Max Bennett (on the history of intelligence)

他们还拥有强大的技术，用来支持市场营销。

And they have an incredible amount of technology to help both with their marketing.

所以他们在个性化营销方面做得非常出色。

So they do an incredible job personalizing marketing.

他们发的邮件，都是根据你在亚马逊上的所有行为来的。

The emails they send are based on all the things that you've done on Amazon.

他们的推荐系统特别聪明。

Their recommendations are incredibly intelligent.

而大多数其他品牌和零售商都没有这样的技术。

And so most other brands and retailers did not have that technology.

这会不会是两个方面的问题？

And is that a two prong issue?

第一，你确实没有这项技术。

One, you don't have the actual tech.

第二，你也没有用户行为的完整数据。

Two, you don't have the database of all their info.

这两个问题都让人工智能的挑战更大了，因为你得让AI系统在数据更少的情况下也能正常运作。

Both of those are problems, which makes the AI problem harder because you need to make an AI system work with less data.

所以Blue Core的目标就是帮助耐克、锐步这样的品牌去和亚马逊竞争。

And so what Blue Core was all about was helping brands like Nike, Reebok compete with an Amazon.

想想耐克居然像是弱者，像大卫对抗歌利亚，这挺有意思的。

Isn't it funny to think of Nike as like the underdog, the David and the Yeah.

还有哥伦比亚。

Columbia

这正是我最初对大脑产生兴趣的原因。

And so that was where my original interest in the brain came.

因为当你在研究机器学习系统时，就会遇到一个叫莫拉维克悖论的现象。

Because when working with machine learning systems, it begets this thing called Moravec's paradox.

汉斯·莫拉维克是一位计算机科学家，他提出了这样一个观察。

So Hans Moravec was a computer scientist who made the following observation.

为什么人类特别擅长某些事，比如打篮球、弹吉他、洗碗，可这些对机器来说却特别难呢？

Why is it the case that humans are really good at certain things like playing basketball or playing guitar or doing the dishes that is so hard for machines to do?

但人类觉得特别难的事，比如快速数数或者做算术，对电脑来说却轻而易举。

And yet things that are really hard for humans, like counting numbers really fast and doing arithmetic, are so easy for computers.

这传统上被称为莫拉维克悖论。

And this is classically called Moravec's paradox.

为什么会这样呢？

Like, why is that the case?

这正是我开始对大脑产生兴趣的起点。

And so that was the beginnings of my interest in the brain.

这本书从一个对我来说最引人入胜的假设开始：我们现在正在做的，就是让计算机拥有能与人类匹敌、超越人类，或者至少相似的智能。

The book starts with the most intriguing premise for me, which is what we're trying to do now in this phase of our technology is to get these computers to have intelligence that either matches ours or exceeds ours or is similar to ours.

但真正有趣的是，我们根本不知道人类的智能到底是怎么运作的。

But what's really funny is we don't know how our intelligence works.

所以我们试图复制一个连自己都不清楚为什么和怎么产生智能的系统。

So we're trying to replicate a system that actually we don't really truly know why and how we're intelligent.

我觉得这特别有意思。

And I think that's really fascinating.

你怎么可能复现一个你根本都不理解的东西呢？

How are you gonna recreate something that you don't actually understand in the first place?

所以这真是个特别吸引人的起点。

So it's just a very intriguing jumping off point.

接着你把这个故事拆开来讲。

And then you break this story.

这是两个相互交织的叙事。

These are coevolving narratives.

一个是人工智能的诞生和我们今天所处的位置，另一个是第一个大脑、第一种智能和第一个神经元的起源。

One is the birth of AI and where we're at today, and then one is the birth of the first brain and the first intelligence and the first neurons.

你把它分成五个很有帮助的阶段，展现这些重大突破是如何发生的。

And you break it up into five really helpful sections of evolution where these big leaps happened.

这样来组织内容，是你本来的想法吗？

Was that an intuitive way to frame this?

其实我最初根本没打算写一本书。

So my original intent was actually never to write a book.

我一开始只是想弄明白大脑是怎么工作的。

I started just by trying to understand how the brain works.

如果没有这份天真，这本书根本不会存在。我当时以为，买本神经科学教材，暑假读一读，就能搞懂大脑的运作了。

The whole book wouldn't exist without a huge dose of being naive because I thought that I was gonna buy a neuroscience textbook, take a summer and read it, and I'd understand how the brain works.

是啊，我本来也觉得会是这样。

Yeah, yeah, I would have thought Yeah, that

对，没错。

yeah, yeah.

读完一本教材后，我才意识到，天啊，我们根本不知道大脑是怎么工作的。

And after reading a textbook, I just realized, wow, we have no idea how the brain works.

于是我又买了一本教材。

So then I bought another textbook.

这个过程持续了一年半左右，不知不觉中我自学了神经科学，可我还是觉得没真正解开大脑运作的谜团。

And this process continued for about a year and a half until I had unintentionally sort of self taught myself neuroscience, and yet I still felt like I had not satisfied the itch of understanding how the brain worked.

所以我开始联系各种神经科学家，想和他们合作，但没人回我邮件——就像你说的，我没啥亮眼的学术背景。

So that led me to start reaching out to various neuroscientists because I wanted to collaborate with them, and no one would respond to my emails, to your point, lack of impressive academic Yeah,

我觉得我

I think I

该为这件事道个歉，

should apologize for that at

现在就该说。

this point.

不，不用在

No, not at

这上面道歉。

all.

确实如此。

It's true.

我本来以为会是哈佛、斯坦福这样的学校，结果

No, I was expecting Harvard and Stanford It's and

这已经是很好的学校了。

a very good school.

我还以为你要说的是亚利桑那州立大学呢。

I thought you were gonna say, like, Arizona State.

你最近可真是一门心思地夸亚利桑那州立大学啊。

Hey, you're on a real run about buying that Arizona State.

太阳恶魔体育场，还有那片飘在校园周围的雾气。

Sun Devil Stadium, Flaky

杰克斯，

Jakes,

听我说，去

listen, have go

漂浮在校园四周的那层薄雾。

a floaty thing around the whole campus.

埃莱西河。

Elysee River.

我觉得我们直接这么说也没关系。

I think it's okay for us to say that anyway.

不好意思。

Sorry.

不是，真的不是。

So, no, no.

抱歉拉你下水。

Sorry to drag you through the mud

针对她。

on her.

所以没人搭理我。

So no one would respond to me.

于是我决定这么做，因为我开始冒出一些跟进化没关系的想法。

So what I decided to do, because I started coming up with these ideas which were not based on evolution.

最初的想法其实是想逆向工程大脑里一个叫新皮层的部位，这个我可以展开说，因为里面有很多特别有趣的东西。

Actually, original ideas were trying to reverse engineer this part of the brain called the neocortex, which I can get into because there's a lot of fascinating things with that.

所以我提出了一个关于它可能如何运作的理论。

And so I had an idea for a theory of how it might work.

所以我决定，唯一能让人们回应我的办法，就是把论文投给科学期刊。

So I decided the only way I was going to get people to respond to me was I'm going to submit it to a scientific journal.

因为他们肯定会拒稿，毕竟我是谁呢？

Because they'll reject it, obviously, who am I?

但他们至少得读一遍，才能告诉我哪里有问题。

But they'll have to at least read it to tell me what was wrong with it.

而正是从这里，这段不可思议的旅程开始了——让我没想到的是，论文居然被接受了。

And this is where this whole unlikely journey began, because to my surprise, it actually got accepted.

他写了一个理论，经过同行评审后发表了。

He wrote a theory, and it was peer reviewed, and it was published.

我还中了审稿人的‘彩票’。

And I won what's called the reviewer lottery.

碰巧，其中一位审稿人是位非常著名的神经科学家，叫卡尔·弗里斯顿。

Just by luck, one of the reviewers is a very famous neuroscientist neuroscientist named Carl Friston.

他后来成了我的非正式导师。

He became sort of an informal mentor of mine.

然后，这件事就像雪崩一样，让我走上了一条自我驱动的学术道路。

And then sort of this started cascading into this self directed academic pursuit.

那是多久以前的事了？

How many years ago was that?

四年前的事了。

That was four years ago.

天啊，这进展也太快了吧。

My God, this is so accelerated.

你今年多大了？

How old are you?

我34岁。

I'm 34.

哎呀，糟了。

Oh, no.

我是不是该更年轻一点，或者更年长一点？

Was I supposed to be younger or older?

哦，大得多呢。

Oh, way older.

斯坦福啊，年纪大了。

Stanford, old.

比我还年轻，这让我有点受不了。

Younger than me is upsetting.

是啊，你啊

Yeah, you're

是个失败者。

a loser.

对。

Yeah.

大男人了。

Big boy.

哦，这可真不容易。

Oh, that's rough.

三十四岁。

34.

哇，那你当时三十岁。

Wow, so you were 30.

不过你现在已经发表了几篇文章了。

But you've had a few articles published now.

你那个独特的理论到底是什么？

What was your novel proprietary theory?

关于新皮层，很大程度上我觉得这种做法其实效果不会太好，所以我才转向了进化论的方向。

So with the neocortex, and in large part, don't think this approach is actually going to work that well, which is why I pivoted to the evolutionary one.

但总体来说，这个理论稍微复杂一点——新皮层特别有意思，如果你记得大脑的图像，所有那些褶皱，整个部分都是新皮层。

But the broad theory, it's a little nuanced, but the neocortex is so fascinating because if you look at a human brain, all the folds that you see, if you remember an image of a brain, that whole thing is neocortex.

它实际上就像一层薄膜。

And what it actually is, is this sort of film.

像是被揉皱后叠在一起的一张纸。

It's like a sheet that's bunched together.

我们一直以为，大脑皮层的不同区域负责不同的功能。

What we've always thought is different regions of the neocortex do different things.

你大脑后部的是视觉皮层。

So the back of your brain is your visual neocortex.

如果那里受损，人就会失明。

If that gets damaged, you become blind.

大脑皮层有个区域，一旦受损，人就无法说话。

There's a region of neocortex that if it gets damaged, you can't speak.

还有一个区域，如果受损，人就无法行动。

It's a part of neocortex that if it gets damaged, you can't move.

所以人们自然会觉得，这根本不是单一的结构。

And so one would think that this isn't really one structure.

它实际上是由许多不同的部分组成的。

It's actually a lot of different structures.

每个部分都有独特的结构，来支持视觉或听觉等功能。

Each have a unique structure that would facilitate the scene or the hearing.

没错。

Exactly.

对。

Yeah.

最让人震惊的是，如果你在显微镜下观察，大脑皮层各个部位的结构完全一样。

And what is mind blowing is if you look under the microscope, the neocortex looks identical everywhere.

什么？

What?

可就这么一个结构，却能完成这么多不同的功能。

And somehow this one structure does all of these different things.

就是从那时起，我开始对这个话题产生了浓厚的兴趣。

And this was when I started becoming really fascinated with this topic.

而且这些功能还能被重新分配、转移，对吧？

And can be relocated and moved, right?

你可以让其他区域接管任务，大脑会自行调整。

You can co opt areas It's of the to change.

如果某个区域受损了，大脑可以把这项功能转移到另一个新皮层区域来完成。

Well, if one area is destroyed, you can relocate into another section of neocortex to perform that task.

你居然还记得那本书，真让人佩服。

You clearly remember the book, I'm impressed.

哦，对，想起来了。

Oh, right.

我们稍后会谈到这些皮层柱。

Well, we'll get into the columns at some part.

最能证明这一点的是一项非常著名的研究，研究人员把雪貂的视觉信号从视觉皮层重新引导到了听觉皮层。

So the best study that demonstrates this, it's a very famous study where they took ferrets and they rerouted their visual input from the visual cortex into their auditory cortex.

如果这些区域真的有本质区别，那它们本不该能正常‘看见’东西。

So if it is the case that these regions are different, then they shouldn't be able to see appropriately.

但结果发现，它们看东西几乎和正常时一样好。

But what they found is they could see pretty much just fine.

这些原本负责听觉的区域，就这样被重新用来处理视觉信息了。

And these areas of auditory cortex were just repurposed to process vision.

这也解释了为什么中风后，人们还能重新恢复某些能力。

And this is also why after strokes, you can regain abilities.

受损的大脑皮层区域并不会真正再生。

The region of neocortex that's damaged doesn't actually grow back.

也没有修复过程，只是功能重新分配。

And there's no repair, it's just relocation.

没错。

Correct.

稍等一下。

Take a second.

大脑皮层真是了不起——现在看来它的作用太关键了。

What a fucking thing the Neocortex- It's important use now.

我和你，我的意思是。

Me and you, I mean.

Max 已经出了四年名了。

Max is known for four years.

我知道我又要说认识多久了，但其实我就是刚学，之前那都是装谦虚。

I know I'm gonna say how long I'm known, but we'll just say, yeah, I'm just learning now because was a faux humble.

这么说，这真勾起了你的兴趣。

Okay, so that really sparked your curiosity.

那你基于这个写了哪篇论文？

And what paper do you write based on that?

我写了一篇论文，想弄清楚皮质柱到底是怎么回事——我们现在认为，大脑皮层其实是由许多重复的柱状结构组成的，也就是所谓的微电路。

So I wrote a paper that was trying to understand how the neocortical column, so what we now think, what the neocortical sheet actually is, is a bunch of repeating columns called a micro circuit.

但这里面有个大谜团：这个微电路到底有什么功能，让它能这么灵活地被重新利用？

And there's a big mystery to figure out what does this micro circuit do that makes it so repurposable?

目前有几种理论可以聊一聊。

And there's a few theories that we can go into.

但我特别想知道的是，它怎么记住序列？

But my question in particular was how does it remember sequences?

比如说，你听一段音乐，只听一遍，就能原样复述出来，这到底是怎么做到的？

So how is it the case that you can hear a sequence of music one time and you can it back to someone?

我提出了一种数学框架，用来解释我如何做到这一点。

There's a sort of mathematical framework where I theorize how am I do that.

这部分依然是个巨大的谜题，对吧？

And that part is still a great mystery, right?

我们现在也开始谈到记忆了。

We're getting into memory too.

记忆这个问题，目前进展到哪儿了？

Where is that at?

我们没法直接观察到它。

We can't observe it.

我们看不到任何东西被刻在神经元上。

We can't see anything etched into a neuron.

大概率是形成了新的连接，以及新的连接模式。

Presumably there's new connections and a new pattern of connections.

正是这种连接模式， somehow 记录下了我们听到的那段声音。

That pattern of connections somehow represents that sound we heard.

神经科学界普遍认为，记忆的物理基础在于突触权重的变化。

There's pretty strong consensus in the neuroscience community that the physical instantiation of memory is in changing the synaptic weights.

神经元之间的连接要么被削弱，要么形成新的连接，或者连接的强度发生改变。

So the connections between neurons either delete or more get formed or the strength changes.

但关键问题是，对于某个具体记忆——比如童年的回忆，或者如何做某件事的记忆——

But the big question is for a specific memory, like a memory of your childhood or the memory of how to do something.

它究竟存放在哪里，目前还是个未解之谜，因为似乎并没有固定在某一个地方。

Where that lives is more of an open question because it doesn't seem to live in any one place.

它更像是分布在大脑的多个不同区域之间。

It seems to be distributed across many different regions of the brain.

记忆其实包含多个组成部分，对吧？

And there's components to a memory, right?

比如情绪层面。

There's an emotional component.

可能还有气味、颜色，这些元素都来自大脑的不同区域，最终汇聚成一个完整的记忆。

There might be a smell, there might be a color, and those are all drawing from all areas of your brain to come together for this one memory.

我们很有信心，大脑和电脑是完全不一样的。

And one way we feel there's a lot of confidence that the brain is different than computers.

在电脑里，有一种叫作寄存器寻址的内存机制。

In computers, there's something that could be called register addressable memory.

你要想调出一段记忆，就得知道它在芯片上的具体地址。

The way you get a memory is you need to know the code for its location on a computer chip.

一旦丢了地址，记忆就找不回来了。

You lose the code, you lose the memory.

但大脑用的是内容寻址的机制——你只要想起一点点片段，整个记忆就会自动补全。

In the brain, there's something called content addressable memory, which is the way you remember something is you give it a tiny piece of the overall memory, then the rest is filled in.

天啊，原来如此。

Oh, wow.

这就是为什么闻到某种气味，突然就想起一段往事；或者你忘了怎么弹一首歌，但只要一弹第一个和弦，后面的旋律就自然而然涌出来了。

And this is why when you smell something, a memory pops back into your head or you forget how to play a song, but you start playing the first chord on a guitar and then the rest starts flowing to you.

好吧，我得翻翻我的笔记了，本来我不想这么干的，但事到如今也只能这样了。

Okay, so, and now I'm gonna have to consult my notes, which I try not to do, but here we are.

我们先聊聊大脑的复杂性，它有八百六十亿个神经元，连接数超过一千万亿个。

Let's first talk about the complexity of the brain, which is 86,000,000,000 neurons and over a 100,000,000,000,000 connections.

这真的很难想象。

I mean, that's hard to really comprehend.

在每一立方毫米的脑组织里——就像硬币上一个字母那么小的区域——就有一百亿个神经连接。

Within a cubic millimeter, so if you look at a penny, one of the letters, there is over 1,000,000,000 connections in the brain.

不可能吧。

No.

所以，我写这本书的思路是，不急于去逆向破解大脑的工作原理，然后照搬给人工智能。不如从最基础的开始，一步步往上走。

It's boggling, Okay, so the way you decide to lay out the book is instead of trying to reverse engineer how the brain works so that we can apply it to how the AI should mechanistically work, Let's start at the very beginning, and then just take the ride up the ladder.

因为从进化角度看，所有过去出现过的生物——从单细胞生物到动物——那些能力都还在我们身上保留着，对吧？

Because everything that's already happened in evolutionary terms to previous iterations, animals, amoebas, it's all still here, right?

我觉得，这个观点可能会让很多人感到震惊。

I think that'd be the first concept that might shock people.

六亿年前发生的事，我们今天依然保留着，之后才一层层叠加了新的功能。

Virtually everything that happened six hundred million years ago, we still have that, and then we've added things on top of that.

进化受到很多限制。

Evolution is very constrained.

每一次进化步骤——也就是下一代——都没法从头重新设计。

So every iteration of an evolutionary step, in other words, another generation, it can't redesign things from scratch.

它只能在现有的材料上做调整。

It has to tinker with the available building blocks.

所以你看，我们的大脑跟鱼其实没太大差别，尽管我们和鱼类的共同祖先早在五亿年前就分道扬镳了。

And that's why you see that our brains aren't that different than fish, even though our common ancestor with fish is 500,000,000 years ago.

因为一旦这些结构定型了，进化就很难反过来把它们彻底推翻重来。

Because once a lot of these structures are in place, it's very hard for evolution to find solutions that reverse engineer them.

就连鸟的翅膀，其实也是从手臂改造过来的。

So usually, just tinkers with even bird wings are repurposing of arms.

它不是重新发明了翅膀，而是慢慢把手臂改成了翅膀。

It didn't just reinvent wings, repurposed an arm over time.

你甚至可以追溯到，鳞片变成了脚，脚又慢慢变成了翅膀。

You even go back into say scales become the feet and the feet become the wings.

这一切都可以追溯到一条连续的线索。

That's all one steady line we can follow back.

那么，我们所谓的大脑，最早的样子是什么样的呢？

So what is the first version of what we might think of as a brain?

哪种动物是第一个拥有神经元的？

What's the first animal that has neurons?

有意思的是，在大脑出现之前，神经元就已经存在了。

Okay, so interestingly, there were neurons before there was a brain.

比如海葵，顺带一提，因为我大部分知识都是自学的，刚进入神经科学领域时，我经常把单词念错，这还挺搞笑的。

So if you look at a sea anemone, which side note, because most of my learning was self taught, One of the most funny parts of me entering the neuroscience community is I pronounce words wrong

一直如此。

all time.

天哪。

Oh, shit.

哦，太好了。

Oh, Good.

我也是。

I do too.

对。

Yes.

因为你只在书上见过这个词。

Because you've only seen it in print.

因为你只在书上见过这个词。

You've only seen it in print.

但这个我真的没理由辩解。

But this one's actually I don't have a good excuse for.

我以前一直念成anemone，直到朋友问我：你没看过《海底总动员》吗？

I used to say anemone until my friend was like, did you ever watch Finding Nemo?

不应该是anemone。

It's not anemone.

是anemone。

Anemone.

海葵、水母这些生物。

A sea anemone, jellyfish.

它们有神经元，却没有大脑。

These are creatures that have neurons but no brain.

它们拥有一种叫做神经网的结构。

And they have what's called a nerve net.

它们的皮肤上布满神经元网络，负责执行反射动作，但并没有一个中心区域来做决策。

Their skin has a web of neurons that are implementing a reflex, but there's no central location where decisions are made.

每一条触手基本上都是独立运作的，互不干扰。

Each arm or tentacle is independent largely from the other tentacles.

第一个问题是，神经元是怎么进化出来的？

The first question is why did neurons evolve?

真菌和动物之间有一个特别有趣的差异，我觉得特别有意思。

And there's a really interesting diversion between fungi and animals that I find fascinating.

是的，这个话题太棒了。

Yeah, love this.

因为我们其实和真菌没多大区别，尽管它们看起来完全不同。

Because we are actually not that different than fungi, even though they look very different.

因为从根本上说，我们获取能量的方式都是通过呼吸。

Because at the core, the way we get energy is through respiration.

也就是说，我们吃植物或其他动物提供的糖分，吸入氧气，排出二氧化碳。

In other words, we eat sugar from plants or other animals and we breathe oxygen and release carbon dioxide.

我们把糖和氧气结合，转化成能量。

We combine sugar and oxygen to make fuel.

没错。

Exactly.

所以我们没法自己制造能量。

So we can't make our own fuel.

植物却可以自己制造能量。

Plants can make their own fuel.

它们只需要阳光和水，就能活得很好。

They just need sunlight and water and they're fine.

如果没有植物为我们制造糖分，我们就无法生存。

We can't survive without plants creating sugar for us.

所以早在进化早期，真菌和动物就采取了不同的方式来获取糖分。

So fungi and animals very early in evolutionary time took diverging strategies to get sugar.

真菌选择等待生物死亡后再利用。

Fungi took the strategy of waiting for things to die.

我想说一点

I just wanna say one

关于这一点，我觉得特别有意思。

thing on this because it's really fascinating.

这标志着捕食行为的诞生。

It's the birth of predation.

在那之前，地球已经有四十五亿年的历史了。

So prior to this and the Earth is four and a half billion years old.

而这件事发生在仅仅几亿年前。

This is only hundreds of millions of years ago.

那时候活着的生物，只是活着，靠阳光获取能量，根本不吃别的东西。

Organisms that live, they just lived, and they got energy from the sun, and they didn't eat anything.

真菌是第一个必须依赖其他生物才能生存的生物。

Fungi is like the first thing that it's going to need other organisms to exist.

它得靠吃掉其他生物来获取养分。

It's going to consume other organisms for its source.

居然真的发生了这种事，简直太不可思议了。

That's really crazy that that even happened.

没错。

Totally.

虽然在氧气出现之前，可能已经有一些初级的捕食形式，但那种不需要氧气的呼吸方式效率低得多。

And so even though there might've been some minor forms of predation before oxygen was around, what's called anaerobic respiration is much less energy efficient.

就是说，没有氧气的情况下进行呼吸。

So trying to do respiration without oxygen.

但氧气一出现，通过吃掉其他生物的糖分，就能获得多得多的能量。

But when oxygen came around, you can get way more energy by consuming the sugar of others.

这正是捕食行为真正诞生的时候。

And that's really when you got the birth of predation.

真菌靠无处不在的孢子生存。

So fungi, they survive by having fungal spores all over the place.

它们时刻存在于我们周围。

They're around us all the time.

一旦有生物死亡，它们就会迅速生长繁殖。

And when something dies, it then flourishes and grows.

这就是为什么面包放久了，到处都是真菌孢子，接着那些恶心的菌丝就开始冒出来。

And this is why when you leave bread out, there's fungal spores everywhere and we'll just start eating them, the gross filaments emerge.

动物走的却是另一条路。

Animals took a very different route.

它们的方式是主动杀死其他生物。

Their route was to actively kill other things.

因此，神经元在这里变得至关重要——即使最早的动物，我们也不知道它们具体长什么样，虽然有各种推测，但目前最好的研究模型是海葵。

And so this is where neurons are really important because even the very first animals, and we don't know what the first animals actually looked like, there's theories, but the best model organism is a sea anemone.

它们大概一直待在原地，长着触手，等着猎物游过，然后一把抓住，送进肚子里。

They probably sat in place, they had tentacles, and they waited for things to fly by, and then they would just capture them and bring them into their stomach.

它们就等着。

They just waited.

它们看不见，听不到，闻不到，什么感官都没有。

They can't see, they can't hear, they can't smell, nothing.

没有任何信息传进来，只能指望有东西撞上来，那就行了。

There's no data coming existing and hopefully things will bump into them and they'll Right, be

它们有触觉这种信息。

they have data which is touch.

只要有什么碰到触手，它们就会立刻拉进来。

So like if something touches their tentacle, then they pull it in.

这就是我们拥有神经元的主要原因之一。

So that's one of the main reasons we have neurons.

说到这点，也印证了你关于进化受限的看法，有很多共通之处。

And with that, there's a bunch of commonalities to your point about how evolution is constrained.

我们的神经元和水母的神经元其实没多大区别，这简直让人难以置信——我们和水母之间，智能的基本单位几乎没有差别。

Our neurons are not that different than jellyfish neurons, which is kind of mind blowing, which is the unit of intelligence between us and a jellyfish is not really much different.

真正不同的，只是这些神经元是怎么连接在一起的。

What's different is just how it's wired together.

神经元有一些基础特性，比如分兴奋性和抑制性两种。

Some of the foundational aspects of neurons, which are they come in excitatory and inhibitory versions.

有些神经元会抑制其他神经元的兴奋，另一些则会促使其他神经元兴奋。

So some neurons inhibit other neurons from getting excited and other neurons excite other neurons from getting excited.

这种现象在海葵身上也能看到。

You see that in a sea anemone.

那为什么海葵也需要这种机制呢？

Why does that exist in a sea anemone?

因为即使是简单的反射，也需要‘如果是这样，就别那样’的规则。

Because even in a basic reflex, you need to have an if this, not that sort of rule.

当我张开嘴去抓食物时，必须让闭嘴的肌肉放松下来。

If I'm going to open my mouth to grab food into it, I need to relax the closed mouth muscles.

如果我要合上嘴，就得让张嘴的肌肉放松。

If I'm going to close my mouth, need to relax the open mouth muscles.

所以你需要这种神经元之间的抑制逻辑——一组神经元抑制另一组。

So you need this sort of logic of one set of neurons inhibiting another set of neurons.

这非常二元化，对吧？

It's very binary, right?

就像非黑即白，要么兴奋，要么抑制。

It's like on or off, exciting or calming.

这跟计算机直到量子计算出现之前的运作方式，倒是很像。

And that's a fun parallel with just how computers up until whatever quantum happens, I guess.

但这也正是计算的基础，对吧？

But that is also the basis of computing, right?

就是一和零。

It's ones and zeros.

就是开和关。

It's on and off.

这就是二进制。

It's binary.

在计算理论和计算神经科学领域，关于神经元和计算机在多大程度上以相似方式编码信息，一直存在丰富的讨论。

There's a really rich history in computational theory and computational neuroscience about the degree with which neurons and computers encode things similarly.

这方面争议很大。

There is a big debate around that.

我们已经知道，有些神经元是通过放电频率来编码信息的。

Some neurons we know encode things in their rates.

所以它并不是简单的‘是’或‘否’，而是放电的快慢。

So it's not really ones or zero, it's how fast it's firing.

压力就是这样的例子。

So pressure is like this.

你能感受到压力大小的不同，是因为大脑接收到了你手指上某个感觉神经元的放电频率变化。

So the reason you can tell the difference in pressure is because your brain is picking up the rate that a specific sensory neuron in your finger is firing.

但我们也知道，大脑的其他部分显然不是只靠放电频率来编码信息的。

But we also know other parts of the brain are clearly encoding things in a different way than just rates.

大脑的其他部分似乎更像计算机那样处理信息。

Other parts of the brain seem to be encoding things more like a computer.

所以这一切是怎么整合在一起的，至今还是个大谜团，但大脑很可能两种方式都在用。

And so it's a big mystery sort of how it all comes together, but the brain's probably doing both.

明白了，真有意思。

Okay, fascinating.

当生物开始捕食其他生物时，有些生物只是被动等待，等着食物自己撞上来再吃掉；但有些生物却不一样，它们决定主动出去寻找食物。

So now as organisms are going to consume other organisms and there are versions just waiting to get lucky and have food bounce into them and then eat it, somehow some new organisms develop where they decide, no, we're gonna go in search of the food.

这就是第一个大脑的由来。

And this is the first brain.

第一个大脑。

The first brain.

而这个最早的大脑，核心功能就是移动和导航。

And so the first brain is all about locomotion and steering.

没错。

Bingo.

对，给我们详细讲讲。

Yeah, break that down for us.

有趣的是，学术界把这个算法叫做趋化性导航。

One thing that's fascinating about this is this is an algorithm that the academic community would call taxis navigation.

我觉得用‘定向移动’这个词更简单易懂。

I think steering is a simpler word for it.

这种算法甚至存在于单细胞生物身上。

And this algorithm exists in single celled organisms.

那细菌是怎么找到食物的呢？

So how does a bacteria find food?

细菌没有眼睛，也没有耳朵。

Bacteria has no eyes or ears.

它没有复杂的感官器官。

A bacteria doesn't have complex sensory organs.

它身上只有一种非常简单的蛋白质装置。

All it does is it has a very simple sort of protein machine on its skin.

如果它朝着喜欢的方向移动，就会一直往前走。

And if it's going in the direction of something that it likes, it keeps going forward.

如果察觉到喜欢的东西浓度下降，就会随机转向。

And if it detects the decrease in something that it likes, it turns randomly.

这利用了物理世界中的一个现象：我在培养皿里放一点食物，食物释放的化学物质会形成一股扩散的气味云，越靠近食物，气味浓度就越高。

And this takes advantage of something in the physical world, which is I place a little food in a petri dish, the chemicals from the food make a plume around it, and the concentration of those smells are higher closer to the food.

所以，只要任何装置能检测到这种食物浓度在上升，一直朝那个方向走，最终一定能找到它。

So if I have any sort of machine that detects an increasing concentration of this food, if I just keep going in that direction, eventually I find it.

对。

Right.

如果你转个方向，发现浓度迅速下降，就知道方向错了。

And if you turn and it decreases very quickly, you'll know wrong direction.

再转一次，浓度又下降了，那就再转一次。

You turn again, oh, it decreases again, I turn again.

很快，你就能通过不断试错找到正确路径。

Really quickly, you're going to run through the trial and error of it.

没错。

Exactly.

就像玩热冷游戏一样。

It's like playing a hot cold game.

这套机制在单个细胞层面是可行的，但对像线虫这样由百万个细胞组成的生物来说——虽然对我们而言不算大，但在细胞尺度上已经非常庞大了——单个细胞那种小蛋白螺旋根本没法推动这么复杂的生物体。

So that works on a single cellular level, but on a large, not large to us, but from a single cell, very large scale of even a small nematode, which is very akin to the very first animals with brains 600,000,000 ago, you can't use the same machinery that a single cell uses, because they have these tiny little protein propellers that can't move an organism with a million cells.

于是，这套算法被重新在神经网络中实现了。

So this algorithm was recapitulated or recreated in a web of neurons.

同样的行为模式依然存在，只不过不再依赖细胞内的蛋白质级联反应。

So the same thing happens, but it doesn't happen in protein cascades within a cell.

而是靠线虫头部的一群神经元，检测食物气味浓度是否上升，并驱动它向前移动。

It happens with neurons around the head of nematode or worm like creature that detects when the concentration of a food smell increases and drives it to go forward.

另一组神经元则负责感知气味浓度下降，并随机转向。

And another set of neurons that detect it decreasing and turn randomly.

最妙的是，最早的脑部功能，就是要把所有这些输入整合起来，做出一个统一的决定。

And what's so cool is the point of the very first brain is you have to integrate all this input to make a single choice.

他们用线虫做过实验，这种生物特别适合研究早期大脑的样子，因为结构太简单了。

And they've tested nematodes, which are good organisms to experiment with what early brains were like because they're so simple.

你可以弄个培养皿，把一群线虫放在一边，中间放一条铜线。

And you can create a little Petri dish and you put a bunch of nematodes on one side and you put a copper line in the middle.

线虫特别讨厌铜，因为铜会破坏它们的表皮，这是个化学原因。

And nematodes hate copper for an isoteric reason that it messes with their skin.

然后在对面放点食物。

And put on the opposite side some food.

它们会冒险穿过铜线吗？

Does it decide to cross the copper barrier?

所以，最早的大脑就得做这种取舍。

And so the very first brain had to make these choices.

而这种决定取决于两个因素，其实都很合理。

And it depends on two things, which would make a lot of sense.

一个是两者的相对浓度。

One is the relative concentration of each.

所以食物越多，愿意穿越的线虫就越多，取决于它们有多饿。

So the more food, the more of them are willing to cross, and how hungry they are.

它们越饿，就越愿意穿越铜线。

The hungrier they are, they're more willing to cross.

是的。

Yeah.

如果另一个选择是死路一条的话。

If the other option's death at that point.

如果我吃饱了，就没必要冒险过去。

If you're full, I don't need to cross.

没错。

Exactly.

如果我快撑不住了，反正也没什么可失去的。

If I'm about to die, there's nothing at risk.

它们会因为铜线而死吗？

Are they gonna die anyway from the copper?

不是。

No.

铜只是让人不舒服。

The copper is just uncomfortable.

我明白了。

I see.

对。

Yeah.

真没想到。

Wow.

让他们变得烦躁。

Makes them cranky.

是的。

Yeah.

这是医学和科学上的正式说法。

It's the medical and scientific term.

想继续听《沙发专家》吗？你敢的话就别走开。

Stay tuned for more Armchair Expert, if you dare.

好的。

Okay.

听起来好像挺简单的。

So that sounds so simple.

我想聊聊扫地机器人，这个也很有意思。

I wanna talk about Roomba because that too is very interesting.

我觉得就属于这一类。

I think it's in this realm.

但刚一开头，我觉得你刚才说的其实有更深的含义——我探测到更多食物，就会朝它移动。

But just right out of the gates, I would say there's a much bigger implication to what you just laid out, which is I'm detecting more of that food, I move towards it.

探测到的少了，我就转向。

I'm detecting less, I turn.

无意中，这就造就了好的一面和坏的一面。

Inadvertently, that's created good and bad.

好坏的最根本基础。

The very foundation of good and bad.

我总是感慨，我们太容易被二元对立吸引，觉得非黑即白特别有吸引力，还追随那些似乎能明确指出哪个选项才是对的人。

And the way that I'm constantly lamenting that we're so drawn to binary, it's so appealing, and we follow people who seem to know which of the two binary options is the correct one.

这简直就是我们的软肋，天生就带着二元思维的烙印。

It's like our Achilles, our binary wiring.

从一开始就是这样。

It's from the very jump.

好和坏。

Good and bad.

要么食物多，要么食物少。

There's less food or there's more food.

世界就是这么简单。

That's how simple the world is.

我们继承了进化过程中这些神经元的全部运作方式。

And we're inheriting all of those neurons in that evolution.

确实，要超越这一点挺难的。

And yeah, that's a hard thing to transcend.

没错，随着进化时间推移，人类身上有些特质其实并不这么非黑即白。

Yep, we'll see as we go forward in evolutionary time, there are aspects of humanity that are not binary in that way.

但进化似乎确实是从这种二元模式开始的。

But evolution does seem to have started in a binary way like this.

关于线虫的大脑，有个挺有意思的现象，线虫中最著名的一种叫C.

And one interesting thing about the brain of a nematode, so a nematode, the most famous one is called C.

elegans，很多研究都是拿这种线虫做的。

Elegans where people study this nematode a lot.

我家墙上还贴着他的海报呢。

We have a poster of him on my own.

明白了。

Okay.

对。

Yeah.

他是个

He's a

很了不起的人物，特别有名。

great He's very famous.

哦。

Oh.

不是。

No.

哦。

Oh.

别耍我啊。

Don't trick me.

我脑子已经在转了。

My brain is already working already.

你刚才还说他是线虫界的布拉德·皮特呢。

You were setting it up though like that he was the Brad Pitt of the anemotope.

你抓住他了，C。

You got him, you C.

Elegans。

Elegans.

天啊，他C。

He oh my god.

他C。

He C.

Elegans有一個神經元。

Elegans have a neuron.

有意思的是，當我們看到世界上的東西時，眼睛裡的神經元會對外界信息進行編碼。

So what's interesting is when we see things in the world, our neurons in our eyes encode information in the world.

這些信息傳到大腦，然後在其他區域，我們會判斷它是不是所謂的‘好’或‘壞’。

It goes into our brain, and then elsewhere in the brain, we decide if it's quote unquote good or bad.

人類的大腦要複雜得多。

Human brain is much more complicated.

但在线虫的大脑里，某样东西是好是坏，直接由感觉神经元发出信号。

But in a nematode brain, whether or not something is good or bad is directly signaled by the sensory neurons themselves.

所以，负责探测气味的神经元会直接连接到控制向前移动的运动神经元，并且能直接感知线虫的饥饿程度。

So the neuron that detects smell directly connects to the motor neuron for going forward and is directly sensitive to how hungry the nematode is.

在这些早期的大脑里，只有好与坏之分。

So in these early brains, there was only good or bad.

当时还没有脱离‘好或坏’这个滤镜去衡量世界的方式。

There was no measuring the world in the absence of putting it through the lens of whether it's good or bad.

没错。

Right.

那在这种阶段，情绪是怎么产生的呢？

So how does emotion originate in this phase?

这部分研究其实是我觉得最有趣的地方。

So this is one of the most fun parts of doing this research.

我发现，这种机制特别精妙，但在比较心理学领域，很少有人提到过类似的现象。

I found this to be one of the most elegant things that I had not seen discussed in the sort of comparative psychology world of studying other animals.

在人类身上，我们大概都听说过两种神经调质：多巴胺和血清素。

In humans, there's two neuromodulators we're probably all familiar with, dopamine and serotonin.

关于这两种神经调质，我们经常听到各种说法。

And we hear a lot about these two neuromodulators.

它们在人脑里做的事情非常复杂。

They do very complicated things in human brains.

但总的来说，多巴胺主要跟探索、追求、渴望有关，而血清素则更多关联着满足、放松，让人觉得一切都没问题。

But by and large, we know that dopamine tends to be the seeking, pursuit, want more chemical, and serotonin is more of the satiation, relaxation, things are okay chemical.

那我们再回到线虫身上，看看能不能通过观察它大脑里的这两种化学物质，来理解这些神经调质的起源。

So let's go back to a nematode and see if we can learn anything about the origin of these neuromodulators by seeing those two chemicals in their brain.

你会发现一个特别美妙的现象。

And what you see is something completely beautiful.

它们的多巴胺神经元直接从头部延伸出来，探测体外是否存在食物。

Their dopamine neurons directly come out of their head and detect the presence of food outside of the worm.

一旦检测到外面有食物，就会让大脑大量释放多巴胺，促使线虫快速转向，去寻找附近的食物。

So it detects food outside, floods the brain with dopamine, and drives a behavioral state of turning quickly to find food nearby.

血清素存在于它的喉咙里，负责感知食物被摄入，并促使身体产生饱足感、停止进食和入睡。

Serotonin lives in its throat, and it detects the consumption of food, and it drives satiation and stopping and sleep.

消化。

Digestion.

消化。

Digestion.

所以这两种神经调质之间的对比——一个告诉你附近有好东西，赶紧去拿；另一个告诉你一切安好，可以放松了——在最原始的生物身上就已经显现出来了。

So this dichotomy between these two neuromodulators, one for there's something good nearby, quickly get it, and another for everything is okay, you can relax, you see in the very first

停顿。

break.

哇哦。

Woah.

这太惊人了。

That is crazy.

这真的很不可思议，对吧？

That's wild, right?

哇哦。

Wow.

现在正好聊聊扫地机器人。

Good time to talk about Roomba.

咱们来谈谈扫地机器人吧。

Let's talk about Roomba.

因为扫地机器人真的太厉害了。

Because Roomba is so impressive.

你第一次看到扫地机器人在别人家里吸尘时，

The first time you see a Roomba, it's vacuuming someone's home.

谁在控制它？

Who's in charge?

它怎么知道该去哪儿？

How does it know?

之前其实有人试过很多复杂的方案，想做出能自动吸尘的设备，但都没成功。

And there were really complicated attempts, right, to create this self vacuuming device that didn't work.

简单讲讲扫地机器人的发展历史吧。

Give me the history of Roomba a little bit.

扫地机器人

So Roomba

跟我聊聊扫地机器人。

Tell me about Roomba.

扫地机器人是由罗德尼·布鲁克斯创立的，他其实是麻省理工学院的一位计算机科学家兼机器人专家。

So Roomba was founded by this guy Rodney Brooks who actually is a computer science MIT roboticist.

他写了很多关于通过逆向工程大脑来构建AI系统所面临的挑战的文章，正好对应了我们刚才讨论的那些问题，因为大脑实在太复杂了。

Did a lot of writing about the challenges with trying to implement AI systems by reverse engineering the brain for exactly the problems we were talking about, which is the brain is so complicated.

他举了一个特别棒的比喻。

He gives this great allegory.

我可能复述得不太准确。

I'm gonna botch it a little bit.

但大意是这样的：假设有一群来自1900年的航空航天工程师，通过一个小传送门穿越到了现代，被允许在飞机上坐五分钟，然后就被送回去了。

But the general idea is suppose a bunch of aerospace engineers from the year 1900 went through a little portal and woke up in the modern world and were allowed to sit in an airplane for five minutes, and then they were sent back.

他们会准确识别出飞机中实现飞行的关键特征吗？

Would they correctly identify the features of an airplane that enables flight?

他认为不会，因为你只会注意到塑料材质、机翼边缘这些表面东西，反而搞不清楚真正让飞机飞起来的核心原理。

And he argues no, because you would look at the plastics, you would look at the material on the edge of the plane, and you'd be confused by the actual features that make flight possible.

所以他觉得，我们现在研究大脑时遇到的问题也是一样的。

And so he thinks that's the same problem with peering into the brain as it is now.

我们很容易误以为某些东西很重要，其实它们根本不是关键。

We're at risk of thinking certain things are important when they're in fact not.

对。

Right.

对。

Right.

对。

Right.

对。

Right.

所以他一开始的想法挺有意思，我觉得他本人可能没这么想过，但这种偶然性确实耐人寻味。

So he started interestingly, I don't think he thought about it this way, but the serendipity is interesting.

他的机器人扫地机器人，工作原理和最早的脑部机制其实差不了多少，因为他选择把问题大大简化。

His robot, the Roomba, works not that differently than the very first brain because he decided to simplify the problem dramatically.

他发现，要造出一台出色的吸尘器，根本不需要做那么多复杂的事。

And he realized you don't really need to do many complicated things to create a great vacuum cleaner.

它只需要随机转向，碰到墙再随机转个方向，就这么一直重复下去。

All it has to do is turn around randomly, and when it hits the wall, it turns randomly, it'll keep going, and it'll just keep doing that.

慢慢地，它就能把整个房间都打扫一遍。

And then eventually, it'll get around the full room.

其实它用的算法稍微复杂一点，但大体上还是这么个逻辑。

Mean, there's slightly more complicated algorithm it uses, but by and large, it's the same thing.

最有趣的是，新一代的扫地机器人多了个叫‘污垢检测’的功能，一旦发现脏污，就会改变行为，大概持续一分钟，专门在那片区域来回转圈。

And most interestingly, the more modern Roombas have something called dirt detect, where if it detects dirt, it actually changes its behavior for a minute or so, and it turns around randomly.

为什么呢？

Why?

因为世界上的脏东西都是成块分布的。

Because the world is clumpy.

如果你在一个地方发现脏污，那附近很可能也有。

If you detect dirt in one area, it's likely there's dirt nearby.

这正是线虫的做法。

This is exactly what nematodes do.

它们会向大脑释放大量多巴胺，然后迅速改变行为，在附近区域来回搜索。

They flood

它会把多巴胺充斥大脑，然后立刻改变行为，快速转向，搜寻周围区域。

the brain with dopamine and they change their behavior to turn quickly and search the local area.

我现在得再多看看关于扫地机器人的纪录片，但我猜最早研究这个问题的人，都以为它得先画出房间的地图，再设计一套高效、不重复的清扫路线。

Now I need to watch more docs on Roombas, but I would imagine the first people trying to crack this were thinking the thing has to map the room first, and then it has to come up with a very efficient protocol for going back and forth and not overlapping.

也许他们还希望它能彻底清洁。

Maybe the goal also was to steep.

很明显，Roomba 会反复清扫同一个地方。

The Roomba is clearly going over stuff it's vacuumed a bunch.

真是这样吗？

Is that the case?

是的。

Yes.

可谁在乎呢？

But who gives a fuck?

反正迟早能打扫完。

It'll eventually get it all.

没错。

Exactly.

因为你一开始可能设定的阈值太高了。

Because you could also start with maybe too high of a Yeah.

它得特别高效才行。

It's like it needs to be so efficient.

对。

Yeah.

它的运作方式是必须掌握每一个细节，然后据此规划出最高效的路径，这就是它的运行原理。

Like the way this must work is it must know every inch and then it'll design the most efficient route through here and that's how it'll work.

但生物的实际运作方式并不是这样的。

But that's not how the organisms were working.

好，接下来是第二个突破：脊椎动物最早出现的强化学习。

Okay, now breakthrough number two, reinforcement learning in the first vertebrates.

你之前说要告诉我这些是什么时候出现的。

And you were gonna tell me when those came along.

顺便也讲讲寒武纪大爆发吧。

And maybe tell us about the Cambrian explosion.

是的，在第一批大脑出现之后，这种类似线虫的生物的后代演化出了许多不同的路径。

Yeah, so in the aftermath of the first brains emerging, the descendants of this nematode like creature went down many, many different evolutionary trajectories.

最终，它们陷入了捕食者与猎物之间的相互博弈中。

And eventually they get caught in this predator prey feedback loop.

正是在这里，进化速度突然加快了——因为捕食者每进化出一项新的捕猎优势，都会给猎物带来更大的压力，迫使它们发展出更好的逃生能力；而猎物的改进又反过来推动捕食者继续进化，形成这种不断升级的循环。

And so this is where evolution really accelerates because for every new adaptation that a predator gets to get a little bit better at catching prey creates more pressure for the prey to get another adaptation to be better at getting away from the predator, which puts more pressure on the predator, and you get this back and forth feedback loop.

寒武纪大爆发发生在大约五亿到五亿五千万年前。

So the Cambrian explosion is around 500 to 550,000,000 years ago.

当时海洋里突然出现了大量动物。

We see an explosion of animals in the ocean.

在此之前，埃迪卡拉纪的化石记录非常稀少。

Before that, the Ediacaran period, very sparse fossils from that period.

而之后，动物就到处都是了。

Afterwards, you see animals everywhere.

那个时代主要是节肢动物主导的。

That era is mostly run by arthropods.

如果我们能坐上一艘小潜艇回到寒武纪，我们的祖先根本很难被发现。

So if we went back in time and we were in a little submarine around the Cambrian, our ancestors would be hard to spot.

那时候满眼都是巨大的昆虫和像甲壳类的生物。

It was mostly these huge insects, crustacean like looking creatures.

我们的祖先只是一种小小的、大概四英寸长的鱼形生物，它们是最早的脊椎动物。

And our ancestors was a small, probably four inch long fish looking creature, which were the first vertebrates.

那里还出现了不少大脑结构。

And there's a bunch of brain structures that emerge there.

我们也可以在这儿提一下马文·明斯基，因为正如我们刚才说的，这本书精妙地呼应了不同的演化阶段。

And maybe we set up Marvin Minsky at this moment too, because as we were saying, the book so elegantly parallels the different evolutions.

1951年，我们第一次真正听到了‘人工智能’这个词。

So really 1951 is the very first time we hear artificial intelligence.

那马文·明斯基当时在做什么呢？

So what's Marvin Minsky up to?

马文·明斯基是人工智能领域的奠基人之一。

Marvin Minsky is one of the sort of founding fathers of AI.

五十年代，人们特别热衷于用计算机来完成智能任务。

In the 50s, there was a lot of excitement around using computers to do intelligent tasks.

当然，后来迎来了一段低谷期，被称为‘人工智能寒冬’，因为五十年代的很多承诺都没能实现。

And of course, there was a winter after this, what's called an AI winter because a lot of the promises from the 50s didn't come to fruition.

但就脊椎动物而言，马文·明斯基曾尝试过用强化学习来训练某种东西。

But in relation to vertebrates, one of the interesting things that Marvin Minsky tried was training something through reinforcement.

所以从20世纪初开始，心理学界就在讨论这个问题了。

And so this had been discussed in psychology since the 1900s.

我们知道，只要给鸡一点奖励，就能训练它完成很多非常复杂的动作。

We know that you can train a chicken to do lots of really complicated things by giving it a treat when it does what you want.

就像巴甫洛夫那样。

Like Pavlovs.

利用奖励机制。

Harnessing the reward system.

奖励机制。

The reward system.

巴甫洛夫研究的是那种习得的反射。

So Pavlov was about sort of learned reflexes.

桑代克则发现，你可以教动物做一些非常有趣的事情，而不仅仅是产生反射。

Thorndyke was the one who saw that you could teach animals to do really interesting things, not just have a reflex.

哦，原来是这样。

Oh, see.

好的。

Okay.

桑代克著名的实验是那些迷箱。

So Thorndyke's famous experiments were these puzzle boxes.

你把一只猫放进箱子里，然后在箱子外面放上美味的食物，比如三文鱼。

You put like a cat in a box, and you put a yummy food treat like some salmon outside of it.

它要想逃出来——箱子是全透明的——就必须做出某种行为。

And the only way it can get out of the box, it's fully see through, is it has to do some behavior.

比如说，拉一下杠杆，或者用鼻子去碰某个装置。

Let's say, it has to pull a lever or has to poke its nose in something.

桑代克原本以为，这些动物是通过模仿来学习的。

Thorndyke theorized that these animals would learn through imitation learning.

所以一旦有动物搞明白了，他就让另一只动物在一旁看着。

So once one animal figured out, he let the other animal watch it.

但结果并没有发生。

And this did not happen.

它们根本不是通过模仿学会的。

They never learned through imitation.

那是灵长类动物才有的行为。

That's a primate thing.

没错。

Right.

但他发现了别的东西。

But he found something else.

他发现，奇怪的是，随着时间推移，动物逃出迷箱所需的时间慢慢变短了。

He found, oddly, that over time, the speed to get out of the puzzle box just slowly went down.

他意识到，真正发生的是试错学习。

And what he realized what was happening is it was just trial and error.

每次成功逃出来，它们下次更有可能重复之前那个让它脱困的动作。

When they got out, they just became slightly more likely to do the behavior that happened before that got them out.

它们其实并不明白自己是怎么逃出去的，只是通过反复尝试，慢慢更倾向于做那些有效的行为。

They didn't really understand how they were getting out, but through trial and error, they just slowly were getting more likely to do these things.