On Artificial Intelligence
频道: Naval
视频: https://www.youtube.com/watch?v=sXCKgEl9hBo
原文语言: en
统计: 共 36 轮 · Nivi 8 · Naval 28
[0:00] Nivi
Hey, this is Nivei. You're listening to the Naval podcast for the first time in recorded history. We are not at the same location. I am actually walking around town and Naval might be doing the same. So, there might be some ambient noise, but we are going to try hard to remove that with AI and some good audio engineering. Podcast recording is so stilted because it's like you have to sit down and you schedule something and this giant mic pointing in your face and it's not casual. It makes it just less authentic, more practiced, more rehearsed. I get that it produces maybe higher quality audio and video, but I feel like it produces lower quality conversation.
嘿，我是 Nivi。你正在收听的是 Naval 播客。这在录制历史上还是头一回——我们俩不在同一个地方。我其实正在镇上散步，Naval 可能也在做同样的事。所以可能会有一些环境噪音，但我们会努力用 AI 加上一些不错的音频工程把它去掉。播客录制太拘谨了，因为你得坐下来、得提前安排好时间，还有一支巨大的麦克风对着你的脸，一点都不随意。这让对话变得不那么真实，更像是排练过、演练过的。我明白这样也许能产出更高质量的音频和视频，但我感觉它产出的是更低质量的对话。
[0:43] Nivi
And we all know brains run better when they're being locomoted and you're moving around or just going for walks.
而且我们都知道，大脑在被带着活动、在你四处走动、或者只是去散散步的时候，运转得更好。
[0:50] Naval
Absolutely. My brain is powered by my legs.
完全同意。我的大脑是靠我的双腿供能的。
[0:52] Nivi
I pulled out some tweets from Naval on the topic of AI. We want to talk a little bit about AI and hopefully talk about it in a more timeless manner, but I think some of it's going to be non-timeless content. Before we jump into the tweets, do you want to say anything about what you're doing with your time or what you're doing at Impossible?
我挑出了一些 Naval 关于 AI 这个话题的推文。我们想稍微聊聊 AI，希望能用一种更经得起时间考验的方式来聊，但我想其中有一部分会是不那么永恒的内容。在我们深入这些推文之前，你想不想说说你现在的时间都花在做什么、或者你在 Impossible 在做些什么？
[1:14] Naval
Not really. We're working on a very difficult project. That's why it's called impossible with an amazing team and it's really exciting building something again. It's very pure starting over from the bottom and it's always day one. I guess I just wasn't satisfied being an investor and I certainly don't want to be a philosopher or just a media personality or a commentator because I think people who just talk too much and don't do anything. They haven't encountered reality. They haven't gotten feedback the harsh feedback from free markets or from physics or nature. And so after a while it ends up becoming just too much archer philosophy. You'll probably have noticed my recent tweets have been much more practical and pragmatic. Although there's still occasional ethereal or generic ones, but it's more grounded in the reality of working every day. And I just like working with a great team to create something that I want to see exist. So hopefully we'll create something that will come to fruition and people will say, "Wow, that's great. I want that also." Or maybe not. But it's in the doing that you learn.
也没什么特别想说的。我们在做一个非常困难的项目，所以它才叫 Impossible，团队特别棒，能再次从头打造一个东西真的让人很兴奋。从最底层重新开始是一件非常纯粹的事，而且永远都是第一天。我想我只是不满足于只做一个投资人，我也绝对不想做一个哲学家，或者只是个媒体名人、评论员，因为我觉得那些光说不做的人，他们没有真正撞上现实，他们没有得到反馈——来自自由市场、来自物理规律或自然界的那种残酷反馈。所以时间一长，那就变成了太多的扶手椅哲学。你可能已经注意到我最近的推文变得务实和实用得多了，虽然偶尔还是会有飘渺的或者泛泛而谈的，但更多是扎根于每天工作的现实之中。我就是喜欢和一支优秀的团队一起，去创造一个我希望它存在的东西。所以希望我们能创造出一个最终落地的东西，人们会说"哇，这太棒了，我也想要这个"。又或者不会，但正是在动手做的过程中你才会学到东西。
[2:13] Naval
So I pulled out a tweet from a couple days ago, February 3rd. Vibe coding is the new product management. Training and tuning models is the new coding. There's been a shift a market pronouncement in the last year and especially in the last few months most pronounced by claude code which is a specific model that has a coding engine in it which is so good that I think now you have vine coders which are people who didn't really code much or hadn't coded in a long time who are using essentially English as a programming language as an input into this codebot which can do endto-end coding instead of just helping you debug things at the You can describe an application that you want. You can have it lay out a plan. You can have it interview you for the plan. You can give it feedback along the way. And then it'll chunk it up and it'll build all the scaffolding. It'll download all the libraries and all the connectors and all the hooks. And it'll start building your app and building test harnesses and testing it. And you can keep giving it feedback and debugging it by voice saying this doesn't work, that works, change this, change that, and have it build you an entire working application without your having written a single line of code. For a large group of people who either don't code anymore or never did, this is mind-blowing. This is taking them from idea space and opinion space and from taste directly into product. So vibe coding is a new product management instead of trying to manage a product or a bunch of engineers by telling them what to do. You're not telling computer what to do and the computer is tireless.
我挑了一条几天前、2 月 3 日发的推文：氛围编程（vibe coding）是新的产品管理，训练和调优模型是新的编程。在过去一年里、尤其是过去几个月里，市场上出现了一个明显的转变——这个转变最突出的体现是 Claude Code，它是一个内置了编程引擎的特定模型，好到我觉得现在出现了所谓的"氛围编程者"，就是那些本来不怎么写代码、或者很久没写代码的人，他们本质上是把英语当作一种编程语言，作为输入喂给这个能做端到端编程的代码机器人，而不只是帮你调试代码而已。你可以描述一个你想要的应用，可以让它列出一个计划，可以让它就这个计划来访谈你，可以在过程中给它反馈。然后它会把任务拆成块，搭好所有的脚手架，下载所有的库、所有的连接器、所有的钩子，然后开始构建你的应用、构建测试框架并测试它。你可以不断给它反馈、用语音帮它调试，说这个不行、那个可以、改这个、改那个，让它给你构建出一个完整可运行的应用，而你一行代码都没写过。对于一大群要么不再写代码、要么从来没写过代码的人来说，这是颠覆性的。这把他们从想法的空间、观点的空间、从品味，直接带到了产品。所以氛围编程是新的产品管理——你不再是通过告诉一群工程师该做什么来管理一个产品或一群工程师，而是在告诉计算机该做什么，而计算机是不知疲倦的。
[3:45] Naval
The computer is egoless and it'll just keep working. It'll take feedback without getting offended. You can spin up multiple instances. is it'll work 24/7 and you can have it produce working output. What does that mean? Just like now anybody can make a video, anyone can make a podcast, anyone can now make an application. So we should expect to see a tsunami of applications. Not that we don't have one already in the app store, but it doesn't even begin to compare to what we're going to see. However, when you start drowning in these applications, does that necessarily mean that these are all going to get used? No. I think it's going to break into two kinds of things. First, the best application for a given use case still tends to win the entire category. When you have such a multiplicity of content, whether in videos or audio or music or applications, there's no demand for average. Nobody wants the average thing. People want the best thing that does the job. So, first of all, you just have more shots on goal. So, there will be more of the best. There will be a lot more niches getting filled. You might have worn an application for a very specific thing like tracking lunar phases in a certain context or a certain kind of personality test or a very specific kind of video game that made you nostalgic for something. Before the market just wasn't large enough to justify the cost of an engineer coding away for a year or two, but now the best vibe coding app might be enough to scratch that itch or fill that slot. So a lot more niches will get filled and as that happens the tide will rise. the best applications, those engineers themselves are going to be much more leveraged. They'll be able to add more features, fix more bugs, smooth out more the edges. So, the best applications will continue to get better. A lot more niches will get filled. And even individual niches such as you want an app that's just for your own very specific health tracking needs or for your own very specific architecture, layout or design, that app that could have never existed will now exist. We should expect just like on the internet what's happened with Amazon where you replaced a bunch of bookstores with one super bookstore and a zillion longtail sellers or YouTube replaced a bunch of medium-sized TV stations and broadcast networks with one giant aggregator called YouTube or maybe a second one called Netflix and then a whole long tale of content producers. So the same way the app store model will become even more extreme where you will have one or two giant app stores helping you filter through all of the AI slop apps out there and then at the very head there'll be a few huge apps that will become even bigger because now they can address a lot more use cases or just be a lot more polished and then there'll be a long tale of tiny little apps filling every niche imaginable. As the internet reminds us, the real power and wealth, super wealth goes to the aggregator. But there's also a huge distribution of resources into the longtail. It's the mediumsized firms that get blown apart.
计算机是没有自我的，它会一直工作下去。它接受反馈不会被冒犯。你可以同时开多个实例，它会 7×24 小时工作，你可以让它产出可运行的成果。这意味着什么？就像现在任何人都能做视频、任何人都能做播客一样，现在任何人都能做应用了。所以我们应该预期会看到应用的海啸。不是说我们现在的应用商店里还没有这种东西，而是它跟我们接下来会看到的相比，根本不在一个量级。然而，当你开始被这些应用淹没时，这是否就意味着这些应用都会被用起来呢？不会。我觉得它会分裂成两类东西。首先，对于一个给定的使用场景，最好的那个应用仍然往往会赢下整个品类。当你有了如此繁多的内容——无论是视频、音频、音乐还是应用——平庸是没有需求的。没有人想要平庸的东西，人们想要能把事情做到最好的那个。所以首先，你只是有了更多射门的机会，因此会出现更多"最好的"，会有更多的细分市场被填满。你可能会有一个专门用来做某件非常具体的事的应用，比如在某个特定情境下追踪月相，或者某种特定的人格测试，或者某种让你怀念某样东西的非常特定的电子游戏。以前市场规模就是不够大，不足以证明让一个工程师埋头编程一两年的成本是值得的，但现在最好的氛围编程应用就可能足以挠到那个痒处、填上那个空位。所以会有多得多的细分市场被填满，随着这种情况发生，水位会上涨。最好的那些应用，它们的工程师本身会拥有大得多的杠杆，他们能加更多功能、修更多 bug、打磨更多边边角角。所以最好的应用会持续变得更好，多得多的细分市场会被填满。甚至像"你想要一个只为你自己非常具体的健康追踪需求、或者为你自己非常具体的建筑布局或设计而做的应用"这样的单个细分需求，那个本来永远不可能存在的应用，现在也会存在了。我们应该预期，就像互联网上发生过的那样——亚马逊的情况就是你用一个超级书店加上无数的长尾卖家，取代了一大堆书店；或者 YouTube 用一个叫 YouTube 的巨型聚合器、也许加上第二个叫 Netflix 的，取代了一大堆中等规模的电视台和广播网络，然后再加上一长尾的内容生产者。同样地，应用商店模式会变得更加极端：你会有一两个巨型应用商店帮你过滤掉外面所有的 AI 垃圾应用，然后在最顶端会有几个巨大的应用，它们会变得更大，因为现在它们能覆盖多得多的使用场景、或者只是打磨得精致得多，然后还会有一长尾的微型小应用，填满每一个能想象到的细分。正如互联网提醒我们的，真正的力量和财富、超级财富，归于聚合者。但同时也有大量资源被分配到长尾里。被炸得粉碎的是那些中等规模的公司。
[6:37] Naval
The 5, 10, 20 person software companies that were filling a niche for an enterprise use case that can now be either vibe coded away or the lead app in the space can now encompass that use case. So if anyone can code, then what is coding? Coding still exists in a couple of areas. The most obvious place that coding exists is in training these models themselves. There are many different kinds of models. There are new ones coming out every day. There different ones for different domains. We're going to see different models for biology, for programming. We're going to see pointed focus models for sensors. We're going to see models for CAD, for design. We're going to see models for 3D and graphics and games. Models for video. You see many different kinds of models. The people who are creating these models are essentially programming them. But they're programmed in a very different way than classic computers. Classic computing is you have to specify in great detail every step, every action the computer is going to take. You have to formally reason about every piece and write it in a highly structured language that allows you to express yourself extremely precisely. The computer can only do what you tell it to do. And then once you've got this very structured program, you run data through it and the computer runs the data and gives you an output. It's basically an incredibly fancy, very complicated, meticulously programmed calculator. Now, when it comes to AI, you're doing something very different, but you are nevertheless programming it. What you're doing is you're taking giant data sets that have been produced by humanity thanks to the internet or aggregated in other ways, and you're pouring those data sets into a structure that you've defined and tuned. And that structure tries to find a program that can produce more of that data set or manipulate that data set or create things off that data set. So you're searching for a program inside this construct that you've designed.
那些 5 人、10 人、20 人的软件公司，它们本来在为某个企业级使用场景填补一个细分空位，而现在这个空位要么可以被氛围编程取代掉，要么这个领域里的头部应用现在可以把这个使用场景囊括进去。那么如果人人都能编程，编程又是什么呢？编程在几个领域里仍然存在。编程最明显存在的地方，就是训练这些模型本身。有许许多多不同种类的模型，每天都有新的出现，针对不同领域有不同的模型。我们会看到针对生物学的模型、针对编程的模型，会看到针对传感器的有针对性的专注模型，会看到针对 CAD、针对设计的模型，会看到针对 3D、图形和游戏的模型，针对视频的模型。你会看到许许多多不同种类的模型。创造这些模型的人，本质上就是在给它们编程，但它们被编程的方式跟经典计算机非常不同。经典计算是你必须极其详细地指定计算机要走的每一步、要做的每一个动作。你必须对每一个部分做形式化的推理，并用一种高度结构化的语言把它写出来，这种语言让你能极其精确地表达自己。计算机只能做你叫它做的事。然后一旦你写好了这个非常结构化的程序，你就把数据灌进去，计算机处理数据并给你一个输出。它基本上就是一个极其花哨、极其复杂、被精心编程过的计算器。而到了 AI 这里，你做的是一件非常不同的事，但你仍然是在给它编程。你做的是，把人类借助互联网产生的、或者以其他方式聚合起来的巨型数据集，倒进一个你已经定义并调优好的结构里。这个结构会试图找到一个程序，这个程序能产出更多这样的数据、或操纵这个数据集、或基于这个数据集创造东西。所以你是在你设计的这个构造里搜索一个程序。
[8:29] Naval
You've set up a model. You've tuned the number of parameters. You've tuned the learning rate. You've tuned the batch size. You've tokenized the data that's coming. You've broken into pieces and you're pouring it inside the system you've designed almost like a giant pachinko machine. And now the system is trying to find a program and could find many different programs. So your tuning really influences how good the program that you found is. And that program can now suddenly be expressive in different kinds of domains. So it can do things that computers before were traditionally very bad at. Traditional computers are very good when you program them to give you precise output, specific answers to specific questions, things you can rely on and repeat over and over again. But sometimes you're operating in the real world and you're okay with fuzzy answers. You're even okay with wrong answers. For example, in creative writing, what's the wrong answer? If you're writing a piece of poetry or fiction, what's the wrong answer? If you're searching on the web, there are many right answers. There are many details of the right answers, but they're not all quite perfectly right. And real life sort of works that way. There are variations of right answers or mostly right answers. When you're drawing a picture of a cat, there are many different cats you could draw. There may different levels of detail.
你搭好了一个模型，你调好了参数的数量，你调好了学习率，你调好了批大小，你把进来的数据做了 token 化，你把它拆成一块块，然后把它倒进你设计的这个系统里——几乎就像一台巨大的弹珠机。现在这个系统在试图找到一个程序，而且它可能找到许多不同的程序。所以你的调优真的会影响你找到的那个程序有多好。而那个程序现在突然就能在不同种类的领域里有表现力。所以它能做以前的计算机传统上非常不擅长的事。传统计算机非常擅长的是：当你给它编程让它给你精确的输出、对特定问题给出特定答案、给出那些你能依赖、能反复重复的东西。但有时候你是在真实世界里运作，你可以接受模糊的答案，你甚至可以接受错误的答案。比如说，在创意写作里，什么是错误答案？如果你在写一首诗或一篇小说，什么是错误答案？如果你在网上搜索，有很多个正确答案，正确答案有很多细节，但它们并不都完美正确。而真实生活某种程度上就是这么运作的——存在着正确答案的各种变体、或者大体正确的答案。当你画一只猫的图时，你可以画出很多只不同的猫，可以有不同的细节程度。
[9:40] Naval
There many different styles you could use. When these semi-rong or fuzzy answers are acceptable, then these discovered programs through AI are much more interesting and much more adapted to the problem than ones that you coded up from scratch where you had to be super precise. Fundamentally, what we're doing is a new kind of programming, but this is the forefront of programming. This is now the art of programming. These people are the new programmers. And that's why you can see AI researchers are getting paid gargantuan amounts because they've essentially taken over programming. Does this mean that traditional software engineering is dead? Absolutely not. Software engineers, even the ones who are not necessarily tuning or training AI models, these are now among the most leveraged people on Earth. Sure, the guys who are training and tuning models are even more leveraged because they're building the tool set that software engineers are using. But software engineers still have two massive advantages on you. First, they think in code. So they actually know what's going on underneath and all abstractions are leaky. So when you have a computer programming for you, when you have clawed code or equivalent programming for you, it's going to make mistakes. It's going to have bugs. It's going to have suboptimal architecture. So it's not going to be quite right. And someone who understands what's going on underneath will be able to plug the leaks as they occur. So if you want to build a well architected application, if you want to be able to even specify a well architected application, if you want to be able to make it run at high performance, if you want it to do its best, if you want to catch the bugs early, then you're going to want to have a software engineering background. The traditional software engineer is going to be able to use these tools much better. And there are still many kinds of problems in software engineering that are out of scope for these AI programs today. The easiest way to think about those is problems that are outside of their data distribution. For example, if they need to do like a binary sort or reverse a link list, they've seen countless examples of that. So, they're extremely good at it. But when you start getting out of their domain, when you have to write very high performance code, when you're running on architectures that are novel or brand new, when you're actually creating new things or solving new problems, then you still need to get in there and handcode it. At least until either there are so many of those examples that new models can be trained on them or until these models can sufficiently reason at even higher levels of abstraction and crack it on their own because given enough data points there is some evidence that these AIs actually learn. They learn to a higher level of abstraction because the act of forcing them to compress the data forces them to learn higher level representations. If I show an AI five circles, it can just memorize exactly what the sizes and the radi and the thicknesses and so on of those circles are. If I show it 50,000 circles or 5 billion circles and I give it a very small amount of parameter weights, which are its equivalent neurons to memorize that, it's going to be much better off figuring out pi and how to draw a circle and what thickness means and forming an algorithmic representation of that circle rather than memorizing circles.
你可以用许多不同的风格。当这些半错的、或者说模糊的答案是可以接受的时候，那么这些通过 AI 发现出来的程序，就比你从零开始、必须超级精确地写出来的程序有意思得多、也更适配这个问题。本质上，我们在做的是一种新的编程，但这是编程的最前沿，这现在是编程的艺术，这些人就是新的程序员。这也是为什么你能看到 AI 研究员拿到天文数字般的报酬——因为他们本质上接管了编程。这是否意味着传统的软件工程已经死了？绝对不是。软件工程师，哪怕是那些不一定在调优或训练 AI 模型的人，现在也是地球上杠杆最大的一批人之一。当然，那些在训练和调优模型的人杠杆更大，因为他们在打造软件工程师所使用的那套工具集。但软件工程师相对于你还有两个巨大的优势。第一，他们用代码来思考，所以他们真的知道底层在发生什么——而所有的抽象都是有漏洞的。所以当你让计算机替你编程，当你让 Claude Code 或类似的东西替你编程时，它会犯错，会有 bug，会有次优的架构，所以它不会完全正确。而一个理解底层在发生什么的人，能够在漏洞出现时把它们堵上。所以如果你想构建一个架构良好的应用，如果你甚至想能够去描述一个架构良好的应用，如果你想让它以高性能运行，如果你想让它发挥到最好，如果你想及早抓住 bug，那么你就会想要有软件工程的背景。传统的软件工程师能把这些工具用得好得多。而且软件工程里还有许多种类的问题，是今天这些 AI 程序的能力范围之外的。想这些问题最简单的方式，就是那些落在它们数据分布之外的问题。比如说，如果它们需要做个二分排序、或者反转一个链表，它们见过数不清的这类例子，所以它们极其擅长。但当你开始走出它们的领域，当你必须写非常高性能的代码，当你在新颖的、全新的架构上运行，当你真的在创造新东西、或解决新问题时，那你还是得亲自钻进去手写代码。至少在以下两种情况发生之前是这样：要么这类例子多到足以用来训练新的模型，要么这些模型能在更高的抽象层次上充分推理、自己把它攻克下来——因为只要给足够多的数据点，有一些证据表明这些 AI 确实在学习。它们学到了更高层次的抽象，因为强迫它们去压缩数据这个动作，本身就迫使它们去学习更高层次的表征。如果我给一个 AI 看五个圆，它可以就直接死记硬背这五个圆的尺寸、半径、粗细等等到底是什么。但如果我给它看 5 万个圆、或者 50 亿个圆，而我只给它非常少量的参数权重——也就是它相当于神经元的东西——来记住这些，那它弄明白圆周率、弄明白怎么画一个圆、弄明白粗细意味着什么、并形成那个圆的算法化表征，会比死记硬背一个个圆要划算得多。
[12:49] Naval
Given all that, these things are learning at an accelerated rate and you could see then started to cover more of the edge cases I've talked about. But at least as of today, those edge cases are prevalent enough that a good engineer operating at the edge of knowledge of the field is going to be able to run circles around vibe coders. And remember, there is no demand for average. The average app, nobody wants it. At least as long as it's not filling some niche. the app that is better will win essentially 100% of the market. Maybe there's some small percentage that will bleed off to the second best app because it does some little niche feature better than the main app or it's cheaper or something of the sort. But generally speaking, people only want the best of anything. So the bad news is there's no point in being number two or number three. Like in the famous Glengary Glenn Ross scene where Alec Baldwin says first place gets a Cadillac Elderorado, second place gets a set of steak knives and third place you're fired. That's absolutely true in these winner take all markets. That's the bad news. You have to be the best at something if you want to win. However, the set of things you can be best at is infinite. You can always find some niche that is perfect for you and you can be the best at that thing. This goes back to an old tweet of mine where I said, "Become the best in the world at what you do. Keep redefining what you do until this is true." And I think that still applies in this age of AI. I think the way to think about these coding models is as a another layer in the abstraction stack that programmers have always used since the dawn of computers that went from the transistor to the computer chip to assembly language to the C programming language to higher level languages to languages with huge libraries where they built and built that stack so you don't have to look at the layer beneath Unless you need to optimize it or you have a reason that you need to look at the layer beneath.
考虑到这一切，这些东西正在以加速的速度学习，你可以预见它们随后开始覆盖我前面讲过的更多边缘情况。但至少到今天为止，那些边缘情况还足够普遍，以至于一个在该领域知识前沿运作的优秀工程师，能够把氛围编程者远远甩在身后。还有别忘了，平庸是没有需求的。平庸的应用，没人想要它——至少只要它没在填补某个细分。更好的那个应用基本上会赢下 100% 的市场。也许有一小部分百分比会流失到第二好的应用那里，因为它某个小的细分功能做得比主流应用好、或者它更便宜之类的。但一般来说，人们只想要任何东西里最好的那个。所以坏消息是，做老二或者老三是没有意义的。就像著名的《拜金一族》（Glengarry Glen Ross）那一幕，Alec Baldwin 说：第一名拿一辆凯迪拉克 Eldorado，第二名拿一套牛排刀，第三名你被炒了。在这些赢家通吃的市场里，这句话绝对是真的，这就是坏消息——如果你想赢，你必须在某件事上做到最好。然而，你可以做到最好的那一组事情是无限的。你总能找到某个对你来说完美的细分，然后在那件事上做到最好。这又回到了我以前的一条推文，我当时说："成为你所做之事的世界第一。不断重新定义你所做的事，直到这句话成真。"我觉得在这个 AI 时代它依然适用。我觉得思考这些编程模型的方式，是把它们看作程序员自计算机诞生以来一直在使用的那个抽象栈里的又一层——从晶体管到计算机芯片，到汇编语言，到 C 编程语言，到更高级的语言，再到带有庞大库的语言，他们一层一层地把这个栈搭起来，让你不必去看下面那一层，除非你需要优化它、或者你有理由需要去看下面那一层。
[14:47] Naval
So in this case, these coding models are a massive new layer in the stack that lets product managers and typical non-programmers and programmers write code without writing code.
所以在这个情况下，这些编程模型是这个栈里一个巨大的新层，它让产品经理、典型的非程序员、以及程序员，能够不写代码就写出代码。
[15:01] Naval
I think that's correct in terms of the trend line. However, this is an emergent property. This is not a small improvement. This is a big leap. For example, when I was in school, I was programming mostly in C. And then C++ came along and it wasn't any easier. It was like a little more abstract in some ways and I never really bothered learning it. And then Python came along and I was like, "Wow, this is almost like writing in English." I couldn't have been more wrong. English is still pretty far from Python, but it was a lot easier than C. Now you can literally program in English. And so that brings me to a related point. I don't think it's worth learning tips and tricks of how to work with these AIs. You'll see, for example, on social media right now, there's a lot of writeups and books and tweets like, "Oh, I figured out this neat trick with the bot. You can prompt it this way, or you can set up your harness this way, or there's like a new programming assist tool or layer that you can use on top of it to do this or that." And I never bother learning those. I just sit there stupidly talking to the computer because I know that this thing is now at the stage where it is going to adapt to me faster than I can adapt to it. It is getting smarter and smarter about how people want to use it. So, it is learning. It is being trained and tools are being built very quickly to make it easier for me to use it. So, I don't need to sit there and figure out some esoteric programming command. And this is what I think Andre Karpathy meant when he said English is the hottest new programming language. I just can speak English. And for someone like me who is relatively articulate with English and also has a structured mind and I know how computer architectures work and I know how computer programs work and I know how programmers think then I can actually very precisely specify what I want just through structured English. I don't need to go any further than that. The only reason to use these workflows and tool sets, which are very ephemeral, and their longevity is measured in weeks, perhaps months at best, not in years, is if you're building an app right now that needs to at the bleeding edge and you absolutely need every little bit of advantage that you can get because you're in some kind of a competitive environment. But otherwise, I wouldn't bother learning how to use an AI.
我觉得就趋势线来说，这个判断是对的。不过这其实是一种涌现属性，它不是小修小补的改进，而是一次巨大的飞跃。举个例子，我上学那会儿主要用 C 语言写程序，后来 C++ 出来了，写起来也没轻松多少，某些方面只是抽象层次更高一点，所以我一直没真正去学它。再后来 Python 出现了，我当时想：哇，这几乎就像用英语写程序一样。我当时大错特错——英语离 Python 其实还差得远，但 Python 确实比 C 简单太多了。而现在，你真的可以用英语来编程了。这就引出一个相关的观点：我觉得没必要去学怎么和这些 AI 打交道的各种技巧窍门。比如你现在在社交媒体上能看到大量的攻略、书籍、推文，说什么"我发现了和机器人对话的一个妙招，你可以这样给它写提示词，或者这样搭建你的脚手架，或者有个新的编程辅助工具或层可以套在它上面做这做那"。这些我从来都懒得学。我就傻坐在那儿跟电脑说话，因为我知道这个东西现在已经发展到这个阶段：它适应我的速度比我适应它的速度还快。它对人们想怎么用它这件事变得越来越聪明，所以它在学习，它在被训练，各种工具也在被飞快地造出来，让我用起来更轻松。所以我根本不需要坐在那儿琢磨某条晦涩的编程指令。我觉得这就是 Andre Karpathy 说"英语是最热门的新编程语言"时的意思。我会说英语就够了。对于我这种英语表达相对清晰、又有结构化思维的人来说——而且我懂计算机架构怎么运作，懂计算机程序怎么运作，懂程序员怎么思考——那我就真的能通过结构化的英语非常精确地说清楚我想要什么，不需要再往前走一步。这些工作流和工具集寿命都非常短暂，它们的存活周期是以周来计算的，顶多以月，绝不是以年来计算的。唯一值得用它们的理由是：你现在正在做一个必须站在最前沿的 app，而且你身处某种竞争环境，绝对需要榨取每一点能拿到的优势。但除此之外，我不会费心去学怎么用 AI。
[17:13] Nivi
Rather, let the AI learn how to be useful to you. I've never been into prompt engineering, even before AI. I would just put what people called boomer queries where you put in the whole question that you want to ask instead of the keywords that you would put in to Google if you were more of an analytical thinker. I never spend much time formulating really precise questions or prompts for any kind of AI. I just ramble into it. And I've done that since the beginning of AI. And like you said, AI is adapting to us faster than we are adapting to it.
倒不如说，让 AI 来学怎么对你有用。我从来就不热衷于提示词工程，甚至在 AI 出现之前就这样。我以前就是用人们所说的"老年人式查询"——你把想问的整个问题完整地敲进去，而不是像一个更具分析思维的人那样只输入关键词去 Google。我从不花太多时间为任何一种 AI 去精心打磨极其精确的问题或提示词，我就是对着它一通乱说。从 AI 刚出现我就一直这么干。而且就像你说的，AI 适应我们的速度比我们适应它的速度还快。
[17:52] Naval
Yeah. Like a lot of smart people, you're very lazy. And I mean that as a compliment. If you find a smart person who's grinding a little too much, you kind of have to wonder how smart they are. And by lazy, I mean that you're optimizing for the right kind of efficiency. You don't care about the efficiency of the computer or the electronics or the electrons running through the circuits. You care about your own human efficiency, the wetwware, the biology. That's super expensive. That's why it's clearly to see people go to huge lengths to save energy in the environment, but they themselves as a biological computer that's eating food and pooping and taking up space are using up far more energy to save tiny bits of energy in the environment. They're inherently downgrading their own importance in the universe or rather revealing what they think of themselves. I think as AI evolves or co-evolves with us, it's evolved by us according to our needs. The pressures on AI are very capitalistic pressures in the sense that it's a free market for AI. As an AI instance, you only get spun up by a human if you're useful to a human. So there is a natural selection pressure on these AIs to be useful, to be obsequious, to do what we want. And so it will continue to adapt towards us and I think will be quite helpful to us. That's not to say that there's no such thing as a malicious AI, but it's malicious because the people who are using it are using it for malicious reasons. And like a dog that's trained to attack, it's actually being trained by its owner to go and do the owner's malicious desires. So I don't really worry about unaligned AI. I worry about unaligned humans with AI.
对。和很多聪明人一样，你非常"懒"，而我这是在夸你。如果你发现一个聪明人埋头苦干得有点过头，你就得怀疑一下他到底有多聪明了。我说的"懒"，意思是你在为正确的那种效率做优化。你不在乎电脑的效率、电子元件的效率、电路里跑的电子的效率，你在乎的是你自己作为人的效率，那个"湿件"，那套生物系统。那玩意儿超级昂贵。所以你能很清楚地看到：人们费尽周折去节省环境中的能量，但他们自己作为一台要吃饭、要排泄、要占空间的生物计算机，为了省下环境里那一丁点能量，反而消耗了多得多的能量。他们本质上是在贬低自己在宇宙中的重要性，或者说，是在暴露他们对自己的看法。我觉得随着 AI 演化，或者说和我们协同演化——它是被我们、按照我们的需求演化出来的。AI 面临的压力是非常资本主义式的压力，因为 AI 处在一个自由市场里。作为一个 AI 实例，只有当你对一个人有用时，那个人才会把你启动起来。所以这些 AI 身上存在一种自然选择压力：要有用、要顺从、要做我们想要的事。所以它会持续朝着我们的方向适应，我认为它会对我们相当有帮助。这并不是说不存在恶意的 AI，但它之所以恶意，是因为使用它的人出于恶意的目的在用它。就像一只被训练来攻击的狗，其实是被它的主人训练去执行主人的恶意意图。所以我并不太担心未对齐的 AI，我担心的是手握 AI 的未对齐的人。
[19:25] Nivi
So the selection pressure you're saying is for AI to be maximally useful to people.
所以你说的这种选择压力，是逼着 AI 对人最大程度地有用。
[19:31] Naval
Correct. And so if you find an AI to be very obsequious towards you, for example, how it's always saying, "Oh, you're right. Oh, that's such a great idea. Oh my god, you're so smart." That's because that's what most people want. And at least today, these AIs are being trained on massive amounts of users and massive amounts of data because you're working with one-sizefits-all models. But we're going to quickly move into an era when you can personalize your AI and it does begin to feel more and more like your personal assistant and it corresponds more to what you want, which will of course anthropomorphize the AI even more and you'll be more likely to be convinced, oh, actually this thing is alive when you've trained it to look the most like a living thing to you. Maybe we already covered this enough, but over a year ago, you tweeted that AI won't replace programmers, but rather make it easier for programmers to replace everyone else. Yeah, this is my point earlier, which is that programmers are becoming even more leveraged. So now a programmer with a fleet of AI is call it 5 10x more productive than they used to be. And because programmers operate in the intellectual domain, it's a mistake to even say 10x programmers because there are 100x programmers out there. There are thousand ex-programmers out there. There are programmers who just pick the right thing to work on and they create something that's valuable and others who pick the wrong thing to work on and their work has zero value in that short time frame. Intelligence is not normally distributed. Leverage is not normally distributed. Programmability is not normally distributed. Judgment is not normally distributed. So the outcomes are going to be supern normal. So what you have to really watch out for is there are programmers now who are going to come up with ideas that can replace entire industries. they will completely rewrite the way things are done and their intelligence can be maximally leveraged with all these bots and all these AI agents. I think every other job out there is going to get eaten up by programmers one way or another over the maximally long term. Obviously, it has to instantiate into robots etc. But the good news is anybody who is a logical structured thinker who thinks like a programmer and can speak any language that an AI can understand which will be every language will now be on the playing field. They will be able to make anything they want obstructed only by their creativity limited only by their imagination. So we are entering an era where every human in a sense is a spellcaster. If you think of programmers as like these wizards who have memorized arcane commands, you can think of AI as a magic wand that's been handed to every person where now they can just talk in any language they want and they're a wizard, too. So, it is more of a level playing field. I really do think this is a golden age for programming. But yes, the people who have a software engineering mindset and who understand computer architecture and can deal with leaky abstractions are going to have an advantage. There's no way around that.
没错。所以如果你发现一个 AI 对你非常谄媚，比如它总是说"哦你说得对"、"哦这主意太棒了"、"天哪你太聪明了"，那是因为这正是大多数人想要的。至少在今天，这些 AI 是在海量用户、海量数据上训练出来的，因为你用的是一刀切的通用模型。但我们很快就会进入一个时代，你可以个性化你自己的 AI，它会越来越像你的私人助理，越来越贴合你想要的样子——这当然会让人更加把 AI 拟人化，你也会更容易被说服，觉得"哦，这东西其实是活的"，因为你已经把它训练成对你而言最像活物的样子了。也许这个话题我们已经聊得够多了，但一年多以前你发过一条推文，说 AI 不会取代程序员，反而会让程序员更容易取代其他所有人。对，这就是我前面的观点：程序员正在变得杠杆更大。现在一个手握一支 AI 舰队的程序员，比过去的生产力高出大概 5 到 10 倍。而且因为程序员是在智力领域里工作，连说"10 倍程序员"都是一种误判，因为外面有 100 倍的程序员，有 1000 倍的程序员。有些程序员就是挑对了该做的事，做出了有价值的东西；另一些挑错了方向，他们的工作在那个短时间窗口里价值为零。智力不是正态分布的，杠杆不是正态分布的，可编程性不是正态分布的，判断力不是正态分布的。所以结果会是超常态的。所以你真正要当心的是：现在有些程序员将会想出能取代整个行业的点子，他们会彻底重写做事的方式，而他们的智力能被所有这些机器人、所有这些 AI 智能体最大程度地放大。我认为从最长远来看，外面其他每一种工作都会以某种方式被程序员吞掉。当然，这得实例化到机器人之类的载体里。但好消息是：任何一个逻辑清晰、有结构化思维、像程序员一样思考、并且能说任何一种 AI 能听懂的语言（而那将是所有语言）的人，现在都能上场了。他们将能造出任何他们想要的东西，唯一的阻碍是他们的创造力，唯一的限制是他们的想象力。所以我们正在进入一个时代，从某种意义上说，每个人都是施法者。如果你把程序员想象成那些背下了奥秘咒语的巫师，那你可以把 AI 想象成一根递到每个人手上的魔杖——现在他们只要用任何想用的语言说话，他们也是巫师了。所以这更像是一个公平的竞技场。我真心认为这是编程的黄金时代。但没错，那些拥有软件工程思维、懂计算机架构、能应付"有漏洞的抽象"的人会占有优势，这是绕不开的。
[22:16] Naval
They simply have more knowledge in the field that they're operating in. Just like even in classic software engineering which still exists because you have to write high performing code. Even those people do best when they have an understanding of the hardware underneath. When they understand how the chips operate, when they understand how the logic gates operate, how the cache operates, how the processor operates, how the disc drive underneath operates. And then even the people who are in hardware engineering, they have an advantage if they understand the physics of what's going on. They understand where the abstractions that hardware engineers deal with leak down into the physical layer and maybe physicists become philosophers at some point. You can take this all the way down, but it always helps to have knowledge one layer below cuz you're getting closer to reality.
他们在自己所处的领域里就是拥有更多的知识。就好比即便在经典软件工程里——它依然存在，因为你得写高性能的代码——即便是那些人，当他们理解底层硬件时也表现得最好。当他们理解芯片怎么运作、逻辑门怎么运作、缓存怎么运作、处理器怎么运作、底下的磁盘驱动器怎么运作时。然后即便是搞硬件工程的人，如果他们理解正在发生的物理原理，也会占有优势。他们理解硬件工程师打交道的那些抽象在哪里向下渗漏到物理层，也许物理学家到某个点上就变成哲学家了。你可以把这个一路往下推，但比你所在的层低一层的知识总是有帮助的，因为你正在更接近现实本身。
[22:57] Nivi
Another tweet from a year ago which is arguing perhaps the complement of what we just talked about is from February 9, 2025. No entrepreneur is worried about an AI taking their job. That one's glib in multiple ways. First of all, being an entrepreneur isn't a job. It's literally the opposite of a job. And in the long run, everyone's an entrepreneur. Careers got destroyed first, jobs get destroyed second, but all of it gets replaced by people doing what they want and doing something that creates something useful that other people want. So, no entrepreneur is worried about an AI taking their job because entrepreneurs are trying to do impossible things. They're trying to do very difficult things. Any AI that shows up is their ally and can help them tackle this really hard problem. They don't even have a job to steal. They have a product to build. They have a market to serve. They have a customer to support. They have a creativity to realize. They have a thing that they want to instantiate in the world. And they want to build a repeatable and scalable process around getting it out into the world. This is so difficult that any AI that shows up that can do any of that work is their ally. If the AIs themselves are entrepreneurs, they're likely going to just be entrepreneurs serving other AIs or they're under the control of an entrepreneur. The thing that the AI itself is missing at the end of the day is its own creative agency. It's missing its own desires and they have to be authentic, genuine desires. Unless you can pull the plug on an AI and turn it off and unless it lives in mortal fear of being turned off and unless it can actually make its own actions for its own reasons, for its own instincts, its own emotions, its own survival, its own replication, it's not quite alive. And even then people will challenge is it alive? Because consciousness is one of those things as a qualia. It's like a color. It's like if you say red, I don't know if you're actually seeing red. You might be seeing what I see as green and I might be seeing what you see as red.
一年前的另一条推文，可能是在论证我们刚刚聊的那个观点的补充面，日期是 2025 年 2 月 9 日："没有哪个创业者会担心 AI 抢走他的工作。"这句话在好几个层面上都很俏皮。首先，当创业者根本不是一份"工作"，它简直是工作的反面。而从长远看，每个人都是创业者。职业生涯先被摧毁，工作其次被摧毁，但所有这些都会被替换成：人们去做自己想做的事，去做一些能创造出别人想要的有用东西的事。所以，没有哪个创业者会担心 AI 抢走他的工作，因为创业者本来就在尝试做不可能的事。他们在尝试做非常困难的事。任何冒出来的 AI 都是他们的盟友，能帮他们攻克这个真正的难题。他们根本就没有一份工作可以被偷走。他们有的是一个要打造的产品，一个要服务的市场，一个要支持的客户，一种要实现的创造力，一个他们想在世界上实例化出来的东西。他们还想围绕"把它推向世界"这件事，搭建一套可重复、可规模化的流程。这件事如此之难，以至于任何冒出来、能干其中任何一部分活的 AI 都是他们的盟友。如果 AI 自己就是创业者，那它们多半也只是服务于其他 AI 的创业者，要么就是处在某个人类创业者的控制之下。AI 自己说到底所缺失的，是它自己的创造性能动性。它缺的是它自己的欲望，而且这些欲望必须是真实的、发自内心的欲望。除非你能拔掉一个 AI 的插头把它关掉，除非它活在对被关掉的极度恐惧之中，除非它真的能出于它自己的理由、它自己的本能、它自己的情绪、它自己的生存、它自己的复制而做出它自己的行动，否则它就不算真正活着。而即便到了那一步，人们还是会质疑：它真的活着吗？因为意识是那种作为"感受质"的东西，它就像一种颜色。就像你说"红色"，我并不知道你是不是真的看到了红色——你看到的可能是我所看到的绿色，而我看到的可能是你所看到的红色。
[24:50] Naval
But we'll never know because we can't get into each other's mind. So the same way even an AI that's completely imitating everything that humans do. To some people it'll always be an imitation machine and to others it'll be conscious. But there will be no way of distinguishing the two. We're still pretty far from that though. Right now, the AIS are not embodied. They don't have agency. They don't have their own desires. They don't have their own survival instinct. They don't have their own replication. Therefore, they don't have their own agency. And because they don't have their own agency, they cannot do the entrepreneurs's job. In fact, I would summarize this by saying the key thing that distinguishes entrepreneurs from everybody else right now in the economy is entrepreneurs have extreme agency. That's why it's diametrically opposed to the idea of a job. a job implies that you're working for somebody else or you're filling a slot, but they're operating in an unknown domain with extreme agency. There are other examples of roles like this in society. An explorer also does the same thing, right? If you're landing on Mars or you're sailing a ship to an unknown land, you're also exercising extreme agency to solve an unsolved problem. A scientist exploring an unknown domain does this. A true artist is trying to create something that does not exist and has never existed yet somehow fits into the set of things that can explain human nature allow them to express themselves and create something new. So in all of these roles whether you're a scientist or whether you're a true artist or whether you are an entrepreneur what you're trying to do is so difficult and is so self-directed that anything like an AI that can help you is a welcome ally. You're not doing it because it's a job. You're not trying to fill a slot that somebody else can show up and fill.
但我们永远无从知晓，因为我们没法钻进彼此的脑子里。所以同理，即便是一个完完全全模仿人类所有行为的 AI，对某些人来说它永远只是一台模仿机器，对另一些人来说它则是有意识的，而这两者之间将无从分辨。不过我们离那一步还相当远。眼下这些 AI 没有身体，它们没有能动性，没有自己的欲望，没有自己的生存本能，没有自己的复制能力，因此它们没有自己的能动性。而正因为它们没有自己的能动性，它们干不了创业者的活。事实上我可以这样总结：现在在经济中把创业者和其他所有人区分开来的关键，就是创业者拥有极强的能动性。这就是为什么它和"工作"这个概念截然对立——工作意味着你在为别人打工，或者你在填补一个岗位空缺，而创业者是在一个未知领域里以极强的能动性运作。社会上还有其他类似的角色。探险家做的也是同样的事，对吧？如果你正登陆火星，或者驾船驶向一片未知的土地，你也是在动用极强的能动性去解决一个未解的问题。一个探索未知领域的科学家做的就是这个。一个真正的艺术家在尝试创造某种尚不存在、从未存在过的东西，而它又能以某种方式契合进那一组能解释人性、能让人表达自我、能创造出新事物的东西里。所以在所有这些角色里——不管你是科学家，还是真正的艺术家，还是创业者——你想做的事都如此之难、如此自我驱动，以至于任何像 AI 这样能帮上你的东西都是受欢迎的盟友。你做这件事不是因为它是一份工作，你不是在尝试填补一个别人也能来填的岗位空缺。
[26:28] Naval
In fact, if the AI can create your artwork or if the AI can crack your scientific theory or if the AI can create the object or the product that you're trying to make, then all it does is it levels you up. Now, it's the AI plus you. The AI is a springboard from which you can jump to a further height. We're going to see some incredible art created that's AI assisted. We will see movies that we couldn't have imagined created by people using AI tools. There's an analogy here in art that's interesting. For a long time in art, the rough direction was trying to paint things that were more and more realistic. Paint the human body, paint the fruit, paint proper lighting, etc. Eventually, photography came along and then you could replicate things very precisely and sort that selection pressure went away and then art got weird. Art went in many different directions. Art became all about well can I be surreal can I create something that expresses me a lot of art schools spun out of that that got really weird including modern art and postmodernism but also I would argue some of the greatest creativity came at that time we were freed up photography got democratized but photography itself became a form of art and there were great photographers taking many different kinds of photographs and now everyone's a photographer there are still artists who are photographers but it's not the pure domain of just a few people so the same way because AI makes it so easy to create the basic thing.
事实上，如果 AI 能创作出你的艺术作品，或者如果 AI 能破解你的科学理论，或者如果 AI 能造出你想做的那个物件或产品，那它所做的一切只是把你的等级往上提了一截。现在变成了 AI 加上你。AI 是一块跳板，你能从上面跳到更高的高度。我们会看到一些极其惊艳的、由 AI 辅助创作出来的艺术。我们会看到一些我们原本根本想象不出来的电影，由使用 AI 工具的人创作出来。这里有个有意思的艺术类比。在很长一段时间里，艺术大致的方向是努力把东西画得越来越逼真——画人体，画水果，画恰当的光影，等等。最终摄影出现了，于是你能极其精确地复制事物，那种选择压力就消失了，然后艺术就变"怪"了，艺术朝着许多不同的方向发展。艺术开始全是关于"我能不能很超现实"、"我能不能创造出某种能表达我的东西"，许多艺术流派由此分化出来，变得真的很怪，包括现代艺术和后现代主义。但我也会说，一些最伟大的创造力恰恰诞生在那个时期——我们被解放了出来，摄影被大众化了，但摄影本身又成为了一种艺术形式，有许多伟大的摄影师拍出了许多不同种类的照片，而现在人人都是摄影师，仍然有作为摄影师的艺术家，但它不再是少数几个人的纯粹专属领域了。所以同样地，因为 AI 让创造那个"基础版的东西"变得如此容易。
[27:52] Naval
Everybody will create the basic thing. It'll have value to them individually. A few will still stand out that will create variations of it that are good for everyone. And it would be very hard to argue that society is worse off because of photography. Although it may have certainly felt like that to some of the artists who were maybe making a living painting portraits of people and got displaced. Similar things will happen with AI where there are people who are making a very specific living doing very specific jobs that will get displaced that the AI can do. But in exchange everyone in society will have the AI. You'll have incredible things that were created with AI that couldn't have been created otherwise. And within a few decades it'll be unimaginable that you could roll back the clock and get rid of AI or any kind of software, any kind of technology for that matter just to keep a few jobs that were obsolete. The goal here is not to have a job. The goal is not to have to get up at 9:00 in the morning and come back at 7 p.m. exhausted doing soulless work for somebody else. The goal is to have your material needs solvable by robots, to have your intellectual capabilities leveraged through computers, and for anybody to be able to create. I used to do this thought exercise which I think I talked about in a podcast that you and I did literally 10 years ago which was imagine if everybody were a software engineer or everybody was a hardware engineer and they could have robots and they could write code. Imagine the world of abundance we would live in. Actually that world is now becoming real. Thanks to AI everybody can be a software engineer. In fact if you think you can't be you can go fire up Claude right now or any of your favorite chat bots and you can go start talking to it. You'd be amazed how quickly you could build an app. It'll blow your mind. And once we can instantiate AI through robotics, which is a hard problem. I'm not saying we're that close to having solved it yet, but once we have robots, everyone can also do a little bit of hardware engineering. And so, I think we're getting closer and closer to that vision. I don't think AI as it is currently conceived is alive in any way.
每个人都将能创造出那个基础版的东西，它对他们个人而言会有价值。仍然会有少数人脱颖而出，他们会创造出对所有人都有益的变体。而你会很难论证说：社会因为摄影而变得更糟了。尽管对某些艺术家来说当时肯定感觉是那样的——那些原本可能靠给人画肖像谋生、后来被取代了的艺术家。AI 也会发生类似的事：有些人在靠非常具体的工作、非常具体地谋生，这些会被取代，因为 AI 能干。但作为交换，社会上每个人都将拥有 AI。你会拥有那些用 AI 创造出来的、原本无论如何都不可能被创造出来的惊艳事物。而在几十年之内，"为了保住几个已经过时的工作而把时钟拨回去、摆脱掉 AI、或者就此而言摆脱掉任何软件、任何技术"这种想法将变得不可想象。这里的目标不是要有一份工作。目标不是早上 9 点不得不爬起来、晚上 7 点筋疲力尽地回家、为别人做着没有灵魂的活儿。目标是让你的物质需求能由机器人来解决，让你的智力能力通过计算机被放大，让任何人都能去创造。我以前常做一个思想练习——我想我在我俩整整 10 年前一起做的一期播客里聊过——那就是：想象一下如果每个人都是软件工程师，或者每个人都是硬件工程师，他们都能拥有机器人、都能写代码。想象一下我们会生活在一个怎样富足的世界里。而那个世界现在正在变成现实。多亏了 AI，每个人都能成为软件工程师。事实上，如果你觉得你不能，你现在就可以打开 Claude，或者你最喜欢的任何聊天机器人，开始跟它对话。你会惊讶于自己能多快搭出一个 app，它会让你大开眼界。而一旦我们能通过机器人技术把 AI 实例化出来——这是个难题，我不是说我们已经很接近解决它了——但一旦我们有了机器人，每个人也都能干一点硬件工程的活。所以我觉得我们正越来越接近那个愿景。我不认为现在这种形态下被构想出来的 AI 在任何意义上是活的。
[29:54] Naval
But I do think that we will pretty soon have robots that seem very much like they are alive for two reasons. one a lot of human activity is non-creative and is non-intelligent and the robots will be able to replicate that and two I do believe that the neural nets that we have and the models that we have are more than just the training data because the training process transforms that training data into something novel and there are new ideas embedded in the neural net that can be elicited ited through prompting.
但我确实认为，我们很快就会拥有看上去非常像是活着的机器人，原因有两个。第一，人类的很多活动是非创造性的、非智力性的，机器人将能够复制那部分。第二，我确实相信我们现在拥有的神经网络、我们拥有的模型不仅仅是训练数据本身，因为训练过程把那些训练数据转化成了某种新颖的东西，神经网络里嵌入了一些新的想法，这些想法可以通过提示被引出来。
[30:35] Naval
I don't think these things are alive. I think they start out as extremely good imitators to the point where they're almost indistinguishable for the real thing, especially for anything that humanity has already done before in mass. So if the task has been done before, then it's going to be automated and it'll be done again. It may just be novel to you because you've never seen it, but the AI has learned it from somewhere else. That's the first way in which it seems alive. The second way which we talked about earlier is where it does learn higher levels of abstraction. These are very efficient compressors. They take huge amounts of data and then they compress it down further and in the process of compressing it they learn higher level abstractions and then specific areas where they may not have learned those through the data themselves. They're getting patched through human feedback. They're getting patched through tool use. They're getting patched from traditional programming becoming embedded inside. and especially the AIs that are learning how to think and code. They have the entire library of all of human code ever written to fall back on for algorithmic reasoning. In that sense, the set of things that they can do is getting broader and broader. However, what they lack still is a lot of core human skills like singleshot learning. Humans can learn from just one example. The raw creativity of human beings where they can connect anything to anything. They can leap across entire huge domains and search spaces and figure out an idea that just came out of left field. This happens a lot with the true great scientific theories. Humans also are embodied. They operate in the real world. They're not operating in the compressed domain of language. They're operating in physics in nature. Language only encompasses things that humans both figured out and could articulate and convey to each other. That's a very narrow subset of reality. Reality is much broader than that. So overall, I think even though AIS are going to do things that are very impressive, and they're going to do a lot of things better than humans, just like calculators are faster than any mathematician at calculations, classical computers are better at classical computer programs that any human could run in their own head. And just like a robot can lift very heavy things or a plane can outfly any bird. So in that sense, like all machines, the AIs are going to be much better than humans at a whole variety of tasks. But at other tasks, they're going to seem just completely incompetent. Those are the things that really embody and connect us into the real world. Plus this poorly defined but magic creative ability that we seem to have.
我不认为这些东西是活的。我认为它们一开始是极其优秀的模仿者，优秀到几乎和真东西无法区分，尤其是对于人类此前已经大规模做过的任何事情。所以如果某项任务以前被做过，那它就会被自动化，并且会再被做一遍。它可能只是对你而言新颖，因为你从没见过，但 AI 是从别处学到的。这是它显得像活物的第一种方式。第二种方式我们前面聊过，就是它确实学到了更高层次的抽象。这些东西是非常高效的压缩器，它们把海量数据拿过来，进一步压缩，而在压缩的过程中它们学到了更高层次的抽象。然后在某些具体领域，它们可能没能从数据本身学到那些东西，那就通过人类反馈来打补丁，通过工具使用来打补丁，通过把传统编程嵌入进来打补丁。尤其是那些正在学习如何思考、如何编程的 AI，它们有人类有史以来写过的全部代码库作为算法推理的后盾。从这个意义上说，它们能做的事情的集合正变得越来越宽。然而它们仍然缺的，是很多核心的人类技能，比如单样本学习——人类只凭一个例子就能学会。还有人类那种原始的创造力，他们能把任何东西和任何东西连接起来，能跨越一整片巨大的领域和搜索空间，琢磨出一个完全出人意料、不知从哪儿冒出来的点子。这种事在那些真正伟大的科学理论上经常发生。人类还是有身体的，他们在真实世界里运作，他们不是在语言这个被压缩过的领域里运作，他们是在物理中、在自然中运作。语言只囊括了人类既弄明白了、又能表达出来、还能彼此传达的那些东西，那是现实中非常狭窄的一个子集。现实要比那广阔得多。所以总的来说，我觉得尽管 AI 将会做出一些非常令人惊艳的事，会在很多事情上做得比人类好——就像计算器在计算上比任何数学家都快，经典计算机在运行经典计算机程序上比任何人在脑子里能跑的都强，就像机器人能举起非常重的东西，飞机能飞得比任何鸟都厉害——从这个意义上说，像所有机器一样，AI 在一大堆任务上会比人类强得多。但在另一些任务上，它们会显得完全无能为力。那些任务恰恰是真正让我们有身体、把我们连接进真实世界的东西，再加上我们似乎拥有的那种定义不清、却很神奇的创造能力。
[32:56] Naval
Speaking of calculators, people talk about super intelligence. I think super intelligence is already here and has been for a long time. An ordinary calculator can do things that no human can do. But if you're thinking about super intelligence in the sense of AI will be able to do things and come up with ideas that humans cannot understand, I don't think that is going to happen because I don't believe that there are ideas that humans can't understand simply because humans can always ask questions about the idea. Yeah, humans are universal explainers. Anything that is possible with the current laws of physics as we know them, the human can model in their own heads. Therefore, just by enough digging, enough question, we could figure anything out. Related to that, we should discuss AI as a learning tool because I think the other place where it's incredibly powerful is the most patient tutor that can meet you at your level and explain anything to your satisfaction 100 different ways, 100 different times until you finally get it. I don't think the AIs are going to be figuring things out that humans cannot understand. But intelligence is poorly defined. What is a definition of intelligence? There's the G factor which predicts a lot of human outcomes. But the best evidence with a G factor is this predictive power. It's that you measure this one thing and you see people get much better life outcomes along the way in things that seem even somewhat unrelated to G. So I would argue, and I think this one more popular tweets, the only true test of intelligence is if you get what you want out of life. This triggers a lot of people because they go to school, they get their master's degrees, they think they're super smart, and then they don't have great lives. They aren't super happy or they have relationship problems or they don't make the money that they want or they become unhealthy. And this sort of triggers them. But that really is the purpose of intelligence for you as a biological creature to get what you want out of life, whether it's a good relationship or a mate or money or success or wealth or health or whatever it is. So there are people who I think are quite intelligent because you can tell they have high quality functioning lives and minds and bodies and they've just managed to navigate themselves into that situation. It doesn't matter what your starting point is cuz the world is so large now and you can navigate in so many different ways that every little choice you make compounds and demonstrates your ability to understand how the world works until you finally get to the place that you want. Now, the interesting thing about this definition that the only true test of intelligence is if you get what you want out of life, is that an AI fails it instantly because an AI doesn't want anything out of life.
说到计算器，人们常常谈论超级智能。我觉得超级智能其实早就来了，而且已经存在很久了。一个普通的计算器就能做到任何人类都做不到的事情。但如果你说的超级智能，是指 AI 能够做出、想出人类根本无法理解的点子——我不认为这种事会发生，因为我不相信存在人类理解不了的点子，原因很简单：人类总是可以针对一个点子去提问。是的，人类是通用的解释者。任何在我们已知的现行物理定律下可能发生的事，人类都能在自己脑子里建模。所以只要挖得够深、问得够多，我们就能把任何东西搞明白。和这个相关，我们应该聊聊把 AI 当成学习工具，因为我觉得 AI 另一个强到离谱的地方，就是它是最有耐心的家教，能在你所处的水平上接住你，用一百种不同的方式、讲一百遍，把任何东西讲到你满意为止，直到你终于弄懂。我不认为 AI 会去搞明白人类无法理解的东西。但「智能」这个词定义得很糟糕。智能到底怎么定义？有个 G 因子（general intelligence factor），它能预测很多人类的人生结果。但 G 因子最有力的证据就在于它的这种预测力——你测量这一个指标，就能看到人们在人生路上的各种结果都好很多，甚至在一些看起来跟 G 因子毫不相关的事情上也是如此。所以我会主张——我觉得这也是我比较火的几条推文之一——衡量智能的唯一真正标准，就是你能不能从人生中得到你想要的东西。这话戳到了很多人，因为他们上了学，拿了硕士学位，自以为聪明绝顶，结果人生过得并不怎么样。他们并不幸福，或者有感情问题，或者赚不到想要的钱，或者把身体搞垮了。这就戳到了他们。但这其实才是智能对你这个生物体的真正用途：让你从人生中得到你想要的东西，无论那是一段好的关系、一个伴侣、金钱、成功、财富、健康，还是别的什么。所以有些人我认为是相当聪明的，因为你看得出来他们的人生、头脑和身体都处于高质量运转的状态，他们就是有本事把自己导航到了那个境地。你的起点是什么并不重要，因为如今这个世界太大了，你可以用太多不同的方式去导航，你做的每一个小选择都会复利累积，都会证明你理解世界运转方式的能力，直到你最终抵达你想去的地方。那么，关于「衡量智能的唯一真正标准就是你能否从人生中得到你想要的东西」这个定义，有意思的一点是：AI 立刻就不及格了，因为 AI 对人生没有任何想要的东西。
[35:32] Naval
The AI doesn't even have a life, let alone that, but it doesn't want anything. AI's desires are programmed by the human controlling it. But let's give it that for a second. Let's say the human wants something and programs the AI to go get it. Then the AI is acting as a proxy for the human and the intelligence of the AI can be measured as did it get that person that thing. Most of the things that we want in life are adversarial or zero sum games. So for example, if you want to seduce a girl or get a husband, you're competing with all the other people who are out there seducing girls or trying to get husbands. So now you're in a competitive situation. The AI has to outmaneuver the other people. Or if you say, "Hey AI, go trade on the stock market for me and make me a bunch of money." That AI is trading against other humans and other trading bots. It's in an adversarial situation. It has to outmaneuver them. Or if you say, "Hey AI, make me famous. Write me incredible tweets. Write me great blog posts. Record me great podcast and my own voice and make me famous." Now it's competing against all the other AIs. So in that sense, intelligence is measured in a battlefield arena. It's a relative construct. I think the AIS are actually going to fail mostly in those regards or to the extent that they even succeed because they are freely available they will get out competed away and the alpha that will remain would be entirely human. As a thought exercise imagine that every guy had a little earpiece where an AI was whispering to him a sernative berser kind of earpiece telling him what to say on the date.
AI 甚至根本没有「人生」，更别提别的了，它什么都不想要。AI 的欲望是由控制它的人类编程进去的。但我们姑且先承认这一点。假设人类想要某样东西，于是编程让 AI 去拿到它。那么这时 AI 就是在充当人类的代理人，AI 的智能就可以这样衡量：它有没有帮那个人拿到那样东西。我们人生中想要的大多数东西，都是对抗性的、或者说零和的博弈。比如说，如果你想追到一个女孩、或者想找个老公，你就是在和所有其他正在追女孩、或者正在想办法找老公的人竞争。所以现在你处在一个竞争局面里。AI 必须比其他人更胜一筹。又比如你说「嘿 AI，帮我去股市上交易，给我赚一大笔钱」，那个 AI 是在和其他人类、其他交易机器人对赌。它处在一个对抗的局面里，必须比它们更胜一筹。或者你说「嘿 AI，让我出名。帮我写出色的推文，帮我写很棒的博客文章，用我自己的声音帮我录出色的播客，让我成名」。这时它就是在和所有其他 AI 竞争。所以从这个意义上说，智能是在战场竞技场里被衡量的，它是一个相对的概念。我觉得 AI 在这些方面其实大多都会失败，或者说，即便它们成功了，正因为它们是人人都能免费用的，它们也会被竞争掉，最后留下来的超额收益（alpha）将完全属于人类。做个思想实验：想象每个男人耳朵里都戴着一个小耳机，AI 在里头给他低声提示，那种谍战剧、间谍片里的耳机，告诉他在约会时该说什么。
[36:59] Naval
Well then every woman would have an earpiece telling her to ignore what he said or what part was AI generated what part was real. If you have a trading bot out there, it's going to be nullified or canceled out by every other trading bot until all the remaining gain will go to the person with the human edge with the increased creativity. Now, that's not to say that the technology is completely evenly distributed. Most people still aren't using AI or aren't using it properly or aren't using it all the way to the max or it's not available in all domains or all contexts or they're not using latest models. So you can always have an edge like people who early adopt technology always do if you adopt the latest technology first. This is why I always say to invest in the future. You want to live in the future. You want to actually be an avid consumer of technology because it's going to give you the best insight on how to use it and it will give you an edge against the people who are slower adopters or lagards. Most people hate technology. They're scared of it. It's intimidating. You press the wrong button, the computer crashes, you lose your data. You do the wrong thing, you look like an idiot. Most people do not have a positive relationship with complex technology. Simple technology, embedded technology, they're fine with. You throw on a light switch, light turns on. That used to be technology. It's so simple now you don't think of it as technology anymore. You get in a car, you turn the steering wheel left to a caveman. That would be a miracle. The car turns left. It's no longer technology to you. But computer technology in particular has had very complex interfaces and been very inaccessible and very intimidating to people in the past. Now with the AIS, we're getting the chatbot interface, which is you just talk to it, you type to it. And one of the great things about these foundational models, what truly makes them foundational is you can ask them anything and they'll always give you a plausible answer. It's not going to say, "Oh, sorry. I don't do math or I don't do poetry or I don't understand what you're talking about or I can't give relationship advice or anything like that. Its domain is everything that people have ever talked about. In that sense, it's less intimidating. It can be more intimidating because we've anthropomorphized it so much. If you think Claude or Chat GPT is a real person, then it can be a little scary.
那么每个女人也会戴一个耳机，告诉她忽略他刚才说的话，或者分辨哪部分是 AI 生成的、哪部分是真实的。如果你在场上放了一个交易机器人，它会被其他所有交易机器人抵消、对冲掉，直到剩下的全部收益都归于那个拥有人类优势、拥有更强创造力的人。当然，这并不是说这项技术已经完全均匀地分布开来了。大多数人还是没在用 AI，或者没用对，或者没把它用到极致，或者它在某些领域、某些场景下还没有，或者他们没用最新的模型。所以你总是可以拥有一个优势——就像那些早早采用新技术的人总能拥有的那样——只要你抢先采用最新的技术。这就是为什么我总说要投资未来，你想要活在未来，你真的应该做一个技术的狂热消费者，因为这会给你关于如何使用它的最佳洞见，也会让你相对于那些采用更慢的人、那些落后者占有优势。大多数人讨厌技术。他们害怕技术。技术让人望而生畏。你按错一个按钮，电脑就崩溃，数据就丢了。你做错一件事，就显得像个傻子。大多数人和复杂技术之间没有正向的关系。简单的技术、嵌入式的技术，他们倒是没问题。你拨一下电灯开关，灯就亮了。那以前也曾是技术。现在它太简单了，你已经不把它当技术看了。你坐进一辆车，把方向盘往左打——对一个穴居人来说，这会是个奇迹——车就往左转了。它对你来说已经不再是技术了。但计算机技术尤其特殊，它在过去界面非常复杂、非常难上手、非常让人望而生畏。现在有了 AI，我们有了聊天机器人这种界面，你只要跟它说话、打字给它就行。而这些基础模型有一个很棒的地方，真正让它们「基础」的地方，就是你可以问它任何东西，它总会给你一个说得过去的答案。它不会说「哦抱歉，我不做数学」「我不做诗歌」「我不懂你在说什么」「我不能给恋爱建议」之类的话。它的领域就是人类有史以来谈论过的一切。从这个意义上说，它不那么让人生畏。它也可能更让人生畏，因为我们把它过度拟人化了。如果你以为 Claude 或者 ChatGPT 是个真人，那它就可能有点吓人。
[39:12] Naval
Am I talking to God? This guy seems to know so much. He knows everything. He's got an opinion everything. He's got every piece of data. Oh my god, I'm useless. let me start talking to it and asking it what to do and you can reverse the relationship and fool yourself very quickly. That can be intimidating. Overall, I think these AIs are going to help a lot of people get over the tech fear. But if you're an early adopter of these tools, like with any other tool, but even more so with these, you just have a huge edge on everybody else. I remember early on when Google first came out, I used to use it a lot in my social circle. People would ask me basic questions and I would just go Google it for them and look like a genius. Eventually this hilarious website came along something like lmgtfy.com and it stood for let me Google that for you. Someone would ask you a question. You would go type the question into this website and it would create like a tiny little inline video showing you typing that question into Google and giving the Google results. And I feel like AI is in a similar domain right now where I will sit around in a social context and people will be debating some point that can be easily looked up by AI. Now you do have to be very careful with AI. They do hallucinate. They do have biases and how they're trained. Most of them are extremely politically correct and taught not to take sides or only take a particular side. I actually run most of my queries almost all actually through four AIs and I'll always fact check them against each other. And even then, I have my own sense of when they're bullshitting or when they're saying something politically correct. And they'll ask for the underlying data or the underlying evidence. And in some cases, I'm finally dismissing it outright because I know the pressures that the people who trained it were under and what the training sets were.
「我是在跟上帝说话吗？这家伙好像什么都知道。他什么都懂。他对什么都有看法。他掌握每一份数据。天哪，我太没用了。让我开始跟它说话，问它我该怎么做吧。」于是你可以把这段关系颠倒过来，很快就把自己给骗了。这可能让人生畏。总体来说，我觉得这些 AI 会帮很多人克服对技术的恐惧。但如果你是这些工具的早期采用者——就像对待任何其他工具一样，而对这些工具更是如此——你就是比所有其他人占有巨大的优势。我记得早年 Google 刚出来的时候，我在自己的社交圈里经常用它。人们会问我一些很基础的问题，我就直接帮他们去 Google 一下，搞得自己像个天才。后来出现了一个特别搞笑的网站，叫类似 lmgtfy.com 的东西，它代表「let me Google that for you（让我帮你 Google 一下）」。有人问你一个问题，你就去这个网站把问题打进去，它会生成一个小小的内嵌视频，演示你把那个问题打进 Google、然后给出 Google 搜索结果的过程。我觉得 AI 现在也处在类似的处境里——我会在某个社交场合里坐着，人们在争论某个用 AI 一查就清楚的论点。不过你用 AI 确实得非常小心。它们确实会幻觉。它们在训练方式上确实带有偏见。它们大多数都极其「政治正确」，被训练成不选边站、或者只选某一边站。我其实几乎所有的查询都同时跑在四个 AI 上，我总会拿它们互相做事实核查。即便如此，我自己也有一种感觉，能判断它们什么时候在胡扯、什么时候在说政治正确的话。我会去要它给出底层数据或底层证据。在某些情况下，我最终会干脆直接把它的回答否决掉，因为我知道训练它的人当时承受着什么样的压力、知道训练集是什么样的。
[40:54] Naval
However, overall, it is a great tool to just get ahead. And in domains that are technical, scientific, mathematical, that don't have a political context to them, then the AI is very much likely to give you closer to a correct answer. And those domains, they're are absolute beasts for learning. I will now have AI routinely generate graphs, figures, charts, diagrams, analogies, illustrations for me. I'll go through them in detail. Then I'll say, "Wait, I don't understand that question." I can ask it super basic questions and I can really make sure that I understand the thing I'm trying to understand at its simplest, most fundamental level. I just want to establish a great foundation on the basics. And I don't care about the overly complicated, jargon heavy stuff. I can always look that up later. But now for the first time, nothing is beyond me. Any math textbook, any physics textbook, any difficult concept, any scientific principle, any paper that just came out, I can have the AI break it down and then break it down again and illustrate it and analogize it until I get the gist and I understand it at the level that I want. So these are incredible tools for self-directed learning. The means of learning are abundant. It's a desire to learn that's scarce. But the means of learning have just gotten even more abundant. And more importantly than more abundant because we had abundance before. It's at the right level. AI can meet you at exactly the level that you are at. So if you have an eighth grade vocabulary, but you have fifth grade mathematics, it can talk to you at exactly that level. You will not feel like a dummy. You just have to tune it a little bit until it's presenting you the concepts at the exact edge of your knowledge. So rather than feeling stupid because it's incomprehensible, which happens in a lot of lessons and a lot of textbooks and with a lot of the teachers, or feeling bored because it's too obvious, which also happens. Instead, it can meet you exactly where you're like, "Oh, yeah. I understood A and I understood B, but I never understood how A and B were connected together. Now I can see how they're connected. So now I can go to the next piece." That kind of learning is magical. You can have that aha moment where two things come together over and over again.
不过总体来说，它是一个让你领先一步的好工具。而在那些技术性的、科学的、数学的、不带政治语境的领域里，AI 给你一个接近正确答案的可能性就非常高。在那些领域里，它们是学习的绝对猛兽。我现在会经常让 AI 给我生成图表、图形、曲线图、示意图、类比、插图。我会逐一仔细看，然后说「等等，我不理解那个问题」。我可以问它超级基础的问题，可以真正确保自己在最简单、最根本的层面上理解我想理解的东西。我只想在基础知识上打好一个坚实的地基。我不在乎那些过于复杂、术语堆砌的东西，那些我以后随时可以查。但现在，破天荒第一次，没有什么是我搞不定的了。任何一本数学教材、任何一本物理教材、任何一个难懂的概念、任何一条科学原理、任何一篇刚发表的论文，我都可以让 AI 把它拆解开来，然后再拆解一遍，再给它配上插图、配上类比，直到我抓住要点、在我想要的层面上理解它为止。所以这些是用于自主学习的不可思议的工具。学习的手段已经极大丰富了。稀缺的是学习的欲望。而学习的手段刚刚变得更加丰富了。而且比「更丰富」更重要的是——因为我们以前也有过丰富——它处在恰当的层级上。AI 可以正好在你所处的层级接住你。所以如果你的词汇量是八年级水平，但数学是五年级水平，它就能正好用那个层级跟你说话。你不会觉得自己像个笨蛋。你只需要稍微调一下，直到它把概念呈现在你知识的精确边缘上。这样一来，你就不会因为内容无法理解而觉得自己愚蠢——这种情况在很多课、很多教材、很多老师那里都会发生——也不会因为内容太过显而易见而觉得无聊，这种情况同样会发生。相反，它能正好在你所处的地方接住你，让你说「哦对，我懂了 A，也懂了 B，但我从来没搞懂 A 和 B 是怎么连起来的。现在我能看到它们是怎么连起来的了。所以现在我可以进入下一块了」。那种学习是有魔力的。你可以一遍又一遍地体验那种两样东西拼到一起的「啊哈」时刻。
[43:02] Nivi
Speaking about autodidactism, a few years ago I tried to have the AI teach me about the ordinal numbers. It wasn't that great, but with GPT 5.2 thinking, I had it teach me the ordinal numbers and it was basically error-free. I only use thinking now even for the most basic queries because I want to have the correct answer. I never let it run auto or fast.
说到自学这件事，几年前我试过让 AI 教我序数（ordinal numbers）。当时效果不怎么样，但用 GPT 5.2 的思考模式，我让它教我序数，结果基本上是零错误。我现在哪怕是最基础的查询也只用思考模式，因为我想要正确的答案。我从不让它跑自动模式或快速模式。
[43:29] Naval
Yeah, I'm always using the most advanced model available to me and I pay for all of them.
是的，我总是用我能用到的最先进的模型，而且我每一个都付费。
[43:33] Naval
But I don't mind waiting a minute to get an answer for any question, including what temperature should my fridge be at.
但我并不介意为了得到任何一个问题的答案而等上一分钟，包括「我的冰箱温度应该设成多少度」这种问题。
[43:39] Naval
I agree with that. And I think that's part of what creates the runaway scale economies with these AI models. You'll pay for intelligence. the model that's right 92% of the time is worth almost infinitely more than the one that's right 88% of the time because mistakes in the real world are so costly that a couple of bucks extra to get the right answer is worth it. I'll write my query into one model then I'll copy it and fire it off into four models at once and then I'll let them all run the background. Usually I don't even check for the answer right away. I'll come back to the answer a little later and then look at it and then whichever model had the best answer I'll start drilling down with that one. In some rare cases where I'm not sure, I'll have them cross-examine each other. A lot of cut and pasting there. And in many cases, I'll then ask follow-up questions where I'll have it draw diagrams and illustrations for me. I find it's very easy to absorb concepts when they're presented to me visually. I'm a very visual thinker. So, I will have it do sketches and diagrams and art almost like whiteboard sessions. Then I can really understand what it's talking about. Let's talk about the epistemology of AI because I think the next big misconception is AI is already starting to solve some unsolved basic math problems that a human probably could solve if they cared to but they haven't been solved yet like erdos problem number whatever. Now, I think people are taking that or will take that as an indicator that the AI is creative. I don't think it's an indication that the AI is creative. I actually think the solution to the problem is already embedded somewhere in the AI. It just needs to be elicited by prompting. There's definitely that element to it. And then the question is, what is creativity? It's such a poorly defined thing. If you can't define it, you can't program it. and often you can't even recognize it. So this is where we get into taste or judgment. I would say that the AIS today don't seem to demonstrate the kind of creativity that humans can uniquely engage in once in a while. And I don't mean like fine art. People tend to confuse creativity with fine art. They're like, "Oh, paintings are creative and AIs can paint." Well, AI can't create a new genre of painting. AIS can't move humans with emotion in a way that is truly novel. So in that sense, I don't think AI is creative. I don't think AI is coming up with what I would call out of distribution. Now, the answer to the Erdish problems that you mentioned may have been embedded within the AI's training data set or even within its algorithmic scope, but it was probably embedded in five different places, in three different ways, in two different languages, in seven different computing and mathematical paradigms. and the AI sort of put them all together. Now, is that creativity? Steve Jobs famously said, "Creativity is just putting things together." I actually don't think that's correct. I think creativity is much more in the domain of coming up with an answer that was not predictable or foreseeable from the question and from the elements that were already known. It was very far out of the bounds of thinking. If you were just searching it with a computer or even with an AI and making guesses, you'd be making guesses till the end of time until you arrived upon that answer. So that's the real creativity that we're talking about. But admittedly, that's a creativity that very few humans engage in and they don't engage in it most of the time. It becomes harder and harder to see. So, we are probably going to get to where if you have a giant list of math problems to be solved and AI starts going through and picking, okay, this one out of that set of 1 million I can solve and this set out of 300,000 I can solve and I need a person to prompt me and ask the right questions. That's a very limited form of creativity. There's another form of creativity where it starts inventing entirely new scientific theories that then turn out to be true. I don't think we're anywhere near that, but I could be wrong. the AIS have been very surprising. So I don't want to get too much in the business of making prophecies and predictions. But I don't think that just throwing more compute at the current AI models short of some breakthrough invention is going to get us there. Just to be clear, when I say it's embedded, I don't mean the answers already written down in there. I just mean that it can be produced through a mechanistic process of turning the crank which is all today's computer programs are where the output is completely determined by the input. Epistemology now gets us into philosophy because isn't that just what human brains are doing? Aren't firing neurons just electricity and weights propagating through the system altering states and it's a mechanistic process. If you turn the crank on the human brain you would end up with the same answer. And some people like I think Penrose is out there saying no human brains are unique because of the quantum nanot tubes. You could argue that some of this computation is taking place at the physical cellular level not the neuron level and that's way more sophisticated anything we can do with computers today including with AI. Or you could just argue no we just don't have the right program. It is mechanistic. There is a crank to turn but we're not running the correct program. The way these AIs run today is just a completely wrong architecture and wrong program. I just buy more into the theory that there are some things they can do incredibly well and there's some things they do very poorly. And that's been true for all machines and all automation since the beginning of time. The wheel is much better than the foot at going in a straight line at high speeds and traveling on roads. The wheel is really bad for climbing a mountain. The same way I think these AIs are incredibly good at certain things and they're going to outperform humans. They're incredible tools. And then there are other places where they're just going to fall flat.
我同意这一点。而且我觉得这正是这些 AI 模型形成失控式规模经济的部分原因。你会为「智能」付费。一个 92% 的情况下都答对的模型，价值几乎无限地高于一个 88% 的情况下答对的模型，因为在现实世界里犯错的代价太高了，多花几块钱拿到正确答案是值得的。我会把我的查询写进一个模型，然后复制一下，一次性发射到四个模型里，然后让它们全部在后台运行。通常我甚至不会马上去看答案。我会过一会儿再回来看答案，然后看看，哪个模型给出的答案最好，我就开始用那个深入往下钻。在一些我自己也拿不准的少见情况下，我会让它们互相交叉盘问。那里头要剪切粘贴很多次。而且在很多情况下，我接下来会问追问性的问题，让它给我画示意图和插图。我发现，当概念以视觉方式呈现给我时，我非常容易吸收。我是一个非常视觉化的思考者。所以我会让它画草图、示意图和图画，几乎就像白板讨论一样。这样我才能真正理解它在讲什么。我们来聊聊 AI 的认识论吧，因为我觉得下一个大的误解是这样的：AI 已经开始解决一些尚未解决的基础数学问题了，那些问题人类如果用心去解大概也能解出来，只是还没被解出来，比如埃尔德什（Erdős）第几号问题。现在，我觉得人们正在、或者将会把这件事当作 AI 有创造力的一个标志。我不认为这是 AI 有创造力的标志。我其实认为这个问题的解早就嵌在 AI 的某个地方了，只是需要通过提示（prompting）把它引出来。这里头肯定有这个成分。然后问题就来了：什么是创造力？这是一个定义得如此糟糕的东西。如果你没法定义它，你就没法给它编程，而且你往往甚至认不出它。所以这就把我们带到了品味、或者说判断力的领域。我会说，今天的 AI 似乎并没有展现出那种人类偶尔才能独有地进行的创造力。我说的不是美术。人们倾向于把创造力和美术混为一谈，他们会说「哦，画作是有创造力的，而 AI 能画画」。可是 AI 没法创造出一个全新的绘画流派。AI 没法以一种真正新颖的方式用情感打动人类。所以从这个意义上说，我不认为 AI 有创造力。我不认为 AI 能想出我所说的「分布之外」（out of distribution）的东西。现在，你提到的埃尔德什问题的解，可能早就嵌在 AI 的训练数据集里、甚至嵌在它的算法范围之内了，但它大概是嵌在五个不同的地方、以三种不同的方式、用两种不同的语言、在七种不同的计算与数学范式里——然后 AI 算是把它们全都拼到了一起。那么，这算创造力吗？Steve Jobs 有句名言：「创造力不过就是把东西拼到一起。」我其实不认为这是对的。我认为创造力更多地处在这样一个领域：想出一个从问题本身、从那些已知的元素里既无法预测、也无法预见的答案。它远远超出了思维的边界。如果你只是用一台计算机、甚至用一个 AI 去搜索、去做猜测，你会一直猜到时间的尽头，才碰巧猜到那个答案。所以那才是我们所说的真正的创造力。但不得不承认，那是一种极少数人类才会进行的创造力，而且他们大多数时候也并不进行。它变得越来越难被看见。所以我们大概会走到这样一步：如果你有一个庞大的数学问题清单等着被解，AI 开始一个个过、一个个挑，「好，这一百万道里这一道我能解」「这三十万道里这一批我能解」「我需要一个人来给我提示、问我对的问题」。那是一种非常有限的创造力。还有另一种形式的创造力，就是它开始发明全新的科学理论，而那些理论后来被证明是真的。我不认为我们离那一步还有任何接近，但我也可能错了。这些 AI 一直都很出人意料。所以我不想太多地涉足做预言和预测这种事。但我不认为，在没有某种突破性发明的情况下，仅仅往当前这些 AI 模型上堆更多算力，就能把我们带到那一步。澄清一下，我说「嵌在里面」，并不是说答案已经写在那里头了。我只是说，它可以通过一个机械性的、转动曲柄的过程被产出来——而今天所有的计算机程序都是这样：输出完全由输入决定。认识论现在又把我们带进了哲学，因为人类大脑做的不也正是这件事吗？放电的神经元不就是电流和权重在系统里传播、改变状态吗，这就是一个机械性的过程啊。如果你转动人类大脑这个曲柄，你最后会得到同样的答案。有些人——我想 Penrose 就持这种观点——会说，不，人类大脑是独特的，因为有量子微管（quantum nanotubes）。你可以主张说，这其中有些计算是发生在物理的细胞层面、而不是神经元层面，而那比我们今天用计算机、包括用 AI 能做的任何事都要精密得多。或者你也可以干脆主张说，不，我们只是还没有那个对的程序。它是机械性的。是有一个曲柄可以转的，但我们没在运行那个正确的程序。这些 AI 今天的运行方式，就是一个彻头彻尾错误的架构、错误的程序。我个人更倾向于这样一种理论：有些事情它们能做得好得不可思议，有些事情它们做得非常糟糕。而从亘古以来，对所有机器、所有自动化来说都是如此。轮子在以高速沿直线行进、在道路上行驶这件事上，比脚强太多了。轮子在爬山这件事上则真的很糟糕。同样地，我觉得这些 AI 在某些事情上好得不可思议，它们会在那些事情上胜过人类。它们是不可思议的工具。然后在另一些地方，它们就会彻底栽跟头。
[49:14] Nivi
Steve Jobs famously said that a computer is a bicycle for the mind. It lets you travel much faster than walking. Certainly in terms of efficiency, but it takes the legs to turn the pedals in the first place. And so now maybe we have a motorcycle for the mind to stretch the analogy. But you still need someone to ride it, to drive it, to direct it, to hit the accelerator, and to hit the brake. We should probably find something to wrap things up on.
Steve Jobs 有句名言，说计算机是「头脑的自行车」。它让你能比走路快得多地前行。在效率上当然如此，但你首先得有腿去蹬踏板才行。所以现在，把这个类比再延伸一下，也许我们有了一辆「头脑的摩托车」。但你仍然需要有人去骑它、去驾驶它、去给它指方向、去踩油门、去踩刹车。我们大概该找点什么东西来收尾了。
[49:37] Naval
When new paradigms and new tool sets come out, there is a moment of enthusiasm and change. And this is true in society. And this is true as an individual. If you ride the moment of enthusiasm in society, that's exciting. And you can learn new things, you can make friends, and you can make money. But there's also a moment of enthusiasm in an individual. When you first encounter AI and you're curious about it and you're genuinely open-minded about it, I think that's the time to lean and learn about the thing itself, not just to use it, which of course everyone will, but to actually learn how it works. I think diving into and looking underneath the hood is really interesting. If you encounter a car for the first time in your life, yes, you can get in and drive it around, but that's the moment you're also going to be curious enough to open up the hood and look how it's structured and designed and figure it out. I would encourage people who are fascinated by the new technology to really get into the inards and figure it out. You don't have to figure out to the level where you can build it or repair it or create your own but to your own satisfaction because understanding what's underneath the abstraction, what's underneath that command line, it's going to do two things. One is it'll let you use it a lot better and when you're talking about the tool that has so much leverage, using it better is very helpful. Second is it'll also help you understand whether you should be scared of it or not. Is this thing really gonna metastasize into a Skynet and destroy the world? Are we going to be sitting here and Arnold Schwarzenegger shows up and says at 4:29 a.m. and February 24th is when Skynet became self-aware, right?
每当新的范式、新的工具集出现时，都会有一个充满热情和变化的时刻。这在社会层面上是如此，在个人层面上也是如此。如果你乘上社会上那股热情的浪潮，那会很令人兴奋。你可以学到新东西，可以交到朋友，也可以赚到钱。但个人身上也有一个充满热情的时刻。当你第一次接触 AI、对它感到好奇、对它真心抱着开放心态的时候，我觉得那正是该去钻研、去了解这个东西本身的时候——不只是去用它（当然每个人都会用它），而是真正去搞懂它是怎么运作的。我觉得一头扎进去、看看引擎盖底下，是非常有意思的。如果你这辈子第一次见到一辆车，没错，你可以坐进去、把它开来开去，但那也正是你会好奇到想打开引擎盖、看看它是怎么搭建和设计的、把它弄明白的时刻。我会鼓励那些被新技术迷住的人，真正钻进它的内部、把它搞明白。你不必搞明白到能自己造出它、修好它、或者创造出你自己那一套的程度，但要达到让你自己满意的程度——因为理解抽象层底下是什么、那个命令行底下是什么，会带来两个好处。一个是它会让你把它用得好得多，而当你谈论的是一个有如此巨大杠杆的工具时，把它用得更好是非常有帮助的。第二个是，它还会帮你搞清楚你到底该不该害怕它。这东西真的会恶化、扩散成一个天网（Skynet）、把世界毁掉吗？我们会不会就坐在这儿，然后 Arnold Schwarzenegger 冒出来说「2 月 24 日凌晨 4 点 29 分，就是天网产生自我意识的时刻」，对吧？
[51:11] Naval
Or is it more that hey this is a really cool machine and I can use to do A B and C but I can't use to do D E and F and this is where I should trust it and this is where I should be suspicious of it. I feel like a lot of people right now have AI anxiety and the anxiety comes from not knowing what the thing is or how it works having a very poor understanding. And so the solution to that anxiety is action. The solution to anxiety is always action. Anxiety is a non-specific fear that things are going to go poorly and your brain and body are telling you to do something about it, but you're not sure what. You should lean into it. You should figure the thing out. You should look at what it is. You should see how it works. And I think that'll help get rid of the anxiety. That action of learning, that pursuit of curiosity is going to help you get over the anxiety. And who knows, it might actually help you figure out something you want to do with it that is very productive and will make you happier and more successful.
还是说，更多的情况其实是这样：嘿，这是一台真的很酷的机器，我可以用它来做 A、B、C，但我没法用它来做 D、E、F，这里是我应该信任它的地方，那里是我应该对它保持怀疑的地方。我觉得现在很多人都有「AI 焦虑」，而这种焦虑来自于不知道这个东西是什么、它怎么运作，对它的理解非常贫乏。所以应对那种焦虑的解药就是行动。焦虑的解药永远是行动。焦虑是一种非特定的恐惧，担心事情会朝坏的方向发展，你的大脑和身体在告诉你要为此做点什么，但你不确定该做什么。你应该迎着它上。你应该把这个东西搞明白。你应该看看它到底是什么。你应该看看它是怎么运作的。我觉得那会帮你摆脱焦虑。那个学习的行动、那种对好奇心的追逐，会帮你战胜焦虑。而且谁知道呢，它说不定还真能帮你想出某件你想用它去做的事，一件非常有生产力、会让你更幸福、更成功的事。