外包智能但不能外包理解：Logan Kilpatrick 谈 AI 速度与保持理智

有一句很棒的话，出自 Andre Karpathy：你可以外包智能，但无法外包理解。我觉得当前的模型和系统在帮你保持这种理解水平方面做得还不够。我观察得越久，越确信与模型实验室竞争并没有那么难。模型实验室做了最难的——打造出色的技术供你在此基础上构建，但他们无法捕获所有客户，这是不可能的。因为变化如此之大，我觉得任何人都有机会做出点东西来。我现在会很兴奋地去创办一家编程初创公司。

There's a great quote from Andre Karpathy: you can outsource intelligence, but you can't outsource your understanding. I don't think current models and systems do a good enough job helping you maintain that level of understanding. The longer I've watched this play out, the more convinced I am that competing with the model labs isn't that hard. The model labs do the hard part of building great technology for you to build on top of, and they can't capture all the customers. It's not possible. Because things are shifting so much, it feels like anyone has a real shot at building something. I'd be really excited to start a coding startup right now.

在今天的《First Time》播客中，我们请到了一位几乎从生成式 AI 热潮一开始就身处核心的嘉宾。Logan Kilpatrick 最初为 NASA 编写软件，帮助规划月球车的行驶路线。之后他去了 Apple，将计算机视觉模型部署到全球所有 Apple Store。他曾是 Path AI 的技术倡导者，Daphne 和北美 Django 活动基金会的董事会成员，并一度是 Julia 编程语言的公众面孔。2022 年，OpenAI 向他发出邀请，Logan 成为他们的第一位开发者关系员工。他见证了 OpenAI 从 100 人成长到 1,500 人，期间协助推出了 GPT-4、DALL-E 3、Assistants API 等。2024 年 4 月，Logan 加入 Google，现在领导 Gemini AI Studio、Gemini API 和 Kaggle，作为 Google DeepMind 的一部分。如果你在过去几年中用 AI 做过任何东西，那么你已经熟悉了 Logan 帮助塑造的生态系统。Logan Kilpatrick，非常荣幸欢迎你来到《First Time》播客。

On today's episode of the First Time Podcast, our guest has been at the epicenter of the generative AI boom almost from the very beginning. Logan Kilpatrick started out writing software for NASA, helping plan traversal routes for a lunar rover. He went to Apple, where he shipped computer vision models into every Apple store in the world. He has been a tech advocate with Path AI, board member of Daphne and the Django Events Foundation North America, and was the public face of the Julia programming language for a while. Then in 2022, OpenAI came calling and Logan became their very first developer relations hire. He saw OpenAI grow from 100 to 1,500 people and helped launch GPT-4, DALL-E 3, the Assistants API, and much more. In April 2024, Logan moved to Google and now leads Gemini AI Studio, Gemini API, and Kaggle as part of Google DeepMind. If you've built anything with AI in the past few years, you're already familiar with the ecosystems Logan has helped shape. Logan Kilpatrick, it is an honor to welcome you to the First Time Podcast.

谢谢你邀请我。我很喜欢。现在我得对得起你这段慷慨的介绍，我会尽力而为。

Thank you for having me. I love it. Now I need to live up to this very generous intro. I'll do my best.

我们开始吧。我有个好奇的问题：AI 发展得太快了，和我聊过的每个人，无论他们是在最前沿构建，还是在远离 AI 的某个角落，都觉得跟不上。你在 AI 之外做些什么来保持头脑清醒，不被它完全吞噬？

Let's start. A curiosity of mine. AI is evolving so fast, and everyone I talk to, whether they're building at the bleeding edge or somewhere tucked away outside AI, feels like they're not keeping up. What do you do outside of AI to stay sane and not get overly consumed?

我想深入探讨一下为什么人们感觉跟不上变化。昨天我和别人聊天时分享了一个假设：如果每个人都只工作四天，全球的 AI 采用率可能会更高。这不是显而易见的联系，但构建东西和探索工具确实需要时间。现在大家都很忙，脚下的平台正在转变，很难有空间保持在前沿。所以我必须刻意留出时间。在 AI 的混乱之外，我刚搬回加州，花时间保持健康，和女朋友一起做园艺。种很多植物，避免晒伤，出去摸摸草地，享受 AI 之外的世界。

Yeah, I think we should double click on why people feel they can't keep pace. I was in a conversation yesterday and shared a hypothesis: if everyone worked only four days a week, we'd probably have greater AI adoption. That's not an obvious connection, but it takes time to build things and explore tools. Everyone is so busy with the platform shift happening beneath our feet that it's hard to find space to stay on the forefront. So I have to carve out intentional time. Outside the AI chaos, I just moved back to California and spend time staying healthy, gardening with my girlfriend. Lots of planting, trying not to get sunburned, going out to touch grass and enjoy the world outside AI.

你真的得去摸草地，对吧？

You literally have to touch grass, right?

真的去摸草地，没错。我们后院有一小块草地，还有一台手推式割草机，不是电动的。所以不仅摸草，还要割草，算是很好的锻炼。

Literally touch the grass, yeah. We have a little patch of grass in our backyard and we have a push mower that's manual, not electric. So lots of good exercise trying to not only touch the grass but cut it as well.

不错。希望 AI 还要一段时间才能影响到生活的那个部分。

Nice. Hopefully it'll be a little bit before AI impacts that part of the world.

我觉得 AI 不会影响这个。说实话，我纯粹是出于热爱才这么做的，所以坚持手动。我很想要个小机器人——屋里我们用扫地机器人，因为我不想拖地，但室外已经有割草机器人了。我们手动割草是为了乐趣，这很棒。

I don't think AI will impact that. I think I'm doing it for the love of the game, honestly. That's why we're doing it manually. I would love a little robot—we have robot vacuums inside because I don't want to vacuum the floor, but outside you can already get robot lawnmowers. We're doing it manually for fun right now, which is great.

回到刚才的话题：我在科技行业很久了，听过人们谈论追赶，但大多数人都找到了办法。现在是什么让大家这么难？

So jumping back to the same theme: I've been in technology for a long time. I've heard people talk about trying to keep up, but for the most part they found ways. What is it about right now that's giving people a hard time?

我认为进步的速度非常快，但这并非挑战的全部。另一个方面是实际上有太多机会。我尝试这样看待：工程工作、产品工作和研究发现之间的界限非常模糊。很多修补实际上就是研究发现。你有机会作为个体发现模型中有趣的能力搁置（capability overhang），或者找到构建框架的方法。一个很好的例子是 Peter Steinberger，他创建了 Open Claw。他在一个想法上修补，存在能力搁置，他搭建了框架和产品体验，将它们串联起来，创造了 Open Claw，从而引发了个人 AI 智能体的爆发。我认为他实际上是在做研究，尽管人们不这么理解。归根结底，探索式研究非常有趣，但它是另一种能力。不只是探索，真正有趣的东西更像是研究探索。

I think the pace of progress is very fast, but that's less of the challenge. The other flavor is that there's actually so much opportunity. I've tried to look at it this way: the boundary between engineering work, product work, and research discovery is very blurry. A lot of tinkering is actually research discovery. The chance that you as an individual might discover some interesting capability overhang in the models, or a way to build a harness, is real. A great example is Peter Steinberger, who created Open Claw. He was tinkering with an idea, there was capability overhang, he built the harness and product experience, strung it together, and created Open Claw, which ignited the personal AI agent explosion. I think he was actually doing research, though that's not how people frame it. Ultimately, exploration research is a lot of fun, but it's a different muscle. It's not just exploring; the really interesting stuff is more research exploration.

你如何跟得上今天发生的这一切？你就像是站在前沿，对吧？你是如何跟上所有进展的？

How do you keep up with everything going on today? You're like creating the edge, right? How do you stay step today with all of it?

是的，这确实是我的一个优势：在谷歌内部有很多这类事情发生，所以我能看到早期步骤。但即便如此，世界上也有很多不是由谷歌推动的事情，跟上这些进展非常重要，要对我们在创造和构建的东西有更深入的理解。真正帮到我的是，很多试图跟上进度的人都在强迫自己去做，因为人人都说如果不这样做就会落后。我认为这是错误的方式。我尝试的方式是，我有一堆宠物项目，它们看起来越来越像微型初创公司，只是我没有变现。它们解决我感兴趣的问题，以一种我感兴趣的方式摆弄。这给了我一个极好的试验场，可以把最新的技术、工具、技术和思维直接应用到我真正喜欢花时间做的事情上。这就是魔力所在：如果你能找到自己真正感兴趣的事情，并不断摆弄它，而且这是一个足够大的问题空间，可以应用很多不同的东西——新模型、新工具——直接解决这个问题。这非常有趣。我喜欢这个模式；它让我一直保持动力。我会感到内疚，因为可以构建的东西太多了，但很容易，因为我真的喜欢做这件事。这不是强迫自己。

Yeah, it's definitely a nice advantage that I have: inside of Google there's lots of this stuff happening, so I get to see the early steps. But even still, there are things happening in the world that aren't driven by Google, and staying on top of those is super important, having an even deeper understanding of what we're creating and building. What's actually helped me is that a lot of folks who try to keep up force themselves to do it, because everyone says if you're not doing it you'll fall behind. I think that's the wrong way. The way I've tried to do this is I have a bunch of pet projects that look more and more like miniature startups that I'm just not monetizing. They solve problems I'm excited about, tinkering in a way I'm excited about. It gives me an incredible test bed to apply the latest technology, tools, techniques, and thinking directly to something I genuinely enjoy spending time on. That's the magic: if you can find something you're really excited about and keep tinkering on it, and it's a repeatable problem space big enough to apply many different things—new models, new tools—directly to that problem. It's a lot of fun. I like this model; it keeps me going. I feel guilty because there's so much I could be building, but then it's easy because I actually enjoy doing it. It's not forcing myself.

从和其他人的交流来看，这似乎是正确的方式——追随你的热情，追随你的兴趣，利用你热情所在的任何领域的最新成果，然后让这一切自然地带着你走。你刚才分享的时候，我脑中浮现了一个潜水的画面：你在水下探索并享受其中，而不是站在水面上试图弄清楚如何下水。

That's what it seems like from talking to others—the way to go is to follow your passion, follow your interests, leverage whatever is latest in whatever category you're passionate about, and let that naturally take you where it takes you. As you were sharing, I had a scuba diving vision where you're down there exploring and enjoying it, versus standing above the surface trying to figure out how to get into it.

我认为你还需要一些有野心的事情——一些稍微超出你能力范围的东西。我对一些项目就是这种感觉。不是全部，但有些项目真的不太好用，只是偶尔能工作。我可以看到随着模型不断改进，这个进展过程；我可以更快、更好、更便宜地构建它。看着那个地平线展开真的很有趣。

I think you also need something that's a little bit ambitious—something slightly out of reach. That's how I feel about some of my projects. Not all, but some don't really work that well, they kind of work sometimes. I can see the progression as models keep getting better; I can build it faster, better, cheaper. It's a lot of fun to see that horizon play out.

你一直身处这个领域，看到了世界上很多用于此目的的东西。我很好奇，回顾一下：你是从 NASA 到 OpenAI？还是从 Apple 到 OpenAI？

You've had the benefit of being inside so much of what is out in the world being used for this purpose. I'm curious, going back: you went from NASA to OpenAI? Or Apple to OpenAI?

离开 Apple 后，我加入了一家名为 Path AI 的初创公司，从事数字病理学的机器学习和深度学习。我以机器学习工程师的身份离开 Apple，加入 Path AI，做了一些平台工程和深度学习的工作，包括高性能计算——他们运行自己的数据中心，这非常有趣。然后我离开了 Path AI，去了 OpenAI。

I was at a startup called Path AI doing machine learning and deep learning for digital pathology right after Apple. I left Apple as a machine learning engineer, joined Path AI doing a bunch of platform engineering and deep learning stuff, including high performance computing—they were running their own data center, which was really interesting. Then I left Path AI and went to OpenAI.

当时 AI 领域发生了什么，促使 OpenAI 联系你？你的看法是什么？是什么让你选择加入？

What was happening in the world of AI at the time that OpenAI reached out to you? What was your perspective? What made you choose to join?

那是一个有趣的时刻。过去四年我一直在做 AI 相关工作，训练计算机视觉模型并部署。很明显，AI 的应用已经对很多人来说很有趣了。但在 Apple，我们将 AI 应用于计算机视觉，却几乎无法让它工作并完成我们想要的最基本的事情。它基本上不管用。这种轨迹在整个生态系统中都在上演：狭窄的特定领域应用，模型在做有用的事情，比如 NLP、时间序列、计算机视觉。我在 Apple 的团队提供服务——我们为内部团队解决定制问题，然后将其转化为服务并产品化。所以我看到了技术在哪里有效、哪里无效的广泛范围。计算机视觉的工作让我接触到了 PathAI。后来听说更多关于 OpenAI 在做的事情——那是一种非常不同的方法，试图泛化很多这些东西。

It was an interesting moment. I had spent the last four years doing AI stuff, training computer vision models and deploying them. It was obvious there were applications of AI that were already interesting for a lot of people. But at Apple, we were applying AI to computer vision and could barely get it to work and do the very basic things we wanted. It kind of didn't work. That trajectory was playing out across the ecosystem: narrow domain-specific applications where models were doing useful stuff, like NLP, time series, computer vision. The team I was on at Apple provided services—we'd solve a bespoke problem for an internal team and then turn it into a service and productionize it. So I saw a broad range of where the technology was working and where it wasn't. The computer vision stuff got me exposed to PathAI. Then hearing more about what OpenAI was doing—it was a very different approach to trying to generalize a lot of these things.

在语言方面，看到趋势后，我大量使用了 GitHub Copilot。真正让我触动的是，我当时正在为 Julia 编程语言做很多事情。我从事开发者关系工作，试图推广这种很少有人知道的全新编程语言。我们处在科学计算的早期阶段，试图加速那个生态系统。我成为 Jasper AI 的客户，它是 OpenAI 最早的客户之一。他们做文案营销，我开始用它来制作大量 Julia 内容，纯粹为了测试如何做得更好、更快、更便宜。有趣的是，它实际上效果并不好，但就像我在 Apple 的计算机视觉经历一样，我可以清楚地看到这将是可能实现并且很快就能运作的事情。那是在我知道 Jasper 由 OpenAI 的 API 驱动之前。后来随着我更深入了解 OpenAI 并接触到它，我才知道，我显然很感兴趣。当时我有很多疑问，因为 OpenAI 从组织角度来看非常深奥。从技术上讲，我加入的是一家由非营利组织拥有的初创公司，带有某种奇怪的合作伙伴关系。这是一个非常奇怪的结构，我努力想弄明白这东西怎么可能成功。但让我兴奋的是，内部人员都知道，这发生在 GPT-4 之前，而且我认为 2022 年夏天，第一次 GPT-4 训练已经完成。OpenAI 完成了所有初步产品探索，开始构建 ChatGPT 以及不同领域的许多其他体验，试图弄清楚这项技术在 API 之外、除了让营销人员做文案之外还有什么商业用途。我在面试过程中遇到的每个人都说：‘哦，我们有很棒的新模型，但不能给你看，但真的很令人兴奋。’所以我确实很好奇，想了解更多，看看当时 text-davinci-003 模型和其他东西能做什么。最终，GPT-3.5 带来的变化非常深刻。

For language, seeing the traction, I used GitHub Copilot a bunch. The thing that clicked for me was that I was doing a lot of stuff for the Julia programming language. I was doing developer relations things, trying to bootstrap this brand new programming language that not many people knew about. We were in the very early days of scientific computing and trying to accelerate that ecosystem. I became a customer of Jasper AI, who were one of the original OpenAI customers. They were doing copywriting marketing, and I started using it to create a lot of Julia content, just to kick the tires on how to do it better, faster, cheaper. It was interesting because it actually didn't work well, but it was similar to my computer vision experience at Apple, where I could see that it was clearly going to be something possible and work somewhat soon. That was before I even knew that Jasper was powered by OpenAI's API. I only learned that later as I spent more time understanding what OpenAI was doing and got exposed to it, and I was clearly interested. I had lots of questions at the time because OpenAI was very esoteric from an organizational perspective. I was technically joining a startup owned by a nonprofit with a weird partnership structure. It was a very weird structure, and I was trying to wrap my head around how this thing would ever be successful. But one thing that got me excited was that internally, this was prior to GPT-4, and I think in summer 2022, the first GPT-4 training run had finished. OpenAI had done all the initial product exploration, starting to build ChatGPT and a bunch of other experiences in different domains, trying to figure out what the commercial use of this technology would be beyond putting it in an API and letting marketing people do copywriting. Everyone I met during the interview process said, 'Oh yeah, we have these great new cool models, and we can't show you, but it's really exciting.' So I was definitely intrigued to learn more, seeing what was possible at that time with the text-davinci-003 model and other stuff. Ultimately, what happened with GPT-3.5 was really profound to see the shift.

就像我刚才说的 GPT 时刻。你在此之前就在那里，所以亲眼目睹了这一切。你对那个时刻有掌控吗？它是偶然发生的吗？你促成了它的发生吗？

Like what I just called the GPT moment. You were in there before that, so you saw it happen. Did you have a finger on the button of that happening? Did it happen kind of by accident? Did you contribute to that happening?

是的，模型已经训练好了。我加入 OpenAI 时，更多是关于我们知道 GPT-4……实际上我的面试过程很长。我加入 OpenAI 的第一天，ChatGPT 在推出后一周就达到了百万用户。我从 2022 年春末夏初开始面试，直到 2022 年 12 月才加入。所以有很多这样的情况：他们知道人们会想使用这些模型，只是需要让人们兴奋起来。所以有很多关于用例的问题，以及如何为开发者铺设道路以便他们用这项技术构建产品的讨论。在 GPT-4 发布之前，我花了很多时间与客户合作。同时，我也花了很多时间在评估方面，包括 OpenAI Evals 框架和开源项目，以获取模型反馈，众包评估模型在哪些方面表现好、哪些方面不好、人们想看什么。现在近三年或四年后回看这些早期工作，很有趣的是我们取得了多大进展，但基准测试问题仍然未解决，这非常有趣。如果你感兴趣，我们可以讨论这个话题。

Yeah, the model had already been trained. When I joined OpenAI, it was much more about knowing that GPT-4... actually, I had a very long interview process. The first day I joined OpenAI was when ChatGPT hit a million users a week after it launched. I started interviewing in late spring/early summer of 2022 and didn't end up joining until December 2022. So there was a lot of this: they knew people were going to want to use the models; they just needed to get people excited. So there were many use case questions and discussions about how to lay the tracks for developers to build with the technology. I spent a lot of time doing that prior to the GPT-4 launch, working with customers. I also spent time on the evaluation front, with the OpenAI Evals framework and open source project, to get feedback on the model and crowd-source evals of where the models worked well, where they didn't, and what folks wanted to see. Seeing that early work now, almost three or four years later, it's very interesting how far we've come and yet how unsolved the benchmarking problem remains, which is very interesting. We can talk about that if you're interested.

是的，非常有趣。我们的一位前嘉宾 Anastasia 和 Elena Marina 花了很多时间在基准测试问题上。是的。

Yeah, fascinating. One of our former guests, Anastasia and Elena Marina, are spending a lot of time on the benchmarking problem. Yep.

我和 Anastasia、Waylon 以及 Elena Marina 团队在他们起步时花了大量时间相处。所以他们做得非常出色。

I spent a bunch of time with Anastasia, Waylon, and the Elena Marina team as they were getting started. So they're crushing it.

太棒了。我有一个问题：我们的很多听众是早期工程师或正在考虑加入早期初创公司的工程师。指导我的一点是，百分之百被问到的人会说，如果知道今天谷歌的一切，他们可能会在谷歌还是初创公司时加入，对吧？每个人都说是。而且每个人可能都会说，考虑到今天对 OpenAI 的了解，他们会加入早期 OpenAI。但真正加入并且能预见初创公司能发展成什么样子的人却很少。工程师应该如何看待初创公司？他们应该如何从成功可能性或发展轨迹的角度评估是否加入初创公司？有什么关键想法吗？

Incredible. A question I have: a lot of our audience is early engineers or engineers considering an early startup. One thing that guides me is that 100% of people you ask would say they probably would have joined Google knowing everything they know today about what happened to Google back when it was a startup, right? Everyone says yes to that. And everyone probably would say they would join OpenAI when it was an early startup given everything they know today. But very few people actually do join and can foresee what a startup can grow into. How should an engineer look at a startup? How should they evaluate whether to join a startup from the lens of likelihood of success or trajectory? Any key thoughts?

是的，我要先说明一下，当我加入 OpenAI 时，我有另一个工作机会。所以回到我之前谈到的，我在科学计算和开放科学领域做了很多工作，最终帮助 NASA 制定开放科学战略等。所以我非常身处那个世界。我也在做一些开发者关系的 AI 相关事情。我当时另一个工作机会，如果我没有加入 OpenAI，就是去领导开源——比如 IBM 的全球开源科学技术负责人。那是与 OpenAI 截然不同的另一端。当时，不给自己太多赞誉，对我来说正确的决定并不明显。我当然对 OpenAI 在做的事感兴趣，它看起来很酷，处于前沿，但事情最终会如何发展并不清晰。

Yeah, and I'll caveat this by saying when I was joining OpenAI, I had another job offer. So, going back to what I talked about before, I was doing a lot of stuff in scientific computing and open science, ultimately helping to advise NASA on their open science strategy and other things. So I was very much in that world. I was also doing developer relations AI stuff. My other job offer at the time, had I not joined OpenAI, was going to be to lead open source—like global technical lead for open source science at IBM. That's the other end of the spectrum from OpenAI. At the time, to not give myself too much credit, it was not very obvious to me what the right decision was. I was definitely interested in what OpenAI was doing; it seemed really cool and at the frontier, but it was less clear how things would actually play out.

当然，与 IBM 以及另一家风格迥异的公司之间的权衡，但我从根本上仍然相信开放科学以及所有相关的工作，这与我当时正在做的事情紧密相连，某种程度上是在平衡这两件事。显然，事后看来，我认为我做出了正确的决定，最终去了 OpenAI，但我没有预知未来的水晶球。当时的 OpenAI 还不是今天的样子。我很大程度上是根据对团队成员的判断来做决定的。还有我在面试过程中遇到的人。Andrew Mayne，他后来离开了 OpenAI，现在主持 OpenAI 播客，是最早的提示工程师之一。他可以说是世界上第一个提示工程师，花了很多时间推动模型的发展。一位名叫 Angela 的女性，后来成为 OpenAI 最早的产品经理之一，帮助发布了 GPT-4 和许多其他产品，现在有了自己的创业公司。真是太棒了。还有其他一些人，比如 Sherwin，他领导应用平台团队，以及许多其他优秀的人。这才是真正让我兴奋的地方。我想，‘使命看起来很酷。这似乎是一个做这类工作的好地方。而且面试过程中遇到的人，我之前从未见过，也不认识 OpenAI 的任何人。但在面试过程中，和他们花了很多时间交谈，很明显这是一个非凡的团队。’所以，实际上事后我很庆幸面试过程拖得很长，因为我个人刚离开加州回到芝加哥和女友在一起。所以他们不确定是否要远程招聘等等。所以这是一次有趣的经历，我认为最终让我对团队成员有了更多的了解。

And of course all the trade-offs with IBM and different very different flavor of a company, but ultimately I still fundamentally believe in open science and all that work and it was very connected to the work that I was doing and sort of balancing these two things. And obviously in hindsight I think I made the right decision and ended up at OpenAI, but I didn't have a crystal ball. OpenAI didn't look like the company it looks like today. I was making that decision in large part based on the signal I had of who the people were. And just the folks that I met during the interview process. Andrew Mayne, who now has left OpenAI and runs the OpenAI podcast as sort of the early prompt engineer. He's like the world's first prompt engineer and spent a lot of time pushing on the models. A woman named Angela who ended up running, she was one of the first PMs at OpenAI and helped ship GPT-4 and a bunch of other stuff and now has her own startup. That's incredible. And a bunch of other folks, Sherwin, who leads the applied platform team, and just other amazing people. That was really what got me excited. I was like, 'The mission seems cool. This seems like a great place to potentially do this work. And the people I met during the process, who I never had met before, I didn't know anyone at OpenAI. But just during the interview process, spending a bunch of time with them in lots of conversations, it became clear that it was an incredible team. And so I'm actually glad in hindsight that my interview process was quite drawn out because I had personally just left California and gone back to Chicago to be with my girlfriend. So they didn't know if they wanted to hire somebody remotely, etc. So it was an interesting experience, which I think ultimately benefited me to get more signal about who the people were in the team.

这也是我听别人说过的论点：找到你最愿意与之共事的一群人，考虑他们正在攻克的技术挑战、团队之间的情谊。从早期团队建设的角度来看，我们当然也会遵循这一点——优秀早期项目所在之处的人才吸引力。这一定是个不错的论点。百分之百——你在 OpenAI 也是这么做的。

That's a thesis I've heard others say as well: find the group of people that you would enjoy spending time with the most, the technical challenges they're working on, the camaraderie that exists among that group. And certainly one that we follow from a building early teams perspective, the talent magnetism of where the good early stuff is being built. That has to be a nice thesis. This is 100% — you follow it at OpenAI as well.

嗯，百分之百。

Yeah, 100%.

我很好奇，你当时负责开发者关系，对吧？那么你一定非常深入——跟世界分享一下。在 OpenAI 这样的公司做开发者关系，尤其是在全世界似乎都在基于 OpenAI 构建的时候，这意味着什么？

I'm curious, so okay, you ran developer relations, right? So that means you're very in — share with the world. What does it mean to be in developer relations at a company like OpenAI at a time where it seems like everyone in the world is building on OpenAI?

是的。最初 OpenAI 希望我做的事情是，他们不知道人们会如何使用这些模型。从产品角度看，他们还没有实现产品-市场匹配。API 的客户群非常专注于文案写作和一些小众应用。我认为他们当时的假设是 GPT-4 会带来更广泛的益处。所以他们希望有人加入，将模型的智能和能力直接转化为更多公司和用户能受益的东西。这是一个开发者的挑战——理解这些用例并充分发挥模型的能力。结果呢，我加入 OpenAI 的第一周，第一天，ChatGPT 就席卷了世界。所以突然之间，关于这项技术如何应用的认知不再是问题。问题变成了：我们有大量用户涌来，如何确保我们为他们构建正确的产品？如何确保我们拥有模型能够提供的所有新能力？如何将这些能力整合到产品体验中，充分发挥模型的潜力？所以我花了很多时间构建 OpenAI 的核心开发者平台，做产品工作、工程工作，与客户交流，编写开发者文档，拜访他们。这有点像在传统初创公司环境中身兼数职，以极快的速度前进。最有趣的就是能够同时担任多种角色。我觉得开发者关系真的取决于你所处的组织。最好的情况就是你能身兼多职，做最有影响力的工作，帮助客户成功。这就是我在 OpenAI 所做的工作，非常有趣。

Yeah. Originally what OpenAI wanted me to do was they didn't know how people were going to use the models. They hadn't gotten product-market fit from a product perspective. The set of customers in the API was very focused on copywriting and some esoteric use cases. I think they had the thesis that GPT-4 would be more widely beneficial. So they wanted somebody to come in to translate this intelligence and capability directly into something that many more companies and users would benefit from. That's a sort of developer problem — to understand those use cases and push the models to their full capacity. What ended up happening was my first week at OpenAI, my first day, ChatGPT was taking the world by storm. So all of a sudden the awareness and how this technology could be applied became much less of a problem. It became: we have tons of users showing up. How do we make sure we're building the right product for them? How do we make sure we have all these new capabilities that the model is capable of? How do you wield those into a product experience that brings the model capability to its full extent? So I spent a bunch of time building the core OpenAI developer platform, doing product work, engineering work, spending time with our customers, writing developer documentation, meeting them. And sort of living at the edge, wearing many hats like in a normal traditional startup environment, moving at breakneck pace. That was the most fun — being able to wear all the hats. I feel like DevRel really depends on the organization. The best case is where you just get to wear lots of hats and do the most impactful work to help customers be successful. That's what I got to do at OpenAI, which was a ton of fun.

那么，我好奇的是——你曾有机会构建一个供开发者使用的平台。所以从创始人的视角来看，也许给现在的创始人一些建议。有一种感觉或观点认为，创始人必须在某种程度上与大型前沿实验室的速度竞争，无论是实验室本身的进化还是它们向世界提供的产品。那么，创始人应该如何思考使用现有模型和平台，以保持领先、保持竞争力或保持独特性，同时应对实验室自身的发展速度？

So, what then I'm curious about — just you getting the opportunity to build a platform that developers build on. So from the lens of a founder, maybe speak to founders right now. There's this feeling or perspective from the founder lens that they have to somehow compete with the speed of the big frontier labs with regards to the evolution of the labs themselves and the product they're offering to the world. So how should a founder think about the use of existing models and platforms with regards to how they stay ahead or stay competitive or proprietary with respect to the speed and pace of development of the labs themselves?

嗯，这是个好问题。我观察这一切的时间越长——尽管有一些显著的例外——我就越确信，实际上与模型实验室竞争并没有那么难。我认为这取决于你试图在哪个接入层面和维度上与他们竞争。

Yeah, it's a good question. I mean, the longer I've seen all this play out, and there are a couple notable exceptions to this, the more that I'm convinced that it's actually not that hard to compete with the model labs. I think it depends on what access and dimension you're trying to compete with them.

我觉得如果你想作为一个通用个人助手去与 ChatGPT 和 Gemini 等产品竞争，那会是非常艰难的路，因为这些产品已经有数十亿用户的分发渠道，很高的知名度、采用率和营销力度。你也许能蚕食一点点市场，但问题是你要进入哪个市场。好处是，我的世界观是：很多这些模型创造了大量的能力冗余和净收益，而这些并没有被模型公司捕获。API 产品对我来说本质上是一种正向的价值交换——我们知道自己没有捕获全部价值。关键是让客户去捕获价值，在我们的基础上构建，通过他们的产品创造经济价值。这是我做开发者产品的核心信念：这是正常的基本价值交换。随着模型能力越来越强，公司越来越野心勃勃，估值越来越高，他们可能需要构建更多产品，在不同维度上竞争，所以你需要警惕他们的动向。但即便如此，以编程为例——我现在会很兴奋地去创办一家编程公司。虽然竞争很多，但市场巨大。软件创建的 TAM 扩展是巨大的。随着一切变化，可以构建全新的产品体验，这属于真正的研究前沿工作。对创始人来说，最好的情况是技术正在进步，而地面也在变化。如果事情更稳定，你反而应该更担心。正因为变化如此巨大，任何人都有机会做出东西。编程是个很好的例子：OpenAI 做的第一个有用的东西是 GitHub Copilot，它在两年半前占据了 AI 编程 100% 的市场份额。现在虽然有人说没人用 Copilot，但它其实做得不错——如果你去看数字就知道。与此同时，Cursor、Codex 爆发，Google 的 Anti-Gravity 也获得很大关注，还有 Cognition、Windsurf 等众多产品都做得很好。Cursor 被收购并与 SpaceX 合作。大概还有 25 个其他非常成功的产品，这还没算上其他开发者 AI 编程生态系统或 vibe coding。所以机会巨大，变化很多。我觉得人们过于担心模型实验室会做什么，而不是专注于构建一个受益于更好模型并为客户创造价值的优秀产品。如果你这么做，我非常积极和看好——不管模型实验室做什么都没关系。难点在于构建你可以在此基础上构建的优秀技术，而他们无法捕获所有客户。这是不可能的。所以每个人都会有大量的机会。

I think if you're going to try to compete as a generic personal assistant and compete with ChatGPT and Gemini and all that stuff, it's going to be a really tough road because these products have billion-user distribution at this point, large awareness, adoption, and marketing. Maybe you can chip away a little, but the question is what market you're actually going to enter. The nice thing is my worldview: a lot of these models create so much capability overhang and net benefit not captured by the model companies. The API product to me is about a positive exchange of value – we know we're not capturing all the value. The point is to let customers capture value and build on top of us, creating economic value in whatever their product does. That's a core part of my belief in building developer products: that's the normal fundamental exchange of value. As models become more capable and companies more ambitious with higher valuations, they may need to build more products and compete in different dimensions, so you need to be wary of where they go. But even then, take coding as an example – I'd be excited to build a coding startup right now. There's a lot of competition but such a big market. The TAM expansion of software creation is massive. There's true research frontier work of new product experiences that can be built as everything changes. The best case for a founder is all this technology progress happening but the ground shifting. You should be more worried if things were more stable. Because of how much things shift, anybody has a real shot at building something. Coding is a great example: the first useful thing OpenAI did was GitHub Copilot, which had 100% market share of AI coding 2.5 years ago. They're doing quite well today despite the narrative that nobody uses it – the numbers show it's fine. At the same time, you've seen Cursor, Codex explode, Google's Anti-Gravity getting traction, and tons of other products like Cognition and Windsurf doing awesome. Cursor got acquired to go with SpaceX. There are probably 25 other really successful products, not even covering other developer AI coding ecosystems or vibe coding. So a huge opportunity, lots of change. I think folks are overly worried about what the model labs will do and less focused on building a great product that benefits from better models and creates customer value. If you do that, I'm very positive and bullish – it doesn't matter what the model labs do. The hard part is building great technology for you to ultimately build on top of, and they can't capture all customers. It's just not possible. So there will be tons of opportunity for everyone else.

说得太好了。所以创始人们、投资者们，别再担心了。直接入场吧。

I love it. So founders, investors, stop worrying about it. Just get in there.

我觉得如果你没担心而是直接去做了，你就能得到像 Cursor 那样的退出和成果，还有 Cognition 也一样。那些家伙——有很多人成功的例子，结果就在那里，这很棒。

I think if you hadn't worried about it and just done it, you would have ended up with the sort of Cursor-level exits and outcomes and Cognition again. Those folks – there are lots of great examples of people doing it, and the outcome is there, which is great.

没错，100%。我想跟你聊的太多了。我想确保我们聊聊你的 Google 经历。从你的背景看，从市场角度可能有点反直觉，但当时是什么情况？是什么吸引你去 Google？你在 Google 看到了什么机会？

Yeah, 100%. There's so much I want to talk to you about. I want to make sure we spend time on your Google story. Looking at your background, it might have seemed counterintuitive from the market's perspective, but what was going on? What attracted you to Google? What were you seeing at Google as the opportunity?

我觉得在很多方面人们对 Google 过于严苛了，这在很多方面是没有道理的。但这里面有细微差别：人们假设 Google 位置很好，所以利用这个优势一定很容易。这是最常见的误解。当然，Google 位置很好，但我们必须付出人类所有可能的努力才能利用这个优势，而且这并不容易。我不会因为拥有好的基础设施、好的模型、很多人才和很多钱，就每天醒来觉得生活很容易。这感觉像是一场激烈的竞争，我喜欢这样，因为我好胜。我当时的观察是：第一，我认为 Google 会处于极其有利的位置。第二，我喜欢逆袭的故事。在 OpenAI 的后半段，每个人都一直夸我们 Open AI 多伟大、多成功。那并不是激励我的东西。

I think in many ways folks are quite harsh on Google, which is unwarranted in many ways. But there's nuance: people assume Google is well positioned, so it must be easy to take advantage of that fact. That's the most classic misnomer. Sure, Google is well positioned, but we have to do every ounce of hard work humanly possible to take advantage of that fact, and it's not easy. I don't wake up every day and because we have great infrastructure, great models, lots of people, and lots of money, somehow life is easy. It feels like we're in a great competition, and I like that because I'm competitive. My observation at the time was, first, I thought Google would be extremely well positioned. Second, I love the underdog story. I spent a lot of time in the latter part of my time at OpenAI, and everybody was always patting us on the back about how great OpenAI is and how successful we were. That's not what gets me motivated.

我非常享受从零到一的故事。所有人都认为 Gemini 和 Google 要输了，关于 Google 的业务有各种悲观论调。但看到过去几年我们取得的进步、我们缩小了多少差距，以及我们现在真正领先和获胜的领域，我觉得这太酷了。我确实被震住了——我没有答案。我看到当前这个领域非常艰难，当然 Google 有一切应该和能够获胜的理由，但我被“这将很艰难”这个想法深深吸引。我们必须推石上山。在 OpenAI 时我几乎没体验到足够的艰难，因为那里太爆炸式增长了。虽然扩展规模显然有挑战，但还不够难。所以我真的很享受在 Google 的经历：它很难，但同时看到所取得的进展也超级有回报感。

I very much enjoy the story of going from zero to one. Everyone thought Gemini and Google were going to lose, and there was a lot of doom and gloom about the business Google was operating, etc. But seeing the level of progress we've made in the last few years, how much ground we've closed, and the places where we're truly leading and winning now, I think it's just so cool to see. I was definitely taken aback—I didn't have the answer. I saw that this looked like a pretty rough space right now, and of course Google has all the reasons it should and could win, but I was very attracted to the idea that it's going to be tough. We're going to have to push the rock up the hill. I almost didn't get enough of that at OpenAI because it was so explosive. There were clearly challenging things about trying to scale up, but it almost wasn't difficult enough. So I really enjoyed the Google experience: it's been difficult, and also super rewarding to see the level of progress that's been made.

那么，AI 组织内部的开发文化和重点是什么样的？我除了公开听到的之外对 Google 一无所知，但曾经有段时期，大家都在谈论谢尔盖·布林如何深度参与并影响了开发的焦点和节奏。你当时在场吗？你能分享一下这是什么样的体验吗？

Yeah, what is the development culture and focus like within the AI organization? I don't know anything about Google except for what I hear publicly, but there was a moment where there was talk about how Sergey Brin leaned in and had an impact on the focus and pace of development. Were you around at that time? What perspective can you share on what it's like?

是的，DeepMind 是一个非常特别的地方。谢尔盖确实深度参与了，这非常有趣，因为他有创始人的能量和非常独特的世界观。桑达尔深度参与，德米斯深度参与，科拉伊——我们的 CTO 兼首席 AI 架构师，日常负责 Gemini——也深度参与。所以我们在这一关键时刻有一群杰出的领导者，他们深入一线，推石上山。每个人都有非常不同的视角。谢尔盖的视角作为创始人很独特。桑达尔的视角作为掌舵一艘大船的人很独特，这艘船有很多事情与日常模型竞赛无关，但都深受 AI 进展的影响。德米斯作为开启这条通往 AGI 之路的人，拥有极其独特的视角。而科拉伊作为 DeepMind 的第一位深度学习研究员，对于如何实际推石上山并取得进展有着非常扎实的研究视角。所以这是一个非常特别的环境。我经常回溯的一点是，DeepMind 以及更广泛的 Google 拥有最好、最成功的产品组合。作为一家大公司，我们在做很多事情，我们也有义务这样做。大公司的优势是可以并行下很多赌注。但这意味着资源分配是一个很难时刻做对的问题。例如，我们在 Gen Media 上一直处于前沿，推石并引领该领域。但在编码方面我们稍微落后于曲线，因为在众多赌注中，你可能会在六个月里投入不足。然后你看到一个大机会，却稍微落后了。我喜欢的是 Google 和 DeepMind 内部一旦知道有差距就立即追赶的文化。它非常稳固。DeepMind 团队：你给人们一个可量化的目标去推石上山，我们拥有世界上最优秀的人去做这件事。那种纪律性和人们锁定目标去爬山、提升质量的水平非常酷。它还非常科学驱动，这很酷。德米斯和最初 DeepMind 团队的科学背景渗透到了组织的运作方式中，这非常有趣。

Yeah, DeepMind is a very special place. Sergey has definitely leaned in, which is a ton of fun because he has founder energy and a very unique worldview. Sundar is very leaned in, Demis is very leaned in, and Koray—our CTO and chief AI architect who runs Gemini day-to-day—is super leaned in. So we have a bunch of incredible leaders who are deeply in the weeds, pushing the rock up the hill at this pivotal moment. Each of them has a very different perspective. Sergey’s perspective is unique as a founder. Sundar’s perspective is unique as someone steering a huge ship with many things going on that are not the day-to-day model race, but are deeply impacted by AI progress. Demis, as the person who started this whole path to AGI, has an incredibly unique view. And Koray, as DeepMind’s first deep learning researcher, has a very grounded research perspective on how to actually push the rock up the hill and make progress. So it’s a very special environment. The thing I always come back to is that DeepMind and Google more broadly have the best and most successful portfolio. As a big company, we’re doing a lot of stuff, and we have an obligation to do so. The advantage of a big company is you can take many bets in parallel. But that means resource allocation is a very difficult problem to get right all the time. For example, we’ve been frontier at Gen Media, pushing the rock and leading in that space. But we’ve been slightly behind the curve on coding, because in the plurality of bets, you might underallocate over six months. Then you see a big opportunity and you’re slightly behind. What I like is the culture inside Google and DeepMind of closing the gap once we know there is one. It’s rock solid. The DeepMind team: you give people a quantifiable thing to push the rock up the hill on, and we have the best people in the world to do it. The level of discipline and how locked in people are to hill-climb and push quality up the hill is very cool to see. It’s scientifically driven, which is cool. Demis and the original DeepMind team’s background as scientists permeates the way the organization operates, which is a lot of fun.

似乎每家公司现在都深度投入了，如果没有，他们就在想办法投入。似乎每个之前没有在建设的创始人现在又重新开始建设了。感觉这真是一个不可思议的时代。

It seems like every company is leaned in right now, and if they're not, they're trying to figure out how to get leaned in. It seems like every founder who hasn't been building is back out building. It feels like just an incredible time to be out there right now.

这对世界也很好。我经常思考竞争。作为一个有竞争心的人，我想赢并且拥有最好的东西，但这也很好地反映了这对世界、对创始人、对每个使用这项技术的人有多好，因为每个人都在如此努力。我真的认为开发者生态系统和消费者最终希望能捕获不成比例的价值，因为从技术角度来看，这一时刻竞争如此激烈。我希望这种竞争水平能够长期持续下去。我认为这将是最好的结果。

It's great for the world too. I think about competition a lot. As someone who is competitive, I want to win and have the best stuff, but it's also a great reflection of how great it is for the world, for founders, and for everyone using this technology that everyone is pushing so hard. I really think the developer ecosystem and consumers are hopefully the ones who capture a disproportionate amount of value because of how competitive this moment is from a technology standpoint. I'm hopeful that level of competition keeps up in a sustained way long term. I think that will be the best outcome.

那么你的工作与开发者生态系统紧密相关。你如何看待在当今开发中取得成功需要什么？

So your job works very closely with the developer ecosystem. What's your take on what it takes to be successful in development today?

是的，从哪个角度来说？你是指创办一家初创公司吗？

Yeah, from what perspective? You mean like to build a startup?

也许是从创始人的角度，但有很多讨论是关于开发者应该多大程度上使用 AI 对比基础开发能力。

Maybe that's from a founder perspective, but there's so much talk about how much developers should be using AI versus foundational development capabilities.

从技能发展的角度，人们应该往哪个方向努力？是成为 AI 原生开发者，还是那些没有在 AI 中成长起来的人也能达到？今天成功的开发者是什么样子的？

Where should people be leaning in from a skill development advancement perspective? Is it AI native or could people who haven't really grown up in AI still get there? What does a successful developer look like today?

Andrej Karpathy 有句名言：你可以外包智能，但不能外包你的理解。更准确地说，你可以外包智能，但你不应该外包你的理解。我在 AI 编程中一直思考这一点。最擅长使用这些工具的人对代码中实际发生的情况有深刻理解。有一个宏观担忧：当大型代码库中的大部分代码都是由 AI 编写，而不是由深入了解代码的人亲手编写时，会发生什么？我认为前面有很多有趣的问题，但使用 AI 工具按需获取智能以获得更大的杠杆作用是很棒的。然而，这并不能免除你真正理解代码的责任。这里有一个矛盾：当前的模型并没有帮助你保持这种理解；它们只是快速完成智能外包。这恰恰说明了这些工具中的巨大机会。想象一下未来的编码智能体，它既能成功创建代码，又能帮助你深入理解。那才是我想要的产品。还有很多东西尚未构建。

There's a great quote by Andrej Karpathy: you can outsource intelligence, but you can't outsource your understanding. Or more precisely, you can outsource your intelligence, but you shouldn't outsource your understanding. I think about that in AI coding all the time. The best folks wielding these tools have a deep understanding of what's actually going on in the code. There's a macro worry: what happens when over time the majority of code in large codebases is written by AI, not handcrafted by people who know it deeply? I think there are interesting questions ahead, but using AI tools to get intelligence on demand for more leverage is great. However, it doesn't abdicate the responsibility to truly understand the code. There's a tension: current models don't help you maintain that understanding; they just do the intelligence outsourcing. This points to the opportunity in these tools. Imagine a future coding agent that successfully creates code and helps you have deep understanding. That's the product I want. There's so much yet to build.

对 2026 年有什么大胆预测吗？

Any bold predictions for 2026?

我认为 2026 年将超越“软件创建基本上已经解决”这一普遍观点。这是我们正在走的轨迹，然后我们将开始处理有趣的二阶问题，带来新的问题和机遇。当软件变得丰富且易于获取时，机会并不会消失，反而会多出一千倍。看到这一切发生将会很有趣。它会在今年年底还是明年初成为现实？我们离这个目标只差模型和产品的几次迭代了。

I think 2026 will move beyond the commonly held belief that software creation is more or less solved. That's the trajectory we're on, and then we'll start dealing with interesting second-order questions, creating new problems and opportunities. When software is abundant and readily available, it's not that the opportunity vanishes; it's that there's a thousand times more opportunity. It'll be fun to see that happen. Is it true by the end of this year or early next? We're only a few turns of the model and product crank away from that being possible.

是的，在这个问题上我站 Logan 这边。一旦软件创建得到解决，感觉机会无限。我还有很多想聊的，但为了节省时间，人们在哪里可以联系到你？他们给你的第一条消息应该说什么？

Yeah, I'm on team Logan from that perspective. It feels like limitless opportunity once software creation is solved. There's so much more I want to talk about, but for the sake of time, where can people reach you? What should their first message to you be?

我收到很多消息，所以不确定是否需要更多，但我总是对人们正在构建的酷东西感兴趣。展望未来，编程能力将被解决，但人们真正希望模型接下来能做而今天还做不好的事情是什么？我正在花时间探索客户需求和模型参数能力，以发现下一个需要攀登的鸿沟。如果你是一位创始人或初创开发者，正在挑战模型的极限，并且看到模型当前能力与你期望之间存在巨大差距，请与我联系。让我们共同构建基准测试来量化这个差距。我花了很多时间帮助客户构建基准测试，以便从模型能力角度了解下一步的方向。

I get a lot of messages, so I don't know if I need more, but I'm always interested to hear about cool things people are building. Looking to the future, coding capabilities will be solved, but what are the next set of things folks really want models to do that they can't do well today? I'm spending time exploring customer needs and parametric capabilities to find the next gaps to hill climb. If you're a founder or startup developer pushing models to the limits and see a large gap between what they can do and what you want, get in touch. Let's build benchmarks together to quantify that gap. I spend a lot of time helping customers build benchmarks so we know the next capability directions from a model perspective.

我很喜欢这样。你在 X 上相当开放。我喜欢你向公众提问并与人互动。所以请关注 Logan 的 X 账号，直接和他交流。这是保持联系的好方法。Logan，非常感谢你第一次参加播客与我们交流。我百分之百希望你再回来，因为还有这么多话题想聊。再次感谢你今天做客。

I love that. You're pretty open on X. I love when you ask the public questions and people interact with you. So go follow Logan on X and just talk to him. That's a great way to stay connected. Logan, thank you so much for spending time with us on your first podcast. I 100% want you back because there are so many more topics I wanted to discuss. Thank you again for being with us today.

我很喜欢。感谢你的深思熟虑的问题。这很有趣。

I love it. Thank you for the thoughtful questions. This was fun.