Karpathy：从 Vibe Coding 到智能体工程

Stephanie Zhan

非常高兴请来第一位特邀嘉宾。他参与建造了现代 AI，又把它讲清楚，偶尔还给它改名。他就在这间办公室里参与创办了 OpenAI，当年也是把 Tesla Autopilot 真正跑起来的人。他有一种少见的本事：让最复杂的技术转折听起来既好懂，又像必然会发生。

We’re so excited for our very first special guest. He has helped build modern AI, then explain modern AI, and then occasionally rename modern AI. He actually helped co-found OpenAI right inside of this office, was the one who actually got Autopilot working at Tesla back in the day. And he has a rare gift of making the most complex technical shifts feel both accessible and inevitable.

大家都知道他去年造了「vibe coding」这个词。但就在最近几个月，他说了一件更惊人的事：他从未觉得自己作为程序员如此落后。我们今天就从这里开始。谢谢 Andrej 来。

You all know him for having coined the term vibe coding last year, but just in the last few months he said something even more startling: that he’s never felt more behind as a programmer. That’s where we’re starting today. Thank you, Andrej, for joining us.

Andrej Karpathy

你好。很高兴来，也高兴给大会开场。

Yeah, hello. Excited to be here and to kick us off.

Stephanie Zhan

就在几个月前，你说从未觉得自己作为程序员如此落后。这话从你嘴里说出来特别刺耳。能拆开讲讲吗？那种感觉是兴奋，还是不安？

Okay, so just a couple months ago you said that you’ve never felt more behind as a programmer. That’s startling to hear from you of all people. Can you help us unpack that? Was that feeling exhilarating or unsettling?

Andrej Karpathy

两者都有。跟在座很多人一样，我用智能体工具——Claude Code 一类——大概有一年了。它对大块代码很有帮助，偶尔会搞砸，你得改。然后 12 月对我是一个很清楚的拐点：我在放假，时间多一点，很多人应该也类似。我开始注意到，最新模型吐出来的块就是对的；我再要更多，还是对的。我不记得上次纠正它是什么时候了。然后就越来越信任这套系统，就开始 vibe coding。

Yeah, a mixture of both for sure. First of all, like many of you, I’ve been using agentic tools like Claude Code, adjacent things, for a while — maybe over the last year as it came out. It was very good at chunks of code. Sometimes it would mess up and you have to edit them, and it was kind of helpful. Then I would say December was this clear point where, for me, I was on a break, so I had a bit more time. I think many other people were similar. I just started to notice that with the latest models, the chunks just came out fine, and then I kept asking for more and it just came out fine. I can’t remember the last time I corrected it. And then I just trusted the system more and more, and then I was vibe coding.

这是一次非常陡的转变。我在 Twitter / X 上特别强调过：很多人去年体验到的 AI 还是 ChatGPT 旁边那种东西，但你真的得再看一眼，而且得按 12 月这个时点再看，因为事情从根本上变了，尤其是那种连贯的智能体工作流真的开始能跑。

I do think that it was a very stark transition. I tried to stress this on Twitter and on X because I think a lot of people experienced AI last year as a ChatGPT-adjacent thing, but you really had to look again, and you had to look as of December, because things have changed fundamentally, especially on this agentic coherent workflow that really started to actually work.

就是这个认识让我掉进无尽的 side project 兔子洞。我的 side projects 文件夹塞满了各种乱七八糟的东西，一直在 vibe coding。大概就是 12 月发生的，从那以后我一直在看这件事的后果。

It was just that realization that really had me go down the whole rabbit hole of just infinite side projects. My side projects folder is extremely full with lots of random things, and just vibe coding all the time. So yeah, that kind of happened in December, I would say, and I was looking at the repercussions of that since.

Stephanie Zhan

你讲过很多次：LLM 是一台新计算机，不只是更好的软件，而是一整套新的计算范式。软件 1.0 是显式规则，软件 2.0 是学到的权重，软件 3.0 就是这个。如果这是真的，一个团队在真正信这一天，会把什么建得不一样？

You’ve talked a lot about this idea of LLMs as a new computer — that it isn’t just better software, it’s a whole new computing paradigm. Software 1.0 was explicit rules, software 2.0 was learned weights, software 3.0 is this. If that’s actually true, what does a team build differently the day they actually believe this?

Andrej Karpathy

对。软件 1.0 我在写代码。软件 2.0 我其实是在靠做数据集、训练神经网络来编程，编程变成了排布数据集、目标、网络结构。然后如果你在足够多的任务上训练 GPT / LLM——因为在互联网上训练，你不得不隐式地同时做数据集里所有事——它们在某种意义上就变成一台可编程的计算机。

Software 1.0, I’m writing code. Software 2.0, I’m actually programming by creating data sets and training neural networks. So the programming is kind of like arranging data sets and maybe some objectives and neural network architectures. Then what happened is that if you train one of these GPT models or LLMs on a sufficiently large set of tasks — implicitly, because by training on the internet you have to multitask all the things that are in the data set — these actually become kind of like a programmable computer in a certain sense.

软件 3.0 就是：编程变成写提示，上下文窗口是你撬动这个解释器的杠杆。LLM 在解读你的上下文，并在数字信息空间里做计算。有几个例子把这点钉死了。

Software 3.0 now is kind of about your programming turning to prompting, and what’s in the context window is your lever over the interpreter that is the LLM — interpreting your context and performing computation in the digital information space. I think there’s a few examples of that that really drove it home for me.

比如 OpenClaw 出来的时候，你通常会以为安装是一段 bash / shell 脚本。但为了覆盖各种平台和电脑，这些脚本会胀得极其复杂，而你还困在软件 1.0 里，想把代码写死。实际上 OpenClaw 的安装是一段给你复制、贴给智能体的文本：贴过去，它就会装。

When OpenClaw came out, when you want to install OpenClaw, you would expect that normally this is a bash script, a shell script. But in order to target lots of different platforms and lots of different types of computers, these shell scripts usually balloon up and become extremely complex. But you’re still stuck in a Software 1.0 universe of wanting to write the code. Actually the OpenClaw installation is a copy-paste of a bunch of text that you’re supposed to give to your agent. Copy-paste this and give it to your agent, and it will install OpenClaw.

这更强，因为你已经在软件 3.0 里：不必精确写死安装的每一步。智能体自带智力，看你的环境、你的电脑，做智能动作，在循环里调试。新的编程范式就是：那段该贴给智能体的文本是什么？

The reason this is a lot more powerful is you’re working now in the Software 3.0 paradigm where you don’t have to precisely spell out all the individual details of that setup. The agent has its own intelligence, follows the instructions, looks at your environment, your computer, and performs intelligent actions to make things work. It debugs things in the loop. What is the piece of text to copy-paste to your agent? That’s the programming paradigm now.

更极端的例子是我做 MenuGen 的时候。你去餐厅，菜单通常没图，大概 30%、50% 的菜我根本不知道是什么。我想拍菜单，拿到这些菜大概长什么样的图。我就 vibe coding 了一个应用：上传照片，跑在 Vercel 上，OCR 标题，用图像生成器配图，再渲染菜单。

One more example that is even more extreme is when I was building MenuGen. You come to a restaurant, they give you a menu, there’s no pictures usually, so I don’t know what any of these things are — usually like 30% of the things, 50%. I wanted to take a photo of the restaurant menu and get pictures of what those things might look like in a generic sense. So I vibe-coded this app that lets you upload a photo. It runs on Vercel, re-renders the menu, uses an image generator to OCR the titles, get pictures of them, and then shows it to you.

然后我看到软件 3.0 版本，把我震住了：把照片交给 Gemini，说用 Nano Banana 把东西叠到菜单上。它返回的图就是我拍的那张菜单，但像素里已经画出了那些菜。我那整个 MenuGen 其实是多余的，还在旧范式里干活，那个 App 不该存在。神经网络在做越来越多的活，提示就是图像，输出就是图像，中间不必再有一层应用。

Then I saw the Software 3.0 version of this, which blew my mind, which is literally just take your photo, give it to Gemini, and say use Nano Banana to overlay the things onto the menu. Nano Banana basically returned an image that is exactly the picture of the menu that I took, but it actually put into the pixels — it rendered the different things in the menu. This blew my mind because actually all of my MenuGen is spurious. It’s working in the old paradigm. That app shouldn’t exist. The Software 3.0 paradigm is a lot more raw. Your neural network is doing more and more of the work, and your prompt or context is just the image, and the output is an image, and there’s no need to have any of the app in between.

人们得换框：别把它当成现有东西的加速。现在出现的是以前不存在的新事物。这甚至不只是编程变快，而是更一般的信息处理现在可自动化了。以前的代码作用在结构化数据上；而我的 LLM 知识库项目是让模型给组织或个人做 wiki——这甚至不是一个程序，以前根本不存在「根据一堆事实编译出知识库」的代码。现在你可以把文档重新编译、重排，得到新的 reframing。我觉得更兴奋的是那些以前不可能的机会，而不只是旧事更快。

People have to reframe — not to work in the existing paradigm of what things existed and just think about it as a speed-up of what exists. It’s actually new things that are available now. Going back to your programming question, that’s also an example of working in the old mindset, because it’s not just about programming becoming faster. This is more general information processing that is automatable now. Previous code worked over structured data. With my LLM knowledge bases project, you get LLMs to create wikis for your organization or for you in person. This is not even a program. This is not something that could exist before, because there was no code that would create a knowledge base based on a bunch of facts. Now you can take these documents and recompile them in a different way. I keep trying to get back to: not only what can we do that existed that is faster now, but new opportunities of things that couldn’t be possible before. I almost think that that’s more exciting.

Stephanie Zhan

我喜欢你把 MenuGen 这条演进摊开。如果再往外推：1990 年代建网站、2010 年代做移动应用、上一轮云时代做 SaaS——2026 年那个事后会显得完全显而易见、但今天大多还没建出来的东西是什么？

I love the MenuGen progression. If you extrapolate that further, what is the 2026 equivalent for building websites in the ’90s, building mobile apps in the 2010s, building SaaS in the last cloud era? What will look completely obvious in hindsight that is still mostly unbuilt today?

Andrej Karpathy

还是顺着 MenuGen：很多这种代码不该存在，神经网络在干大部分活。外推看起来会很怪。你可以想象完全神经化的计算机：设备吃进原始视频或音频，进神经网络，用扩散渲染出只属于那一刻的 UI。

Going with the example of MenuGen — a lot of this code shouldn’t exist and it’s just neural networks doing most of the work. I do think that the extrapolation looks very weird, because you could imagine completely neural computers in a certain sense. Imagine a device that takes raw videos or audio into basically a neural net and uses diffusion to render a UI that is unique for that moment.

计算早期，人们其实有点搞不清计算机会更像计算器，还是更像神经网络。五六十年代并不显然会走哪条路。我们走了计算器这条路，建成了经典计算，神经网络现在是跑在现有计算机上的虚拟化。你可以想象这会翻转：神经网络变成宿主机程，CPU 变成协处理器。

In the early days of computing, people were a little bit confused as to whether computers would look like calculators or computers would look like neural nets. In the ’50s and ’60s it was not really obvious which way would go. Of course we went down the calculator path and ended up building classical computing, and then neural nets are currently running virtualized on existing computers. You could imagine that a lot of this will flip, and that the neural net becomes the host process and the CPUs become the co-processor.

我们看过那张图：智能算力、神经网络的 FLOPs 支出会占主导。你可以想象一种很陌生的世界：神经网络做大部分重活，工具调用只是给某些确定性任务留的历史附肢，真正主事的是联网的神经网络。外推会非常外来，但大概会一块一块走到那儿。那条路径现在还 TBD。

We saw the diagram of intelligence compute — neural networks are going to take over and become the dominant spend of FLOPs. You could imagine something really weird and foreign where neural nets are doing most of the heavy lifting. They’re using tool use as this historical appendage for some kinds of deterministic tasks. But what’s really running the show is these neural nets that are networked in a certain way. You can imagine something extremely foreign as the extrapolation, but I think we’re going to probably get there piece by piece. That progression is TBD, I would say.

Stephanie Zhan

想聊聊可验证性：AI 会更快、更容易自动化那些输出能被核验的领域。如果这个框架对，哪些工作会比人们以为的更快被搬走？哪些职业大家觉得安全，其实高度可验证？

I’d like to talk about this concept of verifiability — the fact that AI will automate faster and more easily domains where the output can be verified. If that framework is right, what work is about to move much faster than people realize, and what professions do we have that people actually think are safe but that are actually highly verifiable?

Andrej Karpathy

我写过可验证性。传统计算机容易自动化你能用代码写死的东西；这一轮 LLM 容易自动化你能核验的东西。因为前沿实验室训练这些模型时，它们是巨大的强化学习环境，给的是核验奖励。于是模型长成锯齿状实体：在数学、代码及邻近可验证域里能力出尖，出了这个空间就比较毛糙。

I spent some time writing about verifiability. Traditional computers can easily automate what you can specify in code, and this latest round of LLMs can easily automate what you can verify. When frontier labs are training these LLMs, these are giant reinforcement learning environments. They are given verification rewards, and because of the way these models are trained they end up progressing and creating these jagged entities that really peak in capability in verifiable domains like math and code and adjacent, and stagnate and are a little bit rough around the edges when things are not in that space.

我写这个，是想理解它们为什么这么锯齿。一部分跟实验室怎么训有关，一部分跟实验室往数据分布里塞了什么有关。经济上更值钱的东西会催生更多环境，因为实验室想在那些设定里工作。代码就是好例子。大概还有很多可验证环境没进配料表，只是因为那种能力没那么有用。

The reason I wrote about verifiability is I’m trying to understand why these things are so jagged. Some of it has to do with how the labs train the models, but some of it also has to do with the focus of the labs and what they happen to put into the data distribution. Some things are significantly more valuable in the economy and end up creating more environments because the labs wanted to work in those settings. Code is a good example of that. There’s probably lots of verifiable environments they could think about that happen not to make it into the mix because they’re just not that useful to have the capability around.

以前最爱举的例子是 strawberry 里有几个字母，模型会著名地答错，这就是锯齿。现在大概补上了。新的例子是：我要去洗车，只有 50 米，该开车还是走路？今天的 SOTA 会让你走，因为太近了。SOTA 的 Opus 4.7 能同时重构十万行代码库、找零日漏洞，却叫我走去洗车。这太疯了。

The favorite example for a while was how many letters are in a strawberry, and the models would famously get this wrong — an example of jaggedness. The models now patch this, I think, but the new one is: I want to go to a car wash to wash my car and it’s 50 meters away. Should I drive or should I walk? And state-of-the-art models today will tell you to walk because it’s so close. How is it possible that state-of-the-art Opus 4.7 will simultaneously refactor a 100,000-line codebase or find zero-day vulnerabilities and yet tells me to walk to this car wash? This is insane.

只要它们还锯齿，就说明：一，可能有点不对；二，你得在环里，把它们当工具，得盯着它们在干什么。长话短说，可验证性那组文章就是想找锯齿有没有规律。我觉得是「可验证」加上「实验室在意」的组合。

To whatever extent these models remain jagged, it’s an indication that number one, maybe something’s slightly off, or number two, you need to actually be in the loop a little bit and treat them as tools and stay in touch with what they’re doing. All of my writing about verifiability is just trying to understand why these things are jagged — is there any pattern to it? I think it’s some kind of combination of verifiable plus labs care.

还有一个有用的轶事：从 GPT-3.5 到 GPT-4，大家发现象棋强了很多，以为只是能力在涨。其实据我在网上看到的公开信息，是大量象棋数据进了预训练。有人在 OpenAI 决定加这份数据，于是这项能力就出尖了。我们有点受制于实验室往配料里放了什么。这东西没有说明书，你得自己探：落在 RL 回路里就飞，落在数据分布外就挣扎。不在那些回路里，就得自己微调，别指望开箱即用。

One more anecdote that is instructive: from GPT-3.5 to GPT-4, people noticed that chess improved a lot, and a lot of people thought it’s just a progression of the capabilities. Actually it’s more that — I think this is public information, I saw it on the internet — a huge amount of data of chess made it into the pre-training set. Someone at OpenAI decided to add this data and now you have a capability that just peaked a lot more. We are slightly at the mercy of whatever the labs are doing, whatever they happen to put into the mix. You have to actually explore this thing that they give you that has no manual. If you’re in the circuits that were part of the RL, you fly. If you’re in the circuits that are out of the data distribution, you’re going to struggle. If you’re not in the circuits, then you have to really look at fine-tuning and doing some of your own work, because it’s not going to necessarily come out of the LLM out of the box.

Stephanie Zhan

如果今天你是创始人，想做一件你觉得可解、领域可验证的事，可环顾四周，实验室已经在数学、编程这些最显眼的地方进入逃逸速度。你会对在座创始人说什么？

If you are a founder today and thinking about building a company, you are trying to solve a problem that you think is tractable, a domain that is verifiable, but you look around and think the labs have really started getting to escape velocity in the ones that seem most obvious — math, coding, and others. What would your advice be to the founders in the audience?

Andrej Karpathy

可验证性让一件事在当前范式里可解，因为你可以往上砸大量 RL。即便实验室没有直接盯着它，这一点仍然成立。如果你处在可验证设定里，能造 RL 环境或例子，那就具备自己微调、从中受益的条件。这套技术根本就能用：有大量多样的 RL 环境，拉一下你喜欢的微调框架，就能得到相当能用的东西。我不想在台上 vague post，但确实有一些很有价值、我觉得还没进配料表的强化学习环境。

Verifiability makes something tractable in the current paradigm because you can throw a huge amount of RL at it. That remains true even if the labs are not focusing on it directly. If you are in a verifiable setting where you could create these RL environments or examples, that actually sets you up to potentially do your own fine-tuning and you might benefit from that. That is fundamentally technology that just works. You can pull a lever if you have a huge amount of diverse data sets of RL environments. You can use your favorite fine-tuning framework and pull the lever and get something that actually works pretty well. I don’t want to give away the answer, but there is one domain — sorry, I don’t mean to vague-post on stage, but there are some examples of this.

Stephanie Zhan

反过来，哪些事远远看去好像能自动化，其实还只是远看？

On the flip side, what do you think still feels automatable only from a distance?

Andrej Karpathy

我最终觉得几乎所有东西都能在某种程度上做成可验证，只是有的更容易。写作之类，你也可以想象一群 LLM 评委，大概能弄出还过得去的东西。所以更多是难易，而不是能不能。最终——所有东西都能自动化。

I do think that ultimately almost everything can be made verifiable to some extent, some things easier than others. Even for things like writing, you can imagine having a council of LLM judges and probably get something reasonable out of this kind of approach. So it’s more about what’s easy or hard. I do think that ultimately everything is automatable.

Stephanie Zhan

去年你造了 vibe coding，今天的世界更严肃一点，更像智能体工程。两者差别是什么？你把我们现在所处的阶段叫什么？

Last year you coined the term vibe coding and today we’re in a world that feels a little bit more serious, more agentic engineering. What do you think is the difference between the two and what would you actually call what we’re in today?

Andrej Karpathy

vibe coding 是把每个人能用软件做成的事的地板抬起来。地板升了，人人都能 vibe code 任何东西，这很惊人。智能体工程则是保住以前专业软件的质量条：你不能因为 vibe coding 就引入漏洞；你仍然对软件负责，只是能不能更快。剧透：能，但怎么做才对。

Vibe coding is about raising the floor for everyone in terms of what they can do in software. The floor rises, everyone can vibe code anything, and that’s amazing, incredible. But then I would say agentic engineering is about preserving the quality bar of what existed before in professional software. You’re not allowed to introduce vulnerabilities due to vibe coding. You’re still responsible for your software just as before, but can you go faster? And spoiler is you can, but how do you do that properly?

我叫它智能体工程，是因为我觉得它就是一门工程学科。这些智能体是尖的实体，有点会出错，有点随机，但极强。你怎么协调它们更快，同时不牺牲质量条——把这件事做对，就是智能体工程的地盘。一个抬地板，一个往外推天花板。我看到智能体工程能力的天花板非常高。以前说 10x 工程师，现在放大得更多，你获得的加速不是 10x。在我看来，真正擅长的人峰值远超 10x。

I call it agentic engineering because I do think it’s kind of like an engineering discipline. You have these agents which are these spiky entities. They’re a bit fallible, a little bit stochastic, but they are extremely powerful. How do you coordinate them to go faster without sacrificing your quality bar, and doing that well and correctly, is the realm of agentic engineering. One is about raising the floor, and the other is about extrapolating. What I’m seeing is there is a very high ceiling on agentic engineer capability. People used to talk about the 10x engineer previously. I think that this is magnified a lot more. 10x is not the speed-up you gain. It does seem to me like people who are very good at this peak a lot more than 10x from my perspective right now.

Stephanie Zhan

Sam Altman 去年在 AI Ascent 说过一句记得很牢的话：不同代人用 ChatGPT 的方式不同。三十多岁当搜索替代，十几岁则把它当上网入口。写代码的平行是什么？如果看两个人用 OpenClaw、Claude Code、Codex，一个你觉得一般，一个完全 AI 原住民，差别在哪？

When Sam Altman came to AI Ascent last year, one memorable thing he said was that people of different generations use ChatGPT differently. If you’re in your 30s, you use it as a Google search replacement, but if you’re in your teens, ChatGPT is your gateway to the internet. What is the parallel here in coding today? If we were to watch two people code using OpenClaw, Claude Code, Codex — one you’d consider mediocre at it and one you would consider fully AI-native — how would you describe the difference?

Andrej Karpathy

就是把现有工具吃到最满：用上所有功能，投资自己的 setup。以前工程师会把 Vim 或 VS Code 用到极致，现在是 Claude Code 或 Codex。看起来就是这样。

I think it’s just trying to get the most out of the tools that are available, utilizing all of their features, investing into your own setup. Just like previously all the engineers are used to getting the most out of the tools you use, whether it’s Vim or VS Code, or now it’s Claude Code or Codex. Just investing into your setup and utilizing a lot of the tools that are available to you. I think it just kind of looks like that.

相关的一点是：很多人想招强的智能体工程师，但我看到大多数人还没把招聘改成面向智能体工程能力。还在发解谜题，那是旧范式。招聘得变成：给我一个很大的项目，看人把它做出来。比如说给智能体写一个 Twitter 克隆，做得很安全，再让一些智能体在上面模拟活动，然后我用 10 个 Codex 5.4 xhigh 去攻你部署的网站，它们不该攻得破。大概是看人在这种设定里做更大的项目、把工具用起来。

A related thought is a lot of people are hiring for this, because they want to hire strong agentic engineers. What I’m seeing is that most people have still not refactored their hiring process for agentic engineer capability. If you’re giving out puzzles to solve, this is still the old paradigm. I would say hiring has to look like: give me a really big project and see someone implement that big project. Let’s write, say, a Twitter clone for agents, and then make it really good, make it really secure, and then have some agents simulate some activity on this Twitter. And then I’m going to use 10 Codex 5.4 xhigh to try to break this website that you deployed, and they should not be able to break it. Watching people in that setting and building bigger projects and utilizing the tooling is maybe what I would look at for the most part.

Stephanie Zhan

智能体做的事越多，哪些人类技能会变得更值钱，而不是更不值钱？

And as agents do more, what human skill do you think becomes more valuable, not less?

Andrej Karpathy

现在的答案是：智能体还像实习生。你仍然得负责审美、判断、品味，以及一点点监督。我最喜欢的一个古怪例子是 MenuGen：你用 Google 账号注册，用 Stripe 买额度，两边都有邮箱。我的智能体买额度时，试图用 Stripe 邮箱去对 Google 邮箱，没有持久的用户 ID。你完全可以用不同邮箱，钱就对不上。

Right now the answer is that the agents are kind of like these intern entities. You basically still have to be in charge of the aesthetics, the judgment, the taste, and a little bit of oversight. One of my favorite examples of the weirdness of agents is for MenuGen: you sign up with a Google account, but you purchase credits using a Stripe account, and both of them have email addresses. My agent, when you purchase credits, assigned it using the email address from Stripe to the Google email address. There wasn’t a persistent user ID. It was trying to match up the email addresses, but you could use a different email address for Stripe and Google and it would not associate the funds.

这种事它们仍会搞错：为什么用邮箱去交叉关联资金？邮箱可以任意。人必须负责规格和计划。我其实不太喜欢 plan mode——当然有用——但更一般的是：你得跟智能体一起设计一份很细的规格，也许就是文档，让智能体去写。你管监督和顶层分类，它们做大量底层。你不再操心某些细节。

This is the kind of thing that these agents still will make mistakes about. Why would you use email addresses to try to cross-correlate the funds? They can be arbitrary. People have to be in charge of this spec, this plan. I actually don’t even like the plan mode. Obviously it’s very useful, but I think there’s something more general here where you have to work with your agent to design a spec that is very detailed, and maybe basically the docs, and then get the agents to write them. You’re in charge of the oversight and the top-level categories, but the agents are doing a lot of the under the hood. You’re not caring about some of the details.

神经网络里的数组、张量也一样。PyTorch、NumPy、pandas 那些 API 细节——keepdim 还是 keep_dims，dim 还是 axis，reshape、permute、transpose——我已经不记得了，因为不必记得。实习生记性很好。但你仍得懂：底下有存储，有 view，同一块存储可以变 view，换存储会更慢。你得有基本理解，才不会无谓拷内存。API 细节交出去；你负责品味、工程、设计，确保在问对的东西，确保用唯一用户 ID 把一切拴住。你做设计和开发，智能体填空。我们现在大概就在这里。

As an example also with arrays or tensors in neural networks, there’s a ton of details between PyTorch and NumPy and pandas and so on for all the different little API details. I already forgot about the keep dims versus keep dim, or whether it’s dim or axis, or reshape or permute or transpose. I don’t remember this stuff anymore, because you don’t have to. This is the kind of details that are handled by the intern because they have very good recall. But you still have to know that there’s an underlying tensor, there’s an underlying view, and then you can manipulate a view of the same storage or you can have different storage which would be less efficient. You still have to have an understanding of what this stuff is doing and some of the fundamentals so that you’re not copying memory around unnecessarily. The details of the APIs are now handed off. You’re in charge of the taste, the engineering, the design, and that it makes sense, and that you’re asking for the right things, and that these have to be unique user IDs that we’re going to tie everything to. You’re doing some of the design and development and the engineers are doing the fill in the blanks. That’s currently kind of where we are.

Stephanie Zhan

品味和判断会不会随时间变得不那么重要，还是天花板会一直升？

Do you think there’s a chance that this taste and judgment matters less over time, or will the ceiling just keep rising?

Andrej Karpathy

我希望它会改善。现在改善不了，大概还是因为不在 RL 里：没有审美的代价或奖励，或者还不够好。你真去看代码，有时会心跳漏一拍：并不总是很好的代码，很臃肿，大量复制粘贴，抽象又脆又别扭。能跑，但很恶心。希望未来模型能好一些。

I’m hoping that it improves. I think probably the reason it doesn’t improve right now is again it’s not part of the RL. There’s probably no aesthetics cost or reward, or it’s not good enough or something like that. When you actually look at the code, sometimes I get a little bit of a heart attack because it’s not like super amazing code necessarily all the time, and it’s very bloated, and there’s a lot of copy-paste, and there’s awkward abstractions that are brittle. It works but it’s just really gross. I do hope that this can improve in future models.

另一个例子是我的 microGPT 项目，想把 LLM 训练简化到不能再简。模型讨厌这个，做不到。我不停提示「再简化、再简化」，它就是不能。你会感觉自己在 RL 回路外面，像在拔牙，不是光速。人现在仍得管这件事，但没有什么根本障碍，几乎只是实验室还没做。

A good example also is this microGPT project, where I was trying to simplify LLM training to be as simple as possible. The models hate this. They can’t do it. I kept trying to prompt an LLM to simplify more, simplify more, and it just can’t. You feel like you’re outside of the RL circuits. It feels like you’re pulling teeth. It’s not like light speed. I do think that people still remain in charge of this. But I do think that there’s nothing fundamental that’s preventing it. It’s just the labs haven’t done it yet almost.

Stephanie Zhan

想回到锯齿智能。你写过一篇很刺激的「动物 vs 幽灵」：我们不是在造动物，是在召唤幽灵。这些锯齿智力由数据和奖励函数塑形，没有内在动机、好玩、好奇或赋能——那些是演化出来的。这个框架为什么重要？它会怎样改变你构建、部署、评估甚至信任它们的方式？

I’d love to come back to this idea of jagged forms of intelligence. You wrote a little bit about this with a very thought-provoking piece around animals versus ghosts. The idea is that we’re not building animals, we are summoning ghosts. These are jagged forms of intelligence that are shaped by data and reward functions, but not by intrinsic motivation or fun or curiosity or empowerment — things that kind of came about via evolution. Why does that framing matter, and what does it actually change about how you build and deploy and evaluate or even trust them?

Andrej Karpathy

我写这个，是想搞清楚它们是什么。如果你对它们是什么、不是什么有个好模型，用起来会更胜任。我不确定它有没有真正的「力量」，有点哲学化。但关键是承认：它们不是动物智力。你对它们大喊，不会让它们更好或更差，没有那种影响。底子是预训练的统计模拟回路，上面再螺栓上 RL，把某些凸起加高。更多是一种心态：什么可能管用、什么不管用、怎么改。我没有「让系统变好的五条显然结论」，更多是对它保持怀疑，再慢慢摸。

The reason I wrote about this is because I’m trying to wrap my head around what these things are. If you have a good model of what they are or are not, then you’re going to be more competent at using them. I’m not sure if it actually has like real power. I think it’s a little bit of philosophizing. But I think it’s just coming to terms with the fact that these things are not animal intelligences. If you yell at them, they’re not going to work better or worse. It doesn’t have any impact. It’s all just kind of like these statistical simulation circuits where the substrate is pre-training, so like statistics, and then there’s RL bolting on top, so it kind of increases the appendages. Maybe it’s just kind of like a mindset of what I’m coming into, or what’s likely to work or not likely to work, or how to modify it. I don’t know that I have “here are the five obvious outcomes of how to make your system better.” It’s more just being suspicious of it and figuring out over time.

Stephanie Zhan

你已经深泡在不只会聊天的智能体里：它们有真实权限、本地上下文，会替你行动。等我们都活进那个世界，世界会长什么样？

You are so deep in working with agents that don’t just chat. They have real permissions. They have local context. They actually take action on your behalf. What does the world look like when we all start to live in that world?

Andrej Karpathy

在座很多人大概都兴奋于原生智能体环境：一切都必须重写。现在的东西根本上仍是为人写的，得挪过去。我用的框架、库，文档根本上仍是给人看的。这是我最烦的：为什么还在告诉我该做什么？我不想做任何事。该复制粘贴给我智能体的那一段是什么？每次让我去某个 URL，我都啊一声。

A lot of people probably here are excited about what this native agentic environment looks like, and everything has to be rewritten. Everything is still fundamentally written for humans and has to be moved around. Most of the time when I use different frameworks or libraries, they still have docs that are fundamentally written for humans. This is my favorite pet peeve. Why are people still telling me what to do? I don’t want to do anything. What is the piece of text I should copy-paste to my agent? Every time I’m told, go to this URL or something like that, it’s just like ah.

大家都在想：怎么把该发生的工作拆成对世界的传感器和对世界的执行器，怎么做成 agent-native，先讲给智能体听，再围绕 LLM 容易读的数据结构做自动化。我希望有大量 agent-first 的基础设施。

Everyone is excited about how do we decompose the workloads that need to happen into fundamentally sensors over the world, actuators over the world. How do we make it agent native — basically describe it to agents first — and then have a lot of automation around data structures that are very legible to the LLMs. I’m hoping that there’s a lot of agent-first infrastructure out there.

MenuGen 那篇博客——不知道算不算出名——真正的麻烦甚至不是写代码，是部署到 Vercel：串各种服务，进设置菜单，配 DNS，烦死了。我希望的测试是：给 LLM 一句「做 MenuGen」，我不用再碰任何东西，它就那样部署到网上。那才说明基础设施在变成 agent-native。

For MenuGen, when I wrote the blog post about MenuGen, a lot of the trouble was not even writing the code for MenuGen. It was deploying it on Vercel, because I had to work with all these different services and string them up and go to their settings and the menus and configure my DNS, and it was just so annoying. That’s a good example of: I would hope that I could give a prompt to an LLM, build MenuGen, and then I didn’t have to touch anything and it’s deployed in that same way on the internet. I think that would be a good kind of a test for whether or not a lot of our infrastructure is becoming more and more agent native.

最终我们会走向人和组织都有智能体代表的世界：我的智能体跟你的智能体去把会议细节谈妥。大概是往那儿去，在座也都很兴奋。

Ultimately I do think we’re going towards a world where there’s agent representation for people and for organizations, and I’ll have my agent talk to your agent to figure out some of the details of our meetings or things like that. I do think that that’s roughly where things are going, and I think everyone here is excited about that.

Stephanie Zhan

必须用教育收尾。你大概是世界上最会把复杂技术讲简单的人之一，也认真想过教育该怎么设计。智力变便宜、进入下一阶段 AI 之后，哪些东西仍然值得学深？

I think we have to end on a question about education, because you are probably one of the very best in the world at making complex technical concepts simple and deeply thoughtful about how we design education around it. What still remains worth learning deeply when intelligence gets cheap as we move into the next era of AI?

Andrej Karpathy

最近有一条推把我震住了，我隔天还在想。大意是：你可以外包思考，但不能外包理解。说得真好。我仍然是系统的一部分，信息仍然得进我的脑子。我感觉自己正在成为瓶颈：我们到底要建什么、为什么值得做、我怎么指挥智能体。最终得有东西来引导思考和处理，而这仍然受理解约束。

There was a tweet that blew my mind recently, and I keep thinking about it like every other day. It was something along the lines of: you can outsource your thinking, but you can’t outsource your understanding. I think that’s really nicely put. I’m still part of the system, and information still has to make it into my brain, and I feel like I’m becoming a bottleneck of just even knowing what are we trying to build, why is it worth doing, how do I direct my agents, and so on. I do still think that ultimately something has to direct the thinking and the processing, and that’s still kind of fundamentally constrained somehow by understanding.

这也是我很兴奋 LLM 知识库的原因：那是我处理信息的方式。每次看到信息的另一种投影，我都觉得有洞察。对我来说就是在固定数据上做很多合成数据生成的提示。我读一篇文章，wiki 就从这些文章里长起来，我爱问问题。这些工具是在增强理解。这仍是瓶颈，因为你没法当好导演——LLM 并不擅长理解，你仍然独特地负责这件事。朝这个方向的工具，我觉得极其有意思。

This is one reason I also was very excited about all the LLM knowledge bases, because I feel like that’s a way for me to process information. Anytime I see a different projection onto information, I always feel like I gain insight. It’s really just a lot of prompts for me to do synthetic data generation over some fixed data. Whenever I read an article I have my wiki that’s being built up from these articles, and I love asking questions about things. Ultimately these are tools to enhance understanding in a certain way. This is still kind of like a bit of a bottleneck, because you can’t be a good director — the LLMs certainly don’t excel at understanding. You still are uniquely in charge of that. Tools to that effect I think are incredibly interesting and exciting.

Stephanie Zhan

我很期待几年后再来，看我们是不是已经被完全自动化出环，它们连理解也接管了。非常感谢 Andrej 来。

I’m excited to be back here in a couple years and to see if we’ve been fully automated out of the loop and they actually take care of understanding as well. Thank you so much for joining us, Andrej. We really appreciate it.