339 | 内德·布洛克论意识是否需要生物学基础

所以，你乐于想象，当我们最终对现象意识做出解释时，它可能会是这样的形式：当我体验到成为某种事物的感觉时，这就是大脑中正在发生的事情。

So you're happy to imagine that when we do get an explanation of phenomenal consciousness, it might take the form, when I'm experiencing what it is like to be something, here is what is happening in the brain.

关于解决难题，这里有件事要说。

Well, here's the thing about solving the hard problem.

嗯。

Yeah.

我认为你无法提前断定解释会以何种形式出现，但你提到的这种形式无疑是一个候选方案。

I don't think you can say in advance what the form of the explanation will be, but the form you mentioned is certainly a candidate.

好的。

Okay.

但这就是我想表达的全部意思。

But that that that's all I'm getting at.

我的意思是，我完全同意。

Mean, I completely agree.

我们无法预知解决方案会是什么样子。

We can't say what the solution is gonna be.

我只是在想，什么样的解释才能被接受为解决方案。

I'm just wondering what would be acceptable as a solution.

嗯，如果不听到具体的解决方案，真的很难说。

Well, it's really hard to say without hearing the solution.

我相信，正如你所熟知的，物理学中许多问题的解决方案反而带来了更多问题，引发了更令人困惑的难题，也许这里也会如此。

I'm sure, as you are well aware, physics, a lot of solutions to problems have raised more issues than they've solved, and more puzzling issues than they've solved, and maybe this will do the same.

这完全正确。

That's absolutely true.

没错。

That's true.

但在物理学中，如果我对某个物理现象没有理论，我大致知道这类理论会是什么样子。

But in physics, if I don't have a theory for a certain physical phenomenon, I kind of know what such theories look like.

对吧？

Right?

比如，存在某种状态空间，存在某些动力学方程。

Like, oh, there's some space of states, there's some dynamical equations.

对于心脏问题，我真的不知道。

For the for the heart problem, I just I just don't know.

我的意思是，也许我

I mean, maybe I'm

不，你知道，我们在意识问题上比在物理学上更加茫然。

not You know, we're just We're just at much more at sea on consciousness than probably we ever have been about physics.

是的。

Yeah.

不，我完全

No, that I completely

一切都更加令人费解。

It's just much more everything is self puzzling.

对。

Right.

这非常自然地引出了我的下一个问题：你已经稍微提到过这一点，但你是物理主义者，所以你目前并不倾向于认为心脏问题太难，以至于我们需要扩展世界的本体论。

Which leads very nicely into my next query about you already mentioned this a little bit, but you're a physicalist, so you're not tempted, at least at the moment, by saying that the heart problem is so hard, we need to expand our ontology of the world.

我不认为扩展本体论会有帮助。

I don't see that expanding your ontology is gonna help.

是的。

Yeah.

你可以是二元论者，但这并不会带来任何实质进展，反而只是引入了宗教或神秘主义，并没有真正解决难题。

You you can be a dualist and it doesn't give you any it just introduces some religion or mystery and doesn't solve the hard problem at all.

泛心论也是如此。

And the same with panpsychism.

它只是用所谓的组合问题取代了心物问题。

It just replaces the heart problem with the so called combination problem.

一切事物都有一点意识，但你究竟该如何把它们组合成一个有意识的人类呢？

Everything is a little bit conscious, but how the hell do you put them together into a conscious human?

是的。

Yeah.

我看不出有任何解决方案。

I don't see any solution.

我看那里也没有任何解决方案。

I don't see any solution there.

这看起来就像是，你知道，我的同事们相信这些东西是有他们的理由的。

It just seems like, you know, I mean, you know, my colleagues who believe this stuff have reasons for believing it.

而且，确实有一些非常有趣的论点支持这一点。

And, know, there are some really interesting arguments about that.

但要说解决这个难题，我不这么认为。

But solving a hard problem, I don't think so.

但你之前提到过非还原物理主义，这可能是某种……是的。

But you you've talked about non reductive physicalism as something that might Yeah.

在这里发挥作用。

Come into play here.

所以，是的。

So Yeah.

出于其他原因，我今天一直在思考还原论。

For other reasons, I was just thinking about reductionism, today.

事实上，我觉得，嗯，你提到的菲利普·安德森，就在我们即将讨论的那篇论文里。

In fact, I think, yeah, I think that you mentioned Phil Anderson in your paper that we're gonna talk about.

哦，我提了吗？

Oh, do I?

可能是菲利普·安德森。

Phil Anderson maybe.

我不确定。

I'm not.

我可能把不同的事情搞混了。

I might be mixing up different things.

所以菲利普·安德森写过一篇著名的论文，说的是‘多即是不同’。

So Phil Anderson wrote this famous famous paper saying more is different.

但人们从来不去读那篇论文。

And people people never read the paper.

他们只是引用标题。

They just quote the title.

而且他在论文中对还原论持非常肯定的态度。

And and he's in the paper, he's super affirmative about reductionism.

他非常喜欢还原论。

He loves reductionism.

是的。

Yeah.

他只是不认为

He just doesn't think

当你关心更高层次时，还原论没什么用。

it's useful when you're when you care about the higher levels.

但你得稍微反还原论一点。

But you want to be anti reductionist a little bit.

实际上，我想成为还原论者。

Actually, I want to be reductionist.

好的，很好。

Okay, good.

当然，这取决于你所说的还原主义是什么意思。

Of course, it depends what you mean by reductionist.

当然。

Of course.

所以，你知道，非还原物理主义的通常含义，或者大多数人所想到的，通常是一种功能主义的形式。

So, you know, the usual meaning of non reductive physicalism, or the one that most people have in mind, is usually a form of functionalism.

这意味着对这些现象的正确描述处于更高的功能主义层面，即你指定系统的组织结构，这允许它有多种不同的实现方式。

It's that the right descriptions of these phenomena are at a higher functionalist level, where you specify the organization of a system, and that allows for many different realizations of it,

实现。

implementations.

这不是我的版本。

That's not my version.

我的版本是以物质为中心的版本。

My version is the meat centric version.

很好。

Good.

我指的是那篇著名的短篇小说《他们是由肉组成的》。

I'm referring to this famous short story called They're Made of Meat.

你读过这篇吗？

Have you read that?

没有，我没读过。

No, I haven't.

哦，这篇非常有趣。

Oh, it's very funny.

我发个链接给你。

I'll send you a link to it.

我们会把它放在节目笔记里。

We'll put it in the show notes.

是的，我总是在我的本科课堂上讲这个。

Yeah, I always do it in my undergraduate classes.

所以故事是从一群机器的视角展开的，你知道的，那些由其他机器制造的硅基生命体，而这些机器又是由更早的机器制造的，最初的起源早已湮没在历史中。

So it's from the point of view of a group of machine, you know, silicon beings, machines made by other machines that were themselves made by other machines on the ultimate origin is lost in history.

它们在宇宙中游历，发现其他有意识的生命。

And they go around the universe discovering other conscious beings.

然后它们发现了我们，并说：‘它们是肉做的。’

And then they discover us, and they say things like, they're made of meat.

然后其中一个说：‘可是，它们是怎么思考的呢？’

And then one of them says, well, but how do they think?

答案是：它们只是用肉来思考。

And the answer is they just use meat.

我必须读一下这个故事。

I have to read the story.

是的。

Yeah.

那它们是怎么交流的呢？

And then how do they communicate?

它们挥动自己的肉。

They flap their meat.

他们觉得有这些由肉构成的意识生物实在太可怕了，因此最好隐瞒这个信息，彻底忘记它们。

And they think it's so awful that there are these meat conscious beings that they just better suppress the information and forget about them altogether.

作为肉身确实有它的缺点。

It does have its downsides being made of meat.

我得

I gotta

说。

say.

确实如此。

It does.

是的。

Yes.

是的，确实如此。

Yes, it does.

但好吧，你提到了‘功能主义’这个词。

But okay, you've introduced the word function or functionalism.

所以，你想要稍微反驳一下这种观点。

So that is a perspective you want to sort of push back against a little bit.

是的，是的。

Yeah, yeah.

我认为，在当前对人工智能的思考中，许多极具影响力的人都是计算功能主义者。

I think what's happened in current thinking about artificial intelligence is that a lot of very influential people are computational functionists.

他们认为某些计算对于意识而言既是必要条件也是充分条件。

They think that certain computations are necessary and sufficient for consciousness.

我长期以来一直对此持怀疑态度。

And I have long been a doubter of this.

1997年，我发表了一篇题为《意识研究中的生物学与计算》的论文。

In 1997, I published a paper called Biology versus Computation in the Study of Consciousness.

实际上，那并不是一篇真正的论文，而是一篇BVS回复。

Actually, it wasn't a real paper, it was a BVS reply.

没错。

Yep.

但我一直坚持这个观点很久了，现在事情变得越来越紧迫，这真的很重要。

But I've been pushing this line for quite a while, but things are getting hot now, and it really matters now.

过去这主要是哲学家们在讨论，但现在有很多其他人都参与进来了。

It used to be that it was kind of just philosophers, but now it's a lot of other people.

正如你可能知道的，人工智能安全是一个重大问题，涉及机器是否具有感受能力、是否可能受到伤害，以及是否需要考虑它们的福祉。

And you know, as you may know, there's a big issue of AI safety, which is the extent to which machines have feelings and they can be damaged, and their welfare has to be taken into account.

由Anthropic公司开发的大语言模型Claude，现在已经允许其AI在对话过于不愉快时选择退出。

And one of the large language models, Claude, made by Anthropic, has now allowed its AI to opt out of a conversation if it's too unpleasant.

宁可安全一点，也不要冒险，我想是这样。

Better safe than sorry, I guess.

是的。

Yeah.

我觉得这就是他们的想法。

I think that's the idea.

所以，这个想法是，呃，抱歉。

So the idea of of well, sorry.

功能主义和计算功能主义之间有区别吗？

Is there a difference between a distinction between just functionalism and computational functionalism?

有。

Yes.

是的。

Yeah.

功能主义是一个更广泛的学说，涵盖了其他各种类型的功能和功能角色，基本上就是哲学家们所说的‘功能角色’，即状态之间的因果映射以及它们如何相互影响，诸如此类。

Functionalism is a much broader doctrine that encompasses a lot of other kinds of functions, functional roles, basically, what philosophers have talked to under the heading of functional roles, they're basically causal map of states and how they affect one another, that kind of thing.

但如今真正变得重要的是计算功能主义，因为人们心中真正关心的问题是：这些人工智能所进行的计算，或者未来某些人工智能的计算，是否决定了意识状态？

But the version that has really become important now is computational functionalism, because the question is in people's minds is, do the computations that these AIs make, or some future AIs determine conscious states?

或者像我认为的那样，是否存在某种生物学上的必要条件？

Or as I think, is there some kind of a biological necessary condition?

对，对。

Right, Right.

所以这个想法

So the idea

我应该说一下。

I should say it.

当我这么说时，我的意思是，我认为这种观点同样合理。

When I say I think that, I mean, I think that that is equally plausible.

我不是说我对这一点有很强的证据。

I don't mean I have really strong evidence for that.

所以，我认为计算功能主义的推销说法是：看，大脑显然在进行某种计算。

So the idea of the sales pitch, I guess, for computational functionalism would be, look, the the brain clearly computes some things.

你可以从某种层面上把与人类交流理解为：它接收输入，然后给出输出。

You can at some level think of how you communicate with a human being as it gets some input, it gives some output.

显然，这背后存在某种计算过程。

Clearly, there is a computation underlying that.

是的。

And Yes.

而计算功能主义的观点认为，这正是本质所在。

And the computational functionalist view is that is what it is.

没有什么额外的东西在发生。

There's nothing really extra going on.

是的。

Yeah.

是的。

Yeah.

是的。

Yeah.

所以我认为，如果你希望机器具有意识，你可能需要以某种特定方式实现这些计算。

So what I think is if you want the machine to be conscious, you may need a certain kind of implementation of those, computations.

是的，你在论文中对角色和实现者做了一个很好的区分。

Yeah, you have a nice distinction in the paper about roles versus realizers.

我知道在音频格式中解释术语总是很难，但你能不能试着解释一下，嗯，这个

I know that it's always hard to explain jargon words in an audio format, but why don't you give that a try, explain Yeah, the

意思是，角色是系统的抽象结构，以及什么导致了什么。

so the idea is that the role is the abstract organization of the system, and what causes what.

当你谈论计算系统时，计算是指系统所执行的操作，而实现者则是执行这些计算的实体。

When you're talking about computational system, it's the computations the thing does, and the realizer is what does those computations.

所以，举个著名的例子，一个简单的计算过程，比如加法或乘法。

So, famously, a computational process, just take a simple computational process like adding or multiplying.

它可以通过电子系统、使用继电器和开关的电气系统，或使用齿轮和滑轮的机械系统来实现。

It can be realized in an electronic system, or an electrical system with relays and switches, or a mechanical system with gears and pulleys.

这些是不同的实现方式，以不同方式执行相同的计算。

And these are different realizers that implement the computations in different ways.

当然，我们认为它们只是以不同方式执行相同的计算，但问题是，意识所特有的计算是否需要某种特定的实现形式？

And of course we think that they're just doing the same computation in a different way, but the question is, do the computations characteristic of consciousness require some particular form of realization?

是的。

And yeah.

所以这非常接近关于基质独立性或依赖性的讨论。

So so how does it this is very, very close to the discussion about substrate independence or dependence.

是的。

Yes.

我不是一个依赖基质的人。

Now I am not a substrate dependence person.

好的。

Okay.

我的同行者阿尼尔·塞思认为。

My fellow traveler, Aneel Seth Yeah.

他认为基质、材料、物质才是重要的。

Thinks the substrate, the material, the stuff is what's important.

我关注的是机制。

I focus on the mechanisms.

对。

Right.

例如，我们的神经放电、神经元涉及某些离子。

So, for example, our neuro firing, our neurons involve certain ions.

你知道的，钙、钾、氯离子等等。

You know, calcium, potassium, chloride, etcetera.

也许神经元可以由不同的基质和不同的离子构成。

And maybe neurons could be made out of a different substrate with different ions.

也许可以用硅来实现。

Maybe there's some silicon way of doing it.

我不知道。

I don't know.

我不是化学家，所以我不知道用不同的基质能组合出什么。

I'm not a chemist, so I don't know what could be put together using a different substrate.

但机制方面，从我无知的角度来看，也许可能存在具有相似机制但基质不同的东西。

But the mechanisms, I mean, from my ignorant point of view, maybe there could be something with similar mechanisms, but a different substrate.

我认为重要的是机制，而不是基质。

And I think it's the mechanisms that count, not the substrate.

很好。

Good.

好的。

Okay.

算盘使用与电子计算机不同的过程来加两个数。

So an abacus adds two numbers together using a different process than an electronic computer does.

是的，没错。

And you Yeah.

你会说这种差异可能很重要，但算盘是用铁还是木头做的对你来说并不重要。

You're gonna say that that difference might matter, but whether the abacus is made of iron or wood does not matter to you.

是的。

Yeah.

顺便说一下，有趣的是，那些不用算盘进行心算的人所犯的错误，与十进制计算不同，因为算盘运算涉及五，我不知道具体怎么运作，但它们常出现五的错误。

I should say, by the way, it's interesting that mental abacus calculations people do without the abacus make different errors than base 10 computations because abacus involve, I don't know quite how they work, but they involve fives, they make mistakes of five.

好的。

Okay.

非常有趣。

Very interesting.

所以，算盘的意象不同于数字意象，它不是数字的，而是十进制意象，关键错误通常发生在进位时。

So imagery, abacus imagery differs from digital, it's not digital, but decimal decimal imagery, where the key mistakes are often made in carrying.

嗯。

Yeah.

有趣。

Interesting.

我的意思是，这让我想到，当我们构建这些出色的大语言模型时，它们在模仿人类对话方面表现得非常出色，但即便它们是计算机，却在算术上表现得很差，对吧？

I mean, it reminds me of the fact that, when we built these wonderful large language models, and they do an amazingly good job of mimicking human conversation, but they they become bad at arithmetic even though they're computers, right?

因为显然它们的处理过程是不同的。

Because clearly the processes are different.

是的，事实上，它们在算术上仍然很差。

Yeah, in fact, they're still bad at arithmetic.

GPT-3 在三位数乘法上的准确率大约只有20%。

And GBT-three was only about, I think, 20% accurate on three digit multiplication.

你知道，原始报告展示了准确率水平，之后它们的准确率越来越高。

You know, the original report showed the accuracy levels, they've gotten more and more accurate.

现在，连接到互联网或接入计算器的模型会尝试将这些计算任务发送给其他类型的计算设备，但它们做得并不好，仍然会出错。

Now the models that are hooked up to the internet, or hook up to a calculator, try to send these calculations to some other kind of computing device, but they don't do it very well, and they still make mistakes.

有一个曾在互联网上流传的例子，我认为是GPT-4，在某种计算中，机器认为5.11比5.9大，因为11比9大，这种类似的推理方式。

One that was circulating on the internet for a while was, I think, GPT-four, there was a certain kind of calculation where the machine regarded 5.11 as larger than 5.9, because 11 is bigger than nine and that kind of similarity.

它们在国际象棋规则上也会犯错。

And they also make mistakes with the rules of chess.

一个经常被报道的现象是，当你让它们处理一些非常规的棋盘局面时，棋子会像非马子那样跳跃。

It's a fairly commonly reported thing that when you get them in unusual chessboard situations, pieces will jump that are not knights.

好的。

Okay.

这里的基本观点是加里·马库斯多年前提出的，实际上，史蒂夫·平克才是第一个明确提出这一点的人：它们没有真正的规则。

And the basic point there is one that Gary Marcus made years and years ago, which is they don't have Actually, Pinker, Steve Pinker was the first person to really make this point, which is they don't have rules.

它们可以阅读规则，也能告诉你规则，但它们的基本计算模式并不基于规则。

They could read the rules, and they can tell you the rules, but their fundamental mode of computation is not based on the rules.

尽管它们知道算术规则，却并不使用这些规则。

While whereas they know the rules of arithmetic, they don't use them.

所以，我认为，过去一年与赛斯的交谈，让我从僵化的计算功能主义迷梦中清醒了过来。

So I think that it was, talking with an Seth over the last year, shook me out of my dogmatic computational functionalist slumbers.

我原本完全支持这一点，但现在我更好地理解了它存在的问题。

I I would have been all in favor of it, but I think I understand the problems with it better now.

但我仍然有些担忧。

But so I still have a little bit of a worry.

我正努力确保自己能够拒绝计算功能主义，同时仍保持物理主义立场。

I'm trying to I'm trying to make sure that I can reject computational functionalism but still be a physicalist.

而我的担忧是，我难道不能宽泛地将每一个过程都视为某种计算吗？

And the worry would be, can't I just be expansive and think of literally every process as some kind of computation?

如果是这样，那么任何物理过程都是计算。

And if so, if any physical is computational.

所以你指的是你在邮件中提到的物理版丘奇-图灵论题。

So you're thinking of the physical Church Turing thesis that you mentioned in an email.

也许我该说说什么是丘奇-图灵论题，是的，

Maybe, should I say something about what the Church Turing thesis Yeah,

没错，请讲。

exactly, please.

是的，好的，丘奇-图灵论题认为，机械可计算的函数就是图灵可计算的，反之亦然。

Yeah, okay, so the Church Turing thesis is the thesis that a mechanically computable function is Turing computable, and vice versa.

反之亦然这一点很明显。

The vice versa is pretty obvious.

首先，我应该说，机械可计算性是一个直观的概念。

Well, should say first, mechanical computability is an intuitive notion.

是的，我们并不清楚它到底意味着什么。

Yeah, we don't know what it means.

因此，丘奇-图灵论题是无法被证明的。

So there could be no proof of the Church Turing thesis.

这就是为什么它被称为一个论题。

That's why it's a thesis.

是的。

Yeah.

所以，如果是图灵可计算的，那么它就是机械可计算的。

So if it's Turing computable, then it's mechanical.

这很显然，因为图灵机所做的每一件事都是机械性的。

Well, is just obvious because everything a Turing machine does is a mechanical thing.

你知道，它在磁带上写入，移动磁带，擦除磁带上的内容，就是这类操作。

Know, it writes on the tape, it moves the tape, it erases from the tape, you know, that kind of stuff.

但如果它是机械的，那它就是图灵可计算的——这一点一直引发争议，人们对于这是否真的成立持保守态度。

But if it's mechanical, then it's Turing computable that has been something that has raised, people are conservative about whether that's really true.

我的观点是，这实际上是通过约定成立的，因为每当有人提出反例时，大家都会说：‘这根本不是真正的反例’，然后他们就会细化计算的定义来排除它。

My view is that that is true by stipulation, because whenever anybody comes up with a counterexample, everybody says, Oh, that's not really a counterexample, and then they refine the notion of computation to rule it out.

最著名的例子实际上是图灵提到的计算。

The most notable case is computation mentioned actually by Turing.

由真正随机的过程进行的计算，比如原子衰变。

Computation by a truly random process, like atomic decay.

你可以设置一个像盖革计数器这样的装置，它能计算出一系列数字，从某种意义上说，它在计算一个函数，而这是图灵机无法计算的。

You can set that up so that like a Geiger counter or something, it computes a series of, an infinite series of numbers, in some sense it's computing a function, and that can't be computed by a Turing machine.

所以人们会说：‘这不符合我们所说的机械性。’

So what people say is, Oh, that's not mechanical in the sense in which we meant it.

或者，这不算是我们所指的计算，因为它不可重现。

Or, It's not a computation in the sense in which we meant it because it's not reproducible.

所以他们修正了计算的定义以排除这种情况。

So they refine the notion of computation to rule it out.

同样地，也可以从非随机的计算角度来讨论，比如一条河流计算着河岸侵蚀的速率。

And a similar point could be made by the sense of computation, which is not random, in which a river computes the rate of erosion of its banks.

没错，正是如此。

Exactly, right.

这也不是图灵可计算的，或者至少可以说它不是图灵可计算的。

And that's not Turing computable either, or at least arguably it's not Turing computable.

你可以近似模拟它，但这里涉及一个关于实数及其数值的问题。

You could approximate it, but the There's a whole issue about the real numbers and real numbered values of things.

总之，我认为丘奇-图灵论题本质上是一种规定。

Anyway, so I regard the Church Turing thesis as really a stipulate.

当然，如果它是机械的，那么它就是图灵可计算的——这算是一种规定，但这是一个合理的规定。

Certainly that if it's mechanical, it's Turing computable as a bit of a stipulation, but it's a reasonable stipulation.

我的意思是，好吧，我们所讨论的过程是机械的，它们所计算的函数，如果我们忽略实数值的话——但你可能无法忽略，因为大脑中的一切都是粒子。

Mean, so, but, okay, so, the processes we're talking about are mechanical, and the functions they are computing, ignoring real numbered values, which probably you can't ignore, because all the things in the brain are particles.

所以可能存在一台图灵机来计算这个函数。

So there can be a Turing machine that computes that function.

但我想要提出的问题是，这种计算本身是否在以下意义上是模拟的。

But the problem I would raise is whether that computation is itself analog in the following sense.

如果你认为一个心理过程是模拟的，那是因为对该过程的计算不必保留其心理特性。

You can regard a mental process as analog if a computation of that process need not preserve the mental properties.

好的。

Okay.

推理可以说是一个非模拟的过程，因为如果你有一台机器执行了推理所涉及的计算，那么至少可以说它确实在进行推理。

Reasoning is arguably a non analog process, because if you have a machine that computes the computations, that does the computations involved in reasoning, at least it's a good case that it is reasoning.

但意识是另一回事，在意识这个问题上，它就像重力或一场暴雨。

But consciousness is another matter, and there consciousness is like gravity or a rainstorm.

暴雨的计算模拟并不会变湿。

A computational simulation of a rainstorm isn't wet.

对重力的计算模拟并不会产生任何重力，我的意思是，进行计算的物体具有质量，因此它们自身会产生引力，但计算本身并不会决定任何引力吸引。

A computational simulation of gravity doesn't produce any gravity, I mean, the objects that do the computation have mass, and so they will themselves have gravity, but the computation doesn't itself determine any gravitational attraction.

所以问题在于，意识是否也是如此。

So the question is whether consciousness is like that.

我认为我们真的不知道。

And I think we we just don't know.

因此，丘奇-图灵论题对我们没有帮助。

So the Church Turing thesis doesn't help us.

我多次听过这个例子：模拟飓风并不会让你变湿，也不会产生湿气。

I've heard many times this example of simulating hurricane does not make you wet or does not create wetness.

但我担心这里存在某种语言上的巧妙手法。

But it's a I worry that there's some linguistic sleight of hand going on here.

如果模拟中包含一个人在模拟的雷暴里，那么在准确的模拟中，这个人会说他们被淋湿了。

If if part of the simulation was a person in the simulated thunderstorm, they would that person would say that they were getting wet if they were an accurate simulation.

所以

So

我不该使用‘模拟’这个词。

I shouldn't have used the word simulation.

真正的问题在于，这种计算的实现是否必须保留其心理特性。

What it's really about is whether an implementation of that computation has to preserve the mental properties of it.

所以，我所说的不是模拟，而是指在计算硬件中对它的实现。

So an implement, I'm really not talking about a simulation, I'm talking about an implementation of it in computational hardware.

举个例子，你可以想想冻结这个过程。

And, you know, for many processes you might name, take for example freezing.

冻结是一种液体形成晶体固体的过程。

Can, you know, freezing is a process by which a liquid forms a crystalline solid.

你可以实现这些计算，而无需实际发生冻结。

You can implement those computations without any actual freezing.

因此，许多——如果不是大多数——过程都是这样的。

So, many, if not most processes are like that.

也就是说，这才是我真正应该定义的方式；我本不该用‘模拟’这个词。

That is an So that's the way I should really define I didn't mean to use the word simulation.

我无论如何都会用这个词，所以这没什么用。

I would have used it anyway, so it's not useful.

嗯。

Yeah.

所以如果我，嗯。

So if I yeah.

所以真正的问题在于实现。

So So the real the real issue is about implementation.

嗯。

Yeah.

好的。

Okay.

很好。

Good.

所以，我的意思是，你的论点之一是，当我试图理解阿尼尔·塞思的观点时，读你的论文时，我有点认同他的说法。

And so, I mean, is is part of your thesis that, well, when I was trying to understand what Anil Seth was saying, and I I it it vibed with me a little bit when I was reading your paper.

对于如何对待人类，有一种黑箱视角。

There's a black box view of how one deals with human beings.

正如我所说，输入然后输出。

Like I said, input and then output.

但同时也有一种湿润的、绿色的生物学视角，其中涉及大量在此期间进行的过程。

But then there's also, like, a a wet green biological view where there's a lot of processes going into it, in the in the meantime.

我想说的是，你所表达的意思是，这些过程可能对我们所认为的意识体验至关重要。

And I I wanna say that what you're saying is that those processes maybe matter to what we think of as conscious experience.

是的。

Yeah.

是的。

Yeah.

正是如此。

Exactly.

如果这是真的，那我——正如我所说，过去一年我的眼睛好像被蒙蔽的薄膜脱落了，我非常接受这种可能性。

And so if that's true and I'm and actually, you know, my the the scales have fallen from my eyes in the last year, like I said, and I'm very open to this possibility.

不，你不是，因为你是个幻象论者。

No, you're not because you're an illusionist.

不。

No.

不。

No.

看，这样很好。

See, this good.

现在我们说到点子上了，因为我认为，意识体验是对发生在纯粹物理层面的事物的一种高阶、涌现性的描述方式。

Now we're getting now we're making money here because, I I think that conscious experiences are emergent higher level ways of talking about things that happen at a physical, purely physical level.

但我承认，这种‘涌现’所指的内容取决于潜意识的、亚计算层面的因素，而不仅仅是输入和输出。

But I'm open to the fact that that what the emergence refers to depends on subconscious, subcomputational things, not just the input and the output.

我的意思是

I mean

所以我明白了。

So I see.

所以，有一种版本的这种观点，你是可以接受的。

So so there's a version of this that you can accept.

因为，你知道，我整个观点是……

Because, you know, my whole thing

即使作为一个幻觉主义者，没错。

Even even as an illusionist Exactly.

因为我的整个观点是熵与时间之箭。

Because my whole thing is entropy in the arrow of time.

而且我真的很难过，不记得是谁说的了，但有人在互联网上说过：大型语言模型不会体验时间的流逝。

And and I I I really feel bad that I don't remember who said it, but someone on the Internet said, LLMs do not experience the passage of time.

是的。

Yeah.

我认为这至关重要，但因为我们的细胞确实体验着时间的流逝，而且……

And I think that's crucially important, but because our cells do experience the passage of time, and

它们确实如此。

They sure do.

如果我们认为这与意识体验有关，我们会感到震惊。

We would be shocked if that had something to do with conscious experience.

是的。

Yeah.

这是个好观点。

That's a good point.

是的。

Yeah.

不。

No.

不。

No.

意识体验本质上是时间性的。

Conscious experience is intrinsically temporal.

对。

Right.

是的，是的，是的，我同意这一点。

Yeah, yeah, yeah, I agree with that.

那么我们对它了解多少呢？

So what do we know about it?

我的意思是，你在论文中区分了电化学过程和单纯的电子过程。

I mean, you make a point in the paper distinguishing between electrochemical processes versus merely electronic processes.

是的，这其实只是一种推测，但我们的大脑将纯粹的化学信号转化为电信号，再转回化学信号，也就是细胞间的神经递质、细胞内的电信号，这一点非常引人注目。正如我在论文中提到的，早期关于突触的理论认为它是电性质的。

Yeah, so yeah, that's really just a speculation, but it is kind of remarkable that the way our brains work translate purely chemical signals, into electrical signals, then back to chemical signals, you know, to neurotransmitters between cells, electrical within the cell, and you know, as I mentioned in the paper, the early theories of the synapse were electrical.

是的。

Yeah.

结果发现，似乎有一些证据表明，纯粹的电性神经系统表现并不好。

And it turned out, it looks like there's some evidence that a purely electrical nervous system didn't do very well.

我尝试列举了多种电化学处理可能更优越的方式，并得出结论：这仍然是一个谜。

And I tried mentioning a number of different ways in which electrochemical processing might be superior, and conclude that it's kind of a mystery.

但我们很幸运拥有它，因为这或许正是导致意识的原因。

But we're lucky we have it because maybe that's what led to consciousness.

所以当你说到——

So when you say-

我也应该说——呃，抱歉，你先说。

I should say also- Yeah, sorry, go

请说。

ahead.

抱歉，我只是想说，当你提到‘表现不太好’，你是从进化角度来说的吗？

Sorry, was just going say, when you say didn't do very well, you mean as a matter of evolutionary

是的。

Yeah.

是所谓的栉水母，至少在某一阶段拥有纯粹的电传导神经系统。

Yeah, that's the so called ctenophores that at least at one stage had a purely electrical nervous system.

而它们并没有演化出更复杂的动物。

And they didn't lead to more complex animals.

它们算是进化上的死胡同。

They were kind of an evolutionary dead end.

而电化学通路确实催生了更复杂的动物，包括我们人类。

Whereas the electrochemical pathway did generate much more complex animals, including us.

这是进化生物学家普遍知晓的事实，还是我们最近才发现的？

Is that a well known fact among evolutionary biologists, or is it something that we've noticed?

嗯，

Well,

好的。

okay.

所以，正如我在论文中所说，直到2022年，人们普遍认为海绵是最早的动物。

So as I said in the paper, up to 2022, it was thought that sponges were the first animals, or widely thought that.

但在2023年，一些研究结果表明，栉水母才是最早的动物，并且它们在染色体层面与所有后续动物存在重要差异。

Then in 2023, there were some results suggesting that actually ctenophores were the first, and that they differ from all subsequent animals in an important chromosomal way.

现在我听说——我还没读过那篇论文——有一些新证据可能与此相矛盾。

Now I'm told, I haven't read the paper yet, that there's some new evidence that maybe goes against some of that.

所以我认为目前的情况尚不确定，我不确定他们是否都会接受这一观点。

So I think the situation is in a state of flux, and I don't know if they all would accept that.

我认为，即使栉水母不是最早的动物，我猜人们很可能普遍认为它们是一个进化上的死胡同。

I think even if the tenophores weren't first, I think it would I'm guessing it would be probably pretty widely agreed that they were a dead end.

好的。

Okay.

但我也不确定。

But I I don't know for sure.

那些科学家总是因为新证据而改变他们的想法。

Those scientists always changing their mind about things because of new evidence.

是啊。

Yeah.

该死的科学家们。

Damn scientists.

我们哲学家从不这样。

We philosophers never do that.

所以我在想，如果我说你的观点或核心信息之一是：意识可以以有趣的方式依赖于无意识，依赖于我们并不知晓的事物，你觉得这有多准确？

So I'm wondering how accurate you would you would count it if I said that part of your lesson or your message is that consciousness can depend in interesting ways on subconsciousness, on things that we're not.

嗯。

Yeah.

当然。

Absolutely.

是的。

Yes.

没错。

Exactly.

嗯。

Yeah.

有是的。

Are there Yeah.

我想说什么来着？

What do I want to say?

我们能有潜意识的体验吗？

Are there can we have subconscious experiences?

这是真的吗？

Is that a thing?

当然，我曾主张，大脑皮层中可能存在某些完全与外界隔绝的体验。

Well, of course, I have advocated that there might be experiences in an isolated part of the cortex that are completely cut off from access.

这比我的其他推测更遥远，但我认为这是可能的。

It's a more remote speculation than my other speculations, but I think it's conceivable.

你确实提到了A

You do mention A

很多人觉得这不可能。

lot of people feel it's not conceivable.

好的。

Okay.

你确实提到了一种可能性，即可能存在类似被压抑的记忆，会影响我们的感受，影响我们的现象意识。

You do mention the possibility that there would be something like a repressed memory that could affect our feelings, that could affect our phenomenal consciousness.

是的，被压抑的情感，我试图从弗洛伊德的角度来说明，现象意识可以在没有元意识的情况下存在。

Yeah, well, repressed feelings, I was trying to illustrate the possibility of phenomenal consciousness without excess consciousness from a Freudian point of view.

我指出，弗洛伊德关于压抑的理论是指在获取层面的压抑。

And I pointed out that the Freudian picture of repression is repression in the access sense.

所以我设想的情况是，你经历了一段极其可怕的事件，并将其压抑，但这段经历仍具有非常鲜明的现象品质，也许即使被压抑，它依然保有这些现象品质，只是你无法接触到它们。

So the case I imagined is you had a very terrible experience, and you repress it, but it had very vivid phenomenal qualities to it, and maybe when it's repressed it still has those phenomenal qualities, it's just that you don't access them.

这就是我的想法。

So that was the thought.

而布弗拉特斯认为，由于缺乏获取途径，这属于无意识。

And the Buphrates regard that as unconscious because of the lack of access.

很好。

Good.

是的。

Yeah.

好的。

Okay.

他们肯定会受益于这一区分。

They would have benefited from this distinction, definitely.

好的。

Okay.

所以，是的，我有个问题。

So, yeah, I I had a question.

每个月做播客时，我都会做一期‘问我任何问题’，观众的问题会蜂拥而至，我会尽量回答。

Every every month for the podcast, I do ask me anything, episode where just questions pour in and I try to answer them.

上一次我收到了一个特别好的问题。

And I got a really good one last time.

这听起来真不容易。

That sounds really tough.

我也是

I get

由我来挑选回答哪些问题。

to pick which ones I answer.

这使得整个过程变得可以接受。

That's what makes it, palatable.

所以

So

好的。

Okay.

嗯。

Yeah.

其中一个是，我们如何判断在什么情况下AI会有意识的标准，做得怎么样？

One of them was, like, how well have we done on deciding the criteria for saying when an AI will be conscious?

是的。

Yeah.

我说过

And I said

我觉得我们

I think we're

做得不太好。

not that well.

嗯。

Yeah.

我不知道。

I don't know.

没那么好。

Not that well.

我同意这一点。

I agree with that.

我们真的才刚刚起步。

We're we're really we're really at first base.

我认为最有希望的建议是许多人提出的观点：如果能训练出一个没有基于人类关于其第一人称视角的表述来训练的AI，但它依然表达了第一人称视角，那将比我们现在的情况更有说服力，远更有说服力。

I think the most promising suggestion is one that a lot of people have made, which is that if you could make an AI that isn't trained on people saying things about their first person point of view, and it nonetheless expressed a first person point of view, that would be more convincing than what we have now, way more convincing.

所以意思是，这些语言模型是基于大量人类生成的内容进行训练的，其中包含了各种第一人称视角。

So the idea is, you know, these LMs are trained on a ton of human generated stuff that involves all kinds of first person point of view.

你给它一些书籍，里面就会有很多第一人称视角的内容。

You you give it a number of books, and you're going to have a lot of first person point of view.

也许如果你只用，比如说，《大英百科全书》之类的东西来训练它，嗯。

Maybe if you trained it just on, I don't know, the Encyclopedia Britannica or something or Yeah.

如果你成功地从训练数据中剔除了所有第一人称视角的内容，而它仍然发展出了第一人称视角，我想这会具有一定的说服力。

And if you successfully eliminated everything in the training data that had a first person point of view and it nonetheless developed one, you know, that would some convincing force, I think.

这很有趣。

That's interesting.

我的意思是，但显然，实现方式与我们的仍然大不相同。

Mean, but certainly still the realizers are still wildly different than for our

实现方式截然不同，但这或许表明，这些不同的实现方式可能确实实现了某种体验，或至少是第一人称视角。

Wildly different, but that would be some indication that maybe those different realizers do realize some kind of experience or at least a first person point of view.

很好。

Good.

所以，这与经典的图灵测试有区别，图灵测试只关注输入和输出。

So there's a distinction between like the classic Turing test, really only focused on inputs and outputs.

对吧？

Right?

比如，如果输出听起来像人，我们就会称它为意识。

Like, if the output sounded human, then we're gonna call it consciousness.

而到目前为止，我认为我们已经足够成熟，知道这还不够。

And then by now, we're I think we're mostly sophisticated enough to say that's not quite enough.

现在没人再谈论图灵测试了。

Nobody talks about the Turing test anymore.

对。

Right.

真有趣，你知道吗，仅仅三年而已。

It's really funny how, you know, just in three years Yeah.

或者四年吧，它已经完全消失了。

It's or four years, I guess, it's completely disappeared.

我的‘块头’例子已经彻底驳斥了图灵测试。

I mean, I think my Blockhead example just conclusively refuted the Turing test.

是吗？

Is that?

你知道那个例子吗？

You know that example?

不知道。

No.

哦，这是一个通过暴力方式通过图灵测试的机器的例子。

Oh, it's an example of a brute force Turing test passing machine.

意思是，就像井字棋可以通过树结构完全解决一样，你让对方先走，然后你为每一步 plausible 的走法都预先设定好。

The idea is, just as tic tac toe is just completely solvable through a tree structure, where you put in, you let the other person go first, and then you put in a plausible move.

人类玩家为每一步可能的走法都预先输入，然后对于裁判的每一步后续动作，人类程序员又再输入下一步，这就是人类程序员做的。

The human puts in a plausible move in the program for each of those, and then for each move that the judge makes next, the human puts in another one, that is the human programmer.

你构建一个树状结构，其中程序员为每一步都预先填好了应对方案。

And you make a tree where the programmer has put in every other move.

你知道，这些应对方案不一定要是最佳的，但无论如何，对于每一步，你都可以声称：‘我把这个放进去了。’

You know, they don't have to be good ones, but anyway, for every move, you can say, I put that in there.

这个想法是，对话也可以这样处理。

The idea is, it's like that for a conversation.

意思是，你决定图灵测试要持续多久，比如一小时。

The idea is that you have, You decide how long the Turing test is gonna go, like an hour.

你计算出所有可输入的字符串可能有多长，然后把这些都列出来，也就是说，这个规模会非常庞大。

You compute the typeable strings that can be, how long they can be, and you pull it I mean, would be very large.

斯图尔特·切弗实际上真的做过这个。

Stuart Cheever actually did it.

斯图尔特·切弗是哈佛大学的一位计算机科学家，他发表了一篇哲学文章，计算了这种系统会有多大，以及需要多长时间——答案是，规模真的巨大。

Stuart Cheever is a computer scientist at Harvard, and he published a philosophy article with a calculation of how big it would be and how long you know, the answer is really big.

但关键是，这在概念上是可能的，而且它在每一个回应上都能表现得和人一样好，因为这些回应都是由人预先设定好的。

But the point is, it's conceptually possible, and it would do as well as a person on every, because the people have put those number, those points in.

而且，对于每一个巧妙的回答，总有人能说：‘嗯，我也想到过这个。’

And the thing about it is, is that for every clever response, some person can say, Yeah, I thought of that.

嗯，我非常喜欢这个例子，因为人们可能会忍不住想：那为什么它不能算有意识呢？

Well, I I like this example a lot because, I mean, one might be tempted to say, well, why wouldn't that be conscious?

它只是一个巨大的查找表，但你怎么知道它没有意识呢？

It's just a big lookup table, but how do you know it's not conscious?

比如，也许宇宙就是一个巨大的查找表。

Like, maybe the universe is just a big lookup table.

但这表明了我们至少在非正式层面将意识与生物有机体中的某些特定活动联系在一起，这种方式是

But it points to the fact that we do, at least informally, associate consciousness with kind of some specific set of goings on in a biological organism in a way that it's

至少我们排除了负面情况。

not At least at least we rule out negative.

所以，我用这个例子想表达的是，也许我们不知道哪些内部活动导致了意识，但仅仅是一个查找表的东西，它并没有意识。

The the negative so the point I was making with that example was maybe we don't know what internal goings on are responsible for consciousness, but something that just is a lookup table, that's not conscious.

所以，它一定不仅仅是输入输出。

So it must be more than just the input output.

不仅仅是这些。

More more than that.

是的。

Yeah.

是的。

Yeah.

我的意思是，你对角色和实现者等的区分，能帮助我们稍微解答一下AI何时会具有意识这个问题吗？

So, I mean, does your distinction between roles and realizers, etcetera, help us even a little bit in getting an answer to when the AIs will be conscious?

这个问题正迫在眉睫，非常重要。

Like, this is an important question coming up on us.

是的。

Yeah.

至少它让我们能够明确地提出这个问题。

So well, at least it allows us to formulate the problem.

没错。

Yeah.

这总归是一个进步。

Which is always a a step up.

我的想法是，在判断任何外星生命是否具有意识时，你别无选择，只能从我们自身出发进行推断。

And, you know, I mean, the way I like to think about it is, in deciding whether any alien being is conscious, you really have no choice but to extrapolate from us.

你可以根据我们的计算特性，或者根据我们的亚计算特性进行推断，但我们就是不知道该选哪一个。

And you can extrapolate on the basis of our computational properties, or on the basis of our subcomputational properties, and we just don't know which.

所以我认为，这种区分对于明确这一点很重要。

So I think the distinction is important for formulating that point.

我们不知道到底该选哪一个，也没有理由偏爱其中任何一个。

We don't know which, we don't have any reason to favor one over the other.

所以如果让我来分配概率，我会给每个选项各50%。

So if I had to assign a probability, it would be 50% for each.

你知道道德哲学家们是否已经开始讨论我们该在什么时候对AI友善了吗？

Do you know if moral philosophers have started to weigh in on when we should be nice to the AIs?

哦，是的。

Oh, yeah.

现在这方面的文献已经非常多了。

Oh, there's a huge amount of literature on this now.

两周前我去伯克利参加了一个关于AI安全的会议，那里有很多人在思考这个问题，所有公司也都有人在研究。

I went to an AI safety, which is that, it's about that, in Berkeley two weeks ago, and there were a lot of people thinking about it, and all the companies have people thinking about it.

嗯，他们更多是在思考AI是否应该对我们友善，而不是我们

Well, they're thinking about whether the AI should be nice to us more than whether we

不，不，不，不，不。

should No, no, no, no, no.

他们也在思考我们是否应该对人工智能友善。

They're also thinking about whether we should be nice to the AI.

我在那次会上向一些人提出的一个问题是，为什么这些公司会关注这个问题。

And one of the issues I raised with a number of people at that meeting was why the companies are thinking about this.

一些人给我的回答是，他们想提前做好准备，以防有人开始抱怨虐待机器。

The answer I got from some people is, well, they want to be on top of it in case people start complaining about torturing the machines.

希望拥有一整套相关信息。

Want to have a whole body of information.

我随后提出的一个担忧是，你看，人们思考这个问题所产生出来的大量内容，可能会被用来反对他们。

The worry, what I then raised was, well, look, what you're getting from people thinking about this is a lot of stuff that could be used against them.

是的。

Yeah.

这真的是他们愿意资助的方向吗？

And is that really what they want to fund?

他们难道不应该直接压制这种想法吗？

Shouldn't they be just squelching it?

然后，我想到了一个答案，虽然没人说出来，但其实非常简单。

And then, you know, the answer that occurred to me, although nobody said it, was actually very simple.

这些公司所获得的利润中，超过50%——等等，我纠正一下。

And that is more than 50% of the profit that these companies make more, sorry, take it back.

人们使用这些设备的用途中，超过50%是作为伴侣。

More than 50% of the uses people put these things to are as companions.

因此，只要他们能强化‘机器可能具有意识’这一观念，就对他们的利润有利。

So the extent to the extent that they can boost the idea that maybe the machines are conscious, that is good for the bottom line.

我发现这个群体是一群奇怪的人的混合体：有些人纯粹为了赚钱而不择手段，有些人则完全理想主义，希望创造更美好的未来，他们两者都存在，而且彼此紧密合作。

I find that that community a weird mix of people who are absolutely ruthlessly in it for the money and people who are entirely idealistic about, you know, creating a better future and creating like, they're both there and they're both working hand in hand.

是的。

Yeah.

说得太对了。

That's exactly right.

这是一件非常有趣的事情。

It's it is a very interesting thing.

我们会看到神经科学和人工智能领域发生很多变化。

And we're gonna see a lot of things happening on the the neuroscience side, the AI side.

我认为哲学家们，哦，没错。

I think that the philosophers Oh, yeah.

哲学家们需要赶紧行动起来，给我们一个答案：当我们应该开始善待人工智能时，因为这是我们的责任。

Philosophers need to get on the stick and give us an answer to this when we should start being nice to the AI because it's our job.

我们真正需要的是更多人来研究这个问题。

Well, what we really need is more people studying it.

我认为我们今天讨论的这些问题都非常重要。

Both the issues we're I think the issues we're talking about today are super important.

每个哲学系都应该有人专门研究这类问题。

You know, every philosophy department should have somebody who's working on this kind of thing.

我们谈了很多事情。

Well, talked about a lot of things.

你是说特别指AI，还是意识，或者

Do you mean the AI in particular, or consciousness, or

是的，特别是AI，因为它是一个重大的社会问题。

Yeah, AI in particular, because it's such a major social issue.

但你知道，与动物意识相关的问题也变得非常重要。

But you know, it's also true that the issues having to do with animal consciousness have become significant.

我们甚至都没谈到这一点。

We didn't even talk about it.

天啊，这背后有一个巨大的议题，可以说属于扩展道德圈的范畴。

And boy, that's, there's a huge, I mean, it's under the heading of expanding the moral circle.

但很多人对这个感兴趣，也有很多相关活动正在进行。

But a lot of people are interested in this, a lot of stuff going on.

我们真的需要，你知道，学术界得行动起来

And we really you know, academia has to get on

是的。

Yeah.

对此表示支持。

On on board with this stuff.

你知道，院系被各种条条框框束缚着，不愿做新事情，但它们真的应该去做这些。

You know, departments are so high bound with you know, they don't do new things, but they really should be doing this.

我站在你这边。

I'm on your side.

我的意思是，我们一想到未来500年的人会怎么评价我们，就会有点不舒服。

I mean, we always cringe a little bit to think of what the people 500 in the future are gonna say about us.

是的。

And Yeah.

我们忽视了很多事情，也许应该关注一下。

A lot of things we're not paying attention to, maybe we should.

但是

But

如果有如果有

If there are if there are

人们五百年后，计算机会对我们的所作所为说什么，也许我们该问问。

people five the what the computers will be saying about us, maybe we should be asking.

但这对我们开展这些讨论提供了非常有益的视角。

But that was extremely helpful in in, framing us for having these discussions.

所以，内德·布洛克，非常感谢你做客《Lionscape播客》。

So Ned Block, thanks very much for being on the Lionscape Podcast.

哦，谢谢。

Oh, thank you.

这很有趣。

It was fun.