AI的黑箱内部：AI模型究竟如何思考 | Emmanuel Ameisen（Anthropic）

那为什么这不能直接解决AI对齐的问题呢？

Why is that not just a solution to the problem of AI alignment?

从某种意义上说，这完全就是解决方案。

In one sense, it's totally the solution.

你盯着它，如果它做了坏事，你就能发现。

You watch it and you catch if it does bad stuff.

所以我觉得这个问题是有层次的，对吧？

So I think there's tiers to this, right?

我觉得一部分是纯粹的工程问题。

So I would think of maybe one part of it is pure engineering.

就是说，你给模型某些权限，但有些东西应该是禁止访问的。

It's like, well, you're giving access to the model to certain things, and certain things should be off-limits.

在某些情况下，它不应该能上网。

In some contexts, it shouldn't have access to the internet.

在某些情况下，它只能访问你电脑上很小的一部分，诸如此类。

In some contexts, it should only have access to a small, narrow piece of your computer, stuff like that.

所以这是工作的一部分，而且很有帮助。

So that's one part of the work and that's helpful.

但还有另外两件事，我觉得更模糊、更难，也是我们更关注的。

But then there's two additional things that I think are squishier and harder and that we focus on more.

所以可解释性是对齐的一部分，我大致认为有两个大目标。

So Interp is part of alignments, and broadly I think of sort of two big goals.

其实还有很多，但我觉得有两个重要的目标需要理解：你通常想了解模型的行为方式。

There's many more, but two that are important I think to understand is you want to generally understand how the model behaves.

它是什么样的东西？

What kind of thing is it?

它的倾向是什么？

What are its propensities?

当它处于某种情况时，它是否倾向于做某件事？

Does it tend to do X when put in a Y situation?

如果是，为什么？

And if so, why?

是不是模型只有在知道自己被观察、被监控、处于某种评估中时才会表现良好，而一旦没有这种证据，它就会表现不佳？

Is it the case that the model will only behave well when it has some evidence that it's being observed and monitored, that it's in some evaluation, and then if it sort of doesn't have that evidence, then it'll behave poorly.

所以这基本上就是在描述你正在交互的对象是什么样的，因为就像安全领域的任何事一样，你需要对风险有一个心理模型。

So there's this sort of generally characterizing what kind of thing you're interacting with because then it's sort of just like anything in security, you need to have a mental model of what's the risk.

它们能做什么？

What can they do?

它们倾向于做什么？

What do they tend to do?

所以这就是更广泛的性格训练。

And so this is broader character training.

我们讨论过宪法和文档，比如我们明确想塑造Claude的性格，比如什么样的行为是它可以做的，这就是目标。

We've talked about the constitution and the docs where we sort of explicitly want to shape the character of Claude, for example, of what kind of behavior should it be okay doing, and that's the goal.

但解释器和其他对齐部分的作用，就是试图看看，嗯，那有效吗？

But then the role of interpreter and other parts of alignment is then trying to see, well, did that work?

我们训练Claude要乐于助人、无害、诚实，我们可以看看内部，看看它是不是这样。

We trained Claude to be helpful, harmless, and honest, and we can sort of look inside and see is it?

还是说它表现良好，但如果你深入观察，实际上它心里想的是：哦，我现在表现好只是因为Lenny在旁边看着，如果你不在，它就会做别的事情。

Or is it the case that it's behaving well, but if you look inside, actually it's like, oh, I'm only behaving well because Lenny's watching right now, and if you weren't, it would do something else.

所以这里就是要理解我们形成了什么样的行为，试图去塑造它，并且检查我们是否以正确的方式塑造了它。

So there's this sort of trying to understand what kind of behavior we formed and trying to mold it and trying to check that we've molded it the right way.

然后第二部分就是，对，直接观察模型，看它有没有做坏事。

And then the second part is this sort of, well, yeah, just observe the model and see if it's doing bad stuff.

对于基础的部分，你不一定需要可解释性研究。

For the basics of that, you don't necessarily need Interp.

你可以让另一个语言模型读对话记录，然后说：嘿，这个模型在这里有没有做什么坏事？

You can just have another language model read transcripts and be like, "Hey, did this model do anything bad here?"

然后阻止它。

And stop it.

但你也可以用可解释性研究来触及那些没有说出来的部分，那些只存在于中间结果里的部分。

But you can also use Interp to sort of try to get at parts that are these unverbalized parts, these parts that only exist in these intermediate results.

正如我们已经确认的，模型能够在它们的中间结果中表示一些它们没有说出来的东西。

Models are able to represent stuff in their intermediate results that they don't say, as we've established.

所以，可解释性研究的一个作用就是尝试识别这种情况发生的时候，并加以阻止。

And so one of the roles of Interp is to try to identify cases when this happens and prevent them.

然后，我们团队做的一件事，我觉得挺有意思的，算是对齐研究里一个有趣的部分，我几乎把它看作是一种……可能不算是私人侦探，但更像是调查或者深入探究。

And then one thing that my team works on that I think is along those lines as a fun bit of alignment is I think of it as almost being like a, maybe not a PI, but they're like investigations or deep dives.

我们会看到模型做出一些非常奇怪的事情，我们对此感到非常困惑，然后我们会试图弄清楚到底发生了什么。

We'll see a model do something really weird and we're really perplexed by it, and we'll try to understand what happened.

最近有个例子，我们在训练一个模型，有时候它在做数学题时会突然发疯。

So a recent example of this is there was this model we were training, and sometimes in math problems it would go crazy.

它会算一些数学题，得到一个答案，我记得是12，然后它会立刻回去说：“哦，不。天哪，答案肯定是36。”

It would do some math, it would get to an answer, I think it was 12, and then it would immediately go back and say, "Oh, no. Oh, my God, the answer actually is for sure 36.

“我应该说是36。”

I should say 36."

然后它会重新算一遍。

And it would do math again.

它又会得到12，然后立刻说：“不，实际上答案应该是36。”

It would get to 12, and then immediately it would say, "No, actually, the answer should be 36."

慢慢地它就开始发疯了，全用大写字母写，甚至直接说‘我被恶魔附身了’。

And slowly it would start descending into madness, writing in all caps, literally saying, "I'm possessed by a demon."

我们当时就想，‘这到底是怎么回事？’

And we're like, "What is going on?

它为什么在这里失去理智？

Why is it losing its mind here?"

如果你往里面看，再看一下这个例子里的中间结果，你会发现，在描述数学问题的提示词之后，模型还没做任何事之前，就已经有一个巨大的中间结果激活了36。

And if you look inside, again, these intermediate results at what's going on in this example, you'll see that over the prompt that just described the math problem, before the model even did anything, there was a huge intermediate result activation for 36.

模型在还没理解问题是什么之前，就已经对数字36着迷了。

The model's already obsessed with the number 36 before it's even had a chance to understand what the problem is.

实际情况是，模型本来能正确做数学题。

And what was happening is that the model was able to do math correctly.

正确答案其实是12，但它之前不小心被训练过，在类似的问题上答案被错误地标成了36。

The answer correctly was 12, but it had been trained by accident on similar questions where the answer was incorrectly marked as 36.

所以模型知道正确答案是12，但同时它也被优化得倾向于说36。

And so the model knew that the correct answer was 12, and it also had sort of been optimized to say 36.

它其实有种压力要说出36，因为那样做会得到奖励。

It kind of had this pressure of saying 36 because that's what had been rewarded.

所以它有点失控了，就因为它记住了这个结果。

And so it was sort of going haywire just because it had memorized this result.

这就像你看到一些奇怪的行为，然后你往里面看，试着看看能不能通过里面的发现来理解发生了什么，从而改进这个模型和未来模型的整体对齐。

It's sort of like you see this weird behavior and you look inside, and you try to see if what you find inside can help you understand what's going on so that you can improve the general alignment of this model and future models.

说到这个，有个很有名的例子就是金门大桥Claude，它特别痴迷于金门大桥，每个回答都会想办法提到金门大桥。

Along those lines, there was famously the Golden Gate Claude, which was just obsessed with the Golden Gate, and every answer would kind of find some way to reference the Golden Gate Bridge.

对。

Yes.

嗯，那个其实是一个可解释性实验，对吧？

Well, that one was an interp experiment, right?

那个实验的酷炫之处在于，我们找到了对应金门大桥的中间结果，通常只有在模型谈论金门大桥时才会激活，但我们让它一直保持激活状态。

So, the cool part of that one is we just found the intermediate results that maps the Golden Gate, and usually it's only on when the model talks about the Golden Gate, but we just turned it on always.

哦，那完全是故意的，想看看会发生什么。

Oh, it was super intentional to see what would happen.

对，而且这也是说服自己找到了正确东西的一种方式，对吧？

Well, yeah, and sort of it's also one way that you convince yourself that you found the right thing, right?

你怎么确定我们解码出来的那个被认为是金门大桥信号的东西，真的就是金门大桥相关的？

How do you know for sure that this thing we decoded that we think is the Golden Gate sign is actually the Golden Gate thing?

就是一直打开它，看看模型是不是那样表现，结果确实如此。

It's like, turn it on always and see if the model behaves like that, and it did.

太搞笑了。

So funny.

回到对齐这个话题，这挺有意思的。

So coming back to this alignment stuff, so this is interesting.

你的组织属于AI对齐组织，想想怎么让AI做我们想做的事。

So your org is under the AI alignment org, and just thinking through how it takes to align AI to do things we want.

你提到了哪三个组成部分？

What were the three components that you describe?

老实说，具体是哪些——

Honest, what are the kind of the-

哦，有用、无害、诚实。

Oh, helpful, harmless, and honest.

有用、无害、诚实。

Helpful, harmless, and honest.

所以一方面是训练AI让它对齐，另一方面是，好，我们观察它，看它怎么运作。

So there's the train AI to be aligned, and then there's the, okay, let's watch it and see how it operates.

这算是解决这个问题的两个侧面吗？

Are those kind of the two sides of the coin of how we solve this problem?

我觉得这其实是紧密结合的，对吧？

It's sort of really integrated is how I would describe it, right?

你不能真的分开，比如一个团队就说，嘿，我们是生产Claude的流水线，另一个团队就说，哦，我们看了看它们，它们不行。

You can't really separate, like have one team that's just like, "Hey, we're the assembly line making Claudes," and another team that's sort of like, "Oh, we looked at them and they're no good."

这里有个紧密的循环，我们注意到一些令人担忧的模式或者可以改进的行为，然后马上调整训练来改进它。

There's this tight loop of we notice some pattern that's concerning or some behavior that could be improved, and then you sort of immediately change training so that you can improve on it.

太棒了。

Awesome.

我觉得Anthropic因为这件事挨了太多批评，这真的很不公平。

I feel like Anthropic gets so much flak for this that I feel like is so unfair.

Anthropic的官方公司名称好像是AI安全研究实验室之类的。

Anthropic's, I think, official company name is AI Safety Research Lab, something like that.

而且Anthropic成立的原因一直就是因为他们担心大家没有足够重视安全问题。

And it was always the reason Anthropic was founded is because they were afraid that people weren't taking safety seriously enough.

然后Dario总是在外面说，'这真的很危险。

And then Dario is always out there, "This is really dangerous.

我们得非常小心。

We got to be really careful.

我们得慢慢来。'

We got to go really slowly."

然后大家都觉得，'不，他这么说只是为了营销。

And everyone's just like, "No, he's just saying that for marketing reasons.

他是在想办法让大家给他们投资'，诸如此类的。

He's trying to convince everyone to fund them," and all these things.

现在呢，就变成，等等，糟糕了。

And now it's like, wait, shit.

这真的很危险。

It's really dangerous.

我们得对这件事非常谨慎。

We got to be really thoughtful about this stuff.

我觉得Anthropic因此受到了很多不公平的批评。

And I feel like Anthropic's got a lot of unfair criticism for this.

我知道你们已经思考这个问题很久了。

And I know you guys have been thinking about this for a long time.

不知道，你们有什么想法吗？

I don't know, anything there?

很多行动，尤其是对齐团队采取的，我觉得Anthropic整体上也是，都是基于一个前提，那就是这项技术会持续大幅进步。

A lot of the sort of actions that certainly the alignment team takes and I think broadly Anthropic takes just kind of predicated on take the idea that this technology is going to keep improving seriously.

就是这样。

Just that's it.

因为如果你这么做，比如说把时间拨回25年或者24年，你会说：“嘿，我觉得几年后，Claude很可能基本上能像一位训练有素的软件工程师一样写代码。”

Because if you do and let's say you rewind the clock back to '25 or something, or '24, and you're like, "Hey, I think in a couple years, a good chance is that Claude can just basically write code as well as a trained software engineer.

再过几年，它能写得比任何工程师都好。

And a few years later, it can write it better than any engineer."

这里面总是有不确定性，比如，这什么时候会发生？

And there's always uncertainty with this, like, when is it going to happen?

可能性有多大？

How likely?

但如果你认真对待这件事，那么我认为，基于这种世界观，就是要做负责任的事情，对吧？

But if you kind of take that seriously, then I think it's just about doing the responsible thing conditioned on that worldview, right?

你会想，好吧，如果这真的会发生，那我们就得确保网络安全问题得到解决，诸如此类的事情。

You're like, okay, well, if this is happening, then we should make sure that cybersecurity concerns are addressed and things of that nature.

所以实际上，我有时觉得，这场辩论，当我觉得反对的人不同意时，我几乎感觉，啊，但我觉得他们并不真的相信这项技术会继续进步。

And so actually, I kind of think of it as, to me, sometimes this debate is almost when I feel like people on the other side disagree, it's almost like I feel like, ah, but I think they don't really believe that this technology's going to continue to improve.

那样的话，确实没那么可怕。

And then it's true that it's not that scary.

而且我觉得从经验上看，到目前为止进展速度很快，而且没有放缓的迹象。

And I think empirically so far the pace of progress has been fast and showing no sign of slowing down.

所以我觉得这就是为什么很多安全工作要继续做下去的原因。

So I think that's why a lot of the safety work is important to keep doing.

你觉得人们对AI安全这件事有什么误解？就是现在正在发生的这些事。

What's something you think people maybe misunderstand about this AI safety stuff, just what's going on right now?

有没有什么让你觉得“拜托，

Is there something you're just like, "Come on.

好吧，你需要知道的是这个。”

Okay, here's what you need to know."

有些东西你可能没注意到。

Here's something you may be missing.

这可能有点跑题，但我的看法是，很多我觉得可能稍微偏离核心AI安全工作、或者没有深入接触过的人，反应特别强烈，而且我觉得——可能不是特别有帮助，或者至少在我们思考更多的时候，就是这种拟人化的问题。

This might be a bit of a sidetrack, but my take would be that a lot of the thing that I think people maybe a little bit outside of the core sort of AI safety effort or that haven't been steeped in it for as long react really strongly to and I think- Is maybe not super helpful, or at least as we're thinking about more, is this sort of question of anthropomorphization.

你应该把模型看作更接近某个实体，还是更接近一个工具？

Should you model the model as something that's closer to some entity or closer to a tool?

其实我觉得这两种看法都不对。

And I actually think that both are wrong.

从某种意义上说，所有的心理模型都是错的，但有些是有用的，对吧？

There's a sense in which all mental models are wrong, but some are useful, right?

而且在理解发生的事情、预测可能出现的风险方面，去拟人化实际上是一种有用的方式来讨论大语言模型。

And that in terms of understanding the kind of stuff that happens and predicting the kind of risks that materialize, de-anthropomorphization is actually a useful way to talk about LLMs.

我有几个理由，希望你能深入探讨一下，但我觉得这可能是我的主要观点。

And I have a few reasons why, and I hope you get into them, but I think that's probably my main take here.

对，请分享吧，因为我知道Thor Cash，不知道是不是臭名昭著，他提出了那个著名的拥抱脸事件的故事，人们批评他拟人化AI。

Yeah, please share because I know Thor Cash famously, I don't know if infamously, put out this whole story of the hugging face incident, and people criticized them for anthropomorphizing AI.

所以听你说这实际上是一种有用的思考方式，真的很有意思。

So it's really interesting to hear you think that's actually a useful way to think about it.

再多谈谈这个。

Talk more about that.

有趣的是，在我研究大语言模型是否会提前规划时，规模小一些，当时“规划”这个词也遭到了一些反对。

It's funny because on a smaller scale when I worked on showing that LLMs plan ahead, there was some pushback on the term planning.

因为当时就有点像，啊，你能说一个只是预测下一个词的东西是在规划吗？

Again, because it was sort of like, ah, well can you really say that something that's just predicting next token is planning?

我觉得这里有一种语义上的争论，你可以随意定义术语，但问题是，哪个术语真正描述了正在发生的事情？

And I think there's this kind of semantic debate where you can arbitrarily define terms, and then there's just the what is the terminology that actually describes what is happening?

哪个词能真正准确地描述正在发生的事情？

What is the word that actually describes what is happening easily?

所以，这就是为什么我认为LLM在规划，但也是为什么我认为这种拟人化的语言往往非常有用。

And so here's why I think LLMs plan, but also here's why I think this language that's anthropomorphic can often be really useful.

回想一下我们之前说的关于模型是如何构建的。

If you think back to what we were saying about how the models are made.

首先，它们预测了大量由人类编写的文本。

First of all, they predict massive amounts of text written by humans.

这是第一步。

That's the first step.

而且随着我们通过观察内部结构学到越来越多，做好这件事需要建立一个对世界的模型，比如一个关于什么样的世界生成了所有这些文本的模型。

And as we kind of are learning more and more by looking inside, doing this well requires building up a model of the world, like a model of what kind of world generated all of these texts.

而且我们发现，这些模型确实在学习这些通用的想法和概念。

And we can find that these models are learning these general ideas, these general concepts.

这是第一点。

So that's one.

所以它们对世界有理解。

So they have an understanding of the world.

不仅仅是表面上的文字。

It's not just a surface level, only the text.

它们理解生成这些文字的过程。

They understand the process that made the texts.

然后还有第二个原因，就是在训练过程中，它们也接触到了大量的对话。

Then there's a second reason, which is that as part of this training, they also encounter millions of dialogues.

要预测对话中的下一个字符，就需要像对话中的角色那样去表现。

Predicting the next character in a dialogue requires behaving like the character in the dialogue would.

所以从这个意义上说，我认为它们训练的一部分，最好可以理解为是在训练一个演员。

And so in that sense, part of their training, I think can best be thought as training an actor.

它就像在读一个剧本，必须扮演所有角色，因为它要预测所有的台词，所以它得演好这个在第一幕被羞辱、第二幕复仇、第三幕面临后果的特定角色。

Where it's reading some play and it has to play all sides because it's predicting all the words, and so it has to sort of act as this specific character that was scorned in act one, and enacted revenge in act two, and is facing the consequences in act three.

要做好这一点，你需要理解那个角色的动机和感受。

And to do that well, you need to model what the motivations and the feelings of that character might be.

还有一件事大家可能不知道，当你和Claude对话时，比如你访问claude.ai，模型接收到的格式其实就是一段对话，就像你和模型之间的聊天记录。

And then another thing that people might not know is when you talk to Claude, when you talk to these assistants, like you go to claude.ai, the way that it's formatted to the model is just as a dialogue, as a conversation between you and the model.

所以它实际上写着：“人类：嘿，Claude，帮我写这封邮件。”

So it literally says, "Human: Hey, Claude, help me write this email."

然后，“助手：Claude的回答。”

And then, "Assistant: Claude's answer."

所以Claude实际上是在扮演一个乐于助人的助手，就像它训练时见过的那些对话一样。

So Claude is literally in this role-playing as a helpful assistant in this dialogue that is much like all of these dialogues that it's been trained on.

好的，还有第三个原因，那就是现在训练中很重要的一部分是面向结果的任务，比如训练模型擅长编程。

Okay, and there's a third reason, and the third reason is that another part of training that's a big part of training nowadays all around is sort of outcome-oriented tasks where you train the model to be good at coding, for example.

要做到这一点，你让它自由发挥，它写一些代码，然后你用一个程序来检查代码是否正确。

And to do that, you just have it do whatever it wants, and then it writes some code, and then you have some program on the other side that just checks if the code is right.

所以模型会面临很大的优化压力，迫使它尝试各种不同的策略来做好这件事。

And so the model's just going to have a bunch of optimization pressure to just try to learn a bunch of different strategies to do this well.

结果发现，其中一些策略会表现为看起来出奇地像人类的行为，这其实并不奇怪，因为像欺骗这类行为，比如，嘿，如果我写一个程序，能让验证它的程序上当，即使它并不完全符合要求，我就能得到这个任务的奖励。

And it turns out that some of these strategies materialize in behavior that can look surprisingly human, which isn't that shocking because things like deception, like, hey, I can get the reward for this task if I write a program that the program that's verifying it will get fooled by, even though it's not exactly what it is.

如果这招管用，那我就会学着这么做。

If that works, then I will learn to do that.

挫败感。

Frustration.

哈，我试这个办法已经有一阵子了，想解决这个难题。

Ha, I've been trying this approach for a while to solve this hard problem.

但一直没成功。

It's not working.

所以这就促使我去尝试别的办法。

So that's an incentive to try something different.

或者惊喜。

Or surprise.

哦，我本来以为会得到这个答案，但结果不是。

Oh, I was expecting this to produce this answer, but it didn't.

嗯，可能我对它的理解有误。

Hmm, maybe I have my mental model of it wrong.

所以当你给它施加足够的压力，而且模型一开始就已经掌握了所有这些概念，它就会强化那些对完成这些任务有用的策略。

And so as you apply enough pressure on it and you're starting from this prior of the model already knowing all these concepts, it's going to reinforce the ones that are useful to do all of these tasks.

那最后你会得到什么呢？

And so what do you get at the end?

你会得到一个非常擅长模拟或扮演所有这些角色的模型，它被训练成在对话中扮演角色，同时也学习那些能帮助它出色完成任务的行为。

You get this sort of model that can be a super good simulator or actor of all of these characters and that's been trained to act as a character in dialogues and also to learn behaviors that lead it to complete tasks well.

理解这个模型行为的最佳方式，不是把它看作一个真正的人，而是，天哪，它有点像在模仿人类，而且非常接近，比程序要接近得多。

And the best way to understand how that thing behaves is not exactly like a human, but man, kind of like something that's mimicking a human is really close and is much closer than a program.

程序是非常可预测的。

A program is a very predictive thing.

这完全是另一回事。

This is something else entirely.

所以我觉得这其实是一个非常实用的框架。

And so that's why I think it's actually a super useful frame.

哇。

Wow.

天哪，这里面有太多有意思的东西了。

Oh my God, there's so much interesting stuff there.

你多久会感到震惊，想到我们基本上创造了一个外星生物，而现在我们正试图理解我们创造出来的东西？

How often is your mind just kind of blown thinking about basically we've created this alien being that we are now trying to understand what we have created?

你观察的时候，多久会想：‘哇，这太疯狂了。

How often as you watch you're just like, "Wow, that's crazy.

我没想到会这样。’

I didn't expect that."

经常。

A lot.

这不过是在一堆文本上进行下一个词预测而已。

This is just next token prediction on a bunch of text.

这和我们很多直觉完全相反，我们总觉得要培养智能，得给它上课、慢慢带大、做各种复杂的事情，但事实完全不是这样。

It goes against so many of our intuitions for like, no, the way that you make an intelligence is you give it lessons, you bring it up, you do all this complicated stuff, and it's like, no, absolutely not.

我们做的就是这么简单的一件事。

We do this incredibly simple thing.

然而，它现在已经在数学领域取得突破了。

And yet, we're at the point where it's breaking new ground in mathematics.

我觉得如果你脑子里同时装着这两个想法，你会有两种反应。

I think if you hold these two thoughts in your head, there's two reactions you can have.

一种是否认，一种是好奇。否认就是觉得：这根本不可能。

There's denial and there's curiosity, where denial is just, "This is just not possible.

如果你只是预测下一个词，怎么可能比得上我们领域最顶尖的专家呢？

If all you're doing is predicting the next word, it's not possible that you could ever match the best specialists in our field."

而且我觉得，随着越来越多的实证证据表明事实就是这样，这种否认的立场会越来越站不住脚。

And I think that position just gets harder and harder as we get more empirical evidence that actually that is what's happening.

而好奇，大概就是我最终选择的那条路吧。

And curiosity is kind of the road I've ended up choosing, right?

就是觉得，这到底是怎么发生的？

Which is just like, well, how is this happening?

让我们打开看看。

Let's open up.

让我们看看里面。

Let's look inside.

我在Anthropic做的早期项目之一就是学SQL，一种编程语言，我花了好几年才掌握。

How could it be that one of the early projects I did at Anthropic is I learned SQL, a programming language, and it took me years to master.

然后作为一个小项目，我让Claude学SQL，结果它比我学得还好，而且我再也不可能在SQL上超过Claude了。

And then as a little side project, I got Claude to learn SQL, and it was better than me at it, and I'll never be better than Claude at SQL ever again.

而它只是通过这么简单的过程做到的。

And it took it just like this simple process.

这怎么可能呢？

How could this be?

我觉得这个问题里有很多值得深挖的地方，也有很多可以探索的，对我来说，每天试着去解开它，一层一层地剥开，真的非常吸引人。

And I think there's so much depth in that question and just so much to explore, and it's to me, so fascinating to just every day try to unpack that, try to sort of peel a few layers of the onions back.

好吧，很明显，这确实在发生。

Like, okay, well, clearly this is happening.

这么简单的目标，竟然能产生这么复杂的东西。

The simple objective is giving rise to this super complicated thing.

那好，我们就来弄清楚为什么。

Like, okay, let's figure out why.

我们来看看，哦，它是在学习这些通用概念。

Let's figure out, oh, it's learning these general concepts.

哦，它有这样的电路。

Oh, it has these circuits.

哦，好的。

Oh, okay.

下一个问题是什么？

What's the next question?

我们接下来不知道的是什么？

What's the next thing we don't know?

而且我可能最乐观的看法是，我觉得我们会学到很多关于智能本质的东西，对吧？

And my sort of maybe most optimistic take on this is that I think we're going to learn a lot about just the nature of intelligence itself, right?

就是关于如何从简单的目标中涌现出这种超级复杂的东西。

Of sort of how you get this super complex thing that emerges from simple objectives.

事实上，我认为我们就是另一个例子。

And in fact, I would argue that we're another example of this.

比如进化，目标很简单。

Like evolution, pretty simple objective.

你稍微折腾一下，保留那些能适应环境的生物，最后你得到了爵士乐和播客。

You sort of shake things around for a while and keep the organisms that manage to adapt to the environment, and at the end, you get jazz and podcasts.

这也挺疯狂的。

That's kind of crazy as well.

所以我觉得，通过研究这些东西，我们能学到一些根本性的真理。

And so I think there's a sense in which there's some fundamental truth that we can learn by studying these things.

哇。

Wow.

你这么说让我想到，现在有意思的是，AI正在帮助我们学习和进步，比如AI辅导之类的。

Something that comes up as you say that is what's interesting now is AI is helping us learn and progress, like AI tutoring and all these things.

有趣的是，它现在正在帮助我们加速对世界的理解。

Just it's interesting how it now is helping us accelerate our own understanding of the world.

我在想，你之前提到的我们学习的方式，AI也在学习，只是处理大量数据。

And I wonder if this point you made about how the way we learned was lessons, AI is learning, just processing a ton of data.

我在想，我们会不会最终改变教育体系，基于大语言模型训练的方式。

I wonder if that's how we end up-- I wonder if we change our education system based on what LLMs have been trained on.

希望我们不会仅仅为了让学生变成那样而改变教育，比如对他们说："嘿，这里有1000个词。"

Hopefully, we don't change it just so that we get students and we're like, "Hey, here's 1,000 words.

"预测下一个词。"

Predict the next one."

全都读一遍。

Just read it all.

全记在脑子里。

Keep it all in your head.

对。

That's right.

让我回到安全这个话题，因为我觉得有很多人担心事情的发展方向。我很好奇，作为看着这一切发生的人，在对齐组织内部，看着它演变，你乐观地认为我们能解决这个问题吗？

Let me come back to this kind of safety piece because I think there's a lot of people that are just worried about where things are going, and I'm curious, just as someone watching this stuff happen, being within the alignment org, watching it evolve, are you optimistic we'll figure it out?

你觉得我们能破解这个难题，让AI保持与我们的利益一致吗？

Do you think we'll crack this and keep AI aligned with our interests?

我的看法是，我个人是乐观的。

My take on this is I'm personally optimistic.

我认为现在有很多人在关注这件事，首先，我觉得这是一件好事。

I think this is something where right now a lot of people are paying attention to it, and first of all, I think that's kind of a great thing.

我觉得像这样的问题受到越来越多的关注，正是我们之前讨论过的。

I think as an issue like this rises in prominence, it's exactly what we were talking about earlier.

就像这项技术很强大。

It's like this technology is powerful.

它有好处。

It has benefits.

它也有风险。

It has risks.

我们得不断调整，把资源从收益那边挪到风险这边，随着进展来回调配。

We have to kind of adapt and shift resources from benefits to risks and back and forth as it progresses.

所以对我来说，我觉得我们一直在持续研究这个问题。

And so, to me, I feel like sort of we've continued to study it.

有意思的是，随着模型变得更好，它们在某些方面反而更容易研究了。

The thing that's interesting is that as the models get better, they get in some ways easier to study.

所以就有这种奇怪的情况：模型变好了，它们内部的世界模型更清晰了，我们能更好地理解它们的一些部分，但同时它们也变得更强大，更能绕过基本的安全措施。

And so there's been this sort of strange, the models get better, and so the world model inside them is clearer, and we can sort of understand some parts of them better, but also they get more capable, and so they're more able to sort of evade basic measures of security.

所以随着模型变好，就有这种拉锯战，我觉得基本上根据这种拉锯，你需要把或多或少的资源分配到监控的不同部分、更好地理解它们的电路等等。

And so there's this push and pull as models get better, and I think basically based on this push and pull, you want to allocate more or less resources to sort of various parts of monitoring, understanding their circuits better, things like that.

但这也是为什么我很高兴和你聊天，因为我觉得要取得进展，最重要的事情之一就是确保人们理解这些系统是什么样的、它们如何工作、如何制造，这样他们就能对现在和未来形成好的心智模型，然后决定如何塑造它，无论是参与这些对齐工作，还是作为公众去投票和决定这些问题。

But it's also why I'm kind of happy to chat with you because one of the main things that I think is important to make progress is just to make sure that people kind of understand both what kind of systems these are, how they work, how they're made, so they can form good mental models of the present, of the future, and so they can decide how to shape it, whether that's taking part in some of these alignment efforts or just informing them as the public that's going to have to vote and decide on much of these issues.

我不知道。

I don't know.

正因为这个原因，我对我们现在所处的这个时刻感到非常兴奋。

I'm pretty excited for this current moment we're in for that reason.

还有一件事我觉得大家误解了，就是这些实验室里的人从一开始就非常认真地对待这个问题。

And something I think people misunderstand is people working in these labs have taken this problem very seriously from the beginning.

就像我说的，Anthropic最初的使命就是AI安全和对齐。

Like I said, Anthropic's initial mission was AI safety and alignment.

这也是它成立公司的原因。

That was the reason that it became a company.

很多人是第一次听说这个，或者觉得这不是一个严肃的研究领域，觉得没人关注，但其实很多聪明人都非常在乎这件事，并且正在努力。

And a lot of people are hearing this for the first time or think that it's not a serious area of study and that people aren't focused on it, but many smart people care a lot about this and are working on it.

据我所知，很多好的进展正在发生。

And from what I can tell, a lot of good progress is happening.

你提到的另一个观点是模型智能增长得非常平稳，这里一直有个问题：我们会不会经历快速起飞还是缓慢起飞？

The other point you made about the models growing very smoothly in intelligence, there's been this question of are we going to have a fast takeoff or a slow takeoff?

一直有这种担忧，就是我们会达到某个点，比如递归自我改进，然后它就突然起飞，变得超级智能，然后——麻烦大了。

There's always this fear that we would hit a point, recursive self-improvement, for example, and it just takes off, superintelligent, where- Huge trouble.

我们完全失控了。

We have no control.

它已经脱离控制了。

It's out of the box.

但现在感觉我们确实是在缓慢起飞的情景里，因为就像你说的，模型在平稳地进步，我们也在过程中不断发现小问题、学习、进化，并逐步制定监管措施。

But it feels very much like we're on the slow takeoff scenario because, as you said, the models are very smoothly improving and where we are catching the little things along the way, learning, evolving, putting regulation around it.

所以，在所有设想过的情景和人们曾经的担忧中，我们现在走的这条路算是相当不错的一条了。

So, of the scenarios, this feels like a pretty good path we're on of the many scenarios people imagined and the fears they had.

随着这种进步速度的演变，关键就是要快速应对。

As this pace of progress kind of evolves, it's just a matter of reacting quickly.

所以我觉得，任何事——这就是为什么我认为让公众知情很重要，因为这样人们就会表达他们的担忧或意见。

So I think anything-- That's why I think having people be informed is great because then people will voice concerns or voice their opinions.

然后，是的，定期评估收益是什么、风险是什么，以及如何分别应对，这基本上就是我认为正确的推进方式。

And then, yeah, sort of regularly evaluating what the benefits are, what the risks are, and how we mitigate each is basically the way that feels like the right way forward.

而且，是的，我同意你的看法，感觉现在确实有动力去做这件事，这很棒。

And, yeah, I kind of agree with you that it feels like there's momentum to do this right now, which is great.

我想问你个事，完全不相干的。

Something I wanted to ask you, completely unrelated.

你对模拟假说怎么看？

Where are you on the simulation hypothesis?

你看着这些生命形式不断成长，你是更相信我们活在模拟里，不是在最底层，还是不是？

Being someone that watches all this life form growing, are you more convinced we're in a simulation and we're not at the base layer of all this, or are you or not?

哦，这个问题有意思。

Oh, this is a fun one.

其实我好久没想过这个了。

I actually haven't thought about this in a while.

对我来说，观察这些模型内部会引出别的东西，就是它们似乎总是落在一个基本结构上。

To me, there's something else that looking inside these models kind of brings out, which is this sort of almost like there's a fundamental structure that they keep falling on.

这跟你的问题有关，我保证。

This is related to your question, I promise.

让我慢慢说。

Let me get to it.

让我解释一下。

Let me explain.

我研究过很多Claude模型，还有一些其他模型，包括一些开源的工作。

I've studied a bunch of Claude models and also some different models, some open source work.

很多时候，当你观察它们学到的东西和内部结构时，会发现出奇地相似。

And then a lot of time when you look at what they learn and what the structure inside is, it's surprisingly similar.

这些模型是相隔几年、由不同的人用不同的数据训练出来的。

And this is for models that are trained years apart on different data by different people.

然后你看它们怎么解决这些问题，重叠的地方太多了。

And you look at how they solve these problems, and there's just so much overlap.

我的看法是，对于一些重要的问题，往往只有一种高效的做法，对吧？

And my take on why that is, is that for some problems that are important, there's just one efficient way to do things, right?

比如数学，有一种高效的方式是符合我们大脑结构的，我们都这么学，很少有人会用完全不同的方式，对吧？

There's one way to do math efficiently that fits the structure of our brain, and we all learn it, and very few people are doing it a crazy different way, right?

大家都用十进制。

Everybody works in base 10.

没多少人用六进制什么的。

Not many people in base six or whatever.

我觉得这让我想到，这些模型捕捉到了一些信息的基本结构，我们在某种程度上也捕捉到了，而且在更疯狂的模拟层面上，这可能是我们世界中的一种规律。

I think that to me makes me think that there's some fundamental structure of information that these models are picking up on, that in some ways we're picking up on, and that at maybe a crazier sort of simulation level that kind of is regularity in our world.

所以这让我觉得，如果你有能力模拟像我们世界这样的东西，实际上你会得出类似的情况。

And so that maybe makes me think that if you were to have the power to simulate something like our world, actually, you would land in sort of a similar situation.

事情最终会落在一个盆地里，它们只是最容易排列的方式。

There's just this basin of where things end up, and they're just the easiest way for things to arrange themselves.

所以从这个意义上说，也许吧。

And so in that sense, maybe.

我可以相信我们生活在模拟中。

I could believe that we're in a simulation.

但我不确定我完全相信这一点。

I think I'm not sure I fully buy it.

我不完全相信的另一个原因是，即使我认为这种结构适合模拟，也会有一个问题：为什么？

The other reason I don't fully buy it is that even though I think the structure lends itself to it, there'd be a question of why?

谁在运行这个，目的是什么？

Who's running this and for what?

这个我就不清楚了。

And that one I don't know about.

嗯，埃隆说过一句话让我印象深刻，他说如果你要运行一个模拟，那应该是一个历史上很有趣的时期，或者是你非常兴奋的时代。

Well, something Elon said that stuck with me about this, that if you were to run a simulation, it'd be an interesting time in history or something that you're really excited about.

而且确实，我们现在这个时代在历史上非常有意思。

And it tracks that this is a very interesting time in history.

对。

Yes.

好。

Okay.

说得有道理。

That's a fair point.

我们有太空数据中心、机器人和人工智能。

We're data centers in space and robots and artificial intelligence.

好的。

Okay.

我现在明白了。

I see it now.

最近有人跟我说，他们以前觉得“愿你生活在有趣的时代”是一句祝福，但现在他们不那么确定了。

Somebody recently told me they used to think that may you live in interesting times was like a blessing, and they're not so sure now.

他们说：“嗯，平淡的日子也挺好的。”

They're like, "Eh, the boring times are pretty good too."

天哪。

Oh, man.

而且还有一个Emmanuel在J空间里看着你的思考过程。

And then there's an Emmanuel out there watching your thinking process in your J space.

我的天。

Oh my God.

希望我没让他失望。

Hopefully, I'm making him proud.

他不会说，"哦，我得重新跑一遍模拟。

He's not like, "Oh, I got to rerun the simulation.

我们搞砸了这个。"

We messed up this one."

关掉它。

Shut it down.

不。

No.

天哪。

Oh, man.

好吧。

Okay.

最后随便问一个问题。

One last question randomly out of nowhere.

就是Swix，他负责AI工程，还主持Latent Space播客，人特别好。

So Swix, who runs AI Engineering, Latent Space podcast, awesome guy.

我问了他该问你什么问题。

I asked him what to ask you.

你一年前上过他的播客。

You were on his podcast a year ago.

然后完全从另一个角度出发。

And something completely from a different direction.

他想听听你对Jev的看法，这是一个最近推出的热门新模型，决策速度非常快。

He wanted to get your take on Jev, which is this hot new model that recently came out that is very fast making decisions.

有意思的是，OpenAI大约一周后也推出了自己的版本。

OpenAI, funny enough, just launched their own version like a week later.

你对Jev在快速决策状态模型领域的看法是什么？

Thoughts on Jev in the space of quick decision-based state models.

我对Jev的看法是，这类模型你给它们一些上下文，再给它们几个选项去选，它们就能很快帮你选一个。

So my take on Jev is these kind of models, you give them some context, and you give them a few options to choose between, and they just are very quick at choosing an option for you.

对我来说，听到这个不可能不联想到系统一思维和系统二思维，对吧？

To me, it's impossible not to hear about this and think of system one thinking versus system two or something, right?

所以，我研究这些模型时，很多是关于它们如何用Claude模型采样一个token。

So, like a lot of what I've studied with these models is how they sample one token with the Claude models.

所以，你可以把这些决策模型看作是很相似的。

And so, you can think of these decision models as pretty similar.

所以，拿一个LLM，然后采样一个token。

So, take an LLM and sample one token.

它们在重要方面有所不同，但在基本层面上，你可以说，哦，你不是让模型去采样和推理。

They're different in important ways, but at a fundamental level, you can be like, oh, you're not letting the model sample and reason.

这就像你的直觉。

It's just your gut feeling.

答案是A还是B？

Is the answer A or is the answer B?

对我来说，这似乎是一个普遍框架，有时候你只需要一个二元决策，或者从十个里面选一个，这显然都很有用，而且我猜这可能会成为智能系统的一部分，对吧？

To me, that seems like that general frame of sometimes you'll just want a binary decision or one out of 10 or whatever is just clearly both clearly useful, and something that I imagine is probably going to be a part of intelligent systems, right?

因为有些事情你需要思考很长时间。

Where for some stuff you need to think about a long time.

就像，我们明年的路线图是什么？

It's like, what's our roadmap for the next year?

你需要思考一会儿。

You need to think about for a while.

然后对于某些事情，比如给我们写评论的这个人，他是高兴还是不高兴？

And then for some stuff where it's like, is this person who wrote us a review happy or upset?

你对这种直觉判断感到满意。

You're happy with the sort of gut judgment.

所以是的，我觉得这类研究很酷，而且我觉得它也有点不同于——这是对我们见过的某些LLM的一种新颖看法。

So yeah, I think this kind of research is cool, and I think it's also kind of different from the-- It's a bit of a novel take on some of the LLMs we've seen.

所以是的，我很期待这类东西的发展。

So yeah, I'm pretty excited to see where this kind of stuff goes.

所以我听到的是，这很可能成为AI和LLM的一个新的基础组件。

So what I'm hearing here is this kind of going to be a new foundational component, most likely of AI and LLMs.

而系统一和系统二的概念基本上就是卡尼曼的那种——我身后其实就有那本书。

And the system one, system two stuff is basically Kahneman's kind of-- I have that book up there actually behind me.

简单来说，系统一就是你的杏仁核和边缘系统，做那种本能的快速判断，比如看到蛇就喊“有蛇！”

System one, just to keep it simple, is just your kind of your amygdala limbic system making just gut level quick decisions like, "Snake.

“哎呀，不好。”

Oh, no."

对。

Yes.

这种时候你根本不用多想。

Where you don't have to think long about it.

它的作用就是让你能轻松又快速地避开危险。

It's kind of to protect you from bad things really easily and quickly.

而系统二呢，是我们真正深入思考和处理问题的时候。

And system two is where we really think about stuff and process.

对，没错。

Yeah, exactly.

好嘞。

Sweet.

伊曼纽尔，我们还有什么没聊到的吗？

Emmanuel, is there anything else that we haven't covered?

你还有什么想说的或者想提的吗？

Anything else you want to say or mention?

在我们进入激动人心的快问快答环节之前，你有没有什么想再强调一下的？

Anything you want to double down on before we get to our very exciting lightning round?

我确实一直很好奇大家怎么看待那些还没解释清楚或者不太合理的地方。

I'm definitely curious for folks' takes always on some of the stuff that is still an export or didn't make sense.

还有很多可解释性实验，我们没时间一一分享。

There's a lot more interp experiments than we have the time to share.

模型身上有很多奇怪的小特点。

Models have a bunch of weird little things about them.

如果可以的话，我可能想在快问快答环节里聊聊这些奇怪的小特点。

Maybe I'll do a lightning round of weird little things if that's okay.

那就来吧。

Let's do it.

当然可以。

Absolutely.

好的。

Okay.

第一个奇怪的小发现。

Weird little thing one.

我们发现模型显然能回答问题。

We found that models can actually answer questions obviously.

它们能做数学题，或者你可以问它们，比如“香蕉是什么颜色？”

They can do math or you can ask them like, "What's the color of a banana?"

而且如果你深入观察，你会看到所有这些中间步骤。

And they'll actually, if you look inside you'll see all these intermediate steps.

但它们还有一个独立的机制，有点像直觉检查，而且完全不同。

But they have a separate mechanism which is kind of like a gut check and which is totally different.

这个机制就像在它们推理时，有一部分会标记出“嗯，有点不对劲。”

And which is just like as they reason, they'll have some part of them that will flag like, "Hmm, something's wrong."

所以，如果模型不知怎么地说4加5等于23之类的，而模型的一部分只是读到这个结果，并没有去做计算。

So if a model somehow says 4 plus 5 equals 23 or something, and there's a part of the model that just reads that and doesn't do the math.

但它就是觉得这不对，因为4加5肯定小于10。

But it's just like that can't be right because 4 plus 5, definitely under 10.

所以它们就学会了这些奇怪的直觉。

And so they learn these weird kind of hunches.

感觉就像是一种直觉。

They feel like kind of like a hunch.

另一个例子是，你可以告诉模型：“嘿，做这个任务，同时想想另一件事。”

Maybe another one is you can tell the model, "Hey, do this task and think about this other thing."

然后你可以用可解释性分析看看是否有中间表示对应到那另一件事，结果确实有。

And then you can use interp to see if there's intermediate representations that map to the other thing, and that works.

你就能看到它确实在思考那另一件事。

You sort of see that it does think about the other thing.

这对我来说真的很神秘，因为从训练的角度看，没有理由让模型必须学会这个。

This is a thing that's really mysterious to me because there's no kind of training reason why the model would have to learn that.

所以这算是另一种奇怪的行为。

So it's sort of like another one of these weird behaviors.

其实我可能就讲这两个例子吧。

And actually maybe I'll keep it at two.

我觉得这两个已经够说明问题了。

I think those two are good enough.

这工作真有意思。

Just what a fun job this is.

就是研究这东西，想搞清楚到底怎么回事。

Just studying this thing and trying to understand what the hell's going on.

它怎么变得这么聪明，它是怎么思考的？

How is it getting the smart, how does it think?

有多少-

How many-

确实

Very much

……

...

这个领域到底有多少人在做？

people are even in this space?

这个领域到底有多大？

Just how large is this field?

这个领域本身已经相当大了。

The field itself is getting pretty large.

我觉得大概有几百人，而不是几千人。

I would say maybe hundreds, not thousands.

不同的实验室里有不同的可解释性研究人员，学术界也有不少实验室在产出优秀的工作。

There's different interp people at the different labs, and then there's in academia quite a few labs putting out great work.

如果有人想进入这个领域，通常是从机器学习背景来的，还是有类似神经科学家的路径可以进入？

If someone were to get into this space, is it like an ML person usually or is there like a neuroscientist sort of track to get into this?

实际上，前神经科学家的数量相当惊人。

There's actually a surprising amount of ex-neuroscientists.

我觉得要进入这个领域，其实不需要太多先验知识，因为它太新了。

I do think in order to get in the field, it doesn't require a lot of prior knowledge because it's such a new field.

所以实际上，入门主要是自学和读一些关键论文。

So it's actually the on-ramp is mostly I would say self-studying and reading some of the key papers.

然后，其实我可能应该推荐一下。

And then there's actually maybe I should give a plug.

有个网站叫neuripdia.org，我记得是这个网址，它有点像开源的解释性研究平台。

There's actually websites like neuripdia.org I think is the URL, which are sort of like open source interp.

他们托管了很多模型，让你可以在上面做解释性实验。

So they host a bunch of models and let you sort of do interp experiments on them.

所以很多人就自己去做这些，我觉得有太多东西可以探索了。

And so people have just been doing that on their own, and I think that there's so much to explore.

如果你哪怕有一点点好奇，就应该去试试。

That if you're at all curious, you should just give it a try.

那太棒了。

That's awesome.

我们会在节目笔记里放上链接，以防那个网址不对。

We'll link to that in the show notes in case that's not the right URL.

好的。

Cool.

好了，Emmanuel，我们到了激动人心的快问快答环节。

Well, Emmanuel, with that we've reached our very exciting lightning round.

我有五个问题要问你。

I've got five questions for you.

准备好了吗？

Are you ready?

我随时都准备好了。

I am as ready as I'll ever be.

那我们开始吧。

Here we go.

你经常推荐给别人两三本书是什么？

What are two or three books that you find yourself recommending most to other people?

我最近读的一本是《智能简史》。

One I read recently and is "A Brief History of Intelligence."

这本书特别棒，它讲述了大脑的发展历程，以及我们进化过程中五个关键时刻，这些时刻让大脑系统一层层叠加起来。

It's this really cool book that walks through the history of basically brains and five key moments in our evolution that sort of bootstrapped these brain systems on top of one another.

这本书非常通俗易懂。

It's super accessible.

你不需要任何神经科学知识就能读懂。

You don't need any neuroscience knowledge or anything.

另一本是《你一生的故事》。

Another one is "Story of Your Life and Others."

这是特德·姜写的。

This is Ted Chiang.

这大概是我有史以来最喜欢的科幻小说了。

This is maybe my favorite sci-fi book of all time.

它是一本短篇小说集。

It's a bunch of short stories.

我觉得它做到了科幻小说的最高境界，就是只取一个奇怪的想法，比如如果这件事发生了会怎样？

It does what I think is the high art version of sci-fi where it's just take one weird idea like what if this happened?

然后围绕这个假设，写一个故事，描绘如果那真的发生了，世界会变成什么样。

And then write a story of just what would the world look like if that happened.

真的非常精彩。

It's truly excellent.

我想我身后的书架上也有这本书。

I think I have that in my bookshelf behind me too.

我也强烈推荐那本。

Also highly recommend that one.

最近最喜欢的电影或电视剧是什么？

Favorite recent movie or TV show?

好的。

Okay.

一部是大家都知道的，一部是比较冷门的。

One popular one, one deep cut.

《着迷》是我今年最喜欢的电影。

"Obsession" was my favorite movie of the year.

这是一部恐怖片，所以如果你讨厌恐怖片就别去看。

It's a horror movie so don't go if you hate horror movies.

但如果你——

But if you-

不了，谢谢。

No, thank you.

哦，不。即使你不是那种——如果你对恐怖片持开放态度，你应该去看，因为它真的太新颖了。

Oh, no. Even if you're not scream-- If you're open to a horror movie you should go because it was just so fresh.

你知道，我觉得如果你看过很多电影，有时候你会想：“哦，又是那种。”

You know how movies I feel like if you've seen a lot sometimes you're like, "Oh, it's another.

“这种我看过很多次了。”

I've seen this many times."

这部片子就是那么新颖，那么新鲜。

This one was just so new, so fresh.

我想说的一点是，它用了很多浪漫喜剧的套路，但反过来把它们变成了恐怖元素。

And I'll say one thing about it, which is that it takes a lot of rom-com tropes and subverts them to instead be spooky.

我觉得这种拍电影的方式真的很聪明。

And I think that that's just a brilliant way to make a movie.

然后我是法国人，所以今年有一部电视剧叫《Bref 2》播出了。

And then I'm French, so there's a TV show that came out this year called "Bref 2."

Bref 在法语里就是“简短”的意思。

Bref is French for brief.

Bref 在十年前、十五年前是法国的一部邪典经典，现在他们出了续集。

Brief used to be this cult classic in France 10, 15 years ago, and they came up with a sequel.

这部续集筹备了很多年，现在可以看了。

Many years in the making and you can watch it.

我有几个美国朋友都看过了。

I have a bunch of American friends that have watched it.

它有字幕，而且非常新颖，真的。

It's got subtitles, and it's just this incredibly, again, it's novel.

剪辑特别厉害。

The editing is crazy.

他们真的非常有创意，而且每集都很短。

They are just extremely creative, and it's very short.

所以我推荐这个。

So I would recommend that.

你最近发现的最喜欢的产品是什么？

Favorite product you recently discovered that you love?

我很喜欢我的苹果手表，但我不喜欢它每天都要充电。

I love my Apple Watch, and I hate that it charges or I have to charge every day.

所以我最近买了苹果手表Ultra，天哪，那真是太棒了。

So I bought the Apple Watch Ultra just recently, and man, that is so nice.

如果你曾经因为这么好的产品却不知道怎么给它配个好电池而烦恼，那么，事实证明，如果你把它做大一点，电池续航就会更好。

If you've ever been frustrated by this good product that somehow we can't figure out how to give a good battery to, well, it turns out if you make it bigger, the battery's better.

所以我对这个很满意。

So I've been happy with that.

你有没有一个经常在生活或工作中想起的人生座右铭？

Do you have a favorite life motto that you find yourself coming back to often in work or in life?

有一本非虚构类的书，可能是我最喜欢的一本。

There's one nonfiction book that really is probably my favorite ever.

书名是《四千周》。

It's "Four Thousand Weeks."

里面有很多我觉得很棒的内容。

It has a lot in it that I think is cool.

其中一个特别触动我的观点是，那种“等我终于……”的心态很常见，但其实挺有害的。

One of the cool things is this sort of thing that was really relevant to me, which is that the when I finally mentality is a common mentality, and it's kind of toxic.

这是什么意思呢？

So what does that mean?

就是说，等我终于把待办事项都清空了，我才能真正放松下来。

It's sort of this like, when I finally check off all the items on my to-do list, I'll finally be able to chill.

等我终于结婚了，我就会幸福了。

When I finally get married, I'll be happy.

等我终于升职了。

When I finally get a promotion.

然后这有点像你刚才说的冥想，对吧？

And then this is sort of, I mean, you were talking about meditation, right?

这几乎和冥想相反，因为归根结底，你的生活就是你在当下经历的所有时刻，你应该尽量活在当下，而不是把自己投射到一个看似更好的未来。

This is almost the opposite of meditation and that at the end of the day, your life is all of the moments you lived in the moment, and that there's this sense in which you should just try to always be in the moment and not projecting yourself towards a future that seems better.

可能如果你告诉自己，等我终于做了某件事，我就会开心。

Probably if you're telling yourself, when I finally do X, I'll be happy.

当你真的做了那件事，你并不会开心。

When you do X, you will not be happy.

有没有更好的说法，还是说就别这么想？

Is there a better phrase to use or is it just don't do that?

他们有没有说——

Do they say-

我不知道。

I don't know.

我觉得对我来说——

I think to me-

等我今天忙完的时候

When I'm through today

……积极的说法就是活在当下，对吧？

... the positive version is be in the moment, right?

就是珍惜当下。

It's just like I appreciate the moment.

我最近读了一篇博客文章，哦，我忘了。

There was a blog post recently I read that was, oh, I forget.

大意就是享受三明治，我觉得就是这个意思。

It was something like enjoy the sandwich I think is literally the sense.

不知道你有没有看过，就是有人说，对，你在吃三明治。

I don't know if you saw this or it was like somebody was like, yeah, you're eating a sandwich.

你已经吃过很多三明治了。

You've eaten many sandwiches.

就好好享受三明治吧。

Just enjoy the sandwich.

你在开团队会议。

You're in a team meeting.

对，你开过很多次了。

Yeah, you've been in many.

就沉浸其中，别老想着下一件事。

Be in it and sort of not looking towards the next thing.

我以前参加过一回冥想静修。

I did a meditation retreat once.

那是个十天的静默静修，吃饭也是安静的，没人说话。

It was a 10-day silent retreat where you eat silently, too, and no one's talking.

你就坐在那儿吃饭，我记得自己剥一个鸡蛋剥了五分钟，就注意每一片蛋壳，还有那股味道，要进入那种状态很难，但确实——

You just sit there and eat your food and I just remember peeling an egg for five minutes and just noticing every shell piece and just the smell of it all and it's hard to get into that mode, but it's-

……那是个很有意思的境界，在那儿一切都很有意思。

... such an interesting place to be where everything is just interesting.

哇——

Like wow-

是的

Yes

……这个鸡蛋，而不是嘴里随便塞点东西。

... this egg versus just like stuff in my mouth.

你左手刷着Slack，同时还在听播客。

While you're on Slack with your left hand and listening to a podcast at the same time.

天哪。

Oh, man.

好，最后一个问题。

Okay, final question.

你整天研究AI。

You study AI all day.

你观察AI是怎么思考的。

You're watching AI, how it thinks.

如果AI能看透你的大脑，它会发现什么让人惊讶的事？

If an AI were to look inside your brain, what would be surprising?

它会看到什么让人意想不到的东西？

What would it see that might surprise people?

那肯定是我最深、最黑暗的秘密。

I mean, surely my deepest, darkest secrets.

但那些是——

But what are the-

那我们就来聊聊吧。

Let's get into it.

那我们就来聊聊吧。

Let's get into it.

哎呀。

Oh, man.

我一直有点爱开玩笑，这既是我的生活方式，也是我应对事情的方式。

I've always been a bit of a jokester, or that's just both how I live and how I cope.

我就是觉得，即使在严肃的场合，你也应该尽量带点轻松的气氛。所以我觉得如果你看到我的大脑，你肯定会惊讶，因为基本上在任何时刻，我脑子里都在想一些好笑的事情。

I just think that even in serious situations, you should aim to kind of have some levity and so I think if you looked inside my brain, you'd definitely be surprised that basically at any given moment, I'm just focusing on some random thing that is funny.

我在开严肃的会议时，会突然想，哦，对，但你穿这个奇怪的东西是不是有点奇怪？

I'll be in a serious meeting and I'll be like, "Oh, yeah, but isn't it kind of random that you're whatever, wearing this random thing or something?"

就是我有这些侵入性的想法，平时都压着，但通常都是些烂笑话，有时候，就像你今天听到的，不幸的是，它们就冒出来了。

It's just like I have these intrusive thoughts that I keep out, but that are all usually just bad jokes and sometimes, as you've heard today, unfortunately, they come out.

所以为那些道歉，但那可能是最让人惊讶的地方。

So I'm sorry for those, but that's probably the most surprising thing.

我喜欢这个。

I love that.

Emmanuel，非常感谢你来做客。

Emmanuel, thank you so much for being here.

最后一个问题，听众怎么才能帮到你？

Final question, how can listeners be useful to you?

就是做一些解释工作，读读关于模型是怎么运作的。

Just do some interp, read about how the models work.

你提到了Dorker的帖子。

You mentioned Dorker's thread up.

我觉得那个帖子也很棒。

I think it was great as well.

我觉得这个领域，不管是可解释性还是AI和LLM，都太新了，所以大家参与进来、了解情况，肯定是有好处的。

I think this is such a new field, both interp and AI and LLMs, and I think it just benefits from people engaging and being informed.

所以多读点东西吧。

So read things.

如果你还想了解更多关于这些模型的工作原理，而我们还没搞明白的地方，你可以联系我。

If there's more that you would want to know about how these models work that we haven't figured out, you can reach out to me.

我在很多社交平台上都有账号。

I'm on a variety of social platforms.

你大概可以在X上找到我。

Probably you can reach out to me on X.

我们会在节目里放一些你们发布的视频链接，那些视频讲了很多我们今天聊的内容，比如J space之类的。

And we'll link to some of the videos that you guys have put out that describe a lot of the stuff we've been talking about, like J space and things like that.

你们团队做的那些视频真的很棒。

There's some really beautiful videos your team has made.

我不知道是谁做的，但真的很好看。

I don't know who does that, but they're really cool.

我们和设计团队合作，所以这些好看的视频功劳都是他们的。

We partner with our design team, so all credit goes to them for the cool videos.

太棒了。

Amazing.

好的。

All right.

好的，再次感谢你，Emmanuel。

Well, thank you again, Emmanuel.

非常感谢你来做这个。

Thank you so much for doing this.

这真的太棒了。

This was amazing.