Speaker 100:00 - 00:03
You can wave your hands and say, no. This is like electricity. Okay. Fine. It's like electricity. Speaker 100:00 - 00:03
你可以挥挥手说,不。这就像电。好吧,行吧。它就像电。
Speaker 100:03 - 00:09
Well, what happened with electricity? And there was a time before electricity, and people before electricity weren't dumb either. Do you Speaker 100:03 - 00:09
那电后来发生了什么?而且在有电之前也有一个时代,而生活在电力出现之前的人也并不蠢。你会不会觉得——
Speaker 200:09 - 00:12
think it's irrational that, like, the heads of these labs are talking so much about job loss? Speaker 200:09 - 00:12
——像这些实验室的负责人一直在大谈工作流失,这种事是不理性的?
Speaker 100:12 - 00:20
We get deep into logical fallacies here. So we've got Jeff Hinton, was ten years ago, saying stop training radiologists. And the problem here here is he doesn't actually understand what radiologists do. Speaker 100:12 - 00:20
我们在这里会深入谈到逻辑谬误。所以我们看到 Jeff Hinton 在十年前就说过,别再培养 radiologists(放射科医生)了。而这里的问题在于,他其实并不真正理解 radiologists 做什么。
Speaker 200:20 - 00:24
Given this discussion we've had on the foundation models, would you be doing anything different in terms of, like, focus? Speaker 200:20 - 00:24
基于我们刚才关于 foundation models(基础模型)的这番讨论,你会不会在关注重点之类的方面做一些不同的调整?
Speaker 100:26 - 00:28
I have a lot of sympathy for Sam Walton. Speaker 100:26 - 00:28
我对 Sam Walton 很有共鸣。
Speaker 200:28 - 00:46
I'm Jacob Efron from Redpoint, and this is Unsupervised Learning. A podcast where we probe the sharpest minds in AI about what's happening in the ecosystem today and where the future's headed. Today, we had an awesome conversation with Benedict Evans. He's one of my favorite thinkers in the space. I love reading his newsletters, presentations, and had the chance to sit down with him and talk through a bunch of things. Speaker 200:28 - 00:46
我是来自 Redpoint 的 Jacob Efron,这里是 Unsupervised Learning。一档 podcast(播客),我们会就当下生态系统里正在发生的事,以及未来将走向何方,向 AI 领域最敏锐的头脑发问。今天,我们和 Benedict Evans 进行了一场非常精彩的对话。他是这个领域里我最喜欢的思考者之一。我很喜欢读他的 newsletters、presentations,也有机会和他坐下来,把很多事情都聊了一遍。
Speaker 200:46 - 01:10
We talked about where value will accrue in AI, as well as what he thinks the ultimate valuations of the foundation models might look like. We talked about why consumer AI applications aren't really working today, and the kind of jagged edge of capabilities and their implications on enterprise adoption overall. And we hit on a lot of his really interesting takes on job usage, whether this is an enabling technology or not. Just awesome to get to talk with one of my favorite thinkers about a bunch of these things. I think folks will really enjoy the conversation. Speaker 200:46 - 01:10
我们聊了在 AI 领域里价值会沉淀到哪里,也聊了他认为 foundation models(基础模型)最终的估值大概会是什么样。我们聊了为什么面向消费者的 AI 应用今天其实还不太行,以及能力那种参差不齐的“锯齿状前沿”及其对企业整体采用的影响。我们还谈到了他很多非常有意思的看法,比如工作的使用方式,以及这到底是不是一种赋能型技术。总之,能和我最喜欢的思考者之一聊这么多话题,真的特别棒。我觉得大家会非常喜欢这场对话。
Speaker 201:10 - 01:23
Without further ado, here's Bennett. Well, very excited to do this. I feel like you're one of the most thoughtful people on tech. I love reading the presentations you do. You've obviously have a huge influence on the ecosystem at large. Speaker 201:10 - 01:23
闲话少说,下面请 Bennett。嗯,非常高兴来聊这个。我觉得你是科技领域里最有洞见的人之一。我很喜欢读你做的那些演示材料。显然,你对整个 ecosystem(生态系统)都产生了巨大的影响。
Speaker 201:23 - 01:50
And today, excited to dig into your opinions on the ecosystem as we have it today, the future of the model labs, where value accrues. But where I really wanna start with is the hype in the current moment, because I feel like you have maybe a more subdued take. It's funny to call a subdued take being that this is as big a deal as the internet and mobile, but maybe not civilization altering like the industrial revolution, but I wanna start there. What do you like paying attention to that might change your mind on that? Because I'm sure a lot of our listeners will disagree. Speaker 201:23 - 01:50
今天也很期待深入聊聊你对当下这个 ecosystem 的看法、model labs(模型实验室)的未来,以及价值会沉淀到哪里。但我真正想先切入的,是此刻这波 hype(热潮),因为我感觉你可能有一个相对更克制的判断。说“更克制”其实也挺有意思,因为你的意思仍然是:这件事的重要性至少和 internet、mobile 一样大,只是也许还没有到像 industrial revolution 那样改变整个人类文明的程度。我想先从这里开始。你现在最关注哪些事情,可能会让你改变这个判断?因为我相信很多听众都会不同意。
Speaker 101:50 - 02:23
I'm not sure how analytic you can get about saying, well, is this as big as mobile or as big as the internet or as big as computer Were mobile bigger than PCs? I struggle to kind of How you can kind of quantify those things? I think the place that I'd come from is just to say, look, the internet was kind of a big deal and did absolutely change everything. And so was mobile and so was PCs. And yes, I think you certainly kind of can say that the web was bigger than PCs because PCs was like 50 or a 100,000,000 devices. Speaker 101:50 - 02:23
我不太确定,关于“这到底和 mobile 一样大,还是和 internet 一样大,还是和 computer 一样大”,这种说法到底能分析到什么程度。mobile 比 PC 更大吗?我一直有点难以理解,这些东西到底该怎么量化。我的出发点大概是这样:你看,internet 本身就是件大事,而且确实彻底改变了一切;mobile 是,PC 也是。是的,我觉得你当然可以说 web 比 PC 更大,因为 PC 那时大概也就是 5000 万到 1 亿台设备。
Speaker 102:24 - 02:48
When Marc Andreessen launched Netscape, there were like 75,000,000, a 100,000,000 PCs in the world and now there's 5,000,000,000 smartphones. So clearly it's got bigger over time. I suppose the thing I sort of struggle with is that nothing like this has ever happened before. And we've not had technology wiping out great swathes of jobs before. And we're automating stuff in a way that this has never been done before. Speaker 102:24 - 02:48
当 Marc Andreessen 推出 Netscape 时,全球大概有 7500 万到 1 亿台 PC;而现在有 50 亿部 smartphone。很明显,这个规模是在随着时间不断变大的。我想我真正有点纠结的是:以前从来没有发生过完全一样的事情。我们过去也没有见过技术这样大面积地替代工作岗位。我们现在正在以一种前所未有的方式把事情自动化。
Speaker 102:48 - 03:08
And I kind of, I was talking to somebody the other day at a big company and they were, you know, very smart guy, PhD in all sorts of computer science y stuff. And they said, you know, big companies have never had to deal with people building tools and systems for themselves before. And I thought, so have you heard of Shadow IT? Have you heard of Excel? Yes, this is different, but it's always different. Speaker 102:48 - 03:08
前几天我和一家大公司的某个人聊过,他是个非常聪明的人,在各种偏 computer science 的领域都有 PhD。然后他说,大公司以前从来不需要面对“人们自己给自己搭工具和系统”这种事。我当时就在想:那你听说过 Shadow IT 吗?听说过 Excel 吗?没错,这次确实不一样,但每一次本来都不一样。
Speaker 103:09 - 03:34
And we do go through these big changes and they're always different and they're always bigger, but that doesn't mean you shouldn't kind of go back and look and say, well, what is it that happened the last five times we had a world changing piece of technology? I don't think there's much percentage in arguing about which one is bigger or smaller. I think what's interesting is to say, well, what happened the last five times we had something like this? And yes, you can wave your hands and say, no, this is like electricity. I'm like, okay, fine. Speaker 103:09 - 03:34
我们确实会经历这种巨大的变化,而且每次都不一样、每次都更大;但这并不意味着你就不该回过头去看一看:过去五次出现改变世界的技术时,到底发生了什么。我不觉得争论哪一个更大、哪一个更小有太大意义。更有意思的问题是:上一次、前几次,当我们遇到类似的东西时,发生了什么?当然,你也可以泛泛地说,不,这次像 electricity(电力)。我会说,好,可以。
Speaker 103:34 - 03:52
It's like electricity. Well, what happened with electricity? And there was a time before electricity and people before electricity weren't dumb either. And people were kind of scratching their heads and trying to work out what to do with this. And so it's just useful not to imagine that nothing ever has happened, never happened before, And to kind of sit and scratch your heads and think, well, what are the patterns that you can see here? Speaker 103:34 - 03:52
如果这像 electricity,那 electricity 当年又发生了什么?在 electricity 出现之前也有一个时代,而那个时代的人也并不傻。他们当时同样在挠头,试图搞清楚这东西到底该怎么用。所以,别把现在想象成一种“以前什么都没发生过”的局面,这一点是有帮助的。更有价值的是坐下来认真想一想:这里面到底有哪些 pattern(模式)是你能看出来的?
Speaker 103:52 - 04:09
I wrote something at the end of last week talking about token pricing. And one of the sort of blocks of the essay was, well, how did mobile work? How did semiconductors work? How did fiber work? How did operating systems work? Speaker 103:52 - 04:09
我上周末写了一点关于 token 定价的东西。那篇文章里有一个核心板块就是:mobile 当年是怎么运作的?semiconductors(半导体)是怎么运作的?fiber(光纤)是怎么运作的?operating systems(操作系统)又是怎么运作的?
Speaker 104:09 - 04:32
And none of those have predictive power. Like you can't prove that this is going to work in a certain way by arguing about how close it is to mobile. But you should go back and look at what happened at mobile and think, okay, so what does this tell us about what might happen? So you can point to semiconductors and say you had this escalating cost structure, Wach's law, that said that the cost of a cutting edge fab doubles every four years. Yeah. Speaker 104:09 - 04:32
而且这些都没有预测力。也就是说,你不能因为它和 mobile 很像,就证明这件事一定会以某种方式奏效。但你应该回头看看 mobile 当年发生了什么,然后想一想:好,这会告诉我们什么,关于接下来可能发生的事?所以你可以指向 semiconductors,说当时存在一种不断升级的成本结构,Wach's law 认为,建设一座最先进 fab 的成本每四年翻一倍。对。
Speaker 104:32 - 04:48
And that's and foundation models look a little bit like that. You can look at mobile and say, well, mobile had marginal cost. Like, adding more a lot of people don't take don't understand this. So mobile networks have marginal cost. You add have you have if you double the number of traffic, you have to build a bunch more base stations and put a bunch more equipment on the base stations. Speaker 104:32 - 04:48
而 foundation models 看起来有点像那样。你也可以看 mobile,说 mobile 是有 marginal cost(边际成本)的。再增加更多用户——很多人没有意识到这一点,也不理解——所以 mobile networks 是有 marginal cost 的。你如果把流量翻倍,就得再建一大堆 base stations,还要在这些 base stations 上加装一大堆设备。
Speaker 104:48 - 05:14
And so you have marginal cost. And that looks a lot like the supply crunch we've had in the last six months where everyone got an iPhone and started watching YouTube and the networks fell down and they had to change pricing schemes. But the interesting part of that comparison is that mobile data traffic in the last fifteen years has gone up by like one or 2,000 times. And it's a trillion dollar industry and they spend $200,000,000,000 a year on CapEx and they didn't make any money. I mean, they make a bit of money, but like all the value is other people. Speaker 104:48 - 05:14
所以你就有 marginal cost。这看起来很像我们过去六个月经历的 supply crunch(供给紧张):每个人都买了 iPhone,开始看 YouTube,结果网络崩了,他们不得不改变定价方案。但这个类比里有意思的地方在于,过去十五年里,mobile data traffic 大概增长了 1,000 到 2,000 倍。这是一个万亿美元级产业,他们每年花 200,000,000,000 美元做 CapEx(资本开支),但他们并没赚到什么钱。我的意思是,他们也赚了一点,但基本上所有价值都被别人拿走了。
Speaker 105:14 - 05:20
They didn't get to do Uber. They don't run your banking. They don't do YouTube. That's all other people. And all the value went up stack. Speaker 105:14 - 05:20
他们没法去做 Uber。他们不运营你的 banking。也不做 YouTube。那些全都是别人做的。所以所有价值都流向了更上层的 stack(技术栈)。
Speaker 105:20 - 05:34
And obviously, that's one of the questions about AI. You can point to cloud. You can point to operating systems. Sam Altman said, we're gonna be like Windows, but like LLMs don't have network effects. So it doesn't look like they're gonna turn out like Windows. Speaker 105:20 - 05:34
而这显然也是 AI 的问题之一。你可以指向 cloud,也可以指向 operating systems。Sam Altman 说,我们会像 Windows 一样;但 LLMs 并没有 network effects(网络效应),所以看起来它们最终未必会变成像 Windows 那样。
Speaker 105:34 - 05:45
But the point is to say, look, there's a bunch of different competitive dynamics, industry dynamics. What happens at different layers of the stack? Can you compete up? Can you compete down? And you can learn from all those. Speaker 105:34 - 05:45
但重点是要说,你看,这里面有很多不同的 competitive dynamics(竞争动态)、industry dynamics(产业动态)。在 stack 的不同层会发生什么?你能不能向上竞争?你能不能向下竞争?这些东西你都可以从中学到。
Speaker 205:45 - 06:08
Yeah. I think one thing that makes it challenging to figure out like the long term macro impact of all this is just, obviously the capabilities of these models are not static, but you pointed to some of the big outstanding questions in the space. It feels like the biggest is just, you know, what capabilities do we ultimately get to on the model side? I'm curious, like, what would change your mind that actually this is a bigger deal than that's something more akin to the industrial revolution? Is it a model capability? Speaker 205:45 - 06:08
对。我觉得,让人很难判断这一切长期 macro impact(宏观影响)的一点在于,很明显,这些模型的能力不是静态不变的;而你刚才也提到了这个领域里一些悬而未决的大问题。感觉最大的问题就是,最终在 model 这一侧,我们到底能获得什么样的能力?我很好奇,什么会改变你的看法,让你觉得这件事其实比你现在认为的更重要,更接近 industrial revolution(工业革命)那种级别?会是 model capability(模型能力)上的变化吗?
Speaker 106:08 - 06:33
Well, so I think the you could make a very obvious kind of counter argument, which is just to point out that each of these compound. Part of that is that's why it's happening quicker. But like mobile internet didn't need to wait for the internet. Right. And the internet didn't need to wait for PCs and PCs didn't need to wait for semiconductors and semiconductors didn't need to wait for, I don't know, specialty chemicals or whatever you think the enabling technology is, photolithography. Speaker 106:08 - 06:33
嗯,所以我认为,你可以提出一个非常直接的反驳观点:只要指出,这里面的每一层其实都是复合叠加的。其中一部分也解释了为什么这次发生得更快。但比如说,mobile internet 并不需要等 internet 先出现,对吧。而 internet 也不需要等 PCs,PCs 也不需要等 semiconductors,semiconductors 也不需要等——我不知道,随便你认为哪种是使能技术——specialty chemicals 之类的,或者 photolithography。
Speaker 106:34 - 06:48
And that didn't need to wait for electricity. And electricity had a whole manufacturing industry that already existed and so on. So there's always this kind of compounding effect. And, you know, I mean, another way of saying this is big as the Internet is to say, well, the Internet was 10 times bigger than PCs. Right. Speaker 106:34 - 06:48
而且那甚至不需要等到 electricity(电力)普及。并且 electricity 还有一个早已存在的完整制造业作为基础,等等。所以这里总会有一种复利式的叠加效应。换句话说,形容它有多大的一种说法是:Internet 已经很大了,但 Internet 比 PC 大 10 倍。对吧。
Speaker 106:48 - 07:30
So maybe this is 10 times bigger than the internet. The big difference, the actual tangible difference to previous platform shifts is that we don't know the physical limits of this. So we do not have a good scientific understanding of why the models work so well and what will happen next, which was, whereas with, like you go back to 2010, you didn't know what the next iPhone would be, but you knew it wouldn't have like a retinal implant and a one year battery life and flight and cost $5 And the internet, like 1995, you didn't know, you had no idea how the internet was gonna work. You probably thought you did, but you didn't. But you knew that PCs are like 2,000 and $3,000 and most people don't have one and telcos aren't gonna give every household on earth fiber to the home next month. Speaker 106:48 - 07:30
所以也许这会比 Internet 大 10 倍。与以往平台迁移相比,真正巨大而且可感知的差别在于,我们不知道它的物理极限在哪里。也就是说,我们并没有很好的科学理解,来解释为什么这些 model(模型)表现得这么好,以及接下来会发生什么。相比之下,比如回到 2010 年,你不知道下一代 iPhone 会是什么样,但你知道它不可能带什么 retinal implant(视网膜植入)、一年续航、还能飞、而且只卖 5 美元。再比如 Internet,在 1995 年时,你并不知道 Internet 最终会怎么运作。你大概以为自己知道,但其实并不知道。不过你知道 PC 要 2,000 到 3,000 美元,而且大多数人并没有一台;你也知道 telcos(电信运营商)不可能下个月就给地球上每个家庭都铺上 fiber to the home(光纤到户)。
Speaker 107:30 - 07:46
Yeah. Like you kind of knew the basic kind of physical parameters. And so that's clearly a kind of a conceptual difference. But I think then you can kind of say, okay, we've got this stuff we don't know, fine. Then you can kind of break that apart. Speaker 107:30 - 07:46
对。你大致知道那些最基本的物理参数。所以这显然是一种概念层面的差异。但我想接下来也可以说,好,我们承认这里有一些自己不知道的东西,没问题。然后你就可以把这个问题拆开来看。
Speaker 107:46 - 08:32
So maybe there's one kind of binary question, which is, there's an urban legend, may or may not be true, that during the Cuban Missile Crisis, a rumor starts that the missiles have launched and everyone on stock exchange starts selling and one veteran trader goes out and starts buying. And he says, well, it's binary. Either the rumor is true, in which case we're all dead anyway, or it's not, in which case the stocks are cheap. And if this stuff, like if AGI really does happen, an ASR, whatever, pick your acronym and your thought concept, if all of that happens, then we've got bigger problems than worrying about middle class unemployment, like rural pets, fine. Otherwise, then, okay, let's worry about let's try and work out what this means for enterprise software. Speaker 107:46 - 08:32
所以也许这里有一个二元问题。流传着一个都市传说,未必是真的:在 Cuban Missile Crisis(古巴导弹危机)期间,有谣言说导弹已经发射了,于是证券交易所里的所有人都开始抛售,而一位老交易员却走出去开始买入。他说,这件事是二元的:要么谣言是真的,那我们反正都要死了;要么它不是真的,那股票现在就很便宜。如果这些事真的发生了——如果 AGI 真会出现,或者 ASR,不管你选哪个 acronym(缩写)和对应的概念——如果这一切都发生了,那我们要面对的问题可比担心中产阶级失业大得多,像“乡村宠物”这种事,随它去吧。否则的话,那好,我们再来担心,再来试着搞清楚这对 enterprise software(企业软件)意味着什么。
Speaker 108:32 - 08:35
Yeah. Those are kind of those I love the comparison. Speaker 108:32 - 08:35
对,就是这类问题。我很喜欢这个类比。
Speaker 208:35 - 08:51
It's like you're not even Yeah. It's like it's kind of unknowable right now what the shape of model progress looks like. There's some all the shape that we're worrying about and the implications it would have for everything else, there's so much bigger fish to fry if that's the case. That it's almost like you wouldn't you probably wouldn't be writing about enterprise software in the in the context of which there's Speaker 208:35 - 08:51
这就像,你甚至都还没法——对。现在其实根本无法知道 model(模型)进展的形状会是什么样。我们所担心的那种走势,以及它会对其他一切带来的影响,如果真是那样的话,还有大得多的问题要先处理。以至于几乎可以说,在那种语境下,你大概都不会去写 enterprise software,因为那时存在的是一种
Speaker 108:51 - 08:59
a rapid Exactly. It's like yeah, exactly. Either you can worry about that stuff. A lot of these conversations end up in hunt for metaphors. It's like nuclear weapons. Speaker 108:51 - 08:59
快速——没错。对,就是这样。要么你还能去担心那些事。很多这类讨论最后都会变成一场寻找隐喻的过程。比如说,它像 nuclear weapons(核武器)。
Speaker 108:59 - 09:14
It's like this, it's like that. A lot of these conversations also end up sounding like the stuff you said at 02:00 in the morning if you're a slightly drunk philosophy grad student. It's like, hey, man, have you thought that maybe we don't have structural understanding either? Well, great. Thank you. Speaker 108:59 - 09:14
像这个,像那个。很多这类讨论还会听起来像是一个微醺的哲学研究生凌晨两点说的话。就像,“嘿,哥们,你有没有想过,也许我们同样并没有结构性的理解?” 好吧,太好了。谢谢你。
Speaker 109:15 - 09:23
I can't do anything with that. I can't analyze that. This is like the Descartes fallacy. Descartes thought that you could deduce the nature of God from first principles. No, you can't. Speaker 109:15 - 09:23
我没法拿那个做任何事。我没法分析那个。这有点像 Descartes fallacy。Descartes 以为你可以从第一性原理推导出 God 的本质。不,你做不到。
Speaker 109:23 - 09:35
We don't know. We'll find Meanwhile, let's build some stuff and work out what works. I mean, you get back to and this is a little bit of an echo of this. You go back to you ever read the the John Perry Barlow declaration of independence of cyberspace? Speaker 109:23 - 09:35
我们不知道。我们之后会弄清楚。与此同时,先做点东西出来,再看看什么真的行。我的意思是,这又有一点呼应前面的话。你会回想到——你读过 John Perry Barlow 的《Declaration of Independence of Cyberspace》吗?
Speaker 209:35 - 09:36
No, I haven't. Speaker 209:35 - 09:36
没有,我没读过。
Speaker 109:36 - 09:55
God, you're so young. So John Perry Barlow is like one of the lyricists of The Grateful Dead, I think. Sorry, I'm too young to really know this. But anyway, so he wrote this beautifully poetic thing and it's like, you giants of steel and iron and stone, you have no sovereignty on the internet. And people kind of thought this was gonna end war. Speaker 109:36 - 09:55
天啊,你也太年轻了。John Perry Barlow 好像是 The Grateful Dead 的词作者之一,我想是。抱歉,其实我也太年轻了,不敢说自己真的了解这个。不过总之,他写过一篇非常优美、很有诗意的文字,大意是:你们这些钢铁、铁器与石头的巨人,在 internet 上并没有主权。那时候有些人真的觉得这东西会终结战争。
Speaker 109:55 - 10:08
Yeah. Because everyone understand each other now. And so you have to remember, there was an awful lot of this kind of wild eyed millenarian utopianism around the internet too. Some of which is true, some of which wasn't. Meanwhile, let's build some software and work out what people do with it. Speaker 109:55 - 10:08
对。因为现在大家都能互相理解了。所以你得记住,围绕 internet 也曾经有过大量这种目光狂热、带有千禧年式乌托邦色彩的想象。其中有些是真的,有些并不是。与此同时,还是先做些 software,再看看人们到底会拿它做什么。
Speaker 110:09 - 10:44
And that gets you like, so you set that whole argument aside and say, right, OpenAir kind of looks like Netscape. Chatbots, is this a web browser? Like, and again, you don't wanna be too rigid on the metaphor, but you have this stuff that's amazing and it doesn't work. And I think this is kind of a more interesting way of thinking about where we are now that go back to like trying to use the internet in the mid nineties or trying to do mobile internet, nevermind 2010, 2002. And it's kind of sort of works and it's obviously really cool and amazing and it's gonna change everything. Speaker 110:09 - 10:44
这样一来,你就会先把那整套争论放到一边,然后说,好,OpenAir 看起来有点像 Netscape。Chatbots,这是 web browser 吗?当然,你也不想把这个比喻卡得太死,但你眼前有一种东西,它很惊人,却又不好用。我觉得,用这种方式来理解我们现在所处的位置更有意思:就像回到 mid nineties 试着用 internet,或者别说 2010 了,回到 2002 去做 mobile internet。它算是有点能用,又显然非常酷、非常惊人,而且看起来会改变一切。
Speaker 110:44 - 11:07
And I have this very vivid memory of sitting in a meeting with a mobile software company in like 1999. And I had my Palm five and I had a Nokia phone that had GPRS and an infrared port. And so we were trying to connect my Palm five to the internet over GPRS, over infrared to this Nokia. And we're sitting there going, oh my God, this is so cool. And like, but nobody, it didn't actually work. Speaker 110:44 - 11:07
我对一个场景记得特别清楚:大概在 1999 年,我坐在一家 mobile software company 的会议室里。我手上有一台 Palm five,还有一部带 GPRS 和 infrared port 的 Nokia phone。于是我们在试着通过 infrared,把我的 Palm five 通过这台 Nokia 的 GPRS 连到 internet 上。我们当时坐在那里想,天啊,这也太酷了。可是实际上,根本没人能真正用起来,它也并没有真的行得通。
Speaker 111:07 - 11:34
And no one is actually gonna do that for another ten years. Obviously this isn't another ten years, but it is that sort of moment of like, what are the building blocks here? Where's the value capture? What are the points of leverage? What are the experiences that people are actually going to use to get billions of people to use this every day, as opposed to 10,000,000 people using it all day and a billion people using it once twice a month, which is kind of where we are now. Speaker 111:07 - 11:34
而且实际上,接下来十年里也不会有人真的这么做。显然,现在这个东西不需要再等十年,但它确实处在那种时刻:基础 building blocks 到底是什么?价值会被哪里捕获?杠杆点在哪里?什么样的体验才是人们真的会去用、并且能让几十亿人每天都使用它的?而不是像现在这样,1000 万人整天都在用,同时 10 亿人一个月只用一两次。
Speaker 111:34 - 11:51
I'm exaggerating a bit. We've got like, kind of like ten, fifteen percent of people are daily active users right now. And even daily active is you're you're using this once or twice a day. You're not using it all the time. And then there's another 20 or 30%, 40% of people who are weekly active users or monthly active users. Speaker 111:34 - 11:51
我稍微有点夸张了。现在大概有差不多 10% 到 15% 的人是 daily active users(日活用户)。而且所谓 daily active,其实也只是你一天用一两次。你不是一直都在用它。然后还有另外 20%、30%,甚至 40% 的人,是 weekly active users(周活用户)或者 monthly active users(月活用户)。
Speaker 111:51 - 12:08
And you remember from social, weekly active user is bullshit. So if you're using this once or twice a week, you're clearly not in the dream of Silicon Valley that this changes computing. Yeah. It's a useful new tool sometimes. And how do we bridge that? Speaker 111:51 - 12:08
你也记得,在 social(社交产品)领域里,weekly active user 这种指标是胡扯。所以如果你一周只用这个一两次,那显然离 Silicon Valley 的那个梦想——它将改变 computing(计算)——还差得远。对,它有时是个有用的新工具。那我们怎么跨过这道坎?
Speaker 112:08 - 12:20
And I think this is kind of a more interesting split. Like you've got software development where this has product market fit unquestionably. Yeah, poof, it works. Changes everything in software development, fine. And then we can talk about that. Speaker 112:08 - 12:20
我觉得这里有一个更有意思的分野。比如说在 software development(软件开发)里,这东西毫无疑问已经有了 product-market fit(产品市场契合)。对,啪,一下就成了。它改变了软件开发中的一切,没问题。这个我们可以展开说。
Speaker 112:20 - 12:43
Then you've got like a bunch of people who are kind of like our side, they're knowledge workers who work for themselves, are flexible, free firm, got lots of stuff going on and you're using this a lot. And I think this reminds me quite a lot of people who use Nation. And then you've got all these other people who are like, yeah, it's kind of useful. Yeah, that was cool. But aren't using it all the time. Speaker 112:20 - 12:43
然后还有一批人,有点像我们这边的人,他们是为自己工作的 knowledge workers(知识工作者),比较灵活,不受大公司体系约束,手上同时有很多事情在推进,所以你会大量使用这个。我觉得这很像那些用 Nation 的人。然后还有另外一大批人,他们会说,嗯,挺有用的。嗯,那个挺酷的。但他们并不会一直使用它。
Speaker 112:44 - 12:54
And how does that change? Yes. The analogy I always use here is imagine you're an accountant seeing the first software spreadsheets in the late seventies. This is life changing. Now imagine you're lawyer seeing it. Speaker 112:44 - 12:54
那这种情况会怎么改变?对。我在这里最常用的类比是:想象一下,你是一个 accountant,在 70 年代末第一次看到软件 spreadsheet(电子表格)。这会彻底改变你的人生。现在再想象一下,你是个 lawyer 看到它。
Speaker 112:54 - 13:14
Okay, that's cool. Yeah. I could use that for my timesheet next week, but that's not what I do every day. And I think that's kind of the split and that's, we were talking before we started recording about like being in the Silicon Valley bubble. If you were in the Valley bubble and you've got five Mac minis in a cluster and you're leaving, you're walking around all day, holding your laptop open. Speaker 112:54 - 13:14
你会说,哦,挺酷。对,我下周可以拿这个来做 timesheet(工时表),但这不是我每天真正做的事。我觉得这大概就是那个分野。还有,这也就是我们开始录之前提到的,关于身处 Silicon Valley bubble(硅谷泡泡)这件事。如果你在 Valley bubble 里,弄了五台 Mac minis 组成一个 cluster(集群),然后你整天走来走去,笔记本电脑一直开着,
Speaker 113:14 - 13:36
So Claude Cobb can carry on running. You're not the normal user here. The normal user, you are the guy using Telnet to connect to the internet, connecting to Telnet over to using Telnet on the internet in 1992, or you're using dial up on Netscape one in 1994. Was it Mosaic even before that in 1994? You're not getting the real experience. Speaker 113:14 - 13:36
好让 Claude Cobb 可以持续运行,那你就不是这里的普通用户。普通用户,更像是 1992 年那个用 Telnet 去连接互联网、在互联网上再用 Telnet 的人,或者是 1994 年用 dial-up(拨号上网)配合 Netscape one 的人。还是更早一点,1994 年用的是 Mosaic?你并没有获得真正的体验。
Speaker 213:37 - 14:02
Probably the hardest part of the whole thing is figuring out like just I feel like that makes a ton of sense for current model capabilities, and it's like you probably could have looked two years ago or a year ago, and you would have been like, God, this stuff has product market fit in some product market fit in consumer, but doesn't really work anywhere else. And then the coding stuff starts to work. I don't know what you make of the fact that the people at the labs who are closest to the research all seem super bullish on kind of continued capability improvement that might even start to change some of this. Speaker 213:37 - 14:02
整件事里可能最难的一部分,就是去搞清楚——我只是觉得,这对于当前 model(模型)的能力来说非常说得通。你大概可以回看两年前或者一年前,当时你可能会说,天哪,这些东西在某些 consumer(消费级)场景里有一些 product-market fit,但在别的地方基本还不太行。然后 coding(编程)这块开始奏效了。至于那些离研究最近、就在 labs(实验室)里的人,似乎都对能力持续提升这件事超级看多,甚至觉得它可能会开始改变其中一些局面——这个事实你怎么看,我也不太确定。
Speaker 114:03 - 14:40
So I think one of the questions here, and this is like a UX interface deployment kind of question is, how many of the problems that we might want to automate would you struggle to explain to an actual person? Let me put this another way, presuming the model capability, presuming here's this thing you want to do with global OTT, and it kind of can't do it, or you can't work out how to tell it to do it. Would that work if you had a person at the other end? How hard would it be for you to explain what you want? What would that be like? Speaker 114:03 - 14:40
所以我觉得这里有一个问题,这有点像 UX 界面部署一类的问题:对于那些我们可能想自动化的问题,其中有多少其实是你连向一个真实的人解释都会觉得很费劲的?我换个说法,先假设 model 能力没问题,假设你有这么一件想用 global OTT 去做的事,但它就是做不了,或者你想不出该怎么告诉它去做。那么,如果另一端是个人,这件事会行得通吗?你要把自己想要什么解释清楚,到底会有多难?那会是什么感觉?
Speaker 114:40 - 15:08
Would that be the right way of doing it? I mean, I'm like, I'm not neither the lawyer nor the accountant. I'm like a weird outlier in that I work for myself and I don't have anybody working for me. And partly because of that, I don't have lots of repetitive tasks because I've structured what I do so that I don't have lots of repetitive tasks because I can't do them because I don't have a bunch of interns. So now I get Claude and Copilot and ChatGPT, now I've got this thing, but I don't actually have the tasks to give to it. Speaker 114:40 - 15:08
那会是正确的做法吗?我的意思是,我既不是 lawyer,也不是 accountant。我有点像个奇怪的 outlier,因为我是为自己工作,而且手下没有任何人。也部分正因为如此,我并没有很多重复性任务,因为我把自己的工作方式设计成没有那么多重复性任务——因为我本来也做不了那些事,我又没有一堆 interns。于是现在我有了 Claude、Copilot 和 ChatGPT,我现在有了这个东西,但我实际上并没有任务可以交给它。
Speaker 115:08 - 15:32
What I'm kind of getting at is, is there is the model capability and then there is your ability to isolate and describe the things that you could give to it. Yeah. And that's a different problem. This is, I think it's particularly a problem with voice and chat. But it's quite hard to work out that you have that problem and to work out that you could give that problem to the system. Speaker 115:08 - 15:32
我真正想说的是,这里面一方面有 model capability,另一方面有你自己识别并描述那些可以交给它处理的事情的能力。对,这是另一个不同的问题。我觉得这在 voice 和 chat 里尤其是个问题。但更难的是,你甚至很难意识到自己有这样一个问题,也很难意识到这个问题其实是可以交给系统处理的。
Speaker 115:32 - 16:00
With every new technology, it's harder and it takes longer to make the new things that you could only do with the new thing, as opposed to automate doing more of the old thing. Like with every new technology, you start by taking the thing you're already doing and doing it more and faster with old technology, with the new technology and forcing the new technology to fit the old way of doing it. And it takes longer to think of the completely new thing. Yeah. That's only possible because of this. Speaker 115:32 - 16:00
每一种新技术都会这样:要想出那些只有借助这种新东西才能做到的新事情,会更难,也需要更久;相比之下,把旧事情自动化、做得更多,反而容易得多。就像每一种新技术刚出现时,你总是先把自己已经在做的事情,用新技术做得更多、更快,同时硬把新技术塞进旧的做事方式里。至于去想出那种完全新的东西,就需要更长时间。对,那种东西只有因为这个新技术才成为可能。
Speaker 116:00 - 16:30
And generally the completely new thing isn't automating a whole bunch of stuff you're already doing, it's something else. One of the sort of analysis I did in the presentation I published in May now, which is feels like a lifetime ago. Dog years now. Dog years, exactly, AI years, was number of people working as accountants in The USA in the course of the twentieth century, it's a straight line up into the right. Even as the tech industry progressively automates over and over again, like punch cards and adding machines and mainframes and all the stuff that the tech industry has done, number of accountants keeps going up. Speaker 116:00 - 16:30
而且通常来说,那种完全新的东西,并不是把你已经在做的一大堆事情自动化,而是别的东西。我在今年 5 月发布的那个演讲里做过一个分析——现在回头看简直像上辈子的事。Dog years now。Dog years,没错,AI years——分析的是 20 世纪里 The USA 的 accountant 从业人数,基本就是一条稳定向右上方增长的直线。即使科技行业一次又一次持续自动化,比如 punch cards、adding machines、mainframes,以及科技行业搞出来的所有这些东西,accountant 的人数还是一直在增加。
Speaker 116:31 - 16:58
And there's two things you can say about this. One of them is is now this cliche, the Jevons paradox that everyone wants to look up on Wikipedia last year, which basically is price elasticity. Like if you make it cheaper to do stuff, do you do the same work for less money or do you do more work for the same money or more work for more money because you might have a different ROI? But I don't think that's what that is. I I don't think that if you are, like, you know, five years in at PwC now, you're doing exactly what you'd have been doing if you'd been five years in at PwC in 1970, but more. Speaker 116:31 - 16:58
对这件事你可以有两种说法。其中一种就是现在这个已经成了陈词滥调的 Jevons paradox,去年大家都跑去 Wikipedia 查这个词,本质上说的其实是价格弹性。比如说,如果你把做事的成本变低了,那你是会用更少的钱做同样的工作,还是用同样的钱做更多工作,或者甚至花更多钱做更多工作,因为你的 ROI 可能变了?但我不觉得这里是这么回事。我不认为,如果你现在在 PwC 工作了五年,你做的事情就是 1970 年在 PwC 工作五年的人当时做的那些事情,只不过做得更多。
Speaker 116:58 - 17:07
It wasn't called PwC there anyway. Pricewaterhouse, something else. Anyway, you're not doing what you were doing then, but more. You're doing the stuff that you were doing then is now ten seconds once a week. Yeah. Speaker 116:58 - 17:07
反正那时候也不叫 PwC。是 Pricewaterhouse,或者别的什么。总之,你现在做的并不是当年那些事的“更多版”。你当年会做的那些事情,现在已经变成每周一次、十秒钟就做完的东西了。对。
Speaker 117:07 - 17:20
But this is his problem. It's just making the model better. Fine. Okay. Even if you had a model that actually was AGI or whatever AGI means, and now people are trying to redefine AGI stuff that was working two years ago, but fine. Speaker 117:07 - 17:20
但这就是他的问题。只是把 model 做得更好。行,好。即便你真有一个实际上就是 AGI 的 model,或者不管 AGI 到底是什么意思——现在人们还在试图把 AGI 重新定义成那些两年前就已经能做到的东西,不过好吧。
Speaker 117:20 - 17:32
Presume you have a model that is actually a person. What does that mean? Where does that fit? What do you do with that? How does that fit into an organization? Speaker 117:20 - 17:32
假设你有一个 model(模型),而它实际上就像一个人。这意味着什么?它该被放在什么位置?你会拿它做什么?它又该如何融入一个组织?
Speaker 117:32 - 17:38
And it's not necessarily as obvious and easy to answer that as it seems. Speaker 117:32 - 17:38
而且,这个问题未必像看上去那么明显、那么容易回答。
Speaker 217:38 - 18:04
If you take accounting or consulting or something, and you think about, I think maybe a counterargument to this would be you had tools in the past that could do one specific part of a large set of skills that someone might do, and cool. So, people found other skills that they could go do. You could imagine, and certainly it seems like pretty models may be decently close to being as good at a set of skills as your average college grad would be for who has never worked in a consulting firm. We'll take that, that was my first job out of Speaker 217:38 - 18:04
如果你拿 accounting(会计)或 consulting(咨询)之类的工作来举例,你会想到——我觉得对此的一种反驳可能是——过去你也有一些工具,能够完成一个人所具备的大量技能中的某一个特定部分,这当然很好。于是,人们就去寻找自己还能做的其他技能。你可以想象,而且看起来相当有可能,当前这些相当不错的 models(模型),也许已经相当接近于:在一组技能上,达到一个普通 college grad(大学毕业生)的水平——这个人从来没在 consulting firm(咨询公司)工作过。这个例子我就拿来用了,毕竟那是我大学毕业后的第一份工作的一部分——
Speaker 118:04 - 18:04
college. Speaker 118:04 - 18:04
大学。
Speaker 218:04 - 18:13
So, I think the concern would be anything that might be reinvented into that a 22 year old with no experience in that job might do, the model could just do out of the box. Speaker 218:04 - 18:13
所以,我认为让人担心的是,任何那种可以被重新定义为“一个 22 岁、在这份工作上毫无经验的人也能做”的事情,model(模型)都可能开箱即用地直接完成。
Speaker 118:13 - 18:26
So, the way to frame this is we've been automating higher and higher level human functions. So, we started in the nineteenth century with human beings as beasts of burden. We automate legs. And then we automate arms. And then we automate fingers. Speaker 118:13 - 18:26
所以,理解这个问题的一种方式是:我们一直在自动化越来越高层级的人类功能。19 世纪时,我们把人类当作 beasts of burden(负重劳力)来看待。于是我们先自动化了腿。然后自动化了手臂。再然后自动化了手指。
Speaker 218:27 - 18:29
All we have left is our brains, but it feels like, Speaker 218:27 - 18:29
我们剩下的似乎就只有大脑了,但是感觉上——
Speaker 118:29 - 19:01
But like you've progressively automated everything all the way up to the top of the stack. I think that's kind of an unfalsifiable statement. And I think the problem is you could have said that in 1950 because you kind of didn't, all this other stuff happened that you couldn't see then that it turned out you couldn't automate. And of course we don't know where this stuff is going to go and what it will be able to automate and won't automate. So as I said, there's an unfalsifiable statement that says, the only way you can falsify it by coming back and waiting for five years and seeing what happens. Speaker 118:29 - 19:01
但就像是,你已经一路把所有东西都逐步自动化,一直自动化到了 stack(栈)的最顶层。我觉得这某种程度上是一个 unfalsifiable statement(不可证伪的说法)。问题在于,其实你在 1950 年也可以这么说,因为当时你并不知道——后来又发生了很多你那时看不见的事情,而结果证明,那些东西你其实没法自动化。当然,我们也不知道这些技术接下来会走向哪里,哪些东西它将能够自动化,哪些又不能自动化。所以就像我说的,这是一个不可证伪的说法;你唯一能够证伪它的方式,就是五年后再回来,看看实际发生了什么。
Speaker 119:01 - 19:14
But it may be, yes, that we get a thing that actually can do anything that people can do by definition. Well, clearly then everything changes. But right now we don't have that. We have something that's very jagged. Yeah. Speaker 119:01 - 19:14
但也有可能,没错,我们会得到一种按定义来说实际上能做人类所能做的任何事情的东西。那么,很显然,到那时一切都会改变。但现在我们还没有。我们现在拥有的是一种非常参差不齐、能力很不均匀的东西。对。
Speaker 119:14 - 19:32
And we don't know whether it would go all the way to that. And we certainly can't just presume, I mean, is the thing that drives me crazy about this chart of the AI can do something that takes people seventeen hours. Eight meter chart. Twenty four hours. But there's also stuff that takes people five minutes that it can't do. Speaker 119:14 - 19:32
而且我们不知道它是否真的会一路发展到那一步。并且我们当然不能想当然,我是说,最让我抓狂的就是这类图表:AI 能做某件事,而这件事要人花十七个小时。那张八米长的图表。二十四小时。但与此同时,也有一些只要人花五分钟的事,它却做不了。
Speaker 119:32 - 19:46
Yeah. It's not linear like that. Intelligence isn't linear like that. So, it may be that yes, this goes all the way to do stuff that anybody could possibly do. But at the moment it can't. Speaker 119:32 - 19:46
对。它不是那种线性的关系。intelligence(智能)不是那样线性增长的。所以,确实有可能,它最终会发展到能做任何人可能做的事。但目前它还做不到。
Speaker 119:46 - 20:12
And we don't know when or if it will be able to do that. Go back and look at all the other times that you automated absolutely everything that your 25 year old was doing, and what is it that they're doing now and how did that change? I mean, this is like, you know, going back to the nineteenth century, this is a lump of labor fallacy, that you can always see the jobs that will go away, and you don't know the new jobs. Yeah. And the new jobs will be doing things with completely different kinds of skills that it didn't occur to you existed or that you needed or wanted. Speaker 119:46 - 20:12
而且我们不知道它何时会、或者是否会具备那种能力。你回头看看过去那些你把一个 25 岁年轻人所做的事情几乎全部自动化掉的时期,再看看他们现在在做什么,这中间又是如何变化的?我的意思是,这有点像,你知道,追溯到十九世纪,这是一种 lump of labor fallacy(劳动总量谬误):你总是能看见哪些工作会消失,却不知道会出现哪些新工作。对。而这些新工作会涉及完全不同类型的技能——那些技能你此前甚至没想到它们存在,或者没想到你会需要、会想要它们。
Speaker 120:13 - 20:14
These are all unfalsifiable statements. Speaker 120:13 - 20:14
这些说法全都是不可证伪的。
Speaker 220:14 - 20:29
Yeah. But I guess given all that, do you think it's irrational that the heads of these labs are talking so much about job loss? It seems like it's a probability distribution of kind of an unknowable of how much better these models are going to get. And there's some part of the probability distribution where this stuff does really matter, right? Speaker 220:14 - 20:29
对。不过话说回来,考虑到这一切,你觉得这些实验室的负责人如此频繁地谈论 job loss(失业、岗位流失)是不理性的吗?看起来这像是一个关于这些模型究竟会变得多好的、某种难以知晓之事的 probability distribution(概率分布)。而在这个概率分布中的某些部分里,这些事情确实非常重要,对吧?
Speaker 120:29 - 20:44
Yeah. I mean, we get deep into logical fallacies here. So, we've got Jeff Hinton, whatever it was ten years ago saying, stop radio, training radiologists. And the problem here is he doesn't actually understand what radiologists do. So the narrow problem is actually machine learning couldn't do what he thought it could do. Speaker 120:29 - 20:44
对。我是说,这里我们会深入到 logical fallacies(逻辑谬误)里。比如 Jeff Hinton,大概十年前吧,说过“别再培养 radiologists(放射科医生)了”。这里的问题在于,他其实并不真正理解 radiologists 是做什么的。所以狭义上的问题是,machine learning(机器学习)实际上做不到他以为它能做到的事。
Speaker 120:44 - 20:48
And the broader problem is, but that isn't what radiologists did anyway. Speaker 120:44 - 20:48
而更广泛的问题是,反正那本来也不是 radiologists 真正在做的工作。
Speaker 220:48 - 20:49
And Speaker 220:48 - 20:49
还有,
Speaker 120:51 - 21:13
you started in consulting, the client generally isn't buying a PowerPoint. The PowerPoint might be the tangible expression of the project, but that isn't what they're paying you for. They're buying all sorts of other stuff. And the same thing as an accountant, or the same thing as a lawyer. The thing I wonder here is about consumer surplus. Speaker 120:51 - 21:13
你最初做咨询时,客户通常买的并不是一份 PowerPoint。PowerPoint 也许是这个项目有形的表达,但那并不是他们付钱给你的真正原因。他们买的是各种别的东西。做 accountant 也是一样,做 lawyer 也是一样。我在这里想问的是 consumer surplus(消费者剩余)。
Speaker 121:13 - 21:27
So, whatever the enterprise version of consumer surplus is. So for the sake of argument, like in 1980, your lawyer goes and sends their paralegal to the basement, into the library. Yep. And they come back after a day with two precedents. Great. Speaker 121:13 - 21:27
所以,姑且说是 consumer surplus 在企业场景里的对应物吧。为了便于讨论,比如在 1980 年,你的 lawyer 会让他们的 paralegal 去地下室的图书馆。对。然后对方花一天时间回来,带回两个 precedent。很好。
Speaker 121:27 - 21:44
Today, five years ago, your accountant, your associate would have gone to the database and they'd have found 20 presidents. And so did the opposition. Right. And so the client gets billed the same amount of money for the same number of hours. You deliver the same number of pages in the filing. Speaker 121:27 - 21:44
到了今天,或者说五年前,你的 accountant、你的 associate 会去查数据库,然后找到 20 个 precedent。对手也一样。对吧。于是客户为同样的工时被收取同样多的钱。你在 filing 里交付的页数也还是一样多。
Speaker 121:44 - 22:03
Your filing has exactly the same likelihood of winning that it did then. You get kind of the same result you would have got then. Consumer surplus. Like you've done kind of a bunch more stuff for the same price to deliver the same result, which is, this is the extreme, this is the iPhone. The iPhone is $10,000, $100,000 of consumer electronics disappeared. Speaker 121:44 - 22:03
你的 filing 获胜的概率,和当年几乎完全一样。你得到的结果,大体上也和当年差不多。所谓 consumer surplus。也就是说,你为了交付同样的结果,实际上做了多得多的事情,但价格没变。这是一个极端例子,这就像 iPhone。iPhone 一出现,相当于价值 $10,000、$100,000 的 consumer electronics 都消失了。
Speaker 122:04 - 22:24
So there will certainly be some cases where that's what's happened. I think there's another way of looking at this here maybe, which is just to think about how massively variable this will be. So there's a chart I've used for a while. The Marc Andreessen loves software's eating the world thing. The Uber doesn't sell software to taxi companies and Airbnb doesn't sell software to hotels. Speaker 122:04 - 22:24
所以,肯定会有一些情况确实是这样。我想,这里也许还有另一种看法,那就是去想一想,这件事的变化幅度会有多么巨大。我有一张图用了挺久。就是 Marc Andreessen 那句著名的话:software is eating the world。Uber 并不是把 software 卖给 taxi companies,Airbnb 也不是把 software 卖给 hotels。
Speaker 122:25 - 22:42
Really important observation. Fine. What did Uber do to taxis and compare that to what Airbnb did to hotels? Turns out you go and look in the, like, say, in New York, mean, it's a bit fuzzy because it depends by the city, but in New York, taxi market is down by, I forget the number, three quarters. Uber is gone like this. Speaker 122:25 - 22:42
这是个非常重要的观察。好。那 Uber 对 taxis 做了什么,再拿它和 Airbnb 对 hotels 做的事比较一下。结果是,如果你去看,比如说 New York,当然这事有点模糊,因为不同城市情况不同,但在 New York,taxi market 大概跌了——我记不清具体数字了——大概四分之三。Uber 则是这样一路冲上去。
Speaker 122:42 - 22:52
And the result is there's a bigger market with more people. I mean, the market looks different. The average number of rides per driver is different and so on. But the point is it unlocked all this new demand and demolish the yellow cab business. Fine. Speaker 122:42 - 22:52
而结果是,市场变得更大了,参与的人也更多了。我的意思是,这个市场的样子已经不同了。平均每个司机接单的次数变了,等等。但关键在于,它释放出了所有这些新的需求,并且摧毁了 yellow cab 业务。好。
Speaker 122:52 - 23:07
Now look at Airbnb, maybe slowed down the growth of hotels a bit. Yeah. It's maybe ten, five, 15% the size of the hotel market. And the reasons why are, well, hotels and cabs are different. Right. Speaker 122:52 - 23:07
现在再看 Airbnb,也许它只是稍微放缓了一点 hotels 的增长。对。它的规模可能只有 hotel market 的 10%、5%、15%。原因在于,hotels 和 cabs 本来就是不同的。对。
Speaker 123:07 - 23:10
Half of hotel business is business. There's a confidence business. Speaker 123:07 - 23:10
酒店业有一半是商务生意。还有一种“信心生意”。
Speaker 223:10 - 23:12
As long as the models have jagged capabilities like this, where they Speaker 223:10 - 23:12
只要这些 model 还呈现出这种参差不齐的能力,也就是它们
Speaker 123:12 - 23:52
work Yeah, a lot on some businesses but the and not are jagged and then the use cases are jagged and some of the use cases are physical and not just about processing information. So I wrote an essay a couple of weeks ago about like trying to predict job impact, because clearly this will have an impact on employment, but trying to do these star charts of radar charts of which ones, where you say, Ah, well, Opus 8.6 can do 93% of what a first year law firm associate is doing. This is a delusional statement. Because you can't measure what the law firm student is the associate's doing like that. You also can't measure whether the model can do it. Speaker 123:12 - 23:52
的确会工作——对,有些业务上很有效——但能力边界是参差不齐的,然后 use case(使用场景)也是参差不齐的,而且有些 use case 是物理层面的,不只是处理信息而已。所以几周前我写了一篇文章,想讨论如何预测对工作的影响,因为很明显,这会影响就业;但如果试图去画那种星形图、雷达图,说,啊,Opus 8.6 能完成一个律师事务所一年级 associate 93% 的工作——这是一种妄想式的说法。因为你根本没法那样衡量那个 law firm 的 associate 到底在做什么。你也同样没法衡量 model 是否真的能做到这些。
Speaker 123:53 - 24:14
This is actually the expert in a system fallacy, where you people kind of thought that you could measure how you recognize a cat from a dog. And you're saying it's the same mistake. You can't measure what a law associate does. Point I was making was, like, first of all, like, the the the accountancy job changed, which is the content we made earlier. Secondly, there are industries that your analysis would say are completely unaffected that will get demolished by something else like journalism. Speaker 123:53 - 24:14
这其实就是一种“系统中的专家”谬误。以前人们某种程度上以为,你可以衡量自己是如何区分猫和狗的。你的意思是,这是同样的错误:你没法衡量一个 law associate 在做什么。我想表达的是,第一,accountancy 这类工作本身会变化,这就是我们前面做过的内容。第二,有些行业按你的分析看似完全不受影响,但会被别的东西彻底摧毁,比如 journalism。
Speaker 124:15 - 24:28
For the sake of argument, the internet didn't change what it was to be a journalist at all, but it demolished the local advertising business, which wasn't in your analysis at all. Right. The same thing for Uber. Like, you wouldn't have looked at the smartphone. Nobody was I mean, I was in the mobile business. Speaker 124:15 - 24:28
为了方便讨论,假设 internet 完全没有改变“做一个记者”这件事本身,但它摧毁了本地广告业务,而这一点根本不在你的分析里。对吧。Uber 也是一样。比如说,你不会因为 smartphone 就看出这件事。没人会——我的意思是,我当时就在 mobile 行业里。
Speaker 124:28 - 24:39
We were talking about location all the time. Nobody saw that you have opportunity. Right. And so you kind of go and back test these analysis where you say, well, this job is exposed this much, and this job is exposed this much. It's horseshit. Speaker 124:28 - 24:39
我们那时一直在谈 location,但没人看出这里面会有这样的机会。对吧。所以你再回头去检验这类分析,说这个工作暴露了多少,那个工作暴露了多少——都是胡扯。
Speaker 124:40 - 24:55
You know, it's like the joke about the physicists who are predicting which horse is going to win a race, they say, we're gonna presume that the horse is a perfect sphere. You know? It was like, great. If you if you presume that If you define this into something that you can understand, then you can understand it, but it isn't that, it's something else. Speaker 124:40 - 24:55
这就像那个笑话:物理学家想预测哪匹马能赢比赛,于是说,我们先假设这匹马是一个完美的球体。你知道吧?就是这种感觉。很好,如果你先做出这种假设——如果你把它定义成一个你能理解的东西,那你当然就能理解它;但它根本不是那个东西,它其实是别的东西。
Speaker 224:55 - 25:01
Right. I think it's really this question of like, yeah, are there still capabilities that entry level workers will have that models different? Speaker 224:55 - 25:01
对。我觉得真正的问题是:是的,entry-level workers(初级员工)是否仍然拥有一些 model 所不具备、也不相同的能力?
Speaker 125:01 - 25:15
Yeah. Do you ever come across Anselm's proof? Anselm is this medieval theologian. And Anselm says, okay, first proposition, God by definition is that by which nothing could possibly be greater than God. Okay, that seems reasonable. Speaker 125:01 - 25:15
对。你有没有接触过 Anselm 的证明?Anselm 是一位中世纪神学家。Anselm 说,好,第一条命题,按定义,God 就是那个不可能有任何事物比 God 更伟大的存在。好,这听起来还算合理。
Speaker 125:15 - 25:31
Second proposition, something that exists is greater than something that doesn't exist. Okay? Therefore, God exists. And thirty seconds later, everyone in the room says, Anselm, dude, that's fucking bullshit. Because you could prove anything like that. Speaker 125:15 - 25:31
第二条命题,存在的东西比不存在的东西更伟大。对吧?因此,God 存在。然后三十秒之后,房间里的所有人都会说,Anselm,老兄,这他妈完全是胡扯。因为照这种方式,你什么都能证明。
Speaker 125:31 - 25:51
Well, a horse that exists is better than a horse that doesn't exist. Like it would be better if a unicorn existed than not exist. Like you can't prove stuff like that, but it took like a thousand years of people arguing about this to work out why that was wrong. And the reason I mentioned it is like, if you define AGI as necessarily inevitable and necessarily going to kill us all, then AGI is necessarily gonna kill us all. Great. Speaker 125:31 - 25:51
比如,一匹存在的马就比一匹不存在的马更好。比如说,如果 unicorn 存在而不是不存在,那当然会更好。你不能这样去证明事情,但人们花了差不多一千年争论这个问题,才搞清楚它错在哪里。我提这个的原因是,如果你把 AGI 定义成必然不可避免、并且必然会杀死我们所有人的东西,那 AGI 就必然会杀死我们所有人。很好。
Speaker 125:51 - 25:52
But you haven't proved anything. Speaker 125:51 - 25:52
但你其实什么都没有证明。
Speaker 225:52 - 26:14
Yeah. So then what do you pay attention to in the interim to, obviously it feels like almost there's always a set of people that the models can do something more, and it's like, Well, it still can't do X or Y. One thing I really appreciate about you is you have a nuanced perspective on this. You kinda take the facts that are out there. What are the things you're paying close attention to or would start to adjust your view on some of this? Speaker 225:52 - 26:14
对。所以在这期间,你会关注什么?很显然,总好像一直有一群人在说,模型又能多做一些事了;但同时又会有人说,好吧,它还是做不了 X 或 Y。我很欣赏你的一点是,你在这个问题上有一种细致、审慎的看法。你会根据已经摆在那里的事实来判断。你正在密切关注哪些事情,或者说,哪些事情会开始让你调整自己在这方面的看法?
Speaker 126:14 - 26:28
I think there's that I kind of pulling apart, like what are the sort of interesting sets of questions? Firstly, there was like a capital conversation. So, what's going on in chips? What goes on in data centers? When the chips catch up with supply and demand? Speaker 126:14 - 26:28
我觉得这里面可以稍微拆开来看:有哪些比较有意思的问题集合?首先,有一类是资本层面的讨论。那么,chips 这边到底发生了什么?data centers 里又发生了什么?芯片的供给和需求什么时候能重新赶上?
Speaker 126:28 - 26:43
Like what's the CEO of SK Hynix the other day said he reckons supply and demand is out of whack until 2030. Well, was like, well, he would say that, wouldn't he? Good business. Yeah, he's got a good business. So there's all of that stuff, data centers, time to power, grid deployment, etcetera. Speaker 126:28 - 26:43
比如,SK Hynix 的 CEO 前几天说,他认为供需失衡会一直持续到 2030 年。那我的反应是,嗯,他当然会这么说,不是吗?这对他的生意有好处。对,他的生意很好。所以这里面还有一整套问题:data centers、time to power、电网部署等等。
Speaker 126:44 - 27:17
Secondly, there is a sort of fairly, what seems to me really incontestable fact that right now there's something between three and six companies making models and they're all kind of the same. And you know, someone's ahead this week, someone's ahead next month, But the fact that SpaceX is like a paradoxical statement. The fact that SpaceX managed to jump straight back up to almost at the top of the leaderboards after completely flunking out is a really negative signal for SpaceX. How's that for a paradoxical statement? Because what that says to me is not very hard if you're willing to spend a couple of billion dollars and hire the right people. Speaker 126:44 - 27:17
第二,还有一个在我看来相当、而且几乎无可争辩的事实:现在大概有三到六家公司在做模型,而且它们基本都差不多。你知道,这周是某家领先,下个月又变成另一家领先。但问题在于,SpaceX 的情况像是一个悖论式的说法。SpaceX 在彻底失利之后,居然能一下子又几乎回到排行榜顶端,这对 SpaceX 来说其实是一个非常负面的信号。这个悖论式说法怎么样?因为这对我来说说明了一件事:如果你愿意花上几十亿美元,并且雇到合适的人,这事并没有那么难。
Speaker 127:17 - 27:26
This stuff remains they do not appear there are not yet fundamental barriers to entry. Barriers to entry in the sense that they were for search or social or mobile. Speaker 127:17 - 27:26
这种情况依然存在;目前看来,似乎还没有形成根本性的进入壁垒。这里说的进入壁垒,是指像 search、social 或 mobile 当年那种级别的壁垒。
Speaker 227:26 - 27:31
No, the capital, I mean, at this point, there's only so many companies that can bear the cost to Speaker 227:26 - 27:31
不,我的意思是,就资本而言,到这个阶段,能承担这种成本的公司终究只有那么几家,去
Speaker 127:31 - 27:34
train Well, a French there are, but it's not two. Speaker 127:31 - 27:34
进行训练。嗯,French 也有一些,但不是只有两家。
Speaker 227:34 - 27:36
Right. But like, it's gonna keep going up, Speaker 227:34 - 27:36
对。但比如说,这个数字还会继续往上升,
Speaker 127:36 - 27:48
Well, so this was my semiconductor comparison that we made earlier. What happened with semiconductors is it got harder and more expensive with each generation. And so, we shrank from dozens of companies at the cutting edge to just one, and then to one step behind. That may be how this evolves. We're not there yet. Speaker 127:36 - 27:48
对,所以这就是我之前做过的 semiconductor 对比。semiconductor 领域发生的情况是:每一代都变得更难、也更贵。于是,处在 cutting edge(前沿)的公司数量,从几十家缩减到只剩一家,然后再到那些落后一代的公司。这可能就是它接下来的演化方式。我们还没到那一步。
Speaker 127:48 - 27:51
Of course, if the curve slows down, then everything else changes. Speaker 127:48 - 27:51
当然,如果这条曲线放缓了,那其他一切都会随之改变。
Speaker 227:51 - 27:54
Yes. Compute is like the major moat of these businesses. Speaker 227:51 - 27:54
是的。compute(算力)就像是这些业务最主要的 moat(护城河)。
Speaker 127:54 - 28:09
Long as scaling continues, then you should sort of expect the number of cutting edge companies to shrink down. Then there's a then there's a question of, well, what what where do you need to be on what is it? The Pareto curve? Yeah. Like how many use cases need to be absolutely the cutting edge? Speaker 127:54 - 28:09
只要 scaling 还在继续,那么你大体上就应该预期,处在 cutting edge 的公司数量会继续收缩。然后问题就变成了:你到底需要处在什么位置——那叫什么来着?Pareto curve?对。就是说,有多少 use case(使用场景)真的需要绝对处在最前沿?
Speaker 128:09 - 28:19
How many well, put another way. How many use cases need to be there, have an ROI to be there? The other extreme at the end of the curve is like dictation works for free on your phone. Yep. And what's in the middle? Speaker 128:09 - 28:19
到底需要多少个 use case(用例)——或者换个说法——需要有多少个能带来 ROI(投资回报率)的 use case?这条曲线另一端的极端情况是,比如手机上的 dictation(语音听写)可以免费用。对。那中间地带是什么样?
Speaker 128:19 - 28:43
Like, does Amazon where on that curve does Amazon need to be to do how much or what to get what ROI on recommendations or review summarization or whatever? And so there's there will be people all the way along that curve. And presumably, as long as you the further, the cheaper you get, the more commoditized it gets. And you get to like, you've got dynamic real time bidding across 30 neoclass. And so how much stuff is at the head of the curve that needs that? Speaker 128:19 - 28:43
比如说,Amazon 在这条曲线上处于什么位置,才能在 recommendations(推荐)、review summarization(评论摘要)之类的事情上,投入多少、做到什么程度,从而获得相应的 ROI?所以,这条曲线上的各个位置都会有人在。并且可以推测,越往后走、越便宜,就越 commodity(商品化、同质化)。然后就会变成类似你在 30 个什么新类别上做动态实时竞价那样的东西。所以,真正位于曲线前端、需要那种能力的东西,到底有多少?
Speaker 128:45 - 28:57
How much value, how competitive is the head of the curve remain? And of course, there may be a paper that says, Hey, guess what? You can have a model that's 10x the size for a 10x the price. So it may be that you get some price collapse, even as everything scales. Yeah. Speaker 128:45 - 28:57
这些位于曲线前端的价值有多大?那一端还能保持多强的竞争性?当然,也可能会有一篇 paper(论文)说,嘿,猜怎么着?你可以用 10 倍大小的 model(模型),价格也只是 10 倍而已。所以也可能出现这样一种情况:即便一切都在扩张,价格反而还是会塌缩。对。
Speaker 128:57 - 29:17
I mean, we don't know. But then like how much of the value can the model capture? And this is kind of we were talking about earlier about the sort of, how do you explain it to the model question? And the thing I was thinking about, I do a weekly newsletter and I wrote a column about this last night, so it's sort of fresh in my mind is like, most people are three things. Most people aren't tool builders. Speaker 128:57 - 29:17
我的意思是,我们其实不知道。但接下来还有一个问题:model 能捕获多少价值?这有点回到我们前面说的那个问题:你要怎么把问题“解释给 model 听”?我想到的一点是——我每周写 newsletter(时事通讯),昨晚还写了一篇专栏谈这个,所以现在印象很新——大多数人有三个特点。大多数人不是工具的构建者。
Speaker 129:18 - 29:37
Most people don't see the problem that the tool would be solving. And most people aren't in a position to build the tool, even if the other two weren't true. And so I kind of explain what I mean by that. Like if you think about, like most enterprise software, half of enterprise software, you look at it and you don't get it. The customer looks at it and they don't get it. Speaker 129:18 - 29:37
大多数人看不到某个工具到底是在解决什么问题。并且,大多数人即使前两个条件都不成立,也并不处在能够把这个工具做出来的位置上。所以我稍微解释一下我的意思。比如你想想,大多数 enterprise software(企业软件),有一半你看一眼根本不明白是干什么的。客户看了也不明白。
Speaker 129:37 - 29:45
You don't see why that's a problem. You don't see why we need that. We don't understand why we would do that. Think about how many tools we all use every day, where the first time you saw it, you thought, what's the point of that? Yeah. Speaker 129:37 - 29:45
你看不出那为什么算个问题。你看不出我们为什么需要那个。我们也不明白为什么要那么做。想想看,我们每天用的多少工具,第一次见到时你都会想:这玩意儿有什么用?对吧。
Speaker 129:46 - 30:09
So now you're you've got the job. You've got that problem every day. Very often, you don't see that you've got that problem. And then even if you see it, most people the the skill of being like a really great salesperson is completely different to the skill of being really good at designing a new piece of sales enablement software. The skill of being a really good video editor is completely different to the skill of being really good at designing a completely different way you edit video. Speaker 129:46 - 30:09
所以现在你是这样一种情况:你手上有这份工作,你每天都在面对那个问题。可很多时候,你甚至意识不到自己有这个问题。然后即便你意识到了,大多数人也一样——成为一个非常出色的 salesperson(销售人员)的技能,和设计一款全新的 sales enablement software(销售赋能软件)的技能,完全是两回事。成为一个非常优秀的 video editor(视频编辑)的技能,和设计一种全新的视频编辑方式的技能,也完全是两回事。
Speaker 130:09 - 30:29
I mean, the person who is who does do that probably is does know a lot about video editing, but most people aren't that person. So the skill of actually working out what the software should be, what the tools should be, how it should be done, what the right structures and workflows and how the network should work. Think about every great software pitch you've ever seen. They're like, oh wow, that's a really clever way of doing it. Yeah. Speaker 130:09 - 30:29
我的意思是,真正能做出这种事的人,可能确实非常懂 video editing(视频编辑);但大多数人都不是那种人。所以,真正去想清楚 software 应该是什么样、tools(工具)应该是什么样、该怎么做、正确的结构和 workflow(工作流)是什么、以及整个网络应该如何运作,这本身就是一种能力。想想你见过的每一个出色的软件 pitch(产品推介):你的反应都会是,哦,原来这样做这么巧妙。对。
Speaker 130:29 - 30:56
I would never have thought of that. And yet you're expecting like random middle managers in the back office of big companies to dream that up like this afternoon, if you give them thought. And then like the third piece is very often it's regulated data. And even if it's not, there's 1,500 people who are touching this data and it needs to go into a system of record and there's money attached to it. So you can't have random people just building tools and like replacing SAP Yeah. Speaker 130:29 - 30:56
我绝不会想到那个。可你却指望大公司 back office 里的普通中层经理,只要给他们一点时间,就能在今天下午把这种东西凭空想出来。而第三个因素往往是,这些都是受监管的数据。即便不是,也会有 1,500 个人在接触这些数据,它还必须进入一个 system of record(记录系统),而且背后牵涉到资金。所以你不能让随便什么人自己搭工具,然后去替换 SAP,对吧。
Speaker 130:56 - 31:15
Or replacing these systems. And so everything gets more fuzzy. So like there's a sort of there's like an improvised space of CSVs and Excel and Tableau and Perl and so on. And then you've got your big iron horizontal systems and then you've got 400 or 500 SaaS apps inside every big company. And AI kind of shuffles all of those around. Speaker 130:56 - 31:15
或者替换这些系统。所以一切都会变得更模糊。也就是说,会存在一个由 CSV、Excel、Tableau、Perl 等拼凑出来的临时空间;然后你还有那些大型、底层的横向系统;再然后,每家大公司内部还会有 400 到 500 个 SaaS apps。AI 某种程度上会把这一切重新洗牌。
Speaker 131:15 - 31:32
And so like if you're PwC and you hire a thousand graduates a year, 10,000 graduates a year, you use software. If you're Redpoint and you maybe hire a couple of people every year, you've got Google Sheets. Yeah. And then there's something in between. And there's a point where you say, Hey, maybe we should use Notion or maybe we should hot build this software. Speaker 131:15 - 31:32
所以,如果你是 PwC,每年招聘 1,000 个毕业生,或者 10,000 个毕业生,你就会用软件系统。如果你是 Redpoint,也许每年只招几个人,那你用的就是 Google Sheets。对吧。然后中间还存在一个过渡地带。到了某个点,你会说,嘿,也许我们该用 Notion,或者也许我们该快速自己搭这套软件。
Speaker 131:32 - 31:57
Maybe we should get this thing. And then this horrible moment where you see, hey, maybe we should use Workday, which is like the ultimate test for AGI. If it can navigate Workday successfully. Then none of that is binary. And so then, okay, there will be this new SaaS app that somebody built using AI and maybe or maybe not, it does stuff using AI that you couldn't do before, but also you'll be able to do that in SAP now. Speaker 131:32 - 31:57
也许我们该上这个东西。然后会出现那个可怕的时刻:你会发现,嘿,也许我们该用 Workday——这简直像是 AGI 的终极测试。如果它能成功驾驭 Workday,那就说明问题了。而这一切都不是非黑即白的。于是,好吧,可能会出现某个新的 SaaS app,是有人用 AI 做出来的;也许它确实能用 AI 做一些以前做不到的事,也许不能;但与此同时,你现在也能在 SAP 里做这些事了。
Speaker 131:57 - 32:10
And also you'll be able to get Excel to do that. And also maybe you'll be able to use Cowork to do that. And so, which do you choose? Well, how did you choose whether to buy software or use Excel or use SAP before? Same question. Speaker 131:57 - 32:10
而且你也能让 Excel 去做这些事。也可能你还能用 Cowork 去做这些事。所以,你该选哪一个?嗯,你以前在“买软件”、“用 Excel”还是“用 SAP”之间是怎么选的?还是同一个问题。
Speaker 232:10 - 32:29
Yeah. I mean, I thought there's two parts of your argument. I think you've probably said the foundation models are mean, I don't even take it as like not necessarily valuable. It's more just like, we'll not capture most of the value of this ecosystem. I'm still unsure whether you think they'll actually be, whether these long term anthropogenic and open AI go public will be. Speaker 232:10 - 32:29
对,我的意思是,我觉得你的论点里有两层。我想你可能是在说,foundation models(基础模型)——我的意思是,我并不是把你的意思理解成“它们没有价值”。更像是说,它们不会捕获这个生态系统里的大部分价值。至于你是否认为,从长期看,这些像 Anthropic 和 OpenAI 这样的公司上市后到底会怎么样,我还是不太确定你的看法。
Speaker 132:29 - 32:41
Well, so, I mean, again, without arguing by analogy, I will point to analogy. TSMC has a monopoly on the cutting edge of semis. You don't write apps for TSMC. Speaker 132:29 - 32:41
嗯,所以,我还是那句话,不靠类比来论证,但我会指出一个类比。TSMC 在最尖端 semis(半导体)上拥有垄断地位。你不会给 TSMC 写 apps。
Speaker 232:41 - 32:43
Yeah. One of the 10 biggest companies in the world though. Speaker 232:41 - 32:43
对,不过它仍然是世界上市值前 10 的公司之一。
Speaker 132:43 - 32:58
Yes. But their net income last year was half of Apple's. It's a great company, very profitable company. Uber does not write out for TSMC. You have multiple layers in the stack, different abstraction layers, different ways of building that. Speaker 132:43 - 32:58
是的。但他们去年的净收入只有 Apple 的一半。这是一家很棒的公司,也是一家盈利能力很强的公司。Uber 不会替 TSMC 开支票。你在这个 stack(技术栈)里会有多层结构、不同的 abstraction layers(抽象层),以及不同的构建方式。
Speaker 132:58 - 33:14
And the right people to build an enterprise go to market for creativity software, different people to the right people to build. This is why we have a stack. This is why AWS doesn't own the entire tech industry. It's why every app on your iPhone isn't made by Apple. Speaker 132:58 - 33:14
而且,适合为 creativity software(创意软件)打造 enterprise go-to-market(企业级市场进入/销售体系)的人,和适合去做构建的人,是不同的一批人。这就是为什么我们会有一个 stack。这也是为什么 AWS 并不拥有整个科技行业。也是为什么你 iPhone 上的每一个 app 都不是 Apple 做的。
Speaker 233:14 - 33:21
Right. But I think, obviously, if you're the AWS of the AI space and the opportunity is as massive as it appears, that's still quite a valuable business. Speaker 233:14 - 33:21
对。但我认为,很显然,如果你是 AI 领域的 AWS,而这个机会真的像看上去那么巨大,那这依然会是一门相当有价值的生意。
Speaker 133:21 - 33:24
Yeah. Great business. Yeah. And this is thing. Windows is a great business. Speaker 133:21 - 33:24
对。很棒的生意。对。问题就在这儿。Windows 是一门很棒的生意。
Speaker 133:25 - 33:48
TSMC is a great business. AWS is a great business. None of them own the whole stack all the way up to the top. Now you can turn around and say, yes, but Benedict, you'll just be able to go to Claude and say, invent 10 new enterprise software applications and then go build them, and then go and work out the right way of building You could build the whole thing, and again, we're back to our Anselm. You're back to, well, if I define this as something that can do absolutely everything, can it do absolutely everything? Speaker 133:25 - 33:48
TSMC 是一门很棒的生意。AWS 是一门很棒的生意。它们没有任何一家拥有从底层一直到最上层的整个 stack。现在你当然可以反过来说,是的,但是 Benedict,你以后只要去找 Claude 说,发明 10 个新的 enterprise software applications(企业软件应用),然后去把它们做出来,再去找出正确的构建方式——你可以把整套东西都做出来。然后,我们又回到了我们的 Anselm。你又回到了那个问题:如果我把这东西定义成一个什么都绝对能做的东西,那它是不是就真的什么都能做?
Speaker 133:49 - 33:52
Yes. Okay. But that's the question. Speaker 133:49 - 33:52
是的。好吧。但问题就在这里。
Speaker 233:53 - 33:56
Can't win the argument by defining it away. What do you make of what's happened in the coding space? Speaker 233:53 - 33:56
你不能靠把问题重新定义掉来赢下这场争论。你怎么看 coding(编程)领域里已经发生的这些事?
Speaker 133:56 - 34:00
It's funny to see obviously software jobs exploding. Speaker 133:56 - 34:00
很有意思,显然 software(软件)岗位正在爆发式增长。
Speaker 234:01 - 34:05
More just from the model providers competing pretty effectively against application companies. Speaker 234:01 - 34:05
更多还是来自 model providers 与应用公司竞争,而且竞争得相当有效。
Speaker 134:06 - 34:30
Well, so I think it's kind of interesting to think about what happened with SAS here. Whereas with SAS, we went from putting an order of magnitude, maybe several orders of magnitude increasing the amount of software that there was. And so, the PC area, the typical big company has, I don't know, a couple of dozen apps in the data center. And now the typical big American company has 400 to 500 SaaS applications plus several thousand legacy stuff they've accumulated over the last thirty years. And that was just go to market. Speaker 134:06 - 34:30
嗯,所以我觉得拿这里的 SaaS 来想这个问题挺有意思。SaaS 出现后,软件数量增长了一个数量级,也许甚至是好几个数量级。比如在 PC 时代,典型的大公司在 data center 里大概也就有,我不知道,几十个 app。可现在,典型的大型 American 公司会有 400 到 500 个 SaaS application,外加过去三十年里积累下来的几千个 legacy 系统。这还只是 go to market 层面的变化。
Speaker 134:30 - 34:49
And so that meant you could automate all of these things that you could not have justified getting on prem software for before. And so at a minimum, we should presume, okay, it's gonna be way cheaper and way quicker to build new software. Plus it can do this whole class of thing that your software couldn't do at all before. Therefore, there will be more software. Yep. Speaker 134:30 - 34:49
这意味着,很多以前不值得为其部署 on-prem software(本地部署软件)的事情,现在都可以自动化了。所以至少我们应该先假设:好,构建新软件会变得便宜得多、快速得多;再加上它还能完成一整类你以前的软件根本做不到的事。因此,软件会变得更多。对。
Speaker 134:49 - 35:02
And some of that, some of the incumbents will fall away. Some of them will fail to make that jump. Some big complex expert systems will get automated away by some very simple machine learning system. Some stuff will get bundled, unbundled. This is what happened, right? Speaker 134:49 - 35:02
而在这个过程中,一部分 incumbents(既有厂商)会掉队,有些会无法完成这次跃迁。有些大型、复杂的 expert systems(专家系统)会被非常简单的 machine learning system(机器学习系统)自动化取代。有些东西会被 bundled(捆绑),有些会被 unbundled(拆分)。事情一向就是这样,对吧?
Speaker 135:02 - 35:16
Remember what happened to PeopleSoft? What happened to Siebel Systems? Like there will be that kind of platform shift, people die, fine. Second question is, it was the point I was making earlier about, does this mean you have less or more software? Do you How much stuff you no longer need the SaaS app? Speaker 135:02 - 35:16
想想 PeopleSoft 怎么了?Siebel Systems 又怎么了?这种 platform shift(平台迁移)一定会发生,会有人出局,这很正常。第二个问题就是我前面在说的:这是否意味着软件会变少还是变多?会有多少事情让你不再需要那个 SaaS app?
Speaker 135:16 - 35:21
You'll just do it inside Excel with Copilot or you'll do it inside Right. Cloud I mean, I think almost Speaker 135:16 - 35:21
你可能直接在 Excel 里配合 Copilot 来做,或者你会直接在——对,Cloud 里面做。我的意思是,我觉得几乎——
Speaker 235:21 - 35:29
certainly we'd all agree there's more software, but I guess that maybe the point I was trying to make on Cloud Code is like that is an example the model providers moving into the application layer and winning it pretty effectively. Speaker 235:21 - 35:29
几乎可以肯定,我们都会同意软件会变得更多;但我想我在说 Cloud Code 时想表达的一点是,那就是一个 model providers 向 application layer(应用层)推进并且相当有效地赢下竞争的例子。
Speaker 135:30 - 35:48
Yeah. Mean, I suppose this is It's funny. This is like all the conversations about what should be integrated in macOS or what should be integrated in Windows. So, and I wrote stuff about this years ago that you go back to the mid eighties and spreadsheets didn't do printing or didn't do charts, particularly didn't do charts. Yeah. Speaker 135:30 - 35:48
对。我想,这很有意思。这有点像以前那些关于“什么应该被集成进 macOS,什么应该被集成进 Windows”的讨论。所以,我很多年前写过这方面的东西:如果你回看八十年代中期,spreadsheets(电子表格)当时并不具备打印功能,也没有图表功能,尤其是没有图表功能。对。
Speaker 135:48 - 36:10
So charts were a separate program. In fact, can go back to the late early nineties and you can find these group tests of 30 or 40 spell checkouts. The So way it worked is you save your document in your Word or WordPerfect or whatever, and then you open the spell check app and spell check apps were like $203,100 dollars each. And they would have this grid of like, they support Word six or do they only support Word five? And do they support WordPerfect? Speaker 135:48 - 36:10
所以,图表是一个独立的程序。其实,你可以回到 90 年代初期,看到那种对 30 或 40 个拼写检查器做的群组测试。它的工作方式是:你先在 Word、WordPerfect 或别的什么软件里保存文档,然后再打开 spell check app(拼写检查应用),而且这些 spell check app 每个都要卖到 20、30,甚至 100 美元。评测里会有一个表格,列着类似这样的问题:它支持 Word 6 吗,还是只支持 Word 5?它支持 WordPerfect 吗?
Speaker 136:10 - 36:38
And And you'd run that. And of course, over time that gets integrated, and then it becomes wavy red lines. Printing, of course, was like, that was a whole thing. They got integrated into the operating system, and all the DOS apps thought they should do printing, and then it gets integrated into the operating So there is this question of what sits naturally at what should, what naturally gets integrated down into your enabling layer and what should be a separate application. This is of course the antitrust argument about Internet Explorer, but today I could be bought an iPhone and it didn't have a web browser. Speaker 136:10 - 36:38
然后你就运行它。当然,随着时间推移,这些功能会被集成进去,最后变成红色波浪线。打印当然也是一样,那曾经也是一整套复杂的东西。后来它们被集成进 operating system(操作系统)里;再早些时候,所有 DOS 应用都觉得打印应该由自己来做,之后又被集成进 operating system。所以这里就有一个问题:什么东西天然应该放在什么位置,什么功能会自然地下沉并集成到你的 enabling layer(基础使能层)里,什么又应该作为独立应用存在。这当然就是关于 Internet Explorer 的 antitrust(反垄断)争论,但放到今天,如果我买了一部 iPhone,结果它没有 web browser(网页浏览器),那就太离谱了。
Speaker 136:38 - 36:59
That would be insane. So it changes over time. And so there is this question, well, what naturally gets integrated into the lowest level substrates and what naturally gets unbundled into separate things where you really need to understand, have deep vertical domain knowledge and go to market and everything else that builds a vertical software app. And this is the same question of, well, why isn't that part of AWS? And the answer is, well, Speaker 136:38 - 36:59
那会很荒唐。所以这会随着时间变化。于是问题就变成了:什么东西会自然地被集成到底层 substrate(基础底层)里,什么东西又会自然地被拆分成独立模块——在那些场景里,你确实需要理解业务、具备很深的垂直领域知识、做 go-to-market(市场进入),以及构建 vertical software app(垂直软件应用)所需的一切。这和“为什么那不是 AWS 的一部分?”其实是同一个问题。答案是,嗯,
Speaker 236:59 - 37:33
it Maybe to push back for a second, like in the case of coding, like the model to some extent is the product, right? So that feedback loop that you have between the model, the real world usage of the model, then makes the underlying product better. Whereas AWS is just like the compute itself, whether it's used in pharma or whether it's used in a grocery store, it's not like that usage fundamentally changes the product in any way. Compute is compute. In this case, it would be very hard to and it has been very hard for other players, and I think both Windsurf, Kershwar, they've all concluded like, God, we need to own the underlying model layer to have an effective product in this space. Speaker 236:59 - 37:33
不过也许可以先反驳一下:比如在 coding(编程)这个场景里,model(模型)在某种程度上就是产品,对吧?所以,你在 model、model 的真实世界使用情况之间形成的那个 feedback loop(反馈回路),会进一步让底层产品变得更好。而 AWS 只是 compute(算力)本身;不管它被用在 pharma(制药)还是 grocery store(杂货店),这种使用方式都不会从根本上改变产品本身。compute 就是 compute。在这种情况下,要做到这一点会非常难,而且对其他玩家来说也一直非常难。我觉得无论是 Windsurf 还是 Kershwar,他们最后都得出了同样的结论:天啊,我们必须拥有底层 model layer(模型层),才能在这个领域做出真正有效的产品。
Speaker 237:33 - 37:37
And ultimately, just as an application on top, we're gonna get totally hosed. Speaker 237:33 - 37:37
否则,最终如果我们只是上层的一个 application(应用),那我们会被狠狠干掉。
Speaker 137:37 - 37:50
Yeah. But isn't that the conversation that we had with the hyperscalers? Which things should naturally be part of the platform and which parts are dedicated vertical tools? Which parts are naturally part of Windows? Yeah. Speaker 137:37 - 37:50
对。但这不就是我们当初和 hyperscalers(超大规模云服务商)讨论过的问题吗?哪些东西天然应该属于 platform(平台)的一部分,哪些部分应该是专用的垂直工具?哪些部分天然就该属于 Windows?对。
Speaker 137:51 - 37:55
And which should unnaturally better if they're part of the operating system and which parts aren't. Speaker 137:51 - 37:55
以及,哪些部分如果成为 operating system(操作系统)的一部分会不自然地更好,哪些部分则不会。
Speaker 237:55 - 38:00
The one big market we've had so far in LLM world has actually proven to be very good for the model providers Speaker 237:55 - 38:00
到目前为止,LLM world(大语言模型世界)里我们见到的那个唯一的大市场,实际上已经证明这对 model providers(模型提供商)非常有利。
Speaker 138:00 - 38:22
So, to provide think there's an interesting, there's a couple of like things you can ask about the specificity of why is it that it's coding that worked? Yeah. Is it just that it's that all these people are coders? Is it like, well, we fix the problem that we see in front of us? Is it that there's something specific about LLMs and code that makes it work particularly well? Speaker 138:00 - 38:22
所以,我想先提供一个我觉得挺有意思的角度:关于“为什么偏偏是 coding 起作用了”这个问题,其实有几个更具体的点可以问。对吧?只是因为这些人本来都是 coder 吗?还是说,嗯,我们就是去修眼前看到的问题?又或者,是不是 LLM 和 code 之间确实有某种特别契合的地方,所以它才尤其有效?
Speaker 138:23 - 38:44
Is it that you see the problem in front of you and say that's what you work on versus is there something specific about this? Is it my point about tool builders? Is it that the tool builders are building the tool for their task? The one place where the tool builders are developers is development tools. So, I don't know. Speaker 138:23 - 38:44
是因为你能直接看到眼前的问题,然后就说“这就是我要解决的事”,还是说这里面确实有某种更特殊的原因?是我刚才提到的 tool builders 那个点吗?也就是说,tool builders 正在为他们自己的任务打造工具?而 tool builders 恰好就是 developers 的一个场景,就是 development tools。所以,我也不确定。
Speaker 138:44 - 38:54
Is it is it is And I don't think we have a good I'm I maybe I'd love to know. Is there some theoretical There are theoretical arguments that says that code is uniquely well suited to LLMs. Speaker 138:44 - 38:54
是不是……而且我觉得我们现在并没有一个很好的答案。我——我其实也很想知道。有没有某种理论上的解释?确实有一些 theoretical arguments(理论论证)认为,code 对 LLM 来说是独特地适配的。
Speaker 238:54 - 38:59
A 100%. Go to math, very verifiable. Can run them millions of times. Speaker 238:54 - 38:59
绝对是。比如 math,也是非常容易验证的。你可以把它们跑上几百万次。
Speaker 138:59 - 39:25
Well, this is another point. This is another point is where If we're going back to the conversation about automation, where do you have enough training data and where is validation scalable? And the more that you get into complex consulting projects, the harder it is to have scalable repeatable validation. Yeah. Because you're getting a long way away from something that's tangible and specific and objective. Speaker 138:59 - 39:25
嗯,这又是另一个点。另一个点在于——如果我们回到 automation 这个话题——到底哪些地方你有足够的 training data,哪些地方 validation(验证)又是可以规模化的?而你越是进入复杂的 consulting project,就越难拥有那种可规模化、可重复的 validation。对。因为你离那种具体、明确、客观的东西越来越远了。
Speaker 239:25 - 39:48
And I think this is like exactly you know, I think there's lot of folks that have been saying, there's whole Neolab push lately of folks spinning out and saying, god, there's gotta be a more data efficient way to train models. And ultimately, if like the way we did code is the only way to do everything else, maybe we get there over time and we the whole economy is just data labelers, and over time you can even do enough data labeling to get there. But like, feels like that will be a real slog relative to something that's more algorithmically efficient. Speaker 239:25 - 39:48
而且我觉得这正是——你知道——很多人一直在说的点。最近有一整波 Neolab 的推动,很多人出来创业并且在说,天啊,训练 model 肯定应该有一种 data efficient(数据效率更高)的方式。归根结底,如果我们做 code 的那套方法,是做其他一切事情的唯一方法,那么也许随着时间推移我们也能到达那个目标,整个经济体都变成 data labelers(数据标注员);再往后,也许你甚至真的能靠足够多的数据标注把这件事做成。但感觉相比某种在 algorithm 上更 efficient 的路径,那会是一场非常艰难的苦战。
Speaker 139:48 - 40:22
So pushing in a different direction, sort of places that it is, given that we don't know how the models will work, what are places where it's conceptually questionable that they might, whether they could work? And one of them is in the conversation about where there isn't training data, where the data is implicit, where it's implicit knowledge, maybe where it's hard actually to explain how you do it, why you do it, which comes right back to the beginning of machine learning. It's hard to explain where you can't turn it into a statistics problem. Right. Where is it hard to turn this from a logic problem into a statistics problem? Speaker 139:48 - 40:22
所以换个方向来想:既然我们并不知道 model 将会如何起作用,那么有哪些地方,从概念上说,它们是否能起作用本身就是可疑的?其中一个就是前面谈到的:那些没有 training data 的地方,那些 data 是隐含的地方,那些属于 implicit knowledge(隐性知识)的地方,也许甚至是那种你实际上很难解释自己是怎么做、为什么这么做的地方。这个问题其实又直接回到了 machine learning 的起点:困难恰恰出现在你没法把它转化成一个 statistics problem(统计问题)的地方。对吧。哪些地方是很难把它从一个 logic problem(逻辑问题)变成一个 statistics problem 的?
Speaker 140:22 - 40:54
And secondly, where do you not want the average? Because in principle, what all of these things are doing is how would most people probably do this? And in software you are saying, how would most people probably build an app that does X, Y, and Z with these criteria? And you want the way most people would probably do it. And the more that you get into, no, I kind of don't want it the way most people would probably do it. Speaker 140:22 - 40:54
第二,哪些地方你并不想要“平均值”?因为原则上,这些东西做的都是:“大多数人很可能会怎么做这件事?”而在 software 里,你实际上是在问:“大多数人很可能会怎样 build 一个满足这些条件、实现 X、Y、Z 的 app?”而你想要的,恰恰就是大多数人可能会采用的那种做法。但你越往后走,就越会进入一种情况:不,我其实并不想要“大多数人很可能会怎么做”的那种答案。
Speaker 140:54 - 41:24
I'm going to do something different. Then the more you get kind of the models there's a kind of conceptual question for the models. And the way I used to think about this was that if you go back to something like AlphaGo, AlphaGo can do model to do moves that no one's done before, but it's got scalable verifying system. It can check whether the moves win. Whereas if you were to create like a completely new enterprise strategy, how would you check that was completely different from what anyone had ever done before? Speaker 140:54 - 41:24
我打算换个角度说点不一样的。因为当你对这些 model(模型)理解得更深一点时,就会出现一个关于 model 的概念性问题。我过去想这件事的方式是:如果你回到像 AlphaGo 这样的例子,AlphaGo 可以让 model 下出以前从没人下过的棋,但它有一个可扩展的验证系统。它可以检查这些走法是否能赢。可如果你要提出一种全新的 enterprise strategy(企业战略),你该怎么验证它确实不同于以往任何人做过的东西?
Speaker 141:24 - 41:26
How would you check whether that made any sense? Speaker 141:24 - 41:26
你又该怎么验证那件事到底讲不讲得通?
Speaker 241:26 - 41:46
Right. It's all about, mean, people are creating these very basic RL environments today, right? Where it's like, hey, did you go through the Salesforce app and update things correctly and use that to train models? And it feels very clear that over time, I think the only solution to a of this stuff is gonna need to be in simulation, right? You kind of are simulating complex business situations or economies. Speaker 241:26 - 41:46
对。归根结底,现在人们正在创建这些非常基础的 RL(强化学习)环境,对吧?比如说,“你有没有在 Salesforce app 里正确完成更新”,然后用这个来训练 model。很明显,随着时间推移,我觉得这类问题唯一的解决办法大概还是得落在 simulation(仿真)上,对吧?你某种程度上是在模拟复杂的商业情境或经济系统。
Speaker 141:46 - 42:30
Yeah, so just sort of thinking out loud here, there's like the dumb criticism is, the model can't do, well, there's a short term criticism which says the model can't do that. The longer term question is, can you get enough training data for the model to be able to do that and enough feedback loops for the model to be able to do that? The third step is, yes, but do you want the product of the training data? Do you want the average? And then you can push this different places because, I I have the sort of Imagine If slides towards the end of my presentation, the last presentation, and I was sort of thinking like there's a consumer version and an enterprise version. Speaker 141:46 - 42:30
对,我只是顺着这个思路大声想一想。有一种比较粗浅的批评是:model 做不了这个。然后还有一种短期批评,说 model 现在还做不了。再往长期看,问题就变成:你能不能为 model 获取足够的训练数据,让它能做到这件事?以及能不能为 model 建立足够的 feedback loop(反馈回路),让它能做到这件事?第三步则是:可以,但你真的想要训练数据产出的那个结果吗?你想要的是平均值吗?然后你还可以把这个问题往不同方向推进,因为我在演讲最后那几页有一些 “Imagine If” 的 slides,我当时在想,大致可以分成 consumer(消费端)版本和 enterprise(企业端)版本。
Speaker 142:30 - 42:39
Consumer version is, step one, here's a picture of a coat, where can I buy that? That wouldn't have ten Okay, years ago, science fiction. Now that's an angemon. Now that's a product problem. Yep. Speaker 142:30 - 42:39
consumer 版本是这样:第一步,这里有一张外套的照片,我能去哪里买到它?如果是十——好吧,十年前,这还是 science fiction(科幻)。现在这已经是 an angemon。现在这已经是个产品问题了。对。
Speaker 142:39 - 42:46
That would work. And coats may be hard, but yeah, in principle, will work. Here's a here's a vase. Where can I buy this? Here's a lamp. Speaker 142:39 - 42:46
这是可行的。外套可能会比较难,但对,原则上这是能做成的。这里有一个花瓶。我能去哪里买这个?这里有一盏灯。
Speaker 142:46 - 43:06
Where can I buy this? Step two, give me 10 options for a coat like that and explain how I would choose. Okay, that's now solvable. Step three, go look on my Instagram and suggest a coat that matches my look, that would improve my look, that would give me a radical change that I might like. That can probably do now. Speaker 142:46 - 43:06
我能去哪里买这个?第二步,给我 10 个类似这件外套的选项,并解释我该如何选择。好,这现在已经是可解的问题了。第三步,去看看我的 Instagram,然后推荐一件符合我风格的外套,一件能提升我整体形象的外套,或者一件能带来我可能会喜欢的激进变化的外套。这个现在大概也能做。
Speaker 143:06 - 43:56
Yeah. But the further you get to not what's obvious, but good, the harder it is to do that. The enterprise version of this would be something like, step one is listen to our Zoom recordings with clients and tell me if there were new concerns emerging in the last couple of months. Step two be, look at our Step two or three would be, go look at our app telemetry and everything in Salesforce, and all of our client calls, and then do a competitive market review. And then spin up a user survey, automate the questions, work out the list of users, run the conference calls, do all the calls, synthesize the results, pull it all back together again, and then tell me why do I need to change my churn. Speaker 143:06 - 43:56
对。但是,你越是从“什么是显而易见的”走向“什么是好的”,这件事就越难做。它在 enterprise 版本里的对应形态,大概会是这样:第一步,听我们和客户的 Zoom 录音,然后告诉我,在过去几个月里是否出现了新的担忧。第二步会是,看我们的——第二步或第三步会是,去查看我们的 app telemetry(应用遥测数据)、Salesforce 里的所有内容,以及我们所有与客户的通话,然后做一份竞争市场评估。接着再启动一个用户调查,自动化生成问题,整理用户名单,安排 conference calls,完成所有通话,综合结果,把这些重新汇总起来,然后告诉我:为什么我需要调整我的 churn(客户流失)?
Speaker 143:56 - 44:11
How do I change my churn? Now, that's an example, incidentally, of something where you probably do need to be at the head of that pricing curve. It's not a commodity question. That's also something that one random person in the company isn't gonna be able to spin up in Claude. You're gonna that's a one month project just to kick that off, just to get that going. Speaker 143:56 - 44:11
我该怎么降低我的 churn(客户流失)?顺便说一句,这就是一个例子:在这类事情上,你大概率确实需要站在那条 pricing curve(定价曲线)的前沿。这不是一个 commodity(标准化商品)问题。公司里也不是随便找个人就能在 Claude 里立刻搞出来的。你得——光是把这件事启动起来、让它运转起来,就是一个为期一个月的项目。
Speaker 144:12 - 44:34
These are sort of the further you get along that curve, the more you're not doing, this is what anyone would tell me. And the more you're getting to variance. To make that much simpler, like, you can get a machine learning. You can get an AI music model to make you a lot more bad jazz or bad bark or bad pop music. You can get it to make you more stuff that sounds like Taylor Swift. Speaker 144:12 - 44:34
大概就是,沿着那条曲线走得越远,你做的事情就越不是“任何人都会告诉我该怎么做”的那种。你会越来越走向 variance(差异化)。说得更简单一点,比如,你可以用 machine learning(机器学习),你可以用 AI 音乐模型,帮你生成更多糟糕的 jazz、糟糕的 bark,或者糟糕的 pop music。你可以让它生成更多听起来像 Taylor Swift 的东西。
Speaker 144:34 - 44:46
Apologies to any Taylor Swift fans listening. You can make more Taylor Swift, fine. But imagine inventing punk. Imagine inventing hip hop. Speaker 144:34 - 44:46
向正在收听的 Taylor Swift 粉丝们道个歉。你可以做出更多 Taylor Swift 风格的东西,没问题。但想象一下去发明 punk。想象一下去发明 hip hop。
Speaker 244:46 - 44:46
Right. Speaker 144:47 - 44:47
I didn't Speaker 144:47 - 44:47
我不——
Speaker 244:47 - 45:01
think it comes where does that from. Come 99% of things in enterprise probably not inventing hip hop. Like that example you just gave, I think you probably could do with LLMs today. And people are, it's not perfect today, but you have your AI interviewers and you have there's enough iteration on it that it doesn't feel Speaker 244:47 - 45:01
觉得那是从哪儿来的。企业里 99% 的事情,大概都不是在“发明 hip hop”。像你刚举的那个例子,我觉得你今天大概就可以用 LLMs(大语言模型)来做。而且人们也确实在做;今天它还不完美,但你已经有 AI interviewers(AI 面试官)了,而且围绕这个已经有了足够多的 iteration(迭代),以至于它不会让人觉得——
Speaker 145:01 - 45:19
The funny yeah. The funny thing about AI interview is I was was talking to somebody from Bain a while ago who were talking about using this for interviews and for synthetic customers. There's a common pattern I've got it's kind of complete sidebar here. There's a common pattern in people talking about what AI means to x y zed industry. If you're not in that industry, you'd think you'd do this with it. Speaker 145:01 - 45:19
有意思的是——对,关于 AI 面试这件事,有意思的是,我前阵子在和 Bain 的一个人聊天,他们当时在谈把这个用在面试里,以及用在 synthetic customers(合成客户)上。我观察到一个很常见的模式——这里算是完全岔开一下题——就是,人们在谈 AI 对某个 x y zed 行业意味着什么的时候,如果你不在那个行业里,你会觉得你应该会拿它来做这个。
Speaker 145:19 - 45:22
And the people in the industry are like, no. No. No. No. We do this. Speaker 145:19 - 45:22
而那个行业里的人会说,不。不是。不是。不是。我们做的是这个。
Speaker 145:22 - 45:29
This is the thing we can do with it. So like in advertising, there's a whole thing of how you can make the ads with it. And and Martin Sowell is like, no. No. No. Speaker 145:22 - 45:29
这才是我们能拿它做的事。比如在 advertising(广告)行业,大家会有一整套说法,讲你可以怎么用它来制作广告。而 Martin Sowell 会说,不。不是。不是。
Speaker 145:29 - 45:37
No. You use it to automate all the really boring back office crap. Yeah. You automate the faxes. The faxes we're sending to Google to book the ads, almost literally. Speaker 145:29 - 45:37
不是。你是用它来把那些特别无聊的 back office(后台运营)破事都自动化。对,就是把传真自动化。那些我们发给 Google 用来预订广告的传真,几乎真就是字面意义上的传真。
Speaker 145:37 - 45:51
And the same thing in consulting, like, we use this to make PowerPoints? Well, kind of. But what we really use it for is to automate the research, customer research. Right. Because now you can do that query, you can do that survey in a day instead of week. Speaker 145:37 - 45:51
咨询行业里也是一样,比如,我们是用这个来做 PowerPoint 吗?某种程度上是。但我们真正用它做的,是把研究自动化,尤其是 customer research(客户研究)。对,因为现在你可以在一天内完成那个 query(查询)和 survey(调研),而不是花一周。
Speaker 145:51 - 45:58
Yeah. Which is my consumer surplus point. You do the same spend, same survey, but you deliver it in a week instead of in a day instead of a week. Speaker 145:51 - 45:58
对,这就是我说的 consumer surplus(消费者剩余)那一点。你花的是同样的钱,做的是同样的 survey(调研),但你现在可以在一天内交付,而不是一周。
Speaker 245:58 - 46:07
Yeah. It makes sense. I mean, I guess given this discussion we've had on the foundation models, like if you were running one of these companies, like would you be doing anything different in terms of like focus or what they're Speaker 245:58 - 46:07
对,有道理。我是说,考虑到我们刚才关于 foundation models(基础模型)的讨论,如果是你在运营这类公司,你会不会在 focus(重点)之类的问题上做一些不同的选择,或者他们的——
Speaker 146:10 - 46:30
I have a lot of sympathy for Semmelman here. That may be a unique and controversial statement. You've got lot of plates to juggle because you have all I mean, you can make your list. In fact, you could ask Claude George at GPT to make your list of what are all the problems Do you make a device? Oops. Speaker 146:10 - 46:30
我对 Semmelman 这里的处境其实很能共情。这可能是个既独特又有争议的说法。因为你有很多盘子要同时转,要兼顾的事情特别多,毕竟你有——我是说,你自己都能列出一长串。事实上,你甚至可以让 Claude、George 或 GPT 给你列个单子,说说到底有哪些问题。你要不要做 device(设备)?哎,说漏了。
Speaker 146:31 - 46:45
Sorry about that. Okay. Let's go I'm gonna have to spend a lot of time on calls with lawyers now. How do you think about of course, OpenAI and Anthropic don't have infrastructure. So there's an infrastructure question. Speaker 146:31 - 46:45
抱歉,刚才那个。好,继续。我现在大概要花很多时间跟律师打电话了。你怎么理解这个问题?当然,OpenAI 和 Anthropic 都没有 infrastructure(基础设施),所以这里首先有个基础设施问题。
Speaker 146:45 - 46:50
How do we get infrastructure? How do we get the funding for the infrastructure? Do we make our own chips? Yes. Okay. Speaker 146:45 - 46:50
我们怎么拿到 infrastructure(基础设施)?我们怎么为这些基础设施拿到 funding(资金)?我们要不要自己造 chip(芯片)?要。好。
Speaker 146:50 - 47:10
How long is that gonna take? Work your way all the way through those layers of the stack. It's like you're trying to invent. It's like you're Bill Gates and you've got to invent, you've got to build PCs and enterprise software and broadband networks in like 1985 or 1980. You've got to so many different parts of that stack are completely unclear and all happening at once. Speaker 146:50 - 47:10
那这得花多久?你得把 stack(技术栈)的一层一层全都走通。这就像你在试图发明——就像你是 Bill Gates,而你必须在 1985 年或者 1980 年那会儿,同时去发明、去搭建 PC、enterprise software(企业软件)和 broadband networks(宽带网络)。你必须面对的是,这个 stack(技术栈)里有太多不同部分都完全不明朗,而且还在同一时间一起发生。
Speaker 147:11 - 47:19
There's a sort of, well, obviously, this is what you would do. And so you can see the two strategies. You can see the strategies across companies. So if you're an incumbent, you try and make the new thing a feature. Yeah. Speaker 147:11 - 47:19
这里有一种“嗯,显然,这就是你会去做的事”的感觉。所以你能看到两种策略。你也能看到各家公司都在采取这些策略。所以如果你是 incumbent(既有市场主导者),你会试着把这个新东西做成一个 feature(功能)。对。
Speaker 147:19 - 47:36
And so that's what we saw at Google. Spray the whole thing over everything is what you see it with. Probably mostly successfully is what you, if you're Microsoft, well, we don't have this capability. So we'll buy it, rent it, whatever, spread all over everything. One of the steps in enterprise AI deployment was like step one is deploy Copilot. Speaker 147:19 - 47:36
这就是我们在 Google 看到的做法。你看到的是把这整套东西喷洒到所有产品上,大概基本算是成功的。如果你是 Microsoft,那就是:好吧,我们自己没有这种能力。那我们就把它买下来、租下来,或者不管怎样,铺到所有东西上。企业 AI 部署中的一个步骤,几乎就像是第一步先部署 Copilot。
Speaker 147:36 - 47:56
Step two, oops, step three, let's do something else. If you are Apple, clearly they fumbled massively two years ago. If you've played with a beta of the new Siri, it's like, this is as good as ChatGePG was six months, nine months ago. And it's a consumer deployment and use case story there. If you are OpenAI or Anthropic, you don't have a legacy business, You don't have a way to make it a feature. Speaker 147:36 - 47:56
第二步,哎呀,第三步,再做点别的。如果你是 Apple,很明显他们两年前严重失误了。如果你玩过新版 Siri 的 beta(测试版),感觉就像:这也就跟 ChatGePG 六个月、九个月前差不多。这里讲的是一个面向消费者的部署和 use case(使用场景)故事。如果你是 OpenAI 或 Anthropic,你没有 legacy business(存量业务),你也没有办法把它做成一个 feature(功能)。
Speaker 147:56 - 48:11
You don't have distribution. You don't have infrastructure. What you do have is cutting edge science, but you've kind of gotta build all of the rest. And so I think what OpenAI did second half of last year was what can we build on top of a commodity foundation model? Is it a browser? Speaker 147:56 - 48:11
你没有 distribution(分发渠道)。你没有 infrastructure(基础设施)。你真正拥有的是 cutting edge science(前沿科学),但剩下的一切基本都得你自己建。所以我觉得 OpenAI 在去年下半年做的事情是:在一个 commodity foundation model(商品化的基础模型)之上,我们还能构建什么?是 browser(浏览器)吗?
Speaker 148:11 - 48:17
Is it a social video app? Is it an app store? Is it another app store? Is it a third app store? Is it shopping? Speaker 148:11 - 48:17
是 social video app(社交视频应用)吗?是 app store(应用商店)吗?是另一个 app store(应用商店)吗?还是第三个 app store(应用商店)?是 shopping(购物)吗?
Speaker 148:17 - 48:28
Is it ads? Everything. If you're philanthropic, less money, more focused, more worried about evil stuff, fine, whatever. Much more narrowly focused on coding. Stumble into coding is something that works, right? Speaker 148:17 - 48:28
是 ads(广告)吗?什么都试。如果你更 philanthropic(偏公益导向),钱更少,更专注,也更担心那些邪恶的东西,那也行,随便吧。那就会更狭窄地聚焦在 coding(编程)上。结果误打误撞进了 coding,而这件事居然是有效的,对吧?
Speaker 148:28 - 48:43
Okay, everyone else now rushes towards coding. It's like the scene in Monty Python with, in Life of Brian with Jova. It's like you're throwing rocks at coding. Coding! But is that the next thing? Speaker 148:28 - 48:43
好,于是现在其他所有人都开始冲向 coding。就像 Monty Python 的那个场景,在 Life of Brian 里 Jova 那段。感觉就像大家都在朝 coding 扔石头。Coding!但那会是下一个方向吗?
Speaker 148:43 - 48:55
I don't know. I mean, part of the other thing one could have said about coding is coding is actually a pretty small industry. I mean, it's weird whenever I point this out, people in tech get really upset with me. There's millions of people writing software. Fine. Speaker 148:43 - 48:55
我不知道。我的意思是,关于 coding,本来也可以提出另一点:coding 实际上是个相当小的行业。每次我指出这一点都很奇怪,tech(科技)圈的人会对我特别生气。是的,确实有几百万人在写软件。没问题。
Speaker 148:55 - 49:17
Imagine if we had a use case that really, really worked that like a billion people wanted to do. We couldn't. Like the infrastructure couldn't possibly support it. I mean, this is one of the interesting differences from every other previous platform shift is it's more expensive rather than cheap, It has marginal cost. And so you can't make a consumer app and get a 100,000,000 users and then work out the revenue model, which is why it's all enterprise as opposed to all consumer. Speaker 148:55 - 49:17
想象一下,如果我们有一个真的、真的行得通的 use case,而且大概会有十亿人都想用。我们也做不到。因为基础设施根本不可能支撑它。我的意思是,这和之前任何一次平台迁移(platform shift)相比,一个有趣的不同点在于:它不是更便宜,而是更贵;它有 marginal cost(边际成本)。所以你没法先做一个 consumer app(面向消费者的应用),拿到 100,000,000 用户,然后再去琢磨 revenue model(盈利模式);这也是为什么现在基本全是 enterprise(企业端),而不是 consumer(消费端)。
Speaker 149:17 - 50:03
But if you're running a model company, I mean, obviously OpenAI has had all sorts of weird political stuff going but you're shepherding your researchers, you're shepherding capital, you're shepherding infrastructure, you're herding cats in all sorts of directions. You've gotta work out, well, what is the path to sustainable competitive differentiation when every four to six weeks, there's another model at the top of the leaderboard? And there's absolutely no sign of what you would do that would mean you would have sustainable competitive differentiation. Do we have a path to network effects here? Can we build product layers on top of this that mean everyone has to use our product, even though the model underneath doesn't seem to be very different? Speaker 149:17 - 50:03
但如果你在经营一家 model company(模型公司),我的意思是,显然 OpenAI 一直有各种奇怪的政治问题在发生,但你要带着研究人员往前走,要统筹资本,要统筹基础设施,还得在各种方向上“herding cats”(费劲地协调一群很难管的人/事)。你必须想清楚:当每隔四到六周,leaderboard(排行榜)顶部就会换一个新模型时,通向可持续竞争差异化的路径到底是什么?而且完全看不出来,你做什么才能真正形成可持续的竞争差异化。这里会有 network effects(网络效应)可走吗?我们能不能在这之上构建 product layers(产品层),让所有人都必须使用我们的产品,尽管底层模型看起来并没有太大差别?
Speaker 150:04 - 50:27
Can we focus our model in on one capability and try and pull ahead in that capability? So you could argue what Claude has done is focused on getting coding really, really good. And maybe they have a competitive lead there. But you also have to be continually conscious that everything is changing every three to six months. No one really knows what this is gonna look like in three or four years time. Speaker 150:04 - 50:27
我们能不能把模型聚焦在某一种能力上,并试着在这项能力上拉开差距?所以你可以说,Claude 所做的就是专注于把 coding(编程)做到非常、非常好。也许他们在这方面确实有竞争领先优势。但你也必须时刻意识到,一切都在每三到六个月持续变化。没有人真正知道,三四年之后这会是什么样子。
Speaker 150:27 - 50:29
And no one really knows what the building blocks are. So do Speaker 150:27 - 50:29
而且也没有人真正知道,底层的 building blocks(构件)到底是什么。所以你
Speaker 250:29 - 50:33
think that running all these experiments makes sense if you're OpenAI just given how uncertain this all is? Speaker 250:29 - 50:33
觉得,考虑到这一切都这么不确定,如果你是 OpenAI,去跑所有这些实验是有意义的吗?
Speaker 150:33 - 50:58
I think it well, the problem is based strategies make sense. I mean, I think if you're opening AI and you've got all this capital, it kind of made sense to do or try all of that stuff. If you're in Swapic and you had less capital and you're more worried about like this take off and evil stuff happening, then it made more sense just to focus in narrowly on coding. We didn't know that coding was gonna work. I mean, we mean, maybe you could have said in hindsight, obviously it was gonna work, but clearly like it flipped at the beginning of this year from kind of working to actually working. Speaker 150:33 - 50:58
我觉得——嗯,问题在于,基于平台的 strategies(战略)是讲得通的。我的意思是,我认为如果你是 OpenAI,而且你手里有这么多资本,那去做或者尝试所有那些事情,某种程度上是合理的。如果你是 Swapic,资本更少,而且你更担心这种快速起飞以及 evil stuff(糟糕后果)会发生,那更合理的做法就是把焦点很窄地放在 coding 上。我们当时并不知道 coding 会奏效。我的意思是,事后看你也许会说,显然它本来就会奏效,但很明显,今年年初它确实发生了一个转变:从“有点能用”变成了“真的能用”。
Speaker 151:00 - 51:28
And again, we've had this conversation, was that inevitable? Was that the thing? But you come back to this question like the thing that sort of fascinated me about OpenAI in the last couple of days is the new app launches. And if it tells you one thing, it's like writing the code isn't the hard part because it's such a disastrous, chaotic, confused mess. And there's like this weird, like, we're gonna give everybody copilot and also work and work is different to chat. Speaker 151:00 - 51:28
而且,我们之前也讨论过这个问题:那是必然的吗?那就是关键所在吗?但你又会回到这个问题上——过去这几天里,OpenAI 最让我着迷的一点,其实是这些新 app(应用)的发布。如果这能说明一件事,那就是写出代码并不是最难的部分,因为它整个呈现出来的是一种灾难性的、混乱的、令人困惑的大杂烩。还有一种很奇怪的感觉,好像是:“我们要给每个人都提供 copilot,同时还有 work,而 work 又和 chat 不一样。”
Speaker 151:28 - 51:42
And it's called chat, but chat's hidden. And when you do chat, then it appears in this weird pop up, but you can't use this in that. And then you hide it And then this happens and this works on cloud and that doesn't work. I'm like, really? How did you ship that? Speaker 151:28 - 51:42
它叫 chat,但 chat 又被藏起来了。然后当你进入 chat 时,它会出现在一个奇怪的弹出窗口里,但你不能在这个里面用那个。接着你把它隐藏掉,然后这个会发生,那个会起作用;这个在 cloud 上能用,那个又不能用。我就想,真的吗?你们怎么会把这种东西 ship(发布)出去?
Speaker 151:42 - 52:05
How did that happen? There's a step back from all of this, which is like before Microsoft and Google and Meta and so on had the winner takes all effects, they had to execute their way into that. I mean, you kind of knew that what the winner takes all effects would be completely obvious. Like remember when people thought that maybe like social would win in different countries, like Bebo would win in Europe Facebook would in America and Uber would win. It was like, like rideshare was gonna be city by city. Speaker 151:42 - 52:05
这是怎么发生的?要把这一切往回退一步看:在 Microsoft、Google、Meta 等公司拥有赢家通吃效应之前,它们必须先靠执行把自己带到那个位置。我的意思是,你大致知道赢家通吃效应会出现,这是很明显的。比如还记得以前人们会想,也许社交产品会在不同国家各自取胜,像 Bebo 会赢下 Europe,Facebook 会赢下 America;而 Uber 那边,人们也觉得 rideshare 会是一座城市一座城市地打下来。
Speaker 152:05 - 52:24
And so you knew what the, you kind of knew what the network effect was, but you weren't sure, but also you had to execute your way into getting it. And remember that like Myspace was there first. And remember that when Facebook was happening, people, Google did wave, buzz thing, other thing. And Google has all the data. So there's an execution. Speaker 152:05 - 52:24
所以你知道那个 network effect(网络效应)大概是什么,但你并不完全确定;而且你还必须靠执行,才能真正把它拿到手。别忘了,Myspace 是先出现的。也别忘了,Facebook 崛起的时候,Google 也做过 wave、buzz 那些东西。Google 还拥有所有数据。所以这里面有一个执行层面的故事。
Speaker 152:24 - 52:37
You can get very deterministic about, well, I want like the network effects and commoditization and so on. But there is an execution story there as well. Like, you actually have to, like, make the thing happen. Yeah. It's like, you know, Marx said history is not a material person. Speaker 152:24 - 52:37
你可以非常确定地谈论:我想要 network effects、commoditization(商品化)等等。但这里同样也有一个执行的故事。就是说,你真的必须把这件事做成。对吧。这就像 Marx 说的,history 不是一个具体的人。
Speaker 152:37 - 52:46
Like, history isn't a person. Yeah. So you get Marx's whole thing about history works like this, and these are inevitable forces. And then he says, yeah, but but history isn't a person. So even somebody actually has to do it. Speaker 152:37 - 52:46
就是说,history 不是一个人。对。于是你会得到 Marx 的整套观点:history 是这样运作的,这些力量是不可避免的。然后他又说,是啊,但 history 不是一个人。所以终究还是得有人真的去把它做出来。
Speaker 252:46 - 52:47
Yeah. So you have Speaker 252:46 - 52:47
对。所以你必须
Speaker 152:47 - 52:48
to actually make that thing happen. Speaker 152:47 - 52:48
真的让那件事发生。
Speaker 252:49 - 52:57
Yeah. No. I love that. Mean, I guess, you know, on on the opening eye side, think, you know, they have obviously they had all these med execs come in and and I think you've talked about how they kinda tried to run a similar playbook. Speaker 252:49 - 52:57
对,不,我很喜欢这个说法。我的意思是,我猜在 opening eye 那边,你知道,他们显然找来了很多 med execs,我觉得你之前也谈到过,他们某种程度上试图照着类似的 playbook(打法手册)来运行。
Speaker 152:57 - 52:59
Yeah. It felt a bit cargo cult What Do you Speaker 152:57 - 52:59
对,感觉有点像 cargo cult(货物崇拜)。你怎么看?
Speaker 252:59 - 53:01
know what a cult went wrong there? Speaker 252:59 - 53:01
知道吗,那里出了什么“cult(狂热崇拜)”式的问题?
Speaker 153:01 - 53:03
Do you know what a cargo cult is? I was never Is this what this reference? Speaker 153:01 - 53:03
你知道什么是 cargo cult 吗?我以前从来没有——这说的是不是这个典故?
Speaker 253:04 - 53:07
It's something I've said a bunch and I don't actually think I have any idea what the origin is. Speaker 253:04 - 53:07
这是我经常会说的一句话,但说实话,我其实也不太知道它的起源到底是什么。
Speaker 153:07 - 53:44
It's a World War II thing. So, there's war in The Pacific, there are all these like islands in the Pacific where like people have seen a European ship every six months or something and every year or something, but no one, these are not people who are exposed to modern mass produced industrialized society. And the Americans arrive and they build an air base and a base and a ship and a port and a pier and there's all this stuff. And like America in particular, like there's infinite money and so all the supplies, they handle the stuff out and like canned food and like souffles and like everything you want. And then the war ends and like the Americans just leave. Speaker 153:07 - 53:44
这是个 World War II 时期的事。那时候太平洋在打仗,太平洋上有很多这样的岛屿,岛上的人可能每隔六个月、或者一年,才见到一次 European 的船,但总之,这些人并不是那种接触过现代、大规模工业化生产社会的人。后来 Americans 来了,建了 air base、基地、船坞、港口、码头,弄出了这一大堆东西。而且尤其是 America,简直像有花不完的钱一样,物资源源不断地运来,到处分发,canned food 之类的,还有各种你想要的东西。然后战争结束了,Americans 就直接走了。
Speaker 153:45 - 54:07
There's this island somewhere in the Pacific that like there's like 10,000 airplanes and trucks and stuff just dumped into a lagoon because there was just it wasn't worth shipping it all back to America and scrapping it. So there's like, I'm slightly I'm only slightly exaggerating. There's like hundreds of brand new trucks and aircraft dumped in this lagoon. And so the point is, suddenly the airfield's empty and there's no more planes coming. So what do you do? Speaker 153:45 - 54:07
太平洋某个岛上有个地方,据说有差不多 10,000 架飞机、卡车之类的东西,被直接倒进了一个 lagoon,因为把这些东西运回 America 再拆解处理根本不划算。所以那里——我虽然有点夸张,但也只是稍微夸张一点——真的有几百辆全新的卡车和飞机被扔进了这个 lagoon。重点是,突然之间,airfield 空了,再也没有飞机飞来了。那你怎么办?
Speaker 154:07 - 54:30
Well, you go up into the control tower and you speak and you line up in parade and you run a flag up the flagpole because that's what the Americans did and the plane came. I mean, is absolutely true. This is this phenomenon in the fifties well, late forties and fifties. If you didn't actually understand what it was that was causing the cargo to come, so you imitate the forms. Speaker 154:07 - 54:30
那你就走上 control tower,开口讲话,排队列队,把旗子升上旗杆,因为 Americans 当时就是这么做的,然后飞机就来了。我的意思是,这事绝对是真的。四十年代末到五十年代,确实有这种现象。因为如果你并不真正理解到底是什么让 cargo(货物)运来的,你就会去模仿那些形式。
Speaker 254:30 - 54:31
Interesting. Speaker 254:30 - 54:31
有意思。
Speaker 154:31 - 55:00
And I'm going down the rabbit hole again, but the point here is like, joined Meta in 2010, way after it had product market fit. And you're at And so what do you do? Well, you make an app store. You build an ad business, you build an e commerce thing and you know how to You're doing all of those playbooks, but you're not at the part where you know what works. And what you actually want is the people who are meta when it was just Harvard. Speaker 154:31 - 55:00
我又开始越说越偏了,但这里的重点是:有人在 2010 年加入 Meta,那已经是它早就实现 product-market fit 之后的事了。你身处那样的环境,那你会做什么?你会做一个 app store,建立一个 ad business,做一个 e-commerce 业务,你知道该怎么把这些 playbook(一整套成熟打法)一项项做出来。但问题是,你所处的阶段并不是那个“已经知道什么有效”的阶段。你真正想要的,其实是那些在 Meta 还只是 Harvard 的时候就在那儿的人。
Speaker 155:00 - 55:04
Yeah. Not the people who are meta when 5,000,000,000 people were using it. Speaker 155:00 - 55:04
对。不是说在 5,000,000,000 人都在用它的时候,那些本来就在 Meta 的人。
Speaker 255:04 - 55:33
And obviously on the consumer side, I feel like you've talked about how these products have some relatively shallow usage today across the board. I think one of the hard parts that you've discussed before is if you're at one of these labs and you're on the product team, you're kind of waiting for model capabilities to figure out what to go build, and so it's a hard way to build product. And then at the same time, I can't tell whether we're just a few model capabilities away from a completely different set of consumer usage. The killer ones always felt like computer use. If you really got computer use to work, would people just use these products way more? Speaker 255:04 - 55:33
很明显,在消费者这一侧,我感觉你一直在谈这些产品如今整体上的使用都还比较浅。我觉得你之前谈过的一个难点是,如果你在这些 labs 之一、而且你在产品团队里,你某种程度上是在等 model(模型)能力发展出来,再决定到底该做什么产品,所以这是种很难的产品构建方式。同时,我也说不好,我们是不是其实只差少数几个 model 能力,就会迎来一套完全不同的消费者使用模式。那些最有杀手级潜力的场景似乎一直都像是 computer use(计算机使用)。如果你真的把 computer use 做通了,人们会不会就大幅更多地使用这些产品?
Speaker 155:33 - 56:00
I'm not sure about that. Just remember how many people, for how many people their main device is a smartphone. Most people don't use a computer as much as they use a smartphone. So, a little bit hesitant about that one. It does feel very geeky to me, even setting aside security and failure rates, and it might accidentally delete all of your stuff, which again is like, you know, I was using I mean, it's this point about PCs in the early eighties. Speaker 155:33 - 56:00
我不太确定。你只要想想,有多少人的主设备是 smartphone。大多数人使用 computer 的频率并没有高过使用 smartphone。所以,对这一点我会有点犹豫。它给我的感觉还是很 geeky,即便先不谈安全性和失败率,也不谈它可能会误删你所有东西这件事——这又有点像,你知道的,我当时在用,我是说,这就像在讲八十年代早期的 PC。
Speaker 156:00 - 56:22
It's like it was completely normal that you'd be working, even in the mid nineties, late nineties, you'd be working and you look up at your screen and you realize it's frozen and nothing moves. And the your solution is you crawl under your desk and you unplug your PC, and then you wait for it to turn on, and then you hope that you haven't lost more than an hour of work. Yeah. That was just normal, completely normal. I always had and that's that's exactly the same now of, you know, I I told coworker to clean up my inbox, and it deleted my inbox. Speaker 156:00 - 56:22
就像那时候,哪怕到了九十年代中期、后期,你正在工作,抬头一看屏幕,发现它卡死了,什么都不动。而你的解决办法,就是钻到桌子底下,把 PC 电源拔掉,然后等它重新开机,再祈祷自己丢掉的工作别超过一个小时。对,那当时就是常态,完全是常态。我一直觉得——而现在其实也一模一样:比如我让 coworker 帮我清理 inbox,结果它把我的 inbox 给删了。
Speaker 156:22 - 56:34
Hey. It's cleaned up. Great. So the the the the the consumer like, the takeoff here is, like, again, are unfalsifiable statements. Is it just that the models will get better? Speaker 156:22 - 56:34
嘿,确实是“清理”干净了。很好。所以,消费者这边的 takeoff(起飞、爆发)判断,归根结底又像是一些无法证伪的说法:是不是只要 models(模型)继续变好就行?
Speaker 156:35 - 56:58
Or is it that that's not quite what the problem is? Is it that most people don't very often have the kind of things that mesh very well with this? Set aside the enterprise side of like, you're seeing an SAP, it's a pain in the ass to do this thing. Great, you can't use ChatGPT for that because it's not authorized, it's not plugged in, you're not allowed to do that, fine. That's a CIO conversation. Speaker 156:35 - 56:58
还是说,问题并不完全在这里?是不是大多数人其实并不经常遇到那种和这类东西特别契合的任务?先把 enterprise 那边放在一边——比如你在 SAP 里看到,有件事做起来特别折磨人。好,那你没法用 ChatGPT 来做,因为它没有授权、没有接进去、你也不被允许这么做,行。这是个 CIO 会讨论的问题。
Speaker 156:58 - 57:25
You're a consumer, How do you work out what this is? How do you bridge it? You open chat GP to And this, I don't think I know the answer here. I mean, obviously I don't know the answer, but like I'm ambivalent here because I remember like portals on the internet in the late nineties, where you give people a web browser and you can't just give them a browser with a URL bar and a blank screen. You've gotta help them. Speaker 156:58 - 57:25
但如果你是个普通消费者,你要怎么弄明白这到底是什么?你要怎么把它连接起来?你打开 chat GP to——而这一点,我不觉得我知道答案。我的意思是,我显然不知道答案,但我在这件事上是摇摆的,因为我记得九十年代后期互联网上的 portals(门户网站):你把一个 web browser 给用户时,不能只是给他们一个带 URL 栏和一片空白页面的 browser。你得帮他们一把。
Speaker 157:25 - 57:55
And eventually you didn't need to help them. And some of that is that wasn't really about broadband or better tech, that was about the ecosystem kind of maturing and people forming new habits. And so when you see you open a chatbot now and it's got these little tiles of like, you could do this, you could do this, you could do this. Is that like the portal then of like the handholding of helping people work out what to do with this? I don't know an awful lot of what you were doing on the internet with stuff you're already doing. Speaker 157:25 - 57:55
而到了后来,你又不再需要帮他们了。这其中有一部分其实并不真是 broadband 或更好技术的问题,而是 ecosystem(生态系统)逐渐成熟、人们形成了新的习惯。所以现在当你看到,你打开一个 chatbot(聊天机器人),里面有这些小 tiles(卡片),写着你可以做这个、你可以做这个、你可以做这个——这是不是就像当年的 portal,那种手把手带着用户理解这东西能拿来干什么的方式?我不太确定,因为你在 internet 上做的很多事,本来就是你已经在做的事。
Speaker 157:56 - 58:08
Like now buy an air conditioner, book a holiday, buy this thing, find this thing, watch that video. It's stuff you kind of already knew. It's like the spreadsheet point. Like you've already knew about spreadsheets. The thing you already knew you could do. Speaker 157:56 - 58:08
比如现在去买一台 air conditioner,预订一个 holiday,买这个东西,找那个东西,看那段视频。这些基本上都是你本来就已经知道要做的事。这就像 spreadsheet 那个点。就像你本来就知道 spreadsheet 能做什么。也就是那些你原本就知道自己可以做的事。
Speaker 158:08 - 58:39
How much of what you do with an LLM is stuff that you're already doing and you realize, ah, I could do that much better with this. This is obviously the search conversation and the booking holiday conversation or whatever, or my point about find me a coat. How much of it is the other side of that, that an entrepreneur is going to realize, Ah, no, this is what you do. I mean, is Flickr and Instagram and TikTok. That's not like a consumer thinking, Oh, I want a way to share pictures. Speaker 158:08 - 58:39
你用 LLM 时,所做的事情里有多少是你本来就在做的,只是你意识到,啊,我可以用这个把它做得好得多?这显然就是 search 对话、预订 holiday 的对话,或者我说的“帮我找一件 coat”这类事。有多少又是另一面——也就是某个 entrepreneur 会意识到,啊,不,不是这样,这才是你该做的东西。我的意思是,Flickr、Instagram 和 TikTok 就属于这种。这并不是消费者自己在想:“哦,我想要一种分享图片的方式。”
Speaker 158:40 - 58:51
Somebody has to invent that. That's not organically emergent behavior out of consumers. It's not like you just give enough everybody enough broadband and they'll start sharing pictures. Well, I'm talking about in Speaker 158:40 - 58:51
这得有人去发明出来。这不是消费者自然涌现出来的行为。不是说只要给每个人足够的 broadband,他们就会开始分享图片。嗯,我这里说的是在
Speaker 258:51 - 58:59
the enterprise side, you've obviously seen people do this with the labs that we need, these like deployment companies, and like the idea of like going out there and like, do you think that will be successful? Speaker 258:51 - 58:59
enterprise 这一侧,你显然已经看到有人通过我们需要的这些 labs 来做这件事,这类 deployment 公司之类的,还有那种走出去推进落地的思路——你觉得这会成功吗?
Speaker 158:59 - 59:13
So again, we're going in lots of different directions. Like step one of enterprise deployment was give everyone a copilot. Step two was, oh crap, that didn't work. For like a small portion of people, like if you've given them Claude, maybe great. Like my wife lives in Claude, fine. Speaker 158:59 - 59:13
所以还是那句话,我们现在是在朝很多不同方向走。enterprise deployment 的第一步是给每个人一个 copilot。第二步则是,哦,糟糕,这没奏效。对少部分人来说,比如如果你给他们 Claude,也许确实很好。比如我妻子基本就“住”在 Claude 里,没问题。
Speaker 159:13 - 59:37
But most people know, and for the reasons we've talked about, that doesn't make sense. Step two is pilots and everyone has done now, done a whole bunch of pilots and deployed about half of them and they work fine, but that's kind of a one at a time thing. And then the conversation becomes like, is that it? That feels like that's not it. Well, we automated this one pain point one at a time. Speaker 159:13 - 59:37
但对大多数人来说不是,而且原因就是我们刚才谈过的那些——那样做没有意义。第二步就是 pilots,而现在大家都已经做了很多 pilots,也部署了其中大约一半,效果都还不错,但那基本上是一种一次做一个的方式。然后话题就变成了:难道就这样了吗?感觉不该只是这样。好吧,我们只是把这个 pain point(痛点)一个一个地自动化了。
Speaker 159:37 - 59:52
Yeah. It feels like that's not the right way to think about this. How do we actually change structurally how we do things around this new technology? We're still sort of at the point of like we successfully PDFed all of our catalogs and put them on our website. Great. Speaker 159:37 - 59:52
对,感觉那不是思考这件事的正确方式。我们到底该怎样围绕这项新技术,从结构上改变我们的做事方式?我们现在某种程度上还停留在这样一个阶段:我们成功把所有 catalog 都做成了 PDF,然后放到了网站上。很好。
Speaker 159:52 - 59:56
That worked. How many people have downloaded them? Loads. What's our ROI? Our printing bills collapsed. Speaker 159:52 - 59:56
这确实奏效了。有多少人下载了?很多。我们的 ROI 是什么?我们的 printing 成本暴跌了。
Speaker 1 | 59:56 - 1:00:03 We've got way more customers. Great. Okay. But that's probably not the end point. And so that's what all of these things are.
我们现在有更多得多的客户。很好,没问题。但这大概还不是终点。所以这些事情的意义就在这里。
Speaker 1 | 1:00:03 - 1:00:36 And this is of course why consulting exists, Because most people do not have those skills and they don't have it set up as a process and they don't have all those people sitting around not doing anything. And so you hire Bain or BCG or McKinsey, or you hire Accenture, IBM, Cognizant, Deloitte, PwC, and so on. Or you hire WPP Publicis and so on, or you hire Edelman or whoever it is for these different kinds of problems and they come to you and help you work it out. This is why or lawyers. This this is why professional services exist because you don't have all of that that talent in house.
当然,这也是为什么 consulting(咨询)会存在,因为大多数人并不具备这些技能,也没有把这些能力建立成一套流程,而且他们也不可能养着一群人闲坐着等事做。所以你会雇 Bain、BCG 或 McKinsey,或者雇 Accenture、IBM、Cognizant、Deloitte、PwC 等等。又或者你会雇 WPP、Publicis 之类,或者雇 Edelman,或者其他任何适合处理这类不同问题的机构,他们会来帮你把事情理清楚。这就是为什么——还有律师——为什么 professional services(专业服务)会存在,因为你并没有把所有这些人才都放在公司内部。
Speaker 1 | 1:00:36 - 1:00:56 But working out what all of that reimagination would be is a project. It's a big difficult project. This is what used to be called digital transformation and now enterprise transformation. I always used to joke that if you say digital transformation three times, then a partner from Accenture will appear in a puff of smoke. It's like my most reliable laugh line at conferences, especially enterprise conferences.
但要弄清楚所有这些重新想象到底应该是什么,本身就是一个 project(项目)。这是个又大又难的项目。这以前叫 digital transformation(数字化转型),现在则叫 enterprise transformation(企业转型)。我以前总爱开玩笑说,如果你把 digital transformation 连说三遍,就会有一个 Accenture 的合伙人从一团烟雾里现身。这几乎是我在 conference(会议)上最稳定的笑点,尤其是在 enterprise conference(企业会议)上。
Speaker 1 | 1:00:56 - 1:01:41 Like the whole conference grinds for halt for thirty seconds after I say that. But like, yeah, there's a reason why they've got like 800,000 people and that will probably go to half of it, that will probably half or whatever the number is, but like, no company is going to have all of those skills. They're just sitting around ready to reimagine stuff. The one thing that occurred to me there though is that there's a sort of symmetry or an overlap between what an enterprise software company does, a vertical enterprise software particularly does, and what a consulting firm does, in that what they're And then a strategy consulting firm, management consulting like Bain, BCG, McKinsey, and so on. What they're doing is they're kind of going and looking at your company and saying, you're kind of doing it like this, but actually you could do it like this and that would work way better.
比如我一说完这个,整个 conference(会议)都会停顿差不多三十秒。不过,确实,他们之所以会有大概 80 万员工是有原因的,当然这个数字可能会减半,或者变成别的什么数,但重点是,没有哪家公司会拥有所有这些技能的人才,还让他们坐在那里随时准备去重新构想各种事情。不过我想到的一点是,enterprise software company(企业软件公司)——尤其是 vertical enterprise software(垂直型企业软件)——在做的事,和 consulting firm(咨询公司)在做的事之间,其实有某种对称性,或者说重叠。再加上 strategy consulting firm、management consulting,像 Bain、BCG、McKinsey 这些,它们本质上是在走进你的公司后对你说:你现在大概是这样做的,但其实你可以那样做,而那样会好得多。
Speaker 2 | 1:01:41 - 1:01:44 A percent. Mean, that's what a lot of these vertical AI companies are doing today, right? You have to have
百分之百。我的意思是,这不就是今天很多 vertical AI company(垂直 AI 公司)正在做的事吗?你必须得有——
Speaker 1 | 1:01:44 - 1:01:45 it, they
对,他们——
Speaker 2 | 1:01:45 - 1:01:47 call this forward deployed, which is reinvention of.
他们把这个叫作 forward deployed,这就是一种……的再发明。
Speaker 1 | 1:01:47 - 1:02:23 But that's what any vertical software company's doing. Well, at least half of them, that's what they're doing. They're saying, well, you were doing it like this and we've worked out, you could do it like this and that would work way better. And that's also what Bain does, McKinsey or does, or McKibbCT does, but just in a kind of a different modality, which is why all of these companies are sort of sitting and scratching their heads and trying to work out like, I mean, you know, there was a joke that a machine learning scientist is a statistician who lives in Silicon Valley. And I said, must be I haven't quite worked out the right equivalent for a forward deployed engineer, is like an Accenture body shopper who got a job at OpenAI.
但这其实就是任何 vertical software company(垂直软件公司)在做的事。至少有一半是这样。它们会说,你原来是这样做的,而我们已经想清楚了,你其实可以改成那样做,而且效果会好得多。这也是 Bain 在做的事,也是 McKinsey 在做的事,或者说 McKibbCT 在做的事,只不过采取的是另一种模式。这也就是为什么所有这些公司现在都有点坐在那里挠头,想搞明白这到底算什么。我的意思是,你知道,之前有个笑话说,machine learning scientist(机器学习科学家)就是住在 Silicon Valley 的 statistician(统计学家)。而我会说,我还没完全想出 forward deployed engineer(前驻工程师)最准确的对应说法,不过大概就像是一个跳槽去了 OpenAI 的 Accenture body shopper。
Speaker 1 | 1:02:23 - 1:02:37 But they're climbing the mountain from opposite directions. It's like, do you want a software company that's got a bunch of forwarded deployed engineers or do you want a consulting company that's got a bunch of software engineers? Because they're kind of doing the same thing from different directions.
但他们是在从相反的方向爬同一座山。就像是在说,你想要的是一家拥有一大堆被外派部署的 engineers 的 software company,还是一家拥有一大堆 software engineers 的 consulting company?因为他们其实是在从不同方向做同一件事。
Speaker 2 | 1:02:37 - 1:03:00 I think this point of basically, obviously, the enterprise, people you you don't just give people these models and they figure out exactly how to use them, and there's needed to be this big translation layer, and then I think on the consumer side, it's certainly been that to date, and I think the question to the point you're making on Instagram earlier is like, will some great entrepreneurs come along or maybe within these companies and find a way that's a little more that has some latent behavior within people that they wanna do, that the model Yeah.
我觉得这个点基本上是,很明显,在 enterprise 端,你不能只是把这些模型丢给别人,然后他们就自己搞清楚该怎么用它们;中间需要有一个很大的 translation layer(转换层)。然后我觉得在 consumer 端,至少到目前为止也确实是这样。我想,回到你刚才提到 Instagram 的那个点,问题就在于:会不会有一些很厉害的 entrepreneurs 出现,或者也许就在这些公司内部,找到一种方式,更贴合人们本来就潜在想做的某种行为,而模型正好能承接这种行为。对。
Speaker 1 | 1:03:00 - 1:03:17 Everything has to actually be invented. All of the use cases have to be invented. Very few of them are just purely spontaneous grassroots consumer behavior. All, every app on our phone, somebody had to think, well, would be a good idea. Then everybody looked at it and said, that's the dumbest thing I've ever heard.
所有东西其实都得被发明出来。所有 use case(使用场景)都得被发明出来。真正纯粹自发、草根式的 consumer 行为其实很少。我们手机上的每个 app,都得先有人想到:“这可能是个好主意。” 然后所有人看到后都说,这听起来是我听过最蠢的东西。
Speaker 1 | 1:03:17 - 1:03:29 I mean, remember how dumb Uber looked, remember how dumb Instacart looked. I was in the pitch, A16z for Instacart, like, will the economics work? Does this make sense? Will people do this? I mean, you know the story about, what's his name?
我的意思是,想想 Uber 当时看起来有多蠢,想想 Instacart 当时看起来有多蠢。我当时就在 A16z 听 Instacart 的 pitch,大家会想:这个 economics(经济性)能成立吗?这事说得通吗?人们真的会这么做吗?我是说,你知道那个故事吧,那个谁来着?
Speaker 1 | 1:03:29 - 1:03:43 The founder, that he didn't have a car. And so he set up this bare bones website that emailed him orders and then he would get an Uber and go to the supermarket and do the shop and get an Uber to people's home to deliver them.
就是那个 founder,他自己没有车。所以他搭了一个极简的网站,用户下单后会直接发邮件给他,然后他就打 Uber 去超市采购,再打 Uber 去别人家里把东西送过去。
Speaker 2 | 1:03:43 - 1:03:44 Tough for the economics.
对 economics 来说挺难的。
Speaker 1 | 1:03:44 - 1:03:46 To work out whether this would work.
很难判断这到底能不能跑通。
Speaker 2 | 1:03:46 - 1:03:50 Yeah. That's hilarious. Well, what do you make of Sora, like some of these early consumer app drives?
对,太好笑了。那你怎么看 Sora,或者说这些早期由 consumer app 推动的东西?
Speaker 1 | 1:03:50 - 1:04:38 So, I think Sora fits into a classic pattern of consumer social, which is you have to come up with some new mechanics, some new piece of masculine hierarchy, some new way of feeling that gets attention, but then it has to be more than a gimmick. And TikTok found that in one form and Snap found that and name another one. It was really hard actually. Snap from, we've found two or three, like disappearing messages and stories and so on, but like it's work. You could kind of argue that social has flattened out at the top of the S curve the way smartphones did or the way PCs did, that like everything's been invented and like saw was like, I don't know, it was like, all the stupid things.
所以,我觉得 Sora 符合 consumer social 的一种经典模式:你必须提出一些新机制、某种新的 masculine hierarchy(男性化等级秩序),或者某种能吸引注意力的新感觉;但它又不能只是个 gimmick(噱头)。TikTok 以一种形式找到了这个点,Snap 也找到了,你还能举出别的例子。其实这非常难。就 Snap 而言,我们大概只找到了两三个点,比如 disappearing messages(阅后即焚消息)、stories(故事)之类,但这都是很费功夫的。你某种程度上可以说,social 已经像 smartphones 或 PCs 那样,在 S 曲线顶部趋于平坦了——好像该发明的都已经发明完了,而 Sora 就像,我不知道,有点像那些各种傻乎乎的东西。
Speaker 1 | 1:04:38 - 1:05:02 Like the weird experiments people tried with smartphones, like modular smartphones and things like, you clip on this or clip on that, like a second screen or ink screen on the back. There's a moment where like, okay, now all of this social stuff has been done. And then of course, Dator comes along and does something else. But I think that's the frame that I would fit it into that narrative rather than specifically an AI narrative. It was another attempt at a cool social thing.
就像人们当年围绕 smartphones 做过的那些奇怪实验,比如 modular smartphones(模块化手机),或者给手机夹上这个、夹上那个,比如第二块屏幕,或者背面的墨水屏之类。有那么一个时刻会让人觉得:好吧,social 这套东西现在都已经做过了。然后当然,Dator 又冒出来做了点别的。但我觉得,我会把它放进这样一个叙事框架里,而不是特指 AI 的叙事。它更像是又一次想做一个很酷的社交产品的尝试。
Speaker 1 | 1:05:02 - 1:05:12 It turned out that it kind of didn't work. Mean, it's the same thing with Midjourney. You remember like a year or two ago, we all spent a month playing with Midjourney. Yeah. And now Meta's got a new image model, how I tried it.
结果证明,它某种程度上并没有奏效。我的意思是,Midjourney 也是一样。你还记得一两年前吗?我们所有人都花了一个月玩 Midjourney。对吧。现在 Meta 又出了一个新的 image model(图像模型),我都没去试。
Speaker 1 | 1:05:15 - 1:05:36 The image generation, mean, there's another point to make here, which is what happened with drones in three d printing, which is one Christmas, whenever it was ten years ago, we were bought a drone for Christmas. And then three days later we say, okay, I've seen the roof of my house now. Yeah. It's very cool. Same with, like, three d printers.
说到 image generation(图像生成),这里还有另一个点可以讲,就是 drones 和 3D printing 发生过的事:十年前的某个圣诞节,不管具体是哪一年,大家都买了个 drone 当圣诞礼物。然后三天后我们就说:好吧,我现在已经看过我家屋顶了。对。确实很酷。3D printers 也是一样。
Speaker 1 | 1:05:36 - 1:05:42 Okay. Made a little Eiffel Tower. Like, there's no consumer use case for three d printing. There is no consumer use case for drones. There's a hobbyist case.
好吧,做了个小 Eiffel Tower。就像是,3D printing 并没有 consumer use case(消费者使用场景)。drones 也没有 consumer use case。它们有的是 hobbyist case(爱好者场景)。
Speaker 1 | 1:05:42 - 1:05:59 There's a bunch of people who think it's really cool. There was this brief idea, like there's not a thing that everyone has. And so I think that was kind of where video creation ended up. Now somebody may need to pull the whole thing inside out. Now like virtual try ons, if you could get virtual try ons working, that would be image generation, but it wouldn't be mid journey.
有一批人觉得这东西特别酷。曾经短暂有过一种想法,好像这会变成人人都有的东西,但事实并不是。所以我觉得 video creation 最后有点也落到了这个位置。当然,也许有人能把整件事彻底翻过来。再比如 virtual try-ons(虚拟试穿),如果你真能把 virtual try-ons 做出来,那会用到 image generation,但它就不是 Midjourney 那种东西了。
Speaker 1 | 1:05:59 - 1:06:22 Yeah. You know, here's, look at my Instagram and go look at the latest look book from this brand, make 10 videos of me in different looks. I'm like, that's probably completely insane, but like that a product. Would Yeah. Midjourney wasn't a product in the sense that you had to log into it with Discord, but it also wasn't like It had the same problem as Drone.
对。你知道的,比如“去看看我的 Instagram,再看看这个品牌最新的 look book,给我生成 10 个我穿不同造型的视频”。我觉得这听起来可能完全疯了,但那倒像是个 product(产品)。会的,对。Midjourney 不是那种意义上的 product,因为你还得通过 Discord 登录,但它也不只是这个问题,它和 drone 遇到的是同样的问题。
Speaker 1 | 1:06:22 - 1:06:27 It's like it didn't translate into a tangible specific thing that you would do that solved a problem.
也就是说,它没法转化成一个具体、可感知、能解决某个问题的明确事情。
Speaker 2 | 1:06:27 - 1:06:41 Yeah. No, super interesting. Play that's a common theme throughout our conversation. I always like to end with kind of a quick fire round where I just stuff a bunch of things into the end that we haven't discussed. And so maybe to start, I'm curious, I feel like you've been pretty consistent in your view on this technology.
对。不是,特别有意思。这其实是我们整场对话里反复出现的一个共同主题。我总喜欢在结尾来一个快问快答,把前面还没聊到的一堆东西都塞进来。所以也许先从这里开始,我很好奇,我觉得你对这项技术的看法一直都相当一致。
Speaker 2 | 1:06:42 - 1:06:46 What's one thing you've changed your mind on in the last year around AI? The basic thing
过去一年里,关于 AI,你在哪一件事上改变了看法?最根本的那个
Speaker 1 | 1:06:46 - 1:07:28 mistake that I still see people making is not to understand that this is an enabling technology that makes lots of different things possible and still look at it as something that makes text and pictures and not understand, no, this will like make fraud detection work better, or this will solve all sorts of like obscure back office processes inside companies you use every day that you've never thought of. And so you'll get all sorts of better experiences and better products that won't come to you looking as AI and won't come to you saying this was ChatGPT. And you still get people kind of saying, but you know, I asked it a question and the answer wasn't quite right. Well, great. It's like looking at the Internet and saying it's kind of slow and anyone can say anything like, yes.
我现在仍然看到人们在犯的一个基本错误,是没有理解这是一种赋能型技术,能够让很多不同的事情成为可能;他们仍然把它看成只是一个生成文本和图片的东西,没有意识到,不,这会让 fraud detection(欺诈检测)做得更好,或者会解决各种你每天都在使用、但从未想过的公司内部晦涩 back office processes(后台流程)。所以你会得到各种更好的体验和更好的产品,但它们未必会以 AI 的面貌出现在你面前,也不会跑来告诉你这背后是 ChatGPT。可你还是会听到有人说,可是你知道,我问了它一个问题,答案并不完全正确。那又怎样?这就好比看着 Internet,然后说它有点慢,而且谁都可以随便乱说——对,没错。
Speaker 1 | 1:07:28 - 1:07:48 Yeah. But well, more it's like looking at the Internet in like 1995 and saying, but but but like there's an interview you can find online with David Bowie a British journalist called TV journalist called Jeremy Paxman in like 1994 or something. And David Bowie is saying, This is gonna change the world. And Jeremy Paxman says, But it's just like random people talking to each other on chat rooms. This is a mistake.
对。但更准确地说,这就像是在 1995 年看待 Internet,然后说,可是可是可是——你可以在网上找到一段采访,大概是 1994 年左右,David Bowie 接受一位 British journalist、TV journalist 叫 Jeremy Paxman 的采访。David Bowie 说,这会改变世界。而 Jeremy Paxman 说,可这不就是一群随机的人在 chat rooms 里互相聊天吗。这种看法就是错的。
Speaker 1 | 1:07:48 - 1:07:53 It's not the chat rooms, it's the network. And it's the same thing here. It's not the text in the pictures. It's the enabling technology.
关键不在 chat rooms,而在 network(网络)。这里也是一样。关键不在文本和图片,而在这种赋能型技术。
Speaker 2 | 1:07:53 - 1:08:14 Yeah. And I think obviously a lot of what you read about compares what's happening in AI to other technological revolutions for context. And I think one interesting thing is many people at the forefront of AI at these labs are like in their twenties, they don't necessarily have the same context around some of these. A bunch of them listen to the podcast. I guess if you could impart a few messages to them on those eras, what would they be?
对。而且我觉得很明显,人们现在读到的很多内容,都会把 AI 正在发生的事情和其他技术革命做比较,以提供背景。我觉得一个很有意思的点是,很多站在 AI 前沿、在这些 labs 里工作的人都才二十多岁,他们未必对其中一些历史背景有同样的理解。这里面有不少人也在听这个 podcast。我想,如果你可以给他们传达几条关于那些时代的信息,会是什么?
Speaker 2 | 1:08:15 - 1:08:17 So, I think we've talked about a bunch of this stuff that
所以,我觉得我们已经谈到很多这类东西了,就是
Speaker 1 | 1:08:17 - 1:08:40 stuff is unclear and stuff will change and all the acronyms and I've I've the slide in my presentation is here are a whole bunch of acronyms and companies and concepts and ideas and technologies from 1995 that that failed. Same thing from 2010 or 2005. We should just presume a bunch of stuff that we're working on now won't work. And some of it you're very conscious of. Mean, was the conversation about OpenAI last year.
很多事情其实并不清楚,而且很多事情会改变;还有那些缩写词——我在演讲里有一页 slide,上面列了 1995 年一大堆 acronyms、companies、concepts、ideas 和 technologies,它们后来都失败了。2010 年或者 2005 年的也一样。我们应该直接假定,我们现在在做的一大堆东西将来是行不通的。其中有些你自己是非常清楚的。我的意思是,去年关于 OpenAI 的讨论不就是这样吗。
Speaker 1 | 1:08:40 - 1:08:48 Of course, some of this wasn't gonna work, but it was worth trying. I mean, I think just like last week, OpenAI gave up on the web browser. Yeah. And I had that on my slide. Is this a thing that will work?
当然,这里面有些东西本来就不会成功,但还是值得试一试。我的意思是,我觉得就像上周一样,OpenAI 放弃了 web browser。对。我当时也把这个放在我的 slide 里了。这会是个能行得通的东西吗?
Speaker 1 | 1:08:48 - 1:08:58 Probably not. No. We should presume MCP will get replaced by something else or may do and so on. So that's kind of one point. Just presume you're in this stage of radical uncertainty where everything is turning over and changing.
很可能不会。对。我们应该预设 MCP 会被别的东西取代,或者可能会如此,等等。所以这算是一个要点。就是先假定你正处在这样一个极度不确定的阶段,一切都在被颠覆、都在变化。
Speaker 1 | 1:08:59 - 1:09:27 Secondly, like, you know, I always have this slide I show of the so from Goldman, the CIO survey, of the percentage of enterprise workflows that are in the cloud, and the number is like 30% after twenty years. Like, it takes time and it's hard and complicated and people have other priorities. I do a poll on social media every now and then, is machine learning still AI? And some people get very upset like there's this academic definition that they've dreamt up of what AI is, but like AI is whatever is new. It's like technology.
第二点,你知道,我总会放一张 slide,引用 Goldman 的 CIO survey,讲企业 workflow 有多少百分比是在 cloud 上,那个数字是大概 30%,而且这是二十年之后的结果。就是说,这需要时间,而且很难、很复杂,人们还有别的优先事项。我也会时不时在社交媒体上发个投票:machine learning 还算不算 AI?有些人会特别生气,好像他们自己构想出了一套关于 AI 的学术定义,但其实,AI 就是任何新的东西。它就像 technology 一样。
Speaker 1 | 1:09:27 - 1:09:46 It's not technology once it's been around for a while. But the point I was sort of thinking is like, I gave this presentation to, I think it was like a European water company. And I gave the presentation on machine learning. The CEO says, this is very interesting. Maybe innovation should be one of our top five priorities next year.
一样东西存在久了,就不再被叫作 technology 了。但我当时在想的是,我曾给一家欧洲的 water company 做过这个演讲。我讲的是 machine learning。CEO 说,这很有意思。也许 innovation 应该成为我们明年前五大优先事项之一。
Speaker 1 | 1:09:46 - 1:09:58 And if you're in tech, that sounds insane. Then you think, but they're a water company. Let's think about what their priorities might be. Regulation, earthquake, there's lead in the pipes. They need to rip out the Huawei equipment.
如果你是在 tech 行业里,这听起来会很疯狂。然后你会想,可他们是一家 water company。我们来想想他们的优先事项可能是什么。监管、地震、管道里有铅、他们还得把 Huawei equipment 拆掉。
Speaker 1 | 1:09:58 - 1:10:06 There's a repricing coming up next year. They've got other stuff. There's a drought. That reservoir, the dam is leaking. Oh crap.
明年还有一次 repricing 要来。他们有别的事情要处理。还有旱灾。那个 reservoir,那座 dam 在漏水。哦,糟了。
Speaker 1 | 1:10:06 - 1:10:25 We're gonna need to screw the rest of our budget. So like most people have other stuff going on and then maybe a much more tangible thing. Other people's jobs are hard too. Everybody else's job is hard. Every other industry has a bunch of stuff they're trying to work out that you probably don't know about.
我们大概要把剩下的预算全都砸进去。所以,大多数人都有别的事情在发生;再进一步说,一个更具体的事实是:别人的工作也很难。每个人的工作都很难。每个别的行业都有一堆他们正努力解决的问题,而这些问题你很可能根本不知道。
Speaker 1 | 1:10:25 - 1:10:41 I remember having a conversation with the head of one of the world's biggest ad agencies at an event years ago. And I said, trying to be clever, like, are you worrying about? You're worried about machine learning? This is before OpenAI. He said, yeah, I'm also worried that one of our countries doesn't have a good head of creative.
我记得很多年前在一次活动上,我和一家全球最大广告代理公司之一的负责人聊过天。我当时还想显得聪明一点,就说,你在担心什么?你是在担心 machine learning 吗?那还是 OpenAI 出现之前。他说,是啊,但我还担心我们其中一个国家的业务没有一个靠谱的创意负责人。
Speaker 1 | 1:10:41 - 1:11:05 I'm worried that this big client is gonna put us up for review. I'm worried that my head of X, Y, Z is in a conflict with this other person. Everyone has all this other stuff going on and AI isn't the only thing. And if you think about, you know, this is this you know, talking about previous lessons, like you go back and think about what industries the Internet affected and didn't affect. If you're the newspaper, fine.
Speaker 1 | 1:10:41 - 1:11:05 我担心这个大客户会把我们放进评估流程。我担心我这边负责 X、Y、Z 的主管和另一个人有冲突。每个人手头都有各种别的事情在发生,AI 并不是唯一的重要事项。而且如果你想想,结合我们之前谈过的那些经验教训,回头看看 Internet 影响了哪些行业、又没有影响哪些行业。如果你是 newspaper,那没问题。
Speaker 1 | 1:11:05 - 1:11:11 You got demolished. Now imagine you're Caterpillar. What did the Internet change for Caterpillar? Okay. It was really useful.
Speaker 1 | 1:11:05 - 1:11:11 你就被彻底击垮了。现在想象一下你是 Caterpillar。Internet 对 Caterpillar 改变了什么?好吧,它确实很有用。
Speaker 1 | 1:11:11 - 1:11:24 In the end of the day, your business is making big pieces of metal. Yeah. What did the Internet change for an aggregates company? Do you even know what an aggregates company is? Your job is digging sand out of the ground, putting it on trains and then on trucks and then giving it to people.
Speaker 1 | 1:11:11 - 1:11:24 但归根结底,你的业务是制造大型金属部件。对吧。Internet 对一家 aggregates company 改变了什么?你甚至知道 aggregates company 是做什么的吗?你的工作就是把沙子从地下挖出来,装上火车,再装上卡车,然后交给别人。
Speaker 1 | 1:11:24 - 1:11:32 How much did the internet change your industry? Well, in the end, kind of not very much. How much is AI going to change that industry?
Speaker 1 | 1:11:24 - 1:11:32 Internet 对你的行业改变了多少?嗯,说到底,其实并没有太大改变。那 AI 会在多大程度上改变那个行业呢?
Speaker 2 | 1:11:32 - 1:11:34 Depends how well the physical AI stuff works, guess.
Speaker 2 | 1:11:32 - 1:11:34 我猜,这取决于 physical AI 这类东西到底能运作得多好。
Speaker 1 | 1:11:36 - 1:11:52 Yeah, but do you want to put, you're gonna put humanoid robots in the three story high truck or the trucks gonna get connected to GPS and drive themselves around? They probably already are. The trucks will get autonomous fine. Okay. How much does that actually change your industry?
Speaker 1 | 1:11:36 - 1:11:52 对,但你真的会把 humanoid robots 放进那种三层楼高的卡车里吗,还是说卡车会接入 GPS 然后自己开来开去?它们可能本来就已经是这样了。卡车会实现 autonomous,这没问题。好。那么这实际上会在多大程度上改变你的行业?
Speaker 1 | 1:11:52 - 1:12:03 Like, guess what? There were some industries where it's not gonna change anything. And it won't be predictable. There'll be some maybe, there'll be some that you think of now where you're completely wrong. There'll be a bunch of industries where the internet actually didn't change a damn thing.
Speaker 1 | 1:11:52 - 1:12:03 你看,事实就是:有些行业根本不会被改变。而且这不会是可预测的。可能会有一些行业,也许就是你现在想到的那些,但你的判断会完全错掉。会有一大批行业,Internet 实际上根本没改变它们分毫。
Speaker 2 | 1:12:03 - 1:12:12 Well, I wanna make sure to leave last word to you. Anything we haven't talked about here that you think our listeners should take away or anywhere you wanna point them, we'll certainly link to your presentations, but anything else before?
Speaker 2 | 1:12:03 - 1:12:12 好,我想确保把最后的发言留给你。我们这里还有什么没谈到、但你觉得听众应该带走的要点,或者你想把他们引导到什么地方?我们当然会附上你那些 presentations 的链接,不过在结束前还有别的吗?
Speaker 1 | 1:12:12 - 1:12:43 Yeah. Well, my parents had good SEO. A story I was thinking about recently, which I didn't use writing about tokens, which was in like 1999, the bank I worked for was going to look at working on the IPO of a company that sold electronics and software on the internet. They'd previously been a catalog company and it wasn't clear Amazon would do this. And the guy, the founder had started it in his bedroom aged 16, and they'd like turn over a £100,000,000 and they were gonna flight for 7 or 800,000,000, which was insane even then.
对。嗯,我父母的 SEO 很不错。我最近想到一个故事,写 token(代币)时没用上。大概在 1999 年,我当时工作的那家银行准备研究是否参与一家公司的 IPO。这家公司在互联网上销售电子产品和软件,此前它是一家目录邮购公司,而当时还不确定 Amazon 能不能做成这件事。那个创始人 16 岁时就在自己卧室里创业,公司营业额已经做到 £100,000,000,他们打算按 7 亿到 8 亿的估值去上市,这甚至在当时都很疯狂。
Speaker 1 | 1:12:44 - 1:13:09 And so we all go up to Birmingham to see it, and then we're on the train back. And we're all talking about the founder and the board's kind of weird and the valuation. By veteran banker called David Tate, I think he's retired now, I should look him up. David Tate sort of opens one eye and says, it's a low margin reseller, one time sells, and goes back to sleep. I mean, sometimes it is kind of simple.
所以我们一群人去了 Birmingham 看这家公司,之后坐火车回来。大家一路都在讨论那个创始人、那个有点古怪的董事会,还有那个估值。这时有位资深银行家,叫 David Tate,我想他现在应该退休了,我该去查查他。David Tate 半睁开一只眼说:“这就是个低利润率的 reseller(转售商),而且是一锤子买卖。” 说完又睡过去了。我的意思是,有时候事情就是这么简单。
Speaker 1 | 1:13:09 - 1:13:11 It's a low margin reseller, one time sells.
这就是个低利润率的 reseller(转售商),一锤子买卖。
Speaker 2 | 1:13:11 - 1:13:15 This has been so much fun to get to talk to you about all this stuff. I really appreciate you taking the time to come on the pod.
今天能和你聊这些,真的太开心了。我非常感谢你愿意抽时间来上这个 pod。
Speaker 1 | 1:13:15 - 1:13:16 Sure. Thanks for having me.
当然。谢谢邀请我。
Speaker 2 | 1:13:16 - 1:13:42 I'm Jacob Efron, and this has been Unsupervised Learning, a podcast where I get to talk to the smartest people in AI and ask them tons of questions about what's happening with models and what it means for businesses in the world. As I hope is clear, I have a ton of fun doing this. It's a nights and weekends project in addition to my day job as an investor at Redpoint. But our ability to get these incredible guests on really comes from folks like you subscribing to the podcast, sharing it with friends. It's really what ultimately makes this whole thing work.
我是 Jacob Efron,这里是 Unsupervised Learning,一档 podcast,在这里我可以和 AI 领域最聪明的一些人交流,问他们大量问题:model(模型)领域正在发生什么,以及这对现实世界中的企业意味着什么。希望你已经能听出来,我做这件事非常开心。除了我在 Redpoint 做投资人的本职工作之外,这是一个利用晚上和周末时间做的项目。不过,我们之所以能请到这些了不起的嘉宾,真的离不开像你这样订阅这档 podcast、并把它分享给朋友的人。归根结底,正是这些支持让整件事能够运转下去。
Speaker 2 | 1:13:42 - 1:13:46 And so please consider doing that, and thank you so much for your support and listening. We'll see you next episode.
所以也请你考虑这样做,非常感谢你的支持和收听。我们下期见。