Lenny's Podcast
The Godmother of AI on Jobs, Robots & Why World Models Are Next
AI教母谈就业、机器人与为何世界模型是下一步
ImageNet creator Fei-Fei Li discusses spatial intelligence breakthroughs, AI's real impact on jobs, the state of robotics, and why language models alone won't achieve true general intelligence.
ImageNet 创始人 Fei-Fei Li 深度解读空间智能的突破、AI 对就业市场的真实影响、机器人技术的现状与未来、以及为什么纯粹的语言模型不足以实现真正的通用智能。
00:00
00:00
A lot of people call you the godmother of AI.
很多人称你为AI教母。
00:02
The work you did actually was the spark
你当时的工作其实是那束火花，
00:04
that brought us out of AI winter.
把我们带出了AI寒冬。
00:05
In the middle of 2015, middle of 2016,
2015年中到2016年中，
00:09
some tech companies avoid using the word AI
有些科技公司避用AI这个词，
00:12
because they were not sure if AI was a 30-word or 20-17-ish
因为他们不确定AI是30个词还是20到17个词左右——
00:18
was the beginning of companies calling themselves AI companies.
那正是公司开始自称AI公司的起点。
00:22
There's a line I think this was when you're presenting to Congress.
我记得有一句话，好像是你向国会做报告时说的。
00:25
There's nothing artificial about AI.
AI 里没有什么“人工”的成分。
00:26
It's inspired by people, it's created by people
它受人类启发，由人类创造，
00:28
and most importantly, it impacts people.
而最重要的是，它影响着人类。
00:30
It's not like I think AI will have no impact on jobs or people.
我并不是说 AI 对工作或人类没有影响。
00:34
In fact, I believe that whatever AI does,
事实上，我相信无论 AI 做什么，
00:38
currently or in the future is up to us.
现在还是未来，都取决于我们。
00:41
It's up to the people.
取决于人类自己。
00:42
I do believe technology is a net positive for humanity,
我确实相信技术对人类整体是利大于弊的。
00:46
but I think every technology is a double-edged sword
但我觉得每项技术都是一把双刃剑，
00:50
if we're not doing the right thing.
如果我们不走在正确的方向上。
00:52
As a society, as individuals, we can screw this up as well.
作为社会，作为个人，我们同样可能搞砸这一切。
00:56
You have this breakthrough inside of just OK.
你有了这个突破，但也就只是“还行”的程度。
00:58
We can train machines to think like humans,
我们可以训练机器像人类一样思考，
01:00
but it's just missing the data that humans have to learn as a child.
但缺少的正是人类像孩子那样学习所需的数据。
01:03
I chose to look at artificial intelligence
我选择通过视觉智能的视角
01:05
through the lens of visual intelligence
来审视人工智能。
01:08
because humans are deeply visual animals.
因为人类是深度依赖视觉的动物。
01:11
We need to train machines with as much information
我们需要用尽可能多的信息来训练机器，让它们学习物体的图像，
01:14
as possible on images of objects,
但物体本身非常非常难以学习。
01:16
but objects are very, very difficult to learn.
一个物体在图像中可以有无限种呈现方式。
01:20
A single object can have infinite possibilities
为了训练计算机掌握成千上万个物体概念，
01:24
that is shown on an image
你真的需要给它展示数百万个例子。
01:25
in order to train computers with tens and thousands of object concepts.
为了训练计算机理解成千上万个物体概念，你确实需要向它展示数百万个示例。
01:30
You really need to show it millions of examples.
她被称为AI教母。
01:36
Today, my guest is Dr. Fei-Fei Li,
今天我的嘉宾是李飞飞博士，她被誉为AI教母。飞飞主导并身处许多引爆当前AI革命的重大突破的核心。她牵头创建了ImageNet，这基本上源于她意识到AI需要海量干净、带标签的数据才能变得更聪明。
01:38
who's known as the Godmother of AI.
李飞飞一直负责并处于许多引发当前AI革命的重大突破的核心位置。
01:41
Fei-Fei has been responsible for and at the center
她主导了ImageNet的创建，这基本上是她意识到AI需要——
01:43
of many of the biggest breakthroughs that sparked the AI revolution
在推动AI革命的众多重大突破中，
01:46
they were currently living through.
她正亲身经历着这一切。
01:48
She spearheaded the creation of ImageNet,
她主导创建了ImageNet，
01:50
which was basically her realizing that AI needed
这基本上源于她意识到AI需要——
01:53
a ton of clean, labeled data to get smarter.
大量干净、带标签的数据才能让它变得更聪明。
01:56
And that data set became deep breakthrough
那个数据集成为了深度突破，
01:58
that led to the current approach to building and scaling AI models.
从而催生了当前构建和扩展AI模型的方法。
02:01
She was chief AI scientist at Google Cloud,
她曾是Google Cloud的首席AI科学家，
02:04
which is where some of the biggest early technology breakthroughs emerged from.
那里诞生了一些最早的重大技术突破。
02:07
She was director at Sail, Stanford's artificial intelligence lab,
她还担任过Sail的主任，也就是斯坦福大学的人工智能实验室，
02:11
where many of the biggest AI minds came out of.
许多顶尖AI人才都出自那里。
02:13
She's also a co-creator of Stanford Human-Centered AI Institute,
她也是斯坦福以人为本AI研究所的联合创始人，
02:17
which is playing a vital role in a direction that AI is taking.
该机构在AI的发展方向上正发挥着关键作用。
02:20
She's also been on the board of Twitter.
她曾担任Twitter董事会成员。
02:22
She was named one of times 100 most influential people in AI.
她被《时代》杂志评为全球AI领域最具影响力的100人之一。
02:26
She's also the United Nations Advisory Board I could go on.
她还是联合国顾问委员会成员——这些头衔我还能继续列举。
02:29
In our conversation, Fei-Fei shares a brief history
在我们的对话中，李飞飞简要回顾了
02:32
of how we got to today in the world of AI,
AI领域如何发展到今天的历史，
02:35
including this mind-blowing reminder that nine to 10 years ago,
其中有个令人震撼的提醒：九到十年前，
02:38
calling yourself an AI company was basically a death knell for your brand.
自称AI公司基本等于给自己的品牌判了死刑。
02:43
Because no one believed that AI was actually going to work.
因为当时没人相信AI真的能行得通。
02:45
Today, it's completely different.
今天，情况完全不同了。
02:47
Every company is an AI company.
每家公司都是AI公司。
02:49
We also chat about her take on how she sees AI impacting humanity in the future,
我们还聊了她如何看待AI未来对人类的影响，
02:54
how far current technologies will take us,
当前技术能带我们走多远，
02:56
why she's so passionate about building a world model,
她为什么对构建世界模型如此热衷，
02:59
and what exactly world models are.
以及世界模型到底是什么。
03:01
And most exciting of all,
最令人兴奋的是，
03:02
the launch of the world's first large world model, Marble,
全球首个大型世界模型Marble的发布。
03:06
which just came out as this podcast comes out.
就在这期播客发布的同时，这个项目也上线了。
03:08
Anyone can go play with this at marble.worldlabs.ai.
任何人都可以去 marble.worldlabs.ai 体验一下。
03:12
It's insane.
简直太疯狂了。
03:13
Definitely check it out.
一定要去试试看。
03:14
Fei-Fei is incredible and way to enter the radar
Fei-Fei 太厉害了，她终于进入了大众视野，
03:16
for the impact that she's had on the world.
而她一直以来对世界的影响也值得被更多人看见。
03:18
So I am really excited to have her on
所以我真的很高兴能请到她来做客，
03:20
and to spread her wisdom with more people.
把她的智慧分享给更多人。
03:22
A huge thank you to Ben Horowitz and Condoleezza Rice
特别感谢Ben Horowitz和Condoleezza Rice为这次对话提供了话题建议。如果你喜欢这期播客，别忘了在你常用的播客应用或YouTube上订阅关注。接下来，在简短赞助商环节之后，有请李飞飞博士。本期节目由Figma呈现，Figma Make的开发者。
03:25
for suggesting topics for this conversation.
感谢你为这次对话建议话题。
03:27
If you enjoyed this podcast, don't forget to subscribe
如果你喜欢这期播客，别忘了订阅
03:29
and follow it in your favorite podcasting app or YouTube.
并在你常用的播客应用或YouTube上关注它。
03:32
With that, I bring you Dr. Fei-Fei Li
接下来，有请李飞飞博士
03:34
after a short word from our sponsors.
在简短赞助商环节之后。
03:37
This episode is brought to you by Figma,
本期节目由Figma呈现，
03:39
Makers of Figma Make.
Figma Make的开发者。
03:41
When it was a PM at Airbnb,
在Airbnb做产品经理的时候，
03:42
I still remember when Figma came out
我还记得Figma刚出来那会儿，
03:44
and how much it improved how we operated as a team.
它大大改善了团队协作的方式。
03:47
Suddenly, I could involve my whole team
突然间，我能让整个团队都参与设计流程，
03:49
in the design process,
快速对设计概念给出反馈，
03:50
dip feedback on design concepts really quickly,
整个产品开发过程也因此变得有趣多了。
03:53
and it just made the whole product development process
这让整个产品开发过程变得有趣多了。
03:55
so much more fun.
我的团队分布在科罗拉多、澳大利亚。
03:56
But Figma never felt like it was for me.
但Figma对我来说一直不太对味。
03:58
It was great for giving feedback and designs,
它用来给反馈和设计稿提意见确实不错，
04:01
but as a builder, I wanted to make stuff.
但作为一个动手做东西的人，我更想自己创造。
04:03
That's why Figma built Figma Make.
所以Figma做了Figma Make。
04:06
With just a few prompts, you can make any idea or design
只要输入几个prompt，你就能把任何想法或设计
04:09
into a fully functional prototype or app
变成一个完整可用的prototype或app，
04:12
that anyone can iterate on and validate with customers.
任何人都能在此基础上迭代，并拿给客户验证。
04:15
Figma Make is a different kind of vibe coding tool.
Figma Make是一种不同风格的vibe coding工具。
04:18
Because it's all in Figma,
因为所有内容都在 Figma 里，
04:19
you can use your team's existing design building blocks,
你可以直接用团队现有的设计组件，
04:22
making it easy to create outputs that look good and feel real
轻松做出看起来不错、用起来真实，
04:25
and are connected to how your team builds.
并且跟团队开发流程一致的效果。
04:28
Stop spending so much time telling people
别再花那么多时间跟别人讲你的产品愿景了，
04:30
about your product vision and instead show it to them.
直接展示给他们看。
04:33
Make code back prototypes and apps fast with Figma Make.
用 Figma Make 快速生成带代码的原型和应用。
04:37
Check it out at figma.com slash20.
前往 figma.com/20 了解一下。
04:40
Did you know that I have a whole team
你知道吗，我其实有一个完整的团队
04:41
that helps me with my podcast and with my newsletter?
在帮我做播客和写 newsletter。
04:44
I want everyone on my team to be super happy
我希望团队里的每个人都超级开心
04:47
and thrive in their roles.
并且能在自己的岗位上茁壮成长。
04:48
JustWorks knows that your employees are more
JustWorks 明白，你的员工不仅仅是员工，
04:50
than just your employees.
他们是你的伙伴。
04:52
They're your people.
我的团队成员分布在科罗拉多、澳大利亚各地。
04:53
My team is spread out across Colorado, Australia
嘿，嘿，非常感谢你来到这里，欢迎来到播客。
04:55
and Nepal, West Africa and San Francisco.
and Nepal, West Africa and San Francisco.
04:58
My life would be so incredibly complicated
My life would be so incredibly complicated
05:00
to hire people internationally to pay people
to hire people internationally to pay people
05:02
on time and in their local currencies
on time and in their local currencies
05:05
and to answer their HR questions 24-7.
and to answer their HR questions 24-7.
05:07
But with justWorks, it's super easy.
But with justWorks, it's super easy.
05:10
Whether you're setting up your own automated payroll,
Whether you're setting up your own automated payroll,
05:12
offering premium benefits or hiring internationally,
offering premium benefits or hiring internationally,
05:15
JustWorks offers simple software and 24-7 human support
JustWorks 提供简单易用的软件和来自小企业专家的全天候人工支持，为你和你的团队服务。
05:19
from small business experts for you and your people.
他们帮你把人力资源管好，这样你就能对得起你的团队。
05:21
They do your human resources right,
JustWorks，为你的团队而生。
05:23
so that you can do right by your people.
嘿，嘿，非常感谢你来到这里，欢迎收听本期播客。
05:25
JustWorks for your people.
我很高兴能来，Lenny。
05:31
Hey, hey, thank you so much for being here
我很高兴能来，Lenny。
05:32
and welcome to the podcast.
简单介绍一下你，让大家有个了解，
05:34
I'm excited to be here, Lenny.
在开场白里我会分享其他所有精彩的内容。
05:36
I'm even more excited to have you here.
能请你来我真的很兴奋。
05:38
It is such a treat to get to chat with you.
能跟你聊天真是太棒了。
05:40
There's so much that I want to talk about.
我有太多想聊的东西了。
05:42
You've been at the center of this AI explosion
你一直身处这场AI爆发的中心
05:45
that we're seeing right now for so long.
已经很久了。
05:47
We're gonna talk about a bunch of the history
我们会聊很多历史
05:49
that I think a lot of people don't even know
我觉得很多人甚至都不知道
05:51
about how this whole thing started.
这一切到底是怎么开始的。
05:52
But let me first read a quote from Wired
先让我念一段《连线》杂志对你的评价，让大家有个概念。开场白里我会介绍你其他那些厉害成就，不过我觉得用这段话铺垫一下氛围挺合适的。李飞飞属于极少数科学家——这个群体小到可能一张厨房餐桌就能围坐——正是他们推动了AI近年来的惊人突破。
05:54
about you, just so people get a sense
关于你，只是为了让人们有个概念——
05:56
and in the intro I'll share all of the other epic things
在开场白里我会分享其他所有史诗级成就——
05:58
you've done, but I think this is a good way
你刚才做的那些，我觉得是个很好的方式
05:59
to just set context.
来铺垫一下背景。
06:01
Fei-Fei is one of a tiny group of scientists,
李飞飞是极少数科学家之一，
06:03
a group perhaps small enough to fit around a kitchen table
这个群体小到可能一张厨房餐桌就能坐下，
06:06
who are responsible for AI's recent remarkable advances.
正是他们推动了AI近年来的惊人进展。
06:10
A lot of people call you the godmother of AI.
很多人叫你AI教母。
06:13
And unlike a lot of AI leaders, you're an AI optimist.
跟很多AI领袖不一样，你是个AI乐观派。
06:18
You don't think AI is gonna replace us.
你不觉得AI会取代我们。
06:20
You don't think it's gonna take all our jobs.
你不觉得它会抢走我们所有的工作。
06:21
You don't think it's gonna kill us.
你也不觉得它会害死我们。
06:22
So I thought it'd be fun to start there.
所以我想从这儿聊起挺有意思的。
06:24
Just what's your perspective on how AI is going
就说说你的看法吧——AI长期来看会怎么影响人类？
06:27
to impact humanity over time?
随着时间的推移对人类产生影响？
06:29
Yeah, okay, so Lenny, let me be very clear.
好，Lenny，我把话说清楚。
06:32
I'm not a utopian.
我不是那种盲目乐观的人。
06:34
So it's not like I think AI will have no impact on jobs
所以我并不觉得AI对就业或人类毫无影响。
06:38
or people.
实际上，我是个以人为本的人。
06:39
In fact, I'm a humanist.
我相信无论AI现在或将来做什么，
06:41
I believe that whatever AI does
都取决于我们人类自己。
06:45
in currently or in the future is up to us.
无论是现在还是未来，都取决于我们。
06:49
It's up to the people.
取决于人类自己。
06:50
So I do believe technology is a net positive for humanity.
所以我确实相信技术对人类整体是利大于弊的。
06:55
If you look at the long course of civilization,
如果你看文明发展的漫长历程，
06:58
I think we are fundamentally
我觉得从根本上来说，
07:02
we're an innovative species that we, you know,
我们是一个善于创新的物种，你看，
07:06
if you look at from, you know, written record,
从几千年前有文字记录开始到现在，
07:10
thousands of years ago to now,
人类一直在不断自我革新，
07:13
humans just kept innovating ourselves
也在不断革新我们的工具。
07:15
and innovating our tools.
并不断革新我们的工具。
07:17
And with that, we make lives better,
就这样，我们让生活更美好，
07:20
we make work better, we build civilization.
让工作更高效，我们建设文明。
07:23
And I do believe AI is part of that.
我确实相信AI是其中的一部分。
07:26
So that's where the optimism comes from.
所以这就是乐观的来源。
07:29
But I think every technology is a double-edged sword
但我认为每一项技术都是一把双刃剑，
07:34
and if we're not doing the right thing,
如果我们没有做对的事情，
07:38
as a species, as a society,
作为物种，作为社会，
07:41
as communities, as individuals,
作为社群，作为个体。
07:44
we can screw this up as well.
我们也能搞砸这件事。
07:47
There's this line.
有这么一句话。
07:48
I think this was when you were presenting to Congress.
我记得这是你在国会作证时说的。
07:50
There's nothing artificial about AI.
AI 里没有什么“人工”的。
07:51
It's inspired by people, it's created by people
它受人类启发，由人类创造，
07:53
and most importantly it impacts people.
而最重要的是，它影响着人类。
07:56
I don't have a question there, but what a great line.
我这不是在提问，但这句话说得真好。
07:59
Yeah, I feel pretty deeply, you know,
是啊，我对此感触很深，你知道的。
08:02
I started working AI two and half decades ago
我大概在二十五年前开始接触AI，过去二十年里也一直在带学生。几乎每个学生毕业时，我都会提醒他们——你知道，当他们从我的实验室毕业时，你的专业领域叫做人工智能，但这里面其实没什么“人工”的成分。回到刚才提到的那个观点，就是这一切最终走向何方，其实取决于我们自己。
08:07
and I've been having students for the past two decades.
过去二十年里，我一直带着学生。
08:11
And almost every student who graduates, I remind them,
几乎每个毕业的学生，我都会提醒他们，
08:15
you know, when they graduate from my lab,
你知道，当他们从我的实验室毕业时，
08:17
that your field is called artificial intelligence,
你的领域叫做人工智能，
08:21
but there's nothing artificial about it.
但其中没有任何“人工”的成分。
08:23
Coming back to the point just made about how it's kind of
回到刚才提到的观点，即这一切走向何方，
08:25
up to us, about where this all goes.
某种程度上取决于我们。
08:27
What is it you think we need to get right?
你觉得我们最需要把什么做对？
08:28
How do we set things on a path?
我们该怎么让事情走上正轨？
08:30
I know this is a very difficult question to answer,
我知道这是个很难回答的问题，
08:33
but just what should, what's your advice,
但到底应该——你的建议是什么？
08:35
what do you think we should be in mind?
你觉得我们心里该装着什么？
08:36
Yeah, like how many hours do we have?
对，就像我们还有多少时间？
08:38
How do we align AI?
我们怎么让AI对齐？
08:40
There we go, let's solve it.
好，那就开始解决吧。
08:41
Yeah, so I think people should be responsible individuals,
是的，我觉得每个人都应该成为负责任的个体，不管我们做什么。这是我们教给孩子的道理，也是我们作为成年人需要做到的——无论你参与的是AI开发、AI部署还是AI应用的哪个环节。而且很可能我们中的大多数人，尤其是技术人员，都身处其中。
08:45
no matter what we do.
不管我们做什么，
08:47
This is what we teach our children
这都是我们教给孩子的道理，
08:49
and this is what we need to do as grownups as well,
也是我们作为成年人需要做到的，
08:52
no matter which part of the AI development
无论你参与的是AI开发的哪个环节，
08:57
or AI deployment or AI application
还是AI部署或AI应用，
09:01
you are participating in.
都一样。
09:03
And most likely many of us, especially as technologists
而且很可能我们中的许多人，尤其是技术人员，
09:07
were in multiple points,
在多个层面上，
09:09
we should act like responsible individuals
我们都应该像负责任的个体一样行事，
09:12
and care about us, actually care a lot about us.
并且关心我们自己——实际上，要非常关心。
09:16
I think everybody today should care about AI
我认为今天每个人都应该关心AI，
09:19
because it is going to impact your individual life,
因为它会影响你的个人生活，
09:23
it is going to impact your community,
会影响你的社区，
09:25
it's going to impact the society
会影响整个社会，
09:28
and the future generation and caring about it
以及未来的世代，而关心这件事本身，就是重要的。
09:31
as a responsible person is the first
作为一个负责任的人，这是第一步，
09:35
but also the most important step.
但也是最重要的一步。
09:37
Okay, so let me actually take a step back
好，那我先退一步，
09:39
and kind of go to the beginning of AI.
回到AI的起点。
09:42
Most people started hearing and caring about AI
大多数人开始听说并关注AI，
09:46
is what it's called today.
是在它如今被广泛讨论的时候。
09:47
Just like, I don't know, a few years ago
就像，嗯，大概几年前，
09:48
when ChatGP came out, maybe it was like three years ago.
ChatGP刚出来那会儿，可能也就三年前吧。
09:51
Three years ago, almost one more month, three years ago.
三年前，差不多再多一个月，三年前。
09:54
Wow, okay, that was ChatGP coming out
哇，好吧，那是ChatGPT刚出来的时候。
09:56
is that the milestone?
那就是里程碑对吧？
09:57
Yeah, mine, okay, cool.
对，对我来说，好吧，酷。
09:58
That's exactly how I saw it.
我当时就是这么看的。
10:00
But very few people know there was a long, long history
但很少有人知道，其实有一段很长的历史
10:02
of people working on, it was called machine learning
很多人一直在研究，那时候叫机器学习
10:04
back then and there's other terms
还有其他术语。
10:06
and now it's just everything's AI.
现在基本上所有东西都跟AI有关。
10:07
And there was kind of like a long period
中间其实有一段很长的时期，就是很多人一直在埋头研究。
10:09
of just a lot of people working on it
然后大家最初经历了一个所谓的AI寒冬，那时候几乎所有人都放弃了，觉得这个方向行不通。
10:11
and then there's what people were first used
而你当时做的工作，实际上就是关键所在。
10:12
to the AI winter where people just gave up,
会经历AI寒冬——那时候人们直接放弃了。
10:14
almost most people did and just, okay, this idea
几乎所有人都这么想，觉得这个想法行不通。
10:17
isn't going anywhere.
然后你做的研究实际上基本上就是——
10:19
And then the work you did actually was essentially
好吧，我甚至算不上第一代AI研究者。
10:21
the spark that brought us out of AI winter
将我们带出AI寒冬的火花，
10:23
and is directly responsible for the world
直接造就了如今这个
10:26
where now just AI is all we talk about
所有人都在谈论AI的世界——
10:28
as you just said, it's gonna impact everything we do.
正如你所说，它将影响我们做的每一件事。
10:31
So that would be really interesting to hear from you
所以真的很想听你聊聊
10:33
just kind of like the brief history
一段简短的历史：
10:35
of what the world was like before ImageNet
ImageNet出现之前的世界是什么样，
10:38
then just the work you did to create ImageNet
以及你创建ImageNet所做的工作。
10:41
why that was so important
为什么那件事如此重要
10:42
and then just what happened after?
然后之后又发生了什么？
10:44
It is for me hard to keep in mind
对我来说很难时刻记住
10:47
that AI is so new for everybody.
AI对所有人来说都这么新。
10:50
When I lived my entire professional life in AI,
当我整个职业生涯都活在AI领域时，
10:55
it's there's a part of me that it's just,
我内心有一部分觉得，
10:58
it's so satisfying to see a personal curiosity
看到自己刚成年时开始的个人好奇心
11:02
that I started barely out of teenage hood
如今得到实现，真是无比满足。
11:06
and now has become a transformative force
如今已成为推动我们文明变革的力量。
11:11
of our civilization.
它本质上是一种文明级别的技术。
11:13
It generally is a civilization level technology.
这段历程大约有三十年，
11:17
So that journey is about about 30 years
或者说二十多年，二十多年吧，
11:21
or 20 something, 20 plus years
而且真的让人非常满足。
11:24
and it's just very satisfying.
那么，这一切是从哪里开始的呢？
11:27
So where did I all start?
其实，我甚至算不上第一代AI研究者。
11:29
Well, I'm not even the first generation AI researcher.
艾伦·图灵在40年代就超前于时代，大胆地向人类发问：我们能不能造出会思考的机器？对吧。
11:33
The first generation really date back to the 50s and 60s
第一代其实可以追溯到五六十年代，艾伦·图灵在四十年代就超前于时代，大胆地向人类发问：我们能不能造出会思考的机器？对吧。当然，他有一套特定的方法来测试这个“思考机器”的概念，也就是一种对话式的聊天机器人——按他的标准来看，我们现在已经有了会思考的机器。
11:37
and Alan Turing was ahead of his time
当然，他有自己特定的方式。
11:41
by in the 40s by asking daring humanity
从40年代开始，通过向人类提出一个大胆的问题：
11:44
with the question, can we is there thinking machines, right?
“我们能否制造出会思考的机器？”
11:48
And of course, he has a specific way
当然，他有自己特定的方式。
11:51
of testing this concept of thinking machine
测试这个“思考机器”的概念
11:55
which is a conversational chapot
也就是一个对话式的聊天机器人
11:58
which to his standard, we now have a thinking machine
按照他的标准，我们现在确实有了一个思考机器
12:02
but that was just a more anecdotal inspiration.
但这只是一个轶事式的灵感来源。
12:08
The field really began in the 50s
这个领域真正始于50年代
12:11
when computer scientists came together
当时计算机科学家们聚集在一起
12:13
and look at how we can use computer programs
研究如何利用计算机程序
12:17
and algorithms to build these programs
和算法来构建这些程序
12:22
that can do things that have been only capable
使其能够完成那些此前只有
12:27
by human cognition.
人类认知才能做到的事情。
12:29
So, and that was the beginning
所以，那就是开端。
12:31
and the founding fathers, the Dartmouth,
然后就是那些奠基人，达特茅斯，1956年的那个研讨会。你知道，我们有John McCarthy教授，他后来去了斯坦福，也是他创造了artificial intelligence这个术语。从50年代、60年代、70年代到80年代，那都是AI探索的早期阶段，我们有了逻辑系统，也有了expert systems。
12:33
the workshop in the 1956.
1956年的那个研讨会
12:37
You know, we have Professor John McCarthy
你知道，约翰·麦卡锡教授
12:39
who later came to Stanford
他后来去了斯坦福
12:41
who coined the term artificial intelligence.
正是他创造了“人工智能”这个术语
12:45
And between the 50s, 60s, 70s and 80s,
从50年代、60年代、70年代到80年代
12:50
it was the early days of AI exploration
那是AI探索的早期阶段，我们有逻辑系统，也有专家系统。希望不仅仅在于机器本身，而是建立在视觉、感知、空间理解之上，而不仅仅是语言本身。我认为它们是互补的。所以我选择研究视觉智能。而我的博士阶段以及早期当教授的几年里，
12:53
and we had logic systems, we had expert systems.
我们有过逻辑系统，也有过专家系统。
12:58
We also had early exploration of neural network.
我们早期也探索过神经网络。
13:02
And then it came to around the late 80s, the 90s
然后到了80年代末、90年代，
13:07
and the very beginning of the 21st century.
以及21世纪初那段时间。
13:12
That stretch about 20 years
那大概20年的跨度
13:15
is actually the beginning of machine learning.
其实是机器学习的开端。
13:17
It's the marriage between computer programming
那是计算机编程
13:20
and statistical learning.
与统计学习的结合。
13:23
And that marriage brought a very, very critical concept
而这个结合带来了一个非常、非常关键的概念。
13:29
into AI which is that purely rule-based program
进入AI领域，纯粹基于规则的程序
13:34
is not a count for the vast amount of cognitive capabilities
根本无法涵盖我们想象中计算机能实现的绝大部分认知能力。
13:43
that we imagine computers can do.
所以，我们必须让机器去学习模式。
13:46
So, we have to use machines to learn the patterns.
一旦机器能够学习模式，
13:51
Once the machines can learn the patterns,
它就有希望做更多事情。
13:53
it has the hope to do more things.
比如，如果你给它三只猫，
13:56
For example, if you give it three cats,
希望不仅仅是让机器……
13:59
the hope is not just for the machines
希望不仅仅在于机器本身。
14:01
to recognize these three cats.
识别这三只猫。
14:03
The hope is the machines can recognize the fourth cat,
希望在于机器能识别第四只猫、
14:07
the fifth cat, the sixth cat, and all the other cats.
第五只猫、第六只猫，以及所有其他的猫。
14:10
And that's a learning ability that is fundamental to humans
而这种学习能力是人类
14:14
and many animals.
和许多动物所具备的基本能力。
14:16
And we as a field realize we need machine learning.
我们整个领域意识到需要机器学习。
14:21
So that was up till the beginning of the 21st century.
这就是直到21世纪初的情况。
14:26
I enter the field of AI literally in the year of 2000.
我进入AI领域正好是在2000年。
14:30
That's when my PhD began at Caltech.
那时我刚开始在Caltech读博。
14:33
And so I was one of the first generation machine learning researchers.
所以我算是第一代machine learning研究者。
14:37
And we were already studying this concept of machine learning,
我们当时已经在研究machine learning这个概念了，
14:41
especially in your network.
尤其是在neural network领域。
14:43
I remember that was one of my first courses
我记得那是我在Caltech最早修的课程之一，
14:46
in at Caltech is called your network.
就叫neural network。
14:49
But it was very painful.
但过程非常痛苦。
14:50
It was still smack in the middle of the so-called AI winter,
那时还正处在所谓的AI winter的寒冬期。
14:54
meaning the public didn't look at this too much.
也就是说，公众当时并没有太关注这个。
14:57
There wasn't that much funding.
资金也没那么多。
14:59
But there was also a lot of ideas flowing around.
但当时各种想法确实在大量涌现。
15:03
And I think two things happened to myself
而就我个人而言，有两件事
15:07
that brought my own career so close to the birth of modern AI
让我自己的职业生涯如此贴近现代AI的诞生。
15:12
is that I chose to look at artificial intelligence
第一是我选择通过视觉智能的视角
15:17
through the lens of visual intelligence.
来审视人工智能。
15:19
Because humans are deeply visual animals.
因为人类本质上是非常依赖视觉的动物。
15:24
We can talk a little more later, but so much of our intelligence
我们待会儿可以再聊，但人类智能很大一部分是建立在视觉、感知和空间理解上的，而不仅仅是语言本身。我认为它们是互补的。所以我选择研究视觉智能。在我的博士阶段和早期教授生涯中，我和我的学生们一直致力于一个“北极星”问题——解决物体识别这一难题。
15:29
is built upon visual, perceptual, spatial understanding,
它是建立在视觉、感知、空间理解之上的，
15:33
not just language per se.
而不仅仅是语言本身。
15:35
I think they're complementary.
我认为它们是互补的。
15:37
So I chose to look at visual intelligence.
所以我选择研究视觉智能。
15:39
And my PhD and my early professor years,
在我的博士阶段和早期教授生涯中，
15:44
I, my students and I are very committed to a North Star
我、我和我的学生们一直致力于一个“北极星”问题，那就是解决物体识别这一难题。我们环游世界，解读、推理并与世界互动，大致都是在物体层面进行的。我们不会在分子层面与世界互动，也不会像有时那样去互动——但很少会这样，比如，如果你想提起一个茶壶，你不会说：“好，这个茶壶是由100块瓷器组成的。”
15:48
problem, which is solving the problem of object recognition.
问题在于解决物体识别这个难题。我们在这个世界上活动，理解、推理，并与它互动，基本上都是在物体的层面上。我们不会在分子层面与世界互动。我们也不会像有时那样去互动，但很少会这样，比如你想提起一个茶壶，你不会说，好吧，这个茶壶是由100块瓷器组成的。物体识别的问题。
15:52
Because it's a building block for the perceptual world, right?
因为它是感知世界的基础模块，对吧？我们在世界里活动，理解、推理、互动，基本上都是在物体层面。我们不会在分子层面跟世界互动。我们不会像有时候那样跟世界互动，但很少会——比如，如果你想拿起一个茶壶，你不会说，好，这个茶壶由100块瓷器组成，然后让我来处理这100块。
15:56
We go around the world, interpreting, reasoning,
我们环游世界，进行解读、推理，
15:59
and interacting with it, more or less at the object level.
并与世界互动，大致是在物体层面。
16:03
We don't interact with the world at the molecular level.
我们不会在分子层面与世界互动。
16:07
We don't interact with the world as we sometimes do,
我们也不会像有时那样与世界互动，
16:11
but we rarely, for example, if you wanna lift a teapot,
但很少会——比如，如果你想提起一个茶壶，
16:15
you don't say, okay, the teapot is made of 100 pieces of porcelain.
你不会说，好，这个茶壶由100块瓷器组成。
16:19
And let me work on these 100 pieces.
让我来处理这100个样本。
16:22
You look at this as one object and interact with it.
你把它看作一个整体对象，然后与它互动。
16:25
So object is really important.
所以对象这个概念真的很重要。
16:27
So I was among the first researchers
我是最早一批研究者之一，
16:32
to identify this as a North Star problem.
把这个问题定义为“北极星”问题。
16:35
But I think what happened is that as a student of AI,
但我觉得，作为AI的学生，
16:41
and then a researcher of AI, I was working on all kinds
后来成为AI的研究者，
16:45
of mathematical models, including your network,
我一直在研究各种数学模型，
16:48
including Bayesian network, including many models,
包括神经网络、贝叶斯网络，还有很多其他模型。
16:53
and there was one singular pain point,
而且有一个非常突出的痛点，
16:56
is that these models don't have data to be trained up.
就是这些模型没有足够的数据来进行训练。
17:00
And as a field, we were so focusing on these models,
作为这个领域，我们当时太专注于模型本身了，
17:04
but it don't only have human learning,
但问题在于，不仅人类学习需要数据，
17:09
as well as evolution is actually a big data learning process.
进化本身其实也是一个大规模数据学习的过程。
17:15
Humans will learn with so much experience constantly
人类通过不断积累经验来学习，
17:19
and evolution if you look at time,
而进化，如果你从时间维度来看，
17:21
animals evolved with just experiencing the world.
动物也只是通过体验世界来演化的。
17:25
So I think my student and I conjectured
所以我和我的学生推测，让AI真正活起来的一个非常关键却被忽视的因素，就是大数据。于是我们在2006、2007年启动了那个图像数据项目。当时我们野心勃勃，想把整个互联网上关于物体的图像数据都收集起来。当然，那时候的互联网比现在小得多，所以我觉得这个野心至少不算太离谱。
17:29
that very critically overlooked ingredient
那个被严重忽视的关键要素
17:34
of bringing AI to life is big data.
就是让AI活起来的大数据。
17:37
And then we began this image data project in 2006, 2007.
然后我们在2006、2007年启动了图像数据项目。
17:42
We were very ambitious.
当时我们野心很大。
17:44
We wanna get the entire internet's image data on objects.
我们想获取整个互联网上关于物体的图像数据。
17:49
Now granted, internet was a lot smaller than today.
当然，那时的互联网比现在小得多。
17:52
So I feel like that ambition was at least not too crazy.
所以我觉得这个野心至少不算太疯狂。
17:57
Now it's totally delusional
现在完全是在妄想
17:59
to think a couple of grander students
几个研究生加一个教授就能搞定这事
18:03
and the professor can do this.
但当年我们就是这么干的
18:05
But and that's what we did.
我们从互联网上精心筛选了1500万张图片
18:07
We curated very carefully 15 million images
构建了一个包含22000个概念的分类体系
18:11
on the internet, created a taxonomy of 22,000 concepts,
借鉴了其他研究者的成果，比如语言学家对WordNet的研究
18:16
borrowing other researchers work like linguists work on wordnet.
这是一种特殊的词典词汇组织方式
18:23
And it's a particular way of dictionary words.
这是一种特定的字典词汇方式。
18:28
And we combine that into image data
我们将这些整合成图像数据，
18:31
and we open source that to the research community,
并开源给研究社区，
18:34
we held an annual image that challenged
我们每年举办一次图像挑战赛，
18:38
to encourage everybody to participate in this.
鼓励大家参与其中。
18:42
We continue to do our own research.
我们持续进行自己的研究。
18:44
But 2012 was the moment that many people think
但2012年，很多人认为
18:48
was the beginning of the deep learning
是深度学习的起点，
18:50
or birth of modern AI
或者说现代AI诞生的时刻。
18:52
because a group of Toronto researchers
因为一群多伦多的研究人员，在Jeff Hinton教授的带领下，参加了那个图像挑战赛，用了ImageNet大数据和两块NVIDIA的GPU，成功创建了第一个神经网络算法——虽然没能完全解决，但在物体识别问题上取得了巨大进展。
18:54
led by Professor Jeff Hinton
由杰夫·辛顿教授主导，
18:57
participated in image that challenge,
参与了图像识别挑战，
18:59
used the image that big data and two GPUs from NVIDIA
使用了图像大数据和两块NVIDIA的GPU，
19:05
and created successfully the first neural network algorithm
成功创建了第一个神经网络算法，
19:09
that can, it didn't fully solve
虽然未能完全解决，
19:14
but made a huge progress towards solving the problem
但在物体识别问题上取得了巨大进展。
19:18
of object recognition.
这是物体识别。
19:20
And that combination of the trial technology,
而那种试验技术、大数据神经网络和GPU的结合，
19:24
big data neural network and GPU
可以说是现代AI的黄金配方。
19:27
was kind of the golden recipe for modern AI.
再快进到AI的公众时刻，
19:32
And then fast forward the public moment of AI
也就是ChatGPT时刻。
19:39
which is the chat GPD moment.
如果你看看是什么要素
19:41
If you look at the ingredients
让ChatGPT问世，
19:44
of what brought chat GPD to the world,
从技术上讲，它依然使用了这三个要素。
19:49
technically it still used these three ingredients.
技术上来说，它仍然使用了这三个要素。
19:53
Now its internet scale data mostly texts
现在的互联网规模数据主要是文本，
19:57
is a much more complex neural network
神经网络架构比2012年复杂得多，但本质上还是神经网络，
20:01
architecture than 2012 but it's still neural network
GPU数量也多了很多，但本质上还是GPU。
20:05
and a lot more GPUs but it's still GPUs.
所以这三个核心要素依然是现代AI的基础。
20:09
So these three ingredients are still at the core
太不可思议了。
20:14
of modern AI.
我从来没听过完整的来龙去脉。
20:16
Incredible.
太不可思议了。
20:17
I have never heard that full story before.
我从来没听过完整的这个故事。
20:19
I love that it was two GPUs was the first.
我喜欢它最初用的是两块GPU。
20:23
I love that.
我真的很喜欢。
20:24
Yeah.
是啊。
20:25
And now it's set on hundreds of thousands right
而现在它建立在数十万块GPU之上，
20:27
that are orders of magnitudes more powerful.
性能提升了几个数量级。
20:31
And those two GPUs where they were like gaming GPUs
而那两块GPU还是游戏用的显卡，
20:34
they just went to the like the game star right
他们直接去游戏商店买的，
20:35
they were people used for playing games.
就是人们用来打游戏的那种。
20:37
As you said this continues to be in a large way
正如你所说，这在很大程度上仍然是模型变聪明的方式。
20:39
the way models get smarter.
现在全球增长最快的几家公司，
20:41
Some of the fastest growing companies in the world right now
我基本都请来过播客，
20:43
I've had them all mostly on the podcast,
像Merchord、Surgeon scale。
20:45
Merchord, Surgeon scale.
它们就是这么做的，持续为实验室提供更多标注数据，
20:47
Like they do this, they continue to do this for labs
都是它们最看好的方向。
20:49
just give them more and more label data
只要给它们越来越多的标注数据
20:51
of the things they're most excited about.
关于他们最兴奋的那些事情。
20:52
Oh yeah I remember Alex Wang from scale very early days
哦对对，我记得Scale的Alex Wang，很早很早的时候。
20:57
I probably still has his emails when he was starting scale.
我可能还留着他刚创办Scale时发的邮件。
21:00
He was very kind, he keeps sending me emails
他特别客气，一直给我发邮件
21:04
about how you met that inspired scale.
讲他怎么遇到那些启发Scale的事。
21:07
I was very pleased to see that.
看到这些我真的很高兴。
21:09
One of my other favorite takeaways from what you just shared
你刚才分享的内容里，另一个我特别喜欢的点
21:11
is just such an example of high agency
就是这种高度自主性的例子
21:14
and just doing things that's kind of a meme on Twitter
还有那种在Twitter上都快成梗的做事方式。
21:16
just you can just do things you're just like
你就直接做你想做的事，比如
21:18
okay this is probably necessary to move AI
“好，这可能是推动AI发展的必要步骤”，
21:21
and it's called machine learning back then right.
那时候它叫机器学习，对吧。
21:23
Was that the term most people used?
当时大多数人用这个术语吗？
21:25
I think it was interchangeably, it's true.
我觉得是混着用的，确实。
21:27
Like I do remember the companies, the tech companies
我记得那些公司，那些科技公司——
21:31
I'm not gonna name names but I was in a conversation
我就不点名了——但我参与过一次对话，
21:36
one of the early days I think is in the middle of 2015,
那是在早期，大概是2015年年中的时候。
21:41
middle of 2016, some tech companies avoid using the word AI
2016年年中，有些科技公司刻意避开“AI”这个词，因为不确定它是不是个敏感词。我记得当时我还在鼓励大家用“AI”，因为对我来说，这是人类在科学和技术探索中提出过最大胆的问题之一。我为这个词感到非常自豪，但确实，一开始有些人心里没底。
21:47
because they were not sure if AI was a dirty word
因为他们不确定AI是不是一个敏感词
21:51
and I remember I was actually encouraging everybody
而且我记得我当时其实在鼓励每个人
21:55
to use the word AI because to me that is one of the most
使用AI这个词，因为对我来说，这是最……
21:59
audacious question humanity has ever asked
人类在科学与技术的探索中提出的最大胆的问题
22:03
in our quest for science and technology
我对这个说法感到非常自豪，但没错
22:06
and I feel very proud of this term but yes
一开始有些人并不确定
22:10
at the beginning some people were not sure.
好的
22:12
What year was that roughly when AI was already working?
那大概是哪一年，AI 已经开始有用了？
22:14
2016, I think it was less than 10 years ago.
2016 年吧，我想还不到十年前。
22:17
That was the changing like some people start calling it AI
那时候开始有变化，有些人开始管它叫 AI。
22:23
but I think if you look at the Silicon Valley tech companies
但我觉得如果你看看硅谷的科技公司，
22:28
if you trace their marketing term,
追溯他们的营销术语，
22:31
I think 2017-ish was the beginning of companies
大概 2017 年左右，公司才开始
22:37
calling themselves AI companies.
自称是 AI 公司。
22:40
That's incredible, just how the world has changed.
真不可思议，世界变化太大了。
22:43
Now you can't not call yourself an AI company.
现在你没法不把自己叫做AI公司了。
22:46
I know.
我知道。
22:47
Nine-ish years later.
大概九年之后吧。
22:48
Yeah.
对。
22:49
Oh man.
天哪。
22:50
Okay, is there anything else around the history?
好，关于这段历史还有什么别的吗？
22:53
That early history that you think people don't know,
那段早期历史里，你觉得有什么是人们不知道的，
22:54
that you think is important before we chat about
或者在我们聊之前你觉得很重要的？
22:57
where things are going and the work that you're doing?
事情的发展方向和你正在做的工作是怎样的？
23:01
I think as all histories, I'm keenly aware
我觉得，就像所有历史一样，我清楚地意识到
23:05
that I am recognized for being part of the history
自己因为参与了这段历史而被认可，
23:09
but there are so many heroes and so many researchers.
但背后有太多英雄和研究者。
23:12
We're talking about generations of researchers.
我们说的是几代研究者。
23:15
They're in my own world, there are so many people
在我自己的领域里，有太多人
23:19
who have inspired me which I talked about in my book
曾启发过我，我在书里也提到过他们，
23:24
but I do feel our culture,
但我确实感受到我们的文化，
23:27
especially the Silicon Valley,
尤其是硅谷那边，
23:29
tends to assign achievements to a single person.
总喜欢把功劳归到某一个人身上。
23:35
Well, while I think it has value,
嗯，虽然我觉得这也有它的意义，
23:38
but it's just to be remembered,
但说白了，只是为了让人记住而已。
23:41
AI is a field of at this point 70 years old
AI这个领域到现在已经有70年了，
23:44
and we have gone through many generations.
我们经历了好几代人的努力。
23:48
Nobody, no one could have gotten here by themselves.
没有任何人，没有任何一个人能单靠自己走到今天这一步。
23:53
Okay.
为了谋生而赚取薪水
23:54
Let me ask you this question.
让我问你一个问题。
23:56
It feels like we're always on this precipice of AI,
感觉我们好像一直站在AI的悬崖边上，
23:58
this kind of vague term, people throw around,
这个模糊的词，大家随口就说，
24:00
AI is coming, it's gonna take over everything.
AI要来了，它会接管一切。
24:03
How, what's your take on?
你怎么看？
24:05
How far do you think we might be from AI?
你觉得我们离AI还有多远？
24:06
Do you think we're gonna get there on the current trajectory?
你认为按现在的轨迹能到达吗？
24:09
Ron, do you think we need more break views?
Ron，你觉得我们需要更多突破性的观点吗？
24:10
Do you think the current approach will get us there?
你觉得现在的路径能带我们到那儿吗？
24:13
Yeah, this is a very interesting term, Lenny.
嗯，Lenny，这个说法挺有意思的。
24:17
I don't know if anyone has ever defined AI.
我不确定有没有人真正定义过AI。
24:22
You know, there are many different definitions,
你知道，定义有很多种，
24:26
including some kind of superpower for machines
从机器拥有某种超能力，
24:30
all the way to can machines become economically viable
到机器能否成为社会中经济上可行的
24:36
agents in the society in other words,
agent——换句话说，就是赚工资养活自己。
24:41
making salaries to live.
我不知道AI和AGI之间有什么区别
24:43
Is that a definition of AI?
这是对AI的定义吗？
24:45
As a scientist, I take science very seriously
作为一名科学家，我非常认真地对待科学，
24:49
and I enter the field because I was inspired
我进入这个领域是因为我被这个大胆的问题所启发：
24:53
by this audacious question of chemistings,
让机器像人类一样思考和做事。
24:57
think and do things in the way that human can do.
对我来说，这始终是AI的北极星。
25:02
For me, that's always the North start of AI.
从这个角度来看，
25:05
And from that point of view,
我不知道AI和AGI之间有什么区别。
25:07
I don't know what's the difference between AI and AGI.
所以我不想陷入无休止的讨论
25:10
I think we've done very well in achieving parts of the goal,
我觉得我们在实现部分目标上做得很好，
25:16
including conversational AI,
包括对话式AI，
25:18
but I don't think we have completely conquered
但我不认为我们已经完全攻克了
25:21
all the goals of AI.
AI的所有目标。
25:23
And I think our founding fathers, the Alan Turing,
而且我觉得我们的奠基人，比如Alan Turing，
25:27
I wonder if Alan Turing is around today
我在想如果Alan Turing今天还在，
25:30
and you ask him to contrast the AI versus AGI.
你让他对比一下AI和AGI，
25:34
Tim, I just shrugged and said,
Tim，我只会耸耸肩说，
25:36
well, I asked the same question back in 1940s.
嗯，我在1940年代就问过同样的问题。所以我不想陷入定义AI和AGI区别的兔子洞。我觉得AGI更像一个营销术语，而不是科学术语。作为科学家和技术人员，AI是我的北极星，也是我所在领域的北极星，大家爱怎么叫就怎么叫，我无所谓。
25:40
So I don't want to get on to a rabbit hole
所以我也不想钻牛角尖，
25:45
of defining AI versus AGI.
定义AI和AGI的区别。
25:47
I feel AGI is more a marketing term
我觉得AGI更像一个营销术语，
25:50
than a scientific term.
而不是科学术语。
25:52
As a scientist and technologist,
作为科学家和技术人员，
25:55
AI is my North star, it is my field's North star
AI是我的北极星，也是我所在领域的北极星，
25:59
and I'm happy people call it whatever name they want to call it.
我很乐意人们随便怎么称呼它。
26:05
So let me ask you, maybe this way,
那我来问你，也许这样问比较好，
26:07
like you described, there's kind of these components
就像你描述的，有这些组成部分，
26:09
that from ImageNet and AlexNet kind of took us
从ImageNet和AlexNet一路把我们带到了今天，
26:12
to where we're today, GPUs, essentially data, label data,
GPU，本质上就是数据、标注数据，
26:17
just like the algorithm of the model.
还有模型的算法。
26:19
There's also just a transformer,
另外还有transformer，
26:21
feels like an important step in that trajectory.
感觉在这条发展路径上是很重要的一步。
26:24
Do you feel like those are the same components
你觉得这些是同样的组成部分吗？
26:26
that'll get us to, I don't know, 10 times smarter model,
这大概能让我们得到一个，怎么说，聪明10倍的模型，
26:29
something that's like life changing for the entire world?
一个能彻底改变整个世界的东西？
26:32
Or do you think we need more breakthroughs?
还是说你觉得我们需要更多突破？
26:34
I know we're going to talk about world models,
我知道我们接下来要聊世界模型，
26:35
which I think is a component of this,
我觉得这是其中的一部分，
26:37
but is there anything else that you think is like,
但还有没有其他你觉得像是，
26:39
oh, this little plateau, or okay, this will take us,
哦，这个小瓶颈，或者说，好吧，这个能带我们往前走，
26:42
just need more data, more compute, more GPUs?
只需要更多数据、更多算力、更多GPU？
26:44
Oh no, I definitely think we need more innovations.
哦不，我绝对认为我们需要更多创新。我觉得在更多数据、更多GPU以及更大的现有模型架构上的scaling loss方面，仍然有很多工作要做，但我完全认为我们需要更多创新。人类历史上没有任何一个深层的科学学科，会走到一个说“我们完成了，我们不再创新了”的地步。
26:47
I think scaling loss of more data, more GPUs
我认为scaling loss就是更多数据、更多GPU
26:51
and bigger current model architecture
以及更大的现有模型架构。
26:55
is there's still a lot to be done there,
这方面还有很多工作要做，
26:58
but I absolutely think we need to innovate more.
但我绝对认为我们需要更多创新。
27:01
There's not a single deeply scientific discipline
人类历史上没有任何一门深奥的科学学科
27:06
in human history that has arrived at a place
曾达到过“我们完成了，创新结束了”的地步。
27:10
that says we're done, we're done innovating.
即便不是人类文明中最年轻的学科，
27:13
And AI is one of the,
而AI，即便不是人类文明中最年轻的学科，也绝对是最年轻的之一。从科技层面来看，我们仍只是触及皮毛。比如，就像我刚才说的，我们接下来要聊到世界模型。现在你拿一个模型，让它看一段办公室房间的视频。
27:16
if not the youngest discipline in human civilization,
从科学与技术的角度来看，
27:20
in terms of science and technology,
我们也只是刚刚触及皮毛。
27:22
we're still scratching the surface.
我们目前还只是触及皮毛。
27:25
For example, like I said,
比如说，就像我说的，
27:27
we're going to segue into world models.
我们接下来要聊到world models。
27:29
Today you take a model and run it
现在你拿一个模型，让它看一段
27:34
through a video of a couple of office rooms
办公室房间的视频，
27:38
and ask the model to count the number of chairs.
让模型数一下有多少把椅子。
27:41
And this is something a toddler could do,
这事儿一个刚会走路的小孩都能做到，
27:43
or maybe an elementary school kid could do.
或者可能一个上小学的孩子也能做到。
27:48
And AI could not do that, right?
但AI就是做不到，对吧？
27:50
So there's just so much AI today could not do.
所以今天AI做不到的事情实在太多了。
27:53
Then let alone thinking about how did you know,
更别提去思考像艾萨克·牛顿这样的人
27:59
someone like Isaac Newton
是怎么通过观察天体的运动来发现规律的。
28:01
look at the movements of the celestial bodies
观察天体的运动——
28:04
and derive an equation,
并推导出一个方程，
28:08
or a set of equations that governs the movement of all bodies.
或一组支配所有天体运动的方程。
28:13
That level of creativity, extrapolation, abstraction,
这种创造力、外推能力、抽象能力，
28:18
we have no way of enabling AI to do that today.
我们目前没有任何办法让AI做到。
28:23
And then let's look at emotional intelligence.
再来看看情商。
28:25
If you look at a student coming to a teacher's office
想象一个学生走进老师的办公室，
28:30
and have a conversation about motivation, passion,
聊关于动力、热情、
28:34
what to learn, what's the problem that's really bothering you.
该学什么、真正困扰你的问题是什么。
28:41
That conversation, as powerful as today's conversational bots
那段对话，即便现在这些对话机器人已经很强大了，你也无法从今天的AI身上获得那种情感认知智能。所以我们还有很多可以改进的地方。而且我不认为我们的创新已经到头了。Demis最近有个很有意思的采访，来自DeepMind和Sasha Google，当时有人问他，就像这样——
28:46
are you don't get that level
但如今的AI还达不到那种
28:48
of emotional cognitive intelligence from today's AI.
情感认知智能的水平。
28:53
So there's a lot we can do better.
所以还有很多可以改进的地方。
28:57
And I do not believe we're done innovating.
而且我不认为我们的创新已经到头了。
29:00
Demis had this really interesting interview recently
Demis最近有个很有意思的采访
29:02
from DeepMind and Sasha Google
来自DeepMind和Sasha Google
29:03
where someone asked him just like,
当时有人问他，就是
29:04
what do you think, how far are we from AGI?
你觉得呢，我们离AGI还有多远？
29:07
What does it look like?
它看起来是什么样的？
29:08
He went through there.
他当时走过了那条路。
29:08
He had a really interesting way of approaching it
他处理这个问题的方式非常有意思，
29:10
is if we were to give the most cutting edge model
就是如果我们给最前沿的模型
29:13
all the information until the end of the 20th century,
所有直到20世纪末的信息，
29:17
see if it could come up with all the breakthroughs Einstein had
看它能否独立推导出爱因斯坦的所有突破，
29:20
and so far we're never near that.
而到目前为止，我们离那一步还差得很远。
29:22
No, we're not.
不，我们并没有。
29:23
In fact, it's even worse.
实际上，情况更糟。
29:25
Let's give AGI all the data,
我们把所有数据都给AGI，
29:29
including modern instruments data of celestial bodies
包括牛顿当时没有的现代天体观测数据，
29:33
which Newton did not have and give it to that
然后让AGI去推导出17世纪关于物体运动定律的方程组。
29:36
and just ask AGI to create the 17th century set of equations
如今的AGI已经能做到这一点了。
29:42
on the laws of bodily movements.
关于身体运动规律的问题。
29:46
Today's AGI can now do that.
如今的AGI已经能做到这一点了。
29:49
All right, we're away the way.
好的，我们开始了。
29:51
Okay, so let's talk about world models.
好，那我们来聊聊世界模型。
29:53
To me, this is just another really amazing example
对我来说，这又是一个绝佳的例子，
29:56
of you being ahead of where people end up.
说明你总是比大家领先一步。
30:00
So you were way ahead on, okay,
你早就领先于大家，认为，
30:02
we just need a lot of clean data for AGI
我们只需要大量干净的数据给AGI
30:05
and neural networks to learn.
和神经网络去学习。
30:07
You've been talking about this idea of world models
你一直在谈论世界模型这个概念。
30:08
for a long time.
很久以来。
30:09
You started a company to build.
你创办了一家公司，目标是构建。
30:11
Essentially, there's language models.
本质上，就是language models。
30:12
This is a different thing.
这是另一回事。
30:13
This is a world model.
这是一个world model。
30:14
We'll talk about what that is.
我们会聊聊那到底是什么。
30:16
And now, as I was preparing for this,
而就在我为此做准备的时候，
30:18
Elon's like talking about world models,
Elon 也在谈论world models。
30:19
Jensen's talking about world models.
Jensen 在聊世界模型（world models）的事。
30:21
I know Google's working on this stuff.
我知道 Google 也在搞这些。
30:22
You've been at this for a long time.
你在这个领域已经深耕很久了。
30:24
And you actually just launched something
而且你刚刚发布了一个东西，
30:25
that's gonna, we're gonna talk about right
我们打算在这期播客上线前聊聊它。
30:28
before this podcast airs.
说说看，什么是世界模型（world model）？
30:31
Talk about what is a world model?
为什么它这么重要？
30:32
Why is it so important?
为什么这如此重要？
30:33
I'm very excited to see that more and more people
我很兴奋地看到，越来越多的人
30:36
are talking about world models like Elon, like Jensen.
开始讨论世界模型，比如Elon，比如Jensen。
30:43
I have been thinking about really how to push AGI
我这一生都在思考如何真正推动AGI
30:48
forward all my life, right?
向前发展，对吧？
30:51
And the large language models that came out
而过去几年从研究界涌现出来的大语言模型，
30:55
of the research world and then OpenAI and all this,
以及OpenAI等等，
31:00
for the past few years, were extremely inspiring,
即使对我这样的研究者来说，也极其鼓舞人心。
31:04
even for a researcher like me.
即使对我这样的研究者来说也是如此。
31:07
I remembered when GPT2 came out,
我记得GPT2刚出来的时候，
31:11
and that was in, I think late 2020,
大概是2020年底吧，
31:15
I was co-director, I still am,
我当时是联合主任，现在也还是，
31:19
but I was, at that time, full-time co-director
但那时候我是全职的联合主任，
31:22
of Stanford's Human Center AI Institute.
在斯坦福大学以人为本AI研究院。
31:25
And I remember it was, you know,
我记得那时候，
31:28
the public was not aware of the power
公众还没意识到
31:30
of the large language model yet,
大语言模型有多强大。
31:33
but as researchers, we were seeing it,
但作为研究者，我们当时已经看到了，
31:35
we're seeing the future.
我们正在目睹未来。
31:36
And I had pretty long conversations
我和我的自然语言处理同事
31:39
with my natural language processing colleagues
比如Percy Liang和Chris Batting，
31:43
like Percy Liang and Chris Batting.
有过很长的对话。
31:45
We were talking about how critical
我们当时在讨论这项技术
31:48
this technology is gonna be,
会有多关键，
31:50
and Stanford AI Institute, Human Center AI Institute,
还有Stanford AI Institute、Human Center AI Institute。
31:53
AGI was the first one to establish a full research center
AGI 是第一个建立完整研究中心来做 foundation model 的机构。
31:58
on foundation model.
当时是 Percy Liang 和很多研究员一起发表了第一篇关于 foundation model 的学术论文。
31:59
We were Percy Liang and many researchers
所以这对我来说特别有启发。
32:02
led the first academic paper foundation model.
当然，我本身是做 visual intelligence 出身的，
32:06
So it was just very inspiring for me.
当时就在想，
32:10
So of course, I come from the world of visual intelligence
除了语言之外，我们还有很多可以推进的方向。
32:14
and I was just thinking,
我就在想，
32:15
there's so much we can push forward on beyond language
语言之外还有太多可以推进的东西——
32:20
because humans have used our sense
因为人类一直依赖我们的空间智能——一种对世界的理解——来完成那么多事情，而这些是超越语言的。想象一个非常混乱的急救现场，无论是火灾、交通事故还是自然灾害。如果你置身于那样的场景中——
32:28
of spatial intelligence,
空间智能，
32:30
a world understanding to do so many things,
对世界的理解，能做的事情太多了，
32:33
and they are beyond language.
而且这些都超越了语言。
32:36
Think about a very chaotic first responder scene,
想象一下一个非常混乱的应急现场，
32:41
whether it's fire or some traffic accident
不管是火灾、交通事故，
32:44
or some natural disaster.
还是自然灾害。
32:47
And it's, if you immerse yourself in a scene
如果你完全沉浸在一个场景里，
32:52
and think about how people organize themselves
想想人们是如何自发组织起来的——
32:55
to rescue people, to stop further disasters,
去救人、阻止灾害蔓延、扑灭大火，
32:59
to put down fires, to a lot of that is movements,
这些很大程度上是行动，
33:05
is spontaneous understanding of objects worlds,
是对物体世界和人类情境的即时理解。
33:10
human situational awareness.
语言是其中的一部分，
33:15
Language is part of that,
但在很多这类情境中，
33:16
but a lot of those situations,
光靠语言是没法让你去把火扑灭的。
33:18
language cannot get you to put down the fire.
语言无法让你放下火。
33:21
So that is what is that?
所以那到底是什么？
33:24
I was thinking a lot,
我想了很多，
33:25
and in the meantime, I was doing a lot of robotics research
与此同时，我也做了大量机器人研究，
33:29
and it dawned on me that the linchpin of connecting
然后我突然意识到，连接额外智能的关键，
33:35
the additional intelligence,
除了语言和连接具身AI（也就是机器人）之外，
33:39
in addition to language and connecting embodied AI,
连接视觉智能的核心其实是空间智能。
33:44
which are robotics, connecting visual intelligence
那就是机器人技术，连接视觉智能，
33:48
is the sense of spatial intelligence
是空间智能的感知，
33:51
about understanding the world.
关于理解世界。
33:53
And that's when, I think I, it was 2024,
那是在，我想想，2024年，
33:58
I gave a TED Talk about spatial intelligence
我做了一个关于空间智能和世界模型的TED演讲。
34:01
and the world models.
这个想法最早是在2022年开始成型的，
34:03
And I started formulating this idea back in 2022,
基于我在机器人和计算机视觉方面的研究。
34:10
based on my robotics and computer vision research.
当时有一件事让我特别清楚，
34:14
And then one thing that was really clear to me
就是我真的很想和最顶尖的技术专家一起工作。
34:17
is that I really wanna work with the brightest technologist
是我真的很想和最聪明的技术专家合作，
34:23
and move as fast as possible to bring this technology to life.
并且以最快的速度将这项技术变为现实。
34:28
And that's when we found it this company called World Labs.
就在那时，我们发现了这家叫World Labs的公司。
34:32
And you can see the word world is in the title of our company
你可以看到“世界”这个词就在我们公司的名字里，
34:37
because we believe so much in world modeling
因为我们非常相信世界建模
34:39
and spatial intelligence.
和空间智能。
34:41
People are so used to just chat bots
人们已经习惯了聊天机器人，
34:43
and that's a large language model.
而那是大语言模型。
34:44
The simple way to understand a world model is you
理解世界模型的一个简单方式是，你
34:46
basically describe a scene
基本上就是描述一个场景，然后它就能生成一个无限探索的世界。我们会把你要启动的那个东西的链接放出来，这个我们后面会聊到。但这样理解是不是太简单了？这只是其中一部分，Lenny。我觉得理解世界模型的一个简单方式就是，这个模型能让任何人去创造。
34:48
and it generates an infinitely explorer world.
它会生成一个无限探索的世界。
34:52
We'll link to the thing you'll launch,
我们会链接到你即将推出的东西，
34:54
which we'll talk about,
这个我们待会儿会聊到。
34:54
but just is that a simple way to understand it?
但这就是理解它的简单方式吗？
34:56
That's part of it, Lenny.
这只是其中一部分，Lenny。
34:57
I think a simple way to understand a world model
我觉得理解世界模型的一个简单方式
35:01
is that this model can allow anyone to create
是它能让任何人去创造，
35:08
any world in their minds eye
在他们脑海中想象的任何世界里，
35:10
by prompting whether it's image or sentence
通过提示——无论是图像还是句子——
35:15
and also be able to interact in this world,
并且能够在这个世界中互动，
35:18
whether you're browsing and walking
无论是浏览、行走，
35:21
or picking objects up or changing things
还是拿起物体或改变事物，
35:27
as well as to reason within this world.
以及在这个世界中进行推理。
35:30
For example, if the person consuming,
例如，如果消费这个内容的人，
35:34
if the agent consuming this output of the world model
如果智能体在消费这个世界模型的输出，
35:38
is a robot, it should be able to plan its path
是一个机器人，它应该能够规划自己的路径，比如帮忙整理厨房。所以world model是一个基础，你可以用它来推理、交互和创造世界。没错，机器人感觉这可能是AI研究者下一个重要的关注点。
35:41
and help to tidy the kitchen, for example.
比如帮忙整理厨房。
35:47
So world model is a foundation
所以世界模型是一个基础，
35:53
that you can use to reason, to interact
你可以用它来推理、互动
35:57
and to create worlds.
并创造世界。
36:00
Great, yeah.
好的，是的。
36:00
So robots feels like that's potentially
所以机器人感觉上可能是
36:04
the next big focus for AI researchers
AI 研究者下一个大方向，
36:07
and just like the impact on the world
以及这对世界的影响，
36:08
and what you're saying here is,
你刚才说的其实是，
36:11
this is a key missing piece of making robots
让机器人在现实世界中真正运作，
36:14
actually work in the real world,
理解世界如何运转，
36:15
understanding how the world works.
这是一个关键缺失的部分。
36:17
Yeah, well, first of all,
对，首先，
36:18
I do think there's more than robots
我觉得令人兴奋的不只是机器人。
36:20
that's exciting.
这挺让人兴奋的。
36:23
So but I agree with everything you just said.
我完全同意你刚才说的每一点。
36:25
I think world modeling and spatial intelligence
我觉得世界建模和空间智能
36:28
is a key missing piece of embody AI.
是具身智能里缺失的关键一块。
36:33
I also think let's not underestimate
但我也想说，别低估了
36:35
that humans are embodied agents
人类本身就是具身智能体，
36:38
and humans can be augmented by AI's intelligence
而且人类可以被AI的智能增强——
36:43
just like today humans are language animals
就像现在人类是语言动物，
36:46
but we're very much augmented by AI
但我们很大程度上已经被AI增强了。
36:50
when helping us to do language tasks
当它们帮助我们完成语言任务时，包括软件工程。我认为我们不应该低估——或者说我们往往不太谈论——人类作为具身智能体，其实可以从世界模型和空间智能模型中获益良多，就像机器人一样。所以这里的关键突破在于机器人。
36:53
including software engineering.
包括软件工程。
36:55
I think that we shouldn't underestimate
我觉得我们不应该低估——
36:58
or maybe it's we tend not to talk about
或者说我们往往不太会去讨论——
37:03
how humans as an embodied agents
人类作为具身智能体（embodied agents）这一点。
37:06
can actually benefit so much from world models
实际上可以从世界模型和空间智能模型以及机器人中获益良多。所以这里最大的突破口就是机器人。而创造力感觉就像是一种乐趣，如果你看看罗莎琳德·富兰克林捕捉到的最重要的一张照片，那是一张平面的二维结构照片，看起来就像一个十字形。
37:10
and spatial intelligent models as well as robots can.
而空间智能模型以及机器人也能做到。
37:15
So the big on-locks here are robots
这里最大的突破点在于机器人。
37:17
which a huge deal if this works out
这要是能成，那可太了不起了。
37:20
and imagine each of us has robots doing a bunch of stuff
想象一下，我们每个人都有自己的机器人，帮我们干一堆活儿。
37:22
for us goes into, you know,
它们还能帮我们应对灾难，比如游戏里的那些场景。
37:23
they help us with disasters, things like games.
显然，这是个特别酷的例子。
37:26
Obviously is a really cool example
就像那些永远玩不腻的游戏，
37:27
just like infinitely playable games
你脑子里随便一想就能创造出来。
37:30
that you just invent at your head.
然后创造力本身感觉就像是一种乐趣。
37:31
And then creativity feels like just like being fun,
然后创造力感觉就像是一种乐趣，
37:34
having fun, being creative,
享受乐趣，发挥创意，
37:35
thinking of wild new worlds and environments.
构想狂野的新世界和新环境。
37:39
And also design humans design from machines
还有设计——从机器到建筑再到住宅的人类设计，
37:42
to buildings to homes
以及科学发现，对吧？
37:45
and also scientific discovery, right?
有太多例子我喜欢用，
37:47
There is so much I like to use the example
比如DNA结构的发现。
37:51
of the discovery of the structure of DNA.
如果你看其中一个最重要的部分——
37:55
If you look at one of the most important piece
如果你看其中最重要的一张图，
37:59
in DNA's discovery history is the X-ray diffraction photo
在DNA的发现史上，有一张由Rosalyn Franklin拍摄的X射线衍射照片，那是一张平面的2D照片，显示的结构看起来像一个十字，带有衍射条纹，你可以去Google搜那些照片。但凭借那张2D平面照片，人类——尤其是两位重要人物，James Watson和Francis Crick——
38:05
that was captured by Rosalyn Franklin
那是罗莎琳·富兰克林拍下的，
38:07
and it was a flat 2D photo of a structure
它只是一张平面的二维结构照片。
38:11
that looks like it looks like a cross
这看起来像是把邮件和通话功能整合进了他们的产品。
38:13
with diffractions, you can Google those photos.
有了衍射图像，你可以去Google那些照片。
38:18
But with that, 2D flat photo,
但就那张二维平面照片而言，
38:23
humans, especially two important humans,
人类，尤其是两位重要人物——
38:26
James Watson and Francis Crick,
James Watson和Francis Crick，
38:29
in addition to their other information
除了他们掌握的其他信息之外，
38:32
was able to reason in 3D space and deduce
他们能够在三维空间中进行推理，并推导出
38:38
a highly three-dimensional double helix structure
DNA高度立体的双螺旋结构。
38:42
of the DNA.
而那个结构绝不可能是二维的。
38:43
And that structure cannot possibly be 2D.
你无法在二维思维中推导出那个结构。
38:47
You cannot think in 2D and deduce that structure.
你必须用三维空间思维，
38:52
You have to think in 3D spatial,
运用人类的空间智能。
38:56
use the human spatial intelligence.
运用了人类的空间智能。
38:59
So I think even in scientific discovery,
所以我认为即使在科学发现中，空间智能或AI辅助的空间智能也至关重要。这正是一个例子，我记得有评论家说过，下一个大事件刚开始时总会让人觉得像个玩具。当ChatGPT刚出来时，如果我没记错的话，Sam Wampin发了一条推文。
39:02
spatial intelligence or AI assisted spatial intelligence
空间智能，或者说AI辅助的空间智能，
39:06
is critical.
至关重要。
39:07
This is such an example of, I think it was critics
我认为这就是一个例子，批评者们大概会这么说。
39:10
and they have this line that the next big thing
他们有一句话是：下一个大事件
39:12
is gonna start off feeling like a toy.
一开始会让人觉得像个玩具。
39:15
When ChatGPT just came out,
ChatGPT刚出来那会儿，
39:17
if I remember Sam Wampin just tweeted,
我记得Sam Wampin发过一条推文，
39:18
it's like here's a cool thing we're playing with,
就像这样：“嘿，我们正在玩一个很酷的东西，来看看吧。”
39:20
check it out.
现在它成了史上增长最快的产品，改变了世界。
39:21
Now it's the fastest growing product
而往往正是那些看起来“哦，这个挺有意思，玩起来很有趣”的东西，最终最能改变世界。
39:22
in all of history, change the world.
是啊。
39:24
And it's oftentimes the things that just look like,
往往是那些看起来——
39:27
okay, this is cool, that it's a fun to play with
“哦，这个挺酷的，玩起来很有意思”的东西，
39:30
and end up changing the world most.
最终最能改变世界。
39:32
Yeah.
对。
39:33
This episode is brought to you by Sinch,
本集由Sinch赞助播出，
39:35
the customer communications cloud.
客户通信云服务商。
39:37
Here's the thing about digital customer communications,
关于数字客户通信，关键在于——
39:40
whether you're sending marketing campaigns,
无论你是发送营销活动、
39:42
verification codes or account alerts,
验证码还是账户提醒，
39:44
you need them to reach users reliably.
都需要确保它们可靠地触达用户。
39:46
That's where Sinch comes in.
这正是Sinch的用武之地。
39:48
Over 150,000 businesses,
超过15万家企业，
39:51
including eight of the top 10 largest tech companies
包括全球前十大科技公司中的八家，都在用 Sinch 的 API 把短信、邮件和通话功能集成到自己的产品里。而消息领域正发生一件大事，产品团队必须了解——富通信服务，也就是 RCS。你可以把 RCS 想象成 SMS 2.0。不再是收到来自随机号码的纯文本，
39:53
globally use Sinch's API to build messaging,
全球范围内，开发者通过 Sinch 的 API 将短信、邮件和通话功能集成到他们的产品中。而在消息领域，有一个重大变化产品团队需要了解——富通信服务，也就是 RCS。你可以把 RCS 想象成 SMS 2.0。它不再是来自随机号码的纯文本，尤其是 AI 算法的发展。
39:56
email and calling into their products.
而在消息领域，正发生一件大事，产品团队必须了解——
39:58
And there's something big happening in messaging
富通信服务，也就是RCS。
40:00
that product teams need to know about,
你可以把RCS看作是短信2.0。
40:02
rich communication services or RCS.
不再是收到来自随机号码的文本，
40:06
Think of RCS as SMS 2.0.
尤其是AI的算法开发方面。
40:08
Instead of getting text from a random number,
而不是从一个随机数中获取文本。
40:11
your users will see your verified company name and logo
你的用户会直接看到你验证过的公司名称和Logo，
40:13
without needing to download anything new.
完全不需要额外下载任何东西。
40:16
It's a more secure and branded experience.
这是一种更安全、更具品牌感的体验。
40:18
Plus you get features like interactive carousels
而且你还能获得像互动轮播
40:21
and suggested replies.
和智能回复建议这样的功能。
40:22
And here's why this matters.
这就是为什么这件事很重要。
40:24
US carriers are starting to adopt RCS.
美国运营商已经开始采用RCS了。
40:27
Sinch is already helping major brands
Sinch已经在帮助各大品牌实现这一点。
40:29
send RCS messages around the world.
在全球范围内发送RCS消息。
40:31
And they're helping Lenny's podcast listeners
他们正在帮助Lenny的播客听众
40:33
get registered first before the rush hits the US market.
在美国市场热潮到来之前抢先注册。
40:36
Learn more, it gets started at cinch.com slash Lenny.
了解更多，请访问cinch.com/lenny。
40:40
That's s-i-n-c-h.com slash Lenny.
拼写是s-i-n-c-h.com/lenny。
40:45
I reached out to Ben Horowitz, who loves what you're doing,
我联系了Ben Horowitz，他很欣赏你正在做的事情，
40:47
a big fan of yours, their investors, I believe in.
是你的忠实粉丝，也是你们的投资人，我相信这一点。
40:50
Yeah, we've known each other for many years.
是的，我们认识很多年了。
40:54
But yes, right now they are investors of world labs.
是的，目前他们是World Labs的投资方。
40:57
Amazing.
太棒了。
40:58
OK, so I asked him what I should ask you about.
好，所以我问了他，该问你什么问题。
41:00
And he suggested ask you, why is the bitter lessen alone,
他建议我问你，为什么“苦涩的教训”本身，
41:04
not likely to work for robots?
不太可能适用于机器人领域？
41:08
So first of all, just explain what the bitter lessen was
那么首先，请解释一下“苦涩的教训”在AI历史中是什么，
41:11
in the history of AI and then just why that won't get us
然后再说为什么它无法让我们
41:14
to where we want to be with robots.
在机器人领域达到我们想要的目标。
41:16
So first of all, there are many bitter lessons.
首先，有很多惨痛的教训。
41:21
But the bitter lessons everybody refers to is a paper written
但大家常说的那个惨痛的教训，指的是Richard Sutton写的一篇论文，他最近刚得了图灵奖。
41:25
by Richard Sutton, who won the Turing Award recently.
他主要做reinforcement learning。
41:29
And he does a lot of reinforcement learning.
Richard说过，如果你回顾历史，尤其是AI的算法发展历程，
41:31
And Richard has said, right, if you look at the history,
你会发现，最终胜出的总是那些结构简单但数据量巨大的模型，而不是更复杂的模型。
41:35
especially the algorithmic development of AI,
尤其是AI的算法发展，
41:39
it turns out simpler model with a ton of data
结果发现，更简单的模型配上海量数据，最终总是能胜出，而不是更复杂的模型。
41:43
always win at the end of the day, instead of the more
而且我觉得网络视频确实起了作用。
41:50
complex model with less data.
复杂模型用更少的数据。
41:52
I mean, that was actually, this paper came years after image
我的意思是，那篇论文其实是在image之后好几年才出现的。
41:57
that to me was not bitter, it was a sweet lesson.
对我来说那不是苦涩的教训，而是甜蜜的教训。
42:01
That's why I built image that because I believe
这就是我构建image的原因，因为我坚信
42:05
that big data plays that role.
大数据扮演着那个角色。
42:07
So why can bitter lessen working robotics alone?
那么，为什么“苦涩的教训”在机器人领域单独行不通呢？
42:12
Well, first of all, I think we need to give credit
首先，我认为我们需要肯定
42:16
to where we are today.
我们今天的成就。
42:18
Robotics is very much in the early days of experimentation.
机器人技术目前仍处于实验探索的早期阶段。
42:24
It's not, the research is not nearly as mature
它的研究远未成熟，
42:28
as, say, language models.
远不如语言模型那样成熟。
42:30
So many people are still experimenting
所以很多人还在尝试
42:35
with different algorithms.
不同的算法。
42:36
And some of those algorithms are driven by big data.
其中一些算法是由大数据驱动的。
42:40
So I do think big data will continue
因此我确实认为大数据将继续
42:44
to play a role in robotics.
在机器人领域发挥作用。
42:47
And, but what is hard for robotics?
那么，机器人领域的难点在哪里？
42:51
There are a couple of things.
有几个方面。
42:53
One is that it's harder to get data.
第一是数据更难获取。
42:56
It's a lot harder to get data.
数据获取的难度要大得多。
42:58
You can say, well, there is web data.
你可能会说，网络数据是存在的。
43:00
This is where the latest robotics research
这正是最新机器人研究所利用的——网络视频。
43:03
is using web videos.
而且我认为网络视频确实能发挥作用。
43:05
And I think web videos do play a role.
所以，这就实现了完美的对齐——在3D世界中的动作。
43:09
But if you think about what made language model worth a very
但如果你想想是什么让语言模型变得如此有价值——
43:14
as someone who does computer vision
作为一个做计算机视觉、空间智能和机器人的人，
43:16
and spatial intelligence and robotics,
我非常羡慕我那些做语言的同事，
43:19
I'm very jealous of my colleagues in language
因为他们有一个完美的设定：
43:22
because they had this perfect setup
他们的训练数据是单词，后来是token，
43:26
where their training data are in words, eventually tokens.
然后他们产出的模型输出也是单词。
43:31
And then they produce a model that outputs words.
所以，你在输入和输出之间有了完美的对齐。
43:36
So who you have this perfect alignment between what
或者用合成数据来训练机器人。
43:40
you hope to get, which we call objective function,
你希望得到的结果，我们称之为目标函数，以及你的训练数据是什么样的。
43:43
and what your training data looks like.
但机器人领域不同。
43:47
But robotics is different.
甚至空间智能也不同。
43:48
Even spatial intelligence is different.
你希望从机器人那里得到动作输出。
43:51
You hope to get actions out of robots.
但你的训练数据缺乏3D世界中的动作。
43:56
But your training data lacks actions in 3D worlds.
而这正是机器人必须做到的，对吧？
44:02
And that's what robots have to do, right?
在3D世界中的动作。
44:04
Actions in 3D worlds.
基于“苦涩教训”这一假设，
44:06
So you have to find different ways to fit a, what do they call
所以你得想各种办法，把那个所谓的“方钉子”塞进“圆孔”里。
44:14
a square in a round hole.
我们手头有的是海量网络视频。
44:19
What we have is tons of web videos.
那接下来就得讨论如何补充数据了，比如teleoperation数据
44:23
So then we have to start talking about adding
或者synthetic data，这样机器人才能基于“苦涩教训”的假设来训练——
44:27
supplementing data, such as teleoperation data
也就是靠海量数据。
44:32
or synthetic data so that the robots are trained
也就是大量数据。
44:37
with this hypothesis of bitter lesson,
基于这个“苦涩教训”的假设，
44:39
which is large amount of data.
也就是大量数据。
44:41
I think there is still hope because even what we are doing
我觉得还是有希望的，因为即使我们现在做的世界建模，也真的能为机器人解锁大量信息。
44:47
in world modeling will really unlock a lot of this information
但我觉得我们得小心，因为现在还处于早期阶段。
44:52
for robots.
而且苦涩的教训还有待验证，因为我们还没完全搞清楚对应的数据。
44:53
But I think we have to be careful because we're
这是机器人领域苦涩教训的另一部分。
44:55
at the early days of this.
在早期阶段是这样。
44:57
And bitter lesson is still to be tested
而苦涩教训仍有待验证，
45:01
because we haven't fully figured out the data for it.
因为我们还没完全搞清楚它的数据。
45:06
Another part of the bitter lesson of robotics,
机器人领域的苦涩教训还有另一层，
45:09
I think we should be so realistic about is, again,
我觉得我们得现实一点，就是，跟语言模型甚至空间模型比起来，机器人是物理系统。所以机器人更接近自动驾驶汽车，而不是大语言模型。认识到这一点非常重要。这意味着要让机器人真正运作起来，我们不仅需要大脑，还需要物理身体。
45:15
compared to language models or even spatial models,
相比语言模型甚至空间模型，
45:19
robots are physical systems.
机器人是物理系统。
45:22
So robots are closer to self-driving cars
所以机器人更接近自动驾驶汽车，
45:26
than a large language model.
而不是大型语言模型。
45:28
And that's very important to recognize.
这一点非常值得注意。
45:30
That means that in order for robots to work,
这意味着，要让机器人真正运作起来，
45:35
we not only need brains, we also need the physical body,
我们不仅需要大脑，还需要物理身体，
45:40
we also need applications and scenarios.
我们还需要应用和场景。
45:44
If you look at the history of self-driving car,
回顾自动驾驶汽车的历史，
45:49
my colleague Sebastian Thrum took Stanford's car
我的同事Sebastian Thrum开着斯坦福的车，
45:55
to win the first DARPA challenge in 2006 or 2005.
在2005或2006年赢得了第一届DARPA挑战赛。
46:00
It's 20 years since that prototype of a self-driving car
从那个能在内华达沙漠行驶130英里的自动驾驶原型车，
46:07
being able to drive 130 miles in the Nevada desert
到现在旧金山街头的Waymo，已经过去了20年。
46:11
to today's Waymo and on the street of San Francisco.
而且我们还没结束，还有很多路要走。
46:17
And we're not even done yet, there's still a lot.
而且这还远远没完，还有大量工作要做。
46:20
So that's a 20-year journey.
所以这是一个长达20年的旅程。
46:22
And self-driving cars are much simpler robots.
自动驾驶汽车是更简单的机器人。
46:25
They're just metal boxes running on 2D surfaces.
它们只是运行在二维平面上的金属盒子。
46:29
And the goal is not to touch anything.
目标是不碰到任何东西。
46:32
Robot is 3D things running in 3D world.
机器人是三维的东西，在三维世界里运行。
46:37
And the goal is to touch things.
目标是去触碰东西。
46:39
So the journey is going to be, there's many aspects, elements.
所以这个旅程会有很多方面、很多元素。
46:45
And of course, one could say, well,
当然，有人可能会说，那好吧——
46:48
the self-driving car early algorithm
早期的自动驾驶汽车算法
46:51
were pre-deplanning era.
还处于预规划时代。
46:53
So deplanning is accelerating the brains.
而预规划就是给大脑加速。
46:56
And I think that's true.
我认为确实如此。
46:57
That's why I'm in robotics.
这也是我投身机器人领域的原因。
46:59
That's why I'm in spatial intelligence
这也是我专注于空间智能的原因，
47:01
and I'm excited by it.
并且对此充满热情。
47:03
But in the meantime, the car industry is very mature.
但与此同时，汽车行业已经非常成熟了。
47:07
And productizing also involves the mature use cases, supply chains, the hardware.
产品化还涉及成熟的应用场景、供应链和硬件。
47:16
So I think it's a very interesting time to work in these problems.
所以我认为现在正是解决这些问题的有趣时期。
47:20
But it's true, Ben is right.
但确实，Ben说得对。
47:22
We might still be subject to a number of better lessons.
我们可能还会学到更多宝贵的经验。
47:28
During this work, do you ever just feel off
在做这些工作时，你是否有时会感到不对劲——
47:31
for the way that brain works and is able to do all of this for us?
对于大脑能如此运作、为我们完成这一切的方式？
47:35
Just the complexity, just to get a machine to just walk around and not hit things and fall.
光是让一台机器能四处走动、不撞到东西、不摔倒，就已经如此复杂。
47:41
Does it just give you more respect for what we've already got?
这会不会让你对我们已有的东西更加敬畏？
47:44
Totally.
完全同意。
47:45
We operate on about 20 watts.
我们大脑的功耗大约只有20瓦。
47:49
That's dimmer than any light bulb in the room I'm in right now.
这比我此刻所在房间里的任何灯泡都要暗。
47:53
And yet, we can do so much.
然而，我们却能完成这么多事情。
47:57
So I think actually the more I work in AI, the more I respect humanism.
所以我觉得，实际上我在AI领域工作得越久，就越尊重人文主义。
48:03
Let's talk about this product you just launched called Marble, a very cute name.
来聊聊你们刚发布的产品Marble吧，名字很可爱。
48:08
Talk about what this is, why this important I've been playing with it.
说说这是什么，为什么它很重要——我一直在试用，真的太棒了。
48:10
It's incredible.
这简直不可思议。
48:10
We'll link to it and for folks to check it out.
我们会附上链接，方便大家去看看。
48:13
What is Marble?
Marble是什么？
48:14
Yeah, I'm very excited.
对，我非常期待。
48:16
So first of all, Marble is one of the first product that Warlabs has rolled out.
首先，Marble是Warlabs推出的首批产品之一。
48:22
Warlabs is a foundation frontier model company.
Warlabs是一家前沿基础模型公司。
48:25
We are funded by four co-founders who have deep technical history.
我们由四位技术背景深厚的联合创始人共同创立。
48:31
My co-founders Justin Johnson, Christoph Lassner, and Ben Mildenhall.
我的联合创始人Justin Johnson、Christoph Lassner和Ben Mildenhall。
48:37
We all come from the research field of AI,
我们都来自AI研究领域。
48:41
computer graphics, computer vision.
计算机图形学、计算机视觉。
48:43
And we believe that spatial intelligence and world modeling is as important,
我们相信空间智能与世界建模的重要性不亚于语言模型，甚至更为关键，且与语言模型相辅相成。
48:49
if not more, to language models and complementary to language models.
因此我们想抓住这个机会，打造一个深度科技研究实验室，将前沿模型与产品之间的脉络串联起来。
48:55
So we wanted to seize this opportunity to create deep tech.
Marble 就是基于我们前沿模型构建的一款应用。
49:01
Research lab that can connect the dots between frontier models with products.
我们花了一年多时间，打造了全球首个能够输出真正3D世界的生成式模型。
49:08
So Marble is an app that's built upon our frontier models.
Marble 是一个基于我们前沿模型构建的应用。
49:15
We've spent a year and plus building the world's first
我们花了一年多时间，打造了全球首个
49:20
generative model that can output genuinely 3D worlds.
能够生成真正3D世界的生成式模型。
49:26
That's a very, very hard problem.
这是一个非常、非常难的问题。
49:29
And it was a very hard process.
整个过程也非常艰难。
49:34
We have a team of incredible,
我们有一支非常出色的团队，
49:37
founding team of incredible technologists from incredible teams.
创始团队由来自顶尖团队的杰出技术专家组成。
49:45
And then around a month or two ago,
然后大约一两个月前，
49:50
we saw the first time that we can just prompt with a sentence and an image,
我们第一次看到，只需用一句话和一张图片，
49:57
and multiple images and create worlds that we can just navigate in.
或者多张图片，就能生成我们可以自由探索的世界。
50:03
If you put it on cargo, which we have an option to let you do that,
如果你把它放到cargo上——我们提供了这个选项——
50:07
you can't even walk around it, right?
你甚至没法绕着它走，对吧？
50:09
So it was, even though we've been building this for quite a while,
所以，尽管我们开发这个已经有一阵子了，
50:13
it was still just awe inspiring.
它依然令人叹为观止。
50:16
And we wanted to get into the hands of people who needed.
我们想把它交到需要的人手中。
50:20
And then we know that so many creators, designers,
而且我们知道，那么多创作者、设计师、
50:25
people who are thinking about robotic simulation,
正在思考机器人模拟的人、
50:29
people who are thinking about different use cases of navigable,
正在思考可导航、可交互、沉浸式世界的不同应用场景的人，
50:34
interactable, immersive worlds, game developers will find this useful.
还有游戏开发者，都会觉得它很有用。
50:41
So we developed Marble as a first step.
所以我们开发了Marble作为第一步。
50:46
It's again, still very early,
这依然非常早期，
50:49
but it's the world's first model doing this
但它是世界上第一个做到这一点的模型，
50:52
and it's the world's first product that allows people to just prompt,
也是世界上第一个让人们只需输入提示就能使用的产品——
50:58
we call it prompt two worlds.
我们称之为“prompt two worlds”。
51:00
Well, I've been playing around it.
嗯，我一直在玩它。
51:02
It is insane.
简直太疯狂了。
51:02
Like you could just have a little shire world,
比如你可以直接创建一个小小的shire世界。
51:04
where you just infinitely walk around Middle Earth basically
基本上就是无限在中土世界里走来走去，
51:07
and there's no one there yet, but it's insane.
里面还没有任何人，但真的超疯狂。
51:10
You just go anywhere, there's like dystopian world.
你随便走到哪儿，都像是个反乌托邦的世界。
51:12
I'm just looking at all these examples.
我就在看这些例子。
51:14
Yes.
对。
51:14
And my favorite part, actually, I don't know,
其实我最喜欢的部分，我不知道，
51:16
I don't know if there's a feature bug.
不知道这算不算功能bug。
51:17
You can see like the dots of the world before it actually renders
你可以在世界真正渲染出来之前，就看到那些点状的轮廓。
51:21
with all the textures and I just love to,
带着所有的质感，我真的很喜欢，
51:23
like you get a glimpse into what is going on with this model.
就像你能一窥这个模型内部到底在发生什么。
51:26
Basically, I like it.
基本上，我喜欢它。
51:27
That is so cool to hear.
听到这个真是太酷了。
51:28
Because this is where as a researcher, I'm learning
因为作为研究者，这正是我在学习的地方，
51:33
because the dots that lead you into the world
因为那些引导你进入这个世界的点，
51:38
was an intentional feature visualization.
是刻意设计的特征可视化。
51:44
It is not part of the model.
它并不是模型本身的一部分。
51:46
It's the model actually just generates the world.
其实是模型在生成这个世界。
51:50
But we were trying to find a way to guide people
但我们当时想找到一种方式，引导人们进入这个世界，不少工程师尝试了不同版本，最终我们统一采用了那个点。
51:53
into the world and a number of engineers
那么多人里，只有你告诉我们，那种体验有多令人愉悦。
51:56
worked on different versions, but we
这对我们来说真的很有成就感。
51:58
converged on the dot.
收敛到了那个点上。
51:59
And so many people, you're the only one
那么多人里，只有你
52:03
told us how delightful that experience is.
告诉我们那种体验有多愉快。
52:06
And it was really satisfying for us
这对我们来说真的很有满足感。
52:09
to hear that this intentional visualization feature,
听到这个有意的可视化功能，
52:13
that's not just a big hardcore model, actually
并不只是一个硬核的大模型，实际上
52:16
has delighted our users.
让我们的用户很开心。
52:19
Wow, so you add that to make it more like to have humans
哇，所以你加上这个是为了让它更像人类，
52:23
understand more and more to like, wow, that is hilarious.
让人类越来越能理解，然后觉得，哇，这太有趣了。
52:27
It makes me think about a lens in the way
这让我想到一种视角，
52:29
they, it's not the same thing, but they
虽然不完全一样，但他们会
52:30
talk about what they're thinking and what they're doing.
谈论自己在想什么、在做什么。
52:32
Yes, it is.
是的，确实如此。
52:33
It is.
没错。
52:34
It also makes me think about just the matrix.
这也让我想到《黑客帝国》本身。
52:36
Like it's exactly the matrix experience.
就像完全是在《黑客帝国》里的体验。
52:39
I don't know if that was your inspiration.
我不知道这是不是你的灵感来源。
52:41
Well, like I said, a number of engineers worked on that.
嗯，就像我说的，有不少工程师参与过那个项目。
52:44
It could be there in inspiration.
灵感可能就藏在那里。
52:46
It's in their subconscious.
它存在于他们的潜意识中。
52:50
OK, so just for folks that may want to play around with us,
好的，对于想跟我们一块儿玩的朋友，
52:53
maybe use a what's like, what are some applications today
现在有哪些应用是大家今天就能上手的？
52:55
that folks can start using today?
这次发布的目标是什么？
52:57
What's your goal with this launch?
嗯，我们确实相信世界建模（world modeling）是非常横向的，
52:59
Yeah, so we do believe that world modeling is very horizontal,
但我们已经看到一些特别令人兴奋的用例，
53:04
but we're already seeing some really exciting use cases,
比如电影虚拟制作，因为他们需要的是
53:09
virtual production for movies, because what they need
能与摄像机对齐的3D世界。
53:12
are 3D worlds that they can align with the camera.
是那些能与摄像头对齐的3D世界。
53:18
So when the actors are acting on it, they can, you know,
所以当演员在上面表演的时候，他们可以，你知道的，
53:22
they can position the camera and shoot the segments
很好地摆放摄像机并拍摄各个片段。
53:26
really well.
而且我们已经看到了非常棒的应用。
53:27
And we're already seeing incredible use.
事实上，不知道你有没有看过我们的发布视频，
53:31
In fact, I don't know if you have seen our launch video
里面展示了 marble 的效果。
53:35
showing marble.
那是由一家虚拟制作公司制作的。
53:36
It was produced by a virtual production company.
我们和索尼合作了。
53:40
We collaborated with Sony.
我们与Sony合作了。
53:43
And they use marble scenes to shoot those videos.
他们用大理石场景来拍摄那些视频。
53:46
So we were collaborating with those technical artists
所以我们当时和那些技术美术师
53:50
and directors.
还有导演在合作。
53:51
And they were saying this has cut our production time by 40x.
他们就说这已经把制作时间缩短了40倍。
53:57
In fact, it has to be.
实际上，必须得这样。
53:58
40x.
40倍。
53:59
Yes, in fact, it has to, because we only had one month
对，确实必须这样，因为我们只有一个月
54:03
to work on this project.
来做这个项目。
54:04
And there were so many things they were trying to shoot.
他们当时想拍的东西太多了。
54:09
So using marble really, really significantly
所以用Marble真的能非常显著地
54:13
accelerated the production of virtual production
加速虚拟制作流程
54:17
for VFX and movies.
用于视觉特效和电影。
54:19
That's one use cases.
这是一个应用场景。
54:20
We are already seeing our users taking our marble scene
我们已经看到用户把我们的Marble场景
54:25
and taking the mesh export and putting games,
导出网格模型放进游戏里，
54:28
you know, whether it's games on VR or games,
不管是VR游戏还是普通游戏。
54:32
just fun games that they have developed.
只是他们开发的一些有趣游戏。
54:36
We have had, we were showing an example
我们之前展示过一个机器人模拟的例子，因为当我——
54:41
of robotic simulation because when I was,
我的意思是，我现在仍然是一名从事机器人训练的研究员——
54:45
I mean, I'm still a researcher doing robotic training,
最大的痛点之一就是为训练机器人创建合成数据。
54:51
one of the biggest pain point is to create synthetic data
而这些合成数据需要非常多样化，
54:55
for training robots.
它们必须来自不同的环境。
54:56
And these synthetic data needs to be very diverse.
而这些合成数据需要非常多样化。
54:59
They need to come from different environments
它们需要来自不同的环境。
55:01
with different objects to manipulate.
通过不同的物体进行操作。
55:04
And one path to it is to ask computers to simulate.
其中一条路径是让计算机进行模拟。
55:10
Otherwise, humans have to build every single asset
否则，人类必须为机器人构建每一个单独的资产，
55:15
for robots that's just going to take a lot longer.
这只会花费更长的时间。
55:19
So we already have researchers reaching out
所以已经有研究人员联系我们，
55:22
and wanting to use marble to create those synthetic environments.
希望使用Marble来创建这些合成环境。
55:26
We also have unexpected user outreach
我们还收到了意想不到的用户反馈，
55:32
in terms of how they want to use marble.
关于他们希望如何使用Marble。
55:35
For example, a psychologist team called us
举个例子，有个心理学团队联系我们，想用marble来做心理学研究。结果发现，他们研究的一些精神科患者需要了解自己的大脑如何对不同特征的沉浸式场景做出反应，比如杂乱的场景、整洁的场景，或者任何你能想到的场景。
55:40
to use marble to do psychology research.
用大理石来做心理学研究。结果发现，他们研究的一些精神科病人。需要理解他们的大脑如何对不同特征的沉浸式场景做出反应。比如，杂乱的场景或整洁的场景，或者随便你怎么说。而大理石几乎是一种即时的方式。
55:43
It turned out some of the psychiatric patients they study.
结果发现，他们研究的一些精神科患者，
55:48
They need to understand how their brain
需要理解自己的大脑
55:51
respond to different immersive scenes
如何对不同特征的沉浸式场景
55:54
of different features.
做出反应。
55:56
For example, messy scenes or clean scenes
比如，杂乱的场景或整洁的场景，
55:59
or whatever you name it.
或者任何你能想到的场景。
56:01
And it's very hard for researchers to get their hands
研究人员很难获取这类沉浸式场景，
56:04
on these kind of immersive scenes.
创建它们又耗时耗资巨大。
56:07
And it will take them too long and too much budget
而Marble几乎能瞬间将大量实验环境交到他们手中。
56:11
to create.
所以目前我们已经看到多种应用场景。
56:14
And marble is a really almost instantaneous way
大理石几乎是一种即时的方式
56:18
of getting so many of these experimental environments
让这么多实验环境落到他们手里，
56:24
into their hands.
所以我们现在看到多个使用场景。
56:25
So we're seeing multiple use cases at this point.
他当时就在刷TikTok，看大家怎么用它，
56:30
But the VFX, the game developers,
但那些VFX、游戏开发者、
56:33
the simulation developers, as well as designers
模拟开发者，还有设计师们，
56:37
are very excited.
都特别兴奋。
56:39
This is very much the way things work in AI
这跟AI领域的运作方式很像——
56:41
have had other AI leaders on the podcast.
我们播客也请过其他AI领域的领军人物。
56:43
And it's always like put things out there early
他们总是说，要尽早把东西推出去，
56:45
as soon as you can to discover where the big use cases are.
越快越好，这样才能发现真正的重大应用场景在哪儿。
56:48
The head of JetGPT told me how,
JetGPT的负责人告诉我，
56:50
when they first put out JetGPT,
他们刚推出JetGPT的时候，他就在刷TikTok看大家怎么用这个产品，还有大家都在聊些什么。正是这些观察让他们确定了该往哪个方向发力，也帮他们看清了用户真正想怎么用它。我特别喜欢最后那个用于心理治疗的案例。我就在想象，那些有恐高症的人，或者怕蛇、怕蜘蛛的人，看到这个功能时的反应。
56:52
he was just scanning TikTok to see how people were using it
以及大家都在聊些什么。
56:55
and all the things they were talking about.
正是这些让他们确定了该往哪个方向发力，
56:56
And that's what convinced them where to lean in
也帮他们看清了人们实际想怎么用这个产品。
56:59
and help them see how people actually want to use it.
我特别喜欢最后一个使用场景，比如用于心理治疗。
57:02
I love this last use case of like for therapy.
我非常喜欢这个用于治疗的最后一个用例
57:04
I'm just imagining like heights to people
我就在想，就像有人恐高、怕蛇、怕蜘蛛那样
57:07
seeing, dealing with heights or snakes or spiders
但我觉得解释一下可能会有帮助
57:11
at which.
在哪个点上。
57:11
It's amazing.
太神奇了。
57:12
A friend of mine last night literally called me
昨晚我一个朋友直接打电话给我，
57:15
and talked about his height scare
聊了他对身高的焦虑，
57:17
and asked me if marble should be used.
还问我该不该用大理石。
57:20
It's amazing.
太神奇了。
57:21
You went straight there.
你直接就想到那儿去了。
57:24
Because I'm imagining all the exposure therapy stuff.
因为我脑子里全是那些暴露疗法的东西。
57:28
This could be so good for that.
这对那个来说可能非常有用。
57:29
That is so cool.
这太酷了。
57:31
OK, so let me, I should have asked you this before,
好，那我——其实我早该问你这个问题，
57:32
but I think there's going to be a question of just,
但我觉得大家肯定会好奇，
57:35
how does this differ from things like VO3
这跟VO3之类的视频生成模型到底有什么不同？
57:38
and other video generation models?
我自己很清楚区别，
57:40
It's pretty clear to me,
但解释一下可能会更有帮助。
57:41
but I think it might be helpful just to explain
其实，这个世界并不是被动地看视频
57:43
how this is different from all the video AI tools people
这和人们见过的所有视频AI工具有什么不同。
57:45
have seen.
Weren apps的核心观点是，空间智能从根本上来说非常重要。
57:46
Weren't apps thesis is that spatial intelligence
而空间智能不仅仅是关于视频。
57:49
is fundamentally very important.
事实上，世界并不是被动地看着视频从眼前流过，对吧？
57:51
And spatial intelligence is not just about videos.
我很喜欢柏拉图的那个洞穴寓言。
57:58
In fact, the world is not passively watching videos
一晃而过，对吧？
58:02
passing by, right?
我特别喜欢柏拉图那个洞穴寓言
58:04
I love Plato has the allegory of the cave analogy
而不是只创造一个平面的2D世界
58:11
to describe vision.
描述视觉。
58:12
He said that imagine a prisoner tidal his chair,
他说，想象一个囚犯被绑在椅子上，
58:18
not very humane, but in a cave,
不太人道，但在一个洞穴里，
58:23
a watching a full life theater in front of him.
正看着面前的全息生活剧场。
58:30
But the actual life theater that actors are acting
但演员们表演的真实生活剧场
58:34
is behind his back.
其实在他背后。
58:36
It was just lit so that the projection of the action
只是灯光打过来，让动作的投影
58:41
is on a wall of the cave.
落在洞穴的墙壁上。
58:44
And then the task of this prisoner
然后这个囚徒的任务
58:48
is to figure out what's going on.
就是搞清楚到底发生了什么。
58:50
It's a pretty extreme example,
这是个挺极端的例子，
58:52
but it really shows it describes what vision is about,
但它确实说明了视觉的本质——
58:59
is that to make sense of the 3D world
就是从二维信息中理解三维世界
59:03
or 4D world out of 2D.
甚至四维世界。
59:05
So spatial intelligence to me is deeper
所以对我来说，空间智能比仅仅创造平面的二维世界要深刻得多。
59:09
than only creating that flat 2D world.
去产生推理、互动、理解
59:14
Spatial intelligence to me is the ability
对我而言，空间智能就是能够去创造、推理、交互，并理解这个深度空间的世界——无论是2D、3D还是4D，包括动态变化等等。所以World Lab正在专注这个方向。当然，生成视频本身的能力也可以算是其中的一部分。实际上，就在几周前，
59:18
to create reason, interact, make sense
来创造推理、互动、理解
59:23
of deeply spatial world, whether it's 2D or 3D or 4D,
对于深度空间化的世界，无论是2D、3D还是4D，
59:30
including dynamics and all that.
包括动态效果等一切内容，
59:32
So the world lab is focusing on that.
世界实验室正专注于这一方向。
59:35
And of course, the ability to create videos per se
当然，生成视频本身的能力
59:40
could be part of this.
也可能是其中的一部分。
59:41
And in fact, just a couple of weeks ago,
事实上，就在几周前，
59:44
we rolled out the world's first real time demoable,
我们推出了全球首个可实时演示的、在单张H100 GPU上运行的实时视频生成技术。这部分技术已经包含在我们的方案中。但我认为Marble非常不同，因为我们真正希望创作者、设计师和开发者能够掌握一个能生成具有3D结构世界的模型，从而用于他们的实际工作。
59:49
real time video generation on a single H100 GPU.
我们实现了在单张H100 GPU上实时生成视频。
59:54
So part of our technology includes that.
因此，我们的技术也涵盖了这一点。
59:58
But I think Marble is very different
但我觉得Marble非常不同，因为我们真心希望创作者、设计师和开发者能掌握一个能生成具有3D结构的世界的模型，这样他们就能用于实际工作。仅此而已，然后继续推进。顺便说一句，在Marble里，我们当时不允许用户以视频形式导出内容。
60:00
because we really want creators, designers,
因为我们真的希望创作者、设计师、
60:05
developers to have in their hands
开发者手中能拥有
60:09
a model that can give them worlds with 3D structure.
一个可以为他们生成具有3D结构世界的模型
60:15
So they can use it for their work.
这样他们就能将其用于工作
60:17
And that's why Marble is so different.
这就是为什么Marble如此不同。
60:21
The way I see it is it's a platform
在我看来，它是一个充满各种机会的平台，能让你做很多事情。
60:23
for a ton of opportunity to do stuff.
就像你描述的那些视频，这是一个一次性的视频，非常有趣、很酷。
60:27
As you describe videos, here's a one-off video
然后就这样了，你就继续往前走了。
60:29
that's very fun and cool.
顺便说一句，在Marble里，我们不允许用户以视频形式导出内容。
60:31
And that's it, and you move on.
就这样，然后继续前进
60:33
By the way, in Marble, we couldn't allow people
顺便提一下，在Marble里，我们不允许用户以视频形式导出内容，或者什么无限算力之类的。但归根结底，AI是关于人的。当有人问我这个问题时，我的回答是响亮的“是的”——每个人在AI中都有自己的角色。这取决于你做什么、想要什么，没有任何技术应该剥夺人的尊严。
60:36
to export in video form.
以视频形式导出。
60:38
So you could actually, like you said,
所以实际上，就像你说的，
60:40
you go into a world, so let's say it's a hobbit cave.
你进入一个世界，比如说一个霍比特人洞穴。
60:45
You can actually, especially as a creator,
作为创作者，你其实可以用导演脑海中那种非常独特的轨迹来移动镜头，对吧？
60:47
you have such a specific way of moving the camera
然后你可以从Marble里把它导出成视频。
60:54
in a trajectory in the director's mind, right?
要做出这样的东西需要什么条件？
60:57
And then you can export that from Marble into a video.
比如团队规模有多大？
61:02
What does it take to create something like this?
要做出这样的东西需要什么条件？
61:03
Just like how big is the team?
团队大概有多大？
61:05
How many GPUs you work in like anything you can share there?
你们用了多少块GPU？这方面有什么可以分享的吗？
61:08
I don't know how much of this is private information,
我不太清楚这算不算机密信息，
61:09
but just what does it take to create something
但就你们这次发布的产品来说，
61:11
like this that you launched here?
打造这样一个东西需要什么条件？
61:12
It takes a lot of brain power.
需要大量的脑力。
61:16
So we just talk about 20 watts per brain.
所以我们按每个大脑20瓦来算的话，
61:20
It's, so from that point of view,
从这个角度看，
61:22
it's a small number, but it's actually an incredible,
数字其实不大，但实际效果却非常惊人。
61:26
you know, it's a half billion years of evolution
你知道，那是五亿年的进化
61:29
to give us those power.
才赋予了我们这些能力。
61:32
We have a team of 30-ish people now
我们现在有一个大约30人的团队
61:36
and we are predominantly researchers or research engineers.
主要是研究人员或研究工程师。
61:43
And, but we also have designers and product.
不过，我们也有设计师和产品人员。
61:48
We actually really believe that we wanna create a company
我们确实相信，我们想创建一家
61:52
that's anchored in the deep tech of spatial intelligence,
以空间智能的深度技术为核心的公司，
61:56
but we are actually building serious products.
但同时我们也在打造真正的产品。
62:02
So we have this integration of R&D and productization.
所以我们把研发和产品化整合在了一起。
62:11
And of course, we use a ton of GPUs.
当然，我们用了大量的GPU。
62:15
That's a, that's a technical happy to hear.
这听起来技术上挺让人开心的。
62:20
Well, congrats on the launch.
那恭喜你们上线了。
62:21
I know there's a huge milestone.
我知道这是个巨大的里程碑。
62:22
I know this took a ton of work.
也知道这背后付出了很多努力。
62:23
So I just wanna say congrats to you and your team.
所以我想对你和你的团队说声恭喜。
62:26
Let me talk about your founder journey for a moment.
接下来我想聊聊你的创始人历程。
62:29
So your founder of this company started,
所以你们公司的创始人，是多久以前开始的？几年前？两三年？
62:31
how many years ago, a couple of years ago, two, three years ago?
一年前。
62:33
A year ago.
一年前。
62:34
A year ago.
一年，好吧。
62:35
A year, okay.
18个月，对。
62:37
18 months, yeah.
好的。
62:38
Okay.
在开始之前，你希望自己早知道些什么？
62:39
What's something you wish you knew before you started this,
有什么事情是你希望自己开始之前就知道的？
62:42
that you wish you could like whisper into the ear
你希望自己能像悄悄话一样，告诉18个月前的自己什么？
62:43
of faith of 18 months ago?
嗯，我一直希望能预知技术的未来。
62:46
Well, I continue to wish I know the future of technology.
其实我觉得，这恰恰是我们创始团队的一个优势——
62:52
I think actually that's one of our founding advantage
我们通常比大多数人更早看到未来，但即便如此，天啊，眼前的一切依然如此激动人心、如此不可思议，未知和即将到来的东西太令人震撼了。
62:55
is that we see the future earlier in general
不过我知道你问我这个问题是有原因的。
63:00
than most people, but still, man, this is so exciting
虽然比大多数人知道得多，但天哪，这依然让人无比兴奋
63:03
and so amazing that what's unknown and what's coming.
也无比神奇，因为未知和即将到来的东西太多了。
63:09
But I know the reason you're asking me this question
但我知道你问我这个问题是为什么——
63:12
is a lot about the future of technology.
这很大程度上关乎技术的未来。
63:14
You're probably more, you know, look, I did not start
你可能更——你看，我20岁的时候可没创办过这种规模的公司。
63:19
a company of this scale at 20 year old.
我19岁倒是开过一家干洗店，但规模小多了。
63:24
So, you know, I started a dry cleaner when I was 19,
咱们得从那儿说起。
63:27
but that's a little smaller scale.
然后呢，我投资了Google Cloud AI，
63:30
We gotta start with that.
后来又投资了一个InstituS网络。
63:32
And then I, you know, funded Google Cloud AI
然后呢，我投资了Google Cloud AI，
63:35
and then I funded an InstituS network,
后来又投资了InstituS网络。
63:37
but those are different beasts.
但那些是截然不同的猛兽。
63:40
I did feel I was a little more prepared as a founder
作为创始人，我确实觉得自己比可能二十岁的创业者
63:46
of the grinding journey that I compared to maybe,
对这段磨砺之旅准备得更充分一些，
63:51
maybe the 20 year old funders,
但我仍然感到惊讶，有时甚至陷入偏执——
63:58
but I still, I'm surprised and it puts me into paranoia sometimes
AI领域的竞争竟如此激烈，
64:06
that how intensely competitive AI landscape is
无论是模型、技术本身，还是人才。
64:12
from the model, the technology itself, as well as talents.
而且，你知道，当我创立这家公司时——
64:18
And, you know, when I founded the company,
而且，你知道，当我创立这家公司的时候，
64:23
we did not have these incredible stories
我们以前可没听过那些离谱的故事，
64:26
of how much certain talents would cost, you know.
比如某个天才人才要花多少钱，你懂的。
64:32
So, these are things that continue to surprise me
所以这些事还是让我挺惊讶的，
64:35
and I have to be very alert about.
我得时刻保持警惕。
64:40
The competition you're talking about is yeah,
你提到的竞争，没错，
64:41
the competition for talent, the speed
就是人才竞争，还有事情发展的速度，
64:44
at which just how things are moving.
变化实在太快了。
64:45
Yeah.
是啊。
64:46
Yeah.
嗯。
64:48
You mentioned this point that I want to come back to
你提到这一点，我想回头再聊聊——
64:50
that you, if you just look over the course of your career,
就是回顾你的整个职业生涯，
64:53
you were like at all of the major collections of humans
你几乎参与了所有那些重要的人类团队，
64:57
that led to so many of the breakthroughs
正是这些团队促成了如今许多突破性进展。
65:00
that are happening today.
显然我们也会谈到ImageNet，还有斯坦福的销售团队，
65:01
Obviously we talk about ImageNet also just sale at Stanford
那里也是很多成果诞生的地方。
65:04
is where a lot of the work happened.
大部分工作都是在那里完成的。
65:06
Google Cloud, which a lot of the breakthroughs happened,
Google Cloud，很多突破性进展都发生在这里，
65:09
what brought you to those places?
是什么吸引你去了那些地方？
65:11
Like for people looking for how to advance in their career,
对于那些想推动职业发展、站在未来中心的人来说，
65:16
be at the center of the future,
这其中是不是有一条主线——
65:17
just like is there a through line there
是什么让你从一个地方到另一个地方，
65:19
of just what pulled you from place to place
又是什么让你加入那些团队，
65:21
and pulled you into those groups
这些经历或许对大家会有所启发？
65:23
that might be helpful for people to hear?
这也许对大家有帮助？
65:25
Yeah, this is actually a good question, Lenny,
嗯，这确实是个好问题，Lenny，
65:28
because I do think about it
因为我确实想过这个，
65:29
and obviously we talked about its curiosity
而且我们之前也聊过，是好奇心和热情让我走进了AI。
65:35
and passion that brought me to AI.
那更多是出于科学探索的起点，对吧？
65:37
That is more a scientific nor start, right?
我当时并不在意AI是不是一个热门领域。
65:40
I did not carry if AI was a thing or not.
所以那是其中一部分，
65:43
So that was one part,
但我最终是怎么选择走上这条路的呢？
65:45
but how did I end up choosing
但我最终是怎么选择的，
65:49
in the particular places I work
在我工作的特定领域，包括创办World Labs？我觉得我非常感谢自己，或者也许感谢我父母的基因。我是一个在智力上非常无畏的人，而且我得说，当我招聘年轻人时，我会寻找这种特质，因为我认为这是非常重要的品质。
65:52
and including starting world labs?
以及为什么包括创办World Labs？
65:55
Is I think I'm very grateful to myself
我觉得我非常感谢自己，
66:01
or maybe to my parents' genes.
或者也许要感谢我父母的基因。
66:04
And I'm an intellectually very fearless person
而且我是一个在智力上非常无所畏惧的人。
66:08
and I have to say when I hire young people,
而且我得说，当我招年轻人时，
66:11
I look for that
我会看重这一点，
66:13
because I think that's a very important quality
因为我认为这在学术界
66:19
if one wants to make a difference,
如果你想有所作为，
66:21
is that when you wanna make a difference,
关键在于当你想要做出改变时，
66:25
you have to accept that you're creating something new
你必须接受自己正在创造新事物，
66:29
or you're diving into something new,
或者投身于全新领域，
66:31
people haven't done that.
而这是前人未曾尝试过的。
66:32
And if you have that self-awareness,
如果你有这份自觉，
66:36
you almost have to allow yourself to be fearless
你就几乎必须让自己无所畏惧，
66:41
and to be courageous.
并且勇敢前行。
66:42
So when I, for example, came to Stanford,
比如我当初来斯坦福的时候，在学术界里，我其实离所谓的终身教职（tenure）非常近——就是在普林斯顿拿到一份永久教职。但我还是选择来了斯坦福，因为我热爱普林斯顿，那是我的母校。只是当时，斯坦福有太多非常出色的人了。
66:49
in the world of academia,
是非常重要的一种品质。
66:52
I was very close to this thing called tenure
我当年差点就拿到tenure了，
66:56
which is have the job forever at Princeton.
就是在普林斯顿拿到终身教职。
67:00
But I choose to come to Stanford
但我选择来到斯坦福，
67:05
because I love Princeton, it's my alma mater.
因为我爱普林斯顿，那是我的母校。
67:09
It's just at that moment,
就在那一刻，
67:11
there are people who are so amazing at Stanford
斯坦福有那么多厉害的人，
67:15
and the Silicon Valley ecosystem was so amazing
硅谷的生态真的太棒了，
67:18
that I was okay to take a risk
所以我愿意冒这个险，
67:22
of restarting my tenure clock.
重新启动我的终身教职计时。
67:26
Going to becoming the first female director of sale,
成为第一位女性销售总监，
67:33
I was actually relatively speaking
说实话，我当时算是
67:35
of very young faculty at that time.
比较年轻的教职人员。
67:38
And I wanted to do that because I care
我之所以想这么做，是因为我
67:40
about that community.
在乎那个群体。
67:41
I didn't spend too much time thinking
我没花太多时间琢磨
67:43
about all the failure cases.
那些失败的情况。
67:46
Obviously I was very lucky
显然我非常幸运，
67:48
that the more senior faculty supported me,
资深的教授们支持了我，
67:50
but I just wanted to make a difference.
但我只是想做出一些改变。
67:54
And then going to Google was similar.
后来去Google也是类似的情况。
67:56
I wanted to work with people like Jeff Dean, Jeff Hinton
我想和Jeff Dean、Jeff Hinton这样的人共事，
68:02
and all these incredible demos, the incredible people.
还有那些令人惊叹的demo，那些了不起的人。
68:11
So the same with world labs, I have this passion
所以和World Labs一样，我怀有这份热情，也相信志同道合的人能做出不可思议的事。这就是我的人生信条。我不会过度纠结所有可能出错的环节，因为那实在太多了。我觉得这点很重要——不是只盯着负面结果。
68:16
and I also believe that people with the same mission
而我也相信，拥有同样使命的人
68:21
can do incredible things.
能做出不可思议的事。
68:22
So that's how I guided my through life.
这就是我一路走来的指引。
68:26
I don't over think of all possible things
我不会过度纠结所有可能出错的事，
68:30
that can go wrong because that's too many.
因为那实在太多了。
68:33
I feel like that's an important element.
我觉得这是很重要的一点。
68:35
This is not focusing on the downside,
这不是在关注负面因素，
68:37
focusing more on the people, the mission,
更关注人本身，关注使命，
68:40
what gets you excited, what do you think?
什么让你兴奋，你怎么看？
68:42
I do want to say one thing to all the young talents
我确实想对所有AI领域的年轻人才说几句——
68:47
in AI, the engineers, the researchers out there
那些工程师、研究员们，
68:49
because some of you apply to world labs
因为你们中有些人会申请world labs，
68:52
if you're very privileged, you consider world labs.
如果你们足够幸运，会考虑world labs。
68:56
I do find many of the young people today
我发现现在很多年轻人
69:00
think about every single aspect of an equation
会把方程式的每一个细节都琢磨透。
69:06
when they decide on jobs.
当他们决定工作的时候。
69:08
At some point, maybe that's the way they want to do it,
某种程度上，也许这就是他们想走的路，
69:12
but sometimes I do want to encourage young people
但有时候我确实想鼓励年轻人
69:15
to focus on what's important
专注于真正重要的事情，
69:18
because I find myself constantly in mentoring mode
因为我在和求职者交流时，
69:24
when I talk to job candidates,
总是不自觉地进入导师模式——
69:27
not necessarily recruiting or not recruiting,
不一定是在招人或者不招人，
69:29
but just in mentoring mode when I see an incredible young talent
就是单纯地进入导师模式，
69:34
who is over-focusing on every minute,
那些过度纠结于工作的每一个细节、维度和方面的人，或许最该问的是：你的热情在哪里？你认同这个使命吗？你相信这个团队吗？专注于你能创造的影响，以及你能与之共事的团队和那种工作方式。
69:39
dimension and aspect of considering a job
也不是考虑一份工作的某个维度或方面，
69:42
when maybe the most important thing is,
也许最重要的事情是，
69:48
where's your passion?
你的热情在哪里？
69:50
Do you align with the mission?
你是否认同这个使命？
69:52
Do you believe it have faith in this team?
你是否相信并信任这个团队？
69:56
And just focus on the impact and you can make
然后专注于你能产生的影响，
70:00
the kind of work in team you can work with.
以及你能与之共事的团队和工作类型。
70:05
Yeah, it's tough for people in the AI space.
做AI这行确实不容易。
70:07
Now there's so much at them,
现在信息量太大了，
70:09
so much new, so much happening, so much promo.
新东西太多，变化太快，宣传也铺天盖地。
70:11
That's true.
确实是这样。
70:12
I could see the stress
我能感受到那种压力，
70:13
and so I think that advice is really important
所以我觉得那个建议真的很重要——
70:14
just like what will actually make you feel fulfilled
就是真正让你在做的事情里感到满足，
70:18
in what you're doing, not just where's the fastest
而不是只盯着哪里发展最快。
70:20
growing company or who's gonna win, I don't know.
公司成长还是谁会赢，我不知道。
70:23
I want to make sure I ask you about the work
我想确保问你关于你现在在斯坦福HCI做的工作。
70:25
you're doing today at Stanford at the HCI.
HCI。
70:28
HCI.
HCI Human Centered AI Institute。
70:29
HCI Human Centered AI Institute.
你在那里做什么？
70:32
What are you doing there?
我知道这是你还在做的一件事。
70:34
I know this is the thing you do on a site still.
所以是的，HCI Human Centered AI Institute。
70:36
So yes, HCI Human Centered AI Institute
是的，HCI Human Centered AI Institute
70:41
was co-founded by me and a group of faculty
由我和一群教授共同创立，比如John H. Mendey教授、James Landey教授、Chris Manning教授，那是在2018年。当时我其实正在谷歌休最后一个学术假。对我来说，这是一个非常非常重要的决定，因为我本可以留在工业界，但在谷歌的经历让我明白了一件事。
70:44
like Professor John H. Mendey, Professor James Landey,
像Professor John H. Mendey、Professor James Landey、
70:49
Professor Chris Manning back in 2018.
Professor Chris Manning，早在2018年。
70:53
I was actually finishing my last,
我当时其实正在结束我
70:55
the last sabbatical at Google.
在Google的最后一次学术休假。
70:58
And it was a very, very important decision for me
那对我来说是一个非常、非常重要的决定，
71:03
because I could have stayed in the industry,
因为我本可以留在工业界，
71:07
but my time at Google taught me one thing
但在Google的经历让我明白了一件事。
71:10
is AI is gonna be a civilizational technology.
AI 会不会成为一项文明级别的技术。
71:14
And it don't know me how important this is to humanity
我其实不太确定这对人类有多重要，
71:19
to the point that I actually wrote a piece
重要到我2018年真的在《纽约时报》上写了一篇文章，
71:21
in New York Times that year 2018
讨论需要一套指导框架来开发和应用AI。
71:24
to talk about the need for a guiding framework
而那个框架必须扎根于人类的善意，也就是以人为中心。
71:28
to develop and to apply AI.
开发和应用AI。
71:33
And that framework has to be anchored
而这个框架必须扎根于
71:35
in human benevolence is human centeredness.
人类的善意，即以人为中心。
71:39
And I felt as Stanford, one of the world's top university
而我认为，作为世界顶尖大学之一的斯坦福，身处硅谷核心地带，孕育了从Nvidia到Google等众多重要企业，理应成为思想引领者，去构建这个以人为本的AI框架，并将其真正融入我们的研究、教育、政策以及生态建设工作中。
71:45
in the heart of Silicon Valley
在硅谷的心脏地带
71:47
that gave birth to important companies
孕育了众多重要公司
71:49
from Nvidia to Google should be a thought leader
从Nvidia到Google，应当成为思想领袖
71:57
to create this human centered AI framework
来创建这个以人为中心的AI框架
72:00
and to actually embody that
并真正将其付诸实践。
72:04
in our research, education and policy
在我们的研究、教育和政策
72:08
and ecosystem work.
以及生态系统工作中，
72:10
So I founded HCI after fast forward,
所以我创立了HCI，在fast forward之后，
72:15
after six, seven years,
过了六七年，
72:17
it has become the world's largest AI institute
它已经成为了全球最大的AI研究机构，
72:21
that does human centered research, education,
专注于以人为中心的研究、教育、
72:27
ecosystem, outreach and policy impact.
生态建设、外联以及政策影响。
72:35
It involves hundreds of faculty across all eight schools
它涵盖了斯坦福所有八个学院的数百名教授，
72:40
at Stanford from medicine to education,
从医学到教育，
72:43
to sustainability, to business, to engineering,
到可持续发展，到商科，到工程。
72:46
to humanities, to law,
到人文学科，到法律领域，
72:49
and we support researchers,
我们支持研究人员，
72:53
especially at the interdisciplinary area
尤其是在跨学科领域，
72:56
from digital economy to legal studies,
从数字经济到法学研究，
73:00
to political science, to discovery of new drugs,
到政治科学，到新药发现，
73:04
to new algorithms to, that's beyond transformers.
到超越transformer的新算法。
73:10
We also actually put a very strong focus on policy
我们实际上也非常重视政策，
73:15
because when we started HCI,
因为当我们开始HCI时，
73:18
I realized that Silicon Valley did not talk to Washington DC
我意识到硅谷没有跟华盛顿特区对话，
73:23
and or Brussels or other parts of the world
也没有跟布鲁塞尔或世界其他地区交流。
73:27
and it's given how important this technology is,
考虑到这项技术的重要性，
73:32
we need to bring everybody on board.
我们需要让所有人都参与进来。
73:35
So we created multiple programs
所以我们创建了多个项目，
73:37
from congressional bootcamp to AI index report
从国会训练营到AI指数报告，
73:43
to policy briefing,
再到政策简报，
73:46
and we especially participated in policy making,
并且我们特别参与了政策制定。
73:51
including advocating for a national AI research cloud bill
包括推动一项国家AI研究云法案，该法案在特朗普第一届政府期间通过，并参与了州级AI监管讨论。所以我们做了很多工作，而我至今仍是领导者之一，尽管在运营上参与少了很多，因为我不仅关心我们创造这项技术，更关心我们以正确的方式使用它。
73:58
that was passed in the first Trump administration
那是在第一届特朗普政府期间通过的，
74:02
and participating in state level regulatory AI discussions.
并且参与了州级的人工智能监管讨论。
74:08
So there's a lot we did and I continue
所以我们做了很多事，而我继续
74:11
to be one of the leaders,
担任领导者之一，
74:15
even though I'm much less involved operationally
尽管我在运营上参与得少了很多，
74:18
because I care not only we create this technology,
因为我不仅关心我们创造这项技术，
74:22
but we use it in the right way.
但我们是以正确的方式使用它。
74:24
Wow, I was not aware of all that.
哇，这些我之前完全不知道。
74:26
Other work you were doing.
你做的其他那些工作。
74:27
As you're talking, I was reminded Charlie Munger
听你这么说，我想起查理·芒格
74:30
had this quote, take a simple idea
有句话，叫“把一个简单的想法
74:33
and take it very seriously.
认真对待到极致”。
74:35
I feel like you've done that in so many different ways
我觉得你在很多方面都做到了这一点，
74:37
and stayed with it and it's unbelievable the impact
并且一直坚持下来，这么多年
74:41
that you've had in so many ways over the years.
在那么多领域产生的影响简直不可思议。
74:44
I'm gonna skip the lightning round
我跳过快速问答环节，直接问你最后一个问题。你还有什么想分享的，或者有什么想留给听众的吗？我对AI Lenny特别兴奋。我想回答一个我环游世界时，每个人都会问我的问题：如果我是个音乐人，
74:46
and I'm just gonna ask you one last question.
最后我再问你一个问题。
74:47
Is there anything else that you wanted to share
你还有什么想分享的，
74:49
or anything else you wanna leave the listeners with?
或者有什么想留给听众的吗？
74:52
I'm very excited by AI Lenny.
我对AI Lenny感到非常兴奋。
74:55
I wanna answer one question that I,
我想回答一个我——
74:58
when I travel around the world,
当我环游世界时，
75:01
everybody asks me is that if I'm a musician,
每个人都会问我的问题：如果我是一名音乐家，
75:05
if I'm a teacher, middle school teacher,
如果我是一个老师，中学老师，
75:09
if I'm a nurse, if I'm a content, if I'm a farmer,
如果我是一个护士，如果我是一个内容创作者，如果我是一个农民，
75:14
do I have a role in AI?
我在AI里能扮演什么角色？
75:16
Or is AI just gonna take over my life or my work?
还是说AI只会接管我的生活或工作？
75:21
And I think this is the most important question of AI.
我认为这是AI最重要的问题。
75:26
And I find that in Silicon Valley,
而且我发现，在硅谷，
75:29
we tend not to speak hard to hard with people,
我们往往不太愿意和那些跟我们不一样的人——
75:34
with people like us and not like us in Silicon Valley,
无论是像我们一样在硅谷的人，还是不像我们的人——坦诚交流。
75:38
but like all of us, we tend to just toss around words
但就像我们所有人一样，我们总爱随口抛出一些词，比如无限生产力、无限闲暇时间，或者无限权力之类的。但归根结底，AI 是关于人的。当有人问我那个问题时，我的回答是响亮的“是”——每个人在 AI 中都有角色。这取决于你做什么、你想要什么，没有任何技术应该剥夺人的尊严。
75:42
like infinite productivity or infinite leisure time
比如无限的生产力，或者无限的闲暇时间，又或者无限的权力之类的。但归根结底，AI 是关于人的。当有人问我这个问题时，答案是响亮的“是的”，每个人在 AI 中都有自己的角色。这取决于你做什么、想要什么，没有任何技术应该剥夺人的尊严。事实上，拥抱 Marvel。
75:47
or infinite power or whatever.
或者无限算力之类的。
75:53
But at the end of the day, AI is about people
但归根结底，AI关乎人。
75:56
and when people ask me that question,
当有人问我这个问题时，
75:59
it's a resounding yes, everybody has a role in AI.
答案是响亮的“是”，每个人在AI中都有角色。
76:03
It depends on what you do and what you want,
这取决于你做什么、想要什么，
76:07
a no technology should take away human dignity.
没有任何技术应该剥夺人的尊严。
76:11
And the human dignity and agency should be at the heart
人类尊严与自主权应当成为每一项技术开发、部署以及治理的核心。
76:15
of the development, the deployment,
所以，如果你是一位年轻的艺术家，
76:19
as well as the governance of every technology.
而你的热情在于讲故事，
76:22
So if you are a young artist
那就把AI当作工具来拥抱吧。
76:27
and your passion is storytelling,
事实上，拥抱Marvel，
76:31
embrace AI as a tool.
我希望它能成为你的工具。
76:32
In fact, embrace Marvel,
事实上，拥抱Marvel，
76:34
I hope it becomes a tool for you,
我希望它能成为你的工具，
76:38
because the way you tell your story
因为你的故事讲述方式独一无二，而世界仍然需要它。
76:41
is unique and the world still needs it.
但你要如何讲述你的故事？
76:44
But how you tell your story?
如何用最不可思议的工具，以最独特的方式讲述你的故事，这很重要。
76:46
How do you use the most incredible tool
而那个声音需要被听见。
76:50
to tell your story in the most unique way is important?
如果你是一位即将退休的农民，AI依然与你有关，因为你是一位公民。
76:54
And that voice needs to be heard.
而那个声音需要被听见。
76:57
If you are a farmer near retirement,
如果你是一位临近退休的农民，
77:01
AI still matters because you're a citizen.
AI依然重要，因为你是一位公民。
77:05
You can participate in your community.
你可以参与到你所在的社区中。
77:07
You should have a voice in how AI is used,
你应该对AI如何被使用、如何被应用有发言权。
77:11
how AI is applied.
你可以和你认识的人一起合作，鼓励大家使用AI
77:13
You work with people that you can, you know,
来让生活变得更轻松。
77:17
encourage all of you to use AI
如果你是一名护士，
77:22
to make life easier for you.
我希望你知道，至少在我的职业生涯中——
77:25
If you're a nurse,
如果你是一位护士，
77:26
I hope you know that at least in my career,
我希望你知道，至少在我的职业生涯中，
77:32
I have worked so much in healthcare research
我在医疗健康研究领域投入了大量工作，
77:34
because I feel our healthcare workers
因为我深感我们的医护人员
77:37
should be greatly augmented and helped by AI technology,
应当得到AI技术的极大增强与辅助，
77:43
whether it's smart cameras to feed more information
无论是通过智能摄像头提供更多信息，
77:48
or robotic assistance because our nurses are overworked,
还是借助机器人辅助——毕竟我们的护士工作超负荷、
77:53
over fatigued, and as our society ages,
极度疲惫，而随着社会老龄化，
77:57
we need more help for people to be taken care of.
我们需要更多帮助来照料他人。
78:01
So AI can play that role.
因此，AI可以扮演这个角色。
78:03
So I just want to say that it's so important
我只想说，这一点非常重要——即使像我这样的技术人员，也真心认为每个人在AI中都有自己的角色。
78:06
that even a technologist like me
多么美好的收尾啊。
78:11
are sincere about that everybody has a role in AI.
这正好呼应了我们开头的话题——
78:17
What a beautiful way to end it.
一切都取决于我们自己，
78:18
Such a tie back to where we started
以及我们每个人对AI将如何影响我们生活所承担的责任。
78:20
about how it's up to us
这件事取决于我们，
78:21
and taking individual responsibility
以及每个人都要承担起自己的责任。
78:24
for what AI will do in our lives.
关于AI将如何影响我们的生活，
78:27
Final question, we can folks find marble,
最后一个问题，大家能找到Marble吗？我们可能会试着加入World Labs，如果他们愿意的话。网站是什么？大家该去哪里？嗯，World Labs的网站是www.worldlabs.ai，你可以在那里看到我们的研究进展。我们有技术博客。你也能在那里找到Marble这个产品。
78:29
we're going to go maybe try to join world labs
我们可能会尝试加入World Labs，
78:32
if they want to.
如果他们愿意的话。
78:33
What's the website, where do people go?
网站是什么？人们去哪里找？
78:34
Well, world labs website is www.worldlabs.ai
World Labs的网站是www.worldlabs.ai，
78:41
and you can find our research progress there.
你可以在那里看到我们的研究进展。
78:45
We have technical blogs.
我们有技术博客。
78:48
You can find marble the product there.
你也能在那里找到Marble这个产品。
78:50
You can sign in there.
你可以在这里登录。
78:51
You can find our job posts a link there.
我们的招聘信息链接也在那里。
78:55
You can, you know, we're in San Francisco.
我们就在旧金山。
78:58
We love to work with the world's best talents.
我们很乐意与全球最优秀的人才合作。
79:02
Amazing, buffet, thank you so much for being here.
太棒了，自助餐，非常感谢你来到这里。
79:04
Thank you, Lenny.
谢谢你，Lenny。
79:06
Bye everyone.
大家再见。
79:09
Thank you so much for listening.
非常感谢你的收听。
79:11
If you found this valuable,
如果你觉得这期内容有用，
79:12
you can subscribe to the show on Apple podcasts, Spotify,
可以在 Apple Podcasts、Spotify
79:15
or your favorite podcast app.
或者你常用的播客应用上订阅我们的节目。
79:17
Also, please consider giving us a rating
另外，也欢迎给我们打个分
79:19
or leaving a review, as that really helps other listeners
或者写个评论，这真的能帮其他听众
79:22
find the podcast.
找到这个播客。
79:23
You can find all past episodes or learn more about the show
你可以在 Lenny'spodcast.com 找到所有往期节目，
79:26
at Lenny'spodcast.com.
或者了解更多关于本播客的信息。
79:29
See you in the next episode.
下集见。