343 | 汤姆·格里菲斯谈思维法则

而且，这又是二十世纪的一种创新。

And, again, that's a kinda twentieth century innovation.

我非常推崇溯因推理。

I'm a huge fan of abduction.

我不确定我是不是喜欢得有点过头了，但在我看来，它最接近科学实际运作的方式。

I'm not sure if I'm more of a fan than I should be, but it it seems to me to be the closest to the way that science actually works.

对吧？

Right?

而且，它有时会被与最佳解释推论混淆，这其实是在承认：这并不是非黑即白、算法化的过程，但我们仍然在做某种独特合理的事情。

And and sometimes it's mixed up with inference to the best explanation, and it's sort of admitting that this is not clear cut and and algorithmic, but still there is something uniquely sensible that we're doing.

是的。

Yeah.

我认为这是对的。

I think that's right.

我的意思是，对于许多这类归纳推理，概率理论为我们提供了很好的描述，说明你应该如何做出这些推理。

I mean, I think for for many of these kinds of inductive inferences, probability theory gives us a good description of how it is that you should make those inferences.

但对于像 abduction 这样的情况，很多工作实际上是在人们如何真正做出这些推断的层面上完成的。

But for something like abduction, a lot of the work is actually done at the level of how people really do make those inferences.

对吧？

Right?

比如，其中一个挑战是你可能在提出一种前所未有的想法。

Like, one of the challenges is you're probably coming up with a kind of thing that no one's thought about before.

对吧？

Right?

你必须先提出一个假设，才能去思考这个假设。

You have to come up with a hypothesis in order to be able to entertain that hypothesis.

这属于一种算法层面的现象。

And that's something which is a algorithmic level phenomenon.

对吧？

Right?

这关乎我们的大脑在做什么，而不是我们被告诉用数学如何去做。

It's something which is about something our brains are doing rather than something that's told we're told how to do using the math.

这其实是在跳到我们对话的未来，但我现在就想提一下。

And this is skipping way ahead to the future of our conversation, but let me just bring it in right now.

我觉得，像我们现在拥有的大型语言模型或某种机器学习算法，在合适的背景下完全有能力求解爱因斯坦的广义相对论方程，但它们很难产生那个最初的创造性时刻——比如提出‘引力是时空弯曲’这样的想法，因为它们只是被训练去学习已经发生过的事情。

Like, I have this feeling that a large language model or some machine learning algorithm that we have right now would be perfectly good at solving Einstein's equations of general relativity in the right context, but it would really struggle to have that first creative moment where it it suggested that gravity is the curvature of space time because simply because they are trained on things that have already happened.

这是不是我太以人类为中心了？还是你觉得这其中确实有道理？

Is this me being anthropocentric, or do you think that there's some truth there?

我觉得这确实有一定道理。

I I think there's some truth to that.

这无疑是一个开放性问题，即这些模型在多大程度上能够做出这类推断性思考。

It's certainly something which I think is an open question about the capacities of these models in terms of the extent to which they're able to make those sorts of extrapolative inferences.

我们预期它们能够做到这一点，前提是它们被训练的内容能为这种能力提供某种基础。

And we would expect that they could do so to the extent that, you know, again, what they've been trained to do provides some kind of, you know, infrastructure for being able to do that.

对吧？

Right?

因此我们发现，这些模型实际上在某些类型的创造性思维上表现得相当不错。

So we find that there are certain kinds of creative thinking that these models can actually do reasonably well.

比如，它们在提出简单的类比时，可能比人类更擅长。

Like, they're maybe better than people that coming up with simple kinds of analogies.

对吧？

Right?

是的。

Mhmm.

你可以把这看作是它们对语言有着极其精细理解的结果。

And you can think about that as being a consequence of having this very fine understanding of language.

对吧？

Right?

这对实现这一点很重要。

That's that's important for doing that.

但类比也是发现的重要组成部分。

But analogy is an important part of discovery too.

有时候，你通过意识到某个领域中的想法可以应用到另一个领域而做出发现。

Sometimes you make a discovery by recognizing that, an idea from one domain applies in another domain.

所以我认为，情况不会像他们无法做到我们所认为的溯因推理那样简单。

And so I think it's not gonna be as simple as they're not able to do this thing that we think of as abductive inference.

我认为，他们会有一些擅长的事情，因为这些事情与他们试图完成的任务相契合，而另一些事情可能对他们来说更难，因为这些事情与他们的训练方向相冲突。

I think it's gonna be something where there's gonna be a sort of set of things that they're gonna be able to do well because they align with the kinds of tasks they're trying to do, and then a set of things that maybe are are harder for them because they push against that training.

是的。

Yeah.

好的。

Okay.

这确实有道理。

That does make sense.

回到概率和信念，我的播客听众经常听我谈论贝叶斯推理。

And back to the probabilities and the beliefs, my podcast listeners hear me talk about Bayesian reasoning all the time.

我们这里讨论的就是这个吗？

Is that what we're talking about here?

我知道贝叶斯本人只是……我印象中他的公式是死后才发表的吧？

And I know that Bayes himself just sort of I get was it even posthumously published his his formula?

所以他并不是这场讨论中的主要人物，但我们还是给他一些赞誉。

So he was not a big player in that discussion, but we give him some credit.

是的。

Yeah.

所以我们之前谈到了一条线索，对吧？那就是从莱布尼茨到布尔的逻辑线索。

So so we we talked about one thread, right, which is the the thread of logic through Leibniz and Boole.

而这本书实际上讲的是三种思维脉络。

And then the book is really about three threads of thinking.

对吧？

Right?

第一条是逻辑。

So one is logic.

第二条是神经网络的基础，我大致将其描述为空间、特征和网络。

One is sort of foundations of neural networks, which I sort of characterize in terms of spaces, features, networks.

对吧？

Right?

所以把思想看作是空间中的一个点。

So thinking about thoughts as corresponding to a point in space.

对吧？

Right?

而我们用来思考空间的一些数学工具，比如微积分等，也可以用来思考思想是如何运作的。

And some of the mathematics that we use for thinking about spaces, like calculus and so on, being a tool for then thinking about how thoughts work.

第三个线索是概率论，我认为它对解释思想运作的各个方面都非常有帮助，并且与其他线索相互补充。

And then the third thread is this thread of probability theory, which I think each of them is very helpful and sort of complementary to the other in in explaining various aspects of of how thinking works.

关于概率论，我在书中重点关注的是十八世纪的这种观点，即概率论可以应用于思想，这是一个非常激进的想法。

And so for probability theory, yeah, the the the origin that I focus on in the book is this sort of eighteenth century idea where really the the radical idea is that probability theory could be applied to thought.

对吧？

Right?

在那之前，人们一直在将概率论发展为一种数学理论。

So before that, people had been developing probability theory as a kind of mathematical theory.

嗯。

Mhmm.

这是一门应用于赌博游戏等事物的数学理论。

And it was a mathematical theory that applied to things like gambling games.

对吧？

Right?

因此，你在概率论的最早起源中就能看到这一点。

So you see this in the very earliest origins of probability theory.

最典型的例子就是吉罗拉莫·卡尔达诺，他是一位数学家，同时也是一位沉迷赌博的人。

It's, you know, it's the the the best example we have is like Juralama Kadana, who is a mathematician, but also an addictive gambler.

是的。

Yep.

对吧？

Right?

因此，出于娱乐和财务方面的考虑，他非常想知道如何思考掷骰子或这类概率事件的结果。

And so he he really wants to know for his recreational, you know, and financial reasons how to think about the outcomes of rolling dice or these sort of probabilistic events.

于是，他推导出了如何计算这些结果的数学方法。

And so he works out the mathematics of how to do that.

在概率论的早期发展中，下一个重要时刻是布莱斯·帕斯卡做了类似的事情。

And there's a few, the next sort of moment that we see in the origins of probability theory is Blaise Pascal doing something similar.

对吧？

Right?

他试图解决一个赌博问题，并由此发展出概率论的一些基础理论。

Sort of like trying to solve a sort of gambling problem and and from that developing some of the the foundations of probability theory.

但那是一种关于掷骰子会发生什么的理论，是的。

But that was a theory of, yeah, what happens when you roll dice.

没错。

Right.

对吧？

Right?

贝叶斯的创新在于提出：也许我们现有的这种数学体系。

And the the innovation that comes with Bayes was saying, well, maybe this mathematical system that we have.

对吧？

Right?

这种描述某个数学对象的公理体系，也同样适用于我们感兴趣的另一个事物。

This sort of set of axioms that characterizes some mathematical object also characterizes another thing that we're interested in.

这不仅仅是关于掷骰子的事情。

It's not just what's going on with dice.

它还关乎我们改变信念时，头脑内部发生的过程。

It's also what's going on inside our heads when we change our beliefs.

对吧？

Right?

所以他当时在思考一些受赌博启发的例子。

So he was thinking about some gambling inspired examples.

对吧？

Right?

比如，如果有一个彩票以某种比率兑奖，你该如何估算它的兑奖率？

So, you know, if there is a lottery which is paying off at some rate, how do you estimate the rate at which it's paying off?

但他设定这个问题的方式，是基于你对这个彩票所持有的信念。

But the way that he sets that up is in terms of the beliefs that you have about that lottery.

比如，基于你迄今为止看到的例子，你对它在下一刻 payout 的概率应该做出怎样的合理估计？

Like, what's the what's a reasonable estimate that you should have for the probability that it's gonna, you know, pay off at the next moment given the the examples that you've seen so far.

随后，皮埃尔·西蒙·德·拉普拉斯进一步发展了这一思想，他独立提出了这一观点，并深入探讨了这种思考方式对更新我们信念的所有后果。

And then that that idea is developed further by Pierre Simon de Plasse who really sort of came up with it independently and really, you know, worked out all of the consequences of of that way of thinking about how to how to update our beliefs.

所以，是否可以大致说，在贝叶斯和拉普拉斯之前，人们可以赌博。

So would it be fair to say that, at least roughly speaking, pre Bayes and Laplace, you could gamble.

当时存在不确定性之类的问题。

There could be uncertainties and things like that.

但在思考时，人们被认为要么对，要么错。

But when it came to thinking, you were supposed to be right or wrong.

而在他们之后，你可以说：我对自己的信念有一定的信心程度。

And after them, you could say, well, I have a certain degree of confidence in my beliefs.

我认为，贝叶斯之前确实存在一个明确的转折点，只是当时还没有系统地发展出这一理论。

It's it's I I you know, I think I think there's a discrete point that happens with Bayes in the past just instead of working working out that that theory.

但在那之前，你确实能看到一些人用这种说法进行讨论。

But you do see hints of people talking in those terms before that.

对吧？

Right?

所以就连威尔金斯，那个提出了一种你永远不会说错话的语言理念的人，也把概率当作表达信念程度的一种方式。

So even Wilkins, you know, who who came up with this sort of idea of, you know, the the language that you could use where you could never say anything false, talked about probability as a way of talking about a degree of belief.

他并没有深入推导出数学上的后果，但他确实使用了这种语言。

He didn't sort of work out the mathematical consequences, but he sort of used that language.

帕斯卡在离开数学界、转而投身宗教思考后， famously 做了一个概率论的论证，这其实也是一个关于信念的论证。

And Pascal famously, you know, after he departed the world of mathematics and and instead started to think about religion, made a probabilistic argument in that setting, which is really an argument about belief as well.

所以在贝叶斯和拉普拉斯之前，你就已经能看到很多这样的迹象了。

So you you see hint lots of hints of this Good.

在贝叶斯和拉普拉斯之前，你就已经能看到很多这样的迹象了。

Prior to Bayes and Laplace, but I think they're the the ones who really developed that into a a theory of what we could call thought.

让我们深入探讨一下这种我们可以称之为思维的理论。

So let's dig into that theory of what we could call thought.

我们拥有一些信念。

We have some beliefs.

它们是概率性的。

They're probabilistic.

你该如何像布尔所建议的那样，用对待真假命题的方式来推理它们？

How do you reason with them in the way that Boole would have had us reasoning with true and false statements?

我认为贝叶斯概率最酷的地方在于，一种非常自然的理解方式就是将其视为逻辑的延伸。

I I think the really cool thing about Bayesian probability is that one very natural way to see it is just an extension of logic.

对吧？

Right?

在逻辑中，我们谈论可能的世界。

So in logic, we talk about possible worlds.

对吧？

Right?

如果你有两个命题 p 和 q，你可以想象所有你可能身处的世界。

So if you have two propositions, p and q, you can imagine all of the possible worlds that you could be in.

你可能处于 p 为真且 q 为真的世界，p 为真而 q 为假的世界，p 为假而 q 为真的世界，以及 p 和 q 都为假的世界。

You could be in a world where p is true and q is true, a world where p is true and q is false, a world where, you know, p is false and q is true, and a world where both p and q are false.

对吧？

Right?

这些就是我们可能身处的可能世界，而逻辑真正关注的是，基于你对自身所处世界的了解，你能确定地得出哪些结论。

Those are the possible worlds we could live And logic is really about what conclusions you can draw with certainty based on the information you have about what world you might be in.

对吧？

Right?

所以，如果你掌握了足够的信息，能够排除某些可能世界，从而确定你所处的世界必然是q为真的那个，那么得出q为真的结论就是合理的。

So if you have got enough information to rule out some of those possible worlds such that it has to be the case that the world you're in is one where q is true, then it's reasonable to conclude the q is true.

对吧？

Right?

因此，你所熟知的经典逻辑论证，其实就是告诉你：你所掌握的信息对可能世界施加了足够的约束，使得你可以合理地得出关于这个世界的结论。

And so your classic logical arguments are arguments that are telling you, oh, the information you have gives you enough constraints on the world that you might be in that this is a reasonable conclusion that you can draw about that world.

概率论更进一步，它为这些可能世界赋予了一个数值。

Probability theory takes one more step, which is to say, for those possible worlds, we're also going to assign a number to them.

这个数值反映了我们对这个世界为真的信念程度或概率。

And that number reflects our degree of belief about the the probability that that world is true.

一旦你这么做了，并开始遵循概率论的规则，你就会根据到目前为止所掌握的信息，更新你对当前可能处于哪个世界的信念。

And then as soon as you do that and you start following the rules of probability theory, you're then updating your beliefs about, you know, what world is it that we're likely to be in based on the information that I've got so far.

因此，逻辑论证仍然有效。

And so the logical arguments still work.

对吧？

Right?

所以，你可以把逻辑论证理解为告诉你某事以概率一为真。

So the you know, you can think about a logical argument is telling you that something is true with probability one.

也就是说，你必然处于一个该事物为真的世界中。

That with certainty, it has to be the case that you're in a world where this thing is true.

但概率论对此进行了扩展，允许我们说：我们没有足够的信息来确定某事必然为真。

But it's generalized by probability theory and allowing us to say, oh, we don't have enough information to determine that this thing is true, you know, with certainty.

我们可以说：这件事有70%的可能性为真。

We can say, oh, well, there's like a 70% chance this thing is true.

但核心理念是一样的：随着你获得更多信息，你可能会排除一些可能的世界。

But the same kinda idea of as you get information, you're maybe ruling out some possible worlds.

也许，正因为如此，你正在改变你所处的其他世界的概率，或者你获得了新的信息，从而改变了你认为自己处于某个世界而非另一个世界的可能性。

Maybe, you know, and as a consequence of that, you're changing the probabilities of the other worlds that you could be in or you're getting information that that changes the chances that you think you're in one world or another.

然后你就只是在做类似的事情。

And then you're just doing the same kind of thing.

只不过你现在是以一种更加细致入微的方式在做这件事。

It's just that you're now doing it in this much more graded way.

就像我们可以质疑人们在多大程度上符合传统的逻辑一样。

And just like we can question how conventionally logical people are.

我们也可以质疑他们在新数据到来时，更新信念的能力有多强。

We can also convent we can question how good they are at updating their beliefs when new data come in.

我的意思是，我们是否应该把完美的贝叶斯推理再次视为思维法则的一种理想目标，还是说它本意是描述人们实际的思考方式？

I mean so maybe should we think about perfect Bayesian reasoning as once again aspirational when it comes to laws of thought, or is this meant to be a description of how people actually think?

不。

No.

它是一种在抽象计算层面上的工具，用来说明：我们究竟应该怎么做？

It's it's a it's a tool at that abstract computational level of saying, what is it that we should be doing?

是的

Yeah.

对吧？

Right?

我们心智所面临问题的理想解决方案是什么？

What's the solution the ideal solution to the problem that that our minds face?

因此，当我们的思维面对归纳问题时，概率理论告诉我们这些问题的理想解决方案是什么样子。

And so when our minds face inductive problems, probability theory tells us what the ideal solution to those problems look like.

当然，有很多方式表明这与人们实际的行为并不一致。

And, of course, there's lots of ways that that doesn't line up with the things that people actually do.

在二十世纪，我们开始探索这些差异。

And in the twentieth century, we started to explore those.

其中一些，我认为可以从这种贝叶斯视角来解释，但我们需要考虑人们实际上在做些与被指示内容略有不同的事情。

Some of those, I think, are things that we can we can explain from this sort of Bayesian perspective, but where we have to think about people doing something slightly different from the thing that they've been told to do.

嗯

Mhmm.

我可以再多谈一点这个。

I can talk a little more about that.

而其中一些则处于这些不同分析层次之间。

And then some of them are in between these different levels of analysis.

对吧？

Right?

所以我们有抽象计算层面，即你该做什么？

So we have that abstract computation level, what should you be doing?

然后是算法层面，即什么是接近目标的好策略？

And then there's the algorithmic level, which is what's a good strategy for trying to get close to that?

你可以在这一层提出一个问题，即在给定资源限制的情况下，你能做到的最好程度是什么。

And you can ask a question at that level, which is about, you know, what's the best that you could do at trying to do the thing you're supposed to be doing with particular constraints on the resources that are available to you.

对。

Right.

当我们明确这些限制条件时，实际上就能推导出其具体表现形式。

And then when we characterize what those constraints are, we can actually work out what that looks like.

在许多情况下，这实际上让我们对人们一些看似奇怪的行为有了更深的理解。

And and in many cases, that actually gives us insight into some of the, you know, what might seem like strange things that people do.

我在这方面做了很多研究，我们几周后将出版一本关于这个理念的书，我们称之为资源理性。

And that's something that I've done a bunch of work on, and we actually have a book coming out in a few weeks, which is about that idea of, we call it resource rationality.

这本书探讨了如何在认识到这些约束后，改变我们对理性的理解。

And the the book explores how to change our notions of rationality as a consequence of recognizing those kinds of constraints.

所以你有两本书将在一个月内相继出版吗？

So you have two books coming out, like, within a month of each other?

这真的很尴尬。

It's it's very awkward.

它们的出版时间只相隔一周。

They're coming out within a week of one another.

其中一本面向普通读者，探讨关于思维规律的这些理念。

So where one one is, you know, more for a general audience and sort of exploring these kinds of ideas about the laws of thought.

另一本则更学术一些，专注于我们该如何利用有限的认知资源。

And then the other is a slightly more academic book, which is focused on, this this idea of, what we should do with our, limited cognitive resources.

你可以同时为两本书签名，举办签售会。

You should you can still do book signings where you sign both at once.

这样也没问题。

That's okay.

这完全说得通。

That makes perfect sense.

换句话说，我们能不能简化一下：我们希望用贝叶斯定理来更新我们的信念，但这很难，而且需要大量资源。

So in other words, can we can we sort of simplify it down to we would like to use Bayes' theorem to update our beliefs, but that's hard, and that takes a lot of resources.

所以，进化和生物学赋予了我们某些捷径吗？

So evolution and biology have equipped us with certain shortcuts?

我认为我可以这样理解：是的。

I think the the way that I would think about it is yeah.

我认为人类认知的一个有趣悖论在于，从心理学家的角度看，我们是容易出错的决策者，会使用这些启发式方法导致偏见。

I think it's that one of the interesting paradoxes of human cognition is that we are both, you know, from the perspective of a psychologist, error prone decision makers who sort of use these heuristics to result in biases.

但从计算机科学家的角度看，我们又是那些具备我们希望AI系统具备的能力的理想型主体。

And from the perspective of computer scientists, these aspirational agents that are doing the kinds of things that we'd like our AI systems to do.

对吧？

Right?

因此，如果你想解决这个悖论，我的做法是说：我们擅长利用现有的资源解决所面临的问题。

And so if you wanna resolve that paradox, the way that I resolve it is to say, we are, you know, good at solving the kinds of problems that we face with the resources that we have.

对吧？

Right?

你可以将这视为不同形式适应的结果。

And you can think about that as being the consequences of the different kinds of adaptation.

包括进化和学习。

So evolution as well as learning.

对吧？

Right?

因此，在我们的一生中，学会更有效地利用我们的认知资源。

So over the course of our lifetime, learning to use our cognitive resources better.

对吧？

Right?

还包括进行某种规划，或者我们称之为元推理，即思考如何恰当地应对我们试图解决的不同类型的问题。

As well as just sort of engaging in some planning or we call it meta reasoning about how to appropriately approach different kinds of problems that we're trying to solve.

我们之前曾邀请卡尔·弗里斯顿做客播客，他谈到了自由能原理。

We did have, Carl Friston on the podcast some time ago talking about the free energy principle.

这是否是一个例子，说明大脑正在以一种高效的方式解决这些难题？

Is that an example of, you know, the brain trying to solve these hard problems in a efficient way?

这是个好问题。

That's a good question.

我还没有从这个角度思考过。

I haven't thought about it in those terms.

我认为他设定这个框架时，更多是从系统所追求的一种目标出发，而不是从资源限制的角度来看。

The I think the way that he sets that up is more in terms of a a kind of objective that the system has rather than in terms of a resource constraint.

而我们的方式则更明确地指出：如果我们想重新定义理性，使其适用于具有有限计算资源的智能体，

And the way that we think about it is a little more explicitly saying, if you are if we want to redefine what rationality is in a way that works for agents with finite computational resources.

这借鉴了人工智能领域斯图尔特·罗素和埃里克·霍维茨的一个观点。

This is drawing on an idea from from the AI literature from Stuart Russell and Eric Horvitz.

对于一个有限资源的智能体来说，定义理性的方式更侧重于：不再关注概率理论告诉你应该采取的行动，而是使用最佳算法来选择你要采取的行动。

The the way to define what rationality is for a a bounded agent is more in terms of, taking, instead of focusing on sort of taking the action that's the action that probability theory and someone tells you you should take, it's using the best algorithm to choose the action that you're going to take.

这就相当于提升了一个抽象层次。

And so it's it's sort of popping up a level of abstraction

好的。

Okay.

从元层面思考理性，将其作为一种工具，用以确定在具体层面——即你所采取的行动中——如何合理运用你的认知资源。我之前写的《算法来生活》一书，实际上是对这些理念的一个很好的大众化解读。

In terms of thinking about, you know, defining rationality at that meta level as a as a tool for then generating, you know, what are the sort of appropriate ways of using your cognitive resources at the what what the the object level, the actions you take in the And and my previous book, Algorithms to Live By, is actually a a pretty good sort of general audience treatment of those ideas.

我们当时并没有用资源理性这一框架来表述，但它核心讲的是：在某种程度上，计算机科学为理性提供了更好的指引。

We we didn't express it in terms of this framework of resource rationality, but it's really about the idea that, you know, in some ways, computer science provides a better guide Right.

比经济学更能指引理性，比如那种基于概率和回报的思维方式。

To rationality than than, you know, economics, right, sort of thinking in terms of probabilities and rewards.

我的意思是，我以前从未这样想过，但当你在编写计算机程序时，资源限制是显而易见的。

I mean, I guess I've never really thought about it this way, but when you are computer programming, the resource limitations are obvious.

你有有限的时间、有限的内存，还有有限的各种数据。

You have a certain amount of time, certain amount of memory, right, a certain amount of whatever data.

但当然，同样的道理也适用于人脑，我们可以想象理想化的思维，但真正实现这些思维可能会消耗过多的资源。

But, of course, the same things are gonna apply to human brains where we can we can imagine ideal thoughts, but actually having them is something that's gonna be probably too resource intensive.

是的。

Yep.

对。

Yep.

我认为这就是我思考这个问题的原因。

I think that's a reason why I think about it.

那么，你刚才提到过一点，但到底我们应该怎么做呢？

And so what so what is you I mean, you said a little bit about this, but what do we do?

我们的策略是什么？

What are our strategies?

生活在一个不完美的世界里，不可能时刻都做到精确的贝叶斯推理，人类实际使用了哪些捷径？

Living in an imperfect world, not being able to be exact Bayesians at all times, do we have what are the shortcuts that actual human people use?

是的。

Yeah.

所以我们使用的一些策略，就是人们所识别出的启发式方法。

So some of the the kinds of strategies we use are the kinds of things people have identified as, you know, heuristics.

对吧？

Right?

启发式方法就是指一种经验法则或解决问题的捷径。

Heuristic just means a rule of thumb or a shortcut for solving a problem.

我认为，从资源合理性角度重新分析这些启发式方法的价值在于，能够说明使用这些启发式方法并不一定是坏事。

I think part of what is valuable about reanalyzing those from the perspective of, you know, resource rationality is is being able to say that using those heuristics isn't necessarily a bad thing.

对吧？

Right?

而这些启发式方法带来的偏见，也可能并非在你现有的认知资源条件下能够避免的。

And the biases that come from those might not necessarily be things that you can avoid given the cognitive resources that you're operating with.

因此，我们可以问一个问题：我们是否在现有认知资源条件下做到了最好？

So instead, we can ask a question like, are we sort of doing the best job we could with the cognitive resources that we have?

然后，我们是否可以通过在特定情境下使用不同的启发式方法，来缓解这些偏见？

And then is there a way that we could mitigate those biases by maybe using a different heuristic in a particular setting or something like that?

因此，我们在资源理性研究中关注的策略包括用于近似贝叶斯推理的采样策略。

And so the kinds of strategies that we focus on in our work on resource rationality are things like sampling strategies for approximating Bayesian inference.

因此，你不必考虑整个概率分布，而只需考虑少数几个样本、少数几种可能的情况。

So instead of thinking about a whole probability distribution, you might think about a few samples, a few possible instances.

对吧？

Right?

当你做决定时，不必考虑所有可能的结果，而只需考虑少数几种可能的结果。

Instead of when you're making a decision considering all of the possible outcomes, you might think about a few possible outcomes.

对吧？

Right?

诸如此类的方法。

And things like that.

我们可以观察人们在这样做时所使用的策略——当他们没有考虑所有可能性，而只考虑部分可能性时，这些策略是否符合在有限资源条件下应考虑的最优选择。

And we can look at the kinds of strategies that people seem to use when they do that, when they don't consider all the possibilities are the ones that they're considering, the ones that they should be considering from the perspective of using limited resources.

我们思考的其他方面包括设定目标，或者子目标。

And then the other kinds of things that we think about are things like setting goals, right, or sub goals.

所以我认为，从认知研究的历史角度来看，艾伦·纽厄尔和赫伯特·西蒙似乎提出了这样一个观点：当我们解决复杂问题时，需要将这些问题分解成若干部分。

So I think from the perspective of the sort of history of thinking about cognition, looks like Alan Newell and and Herb Simon sort of introduced this idea of, you know, we can think about when we're solving challenging problems, we need to decompose those problems into parts.

而聪明的一部分表现，就在于能够以这种方式进行分解。

And part of what it is to be smart is to be able to decompose them in those ways.

但在很多方面，这种分解问题和设定目标的能力，实际上是资源受限的必然结果。

But in many ways, that ability to kinda break down problems and set goals is really a consequence of a resource constraint.

对吧？

Right?

所以，如果你拥有无限的认知资源，就根本不需要设定目标，因为你可以直接推理出你所做选择最终会导致的所有结果。

So, you know, if you had infinite cognitive resources, you would never need to set goals because you can just reason all the way to the end of the trajectory of, you know, whatever is gonna, arise from from the the choices that you're going to make.

因此，设定目标和子目标，是一种在认知资源有限的情况下推进问题解决的工具。

And so setting goals and sub goals and so on is a tool for being able to make progress on problems with finite cognitive resources.

那么我们可以进一步问：哪些目标是好的目标？

And then we can ask, what are the good goals to set?

从资源理性角度来看，什么样的问题结构才是理想的？

What's a good structure to to sort of give to to a problem from that perspective of resource rationality?

抱歉。

So sorry.

从某种意义上说，我们每个人都有一个目标，可以理想化为过上最好的生活。

In some sense, we all have a goal that could be idealized as live the best possible life.

但你说，作为一种实现这一目标的策略，这并不真正合理。

But you're saying that as a strategy for getting there, that's not really reasonable.

我们不可能像《复仇者联盟》中的奇异博士那样，穿越多元宇宙的每一个可能路径，我们必须设定一些中间目标，以获得一条大致不错的前进轨迹。

We can't actually be doctor Strange in the Avengers and go through every possible part of the multiverse, we have to sort of have subgoals along the way that give us an approximately pretty good trajectory.

是的。

Yeah.

然后你可以问，显然你不希望你的子目标太近。

And then you can ask, you know, you obviously don't wanna make your subgoal too close.

对吧？

Right?

你也不希望它太远。

And you don't wanna make it too far away.

因此，人们应该在哪里设定他们的子目标，以及他们是否很好地设定了这些子目标，这个问题我们可以从资源合理性的角度来探讨。

And so the question of where people should set their sub goals and do they do a good job of setting those sub goals is is a question that we can engage with from that from that perspective of resource rationale.

总会有这样一个问题，即贝叶斯推理：因为整体图景是你有一些先验概率。

There's always this question Bayesian reasoning that because the whole picture is you have some prior probabilities.

你获得了一些数据。

You get some data.

你计算出一个似然函数。

You calculate a likelihood function.

你更新了你的先验。

You update your priors.

但那么，这些先验又是从何而来的呢？

But so then where did the priors come from?

这个问题在这里是否也涉及其中？

Is is that question involved here?

也就是说，人类究竟从哪里获得对不同命题可能性的初步直觉？

Like, where do human beings actually have their rough feelings about the plausibility of different propositions?

是的

Yeah.

所以我认为这是一个非常深刻的问题，而且我认为可以从不同角度来提出这个问题。

So this is, I think, a very deep question, and I think there's different different ways that you can ask it.

对吧？

Right?

所以，有一种思考方式是，当你把贝叶斯法则视为描述归纳问题理想解决方案的工具时，这种描述既适用于你在感知中做出的推断，比如当你试图解读落在视网膜上的光线时。

So, there's there's one way of thinking about so so when when you think about Bayes' rule as our tool for describing what ideal solutions to inductive problems look like, That characterization applies both to, you know, an inference that you might make in perception when you're trying to interpret the light that's falling on your retina.

你的大脑必须做一些看起来像是归纳推断的事情，才能弄清楚外部世界的结构，比如解读别人说的一句话，你听到或感受到的声音，会转化为对对方所说内容及其含义的推断。

Your brain has to do something that looks like an inductive inference, right, to figure out the structure of the world out there, to interpreting a sentence that somebody says, right, where you're taking the words that you hear or the, you know, the sound that's hitting your eardrum and sort of turning that into a a an inference about what it is that the person said and maybe what they meant.

但同时也适用于一些根本性的问题，比如我们最初是如何学习语言的？

But also to, you know, like, fundamental things like how do we learn language in the first place?

而且

And

大脑是如何学会解读我们周围物理世界的结构的？

how is it that brains, you know, come to be able to interpret the structure of the physical world around us?

对吧？

Right?

所以所有这些都可以被视为归纳问题。

So all of those things are things you can think about as inductive problems.

因此，在这些不同情况下，先验的来源也会不同。

And so asking where the priors come from is gonna be different in those different cases.

对吧？

Right?

所以更根本的情况，也就是关于我们如何学习语言的情况，如果我们把这看作一个归纳推理问题的话。

So the the the more fundamental case, the one which is about, like, you know, how do we learn language if we think about that as a a problem of inductive inference.

在那里，先验将反映我们天生的学习语言的倾向，同时也包括所有其他非语言输入的信息来源。

The priors there are gonna reflect whatever the innate predispositions we have to learn language, but also all of the other sources of information that we have, you know, that are sort of not the linguistic input.

对吧？

Right?

因此，我们在世界中的经验等等，都会影响我们从听到的言语中学习语言的方式。

So the experience that we have in the world and and so on is is stuff that's gonna inform the way that we learn language from the the utterances that we hear.

因此，这是一个很好的工具，可以帮助我们思考人类与大型语言模型之间的差异，而人类心智和大脑与当今大型语言模型之间的主要区别就在于归纳偏差。

And so that is a good tool for using for thinking about what are differences between, like, humans and large language models, where the big difference between human minds and brains and large language models that we have today is about inductive bias.

它关乎我们人类能够从极少的数据中学习，而大型语言模型则需要海量的数据进行训练。

It's about being able to learn from the small amounts of data that we get as humans relative to the very large amounts of data that our large language models are trained on.

对吧？

Right?

一个孩子只需五年左右的语言接触就能学会使用语言。

So a human child learns to use language in about five years of exposure.

对吧？

Right?

相比之下，训练大型语言模型所用的数据量，相当于5050年连续不断的语音输入。

By comparison, that he used to train large language models is, you know, the equivalent of between 5,050 of of continuous speech.

对吧？

Right?

因此，这仅仅是数量级上的巨大差异。

So it's just sort of orders of magnitude difference.

而构成这一差距的就是归纳偏差。

And the thing that makes up that gap is inductive bias.

它是我们作为人类所拥有的先验分布（广义上）所带来的东西，正是这些使我们能够弥合这一差距。

It's it's the the sort of the thing that comes from our prior distributions broadly construed, right, as human beings that allows us to close that gap.

当我们观察日常生活中所做的推断——比如解释一句话或理解视觉信息这类短期任务时，这些先验正是那些学习过程发挥作用的结果。

When we look at, you know, the the sort of everyday inferences that we make, these sort of short term things like interpreting a sentence or or, you know, making sense of visual information, those priors are things that are really a consequence of those learning processes having worked.

对吗？

Right?

我们已经构建了关于周围世界的模型，这些模型影响着我们解释所经历数据的方式。

It's that we've built models of the world around us that inform the way that we interpret the data that we experience.

这一点上，我认为这些先验的来源没那么神秘，因为它们来自这个世界。

And that's something where, you know, I think it's it's a little less mysterious where those priors come from because they come from the world.

但它们也来自这个世界，再加上我们作为学习者所具备的更普遍的归纳偏差。

But they also come from the world plus, again, whatever are sort of, like, more general inductive biases as learners.

所以我们并不是白板。

So we're not blank slates.

对吧？

Right?

我的意思是，我想从康德到诺姆·乔姆斯基，许多思想家都说过，是的，我们生来就带着一些想法。

I mean, I guess various thinkers from Kant to Noam Chomsky have said that, like, yeah, we're we're born with some ideas in our heads.

而且，我们如今在二十一世纪，显然更擅长分辨哪些想法是与生俱来的，哪些是后天习得的。

And presumably, we're a lot better now in the twenty first century at at teasing out which of the ideas we do come born with and and which we pick up along the way.

是的。

Yeah.

我觉得这已经涉及到二十世纪的认知科学了。

I I think the the sort of this is getting into sort of twentieth century cognitive science.

对吧？

Right?

我们从十九世纪的逻辑和概率理论，跃迁到了二十一世纪对人们是否以正确方式、资源理性等的思考。

So we made a leap from the nineteenth century, you know, which is our logic and probability theory to sort of, like, twenty first century considerations about, you know, are people basing in the right ways and resource rationality and so on.

中间缺失的环节是二十世纪，那时人们开始将这些数学工具用于试图理解人类心智。

The missing chunk there is the twentieth century, which is where people began to use these mathematical ideas as a tool for trying to understand human minds.

对吧？

Right?

因此，二十世纪上半叶，心理学专注于成为一门严谨的科学学科，而它最初的基础其实是内省式的，你会问人们

And so, the first half of the twentieth century, psychology was, really focused on trying to be a a sort of rigorous scientific discipline, having gone from its foundations, which were really sort of introspective, where you'd be asking people

是的。

Right.

他们是否看到了什么、听到了什么，或者对它的感受是什么。

Whether they saw something or heard something or what the impression of it was.

对此产生了一种反拨，行为主义心理学家说，不，不，不。

There was a sort of reaction against that, and and behaviorist psychologists said, no, no, no.

我们看不到思想，也摸不到感觉。

We can't see thought or touch a feeling.

让我们关注那些我们能看到或触摸到的东西，也就是环境以及它们所产生的行为。

Let's focus on the things we can see or touch, which are environments and the behaviors that they produce.

对吧？

Right?

因此，人们不允许将这些心理状态作为解释人类行为的依据。

And so not allowed to really sort of talk about those those mental states as explanatory things in in accounting for human behavior.

随后，计算机的出现，以及用于思考如何在计算机上执行任务的数学框架，还有由此衍生出的逻辑扩展等，为心理学家提供了全新的理论工具，使他们能够构建关于心智运作的严谨理论。

And then the advent of computers, you know, and the existence of mathematical frameworks for thinking about how to do things on computers and sort of, you know, these sort of, like, extensions to logic and so on which came out of that, that provided a new set of theoretical tools that psychologists could use to come up with rigorous theories of how minds work.

对吧？

Right?

所以，你可以谈论思想了。

So you can talk about thoughts.

也许我们还没完全触及情感，但至少可以谈思想了。

Maybe we haven't quite got to feelings yet, but thoughts.

嗯。

Mhmm.

如果你拥有像逻辑这样精确的数学工具，就可以据此提出关于思想作用的假设。

If you have a precise mathematical device like logic for then coming up with hypotheses about what it is that thought does.

因此，这一研究领域的主要成就之一，就是艾伦·纽厄尔和赫伯特·西蒙创造了逻辑理论家，这是一台能够发现数学命题和逻辑证明的机器。

And so the the big sort of successes of that enterprise were Alan Newell and Herbert Simon creating the logic theorist, which was a machine that could discover, you know, proofs for for mathematical propositions and logic.

诺姆·乔姆斯基表明，对形式语言的思考为我们提供了一种工具，用以理解人类所使用的自然语言。

And Noam Chomsky showing that thinking about formal languages gave us a tool for then making sense of sort of natural languages that humans use.

对。

Right.

这些想法确实为科学奠定了基础，但也带来了一些有趣的挑战。

And and those ideas really sort of provided the foundation of science, but they also led to some interesting challenges.

乔姆斯基对语言的研究非常有效地说明了，语言比行为主义者所假设的要复杂得多。

So Chomsky's approach to language was very good in sort of illustrating that language was a much more complex object than behaviorists had assumed.

对吗？

Right?

你需要拥有类似内部结构的东西，比如动词短语、名词短语，以及看起来像语法的结构，才能解释人类语言的结构。

That you needed to have kinda like internal structures and things like verb phrases and noun phrases and and things that sort of look like grammars in order to account for the structure of human languages.

但这又带来了新的问题：人类是如何从有限的数据中学会这些极其复杂的结构的呢？

But then it created this new problem, which was how is it that human beings could possibly learn these very complex objects from the limited data that they get?

因此，这促使乔姆斯基提出：也许他们并不是真正地在学习它。

And so that's the thing that then pushed Chomsky to say, well, maybe they're not really learning it.

是的

Yeah.

你知道，这是通过一些非常强的约束来实现的，这些约束限定了他们能够学习的语言范围，因此只需要相对少量的数据，就能确定：哦，原来这就是我实际所讲语言的特定组合方式。

You know, acquiring it as a consequence of having some very strong constraints on what it is that they can learn as languages, but then only require relatively small amounts of data to determine, oh, okay, it's this particular configuration of, you know, bits and pieces that's that that characterizes the language that I'm actually speaking here.

我们是否应该以类似的方式设计人工智能？

And is there a thought that we should design our AIs similarly?

我的意思是，从连接主义的人工智能方法中得出的教训——这些方法催生了大型语言模型等——是人类已经完成了所有这些工作。

I mean, it seems like the the lesson from connectionist approaches to AIs that have led to large language models, etcetera, is the human beings have done all that work.

我们可以让人工智能成为白板，然后用海量数据来训练它们。

We can just, like, let the AIs be blank slates and and train them on a huge amount of data.

是的

Yeah.

这正是两者的对比。

So that's that's exactly the contrast.

对吧？

Right?

AI模型很好地说明了要解决乔姆斯基所识别的问题，需要多少语言数据。

Is that the the AI models give us a really good illustration of how much language you need to solve the problem that Chomsky had identified.

对吧？

Right?

去学习如此复杂的一个对象。

To learn something that is this very complex object.

对吧？

Right?

所以，如果我们默认AI已经学到了类似人类语言的东西，那么这就可以看作是对乔姆斯基观点的一个证明。

So if we sort of take as given that they've learned something like human language, then you can think about that as a as a proof of the point that Chomsky was making.

要学习类似人类语言的东西，你需要海量的数据。

You're gonna need lots and lots and lots of data to learn something like human language.

对吧？

Right?

所以他说得对，在孩子成长的五年里，他们不可能仅凭自身推导出语言的结构。

So he was right that in the five years that the kid gets, they're not gonna be able to figure out the structural language.

事实上，结果表明，要做得很好，你需要大约5000年或50000年的数据。

And in fact, it turns out you need something more like 5,000 or 50,000 years of data in order to do a really good job.

对吧？

Right?

所以我认为，这是一种非常棒的方式来理解这种差异。

And so I think that's a that's a really nice way of thinking about what that that difference is.

它也为我们提供了一个很好的视角，来思考如何构建在学习能力上更接近人类的系统。

And it also gives us a a good way of thinking about what the challenge is then if you wanted to make systems that are more human like in their ability to learn.

对吧？

Right?

所以你可以这样想。

So you can think about it.

如果你希望仅用五年数据就能学会，而目前却需要五千年数据，那么你就需要弥补四千九百九十五年的差距——也就是在儿童头脑中的归纳偏置或先验分布上做出改进。

If you wanna be able to learn from five years of data and you're currently learning from five thousand years of data, then you've got four thousand nine hundred and ninety five years to make up in terms of the content of that inductive bias or those prior distributions that are inside the child's head.

对吧？

Right?

是的

Mhmm.

因此，你可以从这些其他来源来思考这一点。

And so you can think about that coming from these other kinds of sources.

对吧？

Right?

所以进化是其中之一，还有一些其他因素。

So evolution is one of those, as well as some other things.

对吧？

Right?

因此，孩子在学习语言过程中所拥有的更广泛的经验。

So the the broader set of experiences that the child has as they're learning language.

对吧？

Right?

这有点像构建一个关于周围世界的模型，使得他们所学习的内容不仅仅是任意的词语序列，而是真正与有意义的事物相对应的。

Sort of like building a model of the world around them that means that the things that they're learning aren't just sort of arbitrary sequences of words, but actually things that map onto things that are meaningful.

对吧？

Right?

所以你的大型语言模型必须仅从它所生成的词语序列中，推断出它对世界的所有认知。

So your large language model has to figure out everything that it knows about the world just from the sequences of words that it's saying.

而且，孩子所获得的那些东西，不仅仅源于那些进化带来的约束，也不仅仅源于他们更广泛的经验，还包括他们通过使用语言来产生周围世界中理想结果而获得的内容——也就是说，把语言当作一种工具，而不仅仅是用来预测的东西，对吧？

And also, the kinds the of things that a child is getting, you know, not just as a consequence of whatever those evolved constraints are and not just as a consequence of their their broader experience, but also, you know, the the things that they're getting from, being able to, engage in using that language to produce, you know, desirable outcomes in the world around them, right, using it as a a tool, not just something that you're necessarily learning to predict.

我们在训练大型语言模型的最后阶段，确实也做了一点类似的事情。

And we do a little bit of that in our training of large language models at the end.

有一些基于强化学习的微调之类的方法。

There's some sort of fine tuning about reinforcement learning and so on.

但我认为，对于认知科学家来说，一个非常有趣的研究方向是：去刻画我们所面临的这种差距。

But but but I I think that's a really interesting kind of project for cognitive scientists is thinking about, yeah, how to characterize that that gap that we have.

对吧？

Right?

并把这些模型当作工具，来探索和解决这个问题。

And using these sorts of models as a tool for working that out.

因此，在我的实验室里，我们做了一些工作。

And so in in my lab, we've done a bit of work.

这最近是与现在在耶鲁大学的汤姆·麦考伊合作，研究一种称为元学习的神经网络训练方法。

This is most recently with Tom McCoy who's, now at Yale, looking at an approach that's called meta learning, for training neural networks.

元学习的目标是通过调整神经网络的初始权重，使其能够从更少的数据中学习。

And what meta learning does is it tries to create neural networks where we manipulate the initial weights that the neural network has in such a way that it's able to learn from less data.

因此，这种方法的原理是

And so the way that

这在某种程度上提供了一个起点。

this works a head start in a little sense.

是的。

Yeah.

没错。

That's right.

但同时也更接近于捕捉这些归纳偏差，是的。

But and also coming closer to capturing those sort of inductive biases and Yeah.

先验分布。

Prior distributions.

对吧？

Right?

所以它的原理是，你可以说：我有一系列不同的学习任务需要解决。

So so the way that it works, you say, I've got a bunch of different learning problems I want to solve.

在语言学的情况下，你可以想象成：我要学习很多种不同的语言。

In the linguistic case, you could think about this as I'm gonna want to learn lots of different languages.

我想学习英语。

I'm gonna want to learn English.

我想学习韩语。

I'm gonna want to learn Korean.

我想学习乌尔都语。

I wanna, you know, learn Urdu.

我想学习你希望掌握的每一种语言。

I wanna learn, you know, each each of the languages that you you want to be able to learn.

你知道你只能依靠这五年的输入数据来学习这些内容。

And you know that you're only gonna be able to learn those from, you know, your five years of input.

对吧？

Right?

你会想，我应该在神经网络中设置怎样的初始权重，才能帮助这个网络仅凭这五年的输入数据，利用它自身的学习机制，学会每一种语言？

And you say, what are initial weights that I can put in my neural network that are going to help that neural network learn each of these languages from that five years of input, you know, just using the sort of mechanisms that it has for for learning?

我们使用一种叫做模型无关元学习的算法来解决这个问题，这个算法包含一个外层循环和一个内层循环。

And so the the way that we solve that problem using an algorithm that's called model agnostic meta learning is you you have a learning process which has an outer loop and an inner loop.

内层循环就是学习某一种具体的语言。

And the inner loop is just learning an individual language.

你只是将神经网络的权重从初始值调整，以便学会每一种语言。

So you're just adjusting the weights of the neural network away from those initial weights to learn the the each of those languages.

但外层循环要问的是：当我观察自己在所有这些语言上的表现时，如果我改变初始权重，我的表现会如何变化？

But the outer loop is saying, when I look at my performance across all of those languages that I want to learn, how how how does my performance on those languages change when I change my initial weights?

对吧？

Right?

因此，你可以通过尝试找到那些能提升所有语言表现的初始权重来学习这些初始权重。

And so you can actually learn the initial weights by, know, trying to find initial weights that that help to improve performance across all of the languages.

通过这样做，你就在为神经网络找到一个起点，这个起点能让它们在有限的数据下快速学习，对吧？

And so by doing that, you're finding a starting point for your neural networks, which is one that's going to allow them to learn quickly, right, from the limited data that they're getting.

然后我们可以回头看看这一组初始权重。

And then we can go back and we can look at that set of initial weights.

我们可以问，这能告诉我们人类学习者可能具有哪些偏见？

And we can say, oh, what does that tell us about the biases that human learners might have?

对吧？

Right?

你知道，为了能够从我们实际获得的数据量中学习语言，你需要具备什么样的偏见？

You know, what's what are the kind of biases that you need to have in order to be able to learn language from the amount of data that we actually get?

所以，我对这方面的有限了解可以追溯到AlphaGo和AlphaZero，那是下棋的程序。

So my limited knowledge of this stuff goes back to, you know, AlphaGo and, AlphaZero that was the chess playing program.

我听说，这些程序如果从未接触过人类棋手和围棋手，而只是自己学习，表现会更好。

And I I'm I'm told that those programs did better if they never were exposed to human chess players and Go players and just learned it themselves.

所以，与此类似，你这种通过特定初始化方式给模型一个起步优势的捷径，会不会让它们在某种程度上缺乏创造力呢？

So is there a worry analogously that your version where you can sort of do a little bit of a shortcut to give the models a head start, by by by initializing them in a certain way, will that make them less creative in some way?

我认为这会让它们更不愿意寻找不同于人类解决方案的路径。

I I think it will make them less inclined to find solutions that are not like the human solutions.

对吧？

Right?

这既有好处，也有坏处。

And that's a plus and a minus.

固定答案。

Frozen answer.

是的。

Yeah.

没错。

Exactly.

是的。

Yeah.

没错。

That's right.

因为人类在很多方面其实很糟糕。

Because there are lots of things that humans are really bad at.

对吧？

Right?

我们或许希望打造能够弥补人类短板的AI系统，去完成那些人类不擅长的事情。

And we might wanna be able to make, you know, AI systems that can complement us by being able to do the things that people are bad at.

我认为，这是一种很好的思维方式，可以设想一个人类与AI共存的未来，这种共存对每个人都有好处。

And that's a really, I think, good way of thinking about what a possible future is where humans and AI get to exist side by side in a way which is sort of good for everybody.

我认为，让AI系统具备更贴近人类的归纳偏差，一个很好的优势是，这不仅能让神经网络用更少的数据进行学习，还可能改善训练模型时涉及的能耗等问题。

I think the thing that might be quite good about making AI systems that have inductive biases that are more aligned with people is that it it's going to not only make it possible for those neural networks to learn from less data, and so some of the energy concerns and so on that are involved in training those models might get better.

而且，当你给它们提供五千年的数据时，它们或许还能学会更令人惊叹的东西。

And maybe they'll actually be able to learn even more impressive things when you give them the five thousand years of data.

对吧？

Right?

但这同时也意味着这些系统对人类来说会更有意义。

But it also means that those systems are going to make more sense to humans.

所以，我认为人类思维与当前AI系统之间的两个主要差异之一是归纳偏差，也就是从少量数据中学习的能力。

So one of the kind of the the two things I would say are the the big differences that we see between human minds and our current AI systems are one of these is about an inductive bias, right, ability to learn from small amounts of data.

另一个差异是泛化能力——一个AI系统可能在解决某个问题上非常出色，但在面对邻近问题时却会彻底失败。

And the other is about generalizability where you can have an AI system that's very good at solving one problem and then fails quite spectacularly on a problem that's right next to it.

嗯。

Mhmm.

对吧？

Right?

我想我们都经历过这样的情形：某个AI系统看起来非常聪明，却做出一些非常奇怪的事情。

I think we've all had that experience of, like, you know, some the AI system seems very smart and then does something very weird.

对吧？

Right?

这已经被一些人称为‘破碎的智能’。

And this has been called jagged intelligence by, you know, various people.

而且，你知道，这对人工智能研究者来说是一个有趣的前沿领域，他们正试图弄清楚这些参差不齐的边界是什么样子，以及如何让我们的系统变得更好。

And, you know, that's a that's a sort of interesting frontier for AI researchers trying to figure out is how do we understand what those jagged boundaries look like and how do we make our systems better.

我认为认知科学实际上是非常好的工具，可以帮助回答这类问题。

I think cognitive science is actually a really good tool for trying to answer those kinds of questions.

但可能发生的情况是，如果我们创建出具有更接近人类归纳偏见的人工智能系统，那么当它们在所获得的数据上进行训练时，找到的解决方案也会更类似于人类找到的方案。

But but one thing that might happen is that if we create AI systems that have inductive biases that are more similar to people, then the solutions that they're going to find when they're trained on the data that they get will be more like the kinds of solutions that humans find too.

对吧？

Right?

所以你可以把你的AI系统想象成一块白板。

So you can kind of think about it as your AI system is your blank slate.

好的。

Okay.

它需要数千年的语音数据才能达到你五岁孩子所达到的水平。

It takes, you know, five thousand years of speech to get it to the point where your five year old gets to.

但那可能并不是同一个点。

But that might not actually be the same point.

对吧？

Right?

从外部看，它们可能看起来有点相似，是的。

It might, from the outside, look kinda similar Yeah.

因为它们都在很好地使用和生成语言。

In that they're both doing a good job of sort of using and producing language.

但在内部，可能非常不同。

But on the inside, it might be quite different.

而且它找到了一条奇特的路径，使其能够很好地使用语言。

And there's some weird path that it's found that gets it to the point where it's able to do a good job of using language.

但从我们的角度来看，这确实是一种非常奇特的解决方案。

But it is kind of just a very weird solution from our perspective.

事实上，我们实验室的一些分析表明，情况确实如此。

And in fact, some of the analyses we've done in my lab suggest that that's the case.

因此，如果我们能利用归纳偏差将模型引导至更接近人类的解决方案，这些方案可能也会让我们觉得更合理。

And so if we can use inductive bias to nudge the models towards more human like solutions, they're probably gonna be things that make a little more sense to us as well.

我想多听听你刚才提到的那句插话。

Well, I wanna hear more about that little parenthesis you just said.

我的意思是，我一直以为，生成听起来像人话的句子时，大语言模型的内部机制与人脑的运作方式截然不同。

I mean, I I always presumed that the internal machinations of the LLMs that output a human sounding sentence were very, very different than what goes on in an actual human brain.

那么，我们对这一点了解多少呢？

So what do we know about that?

是的。

Yeah.

所以，我实验室里有一些研究例子，揭示了这种奇特之处。

So, some examples of things that, we've we've done in in my lab that sort of reveal some of this weirdness.

其中一个例子是，大语言模型对它们所生成输出的概率非常敏感。

One of them is that, large language models are very sensitive to the probabilities of the outputs that they're producing.

对吧？

Right?

所以，当人们刚开始对这些模型感到兴奋时，有一篇名为《AGI的火花》的论文指出，GPT-4展现出了一些非凡的能力。

So when people were very excited about these models, there was the the paper, the sparks of AGI paper that came out that said, you know, GPT four sort of exhibits these remarkable abilities.

汤姆·麦科伊和一些同事写了一篇论文，我们称之为‘自回归的余烬’，意思是，尽管顶部会迸发出火花，但底部仍然存在一些余烬，这是这些模型训练方式的结果。

Tom McCoy and and some some colleagues, we we wrote a paper that we we called embers of autoregression, which was saying, much as you're getting sparks at the top, there are still these embers at the bottom, which are a consequence of the way these models are trained.

其中一点是，如果你——要知道，在现代系统中，人们已经用了各种技巧来绕过这个问题。

And so one of these is that if you, you know, again, these are things that in modern systems, there's all sorts of tricks that they've used to sort of get around this.

但如果你拿一个像GPT-4那样的原始语言模型，让它解决一些简单的问题，比如计算字符串中字母的数量，它们的表现会受到它们必须生成的答案的概率影响。

But if you sort of take a raw language model of the kind that we were getting with GPT-four, and you ask it to solve sort of simple problems like counting the number of letters that appears in a string, how well they do on that is influenced by the probability of the answer that they would have to produce.

例如，它们计算含有30个字母的字符串要远比计算含有29个字母的字符串表现更好，因为数字30在互联网上出现的频率高于29。

So for example, they're much better at counting strings that have 30 letters in them than strings that have 29 because the number 30 appears on the internet more often than the number 29.

所以这是一种情况：存在一些相近的、相当不错的答案，而其中一些答案的概率更高。

So it's a situation where there are other nearby answers that are pretty good, and some of those have higher probability.

因此，模型最终生成的是概率更高的答案，而不是它本应生成的那个正确答案。

And so as a consequence, it sort of produces the high probability thing rather than the thing that it's supposed to produce.

这可以说是语言模型的一种奇特的偏见。

And so that's a, you know, sort of like weird idiosyncratic bias of language models.

这是由它们的训练方式所导致的。

It's a consequence of the way that they're trained.

因此，更广泛地说，我对这些系统的看法是，我们应该预期这种现象正是我们从计算层面视角所看到的。

And so more generally, the way that I think about these systems is that, you know, we should expect this is this is applying our computational level lens.

对吧？

Right?

我们应该预期智能系统的行为会受到它们所要解决的问题类型的影响。

We should expect intelligent systems to behave in ways that are shaped by the kinds of problems that they're trying to solve.

当我们设计人工智能系统时，我们明确地选择了它们将要解决的问题类型。

And when we design our AI systems, we're making explicit choices about the kinds of problems that they're going to solve.

对。

Right.

比如预测序列中下一个词或标记的能力。

Things like being able to predict the next word or token that appears in a sequence.

这将会影响它的行为。

And that's gonna be something which influences its behavior.

因此，如果我们训练系统的目标函数与人类大脑在进化过程中所解决的计算问题之间存在差异，那么我们预期它们找到的解决方案也会大不相同。

And so to the extent that there's a difference in the objective function, right, the thing the the goal that we have in training that system and the kinds of computational problems that human minds have evolved to solve, then we're going to expect the kinds of solutions that they find to look quite different.

而这正是我们行为出现不匹配的部分原因。

And that's part of where we get this mismatch in behavior.

你之前提到过一个很有争议的观点，我还没来得及深入探讨，是关于神经网络或神经元的几何结构或空间结构，以及它在思维规律中所起的作用？

You mentioned earlier this provocative thing, which I haven't had a chance to follow-up on about, was it the geometry or the spatial structure of neural networks or or neurons more generally and and the role that that plays in the laws of thought?

所以第三个线索是，我们之前讨论了逻辑和概率理论。

So the the third thread so we talked about logic and probability theory.

对吧？

Right?

所以这里的第三个线索是，把思维看作空间中的点，运用我们用来思考空间的数学方法。

So the third thread here is this kind of idea of thinking about, yeah, thought in terms of, you know, points in space, right, and using the sort of math that we use for thinking about spaces.

这个想法是在二十世纪发展起来的。

And that was an idea that developed in the twentieth century.

对吧？

Right?

所以我说，你知道，纽厄尔和西蒙、乔姆斯基已经证明了逻辑在为我们提供关于思维或语言如何运作的理论表达方式方面非常有效。

So I said, you know, we had Newell and Simon and Chomsky demonstrating that things like logic were really effective for giving us sort of ways of expressing theories about how something like thought or language might work.

但那是五十年代的主要理念，并由此延续下去。

But then, you know, that was sort of like the the big idea of the nineteen fifties and carried forward from there.

到了七十年代，心理学家们开始意识到，这种观点存在一些漏洞。

And then in the nineteen seventies, psychologists sort of started to realize that, you know, there are some some some gaps in this.

对吗？

Right?

于是，我们来到了这样一个问题：当你把这些数学理论（如逻辑）与人类行为进行严格对比时，会发生什么？

And so this is this is where we get to, okay, what happens when you take these mathematical theories like logic and then start comparing them rigorously against human behavior?

当你开始这样做时，就会发现这些有意义的差异。

And when you start to do that, you start to turn up, you know, these sort of meaningful discrepancies.

其中一组差异来自伊莱诺·拉什的研究，她是一位研究人们如何思考类别的心里学家。

And so one of these sets of discrepancies came from the work of Eleanor Rush, who was psychologist who explored how people think about categories.

对吗？

Right?

如果你从逻辑的角度来看类别，你会寻找一个能定义该类别的规则。

If you think about categories from the perspective of logic, you're looking for a rule that characterizes that category.

你寻找的是一种定义，它能明确告诉你成为该类别成员意味着什么。

You're looking for sort of like a definition that tells you, you you know, exactly what it is to be a member of that category.

你必须具备这些属性，不能具备那些属性，不管怎样。

You know, you have to have these properties, you have to not have these properties, whatever it is.

这就是定义，它告诉你成为某个类别的标准是什么。

That's that's the rule that tells you, you know, what it is for belonging to a category.

罗斯发现，实际上很少有人类类别具有这种结构。

And Ross showed that it really seems like, you know, very few human categories have that kind of structure.

如果你观察我们对‘什么构成家具’或‘什么构成交通工具’的直觉，你会发现，根本找不到逻辑学家所期望的那种明确界定标准。

So if you if you look at our intuitions about, you know, what makes something a piece of furniture or, you know, what makes something a vehicle, there aren't definitions that you can find that characterize those in the way that the the the logician would want you to have.

是的。

Yeah.

相反，这些类别似乎具有更模糊的结构，你可以肯定某些东西绝对是家具，比如一把扶手椅。

And And instead, it seems like they're characterized by a much more fuzzy structure where you can say certain things are definitely pieces of furniture, like a, you know, an armchair.

而其他一些东西，比如地毯，则可能是家具。

And other things are maybe pieces of furniture like a rug.

对吧？

Right?

而且从逻辑的角度来看，这种梯度很难被捕捉到。

And there's a sort of gradients that was hard to capture from that logical perspective.

因此，似乎需要一种不同的理论来捕捉这种现象。

And so it seemed like you needed a different kind of theory to be able to capture that.

于是心理学家开始思考，也许可以把物体看作空间中的点，用事物在空间中的接近程度来表征它们属于某个类别（比如家具）的程度。

And psychologists started to think about, well, maybe if you think about objects as points in space, then how close things are in in space as a way of characterizing, you know, their the extent to you know, you think about the you're you have another point that characterizes your category of furniture.

因此，扶手椅离那个点很近，而地毯则离那个点较远。

And so now, you know, an armchair is close to that point, and a rug is further away from that point.

这或许能帮助我们捕捉到这种梯度。

And maybe that gives us a way of capturing that that sort of gradients.

但接着你会遇到一个新问题：如果概念是空间中的点，那如何进行计算呢？

But then you end up with a new problem, which is if concepts are points in space, then how do you do something like computation?

对吧？

Right?

因此，逻辑学催生了数字计算机、图灵机等概念。

So with logic, that translated into the ideas behind digital computers, Turing machines, all of these things.

我们有了理解思维的方式，因为我们可以说：好吧。

We had a way of thinking about what thought was because we could say, okay.

如果你以逻辑方式表示某事物，那么，遵循这种无偏见的理念，我们就能制造出一台机器，它能执行我们的规则，并告诉我们结果是什么。

If you represent something logically, then, you know, fulfilling that idea of blindness, we can then make a machine that, like, executes our rules and sort of tells us, you know, what the consequences are.

但如果一个概念是空间中的一个点，那么接下来我们该往哪里走呢？

But if a concept's a point in space, then, you know, where do we go from there?

那我们该怎么做呢？

What do we do with that?

我们如何学习这些概念？又如何弄清楚它们的后果呢？

How do we how do we learn what those concepts are, and how do we, how do we sort of, like, work out what the consequences are?

而这个问题的答案来自神经网络。

And the answer to that came from neural networks.

人们从20世纪40年代起就开始思考神经网络了。

So people had been thinking about neural networks, you know, since the nineteen forties.

对吧？

Right?

麦卡洛克和皮茨最初做了一些工作，将布尔电路的思想进行了转化。

There was a sort of initial work by McCulloch and Pitts, which was translating the idea of a Boolean circuit.

对吧？

Right?

乔治·布尔所思考的那种逻辑结构。

The kinds of logical structures that George Boole had thought about.

他们提出了一种用神经元之间的运算来表达布尔电路的方法。

They came up with a way of expressing Boolean circuits in terms of operations between neurons.

对吧？

Right?

你可以以某种方式连接神经元，使它们能够表示逻辑与、或、非等操作。

You could connect neurons up in such a way that they could represent logical ands and ors and nots and so on.

由此，你可以构建出复杂的神经网络结构。

And from that, you could build sort of complex neural circuits.