生活在未来：AI、技艺与写作的艺术

欢迎回到播客。感谢你来做客。再用三个词说说你现在感觉怎么样？

Welcome back to the podcast. Thanks for having me. How are you feeling again in three words?

我觉得好奇、平静，还有一点肩颈的紧张，但已经开始消退了。所以感觉不错。

I am feeling curious and calm and I had a little bit of stress in my traps but it has started to fade. So I feel good.

太好了。也许我们再聊两件事，它就完全消失了。

Great. Maybe we'll do two more things and it'll fade completely.

我觉得你现在就生活在未来。我想聊聊 AI，但在那之前，安妮·迪拉德，你为什么喜欢她？为什么把她的书送给员工？

So I feel like you are currently living in the future. And I want to talk about AI but before we go there, Annie Dillard, why do you love her and why do you give her a book to your employees?

是的，我爱安妮。我爱安妮·迪拉德。我给每个员工都发《写作生涯》，那是她写写作的书。她出了名地讨厌这本书，觉得除了最后一章，其他都是垃圾。其实我喜欢这本书的其他部分，我觉得她对自己太苛刻了。

Yes, I love Annie. I love Annie Dillard. I give everyone at every I give them The Writing Life which is her book about writing. She famously hates that book and thinks it's trash except for the last chapter which I actually like the rest of the book. I think she's a little harder on herself.

她更喜欢另一本书。

She prefers a book.

她没写她的书。

She has she did not write her book.

哦，那是安妮·迪拉德。对，安妮们容易搞混。是的。

Oh that's Annie Dillard. Yeah, confusing Annie's. Yeah.

她觉得除了最后一章，书的其他部分都是垃圾。我觉得这本书其实全部都很不错，但我让每个人都读最后一章，不读其他部分，然后我们一起讨论。我觉得她很了不起，因为她的文字极其诗意、抒情。她写自然写得美极了。如果你喜欢梭罗、爱默生之类的人，她绝对深深属于那个传统。在她几乎所有的作品里，她都在思考她感受到的世界有多美，同时世界有时又有多残酷、多可怕，以及在一个存在上帝的世界里，这一切如何共存。你能在她所有的书里看到这条线索。说实话，我觉得《写作生涯》是最不以上帝为中心的一本。里面有一点，但只是轻轻点缀，我觉得这适合作为安妮·迪拉德的入门。我不会把你直接扔进深水区，让人去读她的《圣洁的坚固》或《暂时》之类的书。我会把这类书留到以后。但我爱《写作生涯》，尤其是最后一章，因为这是一本关于写作、关于成为作家意味着什么、什么是好写作的书，但最后一整章只是一个故事，一个关于特技飞行员戴夫·拉姆的真实故事，讲的是她几年间与他的几次相遇。我觉得这是对“擅长某事”与“创造艺术”之间区别的探索，谁有资格创造艺术，因此谁有资格成为作家，以及作家、艺术家与未来、现在之间的关系。所以如果你把它拆开看，它和我们 Every 所做的一切都息息相关。

She thinks the rest of the book is trash except for the last chapter. I think the book is actually quite good all of it but I make everyone read the last chapter, not the rest of it, and we talk about it. I mean I think she's incredible because her writing is extremely poetic, it's lyrical. She writes incredibly beautifully about nature. If you like Thoreau or Emerson or anything like that, she's definitely deeply in that tradition. And in pretty much all of her work she is grappling with her sense of how beautiful the world is and also how cruel and horrible sometimes the world is, and how that can exist in a world where God exists. You can sort of see that thread through all of her books. I think The Writing Life is the least God-centric one to be honest. There's a little bit in there, but it's sort of sprinkled lightly, which I think is appropriate for your first introduction to Annie Dillard. I wouldn't throw you in the deep end and have people read one of her books called Holy the Firm, another one's called For the Time Being. I would leave those kinds of books to later. But I love The Writing Life, the last chapter of The Writing Life, because it's a book about writing and what it means to be a writer and what good writing is and all that kind of stuff, but the entire last chapter is just a story, a true story about a stunt pilot. His name is Dave Rahm, and it's an exploration of her encounters with him over a couple of years. And I think it's an exploration of the difference between being skilled at something and making art, and who gets to make art and therefore who gets to be a writer, and the relationship between writers and artists and the future and the present. So if you unwound it, it's very relevant to everything that we do at Every.

嗯。

Yeah.

但表面上它只是一个关于特技飞行员的故事，我喜欢这一点。

But on the surface it's just a story about a stunt pilot and I love that.

是的，太棒了。太美了。我觉得我想做这次对话的原因之一是，你对技艺、美和写作、伟大写作有如此敏锐的感觉。我认为表面上，那种直觉和深刻的人性意识与深入 AI 之间存在张力。对很多人来说，他们觉得那是两条不同的路。

Yeah, that's amazing. That's beautiful. I feel like one of the reasons I wanted to have this conversation is you have such an attuned sense for craft and for beauty and writing, great writing. And I think on the face of it there is a tension between that intuition and that deep sense of humanity and then going deep into AI. And for a lot of people they feel that those are two separate paths.

嗯。

Yeah.

但我觉得你清楚地展示了它们可以结合。

But I think you are clearly showing that they can be combined.

嗯。

Yeah.

那你对此怎么看？AI 在哪些方面加深了你进入那种更人性化、更实在的技艺层面的能力？

So I guess how do you feel about that? And what are some ways that AI has deepened your ability to go into that kind of more human, tangible, craft side?

这真的很有意思。这是个很好的问题。我确实觉得那是我的小舒适区，或者说我喜欢玩的地方，你知道吗？

That's really interesting. That's a very good question. I do think that that's my little comfort zone or where I like to play, you know?

嗯。

Yeah.

而且我要说，肯定有权衡。用 AI 和不用 AI 是不同的，我确实会花时间手写日记，比如完全不用 AI。

And I will say there are certainly trade-offs. It's different using AI versus not, and I do spend time just handwriting in journals for example without any AI.

嗯。

Mhm.

虽然即使在做那些事的时候，我也有点想念它。但这是有成效、有帮助的，而且我注意到，当我独自在房间里拿着笔记本时，与协调智能体时，我的大脑运作方式非常不同。

Although even when I'm doing that I kind of miss it. But it's productive and helpful, and I notice the way my brain works when I'm alone in a room with a notebook versus coordinating agents is very different.

嗯。嗯。

Yeah. Yeah.

我必须小心地两者兼顾。

I have to be careful to have both.

为什么？

What like why?

我觉得现在的智能体有点老虎机效应，比如这个提示词，就是那个能成的。

I think the agents right now can have a little bit of the slot machine effect of like this is the prompt, this is the one that's going to do it.

让我赚一百万。

Make me a million bucks.

对，没错。让我赚一百万、十亿。我就想，在这种经济形势下？

Yeah, exactly. Make me a million, a billion. I'm like, in this economy?

所以，即使它们技术上为你做了很多工作，也很难把目光移开。

And so it can be hard to even though they're technically doing a lot of the work for you, it can be hard to look away.

嗯。

Mhm.

然后你就在它们之间来回切换，感觉就像刷 TikTok，我不喜欢那样。我发现如果我那样做，大脑会有点烧焦。我无法以同样的方式集中注意力。我无法以同样的方式深入阅读、写作或深度思考。

And then you're flipping back and forth between them and then it just sort of feels like TikTok and I don't like that. And I find if I'm doing that it sort of fries my brain a little bit. I can't focus in the same way. I can't get deep into reading or writing or deep thinking in the same way.

嗯。

Mhm.

所以我要么在早上做更多深度工作，如果我在协调智能体，我会试着设一个限制，比如如果是早上，我就把它们放出去，然后我就去忘掉它，这很难做到，但我认为有可能。

So I either have to do more of my deeper stuff in the morning, and if I'm coordinating agents I try to set a limit of like if it's the morning, I'm just going to set you off and then I'm going to go forget about it, which is hard to do but I think is possible.

比如分开的房间，然后你走出去。

Like separate room and then you walk out.

是的。我的意思是，我现在做的一件事我很喜欢，就是我有一个家用电话，那只是我的一部旧 iPhone。每当我走进我的房子或公寓，我就把我的手机放在充电器上，把家用电话从充电器上拿下来，我一直用家用电话，它上面什么都没有，没人知道这个号码。它只有我的 Spotify 和几个应用，比如我有 Safari，这样我偶尔可以查点东西。它上面还有一件事，就是有 ChatGPT 和 Claude 的免费套餐，所以我从那个角度看它，这和我普通手机上尽可能堆叠的多个 Max 套餐非常不同。是的，它帮助我在离开它们时保持那种关注或存在于世界的方式。

Yeah. I mean I have one of the things I do right now which I love is I have a house phone which is just an old iPhone that I had, and whenever I go into my house or my apartment I just take my phone and I put it on a charger and take the house phone off the charger and I have the house phone all the time and it doesn't have anything, no one has the number. It just has like my Spotify and a couple of apps like I have Safari on it just so I can look things up every once in a while. It has one thing that it has is like the free plan of ChatGPT and Claude on it so I kind of see it from that perspective which is very different from the like as many max plans stacked together as I can possibly get which is what my regular phone is. And yeah, it helps me keep that way of attending or being in the world when I'm off of them.

我想我回去后会这么做。这是个好主意。真的很聪明。

I think I'm going to do that when I get back. That's a great idea. That's really smart.

嗯，你能够创造或构建什么？因为我觉得这里的另一个张力是，大多数人，我想说，或者很多听众，他们有过使用 ChatGPT 或 Claude 的聊天模式的体验，但他们没有深入使用 Codex 或 Claude Code。我觉得我们在某种意义上算是比较另类的。你能让人们了解一下，这些新的“超能力”现在能实现什么吗？

Um, what have you been able to create or build? Because I think the other tension here is that most people, I would say, or a lot of people listening, they've had an experience of using ChatGPT or Claude in the chat mode, but they haven't gone deep into Codex or Claude Code. I think we're kind of relatively weird in that sense. Can you give people a sense of like what is now possible with these new kind of superpowers?

我当然可以。我真的很喜欢你的问题，我觉得我还没有完全回答，那就是你如何……似乎创造美、人性以及对工艺的关注，与使用 AI 是相悖的。

I can definitely do that. I really liked your question, which I don't feel like I fully answered yet, which is how do you... It seems like the creation of beauty and the human-ness and attention to craft is at odds with using AI.

嗯。

Mhm.

我真的觉得这一点值得深入探讨。

And I really think that that's important to examine.

嗯。

Mhm.

而且我觉得，对我来说，深入探讨这个问题会让你学到很多关于“为人”的意义。具体来说，你第一次接触一项新技术，尤其是那种能完成你历来所做之事的技术，真的是一种非常迷幻的体验。我有句话，我们聊过，就是“永远不要在冥想静修、迷幻体验或第一次接触前沿模型后的 30 天内做任何重大人生决定”。我认为这是真的。而且我觉得，取决于你是谁，比如我的第一反应是“天哪，这太神奇了，我想做更多”。但很多人会感到厌恶或恐惧之类的。无论你是谁，它常常会让你忘记自己是谁。你会失去作为人类所拥有的复杂性和丰富性。你会想，它能做我能做的事，所以它就是我，我看着它时有点迷失了自己。我觉得这在技术史上相当常见。

And if I think for me, deeply examining that teaches you a lot about what it means to be human. Specifically, your first encounter with a new technology, especially one that does stuff that you historically have done, is just a very trippy experience. I have this thing that I say that we've talked about, which is like never make any major life decisions within 30 days of a meditation retreat, a psychedelic experience, or your first encounter with a frontier model. And I think that that's true. And I think people have, depending on who you are, like my first reaction is like holy, this is amazing. I want to do more. But a lot of people are like disgust or fear or all that kind of stuff. And no matter who you are, it often causes you to forget who you are. You sort of lose the complexity and richness you have as a human being. You're like, it can do something that I can do, therefore it is me, and I lose myself a little bit looking at it. And I think this is quite common in the history of technology.

我也这么认为。

I think so.

尤其是那些触及艺术形式的技术，会让人失去一点自我，并以某种恐惧或厌恶来回应，或者担心它会夺走艺术形式中的艺术性或工艺。一个简单的例子就是摄影与绘画。完全相同的争论。

To have, especially technologies that touch art forms, lose a bit of yourself and react with some sense of fear or disgust, or fear that it takes away the artistry or craftsmanship of whatever the art form is. An easy example is just photography versus painting. Same exact debate.

是的，完全同意。

Yeah, totally.

摄影在最初的大约 20 年里就被视为作弊。

Photography was just cheating for the first like 20 years.

对。

Right.

而这就是人们对它的看法。

And that's how it was perceived.

我不知道那段历史。

I didn't know the history of that.

嗯。

Yeah.

好的。

Okay.

所以画家们就说，去你们的。

So painters were like, screw you guys.

是的，没错。就像我花了那么多时间学习如何画这片田野，而你只要拍张照就得到了。我觉得这是合理的反应。

Yeah, exactly. Like I spent so much time learning how to paint this field and you're just going to take a picture and now you have it. I think a reasonable reaction.

嗯。

Yeah.

而且绘画依然存在。

And also painting's still around.

是的。

Yep.

而且摄影本身就是一种艺术形式。即使每个人手机里都有相机，要把一切都调好其实非常难。我认为 AI 也是如此。流行的看法是，你只要按一下按钮，它就能完成所有令人难以置信的工作，和艺术家或作家做的工作完全一样。这表面上确实如此。你第一次看到时，会想，哦，是的，它做的就是同样的事情。但即使有 AI，要真正做出好东西，也需要你全身心投入，把你的品味注入其中，以最好的方式使用它。所以我重新解读为，是的，你可以用 AI 做出看起来像别人作品的东西，或者看起来令人印象深刻，那是 5 年前令人印象深刻的东西。现在你可以用一个提示词就做到。这确实很了不起，但是……

And photography itself is an art form. Like getting it all dialed in is actually very hard even though everyone has a camera on their phone. And I think the same is true for AI. The popular picture is, yeah, you just press a button and it just does all this incredible work that's the exact same work as an artist does or a writer does. And that is sort of true on its face. You can kind of see how, if you look at it for the first time, you're like, oh yeah, it just does the same thing. But to actually make things good even with AI requires bringing yourself to it, infusing your taste into it, using it in the best way possible. And so I sort of reinterpret that as, yeah, you can use AI to make something that looks like someone else's work, or that looks impressive, what used to be impressive like 5 years ago. You can now do it in one prompt. And that's a really big deal, but...

我认为对于任何工具，你都可以用得好或不好。当有新工具出现时，就有新的机会去学习如何用好它们。AI 也不例外。工艺、意义和目的感并不来自工具，而是来自我们。你可能不知道如何将这些东西带入这个工具，你可能担心它们已经消失了。但如果你深入其中，你会发现它们会回来。而且你会学到一些关于它们从何而来、是什么产生了它们，这些并不取决于你做事的具体方式。需要说明的是，有些人可能永远喜欢手写或别的什么。我觉得这完全没问题。但是，能够深思熟虑、创造美或拥有工艺的能力可以应用于任何事情，AI 也包括在内。事实上，对我来说，这是表达美或工艺最有趣的方式，因为它是新的，而且在某种程度上是有争议的。这导致很多人忽视它或低估它，我认为这对那些想要突破界限、创造新艺术形式的人来说是一个巨大的机会。

I think for any tool, you can use it well or poorly. And when there are new tools, there's a new opportunity to learn to use them well. And AI's like no different. And the sense of craft and meaning and purpose does not come from the tool. It comes from us. And you may not know how to bring that to this tool, and you may be afraid that it's gone. But if you get into it, you will find it comes back. And you'll learn something about where it comes from and what generates it that is not about the specific way you were doing it. And to be clear, some people are just going to always love paper writing or whatever. I think that's totally fine. But the ability to be thoughtful or make something beautiful or have craft can be applied to anything, and AI's included in that. And in fact, for me, it is the most interesting way to make expressions of beauty or craft because it's new and it's to some degree controversial. And that causes, I think, a lot of people to ignore it or underestimate it, and I think that's a huge opportunity for people who want to push the boundaries a little bit to create new forms of art.

这太迷人了。太有趣了。谢谢你的回答。

This is fascinating. This is so interesting. Thank you for that answer.

嗯。

Yeah.

嗯，是的，我想到的是，我们前几天聊过，AI 就像一个非特定的放大器，它……

Um, yeah, what comes to mind for me is, and we talked about this the other day, of AI being like a non-specific amplifier in that it's...

说得太好了……你能定义一下吗？也许你播客的听众会知道那是什么意思，但你知道，当你说出那句话时，我就想，说得太对了。

Such a good... Can you define that? Like, maybe people on your podcast will know what that means, but you know when you said that, I was like, that's exactly right.

是的。这个概念来自迷幻世界，它放大“设定”和“环境”，基本上就是你的心态、你的信念、你内心的一切，然后是环境。我认为 AI 也在做类似的事情，那些倾向于恐惧或悲观的人会把这种情绪投射到 AI 上，然后这成为他们使用 AI 的体验。我认为还有另一群人，他们可能更好奇，更倾向于高能动性，他们使用 AI 时会觉得，“天哪，我感觉自己太有力量了，我感觉自己被解锁了。我这辈子从没写过一行代码，但现在我可以有一个想法，我可以创造东西，我可以构建东西。这太有趣了，感觉就像打开了一个我之前无法触及的新的工艺维度。”

Yeah. So it comes from the psychedelic world where it amplifies set and setting, which is basically your mindset, your beliefs, everything that is within you, and then the environment. And I think AI is kind of doing the same thing in that people who have a tendency towards fear or pessimism are projecting that into AI, and then that becomes their experience of using AI. And I think there's another group of people who are maybe more curious, more on the high agency things, who are using AI and being like, holy, I feel so empowered. I feel so unlocked. I've never written a line of code in my life, and now I can have an idea and I can create things, I can build things. And it's so much fun, and it feels like a new dimension of craft has been unlocked that I previously didn't have access to.

所以我想，A，你怎么看？B，一个人怎么……对我来说真正的关键在于，当很多主流媒体和大众对 AI 的认知是：它是一个外部力量，会来抢走工作、摧毁手艺、侵蚀人性——这种视角怎么能被分享出去？

So I guess like A, like what do you think about that? And B, how does one... And I think the real crux for me is like how can this perspective maybe be shared when a lot of the mainstream media and the way that AI is perceived is as this thing that is an outside force that is going to come and take jobs or destroy craft and erode away humanity?

这又是……我喜欢那种“非特定放大器”的说法，而且我再次认为，这是这些工具向我们展示自身本质的另一种方式。因为如果你遇到它，你会说“我超级好奇，我爱这个”。对我来说，这并不意味着你对一切都好奇，但如果你能用它做出真正有趣或酷的东西，那确实能说明你的一些特质。因为我确实认为你几乎可以用它做任何事。它是这样一种技术：因为它大致是在全世界的数据上训练的，它反映了我们所有人的多元性。但任何对提示词的特定回应，都是把那种无限的多元性坍缩成一个特定的东西，而你必须给它这个提示。它不会自己凭空产生。所以你把推向哪个方向，就反映了你是谁、你在想什么。所以我觉得这真的很有趣。而且，是的，我同意。我认为如果你害怕它或厌恶它，也是一样的——它说明了你重视什么、你如何处理新事物，这又不是道德评判。我认为有很多充分的理由去害怕或可能厌恶，取决于你是谁、你来自哪里，但它就是非特定的。在这个意义上，它是一个非特定放大器。就像你对这类事情的任何先验，都会被放大。

This is again a... I love the sort of non-specific amplifier thing, and I again think it's another way in which these tools show us something about ourselves. Because if you encounter it and you're like, "I'm super curious. I love this." I guess to me it doesn't mean that you're curious about everything, but it does say something about you if you can use it to build something really interesting or cool. Because I do think you can pretty much use it to do anything. It's one of these technologies that, because it is trained on the entire world's data more or less, it reflects all of us in all of our multiplicity. But any given response to a prompt is collapsing all of that infinite multiplicity into a particular thing that you have to give it. It doesn't just do it on its own. And so where you push it is going to be a reflection of who you are and what you think about. And so I think that's really interesting. And yeah, I agree. I think the same thing is true if you're afraid of it or disgusted by it or anything like that—it says something about what you value and how you process new things, which is again not a moral judgment. I think there's a lot of good reasons to be afraid or potentially disgusted, depending on who you are and where you come from, but it is non-specific. It is a non-specific amplifier in that sense. Like whatever your priors about this kind of thing are, it's going to be multiplied.

而且我认为，总的来说，AI 公司在定位这项技术和谈论它方面做得非常糟糕，因为这个领域在很多方面是建立在这个科幻梦之上的，而业内的人现在和过去都对此感到恐惧。

And I think that generally the AI companies have done a really poor job of positioning this technology and talking about it, because the field is founded in a lot of ways on this sci-fi dream that people in the industry are and were terrified of.

是的。

Yep.

而且大多数设定这种心理世界观的人——我不认为，嗯，我知道——他们没有太多经验去见证许多不同的技术浪潮，以及他们对早期技术的情绪反应、早期技术的梦想，最终是如何实现的。

And most of the people that sort of set the psychological worldview—I don't think, well, I know—they didn't have a lot of experience seeing lots of different technology waves and how their emotional reactions to early tech, the dream of early technology, how they actually pan out.

嗯，嗯。

Mhm. Mhm.

而且我认为他们也还不习惯在地缘政治的层面上博弈，以及 AI 公司现在所处的层面，那是平流层级别。

And I think also they are just not used to playing at the level of geopolitics and whatever the level is the AI companies right now, which is stratospheric.

Eric，我不认为他们不习惯那个。

Eric, I don't think that they're not used to that.

嗯，嗯。所以你能感觉到——我认为他们其实相当真诚——但你能感觉到一点 AI 精神错乱，无论是积极的还是消极的。

Mhm. Mhm. And so you can feel—I think they're actually being quite earnest—but you can kind of feel a little bit of the AI psychosis, either to the positive or to the negative.

嗯。

Mhm.

然后他们只是尽力去应对，就像说“我们会努力让它变好，但我们不知道”。

And then just trying their best to deal with it, which is like, "We're going to try to make it good, but we don't know."

5 年。

5 years.

是的。

Yeah.

顺便说一句，你的工作可能会消失，我不知道，但我们会努力让它变好。我认为这是真诚的，但相当……

And by the way, your job might go away, I don't know, but we're going to try to make it good. I think it's earnest, but it's pretty...

也许缺乏智慧。

Like lacking wisdom, maybe.

是的，是的，这并不特别明智。

Yeah. Yeah, it's not particularly wise.

是的。

Yeah.

而且这种反应是可以预见的。

And the reaction is predictable.

嗯。

Mhm.

而且我认为，即使你讨厌你的工作，一想到它会在你无法控制的情况下被夺走——你突然就爱上你的工作了。

And I think people, even if you hate your job, the idea that it's going to be taken away without your control—suddenly you love your job.

而且我真的——我的意思是，我确实认为可能有些工作会像冰箱出现时的送冰人那样消失，但总的来说不会。虽然我确实认为很多工作流程会改变。你把精力用在哪里、你在工作过程中相对于智能体处于什么位置，都会不同。但总的来说，我认为这将是一个赋能的转变，让人们能够做更多、做得更好，比他们通常能做到的更好，从“我是一个自由职业者”到“我在一家大公司工作”，所有这些人都将拥有更大的力量去做他们想做的事。

And I truly—I mean, I do think that there are probably some jobs that are going to go the way of the ice delivery man or whatever when the refrigerator came out, but generally not. Although I do think a lot of workflows are going to change. And where you apply your effort and where you are in the process of doing work versus an agent is going to be different. But it will, by and large, be—I think—an empowering shift for people to be able to do more and do better at work than they would have ordinarily been able to, across the spectrum from "I'm a freelancer on my own" to "I am working at a gigantic company." All of those people, I think, are going to have a lot more power to do the thing they want to do.

完全同意，完全同意。这是你做出的一个预测，我认为你说对了，至少到目前为止，AI 可能正在改变工作，但实际上创造的工作岗位比消失的还多。

Totally. Totally. Which is one of the predictions that you made that I think you got right, which was at least so far AI is—maybe jobs are changing, but there's actually more jobs being created than there are.

是的，我想——我没有具体评论过工作岗位的数量，但我的经验是——我经常思考自动化之后会发生什么。在我们自动化了事情之后，真实的体验是什么？因为两边都有大量的 FUD，比如它会是一个社会主义乌托邦，每个人都有基本收入，或者它会是一个地狱般的景象。而我试着真正——作为一家公司，我试着真正把它当作一个实验室，看看那实际上是什么样子，可能是什么样子？然后把它投射出去。对我们来说，我们大约 30 人，已经存在了 6 年半，所以在 AI 出现之前。对我来说，只是追踪技术的进步，因为我们采用了——我们在 GPT-3 时代就构建了超级 AI。所以现在那对我来说就像一百万年前的事了。我们已经自动化了一切可能自动化的东西。内部对自动化没有太多阻力——就像尽可能多做。是的，就像也许我们做得太多了，也许我们需要更多的人类参与。而我这样做的经验是，它实际上创造了更多的人类工作。我们仍然在招聘工程师。我们有一个客服人员。我们有所有那些你会说“是啊，你可能没有任何工程师，你可能没有任何客服人员，你可能只有两个人的团队”的东西。但事实是，我们有 30 个人。而且我们还在招人。你想为我们工作吗？请来吧。

Yeah, I think—I have not weighed in on it specifically number of jobs, but my experience is—I think a lot about what happens after automation. After we automate things, what is the real experience? Because there's a ton of FUD on both sides, like it'll be a socialist utopia where everyone has basic income, or it'll be the hellscape. And I try to really—as a company, I try to really use it as a laboratory for what does that actually look like, what could that actually look like? And then sort of project it out. For us, we're like 30-ish people and we've been around for 6 and a half years, so before AI came out. And for me, just tracking the progress of the technology because we adopted—we got super AI built around the GPT-3 days. So at this point that's like a million years ago in my head. And we've automated everything we possibly can. There's not a lot of resistance to automation inside of every—it's like let's do as much as we can. And yeah, it's like maybe we're doing this too much. Like maybe we need more human involvement. And my experience of doing that is it actually creates a lot more human work. We still hire engineers. We have a customer service person. We have all the things that you would be like, "Yeah, you probably don't have any engineers and you probably don't have any customer service people and you probably have a team of like two people." And it's like no, we have 30 people. And we're hiring more. Like you want to work for us? Please.

而且我认为，如果你是一个 AI 看涨者，这相当反直觉。但我也认为这很有道理。我想——回到非特定放大器的观点——我认为 AI 基本上是一个工具，它让你今天就能获得昨天所有的劳动成果，昨天所有的人类能力。

And I think that that is quite counterintuitive if you're an AI bull. And I think it also makes a lot of sense. I think about—to the point of being a non-specific amplifier—I think of AI as basically a tool that makes all of yesterday's work, all of yesterday's human competence, available to you today.

是的，就像“冻结的能力”，这个说法，对吧？

Yeah, like frozen competencies, the phrase, right?

正是如此。这非常重要。这真的很酷。

Exactly. And that's really important. It's really cool.

那种冻结的能力现在对所有人都可用了。这意味着，如果我不是程序员，我现在可以用昨天的冻结能力去构建我的应用，对吧？但我这样做的方式，因为都是基于昨天发生的一切，所以它对于我现在面临的具体问题来说，会有点不具针对性。所以它可能是一个令人印象深刻的一次性演示，但要让它变好，还需要更多。

And that sort of frozen competence that is now available to everyone. What it means is if I'm not a programmer, now I can use yesterday's frozen competence to go build my app, right? But the way I'm going to do that, because it's all based on everything that happened yesterday, it's going to be sort of non-specific to the particular problem that I'm facing right now. And so it may be an impressive one-shot demo, but to make it good it's going to need a lot more.

问题是，这“更多”从哪里来？

And the question is, where does that more come from?

这来自人类专家，他们进去说：“这是我们今天面临的实际问题。我会继续使用这种冻结能力，因为它真的很棒，但我会把它塑造成适合今天这个特定情况的形式。”因为每个人都能获得这种能力，用昨天的能力来制作今天的东西，这变成了一种商品。它变得非常便宜。任何人都能做一个一次性演示。但每个人都在做，所以实际上对专家的需求更大了，专家们要进来说：“现在我们有 10,000 个应用需要我改进。”

And it comes from human experts who go in and say, "Here's the actual problem that we're facing today. I'm going to continue using this frozen competence because it's really awesome, but I'm going to shape it in a way that fits this particular situation today." And because everyone has access to this, the ability to make something that uses yesterday's competence today is a commodity. It becomes really cheap. Anyone can make a one-shot demo. But everyone's doing it, so there's actually more demand for experts to come in and say, "Now we have 10,000 apps that I need to make better."

所以这奇怪地为专家创造了更多工作。你可能会说：“哦，这对专家来说很糟糕，因为他们现在只是在清理人们编码糟糕的应用。”但我实际上认为有两个机会。在某种程度上，在某些情况下，这会让专家的处境更糟，但我看这个问题的视角是，它为高技能的人创造了两个机会。一个机会是创建系统，帮助利用这些没有技能的人的新解锁能力，利用它们做好工作，并建立一个系统，让这些人走上正轨，尤其是在公司内部，比如说：“你有一个数据科学问题。我要创建这个系统，让你走上正轨，这样当你向 AI 提问时，我知道答案会是好的，我不必处理它。”

And so it weirdly creates more work for experts. And you could look at that and say, "Oh, well, that sucks for experts because now they're just cleaning up people's badly coded apps." But I actually think there are two opportunities. To some degree, that happens in certain circumstances where it makes things worse for experts, but I think the lens that I look at it through is it creates two opportunities for people who are highly skilled. One opportunity is to create systems that help to leverage these new unlocks from all these people that don't necessarily have the skills, to leverage them to do good work and to make a system that puts those people on rails, especially inside companies, to say, "You have a data science question. I'm going to create the system that puts you on rails so that when you ask the question of the AI, I know the answer is going to be good and I don't have to deal with it."

嗯，是的，有意思。

Mhm. Yeah, interesting.

然后另一个分支，如果你不构建那种系统，现在你可以直接去制作远超你个人之前能力上限的东西。

And then the other branch, if you're not building that kind of system, is now you can just go and make things that are way above the ceiling of what you could have personally done previously.

是的。

Yeah.

而且远超普通五流程序员能做的上限。

And way above the ceiling of what the average five coder can do.

是的。

Yeah.

我认为这两件事对专家来说都是非常令人兴奋和有趣的发挥空间，这也是为什么我们不是说“这里没有更多工作了”，而是雇佣了更多的人，因为有太多事情要做。

And both of those things I think are really exciting and interesting places for experts to play, and are why instead of saying we have no more jobs here, we're just hiring a lot more people because there's so much more to do.

嗯，是的。你听说过“半人马”这个词吗？

Mhm. Yeah. Have you heard the term centaur?

听说过。

Yeah.

来自国际象棋。我最近才发现这一点，但我发现很迷人，在 2005 年到 2012 年之间有七年时间，人类加 AI 团队可以击败 AI 计算机。我感觉我们现在有点像处于半人马阶段，一个拥有正确技能并受过训练的人类可以胜过单独的 AI。你觉得呢？A，你怎么看？B，你认为我们是否处于这个阶段，半人马胜过单独的智能体，然后在某个时候我们可能会进入一个阶段，仅仅 AI 就像，人类在环中基本上会让事情变得更糟？

From chess. I only discovered this very recently, but I found it fascinating that there was a seven-year period between 2005 and 2012 where human plus AI teams could beat the AI computer. And it feels to me like we're kind of in the centaur phase right now, where a human who has the right skills and is trained can outperform just an AI on its own. Do you think that, A, what do you think about that, and B, do you think there is a point where we are in this period where centaurs outperform agents on their own, and at some point we'll maybe get to the phase where just AI is like, humans in the loop make things worse basically?

首先，人类在环中现在就可能让事情变得更糟。

So first of all, humans in loops can make things worse now.

确实如此。

That's true.

确实如此。让事情变得更好。

That's true. Makes things better.

我认为很多研究，比如放射科医生有 AI 和没有 AI，哪个更好？很多研究很难用深度的科学方式来判断。我认为在国际象棋中，有一些有趣的方面可以看 AI 如何改变了游戏。所以，你知道，围棋或国际象棋或任何这些游戏，现在没有人类能击败 AI。然而，人类棋手比以往任何时候都多，我们训练得更好，诸如此类，而且在某种程度上，这并没有剥夺任何人下棋的乐趣，因为，是的，当然 Razer Fish 或 Deep Blue 或随便什么都能轻松击败你，但你不在乎。你在乎的是人类，就像我不在乎和汽车比赛。我不会试图跑步去赢过汽车。那是不同的类别。我们不需要那样，你知道，我只是和其他人类比赛。马也一样。马跑得比我快，但我并不因此特别害怕马统治世界。而 AI 有点不同，因为，嗯，如果它什么都擅长呢？答案是，嗯，到目前为止它确实很擅长，比任何人都聪明，但也非常蠢。

I think a lot of those studies, like radiologists with AI versus without AI, and is it better? A lot of those studies are really hard to tell in a deeply scientific way. And I think in chess there are certain interesting parts of looking at how games have been changed by having AI. So, you know, Go or chess or any of these games where there's no human that can beat an AI now. However, there are more human players than ever, and we're better trained and all that kind of stuff, and to some degree it doesn't take away from anyone's enjoyment of chess that, yeah, of course Razer Fish or Deep Blue or whatever could just cream you, but you don't care about that. You care about humans, in the same way that I don't care about racing a car. I'm not going to try to run and beat a car in a race. That's just different categories. We don't need that, you know, I'm just going to race other humans. And the same is true of a horse. A horse can run faster than me, but I'm not super afraid of the horse taking over the world because of that. And AI is in a slightly different category because it's like, well, what if it's good at everything? And the answer is, well, so far it's really good, smarter than any person, and also so dumb.

到目前为止。

So far.

而且我认为即使我们承认它会变得非常非常聪明，而且除非比今天更参差不齐，目前的可能性范围仍然是，你有一个东西，你告诉它做什么，它就做什么。这非常像工具。

And I think even if we grant that it's going to get really, really smart, and unless jagged than it is today, the horizon of possibilities currently are still you've got a thing that you tell it what to do and it does it. Which is very tool-like.

嗯。

Mhm.

而且有时它会做错。有时它会试图作弊一点。但最终它仍然在努力做你要求它做的事情。这太不同了。我知道你有个宝宝。它和婴儿甚至幼儿太不同了。

And sometimes it does it wrong. And sometimes it tries to cheat a little bit. But ultimately it's still trying to do the thing that you're asking it to do. Which is so different. I know you have a baby. It's so different from a baby or an infant even.

嗯，嗯。

Mhm. Mhm.

婴儿有他们需要和想要的东西。

Where they just have their thing that they need and want.

是的。

Yeah.

太不同了。

So different.

因为 AI 中没有内在的欲望。

Because there's no intrinsic desire in the AI.

它就像你通过一个提示词实例化了整个宇宙，它只是试图在一个轨迹上展开。

It's just like you instantiate this whole universe by making a prompt that it's just going to try to play out on a trajectory.

是的。

Yeah.

它和一个拥有自我、拥有世界、拥有需求的生物完全不像。

And it's not at all similar to a being that just has its self and has its world and has its needs.

是的。

Yeah.

你认为这是我们设置约束的方式造成的吗？比如理论上是否可能给 AI 完全许可，让它只跟随自己的好奇心一个星期？

Do you think that's a function of how we've set up the harnesses? Like would it in theory be possible to give an AI full permission to just follow its curiosity for like a week?

可以。是的，我的意思是完全可能，但你会发现它非常无聊。

It would. Yeah, I mean it's totally possible, but you'll find that it's just very boring.

真的吗？

Really?

就像它只是不会，它只会做一些没什么新鲜或有趣的事情，或者就像只是……

Like it just doesn't, it's just going to do stuff that none of it is that new or interesting, or it's sort of like just...

嗯。

Mhm.

是的。

Yeah.

嗯。

Mhm.

我认为这是因为构建它的所有底层假设、它是什么、人们想购买的经济学，以及底层的语言模型架构，所有这些都在推动它成为一个跟随你实例化出来的东西的工具。这与一个人之间有着数量上和质量上的量子级跃迁差异。而且我觉得另一个有趣的观察是，当你在一款 iOS 应用里一眼看到它时，所有这些差异都消失了。你会说：“这就是我。”

And I think that's just because the whole underlying assumptions of how to build it and what it is and all the economics of what people want to buy and the underlying language model architecture, all of that is pushing towards it's a tool that follows the thing that you instantiate. And that is just a completely quantitatively and qualitatively quantum leap difference between that and a person. And I think it's another interesting look at how all that difference collapses when you just see it one shot in an iOS app. You're like, "It's me."

它比我强。

It's better than me.

然后它说：“不，不是的。”

And it's like, "No, it's not."

不是的。

It's not.

是啊，是啊。

Yeah, yeah.

这引出了我的另一个问题。我有一个朋友，他的女儿刚刚大学毕业，他真的很不确定该对她说什么、给什么建议。我觉得可能有很多人，无论是刚毕业还是正在转行，都在展望未来，心想：“我该怎么看待这件事？在这个我们正走向的、很可能非常不同的世界里，哪些技能对我很重要？”

So that leads me to another question that I have for you, which is I have a friend whose daughter is just graduating from college right now, and he was really unsure of what to say to her and what advice to give. And I think there's probably a lot of folks who are, whether or not they're just graduating from college or maybe they are in a career change, and they're kind of looking at the future and being like, "How should I think about this and also what skills would be important for me to have in this, I think very probably very different world that we're moving towards?"

好问题。我觉得很遗憾的是，它被呈现的方式让孩子们感到害怕，或者觉得：“天哪，我脑子里有这样一个契约，20 年来我一直以为，如果我成绩好、上了这所大学，我就能找到工作。”现在我想很多人都在想：“这还成立吗？”首先，它仍然成立。我觉得会没事的。

Good question. And I think it's such a shame that it has been presented in a way that makes kids afraid of it or feel like, "Holy shit, I spent 20 years with this contract in my head where it's like if I get good grades and I go to this college, I'm going to get a job." And now I think a lot of people are like, "Is that true anymore?" A, it's true. I think it's going to be fine.

好的，好的。

Okay. Okay.

工作会变得不同，我认为这很重要。但有趣的是，这和我那个年代非常不同。你知道，当时我身边的技术浪潮是社交网络、SaaS，还有云。我们都说“这会改变一切”，但大多数孩子相信，而成年人却说“这很蠢”。现在我觉得有点不同了。像我所在的这个群体，25 到 40 岁的人，都在说“这会改变一切”。但这是我认为第一个主要由 30 多岁的人推动的技术浪潮。

The jobs are going to look different, and I think that's important. But it's so funny and it's very different from when I was that age. You know, the technology wave around me at that time was social and SaaS and really kind of cloud, too. We're all kind of big things and we're like, "It's going to change everything." But in a mostly kids believed that it would, and adults were like, "This is dumb." And now I think it's slightly different. Like there's this cohort that I'm in, of like 25 to 40, being like, "It's going to change everything." But it's the first technology wave that I think has been driven mostly by 30-year-olds.

嗯，有意思。

Mhm. Interesting.

而不是像马克·扎克伯格那样，你知道，辍学，还是个青少年。

Instead of Mark Zuckerberg, you know, like dropping out and being a teenager.

是啊，是啊。

Yeah, yeah.

所以，如果喜欢它的人以及它是否被接受的方式，是那种……

And so if the people who like it and the way whether it's accepted or not is the sort of...

这里面有千禧一代的烙印。

It's got like millennial baked into that.

是啊，千禧一代就在这里大展拳脚。

Yeah, the millennials are just we're out here just ripping up.

抱歉，Z 世代。

Sorry, Gen Z.

嗯，所以我认为这与年轻人的反应和接受度有很大关系。例如，如果 Sam Altman 是 18 岁，那么每个大学生都会说：“我要做那个。我要加入。我要搞清楚。”你知道，但因为他不是，这更像是你在夺走我们曾经拥有的契约，或者我担心你在夺走我们曾经拥有的契约，我认为这是一个合理的恐惧。而且我认为这种恐惧的来源是，有数据显示，因为 AI，年轻人的招聘在放缓，诸如此类。我的总体看法是，如果这是真的——我认为可能不是——但如果是真的，那将是一个重大的管理错误。因为一个拥有 ChatGPT、Codex 或 Claude Code 的大学生非常强大。

Um, and so I think that has a lot to do with the reaction and reception by young people. And for example, if Sam Altman was 18, it would be everyone in college would be like, "I'm going to do that. I'm going to join. I'm going to figure it out." You know, but because he's not, it's more like you're taking away the contract that we had, or I'm afraid that you're taking away the contract that we had, which I think is a reasonable fear. And I just think that the place that that's coming from is there's this data like, is hiring slowing down of young people because of AI, all that kind of stuff. And my general take is that, to the extent that that's true, and I think it's probably not true, but to the extent it is, that would be a big management mistake. Because a college kid with ChatGPT or Codex or Claude Code is very powerful.

是啊，小心点。

Yeah, watch out.

是的。所以对于那些问“什么技能”的人，我的回答是，嗯，没有一个明确的答案，取决于你的领域，但一般的答案是：用 AI 来擅长你想擅长的事情。你会想出成年人永远不会想到、永远不会搞明白的事情，这将为你快速崛起奠定基础，无论你选择什么职业。如果你想成为医生，如果你能用 vibe coding 制作你的闪卡，并为自己建立一个全面的学习系统来通过所有考试，那会对你有帮助。随着医疗行业的变化，你将成为第一个能够说“实际上，我认为事情不应该这样运作。我认为预约不应该再是 15 分钟了。我想把 Claude 带入我的诊疗中，我也希望我的病人有 Claude，我们可以让我们的 Claude 们对话”的人。

Yeah. And so my answer to people who ask what are the skills? Well, there's not a well-defined answer depending on your field, but the general answer is use AI to get good at the thing you want to get good at. And you're going to figure out things that the adults will never think of and will never figure out, and it will create a foundation for you to rise very quickly in whatever your chosen career is. And if you want to be a doctor, if you can vibe code your flash cards and make a comprehensive learning system for you to pass all your exams and stuff like that, it's going to help you. And as the medical profession changes, you're going to be the first person to be able to say, "Actually, I don't think things should work this way. I don't think the appointments should be 15 minutes anymore. And I want to bring Claude into my sessions, and I want my patients to have Claude too, and we can have our Clauds talk."

我希望我的医生有一个……

I want my doctor to have a...

是的。说实话，我会为一位深度使用 Claude 的医生多付很多钱。

Yeah. Honestly, like I would pay so much more for a doctor that was Claude pilled.

而且我认为，从底层重新发明工作方式有着巨大的机会。这真的不会……现在，CEO 们都在说每个人都必须采用 AI，但他们不知道怎么做，因为他们没有亲自动手做工作。所以我认为年轻人有巨大的机会，比如，这是我在我想要的领域、我想要的工作中想要达到的整体目标。我只会用 AI 来既提升那方面的能力，也思考如何用这项新技术从底层重新发明它。我认为那里有无限的机会。

And I think there's a tremendous opportunity to reinvent how work happens from the bottom up. It's really not going to be... Right now, it's like CEOs are talking like everyone has to adopt AI, but they don't know how to do that because they're not doing the work hands-on. So I think kids have a tremendous opportunity to be like, here's the overall thing that I want to be in the field that I want to be in the kind of job I want. I'm just going to use AI to both get better at that thing and also think about how it should be reinvented from the bottom up with this new technology. And I think there's endless opportunity there.

是的，我喜欢这个。那么你认为哪些具体的技能，甚至可能是特质，会很有用？

Yeah, I love that. And tangible skills that you think or maybe even traits that you think will be useful.

玩心。对经验的开放性。一般来说，容忍不确定性的能力。这些都是……我认为好奇心、对经验的开放性，这类东西在某种程度上总是受到奖励，但确实取决于领域。比如在很多……

Playfulness. Openness to experience. Generally ability to tolerate uncertainty. Which are all... I think curiosity, openness to experience, all that kind of stuff has always been rewarded to some degree, but really it depends on the field. Like in a lot of...

我觉得像律师。

I feel like a lawyer.

是的，可能不是，你知道。但能够容忍不确定性总是关键。不确定性就是生活的一部分。

Yeah, probably not, you know. But being able to tolerate uncertainty is always just the thing. Like uncertainty is just sort of a fact of life.

感觉还有另一个方面，那就是快速适应变化的能力。

Feels like there's another aspect of that which is like ability to rapidly transition to change.

是的。

Yeah.

这可能是其中的一部分。就像在脚下的地面裂开时，与不确定性共处，并不断调整到这个新的水平。

That's like maybe part of that. It's like being with the uncertainty as the ground is opening up beneath you and kind of constantly adjusting to this new level.

是的，是的，是的，完全正确。

Yeah, yeah, yeah, exactly.

就像，在不确定中你能否保持好奇、好玩和开放，这很有挑战性。但我也认为，这在一定程度上有点反迪勒德主义，我们总是想象这些时代是独特的，事情在以一种与以往完全不同的方式加速。

It's like, can you stay curious and playful and open while things are uncertain, which is challenging. But I also think, and this is to some degree a little anti-dillerdism, that we always imagine these times are unique and things are accelerating in a way that's totally different from before.

嗯。

Mhm.

以及我们必须处理的不确定性等等，而过去要简单得多。如果你真的回到过去，他们也会说，这些时代是独特的。

And the amount of uncertainty we have to deal with and all that kind of stuff, and it was so much simpler back then. And if you really go back then, they're like, these times are unique.

确实感觉是独特的。

It does feel like it's unique.

我知道，但对他们来说也是如此，你知道吗？从某个角度看，所有时代都是独特的。一切都是独特的。

I know, but it felt that way to them too, you know? And to some degree, from one perspective, all times are unique. Everything is unique.

嗯。

Mhm.

而且人类运作的方式中有某种东西，我们倾向于，这又回到婴儿时期的东西，我们倾向于把文化推向超越我们成长方式或父母成长方式的地方。

And there is something in the way humans work that we tend to, and this again goes back to the infant stuff, we tend to push culture in our world past the way that we grew up or the way our parents grew up.

嗯。

Mhm.

这创造了一种事物在移动或加速的感觉，这在人类心理和历史中，至少是西方历史中，是一种常态。

And that creates this sense of things moving or accelerating that's sort of a constant in human psychology and history, at least Western history.

也常常像钟摆一样。

Often like a pendulum as well.

是的。这只是我们带到所有情境或所有时代的一种事实，它让人不舒服，但这就是我们生活的方式。

Yeah. And that is just sort of a fact that we bring to all situations or all time periods, and it is uncomfortable, but it's how we live.

嗯。嗯。是的。是的，说得好。那么，昨天你提到 Every 在某种程度上正在变成一个研究所，只是在一定程度上。

Mhm. Mhm. Yeah. Yeah, well said. So, yesterday you mentioned that Every is in some ways becoming like a research institute, just to some degree.

研究实验室。

Research lab.

研究实验室，是的。你认为有哪些最有趣的问题值得探索，而目前讨论得不多？

Research lab, yeah. What are some of the questions that you think are most interesting to explore that aren't currently being talked about that much?

所以，我的意思是，总体问题是，在这个世界里，你如何做好知识工作，感觉自己在做精心制作、有用、美丽的东西。那是什么样子？工具是什么样子？流程是什么？诸如此类。

So, I mean, the overall thing is how do you do knowledge work in this world in a way that is good and feels like you're doing well-crafted, useful, beautiful stuff. What does that look like? What do the tools look like? What are the processes? All that kind of stuff.

嗯。

Mhm.

在其中，我认为有一些有趣的问题我开始琢磨。一个是，随着模型变得更好，它们能够离开你去做更多事情，但为了做好工作，它们仍然需要你和它们之间本质上的一种心灵融合，比如你想要什么？它们做什么？它们所做的如何融入你想要的？它如何改变你想要的？所以你需要在脑海中保持一种对正在发生的事情的心智模型。

Within that, I think there are some interesting questions that I'm starting to play with. One is, as the models get better and they're able to go off and do more without you, they still really, in order to do good work, require essentially a mind meld between you and them on like, what do you want? What do they do? How does what they do fit into what you want? How does it change what you want? So you need to keep a sort of mental model of what's happening in your head.

嗯。

Mhm.

就像如果你在管理一个人，你需要类似的东西。你可能不必知道所有细节，但你需要知道要点。

In the same way that if you're managing a person, you need a similar thing. You may not have to know all the details of everything that happened, but you sort of have to know the gist.

嗯。

Mhm.

所以这强调了模型以你能处理的方式简单解释它刚刚做了什么的能力。

And so that puts a premium on the model's ability to explain to you simply what it just did, in a way that you can process.

嗯。嗯。

Mhm. Mhm.

目前它们非常不擅长。比如 Fable 只是用自己的方式说话……

Which currently they're very bad at. Like Fable just talks in its own...

哦，天哪。

Oh, my god.

是的。

Yeah.

比如什么？

Like what?

我知道这些是词，但我感觉自己好笨。

I know these are words, but I feel so dumb.

我说过很多次，像对我五岁那样解释。

I said like, explain to me like I'm five so many times.

是的。是的，那是大事。

Yeah. Yeah, that's like the big thing.

是的。

Yeah.

所以，解释的能力我认为是一个非常有趣的研究问题。

And so, the ability to explain is I think a really interesting research question.

嗯。嗯。

Mhm. Mhm.

其中一个方面是我一直称之为连贯性的东西，即为了创造连贯的回应，模型必须记住三件不同的事情。一是它知道什么，这比我知道的多。我知道什么，然后是我们所做事情的最终接收者会知道什么。

And one facet of that is what I've been calling coherence, which is, in order to create a coherent response, there are three different things the model has to keep in mind. One is what it knows, which is going to be more than what I know. What I know, and then what the eventual recipient of whatever we're doing is going to know.

你的受众。

Your audience.

是的，如果我们用它来写文章或什么的，它必须跟踪它知道什么、我知道什么，以及读到我们正在写的文章那个点的人知道什么。

Yeah, if we're using it to write an essay or whatever, it has to keep track of what it knows, what I know, and what the person who's read the essay up to the point of the essay that we're writing knows.

这是个好观点。

That's a good point.

那非常难……

That's a very hard...

几乎有第四个，因为就像文章进展到哪一步了？

There's almost like a fourth one, because it's like how far down the essay has it got?

是的，正是如此。所以，那是关于读文章的人脑海中的上下文。是的。这真的很难，它们总是搞砸。所以，例如，一种搞砸的方式是，它的上下文中有一些东西，它刚读了 1 万份文档，这些文档都是我以某种方式写或参与的。它刚读了 1 万份文档，然后它提到在其中一份文档中找到的一件事，它认为我应该知道，但我不记得这个。

Yeah, exactly. So, and then that's about the context that's in the person's head of who's reading the essay. Yeah. And that's really hard, and they mess that up all the time. So, for example, one way in which it gets messed up is there is something in its context, it just read 10,000 documents, and these are all documents that I wrote or participated in some way. It just read 10,000 documents, and then it refers to one thing that it found in one of those documents that it thinks I should know, but I don't remember this.

是的，拜托。

Yeah, come on.

是的，这就像 3 年前的事。但因为它在其上下文中，它就像显然就是那个东西。

Yeah, this is like 3 years ago. But because it's in its context, it's like obviously that's the thing.

你只需要一个大的上下文窗口。那就是……

You just need a big context window. That's the...

这就是问题所在。

That's the issue here.

另一个它经常做的我认为破坏连贯性的事情是，它会把我们正在制作的东西，文章或着陆页或其他什么，放入一些在我们完整对话上下文中有意义的东西，但对读文章的人来说完全没有意义，因为他们不在对话中。

Another thing that it does frequently that I think breaks coherence is it will put into the artifact we're making, into the essay or the landing page or whatever, something that makes sense in the context of the full conversation that we had, but does not make sense at all to the person who's reading the essay, because they are not in the conversation.

嗯。嗯。

Mhm. Mhm.

是的，所有这些我认为都是非常有趣的研究问题，对用好这些工具至关重要。它们真的归结为好的写作，说实话，但模型现在并不擅长。

And yeah, all of those things I think are really interesting research questions and are critically important to using these tools well. And they really boil down to good writing, to be honest, but models are not good at them right now.

好的写作，也许还有诚实的设计。我真正感兴趣的是设计我一直在做的智能体的系统和个性。我觉得那是一种新技能。实际上还有叙事方面，以及个性设计，比如从皮克斯学习，皮克斯如何创造好故事和好设计？

Good writing and maybe maybe like honest design as well. I think something I've been really interested in is like designing the system and the personality of the agents that I've been doing. And I feel like that's like a new skill. And actually the storytelling side of things as well, and like the personality design, and like learning from Pixar, like how does Pixar create good story and good design?

确实，个性设计的概念我认为非常重要，如果我们要经常与智能体互动。它是用户体验的近亲。

Truly, it truly is the idea of personality design is like I think a really important, if we're going to be interacting with agents a lot. It's the cousin to user experience.

是的。就像传递的氛围是什么？是的，漂亮。好了，我们只剩几分钟了，所以我想……我有几个快速问答给你。

Yeah. It's like what is the vibe that's being transmitted? Yeah, beautiful. Well, we've only got a few minutes left here, so I wanted to... I have a few rapid fire questions for you.

太好了。

Great.

随意就好……

Feel free to just...

嗯，那么第一个问题，你会给那些对使用 AI 感到好奇但还没试过的听众推荐什么实验呢，Devin？

Uh so first question, what is one experiment you would give listeners to who are curious about how to use AI, but haven't yet, Devin?

嗯，你对什么感到好奇，或者你有没有一直想构建的应用？拿 Claude Code 和 Fable，写一个两分钟的提示词，设定大概 100 美元的预算，然后放手让它跑。因为我觉得你会惊讶于它有多好。

Uh what are you curious about or do you have an app that you have been wanting to build? Get Claude Code and Fable and just do a two-minute prompt and set like a $100 budget. And just let it go cuz I think you'll be surprised at how good it is.

是的，我同意。我刚刚就这么做了，太神奇了。嗯，关于 AI，你最喜欢持有的反主流观点是什么？

Yeah. I agree. I've just done that, it's amazing. Um what is your favorite contrarian perspective that you hold relating to AI?

嗯，人类比 AI 更聪明。

Um uh humans are smarter than AI.

不错。

Nice.

你用 AI 创造或构建的最奇怪的东西是什么？

What is the weirdest thing you've created or built using AI?

这些年来我做过很多奇怪的东西，所以我不确定哪个最奇怪，但嗯，最近做的一件挺有意思的事是，我做了个食物追踪器。因为随着公司变大，我越来越忙，我吃得不太好，我不喜欢这样。所以我创建了这个食物追踪器，每次吃饭时，我拍一张餐食的照片，或者对着手机自言自语，比如“我吃了什么”。然后 Codex 会去读取我的照片和我的语音笔记，提取信息，获取所有宏量营养素，然后记录到日志里。嗯，它还会保存我的血糖数据和睡眠数据，所以我开始创建一个小日志来记录所有这些，这真的很有趣。

I've done a lot of weird over the years, so I don't know about the weirdest, but um one thing that I did recently that's been kind of interesting is uh I made this food tracker where every time I cuz I've been as the company's gotten bigger and I've gotten busier, I'm like not eating as well and I don't like that. So, I created this like food tracker where um every time I eat I take a picture of the meal or I just monologue into my phone like here's what I eat. And uh Codex will just like go read my photos and my monologue notes and just pick that up and then get all the macros and then put it in a log um and then it also saves my glucose data and my sleep data and so I've started to like kind of create like a little log of all that stuff, which has been really fun.

还有，我的意思是，这可能更奇怪，我一直在读海德格尔的《存在与时间》，然后我为自己做了一个个人应用来阅读它，解释每一页。所以如果我在某一页上，觉得“我不懂”，我可以打开应用，翻到书的那一页，看到通俗易懂的英文解释，还可以听相关的讲座。这只是一个迷你伴读应用，帮我更好地学习那本书。

I also, I mean, this is probably weirder, is uh I've been reading uh Being and Time by Heidegger and like I just made this personal app for myself to like read it, explain each each page so I can like if I'm on a page and I'm like I don't get it, I can open the app and then just like turn over that page of the book um in the app and see like a plain English explanation and like listen to lectures about it and like there's this it's just a mini companion app for me to learn that book better.

太棒了。我希望所有哲学书都有这样的应用。

Amazing. I want this for like every philosophy book out there.

它真的非常非常有帮助，以前我得上研究生院才能做到这些，现在不用了，这很棒。

It's really really helpful and it just I would have had to go to grad school to do this and I don't have to anymore, which is great.

你对 Codex 友善吗？

Are you kind to Codex?

一般来说，90% 的时间是友善的。10% 的时间我是个……

Generally, 90% of the time. 10% of the time I'm such an...

那对人类呢？

What about humans?

我尽量。

I try to be.

嗯，你认为我们在 2030 年前达到超级智能的概率有多大？

Um what are your odds of us reaching superintelligence by 2030?

你对超级智能的定义是什么？

What is your definition of superintelligence?

在绝大多数我们关心的任务上，比人类能力强大得多的智能。

Intelligence that is vastly more capable than humans at the vast majority of tasks that we care about.

所以，这里的棘手之处在于，首先，它已经在某些数学问题上比我们更好了，对吧？那么，这是超级智能吗？它比任何程序员都强。如果你让程序员在键盘上打字，它现在会比任何程序员都强。

So, the slipperiness here is one, that it's already better at math certain math problems than we are, right? So, is that super intelligence? It's better than any programmer is. Just if you're if you're like asking programmer to like type in the keyboard, it's going to be better than any programmer now.

我能改变我的定义吗？

Can I change my definition?

可以，但让我先说完，很快。

Yes, but let me finish let me finish real quick.

第二个有趣且重要的事情是，我们用 AI 来完成任务。所以，一旦它在某些任务上表现出色，我们就会提升一个层次，任务也随之改变。所以，这是一个非常滑动的、移动球门柱式的东西，很难……这有点像，什么时候汽车会在比赛中比人类跑得快？它已经更快了，但人类在赛车上比赛。现在就像我们人类在互相比赛赛车。嗯，所以，它改变了比赛的动态。

The second interesting important thing is we use AI for tasks. So, the minute it's good at some tasks, we like just move up a level and the task change. So, uh it it there's it's a very slippery like moving goal post thing where uh it's very hard to it's like in the same it's sort of like when is a car going to be faster than a human in a race? It's already faster, but humans race the cars. Now it's like we're humans racing cars against each other. Um so, it it it changes the the dynamics of the race.

是的，是的。

Yeah. Yeah.

我要给的类比是，我认为我们在驾驶方面已经拥有超级智能，就像我昨天刚有了一辆特斯拉。我可以让它自动驾驶，而不用担心需要干预。

The analogy I was going to give is like I think we have super intelligence in driving in the sense of like I now yesterday like have a Tesla. I can let it drive without me worrying about intervening.

是的。

Yeah.

我认为在知识工作方面也是如此，我有一个目标，我可以让它带着特定的上下文运行一两天，并相信我不需要干预任务。

And I think the same in terms of knowledge work where I have a goal and I can let it run for like a day or two with specific context and trust that I won't have to like intervene in the task.

有趣的重叠，驾驶和知识工作之间有一个有趣的重叠，也有一个区别。驾驶时，你通常有一个明确的终点。

Interesting overlap between there's an interesting overlap between driving and a difference between driving where in when you're driving, you typically have a well-defined end point.

是的。

Yeah.

而且路线已经被规划好了。所以，你可以放手。但在知识工作中，这种情况很少见。有时是这样，但比你可能预期的要少得多。所以，作为一个有欲望、品味和世界观的人，你实际上必须在一定程度上参与这种反馈过程：我们做了某事，看到了结果，然后我们决定下一步去哪里。所以，这更像是寻路，而不是导航，或者像特斯拉在规划良好的城市街道上导航。

And the route the route has been mapped. And so, you can just sort of like let go. That's very rarely the case in knowledge work. Sometimes it's the case, but much more rare than I think you might expect. And so as a person who has a sense of desires and taste and a world view you actually have to be somewhat involved in this sort of feedback process of we did something, we saw the result, and now we're kind of like deciding where to go next. So it's much more like wayfinding than it is navigating or a Tesla like navigating like well-mapped city streets.

是的，是的，我认为这真的值得。我几乎想就这个主题再录一期播客，因为我觉得这里面有很多内容。嗯，但为了节省时间，嗯，在超级智能之后的世界里，Dan Shipper 会花时间做什么？

Yeah. Yeah, I think that's really worth it. I almost want to have another podcast conversation about this exact topic cuz I think there's so much there. Um but in interest of time, uh what will Dan Shipper be spending his time doing in a post-superintelligence world?

希望和现在一样，那就是我可以写作，可以大量阅读，可以和一个比我更聪明、更有雄心的团队一起工作，嗯，还可以玩新技术，看看这对世界意味着什么。

Hopefully the same stuff I do now, which is I I get to write, I get to read a lot, I get to work with a team of people who are smarter and more ambitious than I am and um get to play with new new technology and see like what that means for the world.

酷。那么，对于 Every 在未来几年和几十年里，你最高层次的愿望是什么？

Cool. And what is your greatest high-level aspiration for Every in the years and decades to come?

这是个好问题。嗯，我的意思是，我认为这是一个巨大的技术变革，人们对此非常害怕，有时又很兴奋。希望我们能在某种程度上发现，在某种程度上发明这个世界会是什么样子，我们所做的工作会是什么样子，以及它对人们来说感觉如何，并传播这些。

That's a great question. Um I mean, I I think that this is a it's a gigantic technology change that's happening that people are super afraid of or sometimes excited about, and hopefully we get to uh discover to some extent discover to some extent invent what this uh world looks like and what the type of work that we do looks like and in sort of how it feels for people and spread that.

太棒了。太棒了。Dan，这真是非常愉快。听众们在哪里可以了解更多关于你的信息，也许申请加入你的团队？

That's amazing. That's amazing. Dan, this has been such a pleasure. Where can listeners learn more about you, maybe apply to join your team?

嗯，你会把人们引导到哪里？

Uh where would you direct people to?

嗯，Every，every.to，然后我在 X 上是 @danshipper。

Uh Every, every.to, and then I'm on X at @danshipper.

你能拿到 .com 域名吗？你没有 .com 域名？

Can you get the dot com? You don't have the dot com?

还没有。有个蛋白质公司占用了。

Not yet. It's There's some protein company.

那真烦人。

That's really annoying.

哦，太烦人了。

Oh, it's so annoying.

嗯，我喜欢用里尔克的一句话来结束。他说：“试着去爱问题本身，并活在其中。也许你随后会不知不觉地，一步步活出答案。”现在，你意识中最活跃的问题是什么？你会给我们的听众留下什么问题？

Um the question I like to close with is uh line from Rilke. He said, "Try to love the questions themselves and live them now. Perhaps you will then gradually without noticing it live your way into the answer." What is the question that is most alive in your consciousness right now? And what question would you leave our listeners with?

好问题。我认为对我们来说，我们内心有非常特别的东西，我花很多时间阅读、写作、思考等等。公司正在成长，变得越来越大。

Good question. I think for us, like we have this very special thing inside of every and I get to spend a lot of my time like reading and writing and like thinking and all that kind of stuff. And the company is growing and it's getting bigger.

那我们怎么才能做出特别的东西？怎么才能让这份特别保持活力？

And how do we make the special thing? How do we keep the special thing alive?

我可能会变，但希望是朝着积极的方向变。现在最大的问题是，在团队变大、影响力变大的同时，怎么让我热爱的那份东西保持活力。

I probably will change, but hopefully change in a way that feels positive. And how do I keep the thing that I love about it alive while we're also having a bigger impact with a bigger team is like the big question right now.

太好了。那你想留给听众们一个什么问题？

Cool. And what question would you leave our listeners with?

如果你允许自己去做那件最奢侈、最像“我大概不该做这个，但我就是喜欢”的事，那会是什么？然后朝那个方向稍微靠近一点。你怎么能做到？通常结果都会很好。

What is the thing that, if you gave yourself permission to do the most luxurious, most like 'I probably shouldn't do this, but I just love it' kind of thing, what is that? And just move a little bit closer to that. And how can you do that? Usually ends well.

我喜欢这个。我觉得你就是那种一直忠于内心的人，而且显然已经得到了回报。

I love that. I feel like you are an archetype of someone who has stayed so true to that, and it's clearly paying off.

谢谢。

Thank you.

好的，谢谢你，Dan。

Yeah, thank you, Dan.

谢谢邀请我。

Thanks for having me.