AI 能写代码，为什么软件没有变得更好？

因为我记得我们聊过，大概是2017年吧？

Because I remember talking to you maybe in, was it 2017?

那已经是很久以前的事了。

It was so long.

对，我们确实聊过那个。

We did talk about that, yeah.

是的，而且当时你提到了很多想法，你说数据很重要，你还说想专注于任务本身，所以我就想问，你当时就知道这会发展成一个分类器，还是只是一种直觉，觉得看待整个AI运动有另一种方式？

Yeah, and then a lot of these ideas were in, you were talking about data being important, and you were talking about you want to focus on the task, and so I just, did you know that this was going to end up being a classifier, or was this just an intuition that there was just kind of another way to view this entire kind of AI movement?

你知道吗？

You know?

其实，关于2017年那次演讲中的对话，有个有趣的故事。

So actually, a fun story about that chat in the talk from 2017.

我觉得我那次演讲的主题其实非常相似。

I think my talk was actually on a very similar theme.

我记得标题好像是“AI：模块化理论与——”

I think it was called something like "AI: Modular in Theory and-

……在实践中灵活。

... Flexible in Practice."

这很符合软件系统的视角。

Which is very software systems view.

对，完全同意。

Yeah, totally.

所以我在这一点上还算一致。

So I'm a little bit consistent in that.

我觉得这其实是在ChatGPT出现之前就开始了。

I think that this really started right before ChatGPT.

就在我们发布这些东西的时候，我并没有这种直觉，说实话，RLHF的泛化能力让我非常惊喜。

Right when we released these things, I did not have intuition about this, and honestly, I was very pleasantly surprised by the generalization capabilities of RLHF.

这是什么时候的事？

When is this?

应该是2021年底，大概……

Must be end of 2021, like a-

好的，没问题。

Yeah, okay

……2021年第四季度。

... fourth quarter of 2021.

它真的非常通用。

It was really, really general.

如果你读那篇论文，它不像其他论文那样试图证明自己的观点。

If you read the paper, it's unlike other papers that are trying to prove their point.

我们实际上是用科学方法，试图反驳——

It is us actually, scientific method-ish, trying to disprove-

……这是作弊吗？

... is it cheating?

而且，我最喜欢的问题是：为什么冥想前吃袜子很重要？

And- My favorite query was why is it important to eat socks before meditating?

我们事先确保这个问题在网上不存在。

We'd made sure that was not on the internet beforehand.

而且这些模型居然能给出看起来挺像模像样的答案。

And the models were able to make plausible human-looking answers for this.

对我们团队来说，那一刻就明白了，这不是作弊——在机器学习里，你总是得担心作弊的问题。

And that to us in the team was the thing that clicked, like this is not cheating, which you should always be afraid of cheating in ML.

然后真正让我崩溃的是，我们把模型发布了。

And then what really got me burnt was we released it.

我显然是个特别看重能力的人。

I'm obviously a big capabilities guy.

为了发布那个模型，我做了很多工作。

I did a lot to release that model.

我当时真的觉得那个模型很有可能是AGI，结果不是，那一刻我的整个世界都崩塌了。

I really thought that the model had a decent chance of being AGI, and when it didn't, that was when my whole world came crashing down.

哦，有意思。

Oh, interesting.

然后我就想，为什么？

And I was like, why?

哦，所以你之前有一阵子也是这么想的。

Oh, so you were kind of on the other train for a bit.

你是说那种疯狂的想法吗？还是——

Which, like the crazy train or-

嗯，不是，我只是说——

Well, no, I'm just like-

没有冒犯的意思。

No offense

……强化学习能泛化，也许我们就有通用人工智能了。

... RL generalizes, maybe we have AGI.

基于人类反馈的强化学习泛化得还不错。

RLHF generalizes pretty well.

哦，对。

Oh, yeah.

基于验证奖励的强化学习才是泛化得不太好的那个——

RLVR is the thing that doesn't generalize as well-

……就我目前看到的来说。

... from what I've seen.

然后AGI呢，所以——

And AGI in, so-

其实我刚才想说的是，你之前在做ChatGPT，在做那些早期的GPT模型。

Well, I was just saying more, you were behind ChatGPT, you were behind these early GPTs.

那是一个非常不同的目标，就是做一个能跟人聊天的聊天机器人，而不是给程序员用的工具之类的。

That was a very different goal, which is creating a chatbot that will talk to the human being, was not a programmer's tool, et cetera.

所以我就是有点好奇——

So I'm just wondering like-

哦，其实在2020年初，OpenAI那时候聊AGI，大家经常形容说，AGI就是Ilya加上一堆if语句。

Oh, well, actually, early 2020, OpenAI, when we talked about AGI, people used to describe it as Ilya and every if statement.

所以有点像——

So it's kind of like a-

但这是不是……我们刚才在聊OpenAI的文化。

But is it part, we were talking about OpenAI culture.

部分原因是它故意定义得很模糊，这样就能容纳所有人，大家都能算在里面。

Part of it is that it's intentionally vague, so it's a wide tent so that everyone can be inside of it.

但出于一些微妙的原因，我不认为我们走在RSI的道路上，我到现在依然不这么认为，当时我也是这么想的。

But I am not, for nuanced reasons, I don't think we are on the path of RSI, and I still don't think we're on the path of RSI, and I did to then.

我确实认为OpenAI对AGI的定义是非常可行的。

I do think that what OpenAI defined as AGI is extremely doable.

自动化世界上大部分有经济价值的工作——

Automating most of the world's economically valuable work-

……听起来确实，天哪，这世上的工作可真多。

... actually sounds like a, oh, man, there's a lot of work out there.

其中很多都是非常机械和简单的。

A lot of it is very rote and simple.

而且从数量上看，要外包工作，你需要简单的指令，让普通人都能完成。

And by volume, in order to be able to outsource work, you need simple instructions that basic people can do.

据我所知，这种智能水平在模型里已经存在相当一段时间了。

And as far as I can tell, the intelligence of that has been available in the models for quite a while now.

而且，天哪，我心里一直有个疙瘩，就是为什么这个还没实现？

And my, oh, man, chip on my shoulder is why is this not available?

然后自从RLHF之后，AI行业就分裂成了两个极端，一边是过度承诺，一边是交付不足。

And then since RLHF, the AI industry just kind of bifurcated into gigantic overpromise, underdeliver.

我觉得GPT-3那时候其实还挺靠谱的。

I think GPT-3 was actually quite calibrated back in that day.

但因为人类来评估模型的好坏，看起来效果很好，因为他们是裁判，但我们一直在优化那个裁判，而不是优化自动化本身，这就是缺失的部分。

But because humans evaluate how good the models are, it looks really good because they're the judge, but we've been optimizing that judge instead of the automation part, and that has been the missing thing.

所以我想说，就是那时候我才恍然大悟，为什么这个东西不能更有用呢？

So I would say that it was really then that it hit me, why is this thing not more useful?

所以你觉得我们应该用能自动化多少实际生产任务来衡量？

And so you think that the measure that we should have is to what extent can you automate actual productive tasks?

这就是——

Is that the-

对，没错，我就想要这样。

That's, no, I would like that.

我也希望如此。

I would like that

你说过度承诺和交付不足，你特别指的是自动化任务的能力这个方面吧。

When you say overpromise and underdeliver, that's the dimension in particular you're talking to, the ability to automate tasks.

我心里觉得，这很酷炫科幻。

I think in my heart it's cool sci-fi.

而且我觉得——

And I think that-

完全同意。

Totally

……

...

我觉得这就是酷炫科幻的预警信号。

I think that that is the canary in the coal mine for cool sci-fi.

你真的要告诉我数学问题已经解决了，或者甚至两年前，GPQA，谷歌的证明题问答已经解决了，但我们还是搞不定得来速点餐吗？

Are you really telling me that math is solved or even two years ago, GPQA, that Google proof question-answering is solved, but we still can't handle a drive-through?

这确实是一件很难同时想清楚的事情。

It's a very hard thing to hold in your head at once.

我觉得很多人对这个问题都没有好的答案。

And I think a lot of people don't have good answers to that.

我能先测试一件事吗——

Can I just test one thing which-

当然可以。

Hell yeah

——可能不太合理，但我想——

... which may not make sense, but I want to-

来吧。

Hit me.

不过，难道不是有个说法吗，现实世界的分布和数字世界不一样，对吧？

Isn't there an argument, though, that the distribution of the real world is different than the digital world, right?

它是重尾分布。

It's heavy tail.

有很多例外情况。

There's a lot of exceptions.

我们没有全部的数据。

We don't have all the data.

有没有可能我们在现实世界里没做有生产力的事情，就是因为缺少那个分布的数据，没有在那个分布上训练，所以基本上就被限制在这些低维流形里，比如数学或代码之类的？

And it couldn't it be the case that the reason we're not doing productive stuff in the real world is just like we don't have the data for that distribution, we're not training on that distribution, and this is why it's just been basically relegated to these lower dimensional manifolds, like whatever math or code or...

我个人不完全认同数据这个论点。

I don't entirely buy the data argument, in my opinion.

我确实相信存在长尾，否认这一点就太疯狂了。

I do believe that there's a long tail, for sure, that would be kind of crazy to deny.

而且在我看来，在我的金丝雀煤矿情景里，我们不需要自动化那个长尾。

And I don't think that in my canary in the coal mine situation, we need to automate that long tail.

明白了。

I see.

好的。

Okay.

我觉得我们在所有事情上都需要非常务实。

I think that we need to be incredibly pragmatic on everything.

而且构建可靠的软件总是一项投资，对吧？

And building reliable software is always an investment, right?

程序员三大美德是什么？

What are the three great virtues of a programmer?

懒惰，为了不再做一遍，傲慢，还有第三个。

Laziness to not to do it again, hubris, and there was a third one.

不，我记得。

No, I remember.

这是来自Perl时代的话。

This is from the Pearl days.

还有第三个。

There was a third one.

拉里·沃尔，对。

Larry Wall, yeah.

我真希望我能记得。

I wish I could remember it.

但关键是，花十个小时去完成一个五分钟的任务，就是为了以后再也不用手动做这件事。

But it's about the laziness to spend 10 hours to do the five-minute task instantly and to never have to do it again.

对于喜欢自动化的人来说，这应该是一个投资回报率的决策。

It should be an ROI decision for people who like automate stuff.

我只是希望它能被自动化。

I would just like it to be automatable.

而且我觉得，人们会创造新的工作类型，这就是杰文斯所说的。

And I think that people will just make new kinds of work, hence the Jev in Jevons.

一旦那些事情变得可行，就会出现新的工作。

New kinds of work once that stuff is doable.

但作为一个基准，我觉得看看我们是否真的能自动化那些看起来AI绝对应该能自动化的东西，是很有用的。

But, as a benchmark, I feel like it's useful to see can we actually automate the stuff that it really, really looks like AI should be able to automate.

明白了。

I see.

OpenAI从2020年就开始尝试自动化客户服务了。

OpenAI has been trying to automate customer service since 2020.

但到现在还没成功。

It's not.

这确实挺让人惊讶的。

Which is pretty amazing.

确实很不可思议。

It's wild.

确实很不可思议。

It's wild.

而且，在公司内部，目前自动化的东西其实很少。

Well, and then inside companies, there's very little that's automated right now.

那些项目也没成功。

And the projects haven't worked.

除了编程方面，效果倒是出奇地好。

Other than programming has worked amazing.

你能不能把那些你觉得更容易自动化的问题类型分一下类？

Can you maybe classify the types of problems you think that are easier to automate

现在，因为那还挺有意思的。

Now, because it was kind of interesting.

所以我们其实在生成式AI浪潮之前就看过客服支持这块，那时候就挺有意思的。

So we've actually looked at support before the current generative wave, and it was interesting.

你遇到一家公司，他们会说：“我们处理了95%的客服电话。”

You'd meet a company, and the company would say, "We answer 95% of all help desk calls."

那真的是很多了。

And like, that is so many.

但当你真正去看数据的时候——

But then you actually look at the data-

其实都是同一类问题。

They're all the same

……你会发现全是密码重置。

... and you realize it's all password resets.

但如果按唯一性来算，大概只有50%左右。

But if you did it by uniqueness, it was only something like 50% or something.

所以感觉当你面对人和自然系统时，总会有那么一个很长的尾巴，里面全是例外情况。

So it just feels like when you're dealing with humans and natural systems, there's just this very-

……

...

所以，你在多大程度上预料到了……

long tail of exceptions, and so to what extent did...

大概每小时都有人给我发消息说：“我在用Jev做这个新用途。”

Probably every hour, I have somebody ping me like, "I'm using Jev for this new use case."

我就想：“我完全没想到。”

I'm like, "I had no idea."

那你当初在多大程度上预料到了它会这么广泛地被使用？

And so, to what extent did you even predict the broad range of use case for it?

你当时有预料到这种情况会发生吗？

Did you assume that was going to happen?

那你有没有对此感到惊讶，还是？

And have you been surprised by that, or?

非常惊讶。

Extremely surprised.

完全没想到会发生这种情况。

Did not assume it would happen.

这次发布并不是什么……如果有人预料到这一点，那他们可能疯了。

This launch was not something... If anyone expected this, they are probably insane.

我觉得没人能预料到会出现一个面向开发者的ChatGPT，因为ChatGPT本来是面向普通用户的。

I don't think someone could expect a ChatGPT for developers because ChatGPT was for-

对，确实。

Yeah, sure

……

...

这很奇怪。

the normal users, and it's weird.

其实我甚至不知道Jev社区里有多少人本身就是开发者。

I actually don't even know what percentage of the people who are part of the Jev party are developers themselves.

哦，有意思。

Oh, interesting.

我无法想象非开发者会用它。

I can't imagine non-developers using it.

我不知道他们会怎么用。

I don't know how they would use it.

但即使是我那些非开发者的朋友，也都在社区里，玩推特，发梗图，诸如此类。

But even my non-developer friends are just part of the party and Twitter and memeing and everything like that.

所以第一点，太棒了。

So number one, phenomenal.

第二点，这很难用短短几句话讲清楚，因为这是多年来的心血和汗水。

Number two, this will be hard to convey in this short message because it's been like blood, sweat, and tears for years now.

我对可靠性的重视程度，那是相当高。

The amount I care about reliability is, it's a lot.

可靠性就是这个东西的核心。

Reliability is what this thing is.

如果你不明白这一点，就很难做出一个达到基准的仿制品。

If you don't understand that, it'll be very hard to make a copycat that's benchmaxed.

我觉得每一层的可靠性对每个人来说都非常有价值，即使从市值角度来看它不是最有价值的。

I feel like every nine of reliability is going to be so valuable for everyone, even if it's not the most valuable thing market cap-wise.

因为它会催生新的应用，我们正在追求各种我们甚至还没完全理解的可靠性层级，因为我们才刚刚把这台AI的电动机，也就是智能的电动机——

Because it will just enable new applications, and we are fighting for all sorts of weird nines of reliability that we don't even fully understand because we are just really getting this electric motor of AI, of intelligence-

——装进人们的工作站，让他们自己去发现能拿它做什么。

... into people's workstations, and they can figure out what to do with it.

在这个语境下，可靠性是什么意思？

What does reliability mean in this context?

是指模型的可用性，还是说我调用模型时它返回相同的结果，还是说我该怎么理解可靠性？

Is this availability of the model, or is it like I call the model, and it returns the same thing, or how do I think about reliability?

对，所以——

Yeah, so-

对于本质上有点随机性的东西来说。

For something that's inherently kind of stochastic.

所以，不太是前者。

So, not so much the former thing.

好的。

Okay.

第二个更接近一些。

And the second thing is closer.

我会把第一个描述为类似正常运行时间或服务等级协议。

I would describe the first thing as kind of like uptime or SLAs.

好的。

Okay.

第二个我可能会更倾向于称为确定性。

The second thing I would maybe call closer to determinism.

第三个，我会更倾向于视为稳健性。

Something thirdly, I would consider more like robustness.

所以稳健性，我大概会描述为每次都有相似的智能表现。

So robustness, I would kind of describe as similar intelligence every time.

哦，有意思。

Oh, interesting.

所以，不完全是确定性，因为我觉得确定性对单元测试有用——

So, not exactly determinism, because I think determinism, it's useful for unit tests-

当然。

Sure

……但对真实系统没用。

... but not real systems.

当然。

Sure.

想想看，如果你加一个UUID来采用，它应该是一样的，因为功能上是一样的。

Think about if you add a UUID to adopt, it should be the same because it's the same functionally.

但它并不是完全确定性的。

But it's not exactly deterministic.

我觉得这还有另一层，我还没搞清楚它叫什么。

I think that there's another layer of it that I don't really know what it's called yet.

也许这就是我所说的某种智能——它不需要每次都表现一样，但每次都要聪明。

Maybe this is what I would call some form of intelligence, which is it doesn't have to be the similar function every time, but it needs to be smart every time.

如果你处在那种情况，这对人类来说是不是一个可以理解的想法？

If you were in that situation, would this be an understandable thing for a human to think?

因为开发者可以围绕这个来编程，而且对我来说，可靠性的最高境界就是让人们不用做示例查询就能直接对着Jev编程。

Because a developer can program around that, and actually, to me, the highest honor of reliability will be to get to the point where people can program against Jev without making example queries.

当你完全信任Jev时，你会一直处于心流状态，疯狂地创造软件，而这就是很多……

When you just trust Jev, you'll be in perma flow state, just creating- ... crazy software, and that's where a lot of-

……软件现在的状态，对吧？

... software is today, right?

我觉得这并非完全不现实，但我会为此努力争取。

I don't think it's totally unrealistic, but I'm going to be fighting for it.

顺便说一句，这个问题有点奇怪，所以如果太怪了，你完全可以跳过。

By the way, this is kind of a weird question, so feel free, if it's too weird, just feel free.

但我突然想到，如果你有了这样一个基础工具，像编码代理这类东西的价值反而会下降。

But it occurs to me that actually the value of things like coding agents goes down if you have a primitive like this, in a way.

也就是说，你可以说，随便吧，Codex 帮我写所有代码，但它其实不用 Jev，所以它生成的软件本身是有限的。

Which is, you could be like, whatever, Codex builds all the software for me, but it doesn't actually use Jev, and so the software itself it creates is somewhat limited.

或者你也可以说，好，我作为一个人，我不用编码代理来写软件，但我有这个非常通用的基础工具，让写软件变得更容易。

Or you can be like, okay, I as a human being, I will write the software without using a coding agent, but I've got this very generalized primitive that makes writing software easier.

所以，你觉得未来会是编码代理用 Jev，然后你指挥编码代理，这样会有冗余吗？

So, do you feel, see a future where it's the coding agents using Jev, and then you're telling the coding agents, and then do you have redundancy?

还是你觉得更像是人类直接实现——

Or do you feel it's like humans implement-

这更像是一个关于编码代理的问题，而不是关于 Jev 的问题。

This is more of a coding agent question than it is a Jev question.

对，确实。

Oh yeah, for sure.

所以我的感觉是，我并没有像我希望的那样深入编码领域。

So my vibe is that I'm not in the coding minds as much as I'd like to be.

当然。

Sure.

所以你们俩可能比我更常接触那些。

So you two might be in there more than I am.

这有点遗憾，但我的经验是，它们在语法方面真的很强，而且真的很——

Which is sad, but my experience is that they are really good at syntax and really-

它们在语义方面很差。

They're bad at semantics

……

...

我觉得它们在架构方面差得离谱。

I would say incredibly bad at architecture.

所以对我来说，架构是软件中最具人类创造力的部分。

And so to me, architecture is the most human creative part of software.

所以我非常喜欢用编码代理。

So I love using coding agents.

我觉得Jev几乎肯定不在这个分布范围内。

I think that Jev is almost certainly not in distribution.

如果他们用我们的用户数据训练，那可就太诡异了。

That would be spooky if they trained on our user data.

所以很可能不是这样。

So it's probably not.

但我认为，当它在分布范围内时，让它处理语法部分我觉得完全没问题。

But I think that when it is in distribution, I see no problem with having it do the syntax.

关于架构，也许模型在架构方面不只是糟糕，而是处于第50百分位。

And the thing with architecture is that maybe the models are actually not just crap at architecture, but maybe they're 50th percentile at architecture.

如果你对架构一无所知，那也够用了。

And if you don't know anything about architecture, it would be fine.

所以这些都是灰色地带的权衡，需要你自己去把握，有时候速度就是你的公司或项目需要调整的关键。

So these are all gray area trade-offs in order for you to navigate, and sometimes speed is the knob for your company or project to turn.

你愿意接受第50百分位的架构而不是第60百分位的，因为你想更快推进，让Codex通宵工作之类的。

You're willing to do a 50th percentile architecture instead of a 60th because you want to move faster and have Codex work overnight or something like that.

其实，顺着这个话题说，市场上已经出现了一个有趣的现象：当编程代理刚出来的时候，大家都说SaaS要完蛋了，所有SaaS公司的估值都跌到了谷底。

Actually, kind of along those lines, one of the interesting things or phenomenons in the market already is that when the coding agents came out, it was the SaaSpocalypse, and all their values dropped through the floor.

然后等Jev出来的时候，每家SaaS公司都说：“这是有史以来最棒的东西。”

And then when Jev came out, every SaaS company is like, "This is the greatest thing ever."

确实，太棒了。

It's amazing, yeah.

那这怎么解释呢？

So explain that.

我也不知道该说什么了，对吧？

I don't know what else to say, right?

我觉得这很自然。

I think it's quite natural.

在SaaS末日论里，我觉得最站不住脚的说法就是软件很便宜，而且可能很容易复制。前者我信，后者我不信，因为很多复杂的东西都在表面之下。

In the SaaSpocalypse story, the story that I feel like has panned out really poorly is that software is very cheap and perhaps easy to replicate, which I think I could believe the former, I could not believe the latter, because a lot of the stuff happens beneath the hood.

完全同意。

Totally.

我可能有点太痴迷软件了。

I'm maybe overly a software fanboy here.

我们都是。

All of us.

好吧。

Okay.

我不知道你会怎么看待编程的影响，但是——

I didn't know where you might be the coding impact, but-

我们在这方面有很多历史包袱。

We have a lot of legacy around that.

完全同意，是的。

100%, yeah.

所以我觉得那个说法其实没怎么应验。

So I don't think that really panned out.

所以SaaS看起来可能市场不这么认为，但我觉得SaaS提供的价值跟以前一样。

So SaaS seems like maybe the markets don't agree, but I think SaaS is providing the same value it used to.

也许市场只是害怕了。

Maybe the markets are just scared.

但我认为SaaS将是整个AI游戏的最大赢家之一，我想和那些最大、最无聊、最了解用户问题的SaaS公司紧密合作，因为我认为他们最有能力知道哪些工作流程需要自动化，人们需要什么。

But I think that SaaS will be one of the largest winners of the whole AI game, and I want to work really well with all the biggest, most boring, most in-the-know of user problem SaaS companies because I think that they are the best positioned to know what workflows to automate, what do people need.

这是他们的看家本领。

That's what their bread and butter is.

而且为了投入大...

And to spend the big...

软件总是资本支出，但你提前投入是为了让这个体验变得更好。

Software is always a CapEx investment, but you spend it ahead of time in order to make this experience even better.

然后这个体验会传递给所有用户。

That gets distributed to all of that mass of users.

所以我认为这将是，我不会对金融市场做任何预测——

So I think that it's going to be, I'm not going to forecast anything about the financial markets-

……但我认为就能力游戏而言，这将是一场反向的SaaS末日，我对此感到非常兴奋。

... but I think as far as capabilities games goes, it's going to be like an inverse SaaSpocalypse, and I am so jazzed about it.

我得起个名字。

I should make a name.

你确实该有个名字。

You should have a name.

SaaSpalooza。

SaaSpalooza.

哦。

Oh.

这名字有点太好玩了。

That's a little too fun.

嗯，SaaS应用会突然变得超级有用。

Well, the SaaS applications are going to all of a sudden get dramatically more useful.

而且，顺便说一句，SaaS公司的资本投入，很大一部分其实是花在触达所有客户上。

And by the way, the kind of capital investment, so much of a SaaS company's capital investment is actually getting to all the customers.

所以，如果你已经触达了所有客户，然后你不仅是在SaaS产品上加个聊天机器人，而是真正把软件做得更好，那可就厉害了。

And so if you've gotten to all the customers and then you make, not just put a chatbot on your SaaS product, but actually make the software way, way better, that's a hell of a thing.

我不知道这是不是一个现实的梦想，但我觉得有可能有一天，那些多选表单会完全消失。

I don't know if this is a realistic dream or not, but I think that there's a world where the multi-choice forms just disappear.

我觉得它们就是——

I feel like they are-

……

...

总是把软件里已有的自然语言映射成类似Jev的输出。

always something mapping natural language that usually the software already has into a Jev-like output.

而且我觉得——

And I think-

这真的是80年代的东西。

It's literally from the '80s.

我们以前叫它们第四代语言。

We used to call them 4GLs.

马丁，你还记得吗？

Do you remember that, Martin?

第四代语言。

Fourth generation language.

其实，我觉得这也是80年代的东西。

Actually, also, I think this is from the '80s.

这可能有点冒犯。

This might be an insult.

那时候我还没出生呢。

I wasn't born then.

是的。

Yes.

我觉得“按我的意思做”会被提升到一个全新的高度。

I think that do what I mean is going to be taken to the absolute next level.

当然，对。

For sure, yeah.

如果我要特别提一个Jev应用，我不确定它是否可靠，所以我不能保证什么——

If I could shout out one Jev application, I don't know if it's reliable, so I can't promise anything-

……但那个真的太酷了。

... but it was so freaking cool.

有人用语音控制电脑，它基本上一直在判断，这是命令还是输入文字？

Someone was using a voice to control your computer, and it was basically constantly making decisions on like, is this a command or is it inserting text?

哦，那很酷。

Oh, that's cool.

输入文字在哪里？

Where's inserting text?

听起来简直酷毙了。

That sounds so unbelievably cool.

哦，那确实很酷。

Oh, that is cool.

我觉得界面可能会彻底改变，也许我们得让它更便宜、更快。

I feel like interfaces could just completely change, and maybe we're going to have to make it cheaper and faster.

那你就到了《星际迷航》的水平了。

Then you're at "Star Trek."

哦。

Oh.

嗯……

Well...

这里有一个非常深刻的直觉，就是如果你今天用AI来生成软件，你还是在创建和以前一样的软件。

There's just such a profound intuition here, which is, if you use AI today to generate software, you're still creating the same software that you did before.

但如果你看看大公司平均的PR，大概只有10行。

But if you actually look at the average PR for a large company, it's like 10 lines.

真的，不是开玩笑。

Seriously, no.

我在谷歌待过。

I've been at Google.

我们真的做过这个研究。

We actually did this study.

所以大概就是10行。

So it's like 10 lines.

所以你是在自动化那十行代码，而且顺便说一句，那十行代码可能只是从客户反馈中学到的一部分。

So you're automating 10 lines, and by the way, those 10 lines are part of a learning from a customer or something.

所以你优化的其实是很小的一部分。

So you've kind of optimized something that's actually pretty minimal.

但它并没有给软件带来新的功能，对吧？

But what it doesn't do is provide new capabilities to the software, right?

对。

Yep.

它只是在自动化这个相对次要的东西，而现在实际上有了新的能力，所以有可能软件真的会变得更好。

It's kind of automating this thing, which in the limit ends up being relatively minor, and now there's actually a new capability, and so it could just be the case that just software just actually gets better.

对。

Yep.

而且，甚至在Jev之前，我都没意识到，不管用多少AI编程代理，软件实际上并没有变得更好。

And by the way, even before Jev, it didn't even occur to me that it doesn't matter how much AI coding agents you use, the software actually isn't getting better.

也许你只是写得更快了。

Maybe you're writing it faster.

可以说反而更糟了，因为监督更少了。

It's arguably getting worse just because there's less oversight.

所以我觉得这——

So I think this is-

而且往往更不安全。

And often more insecure.

确实。

For sure.

但现在你确实可以论证，应用会因此获得新功能，因为你提供了这个新的基础组件，它某种程度上能理解自然语言，还能推理。

But you actually now can make an argument like apps will have new functionalities as a result of this because there is this new primitive that you're providing that, in a way, it speaks natural languages and it can reason.

对。

Yep.

但它把这一点和状态机结合起来了。

But it marries that to a state machine.

如果人们能把这个当作核心收获，那将是对我们所做工作的最大赞美。

If people take that as a takeaway, that would be the greatest compliment ever to what we are doing.

我其实觉得，扩展到我们现有的三个逻辑门之外，这个愿景有点太大了。

I actually feel like it's almost too grand of a vision to expand beyond the three logic gates that we have-

……进入……

... into...

我们的类型有点像同一个逻辑门，但里面有个小脑子。

Our types are kind of one of the same logic gate, but one that's a little brain in there.

那将是对TypeSafe遗产的最大赞美，因为那是一个非常……

That would be the greatest compliment to the TypeSafe legacy, because that's a very-

巨大的

Huge

……对世界来说，非常重要、非常重大的事情。

... nontrivial, huge thing for the world.

我不会过度承诺，然后交付不足，但我会为此奋斗。

I'm not going to overpromise, underdeliver that, but I will fight for that.

听着，我觉得有一些相当开放的问题，关于这能深入到什么程度，比如真正严肃的事情，像状态一致性、持久性，或者真正的系统级问题，在这些地方你需要提供强有力的保证。

Listen, there's, I think, pretty open questions to how deep can this get as far as really serious stuff like state consistency or durability or real systems level stuff where you actually need to provide strong guarantees.

所以百分之百，这会改变一些事情，比如分析日志、分析邮件、提供用户界面、与人对话，这些肯定会的。

And so 100%, this will change things like whatever, analyzing logs, analyzing emails, providing a UI, talking to the human, that for sure.

但你可以说，随着时间推移，这会变成一个智能数据库。

But you could argue that over time this becomes a smart database.

对。

Yep.

所以——

And so-

还有——

And also-

空中交通管制系统。

Air traffic control system.

什么都可以，对吧？

Anything, right?

这正是我们非常需要的。

Which we really need.

确实。

True.

有点吓人。

A little scary.

我一直觉得，先自动化简单的工作，再处理困难的工作，这是我的原则。

I think automate the easy work before the hard work is always my philosophy.

但我也觉得，这会开启一个概率编程的全新时代。

But I also think there's going to be an entire era of probabilistic programming that's opened up.

顺便说一句，你知道吗，概率编程有着悠久的历史。

By the way, you know there's a huge history of probabilistic programming.

哦，我也知道。

Oh, I do as well.

它基本上在70年代就消亡了。

It basically died in the '70s.

我对此很熟悉。

I'm familiar with it.

我其实觉得它会变成——同样地——你也可以把Jev称为神经逻辑AI。

I actually think it's going to be- With the same-- You could also call Jev like neurologic AI.

所以你的联合创始人Eric就是那个背景出身的，他跟我说的。

So your co-founder, Eric, came from that background, he was telling me.

哦，不错。

Oh, cool.

他们有过贝叶斯——

They've had a Bayesian-

哦，是的。

Oh, yes.

哦，对。

Oh, yeah.

他做了很多生物学相关的工作。

He did a lot of biology stuff.

这个领域起起伏伏的。

It goes up and down.

但我的意思是更……

But what I mean is a more...

我不太喜欢。

I'm not a fan.

我的风格是务实。

My brand is pragmatism.

非常务实。

Incredible pragmatism.

我完全不看好那些受生物学启发的东西，从来都不看好。

I'm not a fan of biologically inspired stuff- ... at all, ever.

那些从来都没成功过。

It's never worked.

你有没有注意到这一点？

Have you ever noticed that?

我是说——

I mean-

几乎没怎么。

Barely

……

...

我觉得从来就没成功过。

I think it's never worked.

它的用处就是激励那些疯狂的人，让他们几十年如一日地钻研，直到……

It's useful to motivate crazy people to work on things for decades until-

……直到成功，然后他们再把它打磨成工程版本。

... it works, and then they refine it into the engineering version.

人工智能和神经网络的故事肯定就是这样。

The story of AI, neural nets for sure.

对。

Yes.

但很多关于它如何成功的说法其实并不准确，对吧？

But a lot of the stories about how it worked were not accurate, right?

是的。

Yes.

所以，应用的分层特征其实并没有真正奏效，不然ResNets也不会成功。

So, the hierarchical features of applications really didn't end up working, because otherwise ResNets wouldn't have worked.

说来话长。

Longer story.

我确实觉得，从系统角度来看，这会打开新局面，但我不太兴奋，因为这真的——我为世界感到兴奋，而不是为我自己编程这件事，因为听起来太复杂了。

I do think that it opens up, from a systems perspective, I'm not excited about this part, because it's really-- I'm excited for the world, not about me programming this, because it sounds really complicated.

但我觉得，当我们拥有大量不同成本和速度权衡的智能时，那些超级系统型的人会做出权衡，就像Jev对他们来说会聪明一千倍。

But I think that as we have lots of intelligence at lots of different cost and speed trade-offs, the super systems-y types will be making trade-offs at like, Jev's going to be 1,000 times too smart for them.

他们只是想要一个近似的链接，做一个近似的猜测，以便乐观地路由到各处。

They just want an approximate link to have an approximate guess to optimistically route here and there.

这会变得非常疯狂，这——

It's going to be so crazy, this-

……在极端系统端可用的那种东西。

... type of stuff that's available in the extreme systems end.

好消息是我们可以重新构建系统了，这很棒，对吧？

And the good news is we get to rebuild systems again, which is great, right?

不，说真的，我们有了一个新的基础元素。

No, seriously, we have a new primitive.

这有点像一种新的思维方式——

It's kind of a new way of thinking-

……关于如何做软件。

... about doing software.

听着，我们当年在互联网上就这么干过。

Listen, we did this with the internet.

我们也是从大型机转到客户机/服务器架构的。

We did this mainframe to client server.

我是说，我们定期都会这么做。

I mean, we do this periodically.

而且顺便说一句，由于网络安全问题，我们可能不得不重建几乎所有的系统——

And by the way, just because of the cybersecurity issues, we probably have to rebuild almost all the systems-

对，当然。

Yeah, for sure

……只是为了确保它们安全。

... to just make them safe.

我也这么觉得。

I would think.

我觉得很明显，现在没有——

I think it's pretty clear that there is not-

或者至少是关键的——

Or at least the critical-

……基础设施。

... infrastructure.

肯定的。

For sure.

你更多是从应用、SaaS、分析的角度来考虑这个问题，还是从系统、基础架构的角度，或者两者都有？

Do you think about this more in terms of apps, SaaS, analytics, or more in terms of systems, foundations, or all the above?

就我个人的想法，或者怎么说呢——

For what I would think of, or how-

对，就是一般性的应用场景。

Yeah, just general application for this.

当你想到你在做Jev，然后你想象人们会怎么采用它，比如——

When you think about you're working on Jev, and you kind of envision that people are adapting it, like how-

哎呀，天哪

Oh, man

……

...

你心里有想法吗？

do you even have an opinion?

我有一点想法。

I have a little bit.

所以我的思考方式有点深入到TCP的内部机制里。

So the way I think of it is a little deep into the TCP guts.

像UDP/TCP，它不可靠。

Like UDP/TCP, it's unreliable.

你说的正是我想听的。

You're speaking my language.

没错。

Exactly.

而且——所以当我想到AI的时候——

And like- So when I think of AI-

我觉得很自在。

I feel comfortable

……不，不，不。当我想到AI，这就是为什么我在乎每美元能买到多少智能，说清楚点。

... No, no, no. When I think of AI, and this is why I care about intelligence per dollar, to be clear.

我是怎么得出这个结论的，我是从AI驱动的经济革命、AI无处不在、科幻小说里那样，所有软件都充满AI，这样倒推回来的。

When I think, and how I got to this conclusion, so I work backwards from AI-based economic revolution, AI everywhere, sci-fi and everything, all the software has AI all over the place.

然后我问自己一个问题，对AI的调用有多大比例，想象它就像一个函数——

And I ask myself the question, what percentage of the calls to AI, imagine it's like a function-

有多少比例是给人看的，需要讲究风格什么的？

... what percentage are for human consumption, where you need the style and everything?

而不是为了理解。

Versus understanding.

对，那比例会非常高，接近百分之百。

Yeah, and it's going to be many nines.

而且，从同一个问题出发，有多少会发生在最表层——

And actually, from that same question, how many will be at the first layer-

——而不是深入核心内部？

... versus deep in the guts?

我觉得大部分都会在核心内部，但一开始会从表层开始。

And I think that it's going to be many nines in the guts, but it will start at the first layer.

但如果你不瞄准核心，等等，这么说有点奇怪。

But if you don't aim for the guts, wait, that sounds weird.

如果你不瞄准核心，那要花很长时间才能达到那个目标。

If you don't aim for the guts, it's going to take you a while to get there.

我觉得人们不太理解，AI和软件在某种程度上就像是夜航的船，擦肩而过。

I think people don't understand to what extent AI was kind of ships in the night with software.

即使你试图把AI嵌入到软件里，它也不太听话，对吧？

Even if you try to embed AI in software, it kind of didn't behave, right?

因为软件本身不擅长处理自然语言，你得搞各种奇怪的操作。

Because software doesn't really take natural languages, and you do all this weird stuff.

你塞进去一个提示，比如“这是你想要的JSON输出，这是schema”，但它根本不理会。

You stick in the prompt like, "Here's the JSON output that you want, and here's a schema," and it would never listen to it.

所以最后你只能把输出拿过来，交给人工处理。

And so what you ended up doing is just taking the output and giving it to a human.

你心里想：“算了，不管了”，对吧？

You're like, "To hell with it," right?

对。

Yes.

或者交给另一个大语言模型。

Or another LLM.

这就是while循环的本质。

That is what a while loop is.

就像代理的while循环，对吧？

Like the agent while loop, right?

所以从基本原理上讲，它需要人在循环中，也就是聊天。

So it's like from first principles, it needs to be human in the loop, which is the chat.

或者一个代理，也就是while循环，因为自然语言需要反馈到另一个循环中。

Or an agent, which is the while loop, because the natural language needs to be fed back into another.

对。

Yes.

而且我要说，我亲眼目睹过这种情况发生。

And I will say, I have watched this happen.

几乎就像经历了悲伤的五个阶段。

There was almost this kind of five stages of grief.

人们会拿起AI，然后说：“我要把它用在我的软件里”，对吧？

People would pick up AI and like, "I'm going to use this within my software," right?

然后——

And then-

……就会经历各种阶段，先是否认，试着让它工作，然后是愤怒。

... it would go with whatever, denial, try to make it work, and anger.

然后他们进入接受阶段，就像说：“好吧，算了，我干脆把这个交给另一个大语言模型，或者交给真人。”

Then they go to acceptance, which is like, "Okay, never mind, I'm just going to give this to another LLM, to a human being."

所以这就像夜航船一样，我觉得这是我第一次看到——

So it's been very ships in the night, and I think this is the first time I have seen-

……有人会说：“其实你可以把大语言模型、把AI，真正映射到一个状态机上，而且能高效地做到这一点。”

... when someone's like, "Actually, you can take an LLM, you can take AI, and you can actually map it to a state machine, and you can do that productively."

我也希望如此。

And I hope so.

我也不想过度承诺，最后却做不到。

I will not want to overpromise under deliver as well.

我不知道它是否已经准备好应对所有的应用场景——

I don't know if it's ready for all the applications-

哦，当然。

Oh, sure

……那些被过度承诺的应用。

... that have been overpromised.

我真的很希望它能做到，而且我的团队显然会为此努力。

I really, really want it to, and my team will fight for that, obviously.

我们真的非常在意可靠性。

We really, really care about reliability.

我们本来可以更早发布的。

We could've released so much sooner.

我觉得人们没有意识到这一点，而且说实话，我觉得他们也不会意识到。

I don't think people realize that, and I don't think that, honestly, I don't think that they will.

根据我在推特上看到的讨论——

Based on what I see of the Twitter discussion-

……

...

我觉得人们永远不会明白这一点。

I think people will never get it.

但它就是会有那种让人感觉很好的氛围，就像‘哦，我可以信任这个’。

But it'll just have that good vibe of like, "Oh, I can trust this."

所以——

So-

嗯，这就是那个‘防挫败机器’，对吧。

Well, it's the anti-frustration machine, yeah.

是的。

Yes.

我希望如此。

I hope so.

明白我的意思吧？

Do what I mean, right?

是的。

Yes.

对我来说，这就是让世界更顺畅，让所有东西都更顺滑地协同运作，像齿轮一样咬合在一起。

To me, that is about smoothness in the world, having everything just move more smoothly together and interlink like gears.

其实我对AI的乌托邦有很多不同的维度，我真的非常想要，而“懂我意思”是其中非常重要的一部分。

I actually have my whole AI utopia on different axes that I really, really want, and do what I mean is a huge part of this.

想象一下，如果所有技术都能懂你的意思。

Imagine if all technology just did what you mean.

这不是科幻。

That's not sci-fi.

看看现在的AI有多聪明，对吧？

Look how smart AI is, right?

对，真的很厉害。

No, it's amazing.

也许这就是我们该收尾的地方。

And maybe that's the thought to close on.

哦。

Oh.

按我的意思去做。

Do what I mean.

是的，我很喜欢。

Yeah, I love it.

谢谢你，迪奥戈。

Thank you, Diogo.

这次真是——

This has been-

谢谢。

Thank you

——一次很棒的对话。

... a great conversation.

太棒了。

Hell yeah.

真的很享受。

Really enjoyed it.

太好玩了。

So fun.

感谢收听本期a16z播客。

Thanks for listening to this episode of the "a16z" podcast.

如果你喜欢这期节目，记得点赞、评论、订阅，给我们评分或留言，并分享给你的朋友和家人。

If you liked this episode, be sure to like, comment, subscribe, leave us a rating or a review, and share it with your friends and family.

更多节目，请前往YouTube、Apple播客和Spotify。

For more episodes, go to YouTube, Apple Podcasts, and Spotify.

在X上关注我们@a16z，并订阅我们的Substack，地址是a16z.substack.com。

Follow us on X, @a16z, and subscribe to our Substack at a16z.substack.com.

再次感谢收听，我们下期见。

Thanks again for listening, and I'll see you in the next episode.

提醒一下，这里的内容仅供参考。

As a reminder, the content here is for informational purposes only.

这些内容不应被视为法律、商业、税务或投资建议，也不应用于评估任何投资或证券，并且不针对任何a16z基金的投资者或潜在投资者。

It should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any a16z fund.

请注意，a16z及其关联公司可能也持有本播客中讨论到的公司的投资。

Please note that a16z and its affiliates may also maintain investments in the companies discussed in this podcast.

更多详情，包括我们投资的链接，请访问a16z.com/disclosures。

For more details, including a link to our investments, please see a16z.com/disclosures.