World Labs' Fei-Fei Li on Creating Large World Models

李飞飞在 Bloomberg Live 上讲述 World Labs 的核心命题:从 LLM 转向空间智能(Spatial Intelligence),用 Large World Model 让机器真正理解三维物理世界。

来源:Bloomberg Live · 查看原始内容

为什么值得看

这是李飞飞在 World Labs 创立后少有的系统访谈。她明确划清了 World Labs 与 LLM 路线的边界:500 万年的演化证明,智能的起点是"看见并在物理世界中移动",而不是语言。

几个关键论点

  • 空间智能 ≠ LLM:LLM 处理的是 token 序列,世界模型处理的是三维时空中可交互的物理对象。
  • 数据飞轮:互联网有海量文本(喂出了 LLM),但没有等量的"带语义的三维世界数据"。World Labs 的护城河之一就是构建这样的数据管道。
  • 生成与判别统一:同一个世界模型既能"想象"未见过的场景,也能"理解"已存在的场景。
  • 物理 AI 的入口:从机器人、AR/VR、内容创作到自动驾驶,所有需要"理解空间"的赛道都会被这条路线重塑。

我看完后的几个 take

  1. 当所有人都在卷 token,李飞飞押的是"感知 + 行动"这条更古老、但也更基础的路。
  2. 视觉-空间数据是新的瓶颈,谁拥有它谁就赢一半。这跟当年 ImageNet 的逻辑是一致的。
  3. 中国这边讲"具身智能",硅谷这边讲"Spatial Intelligence",本质都在回答同一个问题:机器怎么真正理解世界
  4. 对个人而言:未来的机会不在调 token,而在理解物理世界的数据闭环
查看完整字幕 · 448 条

00:00:03Everyone is focused on LLM's, chat GPT,

00:00:06Claude, large language models, but you

00:00:07have raised a billion dollars to build

00:00:12something different. Large world models.

00:00:16Make the case for us. What is the bet

00:00:18you're making that others aren't?

00:00:22>> Right. So, um this is my uh co-founded

00:00:26startup, World Labs, and uh we are

00:00:29uh all in in spatial intelligence. And

00:00:32uh the means to spatial intelligence is

00:00:34building a large world model. So, what

00:00:37is the case for us? The case for us is a

00:00:41500 million year story. Is that animal

00:00:44intelligence starts with seeing and

00:00:48moving in a physical world. That uh

00:00:52evolution began with us as animals

00:00:54knowing what the world is, knows knowing

00:00:57who we are, knowing how to move around

00:01:01it, interact with it, and uh much of

00:01:04life, human life, human work life, human

00:01:08private life, has a lot to do with uh

00:01:11perceiving, understanding, reasoning,

00:01:14interaction with the world, including

00:01:19imaginary world of creativity, of uh of

00:01:21uh productivity, uh

00:01:25as virtual worlds. So, unlocking that

00:01:28capability in machines, unlocking the

00:01:31capability of generating

00:01:34from any 3D, 4D worlds, unlocking the

00:01:37capability of reasoning within any

00:01:41world, unlocking the capability of uh

00:01:44teaching agents or robots or or

00:01:46assisting humans to interact with the

00:01:49world is what spatial intelligence is

00:01:52about, and that's what we are focusing

00:01:52about, and that's what we are focusing on.

00:01:52on.

00:01:54>> So, what can world models do ultimately

00:02:02Can words put out fires?

00:02:05Can words cook an omelet?

00:02:08I think there's so much, right? So, we

00:02:10for example, creativity.

00:02:14People design. People

00:02:16whether we're designing interior space,

00:02:18we're designing

00:02:20machines, we're designing we're

00:02:22designing homes, we're designing

00:02:25stories. So much of that is beyond

00:02:25stories. So much of that is beyond words.

00:02:26words.

00:02:30We also use agents. Whether we use

00:02:33agents in virtual world, whether it's

00:02:36for entertainment like gaming, or for

00:02:39more serious industrial

00:02:41industrial applications, whether it's

00:02:43digital twin

00:02:47design or or inspection or or

00:02:49or what kind of

00:02:52many kind of optimization tasks. Or we

00:02:56build robots and to help us to do a lot

00:02:58of things from

00:03:02putting out fire to helping health care

00:03:06scenarios to manufacturing. All these

00:03:08are application downstream applications

00:03:12of unlocking spatial intelligence and

00:03:13building world models.

00:03:14>> So, what's the what do you think the

00:03:16chat GPT moment for world models will

00:03:18be? Like how will we know this has

00:03:18be? Like how will we know this has arrived?

00:03:20arrived?

00:03:21>> Yeah, that's a great question, Emily,

00:03:25because chat is such a consumer behavior

00:03:29that chat GPT moment tends to be used to

00:03:32describe a viral

00:03:36public consumer moment of getting so

00:03:40close to what AI can do. In a in a world

00:03:43of world models,

00:03:43of world models, um

00:03:43um

00:03:45the kind of spatial intelligence we're

00:03:51I'm still trying to figure out if there

00:03:54is a corresponding consumer moment

00:03:57because the kind of applications we are

00:03:59talking about

00:03:59talking about um

00:03:59um

00:04:02tends to be first going to the

00:04:04professionals, professional creators,

00:04:06professional designers, professional

00:04:09developers, uh professional researchers

00:04:11and engineers who will use it for

00:04:14robotics and industrial design and all

00:04:18that. So, maybe we will not

00:04:21necessarily have a consumer moment. But

00:04:23maybe we will. And you know, I I would

00:04:26love to design my home in a much easier

00:04:29way and just change the color of the

00:04:31curtain, you know, with a click.

00:04:33>> All right, that sounds pretty cool. So,

00:04:35in the last 6 months, Yann LeCun left

00:04:37Meta to work on world models, Google

00:04:39shipped Project Genie, Nvidia has its

00:04:42own world models, Cosmos. Nvidia's also

00:04:44one of your investors.

00:04:46What do you have that they don't? And

00:04:49which competitors out there worry you

00:04:50the most?

00:04:53>> Yeah, so first of all, we started World

00:04:57Labs in 2024. I still remember when when

00:04:58we were

00:05:00out talking about world models and

00:05:02spatial intelligence,

00:05:05it was just a year after ChatGPT. People

00:05:07were still totally talking about LLMs.

00:05:10So, we we really had a head start and

00:05:12understanding that this is going to be

00:05:15the next frontier of AI. I'm very

00:05:18excited by that. So, what do they have

00:05:20we don't? Well, first of all, I think we

00:05:22have an incredible team. We have the

00:05:22have an incredible team. We have the conviction.

00:05:23conviction.

00:05:24>> They don't have the godmother, that's

00:05:25for sure.

00:05:29>> Um but but the the world is big and and

00:05:32I think this is just like LLMs. I think

00:05:34there will be many companies doing

00:05:36incredible work in world models. Just as

00:05:4024 hours ago, uh I we kind of got fed up

00:05:44that the word world model has been so uh

00:05:47confusing and being used so in so many

00:05:49different ways that we actually put out

00:05:49different ways that we actually put out a

00:05:50a

00:05:55blog just explaining what a functional

00:05:59taxonomy of world model is instead of

00:06:02mushing everything together. And the way

00:06:04I see it is right now there are three

00:06:04I see it is right now there are three ways

00:06:05ways

00:06:08of calling world models when it comes to

00:06:11spatial intelligence. One is what I call

00:06:13a renderer when the model puts beautiful

00:06:16pixels on the screen. Mostly like video

00:06:18generation model. And the consumer is

00:06:22mostly human eyeballs. And while the

00:06:24model commits to beautiful pixels on the

00:06:28screen, it doesn't necessarily commit to

00:06:30physics and dynamics and geo geometric

00:06:30physics and dynamics and geo geometric correctness

00:06:32correctness

00:06:34because that's for just

00:06:36consuming human eyeball consuming not

00:06:39necessarily for computation and and

00:06:42other other tasks. Then another kind of

00:06:45world model is what we call a

00:06:48a planner. That is more for machines,

00:06:52more for robots where it outputs

00:06:54whatever the input is the state of the

00:06:57world or the action. It outputs a

00:07:00correct action to take to the next step.

00:07:02And you see that kind of world model a

00:07:05lot for robotics applications and you

00:07:07hear that in that context. The third

00:07:10kind which I think is the lynchpin of

00:07:13the the three is a simulator. Is that it

00:07:16actually is consumed by humans as well

00:07:19as machines is trying to respect the

00:07:21structure, the physics, and the dynamics

00:07:25of the world and really simulate the 3D

00:07:29and 4D information of the world as well

00:07:32as well as the semantic information. And

00:07:34a simulator could become a renderer. The

00:07:37simulator could become a planner. But

00:07:39this layer is

00:07:43a huge critical path in my opinion, to

00:07:45unlock spatial intelligence. And that's

00:07:48what World Lab is working on.

00:07:50>> All of this rolls up into robotics, so I

00:07:51want to get your take on the field and

00:07:54humanoids in particular. Funding for

00:07:57humanoids hit $6 billion, but, you know,

00:07:59they still can't load my dishwasher as

00:08:00fast as I can. They still can't go get

00:08:02my Amazon packages.

00:08:04Will world

00:08:06models, World Labs, close the gap

00:08:07between hype

00:08:09and reality?

00:08:11>> That's a loaded question, Emily. First

00:08:12of all,

00:08:14>> That is my job.

00:08:17>> I get it. First of all, robotics is

00:08:19going to be one of the most important

00:08:22revolution in human industrialization.

00:08:23$6 billion

00:08:26is too small. Right? If you look at

00:08:28self-driving cars investment, if you

00:08:31look at language models investment, it

00:08:33took way more than $6 billion.

00:08:41I think it will take time

00:08:44to invest, and it will also

00:08:47hopefully not take the hype, but take

00:08:50the thoughtfulness to invest in the

00:08:52right effort. And, for example,

00:08:54unlocking world modeling and spatial

00:08:57intelligence and simulation layer, all

00:09:00this is part of that that

00:09:04important effort. Um

00:09:06are we going to close the gap? I do

00:09:09believe World Labs is working on one of

00:09:13the most critical technology in spatial

00:09:16physical intelligence. And obviously,

00:09:19that's the that's the hope.

00:09:22>> You've been more measured on AI safety,

00:09:24skeptical of the doom narrative, but

00:09:28also of heavy-handed regulation.

00:09:31When you look across the industry, where

00:09:33do you feel real safety work versus

00:09:33do you feel real safety work versus safety

00:09:34safety

00:09:35safety theater?

00:09:36theater?

00:09:39Is anyone getting it right?

00:09:41>> So, in general, I've been just more

00:09:44measured on every every rhetoric makes

00:09:47me very boring, to be honest.

00:09:47me very boring, to be honest. Um

00:09:48Um

00:09:50I think there's just so much hype. There

00:09:53is so much hype. Um

00:09:55obviously, we need to build the right

00:09:57technology. We need to guardrail the

00:09:59technology. Whether you use the word

00:10:02responsible, you use the word safety,

00:10:04you use the word

00:10:08um um trustworthy, um building the right

00:10:12technology and product so that it can

00:10:16empower, enhance, augment humanity, and

00:10:20not harm them is the goal of any any

00:10:22work we do, whether it's AI or not. So,

00:10:25where is it doing right? I really hope

00:10:27every company, every

00:10:30um every product that's being built,

00:10:32that the people behind it are being

00:10:35mindful of that, and are thinking about,

00:10:37you know, what data are we using? What

00:10:40system are we building? What evaluations

00:10:43are we conducting? What guardrails are

00:10:45we putting in? How do we communicate

00:10:47with our with our users and customers?

00:10:50How do we work with regulators so that

00:10:53when the rubber hits the road, that we

00:10:56are um you know, being responsible. I do

00:10:59believe a lot of this work is happening.

00:11:01It's not happening in a theater, to be

00:11:03honest. For example, so many

00:11:06pharmaceutical and health care um

00:11:09industry uh companies are incorporating

00:11:09industry uh companies are incorporating AI.

00:11:10AI.

00:11:13Literally, I just came from the hospital

00:11:16to come to your to to your panel because

00:11:18I have a family member uh about to get a

00:11:22surgery in in the next 1 hour or so. And

00:11:22surgery in in the next 1 hour or so. And I

00:11:22I

00:11:25I was just in Stanford Hospital looking

00:11:27at where AI is already being used and

00:11:30where AI could be used. And it's already

00:11:34happening. Doctors are using AI to to to

00:11:37help them with charting. Radiologists

00:11:40are using AI to assist them reading the

00:11:43the MRI and the CT scans. I do hope that

00:11:46we have more AI to help our nurses, to

00:11:49help family members. I got this long

00:11:51radiology report last night and the

00:11:54first thing I did is send it to a AI so

00:11:56that it can help me to explain it. So,

00:11:59all this is happening. Um safety

00:12:02measures are happening. Um but there

00:12:05needs to be more in a right way, in a in

00:12:08a scientifically grounded way. Um and

00:12:10that's the conversation that should be

00:12:12taking place instead of what you say the

00:12:12taking place instead of what you say the theater.

00:12:13theater.

00:12:14>> Well, thank you for coming and I hope

00:12:17your person is okay. We all we all do.

00:12:17your person is okay. We all we all do. Um

00:12:19Um

00:12:21the backlash is we all it's being called

00:12:23the AI hate wave. I'm sure you've seen

00:12:25the video of former Google CEO Eric

00:12:27Schmidt getting booed at a college

00:12:30graduation. You spend a lot of time with

00:12:33students. What are they saying? And if

00:12:35they're scared,

00:12:38are the fears justified?

00:12:39>> Yeah, I do spend a lot of time with

00:12:41students. Uh

00:12:42to be fair, my students are pretty

00:12:44privileged cuz they're Stanford

00:12:46students. I think it's

00:12:48I think it's even more important and I

00:12:51try to do it myself that we spend time

00:12:53with our teachers, with our nurses, with

00:12:56our our parents, grandparents. And

00:12:59that's actually something I try to do. I

00:13:02try to talk to K-12 educators. I try to

00:13:05go to places and talk to people where

00:13:07they feel that they're not part of the

00:13:10conversation. And

00:13:12even Stanford students reflect some of

00:13:16this mixed sentiment. There is anxiety.

00:13:19There's sense of hope. There is also

00:13:22excitement. There is also confusion.

00:13:24There's also um

00:13:27simultaneously a sense of

00:13:30dignity and agency when AI can help me

00:13:32do things that I couldn't do before and

00:13:35a sense of loss of dignity and agency if

00:13:39AI is is going to take my job. So, I

00:13:39AI is is going to take my job. So, I think

00:13:40think

00:13:42I think the sentiment is mixed and I

00:13:44really want to point out a lot of this

00:13:48sentiment happens when there's a vacuum

00:13:49of thoughtful

00:13:53public discourse. Right now, the oxygen,

00:13:56the air is all sucked into the polarized

00:14:00extreme of doomerism or total utopian.

00:14:03And when hype takes all the oxygen in

00:14:04the room,

00:14:08that void brews the kind of anxiety and

00:14:11it's actually that void we really need

00:14:13to care about because that's where real

00:14:15people live. That's where real people

00:14:17are seeking answers.

00:14:19And I think it's

00:14:19And I think it's uh

00:14:20uh

00:14:23As a scientist and a educator and a

00:14:23As a scientist and a educator and a entrepreneur,

00:14:25entrepreneur,

00:14:29I'm on ground zero with students, with

00:14:29I'm on ground zero with students, with educators,

00:14:30educators,

00:14:33with entrepreneurs

00:14:35and I really do believe it's is one of

00:14:37my responsibility

00:14:39to not hype

00:14:43and try to speak with with both science

00:14:45and humility and

00:14:47and in inspire people to to recognize

00:14:50this is a technology that can truly

00:14:54empower a lot of our work and life,

00:14:57can truly help us, you know, have a

00:15:00better health care system, have better

00:15:03scientific discovery, have better uh

00:15:06uh better environment, better education

00:15:08if we do the right thing.

00:15:08if we do the right thing. >> Mhm.

00:15:09>> Mhm.

00:15:11We're both moms. We both have young

00:15:11We're both moms. We both have young teenagers.

00:15:13teenagers.

00:15:15How do you think AI will change learning

00:15:17in the college experience?

00:15:20>> AI must change learning. AI must change

00:15:24K to 16 learning. I think this is one of

00:15:27the biggest opportunity for humanity in

00:15:30the next decade to come is that

00:15:30the next decade to come is that what

00:15:36the most precious resource of our entire

00:15:39world is human capital.

00:15:43And when we have gotten a technology

00:15:45that can answer standardized tests,

00:15:49whether it's it's a common core kind of

00:15:50test all the way to

00:15:54international Olympiad math exams, when

00:15:57AI can do better than average human,

00:16:01it's not about humans are bad. It's

00:16:03about we need to change the education

00:16:05system. We need to change how we

00:16:08evaluate. We need to change the way we

00:16:12empower teachers to teach to to educate

00:16:15the next generation of students where

00:16:18they can use these tools, be empowered,

00:16:21and do things that we can never imagine.

00:16:23>> So, do you think our kids will still

00:16:23>> So, do you think our kids will still learn?

00:16:24learn?

00:16:26>> Absolutely. If we teach them right, if

00:16:29the society prepares them right, they

00:16:31should not be all of the kids today

00:16:33should not be scared of AI. They should

00:16:37feel the human agency to to lead AI, to

00:16:40use AI in the right way, and to use AI

00:16:42to make the right to make the impact

00:16:46that they want to make for the world.

00:16:49>> Anthropic CEO Dario Amodei has suggested

00:16:51AGI is 2 to 3 years out. We'll get there

00:16:53by scaling the current paradigm. Demis

00:16:56Hassabis says we're at the foothills of

00:16:58the singularity.

00:16:59You've said you don't even engage with

00:17:01the term

00:17:01the term AGI.

00:17:02AGI.

00:17:05Are they wrong, or is the disagreement

00:17:08about what we're calling the goal?

00:17:11>> I don't engage with the term AGI because

00:17:13the founding fathers of artificial

00:17:16intelligence as a scientific field had

00:17:20this dream of thinking and doing

00:17:23machines and that is a scientific quest

00:17:25and that quest has been my lifelong

00:17:29career and I'm still on that quest. Now

00:17:31I'm combining that scientific quest with

00:17:33making products that can make people's

00:17:36life better and that is the field called

00:17:39artificial intelligence and

00:17:39artificial intelligence and um

00:17:42um

00:17:44I'm okay people call it whatever they

00:17:46want they can call it

00:17:48an Apple that's fine.

00:17:48an Apple that's fine. >> [laughter]

00:17:49>> [laughter]

00:17:49>> [laughter] >> Um

00:17:50>> Um

00:17:51I'm focusing on

00:17:56building a technology can that can truly

00:17:58that can truly make a difference in

00:18:01people's lives and at work.

00:18:03>> What's the one thing you'll have shipped

00:18:04this year that we'll be talking about

00:18:06next year?

00:18:09>> I hope that we will be shipping a model

00:18:12for spatial intelligence

00:18:14that will

00:18:14that will inspire

00:18:16inspire

00:18:18incredibly exciting product

00:18:20opportunities that people haven't seen

00:18:20opportunities that people haven't seen before.

Discussion

讨论

正在载入讨论...

Aiden Novak

邮箱登录

验证码登录后可评论、收藏和投稿;管理员邮箱会直接进入管理能力。