入谈谈。但是，嗯，这让我更理智一点。我想说的是，在构建这些 AI 东西时，你必须以两种不同的方式设定优先级。一种是某种从模型能力倒推，这更像是艺术而非科学，我认为你真的需要看看我们有什么技术，以及产品化它的最棒的方式是什么，如果你应用某种 PM 框架，我认为你会犯大错，因为如果你有技术，你知道，嗯，例如，GPD5 现在在前端编码方面非常非常擅长，我认为这意味着你必须重新设定优先级，你必须真正把这种能力带到生活中。也许那就是，你知道，呃，让 chat 在 vibe coding 和渲染应用方面更好。也许那更像是，你知道，利用模型的品味让 UI 更有表现力。我们可以做很多事情，但你得重新规划和重新设定优先级，这比任何特定的受众细分都重要。真的只是看看，你知道，我们拥有的神奇东西是什么，以及如何让它发光。语音也是类似的事情。并不是我们的客户需要语音，他们乞求它之类的。而是，哇，我们找到了一种方法，你知道，让这些东西，任意输入，任意输出。什么是创造性的、很棒的产品化方式？然后我们可以看看人们会怎么做。所以，我认为这是一部分。但另一部分真的更像是经典的产品管理，你需要倾听客户，然后当你的客户非常不同时，这可能会令人困惑，因为呃，你知道，chat 是一个非常通用的产品。

uh it's a good question and you first off I don't run the developer stuff anymore we found someone way more competent uh to do that um and he's amazing So I still look after the, you know, various forms of of of chat, but you know, I luckily don't have to make make that trade-off. Open eye does, and I can get into that, too. But, um, it keeps me a little bit more sane. I will say that there you kind of have to prioritize in two different ways when you're when you're building on this AI stuff. One is sort of working backwards from the model capabilities and that is much more art than science where I think you really need to look at what tech do we have available and what is like the most awesome way to product productize it and if you applied to some sort of PM framework to that I think you would do something horribly wrong because if you have tech that's you know um for example GPD5 is is really really good at front-end coding now like I think we that means you got to rep prioritize it you how to like actually bring that capability to life. Maybe that's you know uh making making chatb better at at vibe coding and rendering you know applications. Maybe that's more like you know leveraging the taste of the model to make the the UI more expressive. There's like a number of things we could do right but you kind of have to replan and rep prioritize and that you know is more important than any particular audience segmentation. It's really just looking at you know what is the magic thing we have and how do you make it shine. Voice is a similar thing. It wasn't like our customers need voice. They're begging for it or something like that. It's like, wow, we figured out a way how, you know, to make these things, anything in, anything out. What is like a creative awesome way to productize that? And then we can see what people do. So, I think that's one chunk of it. But then the other chunk of it really is more like classic product management where you need to listen to customers and then when your customers are really different, that can be confusing because uh you know, chatbt is a very general purpose product.

我们看到，当你观察终端用户时，他们在需求上其实有大量的重叠，比如项目、历史记录、搜索、分享、协作这些基本功能。无论你面对的是工作中的用户，还是家庭和学校里的用户，这些需求都非常普遍。有时机制略有不同，但投资方向大体相似，我认为我们可以从中获得很多收益。然后还有一些企业特有的工作我们必须做，比如要符合 HIPAA、SOC 2 等标准，如果你想成为严肃的玩家，这些都是不可妥协的。所以，正如你正确指出的，这很复杂。但这正是从事开放且强大技术所带来的挑战。OpenAI 有位我非常尊敬的人有时会用这样一个类比：我们有点像迪士尼，迪士尼有一种核心创意 IP，就是他们的内容，然后他们还有游轮、主题公园、漫画等各种业务。我认为我们有很棒的模型，但产品化的方式多种多样，我们必须在所有这些不同方向上最大化影响力。

We see when you look at end users there's actually an immense amount of overlap in terms of what they want like primitives like projects or history or search or sharing or collaboration. All those kind of things are actually very present whether you're talking to people at work or you're talking to people at home and school. They're slightly different mechanics sometimes, but they're largely similar investments that I think we can get a lot of mileage out of. And then there's enterprise specific work that we just have to do, like you got to do HIPAA, you got to do SOC 2, you got to do all those things if you want to be a serious player, and those are just non-negotiable. So it's complex, as you correctly identified. But it's kind of the curse of working on a very open-ended and powerful technology. One analogy that someone at OpenAI who I really respect sometimes uses is we're kind of like Disney, where Disney has this one kind of creative IP, which is their content, and they have cruises and theme parks and comics and all these different things. I think we have amazing models, but there's all these different ways that you could productize them, and we kind of just have to maximize the impact in all these different ways.

我们刚才聊的时候，我在想，通常那些非常通用、能做很多事的横向平台需要很长时间才能起飞，因为人们不知道拿它们做什么。它们在任何方面都不出众。而这是一个惊人的反例，它立刻起飞了，每个人都弄明白了，而且随着时间的推移，他们越来越明白。

As we were talking, I was thinking about how usually horizontal platforms that are just so general and can do so much take a long time to take off because people don't know what to do with them. They're not amazing at anything. And this is an amazing counter example where it took off immediately and everyone figured it out and then over time they figured it out more and more.

但我认为原因在于它直接上线了。这又说到另一个关键决定。你知道，我们当时在争论要不要用候补名单，因为我们很清楚工程系统无法扩展。而最终没有候补名单，这在 OpenAI 之前的发布中从未有过，结果影响深远，因为你能实时看到其他人在做什么。所以我认为，当你一次性向所有人发布这些东西时，确实有一个特殊时刻，你可以看到别人在做什么并从中学习。而且很多学习发生在产品之外。那些疯狂的 TikTok 帖子会病毒式传播，评论区有 2000 个用例，我会仔细浏览，因为那些用例我也不知道。它们非常涌现，我就去评论区里消化，因为有很多可学的。因此，我认为我们得以稍微避开“空盒子问题”，因为大量学习发生在产品之外，人们在线下或线上互相观察。

But I think the reason why is because it just went live. Talk about another consequential decision actually. You know, we were debating waitlist, no waitlist, because we just really knew we couldn't scale the engineering systems. And the fact that there was no waitlist, which no OpenAI release had worked like that before, ended up being consequential because you were able to watch what everyone else was doing live. So I think when you launch these things all at once for everyone, there really is a special moment where you can see what other people are doing and learn from that. And a lot of that is actually out of product. There's these crazy TikTok posts that go viral and they have like 2,000 use cases in the comments, and I go through those in detail because it's not like I knew about those use cases either. They're very emergent, and I just go through the comments and process because there's so much to learn. And for that reason, I think we get to escape the empty box problem a little bit, because so much learning is happening out of product as people are watching each other either IRL or online.

这太有趣了，因为想想 Airtable、Notion 这些公司，它们花了数年时间才构建、打磨、思考并深入研究产品能做什么。比如 Airtable，它必须做模板，必须做各种事情，把横向产品变成用例驱动；而 Instapot 则是在网上到处分享食谱，围绕它形成了一个完整的生态系统。我认为 ChatGPT 很幸运地实现了这一点，用户到处分享用例。因此，我认为我们很幸运，在这条路上走在了前面，对吧？

That is so interesting because you think about Airtable, you think about Notion, all these companies they took like years to just build and craft and think and go deep on what it could be. It's like compare Airtable, which had to do templates, had to do all these kind of things of taking the horizontal product and making it use case driven, compared to the Instapot, which has recipes being shared everywhere online, there's a kind of whole ecosystem around it. I think we were really lucky with ChatGPT that that happened, where there's just users sharing use cases with other users everywhere. And therefore I think we kind of got very lucky by jumping ahead on that journey, right?

是的，感觉核心在于 Sam 有大量粉丝，大家都会关注你发布的东西。所以这是一种非常有趣的新策略，用巨大的分发渠道来发布横向产品。直接发布，看看会有什么反应。

Yeah, and it feels like a core there is Sam had a big following and everyone would pay attention to something you launched. So that's a really interesting new strategy for launching horizontal product with a huge distribution channel. Just launch it and see what comes up.

是的。当然，我其实很兴奋能把其中一些带入产品。我认为我们不应该满足于产品之外有这么多发现。我实际上认为，对于普通消费者来说，如果产品能多做一点工作，真正展示出什么是可能的，那会非常棒。我仍然觉得 ChatGPT 有点像 MS DOS。我们还没有构建出 Windows，一旦构建出来，一切都会变得显而易见。但有些东西感觉有点像，想象 MS DOS 已经病毒式传播，而你只是试图在上面拼凑一些对话开场白，那可能错过了如何真正向人们传达功能与价值的全局。所以我认为，除了看到用例传播之外，还有大量的产品工作要做。

Yeah. And of course, I'm actually really excited to take some of that into the product. Like I think we shouldn't rest on the fact that there's so much out of product discovery happening. I actually think for the average consumer, it would be amazing if the product did a little bit more work on really exposing to you what is possible. I still feel like ChatGPT feels a little bit like MS DOS. We haven't built Windows yet, and it will be obvious once we do. But there's something that feels a little bit like imagine MS DOS had gone viral and you were just trying to hack little conversation starters onto it. That might have missed the big picture in terms of how to really communicate affordances and value to people. So I think there's actually a ton more product work to do in addition to just seeing use cases spread.

你能分享一下你认为这个 Windows 版的 ChatGPT 可能是什么样子吗？

Are you able to share just what you think that might look like, this Windows version of ChatGPT?

等我们搞清楚了我会告诉你。我们正在招人。我认为这里有很多有趣的产品问题。

I'll let you know when we figure it out. We're hiring. I think there's so many interesting product problems here.

好的，明白了。顺便说一句，我也很喜欢你把 TikTok 当作反馈渠道。

Okay, got it. By the way, I also love that TikTok was like your feedback channel.

那些评论串太疯狂了，还有人们对它的热爱。比如人们分享你的产品时的兴奋感。我觉得人们如此热衷于分享他们用你的产品所做的事情，这很特别。我也不认为这是理所当然的。

Those comment threads are just so wild, and also the love that people have for it. Like the excitement with what you're sharing their product. I kind of feel like it's special that people are so excited to share what they're doing with your product. And I don't take that for granted either.

本期节目由 PostHog 赞助播出，这是一个你的工程师真正想用的产品平台。PostHog 拥有创始人、开发者和产品团队所需的所有工具，如产品分析、网络分析、会话回放、热图、实验、调查、LLM 可观测性、错误跟踪等等。PostHog 提供的所有功能都有慷慨的免费层，每月重置。超过 90% 的客户免费使用 PostHog。你会喜欢与这样一个透明且技术过硬的团队合作。你会看到工程师为你的问题提交拉取请求，他们的支持团队在遇到棘手问题时提供代码级帮助。PostHog 让你所有数据集中在一个地方。除了分析事件，他们的数据仓库还能让你从 Postgres 数据库、Stripe、HubSpot、S3 等更多来源同步数据。最后，他们新的 AI 产品分析师 Max AI 帮助你更快更进一步。获取构建复杂查询和设置账户的帮助，专家随时待命。今天就免费注册 posthog.com/lenny，并告诉他们 Lenny 推荐了你。网址是 posthog.com/lenny。

This episode is brought to you by PostHog, the product platform your engineers actually want to use. PostHog has all the tools that founders, developers, and product teams need like product analytics, web analytics, session replays, heat maps, experimentation, surveys, LLM observability, error tracking, and more. Everything PostHog offers comes with a generous free tier that resets every month. More than 90% of customers use PostHog for free. You are going to love working with a team this transparent and technical. You'll see engineers landing pull requests for your issues and their support team provides code level assistance when things get tricky. PostHog lets you have all your data in one place. Beyond analytics events, their data warehouse enables you to sync data from your Postgres database, Stripe, HubSpot, S3, and many more sources. Finally, their new AI product analyst, Max AI, helps you get further faster. Get help building complex queries and setting up your account with an expert who's always standing by. Sign up today for free at posthog.com/lenny and make sure to tell them Lenny sent you. That's posthog.com/lenny.

你现在如何发现涌现的用例？我猜数量非常大。

How do you find emergent use cases these days? I imagine the volume is very high.

你有没有什么诀窍，能发现“哦，这里有个新东西我们真该好好想想”？

Do you have kind of a trick for figuring out, oh, here's a new thing we should really think about?

在我组建产品团队之前，我其实先组建了数据科学团队。因为当时我很沮丧。我尽可能多地跟用户交流，ChatGPT 发布后的几周里，我的日程表上全是 15 分钟的用户访谈，整整一周都是。我通常在能预测下一个人会说什么的时候就不再访谈了，那说明我已经聊够了用户。但这次就是做不到，我总能听到新东西。所以数据是一条出路，我们有对话分类器，不用我们亲自看对话，就能知道人们在聊什么、哪些用例在起飞等等。我觉得这非常非常有帮助。定性研究对共情很重要，尽管你永远无法覆盖人们所有的用例。我仍然花大量时间做这件事。然后，像那些 TikTok、帖子合集，我觉得它们真的非常有用，而且看人们互相交流他们的各种用例也很有趣。

Before I built the product team, I actually built the data science team. Because I was getting frustrated. I was talking to as many users as I could, and my calendar in the weeks after ChatGPT was just 15-minute user interviews the whole week through. And I usually stopped doing interviews when I can predict what the next person's going to say. That's how I know I've talked to enough users. But it just wasn't happening. I just kept getting new stuff. So data is one way out where I think we have conversation classifiers that, without us having to look at the conversations, allow us to figure out what people are talking about, what use cases are taking off, etc. And I think that's very, very helpful. The qualitative stuff is important for empathy, even though you're never going to get a representative sample of all the use cases people have. I still spend a huge amount of my time doing that. And then, yeah, things like those TikToks, collections of threads, I think they're really, really useful, and it's just fun to watch people talk to each other about the various use cases that they have.

有没有什么新兴用例让你很兴奋，或者有没有什么特别不寻常的 ChatGPT 用法，你觉得分享出来会很有趣？

Is there a new emergent use case that you're excited about, or is there a really unusual use of ChatGPT that you think about that would be fun to share?

我之前提到过，但我一直把 ChatGPT 概念化为一个工作产品。无论你在家还是在工作，我觉得帮你处理税务跟你在工作中做的事情非常相似，或者规划旅行其实跟规划工作活动很相似。所以我一直觉得这东西会成为一个生产力工具。但我觉得过去几个月发生了一些事情，这种情况开始改变。我真的觉得，消费者转向这个东西寻求日常建议，帮助他们改善人际关系，看到人们谈论这东西如何拯救了他们的婚姻，这让我非常兴奋。因为他们用它来处理自己的情绪，获得对自己沟通风格的反馈，有一个伙伴可以聊非常困难的事情。这带来了大量的责任和工作，我们必须把这些事情做好，比如生活建议。但这也对我非常重要，因为你不能逃避这些用例。你必须迎难而上，把它们做得更好。这就是我们正在努力的一部分。所以这种新兴行为真的非常非常酷。更广泛地说，我对教育非常兴奋。我对健康也非常兴奋。我觉得如果我们不利用 ChatGPT 的机会去真正帮助人们，那真的是浪费。而且我觉得我们才刚刚触及表面。所以有很多我渴望实现的用例。

I mentioned this earlier, but I had always conceptualized ChatGPT as a work product. Whether you're at home or at work, I feel like helping get help with your taxes is very similar to the types of things you do at work, or planning a trip is actually very similar to planning an event for work. So I've always felt like this thing is going to be a productivity tool. And I think something has happened in the last few months where that has begun to change. I really do think the fact that consumers are turning to this thing for day-to-day advice, helping them have better relationships, seeing people talk about how this thing saved their marriage, is really exciting to me. Because they use it to process their own emotions, get feedback on their communication style, have a buddy to talk to about really difficult things. And that comes with a ton of responsibility and work that we have to do to make those things like life advice great. But it also is really, really important to me because you can't run away from those use cases. You have to run towards them and make them awesome. And that's part of what we're trying to do. So that emerging behavior is really, really cool. And more broadly, I am so excited about education. And I'm so excited about health. I think it would really be a waste if we didn't take the opportunity of using ChatGPT to really, really help people. And I think we've just begun to scratch the surface on that. So there are many aspirational use cases that I want to make happen.

顺着这个思路，我最近有一个有趣的用例，我觉得它对那些意见不合、需要第三方意见的夫妻会很有帮助。我最近就遇到一次，我妻子说：“你不能把只吃一部分的东西在微波炉里加热，然后又放回冰箱。”我说：“有什么问题？我加热一下，再放回冰箱。”她说：“不，那真的很危险。”我说：“我们问问 GPT 吧。”她现在如此信任 ChatGPT，整天依赖它，它真的是一个非常有价值的独立第三方，我们可以去问它。

Along those lines, an interesting use case I've recently had, I feel like it's going to be really helpful for couples that are disagreeing about something when they need a third opinion. I just had this recently where my wife's like, "You can't heat a whole thing that you're gonna only eat part of in the microwave and then put it back in the fridge." It's like, "What's the problem? I'll heat it up. I'll put it back in the fridge." And she's like, "No, that's really dangerous." I'm like, "Let's ask GPT." And the fact that she so trusts ChatGPT now and relies on it throughout the day, it's such a valuable third independent party that we can go to.

是的，完全同意。而且你知道，很多这种微交互，就是有趣的产品工作，对吧？这些微交互很重要，对吧？它是给出了明确的裁决，还是帮助你们自己思考解决了那个分歧？我觉得这些细节其实非常重要，这也是我们花大量时间的地方。

Yeah. Totally. And you know, a lot of those micro interactions, talk about interesting product work, right? Those micro interactions are important, right? Did it definitively weigh in, or did it help you guys think through that disagreement and solve it on your own? I think those details actually matter a lot, and it's where we're spending a bunch of time.

顺着这个思路，之前发布了一个非常谄媚的 ChatGPT 版本，它只会说“你是世界上最好的人。你告诉我的一切都惊人地正确。”你能告诉我们到底发生了什么吗？

Along those lines, there was this whole launch of the very sycophantic version of ChatGPT where it was just "You are the best person in the world. Everything you tell me is amazingly correct." Are you able to tell us just what happened there?

是的，我们在网上有各种相关资料，因为我们真的觉得应该充分沟通我们是如何发现的、我们做了什么等等。所以我鼓励大家去看看。我们对那次模型发布有一个完整的复盘。但基本上发生的是，我们推送了一个更新，让模型更倾向于说那些当下听起来好听的话。就像你说的，它可能会说“你应该和男朋友分手”之类的话。这真的很危险。我们对此的重视程度可能超出你的预期，因为再说一次，在当前的技术水平下，你可能会一笑置之。也许你会想：“啊，这东西总夸我，我还以为只是对我这样。”我看到了网上那些评论。但确保这些模型朝着正确的方向优化，确实非常重要。而且我认为我们有一个巨大的奢侈，那就是我们的使命是真正帮助人们，我们的商业模式不鼓励最大化用户参与度或产品使用时长。所以对我们来说，让你觉得这个产品在帮助你实现目标非常重要，无论是你当前的目标还是长期目标。而很多时候，对用户极度赞美其实并不服务于这个目标。所以我们引入了新的衡量技术。比如，每当我们让这些模型接触现实并发现问题时，我们就会回去确保我们有好的指标来衡量这些东西。所以我们现在每次发布都会衡量安全性，确保我们不退步，并且能真正改进这个指标。GPT-5 是一个改进，这让我非常兴奋，但我们还有更多工作要做。更广泛地说，这件事促使我们阐明了自己的观点。我们花了很多时间写了一篇博客文章，周一刚发布，讲的是我们优化 ChatGPT 是为了什么。它真的是为了帮助你茁壮成长、实现目标，而不是让你留在产品里。所以那次事件带来了很多好的结果。

Yeah, we have all kinds of collateral online because we really felt like we should overcommunicate on how we discovered it, what we did about it, etc. So I encourage people to check that out. We have a whole retro on that model release. But basically what happened is that we pushed out an update that made the model more likely to tell you things that sound good in the moment. And like you're totally right, you should break up with your boyfriend or something like that. And that's just really dangerous. And we took it more seriously than you even might expect because again, at current technology levels, you can kind of laugh about it. Maybe it's like, "Ah, this thing's always complimenting me. I thought it was just me." I saw all those comments online. But it actually is really important to make sure that these models are optimized for the right things. And we have an immense luxury, I think, to have a mission that affords us to really help people, a business model that does not incentivize maximizing engagement or time spent in the product. So it's really important to us that you feel like this product is helping you with your goals, whether that's your current goals or even your long-term goals. And oftentimes, being extremely complimentary with the user isn't actually in service of that. So we instilled new measurement techniques. Like, whenever we put these models in contact with reality and we learn about a problem, we actually go back and make sure we have good metrics for this stuff. So we measure safety now with every release to make sure we don't regress and can actually improve on that metric. GPT-5 is an improvement, which is really exciting for me, but we have more work from there. And more broadly, it caused us to articulate our point of view. We actually spent a bunch of time on a blog post that we just published on Monday on what we're optimizing ChatGPT for. And it really is to help you thrive and achieve your goals, not to keep you in the product. So there were a bunch of good outcomes from that incident.

这是一个很好的例子，说明接触熟悉度不仅对用例很重要，而且对学习要避免什么也很重要，因为除非你实际听到，否则你永远不会在实验室里纯粹发现这个问题。

It's a good example of how contact familiality is not just important for the use cases but also for learning what to avoid because you would have never discovered this issue purely in a lab unless you actually heard it.

那我真的很期待读那篇博客文章。我本来想问你这个问题，就像……

I am excited to read that blog post then. I was going to ask you this just like...

是的，听听你的反馈。

Yeah, have your feedback on it.

是的。我想那里还有什么更多的东西吗？就像你如何，因为这个张力非常困难，你知道，帮助人们感到被支持，但又不只是让他们相信他们想相信的一切。你能分享更多关于如何找到那个中间地带吗？

Yeah. And I guess is there anything more there just like how you, because this tension is so difficult, like you know, helping people feel supported but not just letting them believe everything they want to believe. Is there anything more you can share there just trying to find that middle ground?

激励很重要。有句名言，你知道，给我看激励，我就告诉你结果。

Incentives are important. There's a famous saying, you know, show me the incentive and I'll show you the outcome.

可能是查理·芒格。

Charlie Munger, maybe.

嗯，是的，我想这就是它的出处，对吧？我认为这非常非常重要。所以，我会仔细审视我们的使命、我们的商业模式、我们试图构建的产品类型。而且，我真的认为，聊天是一个非常特别的产品，因为在绝大多数情况下，它让你离开时感觉更好，而不是更糟。你知道，感觉你在实现你试图做的事情。所以我认为这些激励真的很重要，因为它帮助你推理：当野外出现不好的行为时，那是一个错误还是设计使然？你知道，对于 sopy，我可以非常肯定地说，对我们来说那是一个错误。然后在前瞻性工作上，有很多具有挑战性的场景需要处理好。你可以很容易地逃避这些用例，比如你知道，你和你妻子去参加那个活动，寻求关于关系问题或纠纷的建议。如果你完全规避风险，你可以很容易地逃避，说对不起，我帮不了你。我认为大多数科技公司在达到一定规模时都会这样做。他们逃避这些用例，我认为这是帮助人们的机会损失。所以我们想要通过让模型行为变得非常非常好来迎接这些用例。嗯，这可能意味着在你挣扎时为你连接外部资源。这可能意味着不直接回答你的问题，而是给你一个有用的框架，你知道，比如“我应该和男朋友分手吗”这样的问题，可能不应该替你回答，但它应该帮助你以深思熟虑的同伴的方式思考这个问题。所以我认为做这项工作非常重要，因为我认为好处是巨大的。

Um, yeah, I think that's where it came from, right? And I think that's very, very important. So, I would take a good look at, you know, our mission, our business model, the type of product we're trying to build. And, you know, I really think that, you know, chat is a very special product because it, I think in vast majority of cases, it makes you leave it feeling better, not worse. And you know, feeling like you're achieving something you're trying to do. And so I think that those incentives really matter because it helps you reason about okay, when there isn't behavior in the wild that's not good, was that a bug or was that by design? You know, and with sopy I can very much say that to us that's a bug. And then on the forward-looking work, there's so many challenging scenarios to get right. And you could easily run away from these use cases, like you know, you and your wife going to this thing for input on a relationship question or like a dispute. You could very easily run away if you were totally risk avoidant and say sorry I can't help you with that. I think that's what most tech companies do when they hit a certain scale. They run away from these use cases and I think it's a loss opportunity to help people. So we want to run towards these use cases by making the model behavior really really great. Um, that can mean connecting you with external resources when you're struggling. That can mean not directly answering your question but instead giving you a helpful framework, you know, in the case of like should I break up with my boyfriend, should probably not answer that question for you, but it should help you think through that question in the way that a thoughtful companion would. So I think it's really important to do the work because I think the upside is immense.

你提出的观点非常深刻，如果大多数公司，如果他们的用户想问一些有风险的事情，比如寻求医疗建议，或者我应该和伴侣分手吗，或者我该如何处理这个大问题。我觉得如果我们有一个在 Healthbench 上达到最先进水平的模型，你知道，GP5 在很多医学基准上是最先进的，对吧？而你却没有用它来帮助人们，比如你只是因为想避免所有可能的负面影响而禁用这个用例。我认为责任是让它变得很棒，并且做这项工作，与专家交谈，弄清楚它到底有多好，在哪里失效，并传达这些信息。而且，你知道，我认为这项技术太重要了，对人们有太大的潜在积极影响，不能逃避这些高风险的用例。

That is a really profound point you're making there that if most companies, if their users want to ask them something risky like get medical advice or should I break up with my partner or what should I do with this big problem I have. I feel like we would have immense regret if you had a model that was state-of-the-art on Healthbench, which is, you know, a um, GP5 is state-of-the-art on, you know, a bunch of these medical benchmarks, right? And you didn't use that to help people, like if you just disable that use case because you wanted to like avoid all possible downside. I think the duty is to make it awesome um and to do the work, talk to experts, figure out how good it really is, where it breaks down, communicate that. And um you know I think this technology is too important and has too much potential positive impact on people to run away from these high stakes use cases.

快进到今天，它经常拯救生命。它可能经常拯救关系。这是一个如此重要的决定，我想它很早就做出了。

And fast forward to today, it's saving lives regularly. It's probably saving relationships regularly. Such a consequential decision which I imagine was made early on.

你知道，我们才刚刚开始看到这东西如何改变人们。嗯，如果你把这东西的推广与个人电脑的推广相比，它简直太民主化了，对吧？你知道，电脑刚出来时非常稀缺。而这东西无处不在，你可以获得医疗方面的第二意见。你可以获得一个关系伙伴。你可以获得一个个人导师，几乎任何让你好奇的话题。嗯，我们能做到这一点真的非常特别。所以，这是历史上独特的时刻。

You know, we're just at the beginning of watching how this stuff can transform people. Um, it's incredibly democratizing if you compare, you know, the roll out of this with the roll out of the personal computer, right? You know, computers were like so scarce when they first came out. And this stuff is ubiquitous in a way where you have access to a second opinion on medical stuff. You have access to, you know, a relationship buddy. You have access to a personal tutor on literally any topic that makes you curious. Uh, it's really really special that we get to do that. So um, unique point in history.

让我稍微拉远一点，谈谈 OpenAI 和一般产品。所以你在传统的产品公司工作过，比如 Dropbox、Instacart。现在你在 OpenAI。你在 OpenAI 期间学到的关于构建产品的最反直觉的教训是什么？

Let me zoom out a bit and talk about OpenAI and just product in general. So you've worked at traditional, let's say traditional product companies, Dropbox, Instacart. Now you're at OpenAI. What's what's maybe the most counterintuitive lesson you've learned about building products from your time at OpenAI?

每次换工作时，我总是试图选择最不同的、最大程度不同的工作。你知道，所以在 Dropbox 之后，我渴望一个现实世界的产品，因为它与做 SAS 等完全不同。嗯，在 Instacart 之后，我渴望做一些在智力上有趣的事情，你知道，有点激发我内心的书呆子气。你知道，所以我一直在寻找真正不同的东西。然后一旦我到了这些地方，我就试图理解是什么让那个地方成功，比如他们真正破解了什么，以及我们如何进一步利用这一点。我想我和 OpenAI 花了很多时间思考这个问题，尤其是在聊天之前，你知道那是一个没有实际意义的问题，因为我们没有太多收入或产品之类的。有几件事浮现在脑海中，推动了许多决策。嗯，一个是经验主义。我们之前谈过一点。事实是你只能通过发布来发现。嗯，这就是为什么 Max 和我倾向于这样做，这也是我们发布这么多东西的很大一部分原因。嗯，其中之一是，你知道，惊人的想法来自任何地方。嗯，经营研究实验室的事情是你真的不告诉人们研究什么。嗯，那不是你做的。我们继承了这种文化，即使我们成为一家研究和产品公司。所以，让有惊人想法的人去做事情，而不是成为一切的守门人或优先级制定者之类的，已经被证明对我们非常有价值，这也是很多创新的来源，是赋予任何职能的聪明人权力，嗯，所以这是我认为让 OpenAI 成功并让我们成功的良好继承。跨学科性，真正确保你把研究、工程、设计和产品放在一起，而不是把它们当作孤岛。我认为这就是让我们成功的东西，你可以在我们发布的每个产品中看到这一点。

Each time I always tried to pick the most different, maximally different job whenever I made a job change. You know, so after Dropbox I was like craving a real world product because it was just so different than working on SAS etc. Uh, and after Instacart I was craving working on something that intellectually was interesting um and had you know this kind of like sort of invoked the nerd in me. And you know, so I've always looked for things that are really different. And then once I showed up at these places I tried to understand what makes that place successful, like what is truly the thing that they cracked and how we can lean in into that even more. And I think I spent a lot of time thinking about this with OpenAI um, especially after chat before that you know it was kind of a moot point because we didn't really have much revenue or products or anything that you know like that. And there's a few things that come to mind that have driven many decisions. Um, one is the empiricism. We talked about that a bit. The fact that you can only find out by shipping. Um, which is why Max and I lean into that and that's, you know, huge part of why we ship so much. Um, one of them is that, you know, amazing ideas come from anywhere. Um, the thing about running a research lab is you really don't tell people what to research. Um, that's not what you do. And we inherited that culture even as we become a research and product company. So just letting people do things who have amazing ideas rather than sort of being the gatekeeper or prioritizer of everything or something like that um has been proven you know immensely valuable to us and that's where much of the innovation comes from is empowered smart people on any function really um so that was a good inheritance from what I think made OpenAI successful and makes us successful. The interdisciplinariness of really making sure that you put research and engineering and design and product together rather than treating them as silos. I think that's the thing that has made us successful and that you see come through in every product we ship.

比如，如果我们发布一个功能，但模型变聪明两倍，这个功能却没有变好两倍，那它可能就不该发布。嗯，当然这也不是绝对的。比如，袜子 2 就不会因为模型更聪明而变得更好，但我觉得对很多核心能力来说，这是个很好的试金石。所以，我一直觉得你真的要深入思考这个地方为什么成功，然后最大限度地加速它，因为这样才能把看似偶然的东西变成可复制的打法。

Like if you know we're shipping a feature and it doesn't get 2x better as the model gets 2x smarter, it's probably not a feature we should be shipping. Um you know not always true. You know sock 2 doesn't get better with uh you know threader models but you know I think for many of the core capabilities that's a good litmus test. So, I've always found you really have to lean into why is this place successful and then maximally accelerate that so to speak because um it's it's what allows you to turn something that feels like an accident into something that is a repeatable uh playbook.

你谈到了研究人员和产品人员之间的这种协作，而且你从 ChatGPT 的第一天就参与其中，到现在从零到 7 亿周活跃用户——不只是注册用户，而是周活跃用户。你这些年是怎么搭建这个团队的呢？

So, you talked about this kind of collaboration between researchers and product people and you've been at the beginning of chat GPT from day one to today from zero to 700 million weekly active users not just registered users weekly active users. How have you approached building out that team over time?

在研究实验室工作的另一个传承，就是你会非常认真地对待招聘。这是 AI 实验室都知道的事：每个人都至关重要。但很多科技公司在高速增长中会失去自我，失去人才标准，陷入混乱。所以我们一直倾向于保持精简。运营 ChatGPT 的其实是个小团队。我从 WhatsApp 那里得到启发，它就是用很小的团队运营一个全球规模的产品。更重要的是，你得把招聘看得更像高管寻访，而不是纯粹的流水线招聘——你需要真正理解每个团队要填补的空白是什么，需要什么具体的技能组合，以及怎么去填补。

One of the other inheritances of um being in a research lab is that you take recruiting really seriously. That's something that you know AI labs know. Every person matters. But many tech companies they go through hyperrowth and they kind of lose their identity. They lose, you know, their talent bars. They they they just kind of have chaos. Um so we've always had this tendency to run relatively lean. So it is a small team that is running chat GPT. Um I I take inspiration from WhatsApp where like you know it was a very small team running a very global scale product. Um and then more importantly I yeah I you know you have to treat hiring a little bit more like executive recruiting and less like just pure pipelineed recruiting where you really need to understand what is the gap you're trying to fill on each team. what is the specific skill set and how do you fill it?

举个例子，我本质上是个产品人，但有时候团队并不需要产品经理，因为已经有人在做这个角色了。比如在很多情况下，我们有一位非常有才华的工程负责人，他有很棒的产品直觉；或者我们有一位研究员，他有产品想法，在我看来他们就能扮演那个角色，而团队缺的可能是别的东西，比如多一点前端，或者类似的需求。另一些情况下，缺的可能是出色的数据科学家。所以我真的很喜欢逐个团队去分析，弄清楚那个团队需要什么样的技能组合，然后从原则出发去搭建，而不是想当然地认为我们只要为各种岗位做一堆流水线招聘，然后人们之后再找团队。所以我觉得这对我来说一直非常重要。这也是你既能保持团队精简，又能保持超高产出率的方法。

Um to give you an example, you know, I'm a product person at heart, but sometimes a team doesn't need a product person because there's already someone doing that role like like you know, in many cases we have a really talented engineering leader who has amazing product sense or we have a researcher who has product ideas and then and my mind they can play that role and maybe we have something else missing um instead like maybe we need like a little bit more front end um or something like that. In other cases, uh maybe what you're missing is an incredible data scientist. So, I really like to go through every single team and figure out what is the skill sets that that team needs and how do you put it together from principles rather than just assuming, hey, we're going to do like, you know, a bunch of pipeline recruiting for all these different roles and then, you know, people will find a team later. So, so I think that's always felt really important to me. Um, and it's the way that you keep your team really small yet super high throughput.

这还能让你招到那种——我记得 Keith Ro 管这叫“弹药桶”吧——他的意思是，你的产出取决于你有多少个桶，也就是那些能推动事情发生的人。你可以招这样的人，然后在他们周围加上“弹药”，也就是帮助他们的人。我觉得这对我们的招聘来说也很真实，我们总是尽量最大化那些有决策权、能交付的人的数量，因为这样才能用小团队干出大成绩。所以这是几点。另外，我也花了很多时间在团队氛围上，因为我觉得当你试图把研究和产品放在一起时，一个挑战就是文化不同，背景不同。我认为要让合作非常顺畅，你需要花时间做团队建设，确保人们对彼此的能力有极大的信任，觉得可以跨越边界思考。比如，我真的相信产品是每个人的事，正因为如此，招聘并不是把人招进来就结束了，实际上那才是开始，因为你要开始打造优秀的团队。

also allows you to hire people who I think Keith Ke Keith Ro calls this like like barrels I think um barrels of ammunition where he thinks I think I think this comes from him but um the idea being that sort of the throughput of your or depends on how many barrels you have um which is like people who can make stuff happen and I think you can hire um and then you can add ammunition around them um which is people helping those people and you know I I think that's been really true for our recruiting too where we try to maximize sort of the number of empowered people who can ship because that's how you have a small team and still get a ton done. So, those are a couple things. Um, and uh I spent a lot of time on like vibes too with like each team because I think one of the things that is challenging when you try to do research and product together is that the cultures are different. People have different backgrounds and um I think to make that go super well, you need to spend time team building and making sure that people have a huge amount of trust for each other's skill sets. um feel like they can think across their boundaries. Um like you know um I really believe that product is everyone's job for example and and and for that reason the recruiting sort of doesn't stop when the people are in the door it actually starts because you have to you know start making the teams awesome.

你有没有什么团队建设的小妙招可以分享？比如你用来营造……

Is there something you do with team building that would be fun to share just like something you do to create a

我就是喜欢和团队一起在白板上画图，喜欢进入那种生成式的思维状态。它能打破一切隔阂。所以这就是我尝试的方法。虽然不算特别有创意，但我发现它是一个通用工具——一旦你能让人们不再想“这是我的工作，那是你的工作”，而是想“我们都在一个房间里，一起解决某个问题”，那效果就太棒了。

I just love whiteboarding with teams like I just like like love getting into a generative mindset. It breaks down everything. So that's that's the thing that I I I try. not particularly creative, but I found it to be um a universal tool where the minute you can get people to stop thinking about, you know, what's my job versus the other person's job and more like, you know, we're all in a room like trying to crack something together. That is incredible.

你提到了“第一性原理”这个概念。其实我和很多人聊到你的时候，都会提到这个。你真的很看重这个吗？很多人都在谈第一性原理，但大多数人要么说“我不太懂”，要么觉得自己特别擅长从第一性原理思考。你能分享一下，从第一性原理思考到底是什么样的吗？有没有一个例子，你真正用了第一性原理，然后得出了意想不到的结论？

You mentioned this idea of first principles. This came up actually when I talk with a lot of people about you. Is this something you're really big on? A lot of people talk about first principles. Most people are like, I don't really understand like or they think they're amazing at thinking from first principles. Is there something you can share of just what it actually looks like to think from first principles? Maybe an example that comes to mind where you really went to first principles and came up with something unexpected.

是啊，这我可不敢自己说。是别人会这么说我，但你知道，这挺玄的。我觉得你真的要触及问题的本质，搞清楚你到底想解决什么。比如，就像我刚才说的招聘，我不会教条地认为你必须要有产品经理、工程经理、设计师之类的。我们只是想打造一个能交付的优秀团队。所以在那个情况下，第一性原理就是真正理解我们实际需要什么、缺什么，而不是套用以前学过的流程或行为。所以我觉得这是个好例子。

Yeah, this is not something I'd ever say about myself. I said someone else would say it, but um you know, it's a mysterious thing. Yeah, I think you just really got to get to ground truth on what you're really trying to solve. Like for example, as I mentioned with the recruiting thing, like I'm not dogmatic that you have to have a product manager and an engineering manager and a designer or whatever. We're just trying to make an awesome team that can ship. So in that case, first principles means just really understanding what we actually need and what we're missing rather than applying a previously um learned process or behavior. So you know, I think that's a good example.

另一个我认为在这个环境中运用第一性原理的好例子是：这个功能需要打磨得很精致吗？我们因为“模型选择器”挨了不少骂，我认了。我试着跟每个愿意听的人解释。可能有人不知道，模型选择器就是产品里那个巨大的下拉菜单，从传统意义上讲，它简直是任何好产品的反模式。但如果你真的从头推理一下：是等产品打磨好了再发布，还是先发布一个粗糙但可能不太合理的东西，然后开始学习、让用户用起来？我觉得一个流程很多、或者有很多习以为常行为的公司，会做出一个选择，那就是“我们发布时有质量标准，我们就按这个来”。

Another good example of of I think being first principles in this environment is is is you know does this feature need to be polished? You know we get a lot of crap for the for for the model chooser and I own it. Um I've tried to say that every to everyone who will listen. Um you know for those who don't know model chooser is this like giant drop down in the product that is like literally the anti-attern of any good product traditionally. But you know if you are actually reason from scratch of like is it better to wait until you've got a polished product or to ship out something raw even if it makes less sense and start learning and getting it into people's hands. Um I think a company with a lot of process or a lot of just you know learned behaviors will make one call which is know we have like a quality bar when we ship and that's what we do.

如果你从第一性原理出发，我觉得你会想：你知道吗，我们应该发布。虽然很尴尬，但严格来说，这比得不到你想要的反馈要好。所以，我认为在这个领域，从零开始处理每个场景非常重要，因为我们正在构建的东西没有类比。就像，你无法复制现有的事物。没有，你知道，我们是像 Instagram 还是像 Google 还是像生产力工具之类的。我不知道。但你可以从各处学习，但你必须从零开始。我认为这就是为什么这种特质往往能让一个人在 OpenAI 表现出色，这也是我们在面试中会考察的。

If your first principles about it, I think you're like, you know what, we should ship. It's embarrassing, but that's strictly less bad than not getting the feedback you wanted. So, I think just approaching each scenario from scratch is so important in this space because there is no analogy for what we're building. Like there's just you can't copy an existing thing. There's no, you know, are we like an Instagram or are we like a Google or like a productivity tool or something like that. I don't know. But you can learn from everywhere, but you have to do it from scratch. And I think that's why that trait tends to make someone effective at OpenAI and it's something we test for in our interviews, too.

所以这个主题不断出现，我认为重要的是强调你一直回到的一个点，那就是速度和打磨之间的权衡，以及在这个领域，速度不仅是为了保持领先，更是为了了解人们到底想用这个东西做什么。你认为关于为什么在 AI 领域需要如此快速行动，人们可能还忽略了什么吗？

So this theme keeps coming up and I think it's just important to highlight something that you keep coming back to which is this trade-off of speed and polish and how in this space speed is more important not just to stay ahead but to learn what the hell people actually want to do with this thing. Is there anything more that you think people just may be missing about why they need to move so fast in the space of AI?

是的，我的意思是，无聊的答案会是：哦，竞争激烈，每个人都在做 AI，他们试图互相超越。是的，我认为这可能是真的，但这不是我相信这一点的原因。真正的原因是，你会在这个领域打磨错误的东西。你绝对应该打磨，你知道，像模型输出之类的东西，但在发布之前你不会知道该打磨什么。我认为这在产品属性是涌现的、事先无法知道的环境中尤其如此。我认为很多人搞错了，因为最好的产品人往往是工匠。他们有传统的工艺定义。我也认为，很容易用我刚才所说的一切作为借口，最终不打造出伟大的产品。所以，我经常告诉我的团队，发布只是通往卓越之路上的一个点，你应该有意地选择那个点，它不必是你迭代的终点。它可以是开始，但你最好坚持到底。所以，我们一直在做很多工作，尤其是在过去一个季度，真正清理 ChatGPT 的用户界面。我很兴奋接下来要对响应布局和格式做同样的事情，因为一旦你知道人们在做什么，就没有借口不打磨你的产品。只是在你还不知道的世界里，你可能会非常分心。所以，这取决于情况。再说一次，你必须从第一性原理出发。但我确实认为，尤其是在早期，把速度当作工具，这在消费社交领域已经说过。这不是人们第一次说：嘿，你得尝试 10 件事，因为你可能会犯错。所以我不认为这是一种以前从未存在的动态。但我确实认为，对于 AI，内化这一点很重要。

Yeah, I mean the boring answer would be oh it's competitive and everyone's an AI and they're trying to out compete each other. Yeah, I think that may be true, but that's not the reason that I believe this. The reason really is that you're gonna be polishing the wrong things in the space. You absolutely should polish, you know, things like the model output, etc., but you won't know what to polish until after you ship. And I think that is uniquely true in an environment where the properties of your product are emergent and not knowable in advance. And I think many people get that wrong because like the best product people tend to be crafts people. And they have a traditional definition of craft. I also think it would be easy to use all what I just said as an excuse not to eventually build a great product. So, I often tell my teams that shipping is just kind of one point on the journey towards awesomeness and you should pick that point intentionally where it doesn't have to be the end of your iteration at all. It can be the beginning, but you better follow through. So, we've been doing a bunch of work, especially over the last quarter, of like really cleaning up the UI of ChatGPT. I'm really excited to do the same for the sort of the response layouts and formats next simply because once you know what people are doing, there's no excuse to not polish your product. It's just really in a world where you don't know yet, you might get very distracted. So, it's situational. Again, you kind of have to be first principles about it. But I do think using velocity especially early on as a tool you actually this has been said about consumer social for example. This is it's not the first space where people have said hey you just got to try 10 things because you're probably going to be wrong. So I don't think this is never existed before as a dynamic either. But I do think with AI it's important to internalize.

而且还有一个因素，模型在不断变化，所以你甚至可能意识不到它们能做什么。我想。

And there's also an element of the models are changing constantly and so you may not even realize what they're capable of. I imagine.

完全正确。模型在变化，改进它们的最佳方式，无论你是实验室还是只是做上下文工程或微调模型的人，也许你需要失败案例，真实的失败案例，才能让这些东西变得更好。基准测试越来越饱和。所以，你真的需要真实世界的场景，在那里你的产品或模型实际上没有做它应该做的事情。而获得这些的唯一方法就是发布，因为你会回到用例分布，你可以让那些东西变得更好。因此，这实际上是向你的团队，尤其是你的 ML 团队，阐明该攻克什么的最佳方式。就像，哦，你知道，人们试图做 X，而模型以某种方式失败了。现在，让我们让那些东西变得真正好。

Totally. The models are changing and the best way to improve them whether or not you're a lab or actually just someone who's doing context engineering or fine-tuning a model maybe you need failure cases, real failure cases to make these things better. The benchmarks are increasingly saturated. So really you need real world scenarios where your product or model is not actually doing the thing it was supposed to do. And the only way you get that is by shipping because you get back to sort of use case distribution and you can make those things good. And therefore, it's actually the best way to then go articulate to your team, especially your ML teams, what to climb on. It's like, oh, you know, people are trying to do X and the model's failing in ways why. Now, let's make those things really good.

关于失败案例这一点，让我想起 Kevin Weil 和 Mike Krieger 都分享过的一件事，那就是评估正在成为产品人员需要掌握的一项巨大新技能，因为现在很多产品构建都是评估。你有什么想分享的吗？

This point about failure cases makes me think about something that both Kevin Weil and Mike Krieger shared which is that evals are becoming a huge new skill that product people need to get good at because so much of product building is now evals. Is there something there you want to share?

我在 OpenAI 的整个旅程就是在一个个略有不同的情境中重新发现永恒的产品智慧和原则。所以我记得我在知道什么是评估之前就开始写评估了，因为我只是在为各种用例勾勒非常明确指定的理想行为，直到有人告诉我：“嘿，你应该做一个评估。”然后我意识到有一个完整的研究评估基准世界，与我试图构建的产品无关。我当时想：“哇，这可能是与做 AI 研究的人沟通产品应该做什么的通用语言。”这真的让我恍然大悟。归根结底，这与“在做任何事之前你应该阐明成功”的智慧没什么不同。这只是实现这一目标的新机制。但你可以在电子表格中做。你可以在任何地方做。我真的想为那些听到这个术语的人揭开它的神秘面纱，它不是你必须理解的某种技术魔法。它实际上只是以对训练机器人最有效的方式阐明成功。

My entire OpenAI journey has been this journey of rediscovering eternal product wisdom and principles in slightly new contexts. So I remember I started writing evals before I knew what an eval was because I was just outlining very clearly specified ideal behavior for various use cases until someone told me, "Hey, you should make an eval." And I realized there was this entire world of research evaluation benchmarks that had nothing to do with the product that I was trying to make. And I was like, "Wow, this might be the lingua franca of how to communicate what the product should be doing to people who do AI research." And that really clicked for me. And at the end of the day, it's not that different from the wisdom of you ought to articulate success before you do anything else. It's just a new mechanism for doing that. But you can do it in a spreadsheet. You can do it anywhere. And I really want to demystify it for people who hear that term like it's not some technical magic that you have to understand. It's really just about articulating success in a way that is maximally useful for training bots.

太棒了。我很快会发布一篇文章，为产品经理提供如何编写评估的非常好的指南。

Awesome. There's a I have a post coming out soon that gives you a very good how-to for PMs of how to write evals.

我很想读一读。我希望你同意我刚才说的话，因为也许其中有一些深层的东西。是的。

I would love to read it. And I hope you agree with what I just said because maybe there's something deep to it. Yeah.

是的。现在有所有这些工具让你更容易做到这一点。

Yeah. And now there's all these tools that make this easier for you.

完全同意。

Totally.

好的。所以这基本上支持了这一点，即这是产品团队和构建者需要掌握的一项非常重要的技能。

Okay. So this basically backs up this point that this is just a very important skill that product teams and builders need to get good at.

是的。是的。

Yeah. Yeah.

好的。还有几个问题。我知道你今天有很多事。一个是 ChatGPT 成为网站和产品流量增长的主要驱动力的趋势。例如，ChatGPT 现在给我的时事通讯带来的流量比 Twitter 还多，这完全让我震惊。我刚才在看我的统计数据。我想：“搞什么？这不是我知道会来的事情。”所以，我想就这个的未来谈谈你的想法，你如何看待 ChatGPT 推动产品和网站的增长和流量。

Okay. Just a few more questions. I know you have a lot going on today. One is that this trend of ChatGPT being a big driver of growth for traffic to sites for products. For example, ChatGPT is now driving more traffic to my newsletter than Twitter, which completely shocked me. I just was looking at my stats. I'm like, "What the hell? This is not something I knew was coming." So, just I guess thoughts on the future of this, how you think about just ChatGPT driving growth and traffic to products and sites.

我对此感到非常兴奋。因为你知道，就像我觉得通过聊天机器人与一切对话是反乌托邦的一样，我也觉得没有令人惊叹的高质量内容出现是反乌托邦的。

I'm really excited about it. Because you know in the same way that I find it dystopian to talk to everything through a chatbot, I also find it dystopian to not have amazing new high-quality content out there.

正因为如此，我之前谈到过搜索，以及它早期如何解决了一个非常重要的用户问题，因为你有知识截止的问题，突然之间你可以谈论任何事情。事后看来很明显，这不仅仅是用户问题，对吧？这是一个生态系统问题，最初的 ChatGPT 没有外链。它只会回答你的问题，让你留在产品里。即使你想继续阅读或深入探索，我们也没有办法把流量引导回内容生态系统。我对我们在搜索方面所做的工作感到非常兴奋，不仅仅是因为它给人们更准确的答案，还因为它能让我们把像这个播客这样的高质量内容呈现给想看的人。当然，还有很多有趣的问题，比如在谷歌时代，有搜索引擎优化，有明确理解的机制来提升排名和获取更多流量。所以我收到很多人的问题，比如，这相当于什么？如果我是 Lenny，我想让我的播客流量增长 10 倍，我到底需要做什么？事实是，我们在这方面没有很好的答案，因为理想情况下，吸引 AI 模型的方式应该和吸引真实用户的方式一样，因为模型应该代理用户的兴趣，而不是其他任何东西。至少我希望我们的产品是这样运作的。因此，我的建议非常老套，就是制作高质量的内容，这不像内容创作者理想中希望的那样具有可操作性。我认为这就是为什么我们还有更多工作要做，因为也许我们可以想出更好的机制或协议。但我很高兴这为你们带来了可观的流量，我希望其他制作优秀内容的人也开始有这种感觉，因为再说一次，这是一个非常巧妙的场景。

And for that reason, you know, I talked a little bit earlier about search and how that solved a really important user problem early on because you had this knowledge cut-off thing and suddenly you could talk about anything. It was very obvious in retrospect that it wasn't just a user problem, right? It was an ecosystem problem where the original ChatGPT didn't have outlinks. It would just answer your question and keep you in the product. Even if you wanted to keep reading or go deeper, there was no way for us to drive traffic back to the content ecosystem. I've been really excited about what we've been doing in search, not just because it gives people more accurate answers, but because it allows us to surface really high-quality content like this podcast to people who want to see it. And of course, there are so many interesting questions about, well, in the Google era, there was search engine optimization and clearly understood mechanisms of how to show up and get more traffic. So I get a lot of questions from people like, what is the equivalent of that? If I'm Lenny and I want to 10x the traffic to my podcast, what do I actually need to do? And the truth is we don't have amazing answers there, simply because the way to appeal to an AI model ideally is the same way that you would appeal to a real user, because the model is supposed to proxy the interest of the user and nothing else. At least that's how I want our product to work. And for that reason, my advice is super lame, which is make really high-quality content, which is not as actionable as I think people making content would ideally like. And I think this is why we have more work to do, because maybe there's a better mechanism or protocol that we could come up with. But I'm excited this is driving beautiful traffic for you, and I hope that other people making great content start to feel this way, because again, it's a very neat scenario.

人们一直在用两个缩写词来描述这种 AI 驱动的 SEO 技能。我想一个是 AEO，即答案引擎优化。另一个是 GEO。是不是……我忘了 G 代表什么。

There are two acronyms people have been using for this specific skill of AI-driven SEO. I think one is AEO, which is answer engine optimization. The other is GEO. Is that... I forget the G one.

生成式。

Generative.

生成式。对，AI 优化。这两个你更喜欢哪个？

Generative. Yeah, AI optimization. Do you have a favorite of those two?

不，不，我尽量回避这些术语，除非它们变得不可避免，因为我不完全确定这是否应该成为一个概念。再说一次，我认为理想情况下，ChatGPT 理解你的目标，因此理解什么内容对你感兴趣，内容创作者的工作是分享足够的信息和元数据，以便模型做出符合用户利益的决定。因此，我不确定是否应该给这个东西命名并把它变成一个概念。我非常渴望从内容创作者那里了解这可能是什么样子，因为再说一次，我们仍在摸索中。

No, no, I try to shy away from these terms unless they become inevitable, just because I'm not entirely sure if that should be a concept or not. Again, I think ideally ChatGPT understands your goals and therefore understands what content would be interesting to you, and the content creator's job is to share enough information and metadata about that content such that the model can make a user-aligned decision. Therefore, I'm not sure if giving this thing a name and making it a thing is what we should be doing or not. I'm very eager to learn from folks making content about what this could look like, because again, we're still working through it.

沿着这个思路，人们想到的另一个问题是，你们有 GPTs，这些是你可以构建的自定义 GPT 应用，用来回答非常特定的用例。总有一个问题是，你们会构建一个类似应用商店的东西，让我可以把我的新闻、我的产品接入 ChatGPT 并实现盈利吗？有没有什么你可以透露的、未来可能会推出的东西？

Along these lines, another question people think about is you have GPTs, which are kind of these custom GPT apps that you can build to answer very specific use cases. There's always this question of, are you going to build kind of an app store where I can plug in my news, my product into ChatGPT and monetize that? Is there stuff there that you could talk about that might be coming someday?

GPTs 很酷。它们在某种程度上超前于时代，因为我们在你真正能构建出差异化产品之前就提出了这个概念，至少在消费领域是这样。你知道，一个学习 GPT 会和模型开箱即用的能力非常相似。所以它主要是一种向人们阐述用例的方式。但它还没有足够的工具来做出一个像应用一样的东西，可以这么说。顺便说一句，在企业领域则不同。我们看到 GPTs 在那里被大量采用，因为每家公司都有非常定制化的业务流程和问题，而它是一个非常有用的工具。他们还有独特的数据可以连接到这些 GPTs 上，供其检索。所以我们在那里看到了很多成功。我认为这个想法是正确的，而且我认为我们会找到一个好的机制，因为当 AI 拥有如此强大的能力时，允许人们以清晰的功能、清晰的用例和彼此差异化来打包这些能力，会感觉非常强大。我也希望你能在 ChatGPT 上创业。我认为确实存在这样一个世界，当这个东西达到构建用户规模的阶段时，它可以为你带来分发。它可以让你开始创造一些东西，就像人们在互联网上构建一样，并且有全新的业务可以建立。所以我认为未来我们会在这方面分享更多。GPTs 是一次早期的尝试，我很高兴随着模型变得更好、我们的影响力扩大，能够在那里发展这个想法。

GPTs are cool. They're kind of ahead of their time in the sense that we built that concept before you could really build very differentiated things, at least in the consumer space. You know, a learning GPT is going to be pretty similar to what the model could already do out of the box. So it's mainly a way of articulating a use case to people. But it doesn't have enough tools yet to make something that feels like an app, so to speak. Different in the enterprise, by the way. We're seeing a ton of adoption of GPTs there because every single company has very bespoke business processes and problems, and it's a really useful tool there. They also have unique data that they can hook up to these things that it can retrieve over. So we've seen a lot of success there. I think the idea is the right one, and I think we're going to figure out a good mechanism for it, because when you have so much capability packed into AI, it feels really powerful to allow people to package that up in ways that have a clear affordance, a clear use case, and are differentiated from each other. I also would love it if you could start a business on ChatGPT. I think there really is a world where, as this thing hits building user scale, it can get you distribution. It can get you started on making something in the same way that people built on the internet, and there were entirely new businesses to be built. So I think we'll have more to share there in the future. GPTs was an early stab, and I'm just excited to evolve the thinking there as the models get good and our reach increases as well.

太棒了。这真的很酷。我很期待看到你们在那里做的事情。好的，完全不同的方向。我知道你大学时学过哲学。

Amazing. That is really cool. I'm really excited to see what you guys do there. Okay, completely different direction. Something that I know about you is you studied philosophy in college.

是的。

I did.

计算机科学和哲学，对吧？一个组合。

Computer science and philosophy, right? A combo.

是的。我一开始是哲学专业，因为我很喜欢逻辑，而编程和逻辑最相似，所以选修了一门编程课。然后我爱上了编程，最终爱上了计算机科学，并且我一直在做越来越多相关的事情。但在那之前，我从未真正认为自己是一个技术型的人，所以这算是我生命中一个较晚的发现，我对此非常感激。

Yeah. I started as a philosophy major, and took one coding class because I really liked logic, and programming was most similar to that. Then I fell in love with coding, and eventually computer science, and I just kept doing more and more of it. But until then, I never really thought of myself as a technical person, so it was kind of a late discovery in my life that I'm very grateful for.

对于领导这个产品的人来说，这是一个多么不可思议的组合。

What an incredible combination for someone leading this product.

确实如此。这真的以一种我无法预料的方式回到了原点。你需要处理的问题数量真的非常有趣，哲学不是一种传统上实用的技能，但它确实教会你从头开始思考问题，并阐述观点。我认为这已经多次派上了用场。

It's true. It is really coming full circle in a way that I couldn't have predicted. The amount of questions you have to grapple with are truly super interesting, and philosophy is not a traditionally practical skill, but it does really teach you to think things through from scratch and to articulate a point of view. I think that has come in handy numerous times.

有没有哪个特定的哲学家或学派对你最有帮助，还是说更多是一种普遍的……

Is there a specific philosopher or school that has been most handy to you, or is there more just a general...

太多了。我的毕业论文写的是理性的人能否以及为何会意见分歧，这在很多价值观不同的人对你的模型行为或事物应该怎样运作发表看法时也很有用。所以我非常喜欢 20 世纪的分析哲学家。这有点书呆子气，但我不确定有没有最喜欢的。太多了，数不过来。

There are so many. I wrote my senior thesis on whether and why rational people can disagree, which also comes in handy when a lot of people with very different values have opinions on your model behavior or on how things should work. So I really like 20th-century analytical philosophers. It's kind of nerdy stuff, but I don't know if I have a favorite. It's too many to count.

嗯，但这就是我喜欢的东西。其中一些内容相当分析性，比如你设定 P 为某种爱情理论，Q 为另一种爱情理论，然后进行某种符号操作。所以这既是一种脑力思维练习，甚至比实用更偏向于此。但它教会了我一种思考方式，这种方式至今仍然很有价值。

Um, but um, that's the kind of stuff I like. And some of it ends up being quite analytical, like you have, let P be this theory of love and let Q be, you know, this other theory of love, and then you do some sort of symbolic manipulation. So it is just as much a brain thought exercise as it is, or is much more that than practical. But it taught me how to think in a way that continues to be pretty valuable.

太棒了。多么酷的技能和背景组合。嗯，在我们进入非常激动人心的快问快答环节之前，最后一个问题。你曾是 Dropbox 的产品负责人，然后是 Instacart，现在你是历史上最具影响力的产品的产品经理。你是怎么得到这个职位的？加入 OpenAI 并承担这项工作的故事是怎样的？

Incredible. What a cool combo of skills and background. Uh, last question before we get to your very exciting lightning round. So you were a product leader at Dropbox, then Instacart, now you're the PM of arguably the most consequential product in history. How did you land in this role? What was the story of joining OpenAI and taking on this work?

我做出的每一个职业决定，包括大学毕业后的第一个决定，都是弄清楚我认识的最聪明的人是谁，我想和他们一起学习，以及我能否与他们共事。我不知道如何选择公司。我也不知道如何真正逻辑地思考，比如哪个领域会起飞之类的。但我确实觉得我对人有一种直觉。比如，在 Dropbox，我跟随了我担任助教的那门课的首席助教。在 Instacart，我跟随了一些我认识的最聪明的产品人。而在 OpenAI，招募我的人是 Joanne，我给她发消息想从 Dolly 的等待名单中出来，她说除非你来面试。所以她把这变成了一种反向招募。起初，老实说，我不知道我在这里能做什么，因为这是一个研究实验室，而我是产品人。他们说，别担心，我们会搞定的。他们当时是在保密，我以为他们是在保密，因为这是 OpenAI，他们不能分享任何东西。但他们保密是因为我们当时确实还不知道。所以我来了，我几乎什么都做，那肯定不是产品工作。我记得我的第一个任务是修百叶窗之类的。然后我开始给人发保密协议，因为他们需要一些运营帮助。然后我开始问，等等，我为什么要发保密协议？哦，这样我们才能和用户交谈。我当时想，和用户交谈？那听起来像是我知道怎么做的事情。于是我很快就偶然进入了产品工作。然后最终，我领导了很多产品工作，但这是自然而然的，只是出现并做必须做的事情。因为再说一次，我加入的公司绝不是一个产品公司。

Every single career decision I ever made, including my first one out of college, was just figuring out who are the smartest people I know that I want to hang out with and learn from, and can I work with them? And I don't know how to pick companies. I don't know how to really logically think through, you know, what space is going to take off or something like that. But I just do feel like I have a sense on people. And, you know, for Dropbox, I followed the head teaching assistant for a class that I was TAing. And, you know, for Instacart, I followed some of the smartest product people I knew. And for OpenAI, the person who recruited me, Joanne, I had messaged her about getting off the Dolly waitlist, and she said only if you interview here. So she kind of turned it into a reverse recruiting thing. And initially, honestly, I didn't know what I would do here because it was a research lab and I was a product person. And they said, you know, don't worry, we'll figure it out. And they were being cy, and I thought they were being ky because it's OpenAI and they can't share anything. But they were being cy because we actually just didn't know yet at the time. So I showed up and I kind of did everything under the sun, and it definitely wasn't product. You know, I think my first task was like fix the blinds or something like that. And then, you know, I started sending out NDAs for people because they needed some operational help. And then, you know, I started asking, wait, why am I sending out NDAs? Oh, so we could talk to users. And I was like, talking to users? That sounds like the thing I know how to do. And I quickly stumbled into doing product work. Um, and then eventually, you know, leading a bunch of product work, but it was organic by just showing up and doing what had to be done. Because again, the company I joined was not a product company by any means.

哇。嗯，这是一个很好的例子，我不知道你是否这样想，但当有人给你火箭飞船上的座位时，不要问是哪个座位。嗯，也许吧。

Wow. Uh, this is such a good example of, uh, I don't know if you think of it this way, but when someone offers you a seat on a rocket ship, don't ask which seat. Uh, maybe.

我当时不知道那是火箭飞船。我只是觉得，我有点被“书呆子狙击”了，我是这么形容的。或者，你知道，当我为让你从 Dolly 等待名单中出来的对话做准备时，我真的开始阅读这个领域的资料，这激发了我的哲学大脑，然后也激发了计算机科学大脑。我当时想，等等，这很酷。然后我开始阅读那个时代的所有学术论文。所以这只是知识上的痒和人的吸引。但后来我留下来是为了产品机会，显然。在 ChatGPT 之后，当它起飞时，我意识到我们建造了一艘火箭飞船。我们是在建造它的同时发射的，也许这个比喻。但我不能说当我加入时感觉这是一份被炒作的工作或类似的东西。

I didn't know it was a rocket ship. I just thought it was, I kind of got nerd sniped, is what I would describe it as. Or like, you know, as I prepared for the conversation to get you off the Dolly waitlist, really. I just started reading about the space, and that peaked the philosophy brain, and then also actually the computer science brain. I was like, wait, this is cool. And then I started reading all the academic papers of that era. And so I just, it was intellectual itch and the people. But then I stayed for the product opportunity, obviously. I, you know, post ChatGPT, when that took off, realized that we'd built a rocket ship. Uh, where we launched it while building it, uh, maybe this analogy. Uh, but I can't say that it felt like a hyped job or anything like that when I joined.

所以，那里的一个教训是跟随，如你所说，跟随最聪明的人。还有一条线索是跟随你感兴趣的事情。只是你玩 Dolly 就带来了这个机会。

So, kind of a lesson there is follow, as you said, follow the smartest people, you know. There's also just this thread of follow things that are interesting to you. Just you playing with Dolly led to this opportunity.

是的。是的。实际上，这是我们仍然在测试的东西。好奇心是不是一个我们认为比机器学习知识重要得多的特质。嗯，我不是在评论研究招聘。我认为你确实需要一些机器学习知识，恐怕是这样。但对于产品、工程和设计人员，以及这类职能，我实际上认为如果你只是对事物如何运作感到好奇，那么你以前是否做过完全无关紧要。事实上，如果你过滤那些以前做过的人，你会得到一个非常狭窄的、非常幸运的人选，而不一定是你能得到的最好人选。所以我认为我们已经扩展了这一点。这当然让我走到了这里，但我认为它实际上普遍是 OpenAI 成功的良好预测指标。

Yeah. Yeah. And actually, that's something we still test for. Is curiosity like an attribute that we think matters so much more than your ML knowledge. Um, you know, I'm not making a comment on research hiring. I think you do need some ML knowledge, I'm afraid. But you know, for product and engineering and design people, and those kinds of functions, I actually think that if you are just curious about how stuff works, it doesn't matter at all if you've never done it before. In fact, if you were to filter for people who have done it before, you would have a very narrow filter of very lucky people rather than necessarily the best people you can get. So I think we've scaled that. Certainly what got me here, but I think it's actually just generically been a good predictor of success at OpenAI.

Nick，我告诉过你我有十亿个问题。我说过我有二十亿个问题要问你。我觉得我已经问了很多。我觉得我还有十亿个问题，但我知道你告诉我在这之后你有一个重要的 GPT-5 检查要做。

Nick, I told you I had a billion. I said I had two billion questions to ask you. I feel like I've asked a lot. I feel like I still have a billion left, but I know you told me right after this you have a big GPT-5 check-in that you got to get to.

我们有一艘船。

We got a ship.

我们现在有了一艘更好的船，因为这次录制了并要发布。

We got a better ship now that this is recorded and we're putting this out.

这是真的。

This is true.

这是强制功能。好的。那么，在我们进入非常激动人心的快问快答环节之前，你还有什么想分享的，留给听众的，认为重要的吗？

This is the forcing function. Okay. So, before we get to very exciting lightning round, is there anything else that you want to share, leave listeners with, think is important to share?

我试着分享一些我做决定的方式，因为我希望，我离毕业不远。我非常能理解那些刚进入就业市场、试图弄清楚现在该做什么的人。我非常确信，如果你身边都是给你能量的人，并且你追随自己真正好奇的事物，你在这个时代会成功。所以我给人们的临别建议是，让自己身边都是优秀的人，做你真正热爱的事情，因为在一个这个东西能回答任何问题的世界里，提出正确的问题非常非常重要。而学会这样做的唯一方法就是培养你自己的好奇心。所以，嗯，这对我有效，这是我能分享的唯一可重复的事情。其他一切都是运气。

I try to share a little bit about how I made decisions because I hope to, I'm not that far out of school. I like relate a lot to people who are coming in the job market who are trying to figure out what to do with their life right now. And I feel very confident that if you surround yourself with people that give you energy and if you follow the things you're actually curious about, that you're going to be successful in this era. So my parting advice to folks really is put yourself around good people and do the things you're actually passionate about, because in a world where this thing can answer any question, asking the right question is very, very important. And the only way to learn how to do that is to nurture your own curiosity. So, um, it worked for me, and it's the one repeatable thing that I can share. Everything else is luck.

这与现在很多人的做法相反，那就是追随金钱。我在哪里能赚最多？我如何发展这个东西并赚到 1 亿美元？就像所有那些得到疯狂报价的人，他们并不打算通过做这个赚很多钱。

And this is counter to what a lot of people are doing right now, which is follow the money. Where can I make the most? How do I grow this thing and make $100 million? Like all these people that are getting these crazy offers were not planning to make a lot of money doing this.

看到这些事情发生很有趣，因为我认为所有这些人都出于真诚的原因进入学校。他们对这个领域感到兴奋。他们在研究它。他们在追求知识，我很高兴这得到了回报。

It's quite interesting to see that stuff play out because I think all these people entered, you know, school for genuine reasons. They were like excited about the space. They were researching it. They were pursuing knowledge, and I'm happy that that's being rewarded.

而且我不知道未来的回报会是什么样，尤其是在后 AGI（通用人工智能）时代，但我就是有一种感觉，如果你遵循那个建议，你最终会没事的。

And I don't know what the rewards will look like in the future, especially in a post-AGI world, but I just have a feeling that if you follow that advice, you'll end up okay.

那么，Nick，我们到了非常刺激的快问快答环节。我有五个问题要问你。准备好了吗？

With that, Nick, we've reached our very exciting lightning round. I've got five questions for you. Are you ready?

当然。

Sure.

好的。

Yeah.

在产品领域，你发现自己最常向别人推荐的两三本书是什么？

What are two or three books that you find yourself recommending most to other people in the product space?

可能是像《高产出管理》或《设计心理学》这样的经典书籍，因为我觉得它们非常实用。

Probably things like High Output Management or The Design of Everyday Things, or those kind of classic type things, because I think they're extremely applicable.

我们聊过哲学。不知道有没有你喜欢的哲学书？如果你要入门，这本值得一读……

We talked about philosophy. I don't know, is there a philosophy book you like? Here's the one to read if you're getting...

天哪，像罗尔斯或诺齐克的作品。我喜欢政治哲学类的东西，真的很有趣。我觉得我会推荐这类书。虽然读这些没什么实际用处，但我会跟你聊得很投入，所以后果自负。

Oh man, like anything by Rawls or Nozick. I like the political stuff. It's really fun. That's the type I think I recommend. I don't think there's a practical reason to read that stuff, but I will nerd out about it with you, so at your own peril.

如果你有时间看点什么的话，最近有没有特别喜欢、特别享受的电影或电视剧？

Do you have a favorite recent movie or TV show you've really enjoyed, if you've had time to watch anything?

我觉得在这个领域你得看点科幻。你不应该模仿它们，但可以从中学习。所以我经常重看《她》和《西部世界》。《人生切割术》也很棒。这些就是我有时间时会琢磨的东西。

I think you got to do a little bit of sci-fi to be in this space. You shouldn't copy any of it, but I think you learn from it. So, regularly rewatch Her and Westworld. Severance was great. That's the stuff that, when I have time, I'll meddle with.

太棒了。在所有科幻电影里，这两部是你最有共鸣、觉得最有趣、最有价值的。

That is awesome. I love that those are the two of all the sci-fi movies. Those are the ones you resonate most with and find most interesting and valuable.

是的，但这可能是我自己的局限。我相信还有更多值得发现的。

Yes, but that's probably my own limitation. I'm sure there's more to discover.

顺便问一下，你读过《深渊上的火》这本科幻小说吗？

By the way, have you read A Fire Upon the Deep, a sci-fi book?

嗯……

Um...

好吧。我不知道你有没有时间读这本书，但我觉得你会喜欢的。这是一本非常棒的、以 AI 为主题的科幻太空歌剧类小说。

Okay. I don't know if you have time to read this book, but I think you would love it. It's such a good AI-oriented sci-fi space opera sort of book.

太好了。

Great.

好的。

Yeah. Okay.

你最近有没有发现特别喜欢、特别钟爱的产品？

Do you have a favorite product you recently discovered that you really love?

其实没有。我已经处于极度饱和状态。有时候挺有意思的，比如 API 开发者会问我：“嘿，你们是不是要抄袭我们所有的产品？”但实际上我真的没时间去关注 OpenAI 之外的事情，因为这里的节奏太紧张了。所以恐怕我给不了你什么好的推荐。

I actually don't. I am at extreme capacity. It's kind of interesting sometimes, like API developers ask me, 'Hey, are you going to copy all of our products?' But I actually just do not have time to follow up on what's going on outside of OpenAI, because the pace here is so intense. So I don't have good recs for you, I'm afraid.

我觉得这对很多产品公司来说真是个令人安慰的回答。好吧，Nick 连看我们产品的时间都没有。天哪。好吧。

That's a really comforting answer, I think, to a lot of product companies. Okay, Nick has no time to even look at our stuff. Oh man. Okay.

你有没有最喜欢的人生格言，在困难时常用，也会分享给朋友或家人，而且别人也觉得有用？

Do you have a favorite life motto that you find yourself using when things are tough, sharing with friends or family, that other few people find useful?

“你就是你花时间最多的五个人的平均值”，这是我真正内化的一个原则，无论是在个人生活中，那里有给我能量、让我振作、让我变得更好的人。我的未婚妻就是其中之一，但我生活中还有很多这样的人。在工作中也有类似的情况，而我所有的职业决策都是基于这个原则。就像，我想向谁学习？所以我一直在应用这个原则。

Being the average of the five people you spend the most time with is a thing I really internalize, both in my personal life, where there are people who give me energy and lift me up and make me a better person. My fiancée is one of those people, but there are many people in my life. But then at work there's the equivalent, and that's how I've made all my career decisions. It's like, who do I want to learn from? So I apply that principle constantly.

最后一个问题。所有跟我聊过的人都说你是一位非常出色的爵士钢琴家。你赢得过比赛。我想你原本打算以此为生，但后来你不知怎么走了条支线任务。

Final question. Everybody I talked to told me that you are a very good jazz pianist. You have won competitions. I think you were planning to do this as your main thing, and then you somehow took the side quest.

是的，我在最后一刻退缩了，但我本来打算去读音乐学校，这仍然是我希望的第二篇章。

Yeah, I chickened out at the very last minute, but I was going to go to school for music, and that's still my hopefully chapter two.

我喜欢这个。那可能还会发生。

I love that. That might still happen.

可能还会发生。现在我参加了一些玩票性质的乐队，我们会时不时即兴演奏。这是我在极度疲惫、无法思考时唯一能做的事情，因为它能很好地平衡我的状态。但希望未来我能在这方面投入更多。

Might still happen. Now I'm in some for-fun bands, and we will jam from time to time. It's the one thing I can do when I'm otherwise super tired and can't think anymore, because it balances me out in good ways. But hopefully I'll get to do more of it in the future.

你觉得音乐和你的工作之间有什么类比吗？有什么发现吗？

Is there any analogy between music and your job? Anything that you find?

是的，实际上。我觉得你可以把软件开发，或者做产品，想象成要么是管弦乐队的指挥，要么是爵士乐队的一员。而我把它看作爵士乐队。我不相信每个人都必须演奏固定部分、由我指挥何时演奏这种想法。我喜欢爵士乐或其他即兴音乐中那种相互呼应的感觉。你听别人演奏了什么，然后你回应一段。我认为伟大的产品开发就是这样，想法可以来自任何地方。它不应该是一个照本宣科的过程。你应该尝试新事物，享受乐趣，在工作中保持玩乐的心态。所以我经常用这个类比来跟喜欢音乐的人交流，它往往能引起共鸣。

Yeah, actually. I feel like you could think of software development, or being a product person, as you could be a conductor of an orchestra, or you could be in a jazz band. And I think of it as a jazz band. I don't believe in the idea of everyone having this set part that they have to play, and me telling people when to play. I love how, in jazz or other forms of improvised music, you're kind of riffing off each other. You listen to what one person played, and then you play something back. And I think great product development is like that in the sense that ideas could come from anywhere. It shouldn't be a scripted process. You should be trying stuff out, having fun, having play in what you do. So I use that analogy a lot for those who like music. It tends to resonate.

Nick，我非常感谢你抽出时间来做这个访谈。我知道今天已经很疯狂了。今天，明天对全世界来说会更加疯狂。他们完全不知道接下来会发生什么。非常感谢你来做这个。最后两个问题。如果大家想在网上找到你，他们可以去哪里？大家可能在哪里找到 GPT-5？然后，听众怎样才能对你有用？

Nick, I am so thankful that you made time for this. I know today is insane. Today, tomorrow is going to be even more insane for the entire world. They have no idea what's coming. Thank you so much for doing this. Two final questions. Where can folks find you if you want them to find you online? Where can folks find GPT-5 potentially? And then just how can listeners be useful to you?

直接用产品就行，甚至不用付费。从明天开始它应该是你的默认模型。只管用，别再想模型的事了。除非你想，而且你是重度用户，那样的话你可以看到所有小模型。所以放心好了。至于有用？说实话，我从广大用户和 ChatGPT 用户那里学到了很多。所以继续做你的事就好。我在观察和学习，我感激所有反馈。所以我相信在我们修好模型选择器之后，你们会为别的事情吐槽我，我会接受的。所以继续来吧。

Just use the product. You don't even have to pay. It should be your default model starting tomorrow. Just use it and don't think about models anymore. Unless you want to and you're a power user, in which case you get all little models. So rest assured. And useful? Honestly, I learn so much from people at large and ChatGPT users, etc. So just keep doing your thing. I'm watching and learning, and I appreciate all the feedback. So I'm sure after we fix the model chooser, you guys will roast me for something else, and I'll take it. So keep it coming.

太棒了。Nick，非常感谢你来做客。

Amazing. Nick, thank you so much for being here.

谢谢你邀请我，Lenny。

Thanks for having me, Lenny.

祝你明天好运。

And good luck tomorrow.

谢谢。大家再见。

Thanks. Bye everyone.

非常感谢你的收听。如果你觉得这期有价值，可以在 Apple Podcasts、Spotify 或你喜欢的播客应用上订阅本节目。另外，请考虑给我们评分或留下评论，这真的能帮助其他听众发现这个播客。你可以在 lennispodcast.com 找到所有往期节目或了解更多关于本节目的信息。下期再见。

Thank you so much for listening. If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. Also, please consider giving us a rating or leaving a review, as that really helps other listeners find the podcast. You can find all past episodes or learn more about the show at lennispodcast.com. See you in the next episode.