TED | 单词的诞生

00:18

想象一下如果你能记录你的生活，你说的一切，做的一切，就存储在一个完美的你触手可及的记忆库，你可以回到过去，找寻那难忘一刻回顾这一切，或者追寻时间的轨迹，发现在属于你自己的生活模式，那种以前没有发现的规律，而那就是我们全家，5年半前开始的，家庭旅程，这是我妻子和合作者，鲁泊尔，在这一天，从这一刻，我们带着我们第一个孩子走进了这个家，我们美丽的儿子，我们走进了一个，安装了特殊的家庭摄像系统的家。

01:10

(录像)，男人：好。

01:13

戴·罗伊:，这一刻，和其他千万的我们的特殊时刻，在我们家中被捕捉下来，因为这个房子的每个屋子，如果你仰头看，你都可以看见一个摄像机和话筒，而你望下看，你可以俯视整个房间，这是我们的客厅，这是婴儿的房间，厨房，餐厅，这是其余的地方，这些都被装进了一排，为持续拍摄设计的光盘中，这里我们飞快地经历一遍我们家庭的一天，我们从太阳初升的早晨，到亮起电灯的夜晚，最后，熄灯就寝，历经3年的时间，我们每天记录8到10个小时，积累了大约25万小时，的多轨音频和视频资料。

02:07

所以你现在看到的是有史以来，最大的家庭录相集，(笑声)，从个人的角度而言，这些代表了我们家庭的资料，已经产生了巨大的影响，我们还在继续学习其中的价值，无数的时刻，无预兆的，不造作的自然时刻，都记录在这里，我们正开始学习怎样发现和寻找它们。

02:35

而促使这个项目还有一个科学的原因，便是用这些纵向记录的数据，去了解一个，孩子是怎样学习语言的，这个孩子是我的儿子，所以在设置了隐私保护的条件下，每个被记录到的人物都得到保护，我们对我们信任的麻省理工研究团队，公开了部分数据，因此我们可以从这个巨大的，数据资料中排除出一些多余的模式，以此来试图理解社会环境，对语言形成的影响，所以我们在这里看到，我们所做的第一件事情，这是我的妻子和我在厨房做早餐，随着时间的流逝地点的变化，这是厨房里日常生活的轨迹。

03:26

为了转换，这个9万小时的录相，将它变成我们能识辨的东西，我们用行动分析来抽取，我们在时空的移动，我们称之为，时空虫，这个成为了我们工具的一部分，用来观察和辨识，数据中的各种活动，再利用这个办法，去追踪模型，特别是，我儿子在家去过哪些地方，使得我们能够聚焦解读，我儿子学习语言的语境，他从我，我妻子和保姆那里听到的所有词汇，渐渐的，他开始使用的词汇，因此通过技术和数据，在机器的协助下，录制下对话，我们现在已经完成了，超过7万字的家庭言谈的记录，现在，让我带你们，进入这些数据的第一个旅行。

04:22

我相信，你们大家都，看过时间推移的影片，加快时间的推移你可以看见花朵盛开，现在我让你们看看，语言的花朵是怎样绽放的，我的儿子，在他的第一个生日后，会说“gaga“来指水，在这之后的半年里，他渐渐地学会了，成年人说的正确的“水”我们现在来用40秒时间，快速浏览这半年，没有影象，所以你们可以专注听声音，声学上的，这种新的轨迹变化，从“Gaga"到"Water"。

04:59

(声音)婴儿：Gagagagagaga Gaga gaga gaga guga guga guga wada gaga gaga guga gaga wader guga guga water water water water water water water water water。

05:44

戴·罗伊: 他学会了啊，不是吗？

05:53

而他并不只是学会了水，在24个月里，在最初的2年里，这才是我真正关注的，这里有一张图按照时序列出了他所学到的词汇，因为我们有全部的记录，我们为他到两岁前学会的503个单词，都做了辨认和分析，他算是说话早的，所以我们开始分析其原因，为什么有些词他学得早，这是其中的一个研究结果，是一年多前出来的，让我们很吃惊，解读这张看似简单的图表的方式，是横坐标表示，照顾者的话语复杂程度，基于话语的长度，纵坐标代表了时间(演讲者口误)。

06:38

所有的数据，我们都用下述的方法排列：每次我们发现儿子学了一个新的词，我们就会回溯他听过的这个词的，所有的语言记录，然后我们绘制这些语言的长度，我们发现了一个奇特的现象，照顾者的讲话会系统地将语言简化，简化到最简单的程度，然后渐渐地回升到更复杂的句子，而惊奇的事是，这种回升和下降，正好精确的，吻合了每个词的诞生过程，一个词接一个词，很有系统规律，似乎三个主要的照顾他的人，我，我妻子，和我们的保姆，都是有系统的，我想，也是下意识的，重新构建我们的用语，去迎合他的新的词汇的诞生，带他渐渐学习更为复杂的语言，这其中蕴含的--有很多意义，但是我想指出的其中的一个，就是这个过程中必定包涵了一个惊人的反馈循环，当然，我的儿子是，在他的语言环境中学习，但是那个环境也在向他学习，环境，人，都在这个紧密的反馈循环中，并建立了一种类似脚手架的互相支撑关系，这是之前没有被注意到的。

07:57

这是关注讲话的语境来看，若是从视觉环境来看呢？我们现在看到的是，想象这是用我们家做样板做的洋娃娃屋，我们使用环状鱼眼睛摄像机，我们还做了些光学修正，然后我们就可以把它做成三维录像，欢迎到我家来，这是其中的一刻，通过几个录相机拍下的同一时刻，我们这样做是为了创造出终极的记忆机器，你可以用互动的方式前后快速搜寻，然后用这系统体验录像生活，我要做的是，是给你们看一段压缩了30分钟的速放录像，这次也是在客厅，这是我和我儿子在地上，这是影片分析，跟踪我们的移动，我儿子的留下了红色的轨迹，我的是绿色的，我们在沙发上，看着窗外汽车开过，最后，我儿子自己玩他的学步玩具。

08:55

现在定格，30分钟，我们将时间放到垂直轴上，然后我们打开，刚才留下的互动的轨迹，我们看见令人惊讶的结构，这是两种颜色的小结点，我们把它称为社交热点，那些螺旋线，我们称为单一热点，我们觉得这个影响语言学习，我们要做的是，是开始去了解，这些模式与我儿子接触的，语言间的关系，看我们是否能预测，什么时候听到怎样的单词结构，会影响到什么时候学会字词，换句话说，就是，词汇和他们所表示的世界的关系。

09:40

这是我们的解读方法，在这个录像中，同样是跟踪我的儿子，他留下了红色的轨迹，我们的保姆在门边。

09:50

(录像)保姆：你要喝水妈？(宝宝：Aaaa)，保姆：好。(宝宝：Aaaa)。

09:56

戴·罗伊:她给他水，然后两条时空虫，开始移动到厨房拿水，同时我们所做的就和“水”这个词，联系上了，随着一些动作，然后我们用数据的力量，每次我儿子，听到水这个字，以及他看见的情景，我们利用这些来分析整个影片，找到每个跟，“水”字出现时发生的活动，这个数据勾勒出了，这么一幅风景，我们把这个叫做，词景，这是水字的词景，你可以看见大多数行动是在厨房，就是左边的这些高峰，相对，你也可以为其他词汇勾勒词景，比如“goog，bye”(再见)里的，“bye”字，我们放大到房子大门口附近，我们看到，我们发现，你也会想到，一幅相对的景象，在那儿你看到“bye“高频率出现的结构，我们用这些结构，开始预言，学会语言的顺序，这是在持续进行的工作。

11:03

在我麻省理工学院的研究室-就是现在看到，那是在媒体实验室里，这成了我最喜欢的空间，视频制图方法，这个项目的关键人物都在，就是图片里的菲利普·迪坎普,，罗尼·库巴特和布兰登·罗伊，菲利普是一个密切的合作者，你们看到的视觉化功能就是他负责的，还有麦克尔·菲莱舍曼，是我实验室的另一个博士生，和我一起做了家庭视频的分析，是他发表了以下的观点：“我们分析，语言如何于事件相关，这是语言的共同的基础，我们可以把同样的思路带出你的家，戴，我们可以把它用到公共媒体上”所以我们的研究有了个意想不到的转折。

11:49

想到大众媒体，提供共同的基础，你就可以把我们的方法，运用到一个崭新的地方，我们开始分析电视内容，用同样的原则，分析一个电视信号的事件结构，电视剧集，广告，所有的组成事件结构的成分，我们现在，通过卫星电视，抽出分析了，在美国高收视率的电视节目，你不再需要把麦克风装在起居室里来，记录人们的对话，你只要去听公开的社交媒体讯息就可以了。

12:27

我们每个月抽出，大概30亿个评论，奇迹发生了，这中间可以找到事件结构，这些词汇的共同基础，那些从这次电视讯息里透露出的反馈，你得到有关这些，话题的对话，通过语意分析，你们看到的这个是根据我们的数据处理过后，的真实的数据结果，每条黄线显示一个链接，连接着外界的评论，和电视信号发出的事件结构间的关系，这都是用同样的思路，构建起来的，我们得到了这个词汇背景，不过现在词汇不是从我的客厅里来的，取而代之的情境，共同基础活动，是电视内容带动的对话，我们现在看到的这些高耸的结构，都是电视评论，它们跟电视上播放的内容联系着，同样的概念，但是你们看见的是它在不同的领域，展现的交流动态。

13:29

从根本上，而不是，比如，根据收视率衡量内容，这个给了我们观察这些，内容参与性的最基本的资料，就跟我们可以看见家里的，反馈循环和互动一样，我们现在可以利用同样的构想，来观察更大的群体，这是我们资料库里的一个子集，只是几百万信息中的5万条，社交图是和公开资缘，来自于对大众公开的来源，如果你把它们放到平面上，第二个平面是内容活跃的地方，于是我们有了节目，体育活动，广告，所有的链接结构将它们连在一起，形成了内容图表，然后是重要的第三个面向，大家在这里看到的每个连接，是一段内容和有些人评论，和有些人评论，间构成的真实联系，这里的几千万条链，让我们看见了社交图表中的关联组织，和它们跟内容的关系，于是我们可以用有趣的办法来，探索这个结构。

14:44

所以，比如，我们跟踪，某个内容的发展途经，这促使有人对此发表评论，然后我们跟踪这些评论的去向，然后观察整个活跃的社交图，然后又回头追踪查看那个社交图，和内容之间的关系，于是显现出一个非常有趣的结构，我们称之为，共视团体，你可以把它当成一个虚拟的客厅，这里头上演着引人注目的戏剧，它不是单向的，一个内容，一个事件促使某人发表了意见，他们和其他人对话，就驱动了大众传媒的收视行为，于是出现了这样的循环，驱动了整体的收视行为。

15:25

另一个例子--情况很不同，我们的资料库里有一位人士，其实我们可以找到成千上百个例子，我们给这个人一个名字，这是一个专业的媒体评论员，有很多粉丝，很多人都追随他，很有影响力，他们很喜欢讨论电视上在播的东西，于是这个人就是一个关键的链接，将大众媒体和社交媒体联系在了一起。

15:52

这份资料的最后一个例子是：有时确实是一件特别的内容，如果我们回顾这个内容，几个星期前的欧巴马总统，国情咨文演讲，再来看看我们在这组资料中发现些什么，用同样的尺度来衡量，这个内容的可参与属性，真的是很神奇的，整个国家顿时同步，爆发了谈话，是针对广播的东西，当然，通过这些线路，涌现出了结构的语言，我们可以在社交点上，感受一下这个国家即时的动脉，即时的感受，不同的社会圈的社会反应被内容所激活，都展示在社会图表上。

16:40

所以，总结来说，观点是：当我们的世界变得越来越工具化，我们有能力，搜集和链接一个一个小点，将人们的话语，和他们说这些话时所处得环境联系起来，那么呈现的将是洞悉，，社会结构和社交动态的新视野，那是以前我们没有看见过的，这好像是造一个显微镜或者望远镜，展示了我们交流和行为间，的新结构，我觉得其意义是深远的，无论是对科学而言，还是对商业，政府而言，或许更重要的是，对我们每个人而言。

17:20

所以我们把话题回到我的儿子，当我在准备这个演讲时，他在我身后看着，我给他看了这段我今天将要给你们看的录相，我征求他的同意，他同意了，然后我想，“这真是神奇的事情，整个数据库，所有这些录相，我会给交给你和你的妹妹”妹妹是两年后出生的，“你们两个将能够回顾重温，你们生物记忆无法，记得的这些时刻。”那一刻他很安静，我想：”我在想什么啊？他才5岁，他不会理解这些。“，而正当我怎么想着，他抬头对我说：“那等我长大了，我可以给我的孩子们看，是吗？”我想：“哇，这玩意儿真是太强大了。”

18:08

所以，我要给各位，留下最后一个值得回忆的，家庭记忆，这是我儿子第一次，走了迈出两步的情形，拍摄在录像中，我希望你们看的时候，注意到其中的一点，周围有点闹，这是自然的环境，我妈在厨房做饭，就在过道里，我意识到他就要迈步了，大概一两步的样子，因此各位可以听到我在鼓励他，我感到有事要发生，然后妙事发生了，请仔细听，大概在走了三步后，他感到了美妙的事情发生了，令人惊讶的反应循环作用全部启动，他松了一口气，轻轻地说了声：“哇”我也凭着直觉说了同样的话，我们现在回到那一刻，回到那个令人难忘的一刻。

19:08

(录像)戴·罗伊：嗨，过来，你行吗？哇，宝贝，你行吗？宝宝：好，戴1罗伊：妈，他走路了。

19:31

戴·罗伊：谢谢大家。

00:18

Imagine if you could record your life -- everything you said, everything you did, available in a perfect memory store at your fingertips, so you could go back and find memorable moments and relive them, or sift through traces of time and discover patterns in your own life that previously had gone undiscovered. Well that's exactly the journey that my family began five and a half years ago. This is my wife and collaborator, Rupal. And on this day, at this moment, we walked into the house with our first child, our beautiful baby boy. And we walked into a house with a very special home video recording system. 

01:10

(Video) Man: Okay. 

01:13

Deb Roy: This moment and thousands of other moments special for us were captured in our home because in every room in the house, if you looked up, you'd see a camera and a microphone, and if you looked down, you'd get this bird's-eye view of the room. Here's our living room, the baby bedroom, kitchen, dining room and the rest of the house. And all of these fed into a disc array that was designed for a continuous capture. So here we are flying through a day in our home as we move from sunlit morning through incandescent evening and, finally, lights out for the day. Over the course of three years, we recorded eight to 10 hours a day, amassing roughly a quarter-million hours of multi-track audio and video. 

02:07

So you're looking at a piece of what is by far the largest home video collection ever made. (Laughter) And what this data represents for our family at a personal level, the impact has already been immense, and we're still learning its value. Countless moments of unsolicited natural moments, not posed moments, are captured there, and we're starting to learn how to discover them and find them. 

02:35

But there's also a scientific reason that drove this project, which was to use this natural longitudinal data to understand the process of how a child learns language -- that child being my son. And so with many privacy provisions put in place to protect everyone who was recorded in the data, we made elements of the data available to my trusted research team at MIT so we could start teasing apart patterns in this massive data set, trying to understand the influence of social environments on language acquisition. So we're looking here at one of the first things we started to do. This is my wife and I cooking breakfast in the kitchen, and as we move through space and through time, a very everyday pattern of life in the kitchen. 

03:26

In order to convert this opaque, 90,000 hours of video into something that we could start to see, we use motion analysis to pull out, as we move through space and through time, what we call space-time worms. And this has become part of our toolkit for being able to look and see where the activities are in the data, and with it, trace the pattern of, in particular, where my son moved throughout the home, so that we could focus our transcription efforts, all of the speech environment around my son -- all of the words that he heard from myself, my wife, our nanny, and over time, the words he began to produce. So with that technology and that data and the ability to, with machine assistance, transcribe speech, we've now transcribed well over seven million words of our home transcripts. And with that, let me take you now for a first tour into the data. 

04:22

So you've all, I'm sure, seen time-lapse videos where a flower will blossom as you accelerate time. I'd like you to now experience the blossoming of a speech form. My son, soon after his first birthday, would say "gaga" to mean water. And over the course of the next half-year, he slowly learned to approximate the proper adult form, "water." So we're going to cruise through half a year in about 40 seconds. No video here, so you can focus on the sound, the acoustics, of a new kind of trajectory: gaga to water. 

04:59

(Audio) Baby: Gagagagagaga Gaga gaga gaga guga guga guga wada gaga gaga guga gaga wader guga guga water water water water water water water water water. 

05:44

DR: He sure nailed it, didn't he. 

05:53

So he didn't just learn water. Over the course of the 24 months, the first two years that we really focused on, this is a map of every word he learned in chronological order. And because we have full transcripts, we've identified each of the 503 words that he learned to produce by his second birthday. He was an early talker. And so we started to analyze why. Why were certain words born before others? This is one of the first results that came out of our study a little over a year ago that really surprised us. The way to interpret this apparently simple graph is, on the vertical is an indication of how complex caregiver utterances are based on the length of utterances. And the [horizontal] axis is time. 

06:38

And all of the data, we aligned based on the following idea: Every time my son would learn a word, we would trace back and look at all of the language he heard that contained that word. And we would plot the relative length of the utterances. And what we found was this curious phenomena, that caregiver speech would systematically dip to a minimum, making language as simple as possible, and then slowly ascend back up in complexity. And the amazing thing was that bounce, that dip, lined up almost precisely with when each word was born -- word after word, systematically. So it appears that all three primary caregivers -- myself, my wife and our nanny -- were systematically and, I would think, subconsciously restructuring our language to meet him at the birth of a word and bring him gently into more complex language. And the implications of this -- there are many, but one I just want to point out, is that there must be amazing feedback loops. Of course, my son is learning from his linguistic environment, but the environment is learning from him. That environment, people, are in these tight feedback loops and creating a kind of scaffolding that has not been noticed until now. 

07:57

But that's looking at the speech context. What about the visual context? We're not looking at -- think of this as a dollhouse cutaway of our house. We've taken those circular fish-eye lens cameras, and we've done some optical correction, and then we can bring it into three-dimensional life. So welcome to my home. This is a moment, one moment captured across multiple cameras. The reason we did this is to create the ultimate memory machine, where you can go back and interactively fly around and then breathe video-life into this system. What I'm going to do is give you an accelerated view of 30 minutes, again, of just life in the living room. That's me and my son on the floor. And there's video analytics that are tracking our movements. My son is leaving red ink. I am leaving green ink. We're now on the couch, looking out through the window at cars passing by. And finally, my son playing in a walking toy by himself. 

08:55

Now we freeze the action, 30 minutes, we turn time into the vertical axis, and we open up for a view of these interaction traces we've just left behind. And we see these amazing structures -- these little knots of two colors of thread we call "social hot spots." The spiral thread we call a "solo hot spot." And we think that these affect the way language is learned. What we'd like to do is start understanding the interaction between these patterns and the language that my son is exposed to to see if we can predict how the structure of when words are heard affects when they're learned -- so in other words, the relationship between words and what they're about in the world. 

09:40

So here's how we're approaching this. In this video, again, my son is being traced out. He's leaving red ink behind. And there's our nanny by the door. 

09:50

(Video) Nanny: You want water? (Baby: Aaaa.) Nanny: All right. (Baby: Aaaa.) 

09:56

DR: She offers water, and off go the two worms over to the kitchen to get water. And what we've done is use the word "water" to tag that moment, that bit of activity. And now we take the power of data and take every time my son ever heard the word water and the context he saw it in, and we use it to penetrate through the video and find every activity trace that co-occurred with an instance of water. And what this data leaves in its wake is a landscape. We call these wordscapes. This is the wordscape for the word water, and you can see most of the action is in the kitchen. That's where those big peaks are over to the left. And just for contrast, we can do this with any word. We can take the word "bye" as in "good bye." And we're now zoomed in over the entrance to the house. And we look, and we find, as you would expect, a contrast in the landscape where the word "bye" occurs much more in a structured way. So we're using these structures to start predicting the order of language acquisition, and that's ongoing work now. 

11:03

In my lab, which we're peering into now, at MIT -- this is at the media lab. This has become my favorite way of videographing just about any space. Three of the key people in this project, Philip DeCamp, Rony Kubat and Brandon Roy are pictured here. Philip has been a close collaborator on all the visualizations you're seeing. And Michael Fleischman was another Ph.D. student in my lab who worked with me on this home video analysis, and he made the following observation: that "just the way that we're analyzing how language connects to events which provide common ground for language, that same idea we can take out of your home, Deb, and we can apply it to the world of public media." And so our effort took an unexpected turn. 

11:49

Think of mass media as providing common ground and you have the recipe for taking this idea to a whole new place. We've started analyzing television content using the same principles -- analyzing event structure of a TV signal -- episodes of shows, commercials, all of the components that make up the event structure. And we're now, with satellite dishes, pulling and analyzing a good part of all the TV being watched in the United States. And you don't have to now go and instrument living rooms with microphones to get people's conversations, you just tune into publicly available social media feeds. 

12:27

So we're pulling in about three billion comments a month, and then the magic happens. You have the event structure, the common ground that the words are about, coming out of the television feeds; you've got the conversations that are about those topics; and through semantic analysis -- and this is actually real data you're looking at from our data processing -- each yellow line is showing a link being made between a comment in the wild and a piece of event structure coming out of the television signal. And the same idea now can be built up. And we get this wordscape, except now words are not assembled in my living room. Instead, the context, the common ground activities, are the content on television that's driving the conversations. And what we're seeing here, these skyscrapers now, are commentary that are linked to content on television. Same concept, but looking at communication dynamics in a very different sphere. 

13:29

And so fundamentally, rather than, for example, measuring content based on how many people are watching, this gives us the basic data for looking at engagement properties of content. And just like we can look at feedback cycles and dynamics in a family, we can now open up the same concepts and look at much larger groups of people. This is a subset of data from our database -- just 50,000 out of several million -- and the social graph that connects them through publicly available sources. And if you put them on one plain, a second plain is where the content lives. So we have the programs and the sporting events and the commercials, and all of the link structures that tie them together make a content graph. And then the important third dimension. Each of the links that you're seeing rendered here is an actual connection made between something someone said and a piece of content. And there are, again, now tens of millions of these links that give us the connective tissue of social graphs and how they relate to content. And we can now start to probe the structure in interesting ways. 

14:44

So if we, for example, trace the path of one piece of content that drives someone to comment on it, and then we follow where that comment goes, and then look at the entire social graph that becomes activated and then trace back to see the relationship between that social graph and content, a very interesting structure becomes visible. We call this a co-viewing clique, a virtual living room if you will. And there are fascinating dynamics at play. It's not one way. A piece of content, an event, causes someone to talk. They talk to other people. That drives tune-in behavior back into mass media, and you have these cycles that drive the overall behavior. 

15:25

Another example -- very different -- another actual person in our database -- and we're finding at least hundreds, if not thousands, of these. We've given this person a name. This is a pro-amateur, or pro-am media critic who has this high fan-out rate. So a lot of people are following this person -- very influential -- and they have a propensity to talk about what's on TV. So this person is a key link in connecting mass media and social media together. 

15:52

One last example from this data: Sometimes it's actually a piece of content that is special. So if we go and look at this piece of content, President Obama's State of the Union address from just a few weeks ago, and look at what we find in this same data set, at the same scale, the engagement properties of this piece of content are truly remarkable. A nation exploding in conversation in real time in response to what's on the broadcast. And of course, through all of these lines are flowing unstructured language. We can X-ray and get a real-time pulse of a nation, real-time sense of the social reactions in the different circuits in the social graph being activated by content. 

16:40

So, to summarize, the idea is this: As our world becomes increasingly instrumented and we have the capabilities to collect and connect the dots between what people are saying and the context they're saying it in, what's emerging is an ability to see new social structures and dynamics that have previously not been seen. It's like building a microscope or telescope and revealing new structures about our own behavior around communication. And I think the implications here are profound, whether it's for science, for commerce, for government, or perhaps most of all, for us as individuals. 

17:20

And so just to return to my son, when I was preparing this talk, he was looking over my shoulder, and I showed him the clips I was going to show to you today, and I asked him for permission -- granted. And then I went on to reflect, "Isn't it amazing, this entire database, all these recordings, I'm going to hand off to you and to your sister" -- who arrived two years later -- "and you guys are going to be able to go back and re-experience moments that you could never, with your biological memory, possibly remember the way you can now?" And he was quiet for a moment. And I thought, "What am I thinking? He's five years old. He's not going to understand this." And just as I was having that thought, he looked up at me and said, "So that when I grow up, I can show this to my kids?" And I thought, "Wow, this is powerful stuff." 

18:08

So I want to leave you with one last memorable moment from our family. This is the first time our son took more than two steps at once -- captured on film. And I really want you to focus on something as I take you through. It's a cluttered environment; it's natural life. My mother's in the kitchen, cooking, and, of all places, in the hallway, I realize he's about to do it, about to take more than two steps. And so you hear me encouraging him, realizing what's happening, and then the magic happens. Listen very carefully. About three steps in, he realizes something magic is happening, and the most amazing feedback loop of all kicks in, and he takes a breath in, and he whispers "wow" and instinctively I echo back the same. And so let's fly back in time to that memorable moment. 

19:08

(Video) DR: Hey. Come here. Can you do it? Oh, boy. Can you do it? Baby: Yeah. DR: Ma, he's walking. 

19:31

DR: Thank you.