揭露AI骗局的人：AI是世上最大的骗局，他们都知道！| Ed Zitron

接受这个事实，意识到这个世界并不是由精英统治的，这有点让人沮丧。

Reconciling with that and realizing that the world is not controlled by people like a meritocracy is kind of grim.

所以很容易就会想，不。

So it's easy to be like, no.

他们没犯错。

They're not making a mistake.

一定是我漏掉了什么。

I must be missing something.

而这正是他们想要的。

And that's what they want.

所以，你知道，我回顾技术突破的历史，想到各个行业，我最喜欢的一本关于这个主题的书是《创新者的窘境》。

So, you know, I think back through the history of technological breakthroughs, and I think about I mean, you can look at different industries, and one of my favorite books on this subject is The Innovator's Dilemma.

我读过。

I read it.

书中谈到，最终颠覆或改变一个行业的创新，往往一开始表现更差，经济上说不通，也没有客户要求它。

And one of the things it talks about is how the the innovation that ends up taking out or transforming an industry often starts worse, doesn't make economic sense, none of your customers are asking for it.

这通常就是我们最终忽视它的原因。

And this is typically why we end up ignoring it.

比如，在十九世纪，有马车。

So, like, you've got horse and carriages in the eighteen hundreds.

对。

Mhmm.

在十九世纪的人看来，那是很棒的交通工具。

Amazing form of transport according to the eighteen hundreds, you know, people of the eighteen hundreds.

然后，一种叫汽车的东西出现了。

And then you have this thing called cars come along.

问题是，汽车总是抛锚。

Now the problem with cars is they broke down all the time.

有点像现在AI会胡言乱语。

It's kinda like AI hallucinates now.

它们更贵，而且经济上也不划算。

They were more expensive and the the economics of it didn't make sense.

那还不如走路呢，买什么车啊。

You might as well walk than buy a car.

当时有条法律规定，你得在车前面举着红旗走，还得雇个人在前面挥旗。

There was a law at the time that meant you had to walk in front of it with a red flag and wave and someone had to get to employ someone to walk in front of it waving a red flag.

显然，这更糟糕。

Obviously, it's worse.

这就像是个更差的解决方案。

It's like a worse solution.

不过，这些颠覆性创新的东西，它们的增长上限更高。

However, these things that are disruptive innovations, they have a higher ceiling of growth.

所以它们最终会超越马车。

And so they eventually overtake the horse.

当我想到这个类比，再结合现在的情况，我就觉得，好吧，现在它还不完美。

And I when I think about that analogy in the context of all of this, I go, okay, it's imperfect at the at the moment.

经济模式还没有完全理顺。

The economic models aren't perfectly ironed out.

他们还在想办法降低成本、完善基础设施等等。

They're still figuring out how to make it cheaper, the infrastructure, etcetera.

但如果你想想改进的速度和其他东西比，比如说编程，我能训练一个人类程序员提高多少产出，而一个AI代理又能提高多少？

But as if you think about the rate of improvement versus other, you know, let's say coding, how much can I train a human coder to improve and to increase their output versus an AI agent?

你想想，如果你想象这些AI工具每个月有5%的改进速度，总有一天，你知道的，然后你再想象成本降低5%，就像我们对互联网和汽车做的那样。

One would go, if you just imagine any rate of improvement in these AI tools, some point if you just imagine a 5% rate of improvement per month, at some point, it's you know, and then you imagine a 5% reduction in cost, which is what what we did with the Internet, what we did with cars.

对。

Yeah.

但是那个……

But that

摩尔定律。

Moore's Law.

摩尔定律是个理论，而且摩尔定律不适用于GPU。

Moore's Law is a theory, and Moore's Law is not with GPUs.

让我来解释一下。

So let me let me actually explain.

所以，英伟达。

So NVIDIA.

英伟达发明了，我记得是在2000年代。

NVIDIA invented think it was in the two thousands.

他们推出了一个叫CUDA的东西，这是底层的软件库，也是让软件在GPU上运行的方式。

They put out something called CUDA, which is the underlying software library and the way to run software on GPUs.

对。

Mhmm.

他们花了整整十年甚至更久，才让它能用于数据分析，这是早期的应用之一，比如映射之类的。

Took them solid decade or more to make it something where they could do data analytics, one of the early things, mapper and such.

然后当AI出现的时候，他们已经积累了大量经验。

And then when AI came along, they'd had lots of experience with it.

但即便如此，这家公司拥有的资金、关注度、背后的天才，以及专注于提升效率的人，都远超任何人的期望。

But nevertheless, this company has got more money, more attention, more geniuses behind them, more people focused on making their things more efficient than anyone could ever ask for.

英伟达，对于不了解的人来说，就是制造芯片的公司。

And NVIDIA, for anyone that doesn't know, makes the chips.

对，还有我刚才提到的那个CUDA，他们就是靠CUDA让生成式AI发展起来的。

They so and that CUDA thing I mentioned, they were the ones with CUDA, and CUDA allowed generative AI to grow.

好的。

Okay.

所以他们的芯片。

So they're they're chips.

芯片。

Chips.

是的。

Yes.

而且芯片是必需的。

And chips are needed.

这些就是装进数据中心的东西。

Those are the things that go into the data centers.

而且他们的芯片特别之处在于，可以在上面运行AI软件。

And their specific chips are the ones where you can run AI software on it.

所以训练和推理都在用这些芯片。

So the training runs and also the inference.

现在关键来了。

Now here's the thing.

就拿汽车举例。

The the car example.

那时候，可没有几乎所有数学家和科学家都涌进汽车行业。

Back then, you didn't have pretty much every mathematician and scientist going into the car industry.

也没有全世界政府联合起来，整天喋喋不休地讨论这件事。

You didn't have the combined world's governments never shutting up about this.

而且，他们很早就开始看好这个了。

And by the way, giving them credit early.

从2023年起，他们就说这是不可避免的。

Since 2023, they've been saying this is inevitable.

就像你说的，哪怕只是5%的提升。

Even what you said, 5% improvement.

我甚至不知道该怎么衡量这个，因为初级软件工程师还是能从环境、从别人怎么处理问题中获取经验和学习。

I don't even know how you'd measure that because a junior software engineer can still experience things and learn things from context, from how people deal with problems.

而且人们处理问题的方式，不是看看代码或者读几封邮件那么简单。

And the way that people deal with problems is not as simple as looking at the code or reading some emails.

这需要和人交谈时捕捉到的上下文线索。

It's context cues from speaking to a person.

这需要身处不同的环境。

It's being in different environments.

而且，LLM在编程中确实有用武之地。

And there may there are uses for LLMs in coding.

这点我不否认。

I don't dispute that.

但就算说提升5%，那到底是什么意思？

But even saying 5% bit, what does that mean?

是更擅长Rust吗？

Is it better at Rust?

它在C++方面更擅长吗？

Is it better at c plus plus?

我会说

I'd

生产力。

say productivity.

所以就像，对，交付了。

So just like, yeah, shipped.

如果是在编程的背景下，那就是交付的代码。

If we did it in the context of coding, it would be like shipped code.

问题就在这儿。

That's the thing.

那就像说，他是世界上最好的作家，因为他的通讯稿特别长。

That would be like, he's the best writer in the world because his newsletter's really long.

这种评价方式太离谱了。

That's an insane way of evaluating it.

说到编程，我的意思是，这甚至很难评估，因为软件本身是不是更好呢？

With coding, it would be I mean, it's even difficult to evaluate because it's is the software out there better?

这其实是一个很好的评估方式。

Is actually a great way of evaluating it.

我会说，普遍来说不是。

And I would say uniformly not.

我会说，谷歌、微软、亚马逊、Meta，尤其是天哪，Meta简直是个怪物，它们的软件标准更差。

I would say the standard of software across Google, Microsoft, Amazon, meta, especially god meta is a monstrosity, is worse.

GitHub，GitHub，今天早些时候有人在推特上发帖说，我们应该在GitHub恢复时收到通知，而不是在它宕机时，因为那样更可靠。

GitHub GitHub someone posted on Twitter earlier today, we should get a notification when GitHub is up rather than when it's down because that would be more reliable.

微软是世界上最大的公司之一，但在GitHub这件事上，他们连自己的屁股都擦不干净。

Microsoft is one of the largest companies in the world, they can barely wipe their own ass when it comes to GitHub.

奇怪的是，随着越来越多的人使用大语言模型，越来越多的企业要求——我真的是说要求——人们使用这些服务，软件的质量却在下降。

The quality of software is going down, weirdly enough, as more people use LLMs and more businesses demand, and I really do mean demand, that people use these services.

所以关于这一点，如果我们回到马和马车以及汽车的类比，好吧。

So on this point of if we go back to this horse and carriage and car analogy Okay.

假设我们处在今天的某个时间点，如果你想象一下技术有任何改进的速度，自从Trachypity出现以来，我们确实看到了。

Say that we're at whatever point today, if you imagine any rate of improvement in the technology, which we have seen since Trachypity came out Sure.

我记得Trachypity刚出来的时候，我在亚洲，我就在那里给我未婚妻展示它。

I remember when Trachypity came out, and was in Asia, I and was there showing it to my fiance.

我当时就说，看，它能做这个。

I was like, look, can do this.

它偶尔会胡言乱语，出错。

And it was hallucinating once in a while and getting things wrong.

我现在其实已经没有那种体验了。

I actually don't have that experience anymore.

我有时候觉得它的推理能力弱，但我不再有明显的幻觉了。

I have moments where I believe its reasoning is weak, but I don't have outright hallucinations anymore.

看到了吗？

See that?

我不同意。

I I disagree.

那你给我举个例子，说说你所谓的幻觉是什么。

So Give me an example of what you define as a hallucination.

好的。

Okay.

这个例子很好。

Great one.

我有一个彭博终端。

So I have a Bloomberg terminal.

对。

Yep.

上面有个特别有用的功能，叫AskBee。

The very useful thing they have on there is ask me.

比如你在彭博上查一下我们预计英伟达下个季度的收入是多少，它会运行一个叫BQL的东西，那是它自己的编程语言。

So when you do a Bloomberg inquiry to, like, look up what we think NVIDIA's revenue is gonna be next quarter, it runs something called BQL, which is its own programming language.

现在你不用去学那个了，直接输入到AskBee里，它就会帮你生成并运行。

Now instead of having to learn that, you can just type into AskBee and it will generate it and run it for you.

然后数据就给你调出来了，你也知道数据是从哪来的。

And so you get pulled up and you know where the data is coming from.

这样处理幻觉问题就处理得特别好。

Deals with hallucinations real well.

前几天我就想，你知道吗？

The other day, I was like, you know what?

我有点想搞点刺激的。

I'm getting a little spicy.

我打算查一下微软、谷歌、Meta和亚马逊这五年的股票增长率，我记得是五年。

I'm gonna look up the growth rate of stocks of Microsoft, Google, Meta, and Amazon over the course of five years, I think it was.

对。

Yeah.

然后我正准备把它复制粘贴到别的地方，放到Excel里看。

And I was about to I was copy pasted it over to something, looked it in Excel.

我正准备写那篇新闻通讯。

I was about to was writing the newsletter.

我说，微软的股票从来没到过575美元一股？

I went, Microsoft stock's never been $575 a stock?

你知道吗？

You know what?

当它是个小问题，比如股价这种小事，我还能发现。

When it's a cute little thing like, oh, it's a stock price, I kinda caught it.

没什么大碍。

It was no harm, no foul.

那还好。

That's fine.

但如果是医生用的转录工具，或者对冲基金依赖的金融模型，那就有点危险了。

But when you're talking about, I don't know, like a transcribing tool for a doctor or a financial model that a hedge fund is dependent on, at that point, it becomes a little more dangerous.

而且，软件包出现幻觉，比如你在重构代码库，它可能留下安全漏洞，或者直接弄坏东西。

And the thing is, a hallucination with a software package, for example, you're refactoring a code base, and it leaves the door open security wise, or it just breaks something.

然后你，我也不知道。

And you I don't know.

可能你已经用 vibe coding 写了六个月的代码了。

Maybe you've been vibe coding for six months.

你其实已经有一阵子没亲手写代码了。

You haven't really been coding with your own hands for a while.

可能有些东西你已经忘了。

Maybe you've forgotten a few things.

你还有这个漏洞要去找。

You have this slot to look for.

你会说，妈的，我不干了。

You go, fucking, I'm not doing it.

所以问题就变得加倍严重了，我真的不知道你怎么才能让他们改掉这个毛病，而且他们显然也没成功。

And so the problem has become multiplicative, and I don't really know how you train them out of that, and they've certainly not succeeded.

所以一方面，他们确实进步了。

So on one hand, they have got better.

但他们评估进步的主要方式之一，是专门为大型语言模型调整过的基准测试，因为你不能直接让他们做任务。

But one of the main ways they evaluate them getting better are benchmarks that are adjusted specifically for large language models because you can't just have them do tasks.

它们在这方面确实进步了。

They've got better at that.

他们找到了一些可以让它们执行的任务。

They found some tasks they can have them do on.

但即使是METR，就是m e t r，他们也有个东西，像是说，看看这张图表。

But even them like METR, m e t r, they have this thing where it's like, check out this chart.

看看它在执行任务方面进步得多快。

Look how much better it's getting at running tasks.

哇。

Wow.

它能持续运行一个小时。

It can go for an hour.

然后你一看，就会觉得，对。

And then you look at it, it's like, yeah.

而且成功完成任务的概率是50%。

And successfully completing them 50% of the time.

他们有一个幻觉排行榜，而且它主要关注的是基础任务。

They they have a hallucination leaderboard, and it really focuses on basic tasks.

而且它显示，根据Fiktariya幻觉排行榜的历史数据，四年的趋势表明，在简单的摘要任务上，幻觉率已经从四年前的约百分之二十一点八，下降到如今像Gemini和ChatGPT这样的顶级前沿模型的大约百分之零点七。

And it shows that the four year trend according to historical data from the Fiktariya hallucination leaderboard shows that hallucination rates on simple summarization tasks have plummeted from around twenty one percent, twenty one point eight percent four years ago, down to zero point seven percent roughly on today's top frontier models like Gemini and ChatGPT.

再说一次，这里的关键点是这些是在简单任务上的表现，这也符合我自己的体验。

Again, the the point of nuance here is that these are on simple tasks, which is kind of what I've experienced.

我在日常事务中确实感觉到幻觉变少了。

I've experienced it on day to day things that hallucinates less.

再说一次，这是关于改进速度的思考。

Again, rate of improvement thinking.

所以，如果我假设这个趋势继续下去，总有一天幻觉会变得越来越罕见，比现在更少。

So if I just imagine the trajectory to continue, there is gonna become a time where hallucinations become rarer than they are today increasingly.

另外，我想说的是，当我想到其他技术时，还有两点。

And also, what I'd say is when I think about other technologies, there's two more points.

其他技术在刚问世的时候，比如互联网，也遇到过技术上的困难。

Other technologies at their inception when they first came to the world, like the Internet, also had technical difficulties.

我记得小时候用拨号上网，那时候上网就不能同时打电话。

I remember growing up with dial up modems, and I couldn't go on the phone at the same time as going on the Internet.

我得先关掉楼上的《RuneScape》才能去接电话。

I'd have to stop RuneScape upstairs to go on the phone.

你会觉得这玩意儿真烂。

And you thought that this is crap.

这技术真烂，烂透了。

This is crap technology crap.

我不知道，兄弟。

All the I don't know, mate.

我挺喜欢的。

I loved it.

对。

Yeah.

明白。

Know.

那时候感觉就像魔法一样。

You it felt like magic.

现在回头看，你会想，哇。

And then in hindsight, you go, wow.

我现在有星链和手机上的5G网络。

I now have Starlink and five g Internet from my phone.

简直难以置信。

It's unbelievable.

以前你出门根本没法上网。

You couldn't leave the house with Internet before.

这就是我说的进步速度的思维方式。

And that's what I mean by the rate of improvement thinking.

我想说的最后一点是，我们经常把AI和完美作比较，对吧。

I'd say the last point is we often compare AI to perfection Right.

但实际上，那并不是真正的替代方案。

Whereas that's not actually the alternative.

在工作中，比如我想做一个简单的写作任务，我应该把AI和我的其他选择相比，那个选择也要考虑我的时间和我会犯错的能力，毕竟我不是什么都懂。

In the working world, like if I wanted to do, let's say, a simple writing task, I should compare AI to my alternative alternative way of doing that simple writing task, which is both measured in my time, right, and my ability to hallucinate as a person who doesn't know everything at all.

如果我雇一个人，一个实习生，他也可能犯错或者知识有漏洞。

If I'm hiring someone, an intern who might also be prone to hallucination or have gaps in their knowledge.

所以实际上我们不应该把AI和完美相比，而是和其他的选择相比。

So it's not actually like we're comparing we should compare AI to perfection, it's AI to the other alternatives.

如果一个人犯错率只有0.7%，但知道得更多，速度也更快，那总体上可能是个划算的交易。

And if someone hallucinates 0.7% of the time but knows way more and is faster, Maybe on a net basis, that's a good trade.

也许我应该用AI。

Maybe I should use AI.

那我们开始吧。

So let's start

举个例子，一个我非常爱的人。

with an example, someone I love dearly.

马特·休斯，我的编辑。

Matt Hughes, my editor.

对。

Yeah.

利物浦的失败方。

Loser side of Liverpool.

很棒的人。

Wonderful guy.

我付钱给马特·休斯，不是因为他什么都懂。

I don't pay Matt Hughes because he knows everything.

我付钱给他，是因为他有惊人的背景知识和大量的知识储备，而且他愿意扩展这些知识，和我一起工作，给我精神支持。

I pay him because he has incredible context and a ton of knowledge, and he's willing to expand it and work with me and moral support.

他是个很棒的编辑，但他也是个能深入核心、有亲身经历的人。

And he's a great editor, but he's also someone who gets into the guts of it and has the experiences of it.

他是一位资深的科技记者。

He's a decorated tech journalist.

除此之外，他还是个充满爱心的人，对他热爱的事物充满同理心和喜悦，对他讨厌的人则充满绝对的敌意。

And on top of that, a wonderful loving being with empathy and joy in his heart for the stuff he loves and absolute fucking venom for the people he hates.

这是我从大型语言模型那里得不到的。

That's the I can't get that from a large language model.

但除此之外，我对那个假设本身也有异议。

But on top of that, I don't I push back on just the assumption there.

你说它什么都知道，可当它有时候什么都不知道的时候，什么都知道又有什么用呢？

When you say knows everything, what good is something that knows everything when it sometimes doesn't know anything?

而且问题是，你真的会为一些基本的事情去付钱给一个实习生吗？

When it sometimes and on the thing is, are you really paying an intern for something basic?

你真的会去找他们，然后说，对，

Are you really going to them and saying, yeah.

你能查一下今天几号吗？

Can you look up what the date is?

不。

No.

你会在谷歌上查。

You're doing that on Google.

不管是什么任务，你其实也是在培养一个实习生。

Whatever the task is, you are trying to also train an intern.

实习生的意义就在于培养他们，让他们不再是那个什么都不懂的状态。

The point of an intern is to train them and turn them in take them out of Pinocchio status.

但实习生也会学习。

But it's also an intern learns.

实习生会了解背景。

An intern gets context.

实习生会熟悉你的习惯。

An intern learns your habits.

学习，了解背景，然后学习。

Learn gets context and learns.

不对。

No.

它不会。

It doesn't.

它确实不会学习。

It's doesn't learn.

我的意思是，它确实不会。

I mean, it doesn't.

它的学习方式是，你创建一个巨大的Claude。

The way it learns is you create a giant Claude.

一个Markdown文件，它有时候不读，有时候会读。

Md file that it sometimes doesn't read, sometimes does read.

你创建一个框架。

You create a harness.

就像皮威的早餐机器一样，来自《皮威的季后赛》。

You put it's like it's Pee Wee's breakfast machine from Pee Wee's playoffs.

你得做所有这些巧妙的设计来减少幻觉。

You have to do all these contrivances to mitigate the hallucinations.

而且即使到最后，你投入了多少精力呢？

And even then at the end, how much effort have you put in?

不过，好吧。

But so okay.

这是一个非常简化的例子。

This is an extreme simplified example.

如果我现在打开我的Claude，问它我的狗叫什么名字。

If I went on my clawed now and said, what's my dog my dog's name?

嗯哼。

Uh-huh.

它会知道我的狗的名字。

It would know my dog's name.

天哪。

Jesus Christ.

这家公司筹集了950亿美元。

This this company raised $95,000,000,000

今年。

this year.

我是在用一个极其简化的例子来说明它能记住过去的事情。

I'm I'm using an extreme simplified example to show that it can remember things from the past.

当然，它也知道更复杂的东西，我只是拿这个当例子。

Obviously, it knows much more complex things as well, I just use that as an example.

所以我们承认它确实有对过去的记忆。

So we we we accept the fact that it can it does have memory of the past.

它能访问一些存有内容的文件。

It has files it can access that have stuff on it.

对。

Yeah.

但那是……对，没错。

Which But is that's Yeah.

这和记忆不是一回事。

Not the same as memory.

而且那也只是……好吧。

And it's also just okay.

所以它能记住你狗的名字。

So it remembers your dog's name.

它可能记得你的习惯。

It might remember your habits.

它可能能读取你之前说过的话。

It might be able to read things you've said before.

它知道你的情绪吗？

Does it know your moods?

它知道周围世界发生了什么吗？

Does it know what's going on in the world around it?

它有好日子和坏日子吗？

Does it have good days and bad days?

它会在你身边吗？

Is it there for you?

因为它就是个破文本机器。

Because it's just a fucking text machine.

而且关键是，那个实习生的例子。

And the thing is, the intern example.

实习生是可以成长的。

An intern is something that can grow.

是你愿意投入培养的。

It's something that you invest in.

这不是靠喂文件和文字就能做到的。

That's not something you do through feeding files and text to it.

我们储存记忆的方式，我们积累经验的方式，是情感和感受的里程碑，还有……

The way that we store memories ourselves, the way in which we accrue experiences is a a milestone of emotion and feelings and

完全不一样。

Completely different.

所以我觉得这里有两件事。

So I think there's two things here.

一个是事情发生的过程，另一个是结果。

There's the process in which something happens, and then there's the output.

嗯哼。

Uh-huh.

所以这个过程，你刚才描述的是人类记忆的过程。

So the process, you were describing the process of how a human does memory.

对。

Right.

AI记忆的方式不同，但人们真正关心的是输出结果的价值。

The way that an AI does memory is different, But the thing that people care about is their value in the output.

我知道，如果我把所有文件都倒进Claude里，我不太在乎它怎么处理，只要我问它我的收入是多少时，它能给出数字。

I, know, if I dump all of my files into Claude, I don't really care how it processes it as long as when I ask it, what's my revenue?

它知道那个数字。

It has the number.

培训一个人也是同样的道理。

And one could say the same thing about training someone.

你可以说，你教他们，投入很多精力，给他们很多背景信息，教育他们，给他们经验，然后你可能会问他们，顺便问一下，我的收入是多少？

You could say, you teach them, you put lots of effort into them, you give them lots of context, you you educate them and give them experiences, and then you might come and say to them, by the way, what's my revenue?

虽然过程完全不同，但我在乎的是结果。

Now the processes are entirely different, but the outcome is what I care about.

我问他们的时候，他们知道那个收入数字吗？

Do they know the revenue number when I ask them?

所以我觉得这就是我们有时候会迷失的地方，你知道，因为我听过这种争论，比如AI能不能有创造力？

And so I think that's the part that we sometimes get lost and we get you know, because I have I've heard this debate about, like, can AI be creative?

对。

Right.

我觉得，回答这个问题的关键在于结果。

I think, like, the way to fix answer that question is, like, it's about the output.

当我让它做一件有创意的事情时，它能不能给我答案？

When I ask it to do a creative thing, does it give me the answer?

而不是过程是否和人类的过程一样。

Not is the process the same as a human process.

因为实际上，谁在乎呢？人们在乎的是他们为结果和产品买单。

Because actually, no who cares what the people care about they pay for the outcome, the product.

其实我不同意关于过程的看法，因为比如马特·休斯，你的编辑？

I actually disagree about the process because Matt Hughes, for example Your editor?

对。

Yeah.

对。

Yeah.

看着他钻进一个兔子洞，和他一起在那里，反过来他也一样。

Watching him go down a rabbit hole and being there with him, and actually vice versa, him doing the same thing.

我们写了这些，嗯，我是说，我们在做研究。

We wrote these well, I mean, we were working on the research.

我最后坐在那里，像是一整天的写作，写了11000字，他给了我一大堆笔记。

I ended up sitting there for, like, the day long session of writing 11,000 words, and he he had given me a bunch of notes.

光是描述那个过程，我就觉得特别开心，因为就像我们在说，我真不敢相信这些天杀的。

It was actually just even describing that process, I feel so happy because it was like us being like, I can't believe how fucked these Jesus Christ.

他们怎么能这样，就像那些资产管理者，比如黑石，那些可怕又愤世嫉俗的人，了解他们之后就觉得，不可能是这样，然后和他来回讨论。

They can't like, just like the misanthropy of just the horrible cynical people of asset managers like Blackstone just learning about them and being like, it can't be this and having a back and forth with him.

对。

Mhmm.

这根本不一样，因为我们是共同学习的，而且学习的过程本身和产出一样重要。

That is fundamentally different because we were both learning together, and the learning process was as much about creating the output as the output itself.

当你学习的时候，你并不是在生成一个平均结果——那正是这些工具做的事情——从它能找到的文档里取平均值。

When you learn something, you're not creating the average, which really is what these things do, of the documents it could find.

你得到的并不是特别新颖的输出。

You're not getting particularly novel outputs.

如果我需要一份普通的AI生成内容，那当然可以。

If I needed a generic SLOP output, sure.

但我用过一些对冲基金用的高端大语言模型工具，它们生成的东西都一样烂。

But I've I've used some of the higher end LLM harness machines that the hedge funds use, and they all get the same shite.

都是那些千篇一律的报告，同样的“我们注意到了这个分析”。

It's all the the same generic reports, The same, oh, we noticed this analysis.

这些在任何AI生成内容里都能找到。

Things that you can find on any kind of AI slop out there.

你刚才跟我说的，我听到的是，你和Matt在一起的时间有两个价值点。

What you described to me there, what I heard anyway, is there's two points of value you're getting from your time with Matt.

我是说，我是说，还有很多，但对的。

I mean, I mean, there's many more, but Yeah.

你说你在学习，然后你的书也在被编辑。

You said you're you're learning, and then you're getting this book edited.

博客。

Blog.

博客。

Blog.

你在编辑博客，这是输出，你在学习，而且你也在建立联系，所有这些，没错。

You're getting a blog edited, which is the output, and you're getting learning, and you're also really getting connection and all these Exactly.

其他，但当人们考虑AI的价值时，当然，他们可以用它来学习。

Other But when I come to when people sort of think about the value of AI, of course, they could use it to learn.

但在我举的例子中，比如把我的收入数字重复给我，或者做这个数字。

But in the example I gave of, like, repeat my revenue number back to me or do this number.

我只关心结果。

I am I just care about the output.

我可以用它来学习。

I could use it to learn.

我可以说，就目前而言……

I could say as far

那如果收入数字曾经出过错呢？

with What if the revenue number was wrong once?

你应该用确定性的方法来验证这些数字。

You should have defined deterministic ways of knowing those numbers.

对。

Mhmm.

你不应该依赖它们。

You should not rely on them.

即使终端在运行BQL，我也信任它，我还是会反复检查，确保万无一失。

Even with the terminal running BQL, which I trust, I will double triple triple check everything just to be sure.

部分原因也是因为对我来说，学习的过程很重要，我不想要一份报告就那样放着。

Partly because also the process of learning for me, I don't want just a report I go like that.

我想要的是我完全理解的东西，而且还要理解它周围的背景。

I want something that I fully understand and also understand the context around it.

我觉得大语言模型做不到这一点，而且我也不认为它们会在这方面变得更好，因为那根本不是它们的功能。

I don't think that LLMs do that, and I just don't see them getting better in a way that does that because it's it's just not what they do.

而且还有另一个问题，报告越详细，出错的可能性就越大。

And also there's the other problem of the more detailed the report, the more likely there are things to be wrong with it.

比如说，如果你跟马特·休斯合作，我可以相信他做的是对的。

If you are with Matt Hughes, for example, I can trust he's got it right.

我可以相信他理解了，而且我可以相信我能和他来回讨论，这样如果我有遗漏，他会告诉我。

I can trust he understood, and I can trust that I can have a back and forth with him that will inform me if I've missed something.

我可以读他读过的资料，而且真的信任他，因为信任也是很重要的一部分。

I can read the stuff that he's read and actually trust him because there's a big trust part as well.

你对马特的信任是基于什么呢？

What is the basis of your trust in Matt?

会不会是因为他过去的成绩？

Could it be his historical performance?

我是说，是的。

I mean, yes.

好的。

Okay.

而且还有一点，我们是一起学这些东西的。

And also the fact we've learned half of this stuff together.

但是，资历深不一定就代表可信吧，可能有些人你认识十五年了，你也不信任他们。

But but tenure tenure doesn't necessarily there's probably people, you know, for fifteen years who you also don't trust.

对。

Yes.

所以我觉得我在想的是，到底是什么让人信任另一个人或事物？

So I think I was trying to figure out, like, what is the what is the thing that's causing humans to trust another thing?

我想大概就是持续兑现承诺吧。

And I guess it would be continual delivery of a commitment made of sorts.

就拿Claude来说，在简单任务上，就像我们从那个幻觉排行榜上看到的，它一直都能稳定地给出正确结果。

And so with Claude, for example, on simple tasks, as we've seen from this hallucination leaderboard, it continually delivers for people.

这就是为什么我们看到

And that's why we've seen the

最快的增长，我是说，那个排行榜是这么说的吗？

fastest I mean, is that what that board says?

嗯，它是在说，它是不是答错了？

Well, it's it's saying, like, is it getting it wrong?

它是不是在胡编乱造？

Is it hallucinating?

在简单任务上。

On simple tasks.

简单任务。

Simple tasks.

那些任务是怎么定义的？

How are those defined?

我不知道，我不知道。

I don't I don't know.

但问题就在这儿。

But that's the thing, though.

因为这其实很能说明AI行业的现状。

Because this is actually a very illustrative thing of the AI industry.

他们特别擅长转移话题。

They are the what aboutist masters.

他们就会说，你看。

They have like, well, look.

我们有这个基准测试，说明我们在这方面表现不错。

We got this we got this benchmark that says we're good at this.

再看看。

And look.

数字更高了。

The number's higher.

这个数字是什么意思？

What's the number mean?

那又意味着什么？

What does that mean?

我不是在批评你。

And I'm not using this as a critic against you.

就是当你给不出直接答案的时候，你就给个侧面回答。

It's when you can't give a direct answer, you give a side answer.

当你们大语言模型行业想证明自己的价值时，你不能只是说，就用这个产品吧。

When you as the LLM industry want to prove your worth, you can't just be like, just use the product.

第一部iPhone出来的时候，当然，那时候宾州州立大学。

When the first iPhone came out, of course, Penn State at the time.

哦，我感觉自己就像《2001太空漫游》开头那些猿猴一样。

Oh, I felt like the, apes at the beginning of 2001.

就像，靠。

Like, fuck.

官方语音信箱。

Official voice mail.

效果立竿见影。

It was immediate.

我拿给搞技术的朋友看。

And I showed it to tech friends.

也拿给最普通的人看。

I showed it to the most normal people in the world.

每个人都惊呆了。

And everyone's like, holy shit.

他们当时用的还是摩托罗拉刀锋手机。

This is they were on Razors.

用的还是诺基亚3210。

They were on Nokia 32 tens.

价值一目了然。

It was obvious the value.

亚马逊网络服务，也是一样的道理。

Amazon Web Services, same deal.

不过，当时并不明显。

It wasn't obvious, though.

对。

Yes.

确实如此。

It was.

我的意思是，我买了。

I mean, I bought it.

对

To

对你来说，确实如此。

you, it was.

确实。

It was.

我还把它给很多人看了，因为我知道自己一喜欢上电子产品就容易有偏见。

And I also showed it to a bunch of people because I'm aware that I have bias when I just love gadgets.

但我记得微软前CEO史蒂夫·鲍尔默那段著名的采访，他听说iPhone后当场大笑起来。

But but I remember the famous Steve Balmer, who was the CEO of Microsoft interview, where he was told about the iPhone, and he bursts out laughing.

500美元，还要签合约才能享受补贴？

$500 fully subsidized with a plan?

我说，这是世界上最贵的手机，而且对商务客户没什么吸引力，因为它没有键盘，所以作为邮件设备不太好用。

I said, that is the most expensive phone in the world, and it doesn't appeal to business customers because it doesn't have a keyboard, which makes it not a very good email machine.

你现在花99美元就能买到摩托罗拉Q手机。

You can get a Motorola Q phone now for $99.

它功能很强大。

It's a very capable machine.

能听音乐。

It'll do music.

能上网。

It'll do Internet.

它能发邮件。

It'll do email.

它能发即时消息。

It'll do instant messaging.

所以我看着那款手机，心想，嗯，我喜欢我们的策略。

So I I kinda look at that, and I say, well, I like our strategy.

我非常喜欢。

I like it a lot.

你最好开始嘲笑它，因为它太颠覆了。

You better start laughing mocking it Because it was so disruptive.

它贵得多。

It was way more expensive.

对。

Mhmm.

而且它完全不同。

And it was a way different.

没有键盘。

No keyboard.

以前手机贵得离谱，运营商虽然会补贴，但你得签长期合同。

Well, phones used to be insanely expensive, and the carriers would cover them, but you had to sign a long contract.

你照样得花五百美元。

You were still spending $500.

但问题是

But the thing

我想说的是，你不需要跟人解释成本什么的，你只要说，看这多好。

I'm getting at is you didn't have to explain to someone what perhaps you'd to get parts to the cost part, but you could just be like, look how good this is.

然后等到iPhone 3G加上App Store出现后，大家都震惊了。

And then once the app was the iPhone three g with the App Store, people were like, oh, shit.

这真的能改变世界。

This could actually change things.

移动网页一开始虽然很糟糕，但即便如此，它还是那么差劲。

Mobile web, even though it was a monstrosity, was so bad at first.

即使在那时候，你也能收邮件，而且可以直接查看。

Even then, you could get your emails and you could just look at them.

但黑莓手机也很贵，而且当时其实还挺酷的，但它们的操作方式不像消费级软件。

But as BlackBerrys were also expensive and were still actually kind of cool, but the way they worked was not like consumer software.

它们没有经典的图形界面。

They didn't have the classic GUI.

iPhone给人的感觉就不一样。

IPhones felt like that.

它感觉就像一台手机，甚至像一台电脑。

It felt like an a cell phone designed even like a computer.

这一点很明显。

It was obvious.

从一开始就很明显。

It was obvious in the beginning.

当时我在宾夕法尼亚州中部和一个女孩约会，我给她看的每个人都惊叹不已。

Everyone I was I was dating a girl in the center of Pennsylvania at the time, and everyone I showed it to was like, wow.

这太不可思议了。

This is incredible.

对我来说，这是显而易见的事情。

That to me is the obvious thing.

说到人工智能，直到今天，当你说“好吧”的时候。

With AI, to this day, when you're like, okay.

为什么它这么神奇？

Why is it so amazing?

人们仍然犹豫不决。

People still dither.

人们还是那种反应，对。

People are still like, yeah.

你不能完全用它来经营业务，除非有那种复杂的滑轮和杠杆系统之类的。

You can't run a business fully with it without this weird system of pulleys and levers and such.

但是，那为什么当你看到ChatGPT的增长数据时，又会怎样呢？

But what how come then, when you look at the stats around ChatGBT's growth?

嗯。

Mhmm.

上线仅仅六十天，就达到了一亿活跃用户。

A 100,000,000 active users in just the first sixty days after launching.

相比之下，TikTok花了九个月。

For comparison, TikTok took nine months.

Instagram花了两年半，而互联网本身，也就是万维网，大约花了七年才达到这个规模。

Instagram took two point five years, and the Internet itself, the worldwide web, took roughly seven years to reach that scale.

嗯。

Mhmm.

超过60%的美国成年人在AI工具推出后的三年内，将其融入了日常生活的常规流程，覆盖了40%的人口。

Over 60% of The US adults are integrated into AI tools in their daily and regular routines within three years of the launch, reaching at 40 of the population.

而同样的里程碑，互联网花了五年，个人电脑花了将近十二年。

And and that same milestone took the Internet five years and personal computers nearly twelve.

好的。

Okay.

所以，我觉得，这就是让我感到矛盾的地方。

So like, this is the I think this is the part that's giving me dissonance.

就像，当我给我未婚妻看ChatGPT的时候，它其实不太管用。

It's like, when I showed my fiancee chat GBT, it was didn't really work.

但她是个独立创业者，英语不是她的母语，嗯。

But as a sole entrepreneur who English isn't her first language Mhmm.

她需要写很多文字、很多文案，还要生成很多图片，而且之前还得花钱请平面设计师帮她做那些她自己做不了、因为没那个技能的图片。

Who has to write lots of text, lots of copy, and generate lots of images, and was paying a graphic designer to help her make, certain images that she, you know, couldn't make herself because she doesn't have the skills.

她会说这对她的生意来说是革命性的。

She would describe it as being transformative for her business.

嗯。

Mhmm.

我从你这里听到的是，它并不具有革命性，对人们来说没有价值。

What I'm hearing from you is that it's not transformative, and there's no value in it for people.

但如果她认为它具有革命性的话。

But she if she was acting as transformative.

她会按每百万token的费率付费吗？

It would she pay the per million token rate?

她会按实际费率付费吗？

Would she pay the actual rate?

因为这就是关键所在。

Because that's the thing.

如果这东西按真实成本出售，是的。

If this was sold at its honest cost Yeah.

如果人们真的那样反应，而且每次操作都付2.03美元或4美元，并且他们真心满意，那或许还能说得通。

I would actually if and people were reacting like that and they were paying $2.03, $4 every time they did something and they were genuinely happy, that might be an argument.

如果他们没有潜意识里的那些因素，那真实成本会是多少？

What is the what would be the honest cost if they weren't subconsciously?

每百万token的成本，也就是他们应该收取的实际API费用。

Per million token cost, the actual API cost they should charge.

你知道那相对是多少吗？

Do know how much that is relative

哦，天哪。

to Oh god.

这要看具体是哪个模型。

It depends it depends on the model.

不过关于你之前提到的互联网那件事，我其实有个观点想说说。

But there's actually kind of a point I wanna make about the thing you said with the Internet earlier.

我刚开始上网的时候，用的是33.4千比特每秒的调制解调器。

So when I first got on the Internet, I had 33.4 kilobits a second modem.

即使那时候，我就觉得，靠。

Even back then, I was like, fuck.

要是能再快点就好了。

If this was faster.

而且那种想法是立刻就有的。

And that was, like, immediate.

就是觉得，要是能再快点就好了，因为实在太慢了。

It's like, if this was faster because it was slow.

你可能会去像Happy Puppy那样的网站。

You'd go on, like, happy puppy or something.

下载共享软件要花一整天的时间，等得人都要疯了。

Download, take all bloody day waiting for shareware to download.

马上你就会想，要是能更快就好了。

Immediately, you're like, if I could do this faster, it be better.

而且那时候我就觉得，天哪，这玩意儿说不定还能用来搞视频拍摄。

And even back then, I'm like, man, you could probably do video camera stuff with this.

这些后来都实现了。

Stuff that eventually happened.

其实有个叫吉姆·卡维罗的人，是高盛的，他在2024年的一份报告里说，生成式AI投入太多，回报不够。

And, actually, there's this guy called Jim Cavello from Goldman Sachs in a report he did in 2024 that was Gen AI, too much spend for not enough return.

我这是大概转述一下。

Paraphrasing there.

他还指出，在iPhone问世之前，有成千上万的演示都表明，当GSM无线电、蓝牙无线电、WiFi无线电变得越来越小，我们必然会得到像iPhone这样的东西。

And he made the point that in the run up to the iPhone, there was thousands of presentations that when GSM radios get smaller, when Bluetooth radios get smaller, when WiFi radios get smaller, it is inevitable that we will get something like this.

然后他说，AI没有这样的路径。

And then he said that there is no such path for AI.

AI没有路线图，能变成他们承诺的那个东西。

There is no road map to AI becoming this thing that they promised.

我得说清楚。

And I must be clear.

如果这些公司当初站出来说，对，这是有趣的云软件。

If these companies had gone out there and are like, yeah, this is interesting cloud software.

它是生成式的。

It's generative.

它真的很贵。

It's really expensive.

我们不确定能不能完全信任它，不是那种害怕的信任。

We're not sure if we can fully not trust it, not in the oh, I'm scared way.

我的意思是，我们只是不确定这会不会是颠覆世界的东西。

I mean, just like we're not sure that this is going to be a disruptive world changing thing.

它有潜力，但我们会慢慢来。

It has potential, but we're gonna go slow.

它真的很贵。

It's really expensive.

这是一项研发工作。

This is an r and d effort.

我们不会把它暴露给消费者，也不会称它为语言模型或生成式AI之类的东西。

We're not gonna expose consumers to it and actually be like, call them, like, I don't know, language models and no no generative AI stuff.

甚至不叫它，因为它不是自主的。

Just being not even call it because it isn't It's not autonomous.

它不聪明。

It's not smart.

我其实可能会尊重它，但这不是他们从2023年开始就说2022年的事。

I actually might respect it, but this is not they've gone out there since 2023 and said it was 2022.

这是自切片面包以来最好的东西。

This is the best thing since sliced bread.

这正在改变一切。

This is changing everything.

这会帮你做完所有工作。

This is gonna do all your work.

这会抢走你的工作。

This is gonna take your job.

你会跟必应聊天，然后它会让你离开你老婆。

You're gonna talk to Bing, and it's gonna take to leave your wife.

所有这些疯狂的事情。

All of these crazy things.

有趣的是，当那位作者，我记得是凯文·鲁斯，跟微软的CTO凯文·斯科特聊起这件事时。

And what's funny is when the writer, Kevin Roos, I think it was, he was speaking to Kevin Scott, the CTO of Microsoft about it.

凯文·斯科特说，你知道吗，我很高兴我们能进行这次对话。

And Kevin Scott goes, you know, I'm just glad we're having this conversation.

而不是像那样说，冷静点，比维斯。

Instead of being like, settle down, Beavis.

它就是个网站。

It's a website.

网站告诉了你一些东西。

The website told you something.

其实就是大语言模型。

It's just LLMs.

他们把它吹得天花乱坠。

They talked it up.

那是因为大家都在谈论他们希望它成为什么，而不是它实际能做什么。

And that's because everyone is talking about what they wish this was rather than talking about what it can actually do.

这让它让人觉得可怕，而且是故意这么做的。

This makes it scary to people, deliberately so.

这让它变得对环境有破坏性。

It makes it environmentally destructive.

看看那些燃气轮机在毒害黑人社区。

Look at the gas turbines poisoning black neighborhoods.

我觉得那是在路易斯安那州。

I think it's in Louisiana.

那是马斯克的一个数据中心。

It's one of Musk's data centers.

看看那惊人的电力消耗。

Look at the incredible energy draws.

它正在推高电费。

It is raising power bills.

而且，由于大规模的建设，它也在推高所有消费电子产品的价格，造成通货膨胀。

And also, it is creating inflation across all consumer electronics because of the massive ramp.

你知道有趣的是什么吗？

You know what's interesting?

我几乎觉得你说的很多都是对的。

I almost feel like so much of what you're saying is true.

同时，这项技术也确实有可能深刻改变世界。

And also, it can be true that this technology is gonna profoundly change the world.

而且我觉得，你知道，回想互联网早期，也许最接近的类比就是互联网泡沫时期。

And I think like, you know, I think back to the early days of the Internet is maybe the closest analogy we have of, you know, in the .com bubble.

你知道，你写了那篇很棒的文章。

You know, you wrote this great essay.

对。

Yeah.

是的。

Yes.

是的。

Yes.

我觉得那篇文章特别有意思，尤其是那个名字。

Which I found really funny, especially the name.

腐烂经济，你提到了腐烂泡沫。

The Rot economy, and you talked about the Rotcom Bubble.

是的。

Yes.

谈到AI的价值其实没有人们想的那么高。

Talking about how AI is less value than people think.

在互联网泡沫时期，你看到的是巨大的炒作，人们过度吹嘘他们网站和所建东西的能力。

And in that in the sort of .com bubble, what you saw is huge hype, people overselling the capabilities of their websites and what they were building.

但在互联网泡沫破灭后，确实，90%的东西都归零了。

But in the wake of the .com bubble, yes, 90% of stuff went to zero.

对。

Mhmm.

但也有一些划时代的公司诞生，改变了世界。

But you had generational companies born that changed the world.

没错。

Right.

所以我觉得，这就是泡沫的作用。

And so I I do I kind of and that's what bubbles do.

对吧？

Right?

巨大的高潮

Huge high

那是彻底的

It was complete.

过度投资。

Over investment.

投资者变得疯狂妄想。

Investors get crazy delusional.

他们认为一切都会改变。

They think it's everything's gonna change.

与此同时，在这些时刻确实有怀疑者。

At the same time, you do have skeptics in these moments.

互联网泡沫时期，互联网本身就有最大的怀疑者。

The the .com bubble had I mean, the Internet itself had the biggest skeptics.

1998年，诺贝尔经济学奖得主保罗·克鲁格曼说，到2005年左右，人们会清楚地看到，互联网对经济的影响不会比传真机大。

In 1998, Nobel Prize winning economist Paul Krugman said, by 2005 or so, it will become clear that the Internet's impact on the economy has been no greater than the fax machine.

1995年，天体物理学家克利福德·斯托尔，我在书里写过这个，他在《新闻周刊》上写道：我们的电脑专家是不是都缺乏常识？

In 1995, astrophysicist Clifford Stall, famously, I wrote about this in my book, wrote famously in Newsweek, Do our computer pundits lack all common sense?

事实是，没有任何在线数据库能取代你每天看的报纸。

The truth is no online database will replace your daily newspaper.

没有哪张光盘能替代一位称职的老师。

No CD ROM can take the place of a competent teacher.

商业和生意会从办公室和商场转移到网络和调制解调器上。

Commerce and businesses will shift from offices and malls to networks and modems.

胡说八道。

Baloney.

那为什么我当地的商场生意兴隆，而网络商城却无人问津？

So how come my local mall does a roaring business and the cyber mall gets zero business?

然后我再给你举一个克鲁格曼的例子，他是获奖经济学家。

And then I'll give you one more from Krugerman, who was the, award winning economist.

他说互联网的增长将大幅放缓。

He said the growth of the Internet will slow drastically.