本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
我觉得人工智能已经变得像元宇宙这个词一样，当你听到别人说它的时候，你根本不知道他们到底是什么意思。
I feel like AI has almost become like the word metaverse where, like, you don't know what somebody means when they say it.
你能向另一个人解释清楚吗？
Could you explain it to another human being?
你能闭上眼睛，想象一下我该如何解释我想要它做什么吗？
Can you actually kind of shut your eyes and conceptualize how is it that I'm going to explain what it is that I
我希望这个东西能完成这些功能。
want this thing to do.
从某种意义上说，区块链是操作系统研究的前沿之一。
Blockchains are, in a sense, one of the frontiers of operating systems research.
就像有Windows这样的操作系统一样。
Like, the same way, like, there's an operating system like Windows.
是的。
Mhmm.
还有一个浏览器，它本身就是一个操作系统，因为你可以在里面运行应用程序。
There's a browser, which is itself an operating system because you can run apps in it.
它拥有完整的编程语言。
It's got a full programming language.
马克·扎克伯格收购Oculus是因为他认为这是下一代智能手机。
Mark Zuckerberg bought Oculus because he thinks this is the next smartphone.
他收购它不是为了让其成为游戏设备或让一亿人使用它。
He didn't buy it to be a games device or to have a 100,000,000 people using it.
他收购它是因为他认为这是下一代智能手机。
He bought it because he thinks this is the next smartphone.
这个
The
当你最终理解一项技术时，往往就是你该停止关注它的时候。
moment you finally understand a technology is often the moment you should stop paying attention to it.
重要的不是采用的绝对水平，而是变化的速度。
What matters isn't the absolute level of adoption, but the rate of change.
我们最常在技术转型期讨论它，然后就忘了它的存在。
We talk about a technology most during the transition, then forget it exists.
今天，这种转型正在人工智能、加密货币、智能眼镜和机器人技术领域同时发生。
Today, that transition is happening simultaneously in AI, crypto, smart glasses, and robotics.
关于每一项技术的讨论都达到了最高音量，这意味着有趣的问题不在于它们是否重要，而在于每一项技术实际颠覆了什么，又留下了什么。
The conversation about each is at maximum volume, which means the interesting question isn't whether they matter, but what each one actually disrupts and what it leaves standing.
今天，我们将为您带来来自《网络国家》播客的一段对话。
Today, we're bringing you a conversation from the Network State podcast.
主持人巴拉吉·斯里尼瓦桑与本尼迪克特·埃文斯进行了交谈，他是一位独立技术分析师，也是科技界阅读量最高的时事通讯作者之一。
Host Balaji Srinivasan speaks with Benedict Evans, independent technology analyst and one of tech's most read newsletter authors.
我和本尼迪克特·埃文斯在一起。
I'm here with Benedict Evans.
我们十多年前曾在a16z共事过。
We worked together at a six and z more than ten years ago.
本尼迪克特是知名的通讯作者，对观看此视频的观众来说可能无需介绍。
Benedict is, you know, well known newsletter author, probably needs no introduction for people watching this.
我们现在在新加坡。
We're here in Singapore.
你刚来这里参加一个AI会议。
You just came here for an AI conference.
你现在大约四分之一时间写通讯，四分之三时间参加会议或演讲。
You do about one fourth newsletter, three fourths conference nowadays or speaking.
这就是你来这里的原因。
That that's what brought you out here.
差不多是这样。
Pretty much.
是的
Yeah.
对
Yeah.
你说通讯是在一月左右发布的，类似这样吗？
And newsletter sound like January, something like that, you said?
差不多吧
Something like that.
每天都有些波动。
It wobbles a bit from day to day.
你最初是做移动分析的，后来逐渐成了更广泛的科技分析师。
And it started out you you you started as a mobile analyst and you became like a broader tech analyst.
这是你的发展轨迹吗？
Is that is that the evolution?
是的。
Yeah.
这么说也可以。
That's one way to put it.
我的意思是，我觉得我当初是
I mean, I think I was
你在Orange公司。
You're at Orange.
是这样吗？
Is that right?
很久以前。
A long time ago.
是的。
Yes.
很久以前。
Long time ago.
是的。
Yes.
就在它变得极其法国化的时候。
Just when it was all becoming horribly French.
就像我们之前聊天时提到的，我说过，在科技领域，当你真正理解某件事的时候，往往就是你该转向关注其他事情的时候了。我职业生涯起步于互联网泡沫时期，当时是一名股票分析师，负责覆盖移动通信股票。
There's there's like as we we were chatting before this, and I said like, thing in tech is at the point that you understand something is is often the point that you be moving on to pay attention to So something I started my career in the .com bubble as an equity analyst, and I was covering mobile stocks.
那时，移动通信领域充满活力、令人兴奋、很酷且具有颠覆性，但后来却变成了像水务公司一样的企业。
And at that time, mobile was kind of dynamic and exciting and sexy and disruptive, and they turned into water companies.
变成水务公司？
Were gonna Into water companies?
公用事业。
Utilities.
哦，公用事业。
Oh, utilities.
是的。
Yeah.
好的。
Okay.
他们打算连接世界上每一个人，然后他们真的做到了。
Got They were gonna connect everybody in the world, and then they did.
那接下来呢？
And like now what?
就像马克·安德森说的，他们就像追上卡车的狗。
They were like Marc Andreessen's phrase, they were like the dog that caught the truck.
对。
Right.
我后来转行做战略，涉足了一些媒体和电信领域的工作。
I went and worked in strategy and a bunch of media and telecom things.
是的，我当时在分析研究智能手机，因为它突然成为了行业的核心，却没人真正理解它。
And yeah, I was analyzing, looking at smartphones because that was suddenly become the center of the industry and no one understood it.
现在，这一切已经发生了。
Now, like, it happened.
是时候去寻找新的问题了。
Time to look for different questions.
嗯，这很有趣，因为我觉得我们交集的时候，正好处于智能手机红利期，也就是智能手机爆发式增长的阶段。
Well, it's it's funny because I think when we were overlapping, it was right in the middle of the smartphone dividend, the smartphone explosion.
而且，你知道，实际上我们有好几个点要说。
And just to, you know, we actually, there's a few things.
其中之一就是智能手机红利。
One is the smartphone dividend.
这是个很有用的概念，对吧？
That's a useful concept, right?
比如，十亿部智能手机的崛起意味着所有相关组件都变得更便宜，从而推动了VR头显和无人机的发展。
Like that the rise of a billion smartphones meant that everything that went into them became cheaper and that enabled VR headsets, that enabled drones.
对吧？
Right?
所有这些事情。
All this stuff.
是的。
Yeah.
所有从中衍生出来的组件。
All the components that came out of it.
是的。
Yeah.
所以智能手机的销量现在，据我回忆，大约是每年12.5亿到15亿台。
So smartphone sales are now, from memory, like one and a quarter, one and a half billion units a year.
而所有的供应链，所有这些组件，如果你想要购买5000或10000个，现在都可以现成买到。
And all supply chain from that, all of those components is then available off the shelf if you wanna buy 5,000 of them or 10,000 of them.
所有的WiFi芯片、电池、摄像头以及其他所有零部件。
All the WiFi chips and the batteries and the cameras and and all the other bits.
而以前，如果你想在某样东西里装电脑，基本上只能用个人电脑的组件。
And before, if you wanted to put computer into something, you basically need to use PC components.
所以自动取款机等等基本上都是个人电脑。
So ATMs and so on are all basically PCs.
比如电梯基本上也是个人电脑。
Like elevators are basically PCs.
而这带来了尺寸、功耗和成本上的限制。
And that has size and power and cost constraints.
然后智能手机成为主流，所有这些零部件就都变得可用了。
And then smartphones become the thing, and then all those components are available.
所以这就催生了无人机、智能灯泡以及围绕它周边的所有其他零配件。
And so that's what gets you drones and connected light bulbs and all the other bits and pieces around the edge of that.
我认为人们没有意识到的一点是，他们以为比如军方有专门的装备，有自己独特的供应链。
One of the things that I think people don't appreciate is they think, for example, like the consumer they think like the military has like special gear and it's got its own kind of supply chain.
通常，军事供应链往往只是消费级供应链的一个子集，因为你卖出十亿个这样的产品，而军事用途的可能只有十万或一百万个单位。
Often, the military supply chain is often just a subset of the consumer supply chain because you sell a billion units of this and maybe you have a 100,000 or a million units of a military thing.
实际上，现在几乎可以说是反过来了——过去的情况是，在我们出生之前，情报机构会最先获得那些酷炫的新技术。
It's almost it's actually, it's almost kind of the reverse now in that it used to be so the way I think about this is like in the past, like before we were born, the intelligence agencies would get the cool new stuff first.
然后军方会得到它，接着是大公司。
And then the military would get it, and then big corporations would get it.
最终消费者会在三十年之后才接触到它。
And eventually consumers would get it like thirty years after.
后来
Afterwards.
这就像是那个
This is like the That's
版本中
in version.
是的
Yes.
就像微波炉是为NASA发明的。
It's like microwaves are invented for NASA.
没错。
Right.
最终消费者也能用上它们。
And eventually consumers get them.
是的。
Yeah.
或者像GPS最初是为导弹导航而发明的。
Or like GPS was invented to guide missiles.
没错。
Exactly.
而现在它被用来给猫咪照片打标签。
And now it's used for tagging cat photos.
这种转变是因为技术变得足够便宜，可以让消费者使用，而不是需要花费十亿美元才能拥有一个。
And the shift is like a combination of the stuff getting cheap enough that it can be for consumers instead of you needing a billion dollars to have one.
然后一旦价格足够低廉，消费者的规模就会扩大。
And then the scale of consumers once it gets cheap enough.
所以现在的情况是，消费者先获得新技术，而军方要等十年后才能用上，因为
And so now the way it works is the consumers get the new stuff and the military gets it ten years later because that's how long it takes to
官僚机构需要时间来整合，没错。
The bureaucracy to assemble exactly.
第一，官僚主义。
A, the bureaucracy.
B，将其加固并产品化，转变为能够承受发射、冷热环境或其他任何条件的所需形态。
B, to harden it and productize it and turn it into what you need if it's gonna get shot out or it's gonna be cold or hot or warm or whatever it is.
是的。
Yeah.
这很有趣。
That's it's funny.
通过这个过程真的能改进吗？
Does it really improve through that process?
我知道人们认为会这样，但我不确定相对于不使用相当不错的产品与延迟加固所带来的改进相比，是否真的值得。
I know people think it does, but I'm not sure it does relative to the the cost of not using the pretty good product versus whatever improvements come from the delay to harden it.
我不确定它是否真的
I'm not sure if it actually
击中思考还是这个，但显然存在一个过程，你必须把它放进战斗机里。
hits think or this is but there's clearly there's sort of a process of you have to put it into a fighter jet.
你不会每六个月就更换战斗机的航空电子设备。
You don't replace the avionics in a fighter jet every six months.
对。
Right.
嗯，但是
Well, but
不过确实
but yeah.
你知道，这其中的核心在于创新的前沿是为消费者服务的。
You know, that's the kind of the core of it is the the cutting edge of the innovation is for consumers.
然后这会反馈到你展示其他所有内容的地方。
And then that flows back to where you show everything else.
没错。
That's right.
你知道，我本来想说，也许在中国确实如此。
Well, you know, I was gonna say maybe in China you do.
也许在中国，就像消费级无人机那样，他们把四旋翼无人机做得很好。
Maybe in China, like, I think what happened with the consumer drones, they got good at quadcopters.
是的。
Yep.
这促使他们推出了新的产品形态。
And that's led them to their new form.
你见过亿航吗？
Have you seen Ehang?
那就是中国的飞行汽车。
It's like the Chinese flying cars.
哦，好的。
Oh, okay.
我没看过。
I haven't seen it.
你知道，我大概一年、一年半前播放过这段视频，当时人们说，有点像，你知道，那种调侃，就像我们想要飞行汽车，结果只得到了140个字符。
You know, I played this clip like a year, year and a half ago, and people said, kind of, you know, teal one, like we wanted flying cars, we got 140 characters.
我当时就觉得，很多人都在调侃这个，但我想说的是，我们确实想要飞行汽车。
And I was like, a lot of people did a riff on that, but I was like, we wanted flying cars.
我们用中文字符显示的。
We got them in Chinese characters.
明白吗？
Okay?
是的。
Yeah.
关键是当我展示那个的时候，人们会说，那不算汽车。
And the thing is when I put that up there, people are like, that's not a car.
这其实是个直升机，你知道的。
It's a, you know, it's a it's a copter.
对吧？
Right?
这可不是汽车。
That's not a car.
它没有轮子，但以另一种方式解决了问题。
It doesn't have wheels, but it it solved the problem differently.
对吧？
Right?
实际上，这曾经是你说过的一句话：不公平的比较往往是最好的比较。
And actually, think was one of your lines, it's like unfair comparisons are often the best kind of comparisons.
对吧？
Right?
是的。
Yeah.
我记得在Andreessen Horowitz看到过一堆飞行汽车。
I remember seeing a bunch of flying cars when we were at Andreessen Horowitz.
我记得马克·安德森说过，它们都像船屋一样，而船屋既是糟糕的房子也是糟糕的船。
I think Mark Andreessen, he said it was like they were all like houseboats, and a houseboat is a crap house and a crap boat.
是的。
Yes.
说得对。
That's right.
而且，你知道，把它想成飞行汽车其实是个错误的说法。
And, you know, thinking of it as a flying car is like the wrong term.
更好的理解方式是把它看作一种更小、更好、更便宜的直升机，没错。
It's better to think of it as like a small, much better, much cheaper helicopter Yes.
也许吧。
Maybe.
但关键是它们现在，另外一点是它用于短途跳跃，比如城市之间，让你飞越交通拥堵。
But the point is that they now, other the thing is it's done for short hops and like city to city where you fly over the traffic.
顺便说一句，优步在被砍掉之前本来打算做这个的。
And they've got this Uber was going to do this, by the way, before they decapitated Uber.
低空经济是他们当时在考虑的一个方向。
Like the low altitude economy was something they were thinking about.
很多项目在西方被砍掉后，却在中国完整地出现了。
And a lot of things get like cut off in the West and then they appear fully formed in China.
比如消费级无人机，你知道的，克里斯·安德森很早就开始研究无人机，但被美国联邦航空局给拦住了。
Like consumer drones, for example, you know, Chris Anderson, he was very early on drones and that got blocked by the FAA.
所以消费级无人机在美国发展受阻，这就是为什么大疆在中国崛起。
And so consumer drones were hobbled in The US, that's why DGI arose in China.
因此很多事物在西方被阻碍，于是它们就在中国发展起来。
So lots of things get blocked in the West and they arise in China because of that.
总之，回到正题。
Anyway, coming back up.
说到智能手机，我想起你和Horace Didiu，还有Simpo，我前阵子在他的播客里看到过他。
So smartphones, I mean, think you and Horace Didiu, Simpo, who I I saw in his pod a while ago.
我认为你们两位是最顶尖的。
I think you're two of the best.
他好像也是欧洲人什么的。
He's also like European or something like that.
是的。
Yeah.
我知道他以前在诺基亚工作。
Knew he was he was at Nokia.
我是说，这里面有一种挺有意思的信息。
I mean, there's an interesting kind of like information.
你认识他吗？
Do you know him?
是的。
Yeah.
我不认识他。
I don't know him.
他是
He's a
他是个很棒的人。
he's a great guy.
有一部分原因是，当时没多少人
Part of it was it was like and there was a moment in time when there weren't many people
在做移动业务。
Doing mobile.
是的。
Yeah.
真正理解这一点，并且是行业分析师，还能公开讲话的。
Who really understood this and were industry analysts and were able to talk in public.
对。
Right.
是的。
Yes.
所以诺基亚、高盛、贝恩或其他公司内部的人掌握了所有数据，但他们不能公开这些数据，也不被允许在公开场合发表言论。
So there were people inside Nokia or Goldman's or Bain or wherever who had all the data, but they couldn't publish the data, and they weren't allowed to say stuff in public.
对。
Right.
或者如果他们撰写分析报告，那也是面向公开市场投资者之类的分析。
Or if they were writing analysis, it was analysis for public markets investors or something.
没错。
Right.
所以当时很少有人知道，你可以去拿苹果的报告，制作一张销量图表和一张ASP图表，并且知道ASP是什么，或者知道ARPU是什么。
And so there were very few people who were like, knew that you could go and take Apple's reports and make a chart of unit sales and make a chart of ASP and knew what ASP was or knew what ARPU was.
没错。
Right.
现在这种情况就像爆炸式增长。
Now there's like an explosion of this.
所以现在有大量的人在做这件事，特别是如果你现在看人工智能领域，大概有10个人能做出非常出色的200页演示文稿，包含所有可能的人工智能图表。
So there's huge numbers, and particularly if you look at AI now, there's like 10 people who do a really really good 200 page deck of every possible AI chart.
哦，是吗？
Oh, that right?
有意思，是的。
Interesting, yeah.
所以整个情况发生了转变。
And so that whole thing shifted.
但当时，确实只有我、霍拉斯，还有本·巴哈兰。
But at the time, yes, it was me and Horace and, like, Ben Baharan.
我们当时就像是……
We were, like,
唯一且与Stratachari相关的人。
the only And Stratachari kind of adjacent.
是的。
Yeah.
没错。
Exactly.
当时只有少数人懂这个，能做图表，也被允许做图表。
There were, like, you know, a handful of people who understood this and could do the charts and were allowed to do the charts.
所以这有点像是，你知道，恰好在正确的时间出现在正确的地方，让我获得了大量关注。
And so that was sort of, you know, being at the right place, right time got me a lot of attention.
是的。
Yeah.
这很有意思。
It's interesting.
我觉得像你、本·汤普森和Strathecari，我不确定霍勒斯有没有通讯，但你们在Substack将其产品化之前就已经在做通讯了。
I think like you and Ben Thompson and Strathecari, and I'm not sure if Horace had a newsletter, but you guys were newsletters before Substack productized it.
有点像罗根在播客被产品化成为一个类别之前就已经在做播客了。
Sort of like Rogan was podcast before that became productized as a category.
你现在用Substack吗？
And are you on Substack?
还是你之前用的是Ghost？
Or were you on Ghosts?
不是吗？
No?
是的。
Yeah.
不是。
No.
你自己搞了个定制的东西？
I'm You got your own custom thing?
我还在用我老一套拼凑的系统：Mailchimp 加 Memberful 加 Squarespace。
I'm still on my old cobbled together stack of Mailchimp plus Memberful plus Squarespace.
你为什么不想换到别的平台呢？
Why why don't you you don't wanna move to somebody.
搬家太麻烦了。
It's just a pain to move.
换平台是个大工程。
It's a heavy lift to move platform.
然后你坐下来做分析，心想：这值我花上一周的时间。
And you sit and do the analysis, and you're like, this is a good use of like a week of my time.
对。
Right.
也许吧。
Maybe.
也许吧。
Maybe.
有可能。
It might be.
到目前为止，Substack 已经很不错了，但毕竟这是你的，你知道的，你的
At this point, Substack's pretty good, but I mean, it's your, you know, your
目标也很重要。
goal goes also.
是的。
Yeah.
还有一个独立的Substack问题，就是你希望它出现在你的新通讯还是你的Substack上？
There's a separate Substack thing thing, which is do you want it to be on your new your newsletter or your Substack?
是的。
Yes.
没错。
That's true.
对。
Yeah.
因为这是一个平台。
Because it's a platform.
对。
Yeah.
而且你能获得优势，我的意思是，这个我们可以聊聊。
And you get the advantage of I mean, this is something we can talk about.
你知道的，这是克里斯·迪克森说的：为工具而来，为网络而留。
You know, it's Chris Dixon's line of come for the tool state for the network.
对。
Right.
你去Substack，他们会为你带来新订户。
You go on Substack, they will get you new subscribers.
我想它并不会为你带来订户。
Guess it won't get you subscribers.
是的。
Yes.
另一方面，他们现在掌控了你的读者是谁，而你却无法掌控，这正是网络平台一贯的问题。
On the other hand, they now they control who your readers are and you don't, which is always a thing of a network.
我的意思是，它
Well, mean, it
过去会把邮件发给所有人。
used to sell mail out to everybody.
是的。
Yeah.
但接着他们试图让你使用他们的网站和算法来决定谁阅读什么内容。
But then them they're trying to get you to use their website and their algorithm to decide who reads what.
这是真的。
That's true.
所以总是会有这类问题，比如：你是否愿意选择那些能为你提供受众的人？
So there's always these kind of questions like, do you want to go with the people who will give you an audience?
作为交换，他们会决定给你提供受众。
And in exchange for that, they're deciding that they'll give you the audience.
分发总是伴随着权衡。
There's always a trade off for the distribution.
是的。
Yeah.
我认为Ghost是另一个选择。
I think ghost is another option.
对，对。
Yeah, yeah.
我觉得——
I think-
Ghost和Beehive是另外两个选择
Both ghost and beehive are the two others
Ghost这个平台，你知道，我很早就关注它了，当时就觉得它非常出色。
Ghost that people is like, you know, I saw ghost when it was very early and I just thought it was so good.
就其本身而言，它确实很优秀，我的意思是，不仅限于此，作为一个开源产品，它的完成度异常高。
For what it was, it was, I mean, not even for it, it's just a very polished thought For an open source product, it's unusually polished.
约翰·诺兰非常非常出色。
John Nolan's very, very good.
有趣的是，你知道，关于那个，嗯，有很多我们可以讨论的话题，但整个新闻通讯这件事，有时候有些东西比如新闻通讯或播客，在我看来是技术领域的小写字母，然后才变成大写字母。
It's funny, you know, like on that, well, there's a bunch of things we can talk about, but the whole newsletter thing, it's sometimes there's things that are like newsletters or podcasts that are what I consider lowercase in technology before they become uppercase.
比如说Odeo，你知道，就像Twitter在成为Twitter之前是一家播客公司。
Like, for example, Odeo, you know, like what Twitter was Twitter was a podcasting company before it became Twitter.
而时间常数，他们只是搞错了时间常数，你需要——这很难预测——微博会先火起来，然后需要像AirPods这样的东西，以及大家长时间在线，也许，你知道，还需要新冠疫情，播客才能真正爆发。
And the time constant, they just got the time constant wrong where you needed, which is hard to predict, that microblogging would take off first, and then it required like AirPods and everybody being online for a long time, and maybe, you know, COVID before podcasts really exploded.
这个术语当时已经以小写形式存在了。
And the term was around in lowercase.
你甚至可以说，它需要像5G或4G这样的网络。
You could even argue it's needed like five g or four gs.
类似这样的东西。
Something like that.
是的。
Yes.
如果你在车里听，就需要足够快的网络。
If you're listening to it in the car, then you need a half fast enough network.
对。
Yes.
带宽是一个限制因素。
Bandwidth is a constraint.
是的。
Yes.
然后时间就对上了。
And then the time works.
所以你觉得呢
So what do you
你觉得现在哪些是小写的术语，将来会变成大写的？
think is lowercase today that's gonna become uppercase?
比如，科技领域里有哪些东西，人们现在只是说‘哦，这个东西存在’，但将来会大放异彩？
Like, what's up what's what exists in tech that people are like, oh, yeah, that exists, that that's gonna go big?
我有一些想法。
I I have some ideas.
我想听听你的看法。
I wanna hear yours.
有趣的问题。
Interesting question.
我认为，如果我是个顾问，要白板分析这个问题，答案很可能围绕着人工智能。
I think there's probably the answer, if I was a consultant and trying to whiteboard this, is I would be looking around AI.
嗯哼。
Mhmm.
因为那是一个新平台。
Because that's a new platform.
而且，你知道，很多旧的空白领域已经被填满了，现在又出现了一大片新的空白领域。
And, you know, there's a lot all the white the old white space got filled in, and now you've got a whole bunch of new white space.
所以从确定性角度来看，这里应该会出现一大批那样的东西。
So deterministically, there should be a bunch of those things here.
人工智能，我相信我们会讨论到它，确实给人一种九十年代中期的感觉。
AI, which I'm sure we'll talk about, does feel very sort of mid nineties.
就像九十年代中期的互联网一样。
That you're like mid nineties Internet.
就像在问，这是个浏览器吗？
In that like, well, is this a browser?
你要怎么使用它呢？
How do you use it?
它是用来做什么的？
What's it for?
你要怎么访问它？
How would you get to it?
这个怎么运作？
How does this work?
价值在哪里？价值捕获会在哪里发生？
Where's the value where's the value capture gonna be?
我不确定是否有一个答案，可能是我年纪大了，没有花太多时间去寻找那些边缘的奇怪东西。
I'm not sure that there's like maybe one answer is like, I'm too old and I'm not like spending too much time looking for like weird weird stuff around the edges.
我最近一次亲自发现的是Shein。
The last one of these that I spotted personally was Shein.
Shein？
Shein?
是Shein还是Shein？
Is it Shein or Shein?
听那里的人说，应该是Shein。
I'm told it's Shein from speaking to people there.
我没有参与过她的电视保险业务。
I haven't worked on her TV insurance.
那是个有趣的案例，也许可以说是最后一个你能发现的，因为突然间，等等，这个在iPhone应用商店排行榜首的东西是什么？哦，所有的
That was an interesting one that it was maybe you could also say it was the last of the ones that you could spot because suddenly, wait, what is this thing that's at the top of the iPod of the iPhone app store chart Oh, all the
我明白了。
I see.
是的。
Yep.
突然间那个东西就爆发了。
Suddenly that thing exploded.
那可能是全球最大的纯服装零售商。
And that's like probably the probably the largest apparel the pure play apparel retailer on earth.
对。
Yeah.
还有希恩和塔穆。
And like Sheehan and Tammu.
塔穆。
Tammu.
对。
Yeah.
没错。
That's right.
他们现在正受到关税的影响，而且你
They're now getting hit with the tariff stuff and, you
知道，是的。
know Yeah.
关税加上美国的最低限度规则。
Tariffs plus plus a de minimis rule in The US.
而这
And that's
仅仅是美国市场。
only The US market.
而且这还不是，你知道，我不清楚这部分占他们销售额的多少比例。
And that's not, you know, I don't know what fraction of their sales that is.
是的。
Yeah.
大概是他们销售额的三分之一或者四分之一之类的。
Like a third of their sales or a half quarter of their sales or something.
所以那件事确实挺有意思的。
So that was like that was that was a thing that was interesting.
我不太确定。
I'm not sure.
最近我还没注意到有什么值得关注的新情况。
There's not like a new thing that I'm watching that I've noticed recently.
我相信肯定会有新动向的，你知道，我会持续关注。
I'm sure there will be, you know, I I I keep looking.
所以我有几个。
So I have a few.
我们刚才在讨论眼镜。
We we were talking about the glasses.
对吧？
Right?
我觉得智能眼镜就像是iPhone之后最可预测的东西。
Like, I think smart glasses are sort of like the most predictable thing after the iPhone.
那我们是，哦，是的。
Then we're Oh, yeah.
我把这个归到不同的类别里。
That's I put it in a different category.
我一直在想，现在有什么东西人们还没完全注意到，但实际上已经在使用了？
Was sort of thinking like, what stuff that's being used now that people haven't quite noticed is being used yet?
我明白了。
I see.
嗯，我猜是眼镜吧。
Well, I guess Glasses.
那个
That
眼镜绝对是你的下一个目标。
Glasses is definitely your next thing.
当然。
Sure.
所以我想我会把VR头显、AR头显和眼镜归为一类，可以说眼镜就像是护目镜的下一代版本。
So so I guess I would I would sort of bundle VR headsets, AR headsets, you know, like that with glasses and and say that that's just glasses are sort of the next version for goggles.
但是，
But,
所以这是我们能达成共识的一点，但问题在于，正如你所说，会是手表还是手机呢？
So that's one that we'd agree on, except the question is, as you said, is it going be watches or phones?
它的市场规模会有多大？
How big does that get?
对吧？
Right?
是的。
Yeah.
就像播客逐渐扩展到包括视频播客一样，机器狗也很有意思。
Think, just like podcasts grew to mean like a video podcast and so on, the robot dogs are interesting.
它们玩起来很有趣，而且现在便宜多了。
They are fun to play with, and they're getting way cheaper now.
对吧？
Right?
它们从波士顿动力公司那种产品开始。
They went from the Boston Dynamics kind of things.
所以家用机器人作为玩具，我认为可能会越来越受欢迎，最初可能像圣诞礼物一样。
So the home robot as a toy, think I is probably going to become more and more popular, like a Christmas present kind of thing at first.
对吧？
Right?
展开剩余字幕（还有 480 条）
因为我看到孩子们把它们当玩具玩，他们非常喜欢。
Because I see kids playing with them and they just love them just as a toy.
而且，你知道，这有点像机器狗，无人机也开始成为圣诞礼物之类的东西。
And the, you know, it's kind of like the robot dog, the drone as like a starting to become Christmas present kind of thing.
我认为这将会成为一种趋势。
I think that that becomes a thing.
最终，就像我们在阿联酋的未来博物馆讨论的那样，他们给这些机器人穿上外衣，这样它就不只是机器狗的骨架，而是看起来像一只真正的动物。
And eventually, like we were talking about this at the Museum of the Future in The UAE, they clad these so that it doesn't it's not just like a skeleton of a robot dog, but it actually looks like an animal.
而这完全改变了你对它的看法。
And that completely changes your perception of it.
对吧？
Right?
所以我认为这将会成为趋势。
So I think that that'll be a thing.
至于人工智能，我的意思是，我们有AI、比特币、中国、无人机和生物技术这些领域。
And with respect to AI, and so let's do, I mean, there's AI, there's Bitcoin, there's China, there's drones, there's biotech.
我实际上在关注几个不同的领域。
There's actually several different areas that I'm tracking.
我正在关注这些各种各样的拐点，不管怎么说。
I'm tracking a bunch of these various singularities, whatever.
其实并不是技术意义上那种趋向无穷的拐点。
Not really actually singularities in the technical sense of going to infinity.
是曲线。
Was curves.
曲线，曲线。
Curves, curves.
没错。
That's right.
是的。
Yeah.
关于人工智能，你可以这样理解：现在我们已经进入了两年半左右，比如说，我们可以称之为‘聊天机器人普及时刻’，对吧？
With AI, there's, you know, one way of thinking about it is like now we're two and a half years in, let's say, let's call it chat, cheap tea moment, right?
而且
And
这很有趣，因为我认为人们真正高估的是它作为代理智能与增强智能的程度。
it's interesting because I think what people really overestimated was how much it's agentic intelligence versus amplified intelligence.
也就是说，你仍然需要提示它。
Like that to say, you still have to prompt it.
所以，提示就像是更高层次的编程，这是第一点。
So prompting is like higher level programming, number one.
而且你仍然需要验证输出结果。
And you still have to verify the output.
这意味着你需要大致了解自己在寻找什么。
And that means you kind of need to know what it is you're looking for.
比如说，如果它在你不熟悉的数学领域输出一堆数学符号，那你得是陶哲轩才能验证它。
For example, if it spits out a bunch of mathematical symbols in an area of math that you don't know, then you have to be Terrence Tau to verify it.
它可能是胡言乱语也可能是真实的，谁知道呢？
It might be gibberish or might be real, who knows?
对吧？
Right?
所以提示和验证实际上是许多领域的瓶颈。
And so the prompting and verifying are actually the bottlenecks in many areas.
就在大约一周前，卡帕西和我，也就是安德烈·卡帕西，刚刚讨论过这个问题。
Now, Karpathy and I, you know, Andres Karpathy, were just having a discussion on this like a week or so ago.
关于验证的问题是，如果你使用的是我们内置的GPU，并且你在查看图像、视频或前端代码，对吧？
And the thing about verifying is if you're using the GPUs that we have built in, and you're looking at images or video or front end code, right?
就像用户界面一样，你的眼睛能立刻识别出来，并且可以很快地进行验证。
Like a user interface, your eye can just instantly pick out and you can verify pretty quickly.
所以在这些方面，人工智能表现得相当不错。
So for that side of things, AI is quite good.
任何图像、视频相关的内容，你的耳朵也能辨别音频，对吧，还有前端方面。
Anything that's images, video, your ear can also pick out audio, right, and front end.
但当涉及到后端内容时，比如数据库代码、加密技术、数学方程式这些你没有GPU辅助处理的领域。
But when it's back end stuff, right, when it's like database code, when it's like crypto, when it's mathematical equations, that you don't have like GPUs.
你不能光靠眼睛扫一眼就快速判断出来。
You can't just like hit it with your eyes and quickly detect it.
对吧？
Right?
不管它对不对，你都得仔细深入地阅读。
Whether it's whether it's correct or not, you have to deep read it carefully.
对吧？
Right?
所以它能生成大段大段的文字，但你得去仔细检查它。
So it can generate reams of text, but then you have to You have take it.
没错。
Exactly.
正是如此。
That's right.
你对这个问题有什么看法吗？
Maybe you have some thoughts on that.
嗯，这挺有趣的。
Well, so it's funny.
前几天我和约翰·博尔兹韦格聊天，他说：本尼迪克特，你思考时是用幻灯片的。
I was talking to John Bolsweig the other day, and he said, Benedict, you think in slides.
嗯。
Mhmm.
所以我们就
So That we
我们也一样。
do too.
我们都用幻灯片思考。
We both think in slides.
所以我有一张幻灯片。
So I have a slide.
对。
Yes.
也许某种程度上，我会谈谈这张幻灯片，并围绕它做一个观察。
And maybe there's sort of a I'll talk about the slide and this is an observation around it.
我认为很多关于LLM的讨论都在寻找一种正确的概念化方式，就像是在问，什么才是理解它的正确方法？
I think a lot of discussion of LLMs is sort of hunting for the like, what's the right would the right way to conceptualize this?
就像机器学习一样，正确的概念化方式是将其视为模式识别。
It's like with machine learning, the right way to conceptualize it was this is pattern recognition.
嗯哼。
Mhmm.
我们仍在寻找合适的方式来概念化大型语言模型。
And we're still sort of hunting for the right way to conceptualize LLMs.
这张幻灯片说的是，传统软件是确定性的，做的是容易向机器解释的事情。
The slide is that traditional software is deterministic and does things that are easy to explain to machines.
事实上，自动化、机床、缝纫机、打字机、加法机
In fact, automation, machine tools, sewing machines, typewriters, adding machines
没错。
Right.
那些容易向计算机解释的事情。
Things that are easy to explain to a computer.
有些事情对人来说可能非常困难，但它们很容易解释清楚。
There may be things that are very hard for people to to to do, but they're easy to explain.
比如让你心算100次房贷计算可能很难，但写下逻辑步骤来解释如何计算却很容易。
So it's hard for you to drill a hole a 100 times to calculate a mortgage in your head, but it's easy for you to write down the logical steps to explain how you do this.
这就是传统软件，比如数据库、数据处理，以及整个六七十年代的大型机时代。
So that's traditional software, like databases, data processing, the whole sixties, seventies mainframe thing.
机器学习是那些难以向计算机解释的东西。
Machine learning is stuff that's hard to explain to a computer.
所以很难解释为什么那笔信用卡交易很奇怪。
So it's hard to explain why that credit card transaction is weird.
它是
It's
难以解释如何移动你的手之类的
hard to Or how to move your hand or something
比如解释为什么这张图片是一只狗而不是猫。
like to explain why that's a picture of a dog and not a cat.
对。
Right.
你以为这很容易，直到你亲自尝试去做。
You think it's easy until you try and do it.
然后你就像是在试图制造一个机械马。
And then it's like you try to make a mechanical horse.
它总是会倒下，直到……对。
It always falls over until Right.
系统机器人技术出现了。
System robotics comes along.
所以这就是机器学习。
So that was machine learning.
我还想拿个问题考考你，你觉得机器学习还算人工智能吗，还是现在已经只是软件了？
I also think that I as a kind of quiz for you, do you think machine learning is still AI, or is that now just software?
嗯，所以这种方式
Well, so the way that
因为我认为有一个过程是对的。
because I think there's a process that Right.
没错。
Right.
一旦它存在了一段时间，就不再是人工智能了。
Once it's been around for a while, it's not AI anymore.
有意思。
Funny.
所以我认为在这个领域内，从技术上讲，划分标准是机器学习会涵盖所有内容，包括线性逻辑回归、支持向量机等等这类东西。
So I I think like within the field, technically, the division would be machine learning would be, you know, everything up to linear logistic regression, and, you know, SVMs, all that kind of stuff.
然后就在你开始进行深度学习并拥有大型神经网络的那一刻，现在你就开始进入人们所说的现代人工智能领域了。
And then right at the point you start doing deep learning and you have large neural networks, now you start getting into what people would call modern AI.
所以机器学习几乎可以说是可理解性的边界，可以这么说，对吧？
So ML is almost like the boundary of understandability, you might say, right?
在那里你可以写出清晰的方程，并且真正理解正在发生的事情。
Where you can write clean equations and like really understand what's going on.
对我来说，最令人惊讶和困惑的是——我仍然觉得自己不明白这个现象是什么，但我仍然觉得它很神奇——那就是所谓的双重下降问题。
And to me, the most surprising and confusing I still don't feel like I know what the phenomenon is, but I still find it magical, is something called the double descent problem.
你知道那是什么吗？
Do you know what that is?
基本上，通常当你拟合数据时，你希望使用尽可能少的参数，因为可能会过拟合。
Basically, normally when you're fitting to data, you want to have the fewest possible parameters because you can overfit.
对吧？
Right?
所以你的误差会下降，然后在保留集上误差开始上升。
And so your error goes down, and then your error starts going up on the holdout set.
因此你在机器学习中训练模型，希望用最少的参数来解释训练数据并预测测试数据。
So you train your model in machine learning, and you want the minimum number of parameters to be able to explain the training data and predict the test data.
如果你过拟合了，那你就无法预测样本外的数据了。
And if you overfit, then you're no longer predicting out of sampled stuff.
双重下降现象是指当你进行人工智能研究时，实际上在转向高度参数化的模型时会获得第二次提升，误差实际上会再次下降。
The double descent is when you do AI, you get actually a second wind when you start going to a very highly parameterized model and the error actually drops again.
对吧？
Right?
这真是个非常奇怪的现象，已经有相关论文对此进行了研究等等。
And which is just a really weird phenomenon that There's papers on this and so on.
这是整个领域中最反直觉的事情之一——这些巨型参数化模型竟然能很好地泛化，对吧？
And it's one of the most counterintuitive things about the whole thing that just having these gigantically parameterized models would generalize well, right?
因为它违背了这一点，这才是最大的区别。
Because it violates That's the biggest difference.
请继续。
Go ahead.
或者人们可能会说还有其他方面是最大的
Or there's other things people might say is the biggest
区别。
difference.
我的意思是，我觉得人们使用‘人工智能’这个词的方式，就像使用‘技术’这个词一样。
Well, what I'm mean, think there's one of the ways I sort of think about the term AI is that people kind of use it like technology, the word technology.
是的。
Yeah.
没错。
That's right.
任何新东西都被叫做技术。
That anything new is technology.
你父母那一代的东西也被叫做技术。
Anything your parents had is technology.
我非常注重精确性。
I'm a stickler for precision.
所以存在
So there's
有有
there's there's
我们可以用不同的方式来解释AI这个词的含义。
different ways that you can say what do we mean by the word AI.
是的。
Yes.
我觉得AI几乎变得像元宇宙这个词一样，当人们提到它时，你并不清楚他们具体指的是什么。
I feel like AI has almost become like the word metaverse, where, like, you don't know what somebody means when they say it.
但回到我的幻灯片，第一点是存在确定性软件，这类东西比较容易解释。
But to to continue my slide, so this the first point is there's deterministic software, which is stuff that's easy to explain.
对。
Right.
有机器学习，这是些难以解释的东西，基础的机器学习就是如此。
There's machine learning, which is stuff that was hard to explain, which basic machine learning salt is.
而现在的大语言模型或许是那些容易向实习生解释的东西。
And now an LLM is maybe stuff that's easy to explain to an intern.
所以它是这样一种东西：如果你需要离开去开个启动会议，花半小时研究我们如何开展这个项目，那么大语言模型很可能做不到这一点。
It's So it's something where if you had to go away and have a have a, like, a kickoff meeting and spend half an hour working out how we're gonna do this project, then an LLM probably can't do that.
但如果这件事你能在十秒或二十秒内解释清楚，那么大语言模型就能做到。
If it's but if it's something that you could explain in ten seconds or twenty seconds, then an l l l m's gonna be able to do that.
而问题的一部分在于，你甚至能向自己解释清楚吗？
And part of the problem is, are you able to explain it even to yourself?
嗯哼。
Mhmm.
你能向另一个人解释清楚吗？
Could you explain it to another human being?
你是否能够闭上眼睛，在脑海中构思我将如何解释我希望这个东西做什么？
Can you actually kinda shut your eyes and conceptualize how is it that I'm going to explain what it is that I want this thing to do?
你所说的
What you're
非常重要，因为你知道，我想从几个不同的角度来探讨这个问题。
saying is very important because, know, there's like several different angles I want to take off of that.
你知道，从某种意义上说，我们正处在一个短语盛行的时代，对吧？
You know, in one sense, had this tweet, we're living in the age of the phrase, right?
所以AI的提示词，或者140个字符的推文，或者实际上在加密领域，比如14个词、13个词、12个词就可以成为你的加密重置短语，对吧？
So the prompt for the AI, or the 140 character tweet, or actually in crypto, like 14 words, 13 words, 12 words can be your crypto reset phrase, right?
这些都是具有强大功能的短语，对吧？
These are phrases of power, right?
在AI领域、社交媒体、加密领域都是如此，对吧？
In AI, in social, in crypto, right?
就像这些字符串能发挥很大作用，你明白吗？
Like this strings of characters that do a lot, you know?
它们是咒语。
They're spells.
它们是咒语，对吧？
They're spells, right?
关键在于，作为管理者，如果你是一名非常优秀的工程经理，你就很擅长提示AI。
And the thing about it is the CRISPR you are as a manager, like, you know, if you're a really good engineering manager, you're great at prompting AI.
因为关键是你不会只说‘嘿，把这个代码写出来’就完事了。
Because crucially, you don't just say, Hey, code this.
你会说，嘿，你知道，尝试用React来做这个。
You say, Hey, you know, try and use React for this.
你可以用React Native来做iOS和Android界面，用Tailwind，用起来。
You can use React Native for the iOS and Android interfaces, use Tailwind, use it.
在某种意义上，你掌握的词汇术语越多，就越能更好地进行提示。
The more in a sense vocabulary terms you have, the better you can prompt something with.
而且你必须正确使用这些词汇术语。
And you have to use those vocabulary terms correctly.
这意味着，举个例子，我在使用Dali时意识到，你知道，在Chattypie出现之前的那段早期，我就觉得，哇，艺术史现在成了一门应用学科了。
And what that meant is, for example, I realized with Dali, you know, when those first, you know, before the Chattypie moment, I was like, wow, art history is now an applied subject.
了解像塞尚、毕加索以及各种晦涩风格的知识，突然间你就能说，砰，按这种风格来，按那种风格来，它就能做到。
Knowing like Cezanne and Picasso and what, you know, these various kinds of obscure styles, suddenly you can be like, boom, style it like this, style it like this, and it'll do that.
对吧？
Right?
音乐方面也是同样的道理。
You can say the same thing for music.
那么，那里到底在做什么呢？
Like, what exactly is it that's being done there?
我需要一个词来形容这种事。
I need there is a word for that.
嗯哼。
Mhmm.
所以你可以用这个词。
So and and can use that word.
没错。
That's right.
正是如此。
Exactly.
所以你可以上传一段音频，然后问：这是什么风格？
So you can upload a track, and you can say, what style is this?
你会怎么描述这段音乐？
How would you caption this?
对吧？
Right?
你有没有见过那种高档餐厅，菜单上根本不写‘番茄’，而是写一些花哨的词，比如……
Ever seen, you know, like the restaurants with the fancy menus and they don't say tomatoes, they say like, Well,
有个词，有些东西理论上是主观的。
there's a word, there's things that theoretically subjective.
是的。
Yeah.
但在规定范围内是有界定的。
But there are within the provisions.
做这种特定事情时，有一个专门的术语。
There is a particular term for doing that particular thing.
没错。
Exactly.
就像红色和勃艮第红之间的区别，你知道的，深红啊之类的。
It's like difference between like red versus burgundy you know, crimson and what have you.
它们有精确的词汇来表达特定含义，通过这些精确词汇你就能实现更高的表达精度。
They've got precise words which mean something, and then you can summon greater precision with those precise words.
所以我对你观点的理解是——我也曾撰文讨论过——人工智能就像没有文档说明的API接口。
And so a way I was thinking about what you're saying is that, and I've written about this, AI is like undocumented APIs.
对吧？
Right?
正常的API每个功能都有详细说明，就像写着：你可以这样做，也可以那样操作。
So normal API, every function is like written out, and it's like, you can do this and you can do that.
所以我得到了20个功能，所有东西都在这里。
And so I got 20 functions and here's everything is there.
对吧？
Right?
有了人工智能，它能做很多连编写它的人都想不到的事情。
With AI, it can do lots of things that even the people who wrote it up.
所以它能做什么要神秘得多。
So it's much more mysterious as to what it can do.
你只需要尝试各种方法。
You just have to try things.
对吧？
Right?
所以我是从另一个角度思考这个问题，想到了图形用户界面。
So the way I was thinking about this from a different angle was to think about GUIs.
哦，是的。
Oh, yeah.
是的。
Yeah.
图形用户界面正在执行多项功能。
What the what a GUI is doing several things that a GUI is doing.
其中之一是告诉你开发者创建的所有功能。
One of them is it's telling you all the features that the developers have created.
这部分之所以成为一场革命，原因之一是你知道这些功能是什么，而且不需要记忆键盘命令。
And part of the reason that was a revolution is, a, you knew what they were and you didn't need to memorize keyboard commands.
是的。
Yeah.
但另一方面，你可以容纳更多功能，因为你不再受限于自己记下的键盘命令数量。
But b, you can actually have more stuff because you're not constrained by the number of keyboard commands that you've been writing down.
所以你可以有数百个功能，而不是像你只能记住那么几个，你可以直接把更多东西加到菜单里。
So you can have hundreds of functions instead of, like, you know, you only you can just put them all you can just add more shit to the menus.
对。
Yes.
但另一部分是，图形用户界面还在向用户传达大量累积的决策和机构知识，告诉你此时该做什么才是正确的
But the other part of it is that the GUI is telling the user a whole bunch of accumulated decision and institutional knowledge about what the right things to do at this point would
。
be.
对。
Yes.
那
That's
没错。
right.
所以如果你处于一个工作流程中，而不是面对一个空白屏幕，你知道，就像在Photoshop或Excel里那样，情况就不一样了。
And so if you're in a workflow as opposed to just a blank screen, you know, it's it's one thing if you're in, like, Photoshop or Excel.
是的。
Yeah.
它可以在提示中引导你。
It can it can prompt you on the prompt.
但如果你在Salesforce的工作流程中，那么就已经有一个决策决定了：我会在这里给用户提供这五个选项，而不是750个。
But if you're in a workflow in Salesforce, then there's a decision taken that says I'm going to offer the user these five options here and not 750 options.
是的。
Yeah.
而对于一个提示，你完全没有这些。
And with a with a prompt, you don't have any of that.
所以你必须闭上眼睛思考一下，嗯，我在这里会做什么呢？
So you've got to shut your eyes and think for a minute of like, well, what would I do here?
而你并没有那种帮助。
And you don't have that help.
你知道，Karpathy也讨论过这一点，但我确实认为AIOS（人工智能操作系统）有其发展空间。
This is, you know, Karpathy has talked about this also, but I do think there's room for AIOS.
对吧？
Right?
就像，某种意义上——我们可以稍后讨论加密货币，但我认为人工智能和加密货币实际上都是操作系统级别的创新。
Like, a sense, and we can talk about crypto in a second, but I think AI and crypto are both actually operating system level innovations.
举个例子，可能有人会把它做成一个应用程序或可下载的东西，就像在Mac上添加一个覆盖层那样运行。
And for example, it may be someone who just does it as an app or like a downloadable thing and just does as a layer on top of the Mac.
但如果你能全面了解Mac上发生的所有操作，你就可以建议使用哪些应用程序，建议下载哪些应用，或者提示说‘嘿，你可能需要更改这些键盘设置’。
But if you have the full context of all the actions that are happening on your Mac, you can suggest which apps to use, suggest which apps to download, suggest, hey, you probably want to change these keyboard settings.
所以，这么说可能有点有趣，但Clippy（微软Office助手）终于被证明是正确的了。
And so like there's, you know, it's funny to put it this way, but Clippy is finally vindicated.
一个适用于所有场景的Clippy。
Clippy, but for everything.
对吧？
Right?
是的。
Yeah.
因为现在Clippy可以变得非常、非常、非常、非常聪明。
And because Clippy can now be really, really, really, really smart.
对吧？
Right?
你知道，这是安德森说过的话。
You know, was Anderson's line.
就像，科技领域的所有东西都能行得通。
It's like, everything in tech works.
只是什么时候。
It's just when.
对吧？
Right?
关于Clippy这件事有趣的地方在于，有人还提出了一个观点，就是你实际上会想给AI虚拟形象、AI智能体加上人脸。
And even the thing that's interesting about the Clippy thing is somebody also made a point, which is that you actually want to put faces on your AI avatars, on your AI agents.
所以你可以从Clippy或其他十几种形象中选择。
So you could pick from Clippy or 10 other kinds of things.
这么做的原因可能有点反直觉，但像你我这样的人之所以能使用ChatGPT和Claude等工具，是因为我们熟悉界面交互。
The reason you want to do that, this is counterintuitive, but people like you and I can use ChatGPT and Claude and what have you, because we're familiar with interfaces.
但它们之所以对一亿用户来说直观易用，真正原因是这些人已经习惯了与另一端的人类聊天。
But the reason they're actually intuitive to a 100,000,000 people is they're used to chatting with another human on the other side.
所以他们已经在将聊天框建模为类似人类的回应，因为他们长期使用WhatsApp、Facebook Messenger、Instagram聊天或类似工具。
So they're already modeling the chat box as being a human like response, because they've been using WhatsApp or Facebook Messenger or Instagram chat or something like that for a long time.
对吧？
Right?
但当它脱离聊天框环境，像是在屏幕上给出建议时，你会希望弹出一张脸，这样他们就能联想到：好的，这个人之所以提出这个建议是因为这就是他们的身份。
But when it's outside of that chat box environment, and it's like suggesting on the screen, you kind of want a face to pop up so they can associate, okay, this person is suggesting this because that's who they are.
他们某种程度上将那种个性映射到了AI代理上。
They kind of map that personality onto the AI agent.
所以你可以选择不同类型的Clippy助手，它们会提示你该做什么，或者直接帮你完成。
And so you could choose from different kinds of clippies that would give you prompts on what to do, or or it just does it for you.
这算是一种可能性。
That's sort of a possibility.
但我觉得人们不太喜欢它自动执行操作。
But I don't think people like it when it does it.
他们希望能够先审批确认再
They want they want to be able to approve it before they
做吧。
do it.
我认为这里有一种关于人们如何概念化这个东西及其运作方式的感觉。
I think there's a sort of sense in here of how people conceptualize what this thing is and how it works.
我记得谷歌的约翰·普拉瑟罗给我看过一张图表，一张关于‘最好’与‘便宜’的谷歌趋势图。
I remember John Pratherow at Google showing me a chart, a Google Trends chart of best versus cheap.
最好与便宜。
Best versus cheap.
嗯嗯。
Uh-huh.
是的。
Yeah.
所以最好的做这个，便宜的做那个。
So the best does this and cheap does that.
那坐标轴是什么？
And what are the axes?
跨越时间。
Crossing over time.
所以是谷歌趋势。
So Google Trends.
那么“最佳”这个词的搜索频率是多少？
So what's the frequency of the word best?
所以一开始是便宜手机，然后变成最佳手机？
So it starts with like cheap phones and then it goes to best phones?
是的。
Yes.
所以论文的观点是，互联网正在从价格比较工具——即你已经知道自己想要什么，这处于销售漏斗的底端——转变为推荐、策展和建议平台。也就是说，你是在那里寻找建议的。
And so And so thesis was that this was shifting from the Internet as price comparison, where you'd already knew what you wanted, and it's at the top of the and that's at the the bottom of the funnel, to the Internet as recommendation, curation, suggestion Is that where you're looking for suggestions.
这真有意思。
That's so interesting.
让我看看是否
So let me see if
我能理解这种心理。
I can understand the psychology.
所以当...当...所以
So when the when the So
你必须给出它，所以它从2.20美元开始，从2004年开始。
you must give it so it starts from $2.20 it starts from 2,004.
好的。
Okay.
所以在2004年，你上网时已经知道自己想要什么，然后寻找最便宜的x是什么？
So in 2004, you go on the Internet and you already know what you want and you look for the cheap what is the cheap x?
然后你输入一个搜索词或某个产品之类的。
And then you put in a scoop or you put in a product or something.
嗯哼。
Mhmm.
而随着时间的推移，这种情况会减少
Whereas over time, that goes down
而最佳选择的价格上涨了。
And best goes up.
而最佳选择的价格上涨并超过了它。
And best goes up and crosses it.
图表上形成了一个完美的交叉，真不巧。
It's a perfect x on the chart, unfortunate.
但核心观点是你们正在向漏斗更上游发展。
And but the thesis is you're going further up the funnel.
你越来越希望网上有人告诉你什么是最好的X或Y。
You're looking more and more for, I want someone on the internet to tell me the best X or Y.
而以前，你会从杂志、报纸之类的渠道获得这些信息。
Where previously you'd have got that from the magazine or newspaper or something.
关于这一点，有两个方面。
There's two things about that.
首先，你知道安迪·格鲁夫关于双指标的说法吗？
The first is, you know Andy Grove's thing about the paired metrics?
安迪·格鲁夫每当有人优化销售这类指标时，通常会先从优化数量入手。
So Andy Grove, whenever anybody's optimizing like sales, for example, will usually start recruiting, they'll start by optimizing quantity.
但有时候你优化了数量，质量却会下降。
But you can sometimes optimize quantity and then quality drops off.
是的，没错。
Yeah, yeah.
对吧？
Right?
所以数量是容易衡量的。
So quantity is easy to measure.
就像我们雇了多少人这样，但质量是指他们有多优秀，对吧？
It's like just the number of people we hired or whatever, but quality is how good were they, right?
因此第二个配对指标通常是质量指标，数量是廉价的，对吧？
And so that's the second paired metric is usually a quality metric that, and so quantity is cheap, right?
人们从廉价开始，然后质量是最好的，他们就会追求最好的。
And people start with cheap, and then quality is best and they go to best.
所以这是看待这个问题的另一个视角。
So that's another lens on this.
第三个视角，我认为我的解释——也许与实际发生的情况不同——是人们刚开始尝试某个领域时，他们只想先用便宜版本试试看。
A third lens, what I thought my explanation, maybe it's different than what actually happened was when people are just trying out a space, they just want the cheap version to try it out.
一旦他们决定投入某个领域，比如便宜的数码相机、便宜的无人机之类的，他们想先尝试一下。
And once they've committed to a space, like for example, cheap digital camera, cheap drone, something like that, they want to try it out.
对吧？
Right?
他们想以低成本尝试一下，先试后买。
And they want to try it out at low cost, try before you buy.
一旦他们决定进入这个领域，就会想要市场上最好的无人机，因为我想要一款好的
And then once they're committed to a space, then they're like, I want the best drone out there now because I want a good
那么分析就是低成本无人机和顶级无人机的对比。
Well, analysis then would be cheap drone versus best drone.
对。
Right.
但我认为那正是
But I think the That's what
我以为你就是这个意思。
I thought you were saying.
是的。
Yeah.
不是。
No.
但你说的是总体上更便宜就是最好的。
But you're saying cheaper is best overall.
是的。
Yes.
总体上。
Overall.
但我很想按类别来看待这个问题。
But I'd love to see it category by category.
我不会惊讶于看到这种情况按类别逐步发生，但也可能不会。
I wouldn't be surprised to see that happen category by category, but maybe not.
嗯，这里还有另一个不同的观点，有点像我今早在我们小组讨论中谈到的，就是这种无限产品的问题——你怎么知道该买什么？
Well, there's a different point there, which is sort of what I was talking about in our panel this morning, which is this infinite product, so how do you know what to buy?
过去的情况是，你会先看杂志，然后去网上找最便宜的购买渠道，或者你已经知道自己要什么，而现在你是直接上网去寻找该去哪里做这件事。
And it used to be that you'd start with a magazine and then you'd go to the internet to find the cheap place to buy it, or you know what you wanted, and now you go to the internet to find what's the right place to do this.
所以互联网已经变得更加像是一个默认选择了。
So the internet has become much more kind of a default.
但实际上，促使我产生这个想法的是，你也可以去试试谷歌趋势，我做了个图表，研究了像‘嗯’这样的问题。
But actually, the thing that the thought that prompted me to that was you can also go and play with Google Trends, and and I did a chart played with like how Mhmm.
为什么、在哪里、是什么，这类更基础的问题，你真的需要进入谷歌内部才能正确进行那种分析。
Why, where, what, like more kind of basic questions, and you really need to be inside Google to do that analysis properly.
是的。
Yes.
没错。
Exactly.
是的。
Yeah.
但关键在于人们有多少是在谷歌上进行对话式查询，而不是输入关键词。
But it's that sense of how much are people doing conversational queries into Google as opposed to typing keywords into Google.
嗯哼。
Mhmm.
而那些不太适合用谷歌搜索的东西，比如，什么是...好吧，这很有趣。
And things that are not really a Google query, like, what is a is not Well, it's funny.
可能对谷歌没什么帮助，但人们还是这么用。
Probably doesn't help Google, but that's still how people use it.
多年来人们被训练要去掉所有介词，去掉所有那些东西，只做关键词搜索。
People were trained for years to not to to to remove all prepositions, remove all that stuff and just do keyword e's.
而现在我们被训练成相反的方式，写出完整英语句子，就像提示是新的搜索方式，但这完全是不同的行为，对吧？
And now we're trained the opposite, to write full and complete English sentences like prompting is the new searching, but it's a completely different behavior, right?
请继续。
Go ahead.
嗯，我刚才想说这是其中一个...这其实是个有点跑题的观点。
Well, was just gonna say this is one of the There's a such tangential point to that.
人们早期在互联网上利用AI和大型语言模型部署的一个简单明显应用，就是各种类型的自然语言查询，或者更准确地说，是不同类型的查询方式。
One of the early easy obvious things that people have deployed with AI, with LLMs on the Internet is different kinds of is is sort of natural language queries or different not so much natural language, but like different kinds of query.
所以人们常说的经典案例就是沃尔玛推出的功能：现在你可以搜索'我应该买什么去野餐'。
So the the canonical one people talk about is Walmart saying, now you can search for what should I buy to take on a picnic.
嗯。
Mhmm.
这并不是一个数据库查询。
Which isn't a database query.
五年前，对于谷歌、沃尔玛或亚马逊来说，这种搜索根本行不通。
And for Google for for for Walmart or for Amazon five years ago, that search just wouldn't work.
因为除非某个产品被标记为与野餐相关，否则它不会出现。
Because unless there's a product that's, like, tagged with picnic, it's not gonna come up.
而现在，有了一个具备世界模型的大型语言模型，它对如何
Whereas now there's an LLM with a world model that has some sense of how you
回答这个问题有一定的理解。
might answer that question.
这是世界模型吗？
Is it a world model?
至少是一个网络模型。
Is, at least a web model.
但无论如何，这是一种不同的查询方式。
It's a different kind of query anyway.
你不是在做SQL查询，而是在做别的事情。
You're not doing a SQL query, you're doing something else.
没错。
That's right.
而且我认为，你知道，其中一个
And I think, you know, one of the
有趣的事情是，计算机一直非常擅长第一种确定性计算，比如SQL查询和运算。
things that's interesting is computers are, we knew they were, very good at that first kind of deterministic computation, the SQL query, the calculation.
它们的设计本来就是为了进行数学计算。
That's what they're built for doing math.
对吧？
Right?
是的。
Yeah.
而现在，它们在处理概率性任务方面也变得很擅长了，对吧？
And now they've gotten good at probabilistic kinds of things, right?
所以这就像是系统一和系统二的思维方式，对吧？
So this would be like system one and system two thinking, right?
概率性思维就像是快速直觉反应，而这就像是逻辑计算。
Probabilistic is like the quick impression, and then this is like the logical calculation.
所以实际上它在核心能力上很强。
So it's actually good at the heart.
对人类来说更难的是那些冗长复杂的数学方程，而它能毫无差错地完成。
The thing that's harder for humans is the long involved mathematical equation can do that errorlessly.
而现在它也能做另一种事情了。
And now it can also do the other kind of thing.
所以这确实预示着最终会有某种融合，人工智能可以——我的意思是，这就像是AI工具使用之类的，比如它检测到需要进入系统一模式，就开始为此调用Python？
And so it does suggest that there'd be some synthesis of that eventually, where an AI can I mean, this is like AI tool use or what have you, like it detects that it needs to go to system one and it starts invoking Python for that?
这方面正在改善，但两年半过去了，它的表现却出乎意料地不算惊艳，对吧？
And this is getting better, but it's surprisingly not amazing two and a half years in, right?
当它需要进入确定性模式时。
When it needs to go deterministic.
嗯，我上次关于这个话题写的长文是关于审视OpenAI推出的深度研究。
Well, I wrote last long thing I wrote about this was about looking at deep research, which OpenAI launched.
而审视新事物时的一个陷阱就是根据旧事物的重要标准来测试它。
And one of the kind of traps in looking at the news thing is to test it based on what was important to the old thing.
嗯哼。
Mhmm.
所以，你知道，就像看着Apple II问它是否达到大型机的正常运行时间标准一样。
So, you know, to look at the Apple two and say, does this match the main the uptime of a mainframe?
不对。
No.
所以它没用。
So it's useless.
嗯，不是。
Well, no.
但这不是正确的问题。
But that's not the right question.
是的。
Yes.
你能在iPhone上编写构建Excel模型吗？
Can you write an build an Excel model on an iPhone?
不能。
No.
但这不是重点。
But that's not point.
它仍然可以取代个人电脑。
It can still replace PCs.
而我提到这个的原因是，Deep Research（OpenAI推出的这个产品），不管它是什么，每月100美元之类的。
And and the reason I mentioned this is so so so so Deep Research, OpenAir launched this thing, and it's whatever it was, a $100 a month or whatever.
但我后来看了它的营销页面。
So I but then you look at the marketing page.
营销页面上展示了它在做一个关于移动领域的研究项目，正如我们所说，我对这方面很了解。
And the marketing page shows it doing a research project about mobile, which as we said I know a lot about.
而且它的答案错了。
And it got the answers wrong.
这是第一次，那是在验证。
It's first That's the verifying.
你看，你早就知道，你能看出它是
See, you knew, you could tell that it was
错的，但看起来完全像那么回事。
wrong, but it looked plot Exactly.
所以问题是，它在好几个层面上都搞错了。
So this is the thing, and it got stuff wrong in several levels.
人们现在还记得我大约两个月前写的东西。
People can remembering now what I wrote like two months ago.
具体来说，它做了一个表格，显示多个国家的智能手机普及率以及操作系统市场份额。
And so there was a specific it was make a table which shows mobile smartphone adoption in a bunch of countries, and then the operating system market share.
然后这就像一个实习生教学时刻。
And then this is like an intern teaching moment.
首先，采用率具体指什么？
Because first of all, what does adoption mean?
是指设备销量份额、装机基数还是应用商店销售额？
Does that mean unit sales share, installed base, app store sales?
你具体需要我提供哪些指标？
Like, what metrics specific are you asking me for?
对。
Yeah.
然后它为得出的数字提供了一个来源，就是Statista。
Then it had given a source for the number it had come with, which was Statista.
而Statista是一个窃取他人数据并重新包装的聚合平台。
And Statista is an aggregator that steals other people's data and repolishes it.
是的。
Yeah.
当你完成一系列注册流程后，会发现实际数据来源其实是Kantar。
And when you jump through a bunch of registration hoops, you discover that the actual source was, I think, Kantar.
Kantar。
Kantar.
这是一家广告公司。
It's an ad agency.
属于GroupAirMittal的一部分，我以为它是其中一家。
Part of GroupAirMittal, thought it was part of one of that.
这是消费者调查数据。
It's consumer survey data.
所以这是一家正规公司。
So it's a proper company.
所以这是真正的正规消费者调查数据。
So it was actual proper consumer survey data.
但当你查看Canton图表页面时，你会发现Deep Research把数字弄反了。
But the two things that so then when you go to the Canton chart page, you discover that Deep Research had got the numbers the opposite.
所以它把百分比颠倒了。
So it had flipped percentages.
我明白了。
I see.
然后它还说——
And then it had also said-
对。
Right.
因为这是一个错误的统计，
Because it was an ab act,
这里说你可以登录，它从网站上抄错了数据。
it said here that you can log in It had copied them from the website wrong.
我明白了。
I see.
而它提供的另一个来源是StatCounter，而StatCounter
And then the other source it gave was stat counter, and stat counter
只是用了同样的错误数据。
was- Was just using the same wrong data.
这是一种流量指标。
Which is a traffic measure.
对。
Right.
所以这并不能告诉你任何选项。
So that's not gonna tell you an option.
因为高端手机使用得更多，iPhone 使用得也更多。
Because high end phones get used more and iPhones get used more.
所以这里有很多地方，你会觉得，这完全是实习生会犯的错误。
So and there's a bunch of things in here where you'd like, this is what I'd expect from an intern.
对。
Right.
我会回去说，不，这才是我所说的采用率，这才是一个可靠的数据来源，而那个不是。
I would go back and say, no, this is what I mean by adoption and this is a good data source and that isn't.
对。
Right.
这就像是一个很棒的第一个版本。
And it was like a great first version.
问题是，第一，它把数字抄错了，这不是我对实习生的期望，至少不是一个好实习生该犯的错误。
The problem is A, it had copied the number out wrong, which is not what I would expect from an intern, or at least not a good intern.
但其次，我得是个移动分析师才能知道这些
But secondly, I'd have to be a mobile analyst to know any of these
事情。
things.
这正是我们想要验证的核心点。
And that's a verifying thing that was getting at.
而这件事的核心在于，你知道，所有这些人都在深入研究并说，这对于研究你一无所知的事物来说太棒了。
And this is the kind of the core of it is I, you know, all these people were looking at deep research and saying, this is fantastic for researching things you don't know anything about.
我当时就说，不对。
I was like, no.
不，它是
No, it's
如果你需要大量关于你非常熟悉领域的材料，那它确实很棒。
fantastic if you need a bunch of material about something you know a lot about.
完全正确。
Exactly.
所以这就是为什么我认为当前形态的AI更适合被看作增强智能——因为你对某个领域了解越多，就越擅长提问（因为你有更丰富的词汇量），也越擅长验证结果（因为你掌握更多相关事实并能进行交叉验证）。
So that's thing, that's why I think AI in its current incarnation is better thought of as amplified intelligence, because the more you know about a field, the better you are at prompting, because you've got better vocabulary, and the better you are at verifying because you know more facts about it and you have more cross cutting checks.
这一点在视觉领域的适用性稍弱，但认识到这个重要局限性很关键——对于视觉内容我们有一套完全不同的验证系统，也就是我们的眼睛，对吧？
And that is less true for the visual area, but just identifying that is a very important limitation where you have a completely different system you can use for the visual stuff, which is just your eyes, right?
你不必使用那些，你知道，我们只是用不同的硬件来快速查看，比如手部动作之类的。
You don't have to use the, you know, we just different hardware for quickly seeing, you know, this way the hands or something like that.
而如果那是个猴子大脑。
Whereas if that It's a monkey brain.
那是个猴子大脑。
It's a monkey brain.
没错，对吧？
Exactly, right?
所以现在这是个有趣的问题。
So that's now an interesting question.
这就是，你知道，我和卡帕西亚讨论的是，有没有办法将部分或一些非视觉内容转化为视觉提示，让你能立刻看出它是错的？
This is, know, Karpathya I were discussing this is, is there some way to turn some or a subset of the non visual things into visual cues where you could see it was wrong immediately?
那么我给你举个小而简单的例子。
So give me I'll give you a small, simple example.
假设它生成了一个音频文件。
Let's say it generated an audio file.
对吧？
Right?
你知道的，就像音频文件的频谱图。
You know you know like a spectrogram of an audio file.
对不对？
Right?
你或许能立刻看出那里是否有某些人为痕迹。
You could maybe immediately see if there's some artifact there.
对吧？
Right?
这是一个简单的例子。
That's a trivial example.
所以
So
我觉得这个概念非常有趣。
I think this I'm it's a fascinating concept.
我一直在想，这样的划分是不是恰当的。
I'm I'm I would wonder whether that's the right split.
好的。
Okay.
至少目前我发现这种划分是有用的。
It's at least one split I found useful for now.
那你有什么想法？
But what are you thinking?
嗯，我考虑的划分方式是自然语言生成文本这方面已经非常完善了。
Well, the split I was thinking was that the natural language generation to make text is perfect.
确实。
Mhmm.
所以生成的文本在语法上总是正确的。
So the text is always grammatically correct.
这倒是事实。
That is true.
是的。
Yes.
但底层的模型，文本中呈现的事实可能是错误的。
But the model underneath, the facts presented by the nut in the text might be wrong.
对。
Yes.
这某种程度上是在欺骗我们，因为我们会看到文本是正确的，而且它看起来很有信心
And that's sort of deceptive to us because we see the text is correct and it looks confident and
所以，是的。
so Yeah.
那是
That's
对的。
right.
而在图像中，比如你让它生成某个人的图片，所有地方都很完美，只是那个人长了六只手。
Whereas in an image, like, you ask it for a picture of somebody and everything's perfect except the person's got six hands.
嗯。
Mhmm.
我不确定从概念上讲，这种问题到底是什么被简化了？
I'm not sure conceptually what is it that that's flattened?
是你在一层里看到了两个东西吗？
Is it you're seeing two things in one layer?
还是说，你明白我的意思吗？
Or is it that do you see what I mean?
我明白你的意思。
I I see what you mean.
我认为情况是，或者说这可能是另一个层面的问题，也许这里有一个不同的观点，就是如果你要求生成一辆汽车的图片，是的。
I think what it is Or is it that it's a different level of well, maybe then maybe there's a different point here, which is if you ask for an image of a car Yeah.
而我实际上就是这么做的。
And the car I I actually do this.
我要求生成一辆六十年代法国梦幻跑车。
I ask for a fantasy nineteen sixties French sports car.
正确。
Correct.
它会看起来像法国风格。
It will look French.
它会看起来像一辆跑车。
It will look like a sports car.
它会拥有四个轮子。
It will have four wheels.
它可能有两个方向盘。
It might have two steering wheels.
是的。
Yes.
没错。
That's right.
两个方向盘就像是文本生成器中的语法错误或拼写错误。
The two steering wheels is the equivalent of a grammatical mistake or spelling mistake in the text generator.
对。
Yes.
因为然而，也可能这辆车的平衡完全失调，如果试图转弯就会翻车。
Because However, it may also be that the balance of the car is all wrong, and it would flip over if it tried to go around a corner.
但你得是汽车专家才能知道这一点。
But you'd have to be an automotive expert to know that.
所以我说存在不同的
So I'm saying there's different
错误级别。
levels of error.
没错。
That's right.
你的意思是两个方向盘就像拼写错误，但AI很少犯拼写错误，而两个方向盘却是常见错误。
What you're saying is the two steering wheels is like a spelling error, but spelling errors are very rare for AI, whereas the two steering wheels is a common error.
对吗？
Right?
我认为这与扩散模型和Transformer模型的工作原理差异有关。
And I think that has to do with just the way diffusion models work versus how transformers work.
这就像我会给出的一个高层次答案，扩散模型的处理方式更偏向局部化。
That'd be like one high level answer I'd give, where it's doing like kind of, it's more local with the diffusion model.
就像方向盘可能在局部正确，但整体布局却是错误的。
And you can be locally correct with the steering wheel, but globally incorrect.
而拼写的局部正确通常就意味着整体正确。
Whereas locally correct with spelling is usually correct.
这可能是其中一个——是的，我明白了。
That's like maybe one- Yeah, I see.
很有用。
Useful.
这是一个答案。
That's one answer.
第二点是，我认为空间非常有限。
The second is that with There's only a small space, I think.
例如，我们天生擅长识别人脸，所以能检测到面部非常细微的差异。
Like for example, we are optimized to recognize faces, so we can detect very subtle differences in faces.
但如果我给你看五张不同的静态噪点图，对吧？
But if I gave you like five different sheets of like static noise, right?
即使存在非常清晰的数学模式——比如这些都是同一物体的傅里叶变换结果，而我特意加入这张——在你看来它们完全就是噪点。
Even if there are very clear patterns, like mathematically, like these are all like Fourier transforms of the same object, and this is the one I add, they just look like total noise to you.
计算机会认为这12张是相同的，而这张是异类，对吧？
A computer would be like, these 12 are the same, and this one is odd one out, right?
所以在某种意义上，我们的眼睛是针对一个非常低维度的事物集合进行了优化的，这些事物就是现实世界中出现的那些。
So in a sense, our eyes are optimized for a very low dimensional set of things, which are the things that occur in the real world.
就像那些是我们能够识别出来的东西。
Like those are the things we can pick out.
这也说明我们的眼睛——狗在运动感知方面比我们更强。
Which is also that our eyes are dogs are better at motion than us.
是的。
Yes.
没错。
Exactly.
所以即使是眼睛也会因物种而异。
So even eyes are different depending on the species.
说得对。
That's right.
正因如此，我们实际上拥有一种...因为它们无法在静态中检测模式，那就像是维度过高的空间。
So so so because of that, we actually have a like, because they they can't detect patterns in static, that's like too high dimensional space.
我认为文本有点像这样，因为它能描述——关于AI发展最让我惊讶的一点，我们之前讨论过这个问题就是，我惊讶于纯文本竟然能带来如此大的价值。
I think text is kind of like that because it can describe One of the most, I mean, surprising things to me about how AI has evolved, we were talking about this question before is, I was surprised you could get so much mileage out of pure text.
原因是——什么如此之大，先生？
The reason is- So much what, sir?
纯文本带来的巨大价值。
So much mileage out of pure text.
我对此感到惊讶的原因是，你知道，你会觉得——你
And the reason I was surprised by that is, you know, you'd think- You
指的是推理以及所有看起来像推理的东西。
mean like reasoning and all stuff that looks like reasoning.
推理和
Reasoning and
还有空间操作。
also spatial manipulation.
拥有摄像头、拥有眼睛、观察世界、进行推理，就像婴儿那样等等。
Having cameras, having eyes, seeing the world, reasoning about it, like a baby and so on and so forth.
人类用文本为那个世界中的多少事物赋予了机器可读的标签，这真是令人惊叹。
It is amazing how much of that world humans have assigned machine readable labels to with text.
而且这种方式的效果如此之好，确实非常出人意料。
And the way that, it's very surprising how well that worked.
就像语言，我想说的是，用大约40个词就能描述清楚，这就像代码一样。
Like language, what I'm trying to say is in a few, in like 40 words, you can describe, it's like code.
用大约40个词你就能描述许多许多不同种类的事物，对吧？
You can describe many, many, many different kinds of things in like 40 words, right?
而且它更加通用。
And it's just more general.
有时候当你真正深入某个领域时，反而会比局外人更对突破感到惊讶。
It's one of those things where sometimes when you're really close to a space, you're actually more surprised by a breakthrough than if you're farther away.
我应该说，即便目睹了2010年代中期的风格迁移技术、ImageNet数据集以及各种基准测试等等，我仍然惊讶于它能达到...你知道马尔可夫链吗？
And I should say, even looking, seeing all the style transfer stuff in the mid-2010s and seeing ImageNet and seeing the benchmarks and so on and so forth, I was surprised that it got You know where Markov chain is?
如果你看过GPT-3之前的技术成果，那些内容虽然半连贯，但看起来并不像在向某个方向收敛。
Well, if you saw the stuff before GPT-three, right, it was like semi coherent, but it didn't look like it was converging on something.
它看起来只是反复重复，诸如此类。
It just looked like it'd repeat itself many times and what have you.
而它仅仅基于语言就取得了如此突破，这简直反直觉。
And the fact that it broke through to what it did, just based on language, was so counterintuitive.
我认为这是因为语言是一个极高维度的东西。
And I think it's because it's such a high dimensional thing.
它捕捉了世界中的众多不同方面。
It captures so many different aspects of the world.
比如，你能在世界上感知到的任何事物，都有对应的词语来描述它。
Like anything you can perceive in the world, there's a word for it.
甚至有很多词语来描述它。
There's many words for it.
而且我们还有数十亿人，过去二十年来一直在打这些字。
And then we also have billions of people who've been typing those words for two decades.
对吧？
Right?
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。